2019-05-27 06:55:01 +00:00
|
|
|
// SPDX-License-Identifier: GPL-2.0-or-later
|
2020-01-23 13:01:33 +00:00
|
|
|
/* Processing of received RxRPC packets
|
2007-04-26 22:48:28 +00:00
|
|
|
*
|
2020-01-23 13:01:33 +00:00
|
|
|
* Copyright (C) 2020 Red Hat, Inc. All Rights Reserved.
|
2007-04-26 22:48:28 +00:00
|
|
|
* Written by David Howells (dhowells@redhat.com)
|
|
|
|
|
*/
|
|
|
|
|
|
2016-06-02 19:08:52 +00:00
|
|
|
#define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
|
|
|
|
|
|
2007-04-26 22:48:28 +00:00
|
|
|
#include "ar-internal.h"
|
|
|
|
|
|
2024-05-03 15:07:40 +00:00
|
|
|
/* Override priority when generating ACKs for received DATA */
|
|
|
|
|
static const u8 rxrpc_ack_priority[RXRPC_ACK__INVALID] = {
|
|
|
|
|
[RXRPC_ACK_IDLE] = 1,
|
|
|
|
|
[RXRPC_ACK_DELAY] = 2,
|
|
|
|
|
[RXRPC_ACK_REQUESTED] = 3,
|
|
|
|
|
[RXRPC_ACK_DUPLICATE] = 4,
|
|
|
|
|
[RXRPC_ACK_EXCEEDS_WINDOW] = 5,
|
|
|
|
|
[RXRPC_ACK_NOSPACE] = 6,
|
|
|
|
|
[RXRPC_ACK_OUT_OF_SEQUENCE] = 7,
|
|
|
|
|
};
|
|
|
|
|
|
2022-10-06 20:45:42 +00:00
|
|
|
static void rxrpc_proto_abort(struct rxrpc_call *call, rxrpc_seq_t seq,
|
|
|
|
|
enum rxrpc_abort_reason why)
|
rxrpc: Rewrite the data and ack handling code
Rewrite the data and ack handling code such that:
(1) Parsing of received ACK and ABORT packets and the distribution and the
filing of DATA packets happens entirely within the data_ready context
called from the UDP socket. This allows us to process and discard ACK
and ABORT packets much more quickly (they're no longer stashed on a
queue for a background thread to process).
(2) We avoid calling skb_clone(), pskb_pull() and pskb_trim(). We instead
keep track of the offset and length of the content of each packet in
the sk_buff metadata. This means we don't do any allocation in the
receive path.
(3) Jumbo DATA packet parsing is now done in data_ready context. Rather
than cloning the packet once for each subpacket and pulling/trimming
it, we file the packet multiple times with an annotation for each
indicating which subpacket is there. From that we can directly
calculate the offset and length.
(4) A call's receive queue can be accessed without taking locks (memory
barriers do have to be used, though).
(5) Incoming calls are set up from preallocated resources and immediately
made live. They can than have packets queued upon them and ACKs
generated. If insufficient resources exist, DATA packet #1 is given a
BUSY reply and other DATA packets are discarded).
(6) sk_buffs no longer take a ref on their parent call.
To make this work, the following changes are made:
(1) Each call's receive buffer is now a circular buffer of sk_buff
pointers (rxtx_buffer) rather than a number of sk_buff_heads spread
between the call and the socket. This permits each sk_buff to be in
the buffer multiple times. The receive buffer is reused for the
transmit buffer.
(2) A circular buffer of annotations (rxtx_annotations) is kept parallel
to the data buffer. Transmission phase annotations indicate whether a
buffered packet has been ACK'd or not and whether it needs
retransmission.
Receive phase annotations indicate whether a slot holds a whole packet
or a jumbo subpacket and, if the latter, which subpacket. They also
note whether the packet has been decrypted in place.
(3) DATA packet window tracking is much simplified. Each phase has just
two numbers representing the window (rx_hard_ack/rx_top and
tx_hard_ack/tx_top).
The hard_ack number is the sequence number before base of the window,
representing the last packet the other side says it has consumed.
hard_ack starts from 0 and the first packet is sequence number 1.
The top number is the sequence number of the highest-numbered packet
residing in the buffer. Packets between hard_ack+1 and top are
soft-ACK'd to indicate they've been received, but not yet consumed.
Four macros, before(), before_eq(), after() and after_eq() are added
to compare sequence numbers within the window. This allows for the
top of the window to wrap when the hard-ack sequence number gets close
to the limit.
Two flags, RXRPC_CALL_RX_LAST and RXRPC_CALL_TX_LAST, are added also
to indicate when rx_top and tx_top point at the packets with the
LAST_PACKET bit set, indicating the end of the phase.
(4) Calls are queued on the socket 'receive queue' rather than packets.
This means that we don't need have to invent dummy packets to queue to
indicate abnormal/terminal states and we don't have to keep metadata
packets (such as ABORTs) around
(5) The offset and length of a (sub)packet's content are now passed to
the verify_packet security op. This is currently expected to decrypt
the packet in place and validate it.
However, there's now nowhere to store the revised offset and length of
the actual data within the decrypted blob (there may be a header and
padding to skip) because an sk_buff may represent multiple packets, so
a locate_data security op is added to retrieve these details from the
sk_buff content when needed.
(6) recvmsg() now has to handle jumbo subpackets, where each subpacket is
individually secured and needs to be individually decrypted. The code
to do this is broken out into rxrpc_recvmsg_data() and shared with the
kernel API. It now iterates over the call's receive buffer rather
than walking the socket receive queue.
Additional changes:
(1) The timers are condensed to a single timer that is set for the soonest
of three timeouts (delayed ACK generation, DATA retransmission and
call lifespan).
(2) Transmission of ACK and ABORT packets is effected immediately from
process-context socket ops/kernel API calls that cause them instead of
them being punted off to a background work item. The data_ready
handler still has to defer to the background, though.
(3) A shutdown op is added to the AF_RXRPC socket so that the AFS
filesystem can shut down the socket and flush its own work items
before closing the socket to deal with any in-progress service calls.
Future additional changes that will need to be considered:
(1) Make sure that a call doesn't hog the front of the queue by receiving
data from the network as fast as userspace is consuming it to the
exclusion of other calls.
(2) Transmit delayed ACKs from within recvmsg() when we've consumed
sufficiently more packets to avoid the background work item needing to
run.
Signed-off-by: David Howells <dhowells@redhat.com>
2016-09-08 10:10:12 +00:00
|
|
|
{
|
2022-10-06 20:45:42 +00:00
|
|
|
rxrpc_abort_call(call, seq, RX_PROTOCOL_ERROR, -EBADMSG, why);
|
rxrpc: Rewrite the data and ack handling code
Rewrite the data and ack handling code such that:
(1) Parsing of received ACK and ABORT packets and the distribution and the
filing of DATA packets happens entirely within the data_ready context
called from the UDP socket. This allows us to process and discard ACK
and ABORT packets much more quickly (they're no longer stashed on a
queue for a background thread to process).
(2) We avoid calling skb_clone(), pskb_pull() and pskb_trim(). We instead
keep track of the offset and length of the content of each packet in
the sk_buff metadata. This means we don't do any allocation in the
receive path.
(3) Jumbo DATA packet parsing is now done in data_ready context. Rather
than cloning the packet once for each subpacket and pulling/trimming
it, we file the packet multiple times with an annotation for each
indicating which subpacket is there. From that we can directly
calculate the offset and length.
(4) A call's receive queue can be accessed without taking locks (memory
barriers do have to be used, though).
(5) Incoming calls are set up from preallocated resources and immediately
made live. They can than have packets queued upon them and ACKs
generated. If insufficient resources exist, DATA packet #1 is given a
BUSY reply and other DATA packets are discarded).
(6) sk_buffs no longer take a ref on their parent call.
To make this work, the following changes are made:
(1) Each call's receive buffer is now a circular buffer of sk_buff
pointers (rxtx_buffer) rather than a number of sk_buff_heads spread
between the call and the socket. This permits each sk_buff to be in
the buffer multiple times. The receive buffer is reused for the
transmit buffer.
(2) A circular buffer of annotations (rxtx_annotations) is kept parallel
to the data buffer. Transmission phase annotations indicate whether a
buffered packet has been ACK'd or not and whether it needs
retransmission.
Receive phase annotations indicate whether a slot holds a whole packet
or a jumbo subpacket and, if the latter, which subpacket. They also
note whether the packet has been decrypted in place.
(3) DATA packet window tracking is much simplified. Each phase has just
two numbers representing the window (rx_hard_ack/rx_top and
tx_hard_ack/tx_top).
The hard_ack number is the sequence number before base of the window,
representing the last packet the other side says it has consumed.
hard_ack starts from 0 and the first packet is sequence number 1.
The top number is the sequence number of the highest-numbered packet
residing in the buffer. Packets between hard_ack+1 and top are
soft-ACK'd to indicate they've been received, but not yet consumed.
Four macros, before(), before_eq(), after() and after_eq() are added
to compare sequence numbers within the window. This allows for the
top of the window to wrap when the hard-ack sequence number gets close
to the limit.
Two flags, RXRPC_CALL_RX_LAST and RXRPC_CALL_TX_LAST, are added also
to indicate when rx_top and tx_top point at the packets with the
LAST_PACKET bit set, indicating the end of the phase.
(4) Calls are queued on the socket 'receive queue' rather than packets.
This means that we don't need have to invent dummy packets to queue to
indicate abnormal/terminal states and we don't have to keep metadata
packets (such as ABORTs) around
(5) The offset and length of a (sub)packet's content are now passed to
the verify_packet security op. This is currently expected to decrypt
the packet in place and validate it.
However, there's now nowhere to store the revised offset and length of
the actual data within the decrypted blob (there may be a header and
padding to skip) because an sk_buff may represent multiple packets, so
a locate_data security op is added to retrieve these details from the
sk_buff content when needed.
(6) recvmsg() now has to handle jumbo subpackets, where each subpacket is
individually secured and needs to be individually decrypted. The code
to do this is broken out into rxrpc_recvmsg_data() and shared with the
kernel API. It now iterates over the call's receive buffer rather
than walking the socket receive queue.
Additional changes:
(1) The timers are condensed to a single timer that is set for the soonest
of three timeouts (delayed ACK generation, DATA retransmission and
call lifespan).
(2) Transmission of ACK and ABORT packets is effected immediately from
process-context socket ops/kernel API calls that cause them instead of
them being punted off to a background work item. The data_ready
handler still has to defer to the background, though.
(3) A shutdown op is added to the AF_RXRPC socket so that the AFS
filesystem can shut down the socket and flush its own work items
before closing the socket to deal with any in-progress service calls.
Future additional changes that will need to be considered:
(1) Make sure that a call doesn't hog the front of the queue by receiving
data from the network as fast as userspace is consuming it to the
exclusion of other calls.
(2) Transmit delayed ACKs from within recvmsg() when we've consumed
sufficiently more packets to avoid the background work item needing to
run.
Signed-off-by: David Howells <dhowells@redhat.com>
2016-09-08 10:10:12 +00:00
|
|
|
}
|
|
|
|
|
|
2016-09-24 17:05:27 +00:00
|
|
|
/*
|
2024-12-04 07:47:07 +00:00
|
|
|
* Do TCP-style congestion management [RFC5681].
|
2016-09-24 17:05:27 +00:00
|
|
|
*/
|
|
|
|
|
static void rxrpc_congestion_management(struct rxrpc_call *call,
|
2024-12-04 07:46:52 +00:00
|
|
|
struct rxrpc_ack_summary *summary)
|
2016-09-24 17:05:27 +00:00
|
|
|
{
|
2024-12-04 07:46:50 +00:00
|
|
|
summary->change = rxrpc_cong_no_change;
|
2024-12-04 07:47:07 +00:00
|
|
|
summary->in_flight = rxrpc_tx_in_flight(call);
|
2016-09-24 17:05:27 +00:00
|
|
|
|
|
|
|
|
if (test_and_clear_bit(RXRPC_CALL_RETRANS_TIMEOUT, &call->flags)) {
|
|
|
|
|
summary->retrans_timeo = true;
|
2024-12-04 07:46:50 +00:00
|
|
|
call->cong_ssthresh = umax(summary->in_flight / 2, 2);
|
|
|
|
|
call->cong_cwnd = 1;
|
|
|
|
|
if (call->cong_cwnd >= call->cong_ssthresh &&
|
|
|
|
|
call->cong_ca_state == RXRPC_CA_SLOW_START) {
|
|
|
|
|
call->cong_ca_state = RXRPC_CA_CONGEST_AVOIDANCE;
|
2024-12-04 07:46:52 +00:00
|
|
|
call->cong_tstamp = call->acks_latest_ts;
|
2024-12-04 07:46:50 +00:00
|
|
|
call->cong_cumul_acks = 0;
|
2016-09-24 17:05:27 +00:00
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
|
2024-12-04 07:46:50 +00:00
|
|
|
call->cong_cumul_acks += summary->nr_new_sacks;
|
2024-12-04 07:46:51 +00:00
|
|
|
call->cong_cumul_acks += summary->nr_new_hacks;
|
2024-12-04 07:46:50 +00:00
|
|
|
if (call->cong_cumul_acks > 255)
|
|
|
|
|
call->cong_cumul_acks = 255;
|
2016-09-24 17:05:27 +00:00
|
|
|
|
2024-12-04 07:46:50 +00:00
|
|
|
switch (call->cong_ca_state) {
|
|
|
|
|
case RXRPC_CA_SLOW_START:
|
2024-12-04 07:46:51 +00:00
|
|
|
if (call->acks_nr_snacks > 0)
|
2016-09-24 17:05:27 +00:00
|
|
|
goto packet_loss_detected;
|
2024-12-04 07:46:50 +00:00
|
|
|
if (call->cong_cumul_acks > 0)
|
|
|
|
|
call->cong_cwnd += 1;
|
|
|
|
|
if (call->cong_cwnd >= call->cong_ssthresh) {
|
|
|
|
|
call->cong_ca_state = RXRPC_CA_CONGEST_AVOIDANCE;
|
2024-12-04 07:46:52 +00:00
|
|
|
call->cong_tstamp = call->acks_latest_ts;
|
2016-09-24 17:05:27 +00:00
|
|
|
}
|
|
|
|
|
goto out;
|
|
|
|
|
|
2024-12-04 07:46:50 +00:00
|
|
|
case RXRPC_CA_CONGEST_AVOIDANCE:
|
2024-12-04 07:46:51 +00:00
|
|
|
if (call->acks_nr_snacks > 0)
|
2016-09-24 17:05:27 +00:00
|
|
|
goto packet_loss_detected;
|
|
|
|
|
|
|
|
|
|
/* We analyse the number of packets that get ACK'd per RTT
|
|
|
|
|
* period and increase the window if we managed to fill it.
|
|
|
|
|
*/
|
2024-12-04 07:47:05 +00:00
|
|
|
if (call->rtt_count == 0)
|
2016-09-24 17:05:27 +00:00
|
|
|
goto out;
|
2024-12-04 07:46:52 +00:00
|
|
|
if (ktime_before(call->acks_latest_ts,
|
2020-05-11 13:54:34 +00:00
|
|
|
ktime_add_us(call->cong_tstamp,
|
2024-12-04 07:47:05 +00:00
|
|
|
call->srtt_us >> 3)))
|
2016-09-24 17:05:27 +00:00
|
|
|
goto out_no_clear_ca;
|
2024-12-04 07:46:50 +00:00
|
|
|
summary->change = rxrpc_cong_rtt_window_end;
|
2024-12-04 07:46:52 +00:00
|
|
|
call->cong_tstamp = call->acks_latest_ts;
|
2024-12-04 07:46:50 +00:00
|
|
|
if (call->cong_cumul_acks >= call->cong_cwnd)
|
|
|
|
|
call->cong_cwnd++;
|
2016-09-24 17:05:27 +00:00
|
|
|
goto out;
|
|
|
|
|
|
2024-12-04 07:46:50 +00:00
|
|
|
case RXRPC_CA_PACKET_LOSS:
|
2024-12-04 07:46:51 +00:00
|
|
|
if (call->acks_nr_snacks == 0)
|
2016-09-24 17:05:27 +00:00
|
|
|
goto resume_normality;
|
|
|
|
|
|
2024-12-04 07:46:50 +00:00
|
|
|
if (summary->new_low_snack) {
|
|
|
|
|
summary->change = rxrpc_cong_new_low_nack;
|
2016-09-24 17:05:27 +00:00
|
|
|
call->cong_dup_acks = 1;
|
|
|
|
|
if (call->cong_extra > 1)
|
|
|
|
|
call->cong_extra = 1;
|
|
|
|
|
goto send_extra_data;
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
call->cong_dup_acks++;
|
|
|
|
|
if (call->cong_dup_acks < 3)
|
|
|
|
|
goto send_extra_data;
|
|
|
|
|
|
2024-12-04 07:46:50 +00:00
|
|
|
summary->change = rxrpc_cong_begin_retransmission;
|
|
|
|
|
call->cong_ca_state = RXRPC_CA_FAST_RETRANSMIT;
|
|
|
|
|
call->cong_ssthresh = umax(summary->in_flight / 2, 2);
|
|
|
|
|
call->cong_cwnd = call->cong_ssthresh + 3;
|
2016-09-24 17:05:27 +00:00
|
|
|
call->cong_extra = 0;
|
|
|
|
|
call->cong_dup_acks = 0;
|
2024-12-04 07:46:51 +00:00
|
|
|
summary->need_retransmit = true;
|
2024-12-04 07:47:07 +00:00
|
|
|
summary->in_fast_or_rto_recovery = true;
|
2016-09-24 17:05:27 +00:00
|
|
|
goto out;
|
|
|
|
|
|
2024-12-04 07:46:50 +00:00
|
|
|
case RXRPC_CA_FAST_RETRANSMIT:
|
2024-12-04 07:47:07 +00:00
|
|
|
rxrpc_tlp_init(call);
|
|
|
|
|
summary->in_fast_or_rto_recovery = true;
|
2024-12-04 07:46:50 +00:00
|
|
|
if (!summary->new_low_snack) {
|
|
|
|
|
if (summary->nr_new_sacks == 0)
|
|
|
|
|
call->cong_cwnd += 1;
|
2016-09-24 17:05:27 +00:00
|
|
|
call->cong_dup_acks++;
|
|
|
|
|
if (call->cong_dup_acks == 2) {
|
2024-12-04 07:46:50 +00:00
|
|
|
summary->change = rxrpc_cong_retransmit_again;
|
2016-09-24 17:05:27 +00:00
|
|
|
call->cong_dup_acks = 0;
|
2024-12-04 07:46:51 +00:00
|
|
|
summary->need_retransmit = true;
|
2016-09-24 17:05:27 +00:00
|
|
|
}
|
|
|
|
|
} else {
|
2024-12-04 07:46:50 +00:00
|
|
|
summary->change = rxrpc_cong_progress;
|
|
|
|
|
call->cong_cwnd = call->cong_ssthresh;
|
2024-12-04 07:47:07 +00:00
|
|
|
if (call->acks_nr_snacks == 0) {
|
|
|
|
|
summary->exiting_fast_or_rto_recovery = true;
|
2016-09-24 17:05:27 +00:00
|
|
|
goto resume_normality;
|
2024-12-04 07:47:07 +00:00
|
|
|
}
|
2016-09-24 17:05:27 +00:00
|
|
|
}
|
|
|
|
|
goto out;
|
|
|
|
|
|
|
|
|
|
default:
|
|
|
|
|
BUG();
|
|
|
|
|
goto out;
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
resume_normality:
|
2024-12-04 07:46:50 +00:00
|
|
|
summary->change = rxrpc_cong_cleared_nacks;
|
2016-09-24 17:05:27 +00:00
|
|
|
call->cong_dup_acks = 0;
|
|
|
|
|
call->cong_extra = 0;
|
2024-12-04 07:46:52 +00:00
|
|
|
call->cong_tstamp = call->acks_latest_ts;
|
2024-12-04 07:46:50 +00:00
|
|
|
if (call->cong_cwnd < call->cong_ssthresh)
|
|
|
|
|
call->cong_ca_state = RXRPC_CA_SLOW_START;
|
2016-09-24 17:05:27 +00:00
|
|
|
else
|
2024-12-04 07:46:50 +00:00
|
|
|
call->cong_ca_state = RXRPC_CA_CONGEST_AVOIDANCE;
|
2016-09-24 17:05:27 +00:00
|
|
|
out:
|
2024-12-04 07:46:50 +00:00
|
|
|
call->cong_cumul_acks = 0;
|
2016-09-24 17:05:27 +00:00
|
|
|
out_no_clear_ca:
|
2024-12-04 07:46:50 +00:00
|
|
|
if (call->cong_cwnd >= RXRPC_TX_MAX_WINDOW)
|
|
|
|
|
call->cong_cwnd = RXRPC_TX_MAX_WINDOW;
|
2024-12-04 07:46:52 +00:00
|
|
|
trace_rxrpc_congest(call, summary);
|
2016-09-24 17:05:27 +00:00
|
|
|
return;
|
|
|
|
|
|
|
|
|
|
packet_loss_detected:
|
2024-12-04 07:46:50 +00:00
|
|
|
summary->change = rxrpc_cong_saw_nack;
|
|
|
|
|
call->cong_ca_state = RXRPC_CA_PACKET_LOSS;
|
2016-09-24 17:05:27 +00:00
|
|
|
call->cong_dup_acks = 0;
|
|
|
|
|
goto send_extra_data;
|
|
|
|
|
|
|
|
|
|
send_extra_data:
|
|
|
|
|
/* Send some previously unsent DATA if we have some to advance the ACK
|
|
|
|
|
* state.
|
|
|
|
|
*/
|
2022-03-31 22:55:08 +00:00
|
|
|
if (test_bit(RXRPC_CALL_TX_LAST, &call->flags) ||
|
2024-12-04 07:46:50 +00:00
|
|
|
call->acks_nr_sacks != call->tx_top - call->tx_bottom) {
|
2016-09-24 17:05:27 +00:00
|
|
|
call->cong_extra++;
|
|
|
|
|
wake_up(&call->waitq);
|
|
|
|
|
}
|
|
|
|
|
goto out_no_clear_ca;
|
|
|
|
|
}
|
|
|
|
|
|
2022-11-11 13:47:35 +00:00
|
|
|
/*
|
|
|
|
|
* Degrade the congestion window if we haven't transmitted a packet for >1RTT.
|
|
|
|
|
*/
|
|
|
|
|
void rxrpc_congestion_degrade(struct rxrpc_call *call)
|
|
|
|
|
{
|
2024-12-04 07:47:07 +00:00
|
|
|
ktime_t rtt, now, time_since;
|
2022-11-11 13:47:35 +00:00
|
|
|
|
2024-12-04 07:46:50 +00:00
|
|
|
if (call->cong_ca_state != RXRPC_CA_SLOW_START &&
|
|
|
|
|
call->cong_ca_state != RXRPC_CA_CONGEST_AVOIDANCE)
|
2022-11-11 13:47:35 +00:00
|
|
|
return;
|
2022-10-27 10:25:55 +00:00
|
|
|
if (__rxrpc_call_state(call) == RXRPC_CALL_CLIENT_AWAIT_REPLY)
|
2022-11-11 13:47:35 +00:00
|
|
|
return;
|
|
|
|
|
|
2024-12-04 07:47:05 +00:00
|
|
|
rtt = ns_to_ktime(call->srtt_us * (NSEC_PER_USEC / 8));
|
2022-11-11 13:47:35 +00:00
|
|
|
now = ktime_get_real();
|
2024-12-04 07:47:07 +00:00
|
|
|
time_since = ktime_sub(now, call->tx_last_sent);
|
|
|
|
|
if (ktime_before(time_since, rtt))
|
2022-11-11 13:47:35 +00:00
|
|
|
return;
|
|
|
|
|
|
2024-12-04 07:47:07 +00:00
|
|
|
trace_rxrpc_reset_cwnd(call, time_since, rtt);
|
2022-11-11 13:47:35 +00:00
|
|
|
rxrpc_inc_stat(call->rxnet, stat_tx_data_cwnd_reset);
|
|
|
|
|
call->tx_last_sent = now;
|
2024-12-04 07:46:50 +00:00
|
|
|
call->cong_ca_state = RXRPC_CA_SLOW_START;
|
2024-12-04 07:46:31 +00:00
|
|
|
call->cong_ssthresh = umax(call->cong_ssthresh, call->cong_cwnd * 3 / 4);
|
|
|
|
|
call->cong_cwnd = umax(call->cong_cwnd / 2, RXRPC_MIN_CWND);
|
2022-11-11 13:47:35 +00:00
|
|
|
}
|
|
|
|
|
|
2024-12-04 07:46:52 +00:00
|
|
|
/*
|
|
|
|
|
* Add an RTT sample derived from an ACK'd DATA packet.
|
|
|
|
|
*/
|
|
|
|
|
static void rxrpc_add_data_rtt_sample(struct rxrpc_call *call,
|
|
|
|
|
struct rxrpc_ack_summary *summary,
|
|
|
|
|
struct rxrpc_txqueue *tq,
|
2024-12-04 07:47:02 +00:00
|
|
|
int ix)
|
2024-12-04 07:46:52 +00:00
|
|
|
{
|
2024-12-04 07:47:07 +00:00
|
|
|
ktime_t xmit_ts = ktime_add_us(tq->xmit_ts_base, tq->segment_xmit_ts[ix]);
|
|
|
|
|
|
2024-12-04 07:47:05 +00:00
|
|
|
rxrpc_call_add_rtt(call, rxrpc_rtt_rx_data_ack, -1,
|
2024-12-04 07:47:02 +00:00
|
|
|
summary->acked_serial, summary->ack_serial,
|
2024-12-04 07:47:07 +00:00
|
|
|
xmit_ts, call->acks_latest_ts);
|
2024-12-04 07:46:52 +00:00
|
|
|
__clear_bit(ix, &tq->rtt_samples); /* Prevent repeat RTT sample */
|
|
|
|
|
}
|
|
|
|
|
|
2007-04-26 22:48:28 +00:00
|
|
|
/*
|
rxrpc: Rewrite the data and ack handling code
Rewrite the data and ack handling code such that:
(1) Parsing of received ACK and ABORT packets and the distribution and the
filing of DATA packets happens entirely within the data_ready context
called from the UDP socket. This allows us to process and discard ACK
and ABORT packets much more quickly (they're no longer stashed on a
queue for a background thread to process).
(2) We avoid calling skb_clone(), pskb_pull() and pskb_trim(). We instead
keep track of the offset and length of the content of each packet in
the sk_buff metadata. This means we don't do any allocation in the
receive path.
(3) Jumbo DATA packet parsing is now done in data_ready context. Rather
than cloning the packet once for each subpacket and pulling/trimming
it, we file the packet multiple times with an annotation for each
indicating which subpacket is there. From that we can directly
calculate the offset and length.
(4) A call's receive queue can be accessed without taking locks (memory
barriers do have to be used, though).
(5) Incoming calls are set up from preallocated resources and immediately
made live. They can than have packets queued upon them and ACKs
generated. If insufficient resources exist, DATA packet #1 is given a
BUSY reply and other DATA packets are discarded).
(6) sk_buffs no longer take a ref on their parent call.
To make this work, the following changes are made:
(1) Each call's receive buffer is now a circular buffer of sk_buff
pointers (rxtx_buffer) rather than a number of sk_buff_heads spread
between the call and the socket. This permits each sk_buff to be in
the buffer multiple times. The receive buffer is reused for the
transmit buffer.
(2) A circular buffer of annotations (rxtx_annotations) is kept parallel
to the data buffer. Transmission phase annotations indicate whether a
buffered packet has been ACK'd or not and whether it needs
retransmission.
Receive phase annotations indicate whether a slot holds a whole packet
or a jumbo subpacket and, if the latter, which subpacket. They also
note whether the packet has been decrypted in place.
(3) DATA packet window tracking is much simplified. Each phase has just
two numbers representing the window (rx_hard_ack/rx_top and
tx_hard_ack/tx_top).
The hard_ack number is the sequence number before base of the window,
representing the last packet the other side says it has consumed.
hard_ack starts from 0 and the first packet is sequence number 1.
The top number is the sequence number of the highest-numbered packet
residing in the buffer. Packets between hard_ack+1 and top are
soft-ACK'd to indicate they've been received, but not yet consumed.
Four macros, before(), before_eq(), after() and after_eq() are added
to compare sequence numbers within the window. This allows for the
top of the window to wrap when the hard-ack sequence number gets close
to the limit.
Two flags, RXRPC_CALL_RX_LAST and RXRPC_CALL_TX_LAST, are added also
to indicate when rx_top and tx_top point at the packets with the
LAST_PACKET bit set, indicating the end of the phase.
(4) Calls are queued on the socket 'receive queue' rather than packets.
This means that we don't need have to invent dummy packets to queue to
indicate abnormal/terminal states and we don't have to keep metadata
packets (such as ABORTs) around
(5) The offset and length of a (sub)packet's content are now passed to
the verify_packet security op. This is currently expected to decrypt
the packet in place and validate it.
However, there's now nowhere to store the revised offset and length of
the actual data within the decrypted blob (there may be a header and
padding to skip) because an sk_buff may represent multiple packets, so
a locate_data security op is added to retrieve these details from the
sk_buff content when needed.
(6) recvmsg() now has to handle jumbo subpackets, where each subpacket is
individually secured and needs to be individually decrypted. The code
to do this is broken out into rxrpc_recvmsg_data() and shared with the
kernel API. It now iterates over the call's receive buffer rather
than walking the socket receive queue.
Additional changes:
(1) The timers are condensed to a single timer that is set for the soonest
of three timeouts (delayed ACK generation, DATA retransmission and
call lifespan).
(2) Transmission of ACK and ABORT packets is effected immediately from
process-context socket ops/kernel API calls that cause them instead of
them being punted off to a background work item. The data_ready
handler still has to defer to the background, though.
(3) A shutdown op is added to the AF_RXRPC socket so that the AFS
filesystem can shut down the socket and flush its own work items
before closing the socket to deal with any in-progress service calls.
Future additional changes that will need to be considered:
(1) Make sure that a call doesn't hog the front of the queue by receiving
data from the network as fast as userspace is consuming it to the
exclusion of other calls.
(2) Transmit delayed ACKs from within recvmsg() when we've consumed
sufficiently more packets to avoid the background work item needing to
run.
Signed-off-by: David Howells <dhowells@redhat.com>
2016-09-08 10:10:12 +00:00
|
|
|
* Apply a hard ACK by advancing the Tx window.
|
2007-04-26 22:48:28 +00:00
|
|
|
*/
|
2018-10-08 14:46:01 +00:00
|
|
|
static bool rxrpc_rotate_tx_window(struct rxrpc_call *call, rxrpc_seq_t to,
|
2024-12-04 07:47:02 +00:00
|
|
|
struct rxrpc_ack_summary *summary)
|
2007-04-26 22:48:28 +00:00
|
|
|
{
|
2024-12-04 07:46:46 +00:00
|
|
|
struct rxrpc_txqueue *tq = call->tx_queue;
|
|
|
|
|
rxrpc_seq_t seq = call->tx_bottom + 1;
|
2024-12-04 07:47:07 +00:00
|
|
|
bool rot_last = false, trace = false;
|
2007-04-26 22:48:28 +00:00
|
|
|
|
2024-12-04 07:46:47 +00:00
|
|
|
_enter("%x,%x", call->tx_bottom, to);
|
2024-12-04 07:46:46 +00:00
|
|
|
|
|
|
|
|
trace_rxrpc_tx_rotate(call, seq, to);
|
|
|
|
|
trace_rxrpc_tq(call, tq, seq, rxrpc_tq_rotate);
|
|
|
|
|
|
2024-12-04 07:46:51 +00:00
|
|
|
if (call->acks_lowest_nak == call->tx_bottom) {
|
|
|
|
|
call->acks_lowest_nak = to;
|
|
|
|
|
} else if (after(to, call->acks_lowest_nak)) {
|
|
|
|
|
summary->new_low_snack = true;
|
|
|
|
|
call->acks_lowest_nak = to;
|
|
|
|
|
}
|
|
|
|
|
|
2024-12-04 07:46:46 +00:00
|
|
|
/* We may have a left over fully-consumed buffer at the front that we
|
|
|
|
|
* couldn't drop before (rotate_and_keep below).
|
|
|
|
|
*/
|
|
|
|
|
if (seq == call->tx_qbase + RXRPC_NR_TXQUEUE) {
|
|
|
|
|
call->tx_qbase += RXRPC_NR_TXQUEUE;
|
|
|
|
|
call->tx_queue = tq->next;
|
|
|
|
|
trace_rxrpc_tq(call, tq, seq, rxrpc_tq_rotate_and_free);
|
|
|
|
|
kfree(tq);
|
|
|
|
|
tq = call->tx_queue;
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
do {
|
|
|
|
|
unsigned int ix = seq - call->tx_qbase;
|
|
|
|
|
|
|
|
|
|
_debug("tq=%x seq=%x i=%d f=%x", tq->qbase, seq, ix, tq->bufs[ix]->flags);
|
|
|
|
|
if (tq->bufs[ix]->flags & RXRPC_LAST_PACKET) {
|
2016-09-23 11:39:22 +00:00
|
|
|
set_bit(RXRPC_CALL_TX_LAST, &call->flags);
|
2018-10-08 14:46:01 +00:00
|
|
|
rot_last = true;
|
|
|
|
|
}
|
2024-12-04 07:46:51 +00:00
|
|
|
|
2024-12-04 07:47:07 +00:00
|
|
|
if (summary->acked_serial == tq->segment_serial[ix] &&
|
2024-12-04 07:46:52 +00:00
|
|
|
test_bit(ix, &tq->rtt_samples))
|
2024-12-04 07:47:02 +00:00
|
|
|
rxrpc_add_data_rtt_sample(call, summary, tq, ix);
|
2024-12-04 07:46:52 +00:00
|
|
|
|
2024-12-04 07:46:51 +00:00
|
|
|
if (ix == tq->nr_reported_acks) {
|
|
|
|
|
/* Packet directly hard ACK'd. */
|
|
|
|
|
tq->nr_reported_acks++;
|
2024-12-04 07:47:07 +00:00
|
|
|
rxrpc_input_rack_one(call, summary, tq, ix);
|
|
|
|
|
if (seq == call->tlp_seq)
|
|
|
|
|
summary->tlp_probe_acked = true;
|
2024-12-04 07:46:51 +00:00
|
|
|
summary->nr_new_hacks++;
|
|
|
|
|
__set_bit(ix, &tq->segment_acked);
|
|
|
|
|
trace_rxrpc_rotate(call, tq, summary, seq, rxrpc_rotate_trace_hack);
|
|
|
|
|
} else if (test_bit(ix, &tq->segment_acked)) {
|
|
|
|
|
/* Soft ACK -> hard ACK. */
|
|
|
|
|
call->acks_nr_sacks--;
|
|
|
|
|
trace_rxrpc_rotate(call, tq, summary, seq, rxrpc_rotate_trace_sack);
|
|
|
|
|
} else {
|
|
|
|
|
/* Soft NAK -> hard ACK. */
|
|
|
|
|
call->acks_nr_snacks--;
|
2024-12-04 07:47:07 +00:00
|
|
|
rxrpc_input_rack_one(call, summary, tq, ix);
|
|
|
|
|
if (seq == call->tlp_seq)
|
|
|
|
|
summary->tlp_probe_acked = true;
|
2024-12-04 07:46:51 +00:00
|
|
|
summary->nr_new_hacks++;
|
|
|
|
|
__set_bit(ix, &tq->segment_acked);
|
|
|
|
|
trace_rxrpc_rotate(call, tq, summary, seq, rxrpc_rotate_trace_snak);
|
|
|
|
|
}
|
|
|
|
|
|
2024-12-04 07:47:07 +00:00
|
|
|
call->tx_nr_sent--;
|
|
|
|
|
if (__test_and_clear_bit(ix, &tq->segment_lost))
|
|
|
|
|
call->tx_nr_lost--;
|
|
|
|
|
if (__test_and_clear_bit(ix, &tq->segment_retransmitted))
|
|
|
|
|
call->tx_nr_resent--;
|
|
|
|
|
__clear_bit(ix, &tq->ever_retransmitted);
|
|
|
|
|
|
2024-12-04 07:46:46 +00:00
|
|
|
rxrpc_put_txbuf(tq->bufs[ix], rxrpc_txbuf_put_rotated);
|
|
|
|
|
tq->bufs[ix] = NULL;
|
|
|
|
|
|
|
|
|
|
WRITE_ONCE(call->tx_bottom, seq);
|
|
|
|
|
trace_rxrpc_txqueue(call, (rot_last ?
|
|
|
|
|
rxrpc_txqueue_rotate_last :
|
|
|
|
|
rxrpc_txqueue_rotate));
|
|
|
|
|
|
|
|
|
|
seq++;
|
2024-12-04 07:47:07 +00:00
|
|
|
trace = true;
|
2024-12-04 07:46:46 +00:00
|
|
|
if (!(seq & RXRPC_TXQ_MASK)) {
|
2024-12-04 07:47:07 +00:00
|
|
|
trace_rxrpc_rack_update(call, summary);
|
|
|
|
|
trace = false;
|
2024-12-04 07:46:46 +00:00
|
|
|
prefetch(tq->next);
|
|
|
|
|
if (tq != call->tx_qtail) {
|
|
|
|
|
call->tx_qbase += RXRPC_NR_TXQUEUE;
|
|
|
|
|
call->tx_queue = tq->next;
|
|
|
|
|
trace_rxrpc_tq(call, tq, seq, rxrpc_tq_rotate_and_free);
|
|
|
|
|
kfree(tq);
|
|
|
|
|
tq = call->tx_queue;
|
|
|
|
|
} else {
|
|
|
|
|
trace_rxrpc_tq(call, tq, seq, rxrpc_tq_rotate_and_keep);
|
|
|
|
|
tq = NULL;
|
|
|
|
|
break;
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
} while (before_eq(seq, to));
|
2007-04-26 22:48:28 +00:00
|
|
|
|
2024-12-04 07:47:07 +00:00
|
|
|
if (trace)
|
|
|
|
|
trace_rxrpc_rack_update(call, summary);
|
|
|
|
|
|
2024-12-04 07:46:46 +00:00
|
|
|
if (rot_last) {
|
2022-03-31 22:55:08 +00:00
|
|
|
set_bit(RXRPC_CALL_TX_ALL_ACKED, &call->flags);
|
2024-12-04 07:46:46 +00:00
|
|
|
if (tq) {
|
|
|
|
|
trace_rxrpc_tq(call, tq, seq, rxrpc_tq_rotate_and_free);
|
|
|
|
|
kfree(tq);
|
|
|
|
|
call->tx_queue = NULL;
|
|
|
|
|
}
|
|
|
|
|
}
|
2007-04-26 22:48:28 +00:00
|
|
|
|
2024-12-04 07:46:47 +00:00
|
|
|
_debug("%x,%x,%x,%d", to, call->tx_bottom, call->tx_top, rot_last);
|
2016-09-13 21:36:21 +00:00
|
|
|
|
2022-03-31 22:55:08 +00:00
|
|
|
wake_up(&call->waitq);
|
2018-10-08 14:46:01 +00:00
|
|
|
return rot_last;
|
rxrpc: Rewrite the data and ack handling code
Rewrite the data and ack handling code such that:
(1) Parsing of received ACK and ABORT packets and the distribution and the
filing of DATA packets happens entirely within the data_ready context
called from the UDP socket. This allows us to process and discard ACK
and ABORT packets much more quickly (they're no longer stashed on a
queue for a background thread to process).
(2) We avoid calling skb_clone(), pskb_pull() and pskb_trim(). We instead
keep track of the offset and length of the content of each packet in
the sk_buff metadata. This means we don't do any allocation in the
receive path.
(3) Jumbo DATA packet parsing is now done in data_ready context. Rather
than cloning the packet once for each subpacket and pulling/trimming
it, we file the packet multiple times with an annotation for each
indicating which subpacket is there. From that we can directly
calculate the offset and length.
(4) A call's receive queue can be accessed without taking locks (memory
barriers do have to be used, though).
(5) Incoming calls are set up from preallocated resources and immediately
made live. They can than have packets queued upon them and ACKs
generated. If insufficient resources exist, DATA packet #1 is given a
BUSY reply and other DATA packets are discarded).
(6) sk_buffs no longer take a ref on their parent call.
To make this work, the following changes are made:
(1) Each call's receive buffer is now a circular buffer of sk_buff
pointers (rxtx_buffer) rather than a number of sk_buff_heads spread
between the call and the socket. This permits each sk_buff to be in
the buffer multiple times. The receive buffer is reused for the
transmit buffer.
(2) A circular buffer of annotations (rxtx_annotations) is kept parallel
to the data buffer. Transmission phase annotations indicate whether a
buffered packet has been ACK'd or not and whether it needs
retransmission.
Receive phase annotations indicate whether a slot holds a whole packet
or a jumbo subpacket and, if the latter, which subpacket. They also
note whether the packet has been decrypted in place.
(3) DATA packet window tracking is much simplified. Each phase has just
two numbers representing the window (rx_hard_ack/rx_top and
tx_hard_ack/tx_top).
The hard_ack number is the sequence number before base of the window,
representing the last packet the other side says it has consumed.
hard_ack starts from 0 and the first packet is sequence number 1.
The top number is the sequence number of the highest-numbered packet
residing in the buffer. Packets between hard_ack+1 and top are
soft-ACK'd to indicate they've been received, but not yet consumed.
Four macros, before(), before_eq(), after() and after_eq() are added
to compare sequence numbers within the window. This allows for the
top of the window to wrap when the hard-ack sequence number gets close
to the limit.
Two flags, RXRPC_CALL_RX_LAST and RXRPC_CALL_TX_LAST, are added also
to indicate when rx_top and tx_top point at the packets with the
LAST_PACKET bit set, indicating the end of the phase.
(4) Calls are queued on the socket 'receive queue' rather than packets.
This means that we don't need have to invent dummy packets to queue to
indicate abnormal/terminal states and we don't have to keep metadata
packets (such as ABORTs) around
(5) The offset and length of a (sub)packet's content are now passed to
the verify_packet security op. This is currently expected to decrypt
the packet in place and validate it.
However, there's now nowhere to store the revised offset and length of
the actual data within the decrypted blob (there may be a header and
padding to skip) because an sk_buff may represent multiple packets, so
a locate_data security op is added to retrieve these details from the
sk_buff content when needed.
(6) recvmsg() now has to handle jumbo subpackets, where each subpacket is
individually secured and needs to be individually decrypted. The code
to do this is broken out into rxrpc_recvmsg_data() and shared with the
kernel API. It now iterates over the call's receive buffer rather
than walking the socket receive queue.
Additional changes:
(1) The timers are condensed to a single timer that is set for the soonest
of three timeouts (delayed ACK generation, DATA retransmission and
call lifespan).
(2) Transmission of ACK and ABORT packets is effected immediately from
process-context socket ops/kernel API calls that cause them instead of
them being punted off to a background work item. The data_ready
handler still has to defer to the background, though.
(3) A shutdown op is added to the AF_RXRPC socket so that the AFS
filesystem can shut down the socket and flush its own work items
before closing the socket to deal with any in-progress service calls.
Future additional changes that will need to be considered:
(1) Make sure that a call doesn't hog the front of the queue by receiving
data from the network as fast as userspace is consuming it to the
exclusion of other calls.
(2) Transmit delayed ACKs from within recvmsg() when we've consumed
sufficiently more packets to avoid the background work item needing to
run.
Signed-off-by: David Howells <dhowells@redhat.com>
2016-09-08 10:10:12 +00:00
|
|
|
}
|
2007-04-26 22:48:28 +00:00
|
|
|
|
rxrpc: Rewrite the data and ack handling code
Rewrite the data and ack handling code such that:
(1) Parsing of received ACK and ABORT packets and the distribution and the
filing of DATA packets happens entirely within the data_ready context
called from the UDP socket. This allows us to process and discard ACK
and ABORT packets much more quickly (they're no longer stashed on a
queue for a background thread to process).
(2) We avoid calling skb_clone(), pskb_pull() and pskb_trim(). We instead
keep track of the offset and length of the content of each packet in
the sk_buff metadata. This means we don't do any allocation in the
receive path.
(3) Jumbo DATA packet parsing is now done in data_ready context. Rather
than cloning the packet once for each subpacket and pulling/trimming
it, we file the packet multiple times with an annotation for each
indicating which subpacket is there. From that we can directly
calculate the offset and length.
(4) A call's receive queue can be accessed without taking locks (memory
barriers do have to be used, though).
(5) Incoming calls are set up from preallocated resources and immediately
made live. They can than have packets queued upon them and ACKs
generated. If insufficient resources exist, DATA packet #1 is given a
BUSY reply and other DATA packets are discarded).
(6) sk_buffs no longer take a ref on their parent call.
To make this work, the following changes are made:
(1) Each call's receive buffer is now a circular buffer of sk_buff
pointers (rxtx_buffer) rather than a number of sk_buff_heads spread
between the call and the socket. This permits each sk_buff to be in
the buffer multiple times. The receive buffer is reused for the
transmit buffer.
(2) A circular buffer of annotations (rxtx_annotations) is kept parallel
to the data buffer. Transmission phase annotations indicate whether a
buffered packet has been ACK'd or not and whether it needs
retransmission.
Receive phase annotations indicate whether a slot holds a whole packet
or a jumbo subpacket and, if the latter, which subpacket. They also
note whether the packet has been decrypted in place.
(3) DATA packet window tracking is much simplified. Each phase has just
two numbers representing the window (rx_hard_ack/rx_top and
tx_hard_ack/tx_top).
The hard_ack number is the sequence number before base of the window,
representing the last packet the other side says it has consumed.
hard_ack starts from 0 and the first packet is sequence number 1.
The top number is the sequence number of the highest-numbered packet
residing in the buffer. Packets between hard_ack+1 and top are
soft-ACK'd to indicate they've been received, but not yet consumed.
Four macros, before(), before_eq(), after() and after_eq() are added
to compare sequence numbers within the window. This allows for the
top of the window to wrap when the hard-ack sequence number gets close
to the limit.
Two flags, RXRPC_CALL_RX_LAST and RXRPC_CALL_TX_LAST, are added also
to indicate when rx_top and tx_top point at the packets with the
LAST_PACKET bit set, indicating the end of the phase.
(4) Calls are queued on the socket 'receive queue' rather than packets.
This means that we don't need have to invent dummy packets to queue to
indicate abnormal/terminal states and we don't have to keep metadata
packets (such as ABORTs) around
(5) The offset and length of a (sub)packet's content are now passed to
the verify_packet security op. This is currently expected to decrypt
the packet in place and validate it.
However, there's now nowhere to store the revised offset and length of
the actual data within the decrypted blob (there may be a header and
padding to skip) because an sk_buff may represent multiple packets, so
a locate_data security op is added to retrieve these details from the
sk_buff content when needed.
(6) recvmsg() now has to handle jumbo subpackets, where each subpacket is
individually secured and needs to be individually decrypted. The code
to do this is broken out into rxrpc_recvmsg_data() and shared with the
kernel API. It now iterates over the call's receive buffer rather
than walking the socket receive queue.
Additional changes:
(1) The timers are condensed to a single timer that is set for the soonest
of three timeouts (delayed ACK generation, DATA retransmission and
call lifespan).
(2) Transmission of ACK and ABORT packets is effected immediately from
process-context socket ops/kernel API calls that cause them instead of
them being punted off to a background work item. The data_ready
handler still has to defer to the background, though.
(3) A shutdown op is added to the AF_RXRPC socket so that the AFS
filesystem can shut down the socket and flush its own work items
before closing the socket to deal with any in-progress service calls.
Future additional changes that will need to be considered:
(1) Make sure that a call doesn't hog the front of the queue by receiving
data from the network as fast as userspace is consuming it to the
exclusion of other calls.
(2) Transmit delayed ACKs from within recvmsg() when we've consumed
sufficiently more packets to avoid the background work item needing to
run.
Signed-off-by: David Howells <dhowells@redhat.com>
2016-09-08 10:10:12 +00:00
|
|
|
/*
|
|
|
|
|
* End the transmission phase of a call.
|
|
|
|
|
*
|
|
|
|
|
* This occurs when we get an ACKALL packet, the first DATA packet of a reply,
|
|
|
|
|
* or a final ACK packet.
|
|
|
|
|
*/
|
2022-10-06 20:45:42 +00:00
|
|
|
static void rxrpc_end_tx_phase(struct rxrpc_call *call, bool reply_begun,
|
|
|
|
|
enum rxrpc_abort_reason abort_why)
|
rxrpc: Rewrite the data and ack handling code
Rewrite the data and ack handling code such that:
(1) Parsing of received ACK and ABORT packets and the distribution and the
filing of DATA packets happens entirely within the data_ready context
called from the UDP socket. This allows us to process and discard ACK
and ABORT packets much more quickly (they're no longer stashed on a
queue for a background thread to process).
(2) We avoid calling skb_clone(), pskb_pull() and pskb_trim(). We instead
keep track of the offset and length of the content of each packet in
the sk_buff metadata. This means we don't do any allocation in the
receive path.
(3) Jumbo DATA packet parsing is now done in data_ready context. Rather
than cloning the packet once for each subpacket and pulling/trimming
it, we file the packet multiple times with an annotation for each
indicating which subpacket is there. From that we can directly
calculate the offset and length.
(4) A call's receive queue can be accessed without taking locks (memory
barriers do have to be used, though).
(5) Incoming calls are set up from preallocated resources and immediately
made live. They can than have packets queued upon them and ACKs
generated. If insufficient resources exist, DATA packet #1 is given a
BUSY reply and other DATA packets are discarded).
(6) sk_buffs no longer take a ref on their parent call.
To make this work, the following changes are made:
(1) Each call's receive buffer is now a circular buffer of sk_buff
pointers (rxtx_buffer) rather than a number of sk_buff_heads spread
between the call and the socket. This permits each sk_buff to be in
the buffer multiple times. The receive buffer is reused for the
transmit buffer.
(2) A circular buffer of annotations (rxtx_annotations) is kept parallel
to the data buffer. Transmission phase annotations indicate whether a
buffered packet has been ACK'd or not and whether it needs
retransmission.
Receive phase annotations indicate whether a slot holds a whole packet
or a jumbo subpacket and, if the latter, which subpacket. They also
note whether the packet has been decrypted in place.
(3) DATA packet window tracking is much simplified. Each phase has just
two numbers representing the window (rx_hard_ack/rx_top and
tx_hard_ack/tx_top).
The hard_ack number is the sequence number before base of the window,
representing the last packet the other side says it has consumed.
hard_ack starts from 0 and the first packet is sequence number 1.
The top number is the sequence number of the highest-numbered packet
residing in the buffer. Packets between hard_ack+1 and top are
soft-ACK'd to indicate they've been received, but not yet consumed.
Four macros, before(), before_eq(), after() and after_eq() are added
to compare sequence numbers within the window. This allows for the
top of the window to wrap when the hard-ack sequence number gets close
to the limit.
Two flags, RXRPC_CALL_RX_LAST and RXRPC_CALL_TX_LAST, are added also
to indicate when rx_top and tx_top point at the packets with the
LAST_PACKET bit set, indicating the end of the phase.
(4) Calls are queued on the socket 'receive queue' rather than packets.
This means that we don't need have to invent dummy packets to queue to
indicate abnormal/terminal states and we don't have to keep metadata
packets (such as ABORTs) around
(5) The offset and length of a (sub)packet's content are now passed to
the verify_packet security op. This is currently expected to decrypt
the packet in place and validate it.
However, there's now nowhere to store the revised offset and length of
the actual data within the decrypted blob (there may be a header and
padding to skip) because an sk_buff may represent multiple packets, so
a locate_data security op is added to retrieve these details from the
sk_buff content when needed.
(6) recvmsg() now has to handle jumbo subpackets, where each subpacket is
individually secured and needs to be individually decrypted. The code
to do this is broken out into rxrpc_recvmsg_data() and shared with the
kernel API. It now iterates over the call's receive buffer rather
than walking the socket receive queue.
Additional changes:
(1) The timers are condensed to a single timer that is set for the soonest
of three timeouts (delayed ACK generation, DATA retransmission and
call lifespan).
(2) Transmission of ACK and ABORT packets is effected immediately from
process-context socket ops/kernel API calls that cause them instead of
them being punted off to a background work item. The data_ready
handler still has to defer to the background, though.
(3) A shutdown op is added to the AF_RXRPC socket so that the AFS
filesystem can shut down the socket and flush its own work items
before closing the socket to deal with any in-progress service calls.
Future additional changes that will need to be considered:
(1) Make sure that a call doesn't hog the front of the queue by receiving
data from the network as fast as userspace is consuming it to the
exclusion of other calls.
(2) Transmit delayed ACKs from within recvmsg() when we've consumed
sufficiently more packets to avoid the background work item needing to
run.
Signed-off-by: David Howells <dhowells@redhat.com>
2016-09-08 10:10:12 +00:00
|
|
|
{
|
2016-09-23 11:39:22 +00:00
|
|
|
ASSERT(test_bit(RXRPC_CALL_TX_LAST, &call->flags));
|
2007-04-26 22:48:28 +00:00
|
|
|
|
2024-12-04 07:47:07 +00:00
|
|
|
call->rack_timer_mode = RXRPC_CALL_RACKTIMER_OFF;
|
|
|
|
|
call->rack_timo_at = KTIME_MAX;
|
|
|
|
|
trace_rxrpc_rack_timer(call, 0, false);
|
|
|
|
|
trace_rxrpc_timer_can(call, rxrpc_timer_trace_rack_off + call->rack_timer_mode);
|
2024-01-30 21:37:16 +00:00
|
|
|
|
2022-10-27 10:25:55 +00:00
|
|
|
switch (__rxrpc_call_state(call)) {
|
2016-09-23 11:39:22 +00:00
|
|
|
case RXRPC_CALL_CLIENT_SEND_REQUEST:
|
rxrpc: Rewrite the data and ack handling code
Rewrite the data and ack handling code such that:
(1) Parsing of received ACK and ABORT packets and the distribution and the
filing of DATA packets happens entirely within the data_ready context
called from the UDP socket. This allows us to process and discard ACK
and ABORT packets much more quickly (they're no longer stashed on a
queue for a background thread to process).
(2) We avoid calling skb_clone(), pskb_pull() and pskb_trim(). We instead
keep track of the offset and length of the content of each packet in
the sk_buff metadata. This means we don't do any allocation in the
receive path.
(3) Jumbo DATA packet parsing is now done in data_ready context. Rather
than cloning the packet once for each subpacket and pulling/trimming
it, we file the packet multiple times with an annotation for each
indicating which subpacket is there. From that we can directly
calculate the offset and length.
(4) A call's receive queue can be accessed without taking locks (memory
barriers do have to be used, though).
(5) Incoming calls are set up from preallocated resources and immediately
made live. They can than have packets queued upon them and ACKs
generated. If insufficient resources exist, DATA packet #1 is given a
BUSY reply and other DATA packets are discarded).
(6) sk_buffs no longer take a ref on their parent call.
To make this work, the following changes are made:
(1) Each call's receive buffer is now a circular buffer of sk_buff
pointers (rxtx_buffer) rather than a number of sk_buff_heads spread
between the call and the socket. This permits each sk_buff to be in
the buffer multiple times. The receive buffer is reused for the
transmit buffer.
(2) A circular buffer of annotations (rxtx_annotations) is kept parallel
to the data buffer. Transmission phase annotations indicate whether a
buffered packet has been ACK'd or not and whether it needs
retransmission.
Receive phase annotations indicate whether a slot holds a whole packet
or a jumbo subpacket and, if the latter, which subpacket. They also
note whether the packet has been decrypted in place.
(3) DATA packet window tracking is much simplified. Each phase has just
two numbers representing the window (rx_hard_ack/rx_top and
tx_hard_ack/tx_top).
The hard_ack number is the sequence number before base of the window,
representing the last packet the other side says it has consumed.
hard_ack starts from 0 and the first packet is sequence number 1.
The top number is the sequence number of the highest-numbered packet
residing in the buffer. Packets between hard_ack+1 and top are
soft-ACK'd to indicate they've been received, but not yet consumed.
Four macros, before(), before_eq(), after() and after_eq() are added
to compare sequence numbers within the window. This allows for the
top of the window to wrap when the hard-ack sequence number gets close
to the limit.
Two flags, RXRPC_CALL_RX_LAST and RXRPC_CALL_TX_LAST, are added also
to indicate when rx_top and tx_top point at the packets with the
LAST_PACKET bit set, indicating the end of the phase.
(4) Calls are queued on the socket 'receive queue' rather than packets.
This means that we don't need have to invent dummy packets to queue to
indicate abnormal/terminal states and we don't have to keep metadata
packets (such as ABORTs) around
(5) The offset and length of a (sub)packet's content are now passed to
the verify_packet security op. This is currently expected to decrypt
the packet in place and validate it.
However, there's now nowhere to store the revised offset and length of
the actual data within the decrypted blob (there may be a header and
padding to skip) because an sk_buff may represent multiple packets, so
a locate_data security op is added to retrieve these details from the
sk_buff content when needed.
(6) recvmsg() now has to handle jumbo subpackets, where each subpacket is
individually secured and needs to be individually decrypted. The code
to do this is broken out into rxrpc_recvmsg_data() and shared with the
kernel API. It now iterates over the call's receive buffer rather
than walking the socket receive queue.
Additional changes:
(1) The timers are condensed to a single timer that is set for the soonest
of three timeouts (delayed ACK generation, DATA retransmission and
call lifespan).
(2) Transmission of ACK and ABORT packets is effected immediately from
process-context socket ops/kernel API calls that cause them instead of
them being punted off to a background work item. The data_ready
handler still has to defer to the background, though.
(3) A shutdown op is added to the AF_RXRPC socket so that the AFS
filesystem can shut down the socket and flush its own work items
before closing the socket to deal with any in-progress service calls.
Future additional changes that will need to be considered:
(1) Make sure that a call doesn't hog the front of the queue by receiving
data from the network as fast as userspace is consuming it to the
exclusion of other calls.
(2) Transmit delayed ACKs from within recvmsg() when we've consumed
sufficiently more packets to avoid the background work item needing to
run.
Signed-off-by: David Howells <dhowells@redhat.com>
2016-09-08 10:10:12 +00:00
|
|
|
case RXRPC_CALL_CLIENT_AWAIT_REPLY:
|
2022-10-27 10:25:55 +00:00
|
|
|
if (reply_begun) {
|
|
|
|
|
rxrpc_set_call_state(call, RXRPC_CALL_CLIENT_RECV_REPLY);
|
|
|
|
|
trace_rxrpc_txqueue(call, rxrpc_txqueue_end);
|
|
|
|
|
break;
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
rxrpc_set_call_state(call, RXRPC_CALL_CLIENT_AWAIT_REPLY);
|
|
|
|
|
trace_rxrpc_txqueue(call, rxrpc_txqueue_await_reply);
|
rxrpc: Rewrite the data and ack handling code
Rewrite the data and ack handling code such that:
(1) Parsing of received ACK and ABORT packets and the distribution and the
filing of DATA packets happens entirely within the data_ready context
called from the UDP socket. This allows us to process and discard ACK
and ABORT packets much more quickly (they're no longer stashed on a
queue for a background thread to process).
(2) We avoid calling skb_clone(), pskb_pull() and pskb_trim(). We instead
keep track of the offset and length of the content of each packet in
the sk_buff metadata. This means we don't do any allocation in the
receive path.
(3) Jumbo DATA packet parsing is now done in data_ready context. Rather
than cloning the packet once for each subpacket and pulling/trimming
it, we file the packet multiple times with an annotation for each
indicating which subpacket is there. From that we can directly
calculate the offset and length.
(4) A call's receive queue can be accessed without taking locks (memory
barriers do have to be used, though).
(5) Incoming calls are set up from preallocated resources and immediately
made live. They can than have packets queued upon them and ACKs
generated. If insufficient resources exist, DATA packet #1 is given a
BUSY reply and other DATA packets are discarded).
(6) sk_buffs no longer take a ref on their parent call.
To make this work, the following changes are made:
(1) Each call's receive buffer is now a circular buffer of sk_buff
pointers (rxtx_buffer) rather than a number of sk_buff_heads spread
between the call and the socket. This permits each sk_buff to be in
the buffer multiple times. The receive buffer is reused for the
transmit buffer.
(2) A circular buffer of annotations (rxtx_annotations) is kept parallel
to the data buffer. Transmission phase annotations indicate whether a
buffered packet has been ACK'd or not and whether it needs
retransmission.
Receive phase annotations indicate whether a slot holds a whole packet
or a jumbo subpacket and, if the latter, which subpacket. They also
note whether the packet has been decrypted in place.
(3) DATA packet window tracking is much simplified. Each phase has just
two numbers representing the window (rx_hard_ack/rx_top and
tx_hard_ack/tx_top).
The hard_ack number is the sequence number before base of the window,
representing the last packet the other side says it has consumed.
hard_ack starts from 0 and the first packet is sequence number 1.
The top number is the sequence number of the highest-numbered packet
residing in the buffer. Packets between hard_ack+1 and top are
soft-ACK'd to indicate they've been received, but not yet consumed.
Four macros, before(), before_eq(), after() and after_eq() are added
to compare sequence numbers within the window. This allows for the
top of the window to wrap when the hard-ack sequence number gets close
to the limit.
Two flags, RXRPC_CALL_RX_LAST and RXRPC_CALL_TX_LAST, are added also
to indicate when rx_top and tx_top point at the packets with the
LAST_PACKET bit set, indicating the end of the phase.
(4) Calls are queued on the socket 'receive queue' rather than packets.
This means that we don't need have to invent dummy packets to queue to
indicate abnormal/terminal states and we don't have to keep metadata
packets (such as ABORTs) around
(5) The offset and length of a (sub)packet's content are now passed to
the verify_packet security op. This is currently expected to decrypt
the packet in place and validate it.
However, there's now nowhere to store the revised offset and length of
the actual data within the decrypted blob (there may be a header and
padding to skip) because an sk_buff may represent multiple packets, so
a locate_data security op is added to retrieve these details from the
sk_buff content when needed.
(6) recvmsg() now has to handle jumbo subpackets, where each subpacket is
individually secured and needs to be individually decrypted. The code
to do this is broken out into rxrpc_recvmsg_data() and shared with the
kernel API. It now iterates over the call's receive buffer rather
than walking the socket receive queue.
Additional changes:
(1) The timers are condensed to a single timer that is set for the soonest
of three timeouts (delayed ACK generation, DATA retransmission and
call lifespan).
(2) Transmission of ACK and ABORT packets is effected immediately from
process-context socket ops/kernel API calls that cause them instead of
them being punted off to a background work item. The data_ready
handler still has to defer to the background, though.
(3) A shutdown op is added to the AF_RXRPC socket so that the AFS
filesystem can shut down the socket and flush its own work items
before closing the socket to deal with any in-progress service calls.
Future additional changes that will need to be considered:
(1) Make sure that a call doesn't hog the front of the queue by receiving
data from the network as fast as userspace is consuming it to the
exclusion of other calls.
(2) Transmit delayed ACKs from within recvmsg() when we've consumed
sufficiently more packets to avoid the background work item needing to
run.
Signed-off-by: David Howells <dhowells@redhat.com>
2016-09-08 10:10:12 +00:00
|
|
|
break;
|
2016-09-23 11:39:22 +00:00
|
|
|
|
rxrpc: Rewrite the data and ack handling code
Rewrite the data and ack handling code such that:
(1) Parsing of received ACK and ABORT packets and the distribution and the
filing of DATA packets happens entirely within the data_ready context
called from the UDP socket. This allows us to process and discard ACK
and ABORT packets much more quickly (they're no longer stashed on a
queue for a background thread to process).
(2) We avoid calling skb_clone(), pskb_pull() and pskb_trim(). We instead
keep track of the offset and length of the content of each packet in
the sk_buff metadata. This means we don't do any allocation in the
receive path.
(3) Jumbo DATA packet parsing is now done in data_ready context. Rather
than cloning the packet once for each subpacket and pulling/trimming
it, we file the packet multiple times with an annotation for each
indicating which subpacket is there. From that we can directly
calculate the offset and length.
(4) A call's receive queue can be accessed without taking locks (memory
barriers do have to be used, though).
(5) Incoming calls are set up from preallocated resources and immediately
made live. They can than have packets queued upon them and ACKs
generated. If insufficient resources exist, DATA packet #1 is given a
BUSY reply and other DATA packets are discarded).
(6) sk_buffs no longer take a ref on their parent call.
To make this work, the following changes are made:
(1) Each call's receive buffer is now a circular buffer of sk_buff
pointers (rxtx_buffer) rather than a number of sk_buff_heads spread
between the call and the socket. This permits each sk_buff to be in
the buffer multiple times. The receive buffer is reused for the
transmit buffer.
(2) A circular buffer of annotations (rxtx_annotations) is kept parallel
to the data buffer. Transmission phase annotations indicate whether a
buffered packet has been ACK'd or not and whether it needs
retransmission.
Receive phase annotations indicate whether a slot holds a whole packet
or a jumbo subpacket and, if the latter, which subpacket. They also
note whether the packet has been decrypted in place.
(3) DATA packet window tracking is much simplified. Each phase has just
two numbers representing the window (rx_hard_ack/rx_top and
tx_hard_ack/tx_top).
The hard_ack number is the sequence number before base of the window,
representing the last packet the other side says it has consumed.
hard_ack starts from 0 and the first packet is sequence number 1.
The top number is the sequence number of the highest-numbered packet
residing in the buffer. Packets between hard_ack+1 and top are
soft-ACK'd to indicate they've been received, but not yet consumed.
Four macros, before(), before_eq(), after() and after_eq() are added
to compare sequence numbers within the window. This allows for the
top of the window to wrap when the hard-ack sequence number gets close
to the limit.
Two flags, RXRPC_CALL_RX_LAST and RXRPC_CALL_TX_LAST, are added also
to indicate when rx_top and tx_top point at the packets with the
LAST_PACKET bit set, indicating the end of the phase.
(4) Calls are queued on the socket 'receive queue' rather than packets.
This means that we don't need have to invent dummy packets to queue to
indicate abnormal/terminal states and we don't have to keep metadata
packets (such as ABORTs) around
(5) The offset and length of a (sub)packet's content are now passed to
the verify_packet security op. This is currently expected to decrypt
the packet in place and validate it.
However, there's now nowhere to store the revised offset and length of
the actual data within the decrypted blob (there may be a header and
padding to skip) because an sk_buff may represent multiple packets, so
a locate_data security op is added to retrieve these details from the
sk_buff content when needed.
(6) recvmsg() now has to handle jumbo subpackets, where each subpacket is
individually secured and needs to be individually decrypted. The code
to do this is broken out into rxrpc_recvmsg_data() and shared with the
kernel API. It now iterates over the call's receive buffer rather
than walking the socket receive queue.
Additional changes:
(1) The timers are condensed to a single timer that is set for the soonest
of three timeouts (delayed ACK generation, DATA retransmission and
call lifespan).
(2) Transmission of ACK and ABORT packets is effected immediately from
process-context socket ops/kernel API calls that cause them instead of
them being punted off to a background work item. The data_ready
handler still has to defer to the background, though.
(3) A shutdown op is added to the AF_RXRPC socket so that the AFS
filesystem can shut down the socket and flush its own work items
before closing the socket to deal with any in-progress service calls.
Future additional changes that will need to be considered:
(1) Make sure that a call doesn't hog the front of the queue by receiving
data from the network as fast as userspace is consuming it to the
exclusion of other calls.
(2) Transmit delayed ACKs from within recvmsg() when we've consumed
sufficiently more packets to avoid the background work item needing to
run.
Signed-off-by: David Howells <dhowells@redhat.com>
2016-09-08 10:10:12 +00:00
|
|
|
case RXRPC_CALL_SERVER_AWAIT_ACK:
|
2022-10-27 10:25:55 +00:00
|
|
|
rxrpc_call_completed(call);
|
|
|
|
|
trace_rxrpc_txqueue(call, rxrpc_txqueue_end);
|
rxrpc: Rewrite the data and ack handling code
Rewrite the data and ack handling code such that:
(1) Parsing of received ACK and ABORT packets and the distribution and the
filing of DATA packets happens entirely within the data_ready context
called from the UDP socket. This allows us to process and discard ACK
and ABORT packets much more quickly (they're no longer stashed on a
queue for a background thread to process).
(2) We avoid calling skb_clone(), pskb_pull() and pskb_trim(). We instead
keep track of the offset and length of the content of each packet in
the sk_buff metadata. This means we don't do any allocation in the
receive path.
(3) Jumbo DATA packet parsing is now done in data_ready context. Rather
than cloning the packet once for each subpacket and pulling/trimming
it, we file the packet multiple times with an annotation for each
indicating which subpacket is there. From that we can directly
calculate the offset and length.
(4) A call's receive queue can be accessed without taking locks (memory
barriers do have to be used, though).
(5) Incoming calls are set up from preallocated resources and immediately
made live. They can than have packets queued upon them and ACKs
generated. If insufficient resources exist, DATA packet #1 is given a
BUSY reply and other DATA packets are discarded).
(6) sk_buffs no longer take a ref on their parent call.
To make this work, the following changes are made:
(1) Each call's receive buffer is now a circular buffer of sk_buff
pointers (rxtx_buffer) rather than a number of sk_buff_heads spread
between the call and the socket. This permits each sk_buff to be in
the buffer multiple times. The receive buffer is reused for the
transmit buffer.
(2) A circular buffer of annotations (rxtx_annotations) is kept parallel
to the data buffer. Transmission phase annotations indicate whether a
buffered packet has been ACK'd or not and whether it needs
retransmission.
Receive phase annotations indicate whether a slot holds a whole packet
or a jumbo subpacket and, if the latter, which subpacket. They also
note whether the packet has been decrypted in place.
(3) DATA packet window tracking is much simplified. Each phase has just
two numbers representing the window (rx_hard_ack/rx_top and
tx_hard_ack/tx_top).
The hard_ack number is the sequence number before base of the window,
representing the last packet the other side says it has consumed.
hard_ack starts from 0 and the first packet is sequence number 1.
The top number is the sequence number of the highest-numbered packet
residing in the buffer. Packets between hard_ack+1 and top are
soft-ACK'd to indicate they've been received, but not yet consumed.
Four macros, before(), before_eq(), after() and after_eq() are added
to compare sequence numbers within the window. This allows for the
top of the window to wrap when the hard-ack sequence number gets close
to the limit.
Two flags, RXRPC_CALL_RX_LAST and RXRPC_CALL_TX_LAST, are added also
to indicate when rx_top and tx_top point at the packets with the
LAST_PACKET bit set, indicating the end of the phase.
(4) Calls are queued on the socket 'receive queue' rather than packets.
This means that we don't need have to invent dummy packets to queue to
indicate abnormal/terminal states and we don't have to keep metadata
packets (such as ABORTs) around
(5) The offset and length of a (sub)packet's content are now passed to
the verify_packet security op. This is currently expected to decrypt
the packet in place and validate it.
However, there's now nowhere to store the revised offset and length of
the actual data within the decrypted blob (there may be a header and
padding to skip) because an sk_buff may represent multiple packets, so
a locate_data security op is added to retrieve these details from the
sk_buff content when needed.
(6) recvmsg() now has to handle jumbo subpackets, where each subpacket is
individually secured and needs to be individually decrypted. The code
to do this is broken out into rxrpc_recvmsg_data() and shared with the
kernel API. It now iterates over the call's receive buffer rather
than walking the socket receive queue.
Additional changes:
(1) The timers are condensed to a single timer that is set for the soonest
of three timeouts (delayed ACK generation, DATA retransmission and
call lifespan).
(2) Transmission of ACK and ABORT packets is effected immediately from
process-context socket ops/kernel API calls that cause them instead of
them being punted off to a background work item. The data_ready
handler still has to defer to the background, though.
(3) A shutdown op is added to the AF_RXRPC socket so that the AFS
filesystem can shut down the socket and flush its own work items
before closing the socket to deal with any in-progress service calls.
Future additional changes that will need to be considered:
(1) Make sure that a call doesn't hog the front of the queue by receiving
data from the network as fast as userspace is consuming it to the
exclusion of other calls.
(2) Transmit delayed ACKs from within recvmsg() when we've consumed
sufficiently more packets to avoid the background work item needing to
run.
Signed-off-by: David Howells <dhowells@redhat.com>
2016-09-08 10:10:12 +00:00
|
|
|
break;
|
2016-09-23 11:39:22 +00:00
|
|
|
|
|
|
|
|
default:
|
2022-10-27 10:25:55 +00:00
|
|
|
kdebug("end_tx %s", rxrpc_call_states[__rxrpc_call_state(call)]);
|
|
|
|
|
rxrpc_proto_abort(call, call->tx_top, abort_why);
|
|
|
|
|
break;
|
2007-04-26 22:48:28 +00:00
|
|
|
}
|
2016-09-23 11:39:22 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
|
|
/*
|
|
|
|
|
* Begin the reply reception phase of a call.
|
|
|
|
|
*/
|
|
|
|
|
static bool rxrpc_receiving_reply(struct rxrpc_call *call)
|
|
|
|
|
{
|
2016-09-24 17:05:26 +00:00
|
|
|
struct rxrpc_ack_summary summary = { 0 };
|
2016-09-23 11:39:22 +00:00
|
|
|
rxrpc_seq_t top = READ_ONCE(call->tx_top);
|
|
|
|
|
|
2016-09-24 17:05:27 +00:00
|
|
|
if (call->ackr_reason) {
|
2024-01-30 21:37:16 +00:00
|
|
|
call->delay_ack_at = KTIME_MAX;
|
|
|
|
|
trace_rxrpc_timer_can(call, rxrpc_timer_trace_delayed_ack);
|
2016-09-24 17:05:27 +00:00
|
|
|
}
|
|
|
|
|
|
2016-09-23 11:39:22 +00:00
|
|
|
if (!test_bit(RXRPC_CALL_TX_LAST, &call->flags)) {
|
2024-12-04 07:47:02 +00:00
|
|
|
if (!rxrpc_rotate_tx_window(call, top, &summary)) {
|
2022-10-06 20:45:42 +00:00
|
|
|
rxrpc_proto_abort(call, top, rxrpc_eproto_early_reply);
|
2018-10-08 14:46:01 +00:00
|
|
|
return false;
|
|
|
|
|
}
|
2016-09-23 11:39:22 +00:00
|
|
|
}
|
2022-10-06 20:45:42 +00:00
|
|
|
|
|
|
|
|
rxrpc_end_tx_phase(call, true, rxrpc_eproto_unexpected_reply);
|
|
|
|
|
return true;
|
rxrpc: Rewrite the data and ack handling code
Rewrite the data and ack handling code such that:
(1) Parsing of received ACK and ABORT packets and the distribution and the
filing of DATA packets happens entirely within the data_ready context
called from the UDP socket. This allows us to process and discard ACK
and ABORT packets much more quickly (they're no longer stashed on a
queue for a background thread to process).
(2) We avoid calling skb_clone(), pskb_pull() and pskb_trim(). We instead
keep track of the offset and length of the content of each packet in
the sk_buff metadata. This means we don't do any allocation in the
receive path.
(3) Jumbo DATA packet parsing is now done in data_ready context. Rather
than cloning the packet once for each subpacket and pulling/trimming
it, we file the packet multiple times with an annotation for each
indicating which subpacket is there. From that we can directly
calculate the offset and length.
(4) A call's receive queue can be accessed without taking locks (memory
barriers do have to be used, though).
(5) Incoming calls are set up from preallocated resources and immediately
made live. They can than have packets queued upon them and ACKs
generated. If insufficient resources exist, DATA packet #1 is given a
BUSY reply and other DATA packets are discarded).
(6) sk_buffs no longer take a ref on their parent call.
To make this work, the following changes are made:
(1) Each call's receive buffer is now a circular buffer of sk_buff
pointers (rxtx_buffer) rather than a number of sk_buff_heads spread
between the call and the socket. This permits each sk_buff to be in
the buffer multiple times. The receive buffer is reused for the
transmit buffer.
(2) A circular buffer of annotations (rxtx_annotations) is kept parallel
to the data buffer. Transmission phase annotations indicate whether a
buffered packet has been ACK'd or not and whether it needs
retransmission.
Receive phase annotations indicate whether a slot holds a whole packet
or a jumbo subpacket and, if the latter, which subpacket. They also
note whether the packet has been decrypted in place.
(3) DATA packet window tracking is much simplified. Each phase has just
two numbers representing the window (rx_hard_ack/rx_top and
tx_hard_ack/tx_top).
The hard_ack number is the sequence number before base of the window,
representing the last packet the other side says it has consumed.
hard_ack starts from 0 and the first packet is sequence number 1.
The top number is the sequence number of the highest-numbered packet
residing in the buffer. Packets between hard_ack+1 and top are
soft-ACK'd to indicate they've been received, but not yet consumed.
Four macros, before(), before_eq(), after() and after_eq() are added
to compare sequence numbers within the window. This allows for the
top of the window to wrap when the hard-ack sequence number gets close
to the limit.
Two flags, RXRPC_CALL_RX_LAST and RXRPC_CALL_TX_LAST, are added also
to indicate when rx_top and tx_top point at the packets with the
LAST_PACKET bit set, indicating the end of the phase.
(4) Calls are queued on the socket 'receive queue' rather than packets.
This means that we don't need have to invent dummy packets to queue to
indicate abnormal/terminal states and we don't have to keep metadata
packets (such as ABORTs) around
(5) The offset and length of a (sub)packet's content are now passed to
the verify_packet security op. This is currently expected to decrypt
the packet in place and validate it.
However, there's now nowhere to store the revised offset and length of
the actual data within the decrypted blob (there may be a header and
padding to skip) because an sk_buff may represent multiple packets, so
a locate_data security op is added to retrieve these details from the
sk_buff content when needed.
(6) recvmsg() now has to handle jumbo subpackets, where each subpacket is
individually secured and needs to be individually decrypted. The code
to do this is broken out into rxrpc_recvmsg_data() and shared with the
kernel API. It now iterates over the call's receive buffer rather
than walking the socket receive queue.
Additional changes:
(1) The timers are condensed to a single timer that is set for the soonest
of three timeouts (delayed ACK generation, DATA retransmission and
call lifespan).
(2) Transmission of ACK and ABORT packets is effected immediately from
process-context socket ops/kernel API calls that cause them instead of
them being punted off to a background work item. The data_ready
handler still has to defer to the background, though.
(3) A shutdown op is added to the AF_RXRPC socket so that the AFS
filesystem can shut down the socket and flush its own work items
before closing the socket to deal with any in-progress service calls.
Future additional changes that will need to be considered:
(1) Make sure that a call doesn't hog the front of the queue by receiving
data from the network as fast as userspace is consuming it to the
exclusion of other calls.
(2) Transmit delayed ACKs from within recvmsg() when we've consumed
sufficiently more packets to avoid the background work item needing to
run.
Signed-off-by: David Howells <dhowells@redhat.com>
2016-09-08 10:10:12 +00:00
|
|
|
}
|
|
|
|
|
|
2022-10-26 22:43:00 +00:00
|
|
|
/*
|
|
|
|
|
* End the packet reception phase.
|
|
|
|
|
*/
|
|
|
|
|
static void rxrpc_end_rx_phase(struct rxrpc_call *call, rxrpc_serial_t serial)
|
|
|
|
|
{
|
|
|
|
|
rxrpc_seq_t whigh = READ_ONCE(call->rx_highest_seq);
|
|
|
|
|
|
2022-10-27 10:25:55 +00:00
|
|
|
_enter("%d,%s", call->debug_id, rxrpc_call_states[__rxrpc_call_state(call)]);
|
2022-10-26 22:43:00 +00:00
|
|
|
|
|
|
|
|
trace_rxrpc_receive(call, rxrpc_receive_end, 0, whigh);
|
|
|
|
|
|
2022-10-27 10:25:55 +00:00
|
|
|
switch (__rxrpc_call_state(call)) {
|
2022-10-26 22:43:00 +00:00
|
|
|
case RXRPC_CALL_CLIENT_RECV_REPLY:
|
2022-10-27 10:25:55 +00:00
|
|
|
rxrpc_propose_delay_ACK(call, serial, rxrpc_propose_ack_terminal_ack);
|
|
|
|
|
rxrpc_call_completed(call);
|
2022-10-26 22:43:00 +00:00
|
|
|
break;
|
|
|
|
|
|
|
|
|
|
case RXRPC_CALL_SERVER_RECV_REQUEST:
|
2022-10-27 10:25:55 +00:00
|
|
|
rxrpc_set_call_state(call, RXRPC_CALL_SERVER_ACK_REQUEST);
|
2024-01-30 21:37:16 +00:00
|
|
|
call->expect_req_by = KTIME_MAX;
|
2022-10-27 10:25:55 +00:00
|
|
|
rxrpc_propose_delay_ACK(call, serial, rxrpc_propose_ack_processing_op);
|
2022-10-26 22:43:00 +00:00
|
|
|
break;
|
2022-10-27 10:25:55 +00:00
|
|
|
|
2022-10-26 22:43:00 +00:00
|
|
|
default:
|
|
|
|
|
break;
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
|
2022-08-27 13:27:56 +00:00
|
|
|
static void rxrpc_input_update_ack_window(struct rxrpc_call *call,
|
|
|
|
|
rxrpc_seq_t window, rxrpc_seq_t wtop)
|
|
|
|
|
{
|
2022-10-17 10:44:22 +00:00
|
|
|
call->ackr_window = window;
|
|
|
|
|
call->ackr_wtop = wtop;
|
2022-08-27 13:27:56 +00:00
|
|
|
}
|
|
|
|
|
|
rxrpc: Rewrite the data and ack handling code
Rewrite the data and ack handling code such that:
(1) Parsing of received ACK and ABORT packets and the distribution and the
filing of DATA packets happens entirely within the data_ready context
called from the UDP socket. This allows us to process and discard ACK
and ABORT packets much more quickly (they're no longer stashed on a
queue for a background thread to process).
(2) We avoid calling skb_clone(), pskb_pull() and pskb_trim(). We instead
keep track of the offset and length of the content of each packet in
the sk_buff metadata. This means we don't do any allocation in the
receive path.
(3) Jumbo DATA packet parsing is now done in data_ready context. Rather
than cloning the packet once for each subpacket and pulling/trimming
it, we file the packet multiple times with an annotation for each
indicating which subpacket is there. From that we can directly
calculate the offset and length.
(4) A call's receive queue can be accessed without taking locks (memory
barriers do have to be used, though).
(5) Incoming calls are set up from preallocated resources and immediately
made live. They can than have packets queued upon them and ACKs
generated. If insufficient resources exist, DATA packet #1 is given a
BUSY reply and other DATA packets are discarded).
(6) sk_buffs no longer take a ref on their parent call.
To make this work, the following changes are made:
(1) Each call's receive buffer is now a circular buffer of sk_buff
pointers (rxtx_buffer) rather than a number of sk_buff_heads spread
between the call and the socket. This permits each sk_buff to be in
the buffer multiple times. The receive buffer is reused for the
transmit buffer.
(2) A circular buffer of annotations (rxtx_annotations) is kept parallel
to the data buffer. Transmission phase annotations indicate whether a
buffered packet has been ACK'd or not and whether it needs
retransmission.
Receive phase annotations indicate whether a slot holds a whole packet
or a jumbo subpacket and, if the latter, which subpacket. They also
note whether the packet has been decrypted in place.
(3) DATA packet window tracking is much simplified. Each phase has just
two numbers representing the window (rx_hard_ack/rx_top and
tx_hard_ack/tx_top).
The hard_ack number is the sequence number before base of the window,
representing the last packet the other side says it has consumed.
hard_ack starts from 0 and the first packet is sequence number 1.
The top number is the sequence number of the highest-numbered packet
residing in the buffer. Packets between hard_ack+1 and top are
soft-ACK'd to indicate they've been received, but not yet consumed.
Four macros, before(), before_eq(), after() and after_eq() are added
to compare sequence numbers within the window. This allows for the
top of the window to wrap when the hard-ack sequence number gets close
to the limit.
Two flags, RXRPC_CALL_RX_LAST and RXRPC_CALL_TX_LAST, are added also
to indicate when rx_top and tx_top point at the packets with the
LAST_PACKET bit set, indicating the end of the phase.
(4) Calls are queued on the socket 'receive queue' rather than packets.
This means that we don't need have to invent dummy packets to queue to
indicate abnormal/terminal states and we don't have to keep metadata
packets (such as ABORTs) around
(5) The offset and length of a (sub)packet's content are now passed to
the verify_packet security op. This is currently expected to decrypt
the packet in place and validate it.
However, there's now nowhere to store the revised offset and length of
the actual data within the decrypted blob (there may be a header and
padding to skip) because an sk_buff may represent multiple packets, so
a locate_data security op is added to retrieve these details from the
sk_buff content when needed.
(6) recvmsg() now has to handle jumbo subpackets, where each subpacket is
individually secured and needs to be individually decrypted. The code
to do this is broken out into rxrpc_recvmsg_data() and shared with the
kernel API. It now iterates over the call's receive buffer rather
than walking the socket receive queue.
Additional changes:
(1) The timers are condensed to a single timer that is set for the soonest
of three timeouts (delayed ACK generation, DATA retransmission and
call lifespan).
(2) Transmission of ACK and ABORT packets is effected immediately from
process-context socket ops/kernel API calls that cause them instead of
them being punted off to a background work item. The data_ready
handler still has to defer to the background, though.
(3) A shutdown op is added to the AF_RXRPC socket so that the AFS
filesystem can shut down the socket and flush its own work items
before closing the socket to deal with any in-progress service calls.
Future additional changes that will need to be considered:
(1) Make sure that a call doesn't hog the front of the queue by receiving
data from the network as fast as userspace is consuming it to the
exclusion of other calls.
(2) Transmit delayed ACKs from within recvmsg() when we've consumed
sufficiently more packets to avoid the background work item needing to
run.
Signed-off-by: David Howells <dhowells@redhat.com>
2016-09-08 10:10:12 +00:00
|
|
|
/*
|
2022-08-27 13:27:56 +00:00
|
|
|
* Push a DATA packet onto the Rx queue.
|
|
|
|
|
*/
|
|
|
|
|
static void rxrpc_input_queue_data(struct rxrpc_call *call, struct sk_buff *skb,
|
|
|
|
|
rxrpc_seq_t window, rxrpc_seq_t wtop,
|
|
|
|
|
enum rxrpc_receive_trace why)
|
|
|
|
|
{
|
|
|
|
|
struct rxrpc_skb_priv *sp = rxrpc_skb(skb);
|
|
|
|
|
bool last = sp->hdr.flags & RXRPC_LAST_PACKET;
|
|
|
|
|
|
2024-12-04 07:47:01 +00:00
|
|
|
skb_queue_tail(&call->recvmsg_queue, skb);
|
2022-08-27 13:27:56 +00:00
|
|
|
rxrpc_input_update_ack_window(call, window, wtop);
|
|
|
|
|
trace_rxrpc_receive(call, last ? why + 1 : why, sp->hdr.serial, sp->hdr.seq);
|
2022-10-26 22:43:00 +00:00
|
|
|
if (last)
|
|
|
|
|
rxrpc_end_rx_phase(call, sp->hdr.serial);
|
2022-08-27 13:27:56 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
|
|
/*
|
|
|
|
|
* Process a DATA packet.
|
rxrpc: Rewrite the data and ack handling code
Rewrite the data and ack handling code such that:
(1) Parsing of received ACK and ABORT packets and the distribution and the
filing of DATA packets happens entirely within the data_ready context
called from the UDP socket. This allows us to process and discard ACK
and ABORT packets much more quickly (they're no longer stashed on a
queue for a background thread to process).
(2) We avoid calling skb_clone(), pskb_pull() and pskb_trim(). We instead
keep track of the offset and length of the content of each packet in
the sk_buff metadata. This means we don't do any allocation in the
receive path.
(3) Jumbo DATA packet parsing is now done in data_ready context. Rather
than cloning the packet once for each subpacket and pulling/trimming
it, we file the packet multiple times with an annotation for each
indicating which subpacket is there. From that we can directly
calculate the offset and length.
(4) A call's receive queue can be accessed without taking locks (memory
barriers do have to be used, though).
(5) Incoming calls are set up from preallocated resources and immediately
made live. They can than have packets queued upon them and ACKs
generated. If insufficient resources exist, DATA packet #1 is given a
BUSY reply and other DATA packets are discarded).
(6) sk_buffs no longer take a ref on their parent call.
To make this work, the following changes are made:
(1) Each call's receive buffer is now a circular buffer of sk_buff
pointers (rxtx_buffer) rather than a number of sk_buff_heads spread
between the call and the socket. This permits each sk_buff to be in
the buffer multiple times. The receive buffer is reused for the
transmit buffer.
(2) A circular buffer of annotations (rxtx_annotations) is kept parallel
to the data buffer. Transmission phase annotations indicate whether a
buffered packet has been ACK'd or not and whether it needs
retransmission.
Receive phase annotations indicate whether a slot holds a whole packet
or a jumbo subpacket and, if the latter, which subpacket. They also
note whether the packet has been decrypted in place.
(3) DATA packet window tracking is much simplified. Each phase has just
two numbers representing the window (rx_hard_ack/rx_top and
tx_hard_ack/tx_top).
The hard_ack number is the sequence number before base of the window,
representing the last packet the other side says it has consumed.
hard_ack starts from 0 and the first packet is sequence number 1.
The top number is the sequence number of the highest-numbered packet
residing in the buffer. Packets between hard_ack+1 and top are
soft-ACK'd to indicate they've been received, but not yet consumed.
Four macros, before(), before_eq(), after() and after_eq() are added
to compare sequence numbers within the window. This allows for the
top of the window to wrap when the hard-ack sequence number gets close
to the limit.
Two flags, RXRPC_CALL_RX_LAST and RXRPC_CALL_TX_LAST, are added also
to indicate when rx_top and tx_top point at the packets with the
LAST_PACKET bit set, indicating the end of the phase.
(4) Calls are queued on the socket 'receive queue' rather than packets.
This means that we don't need have to invent dummy packets to queue to
indicate abnormal/terminal states and we don't have to keep metadata
packets (such as ABORTs) around
(5) The offset and length of a (sub)packet's content are now passed to
the verify_packet security op. This is currently expected to decrypt
the packet in place and validate it.
However, there's now nowhere to store the revised offset and length of
the actual data within the decrypted blob (there may be a header and
padding to skip) because an sk_buff may represent multiple packets, so
a locate_data security op is added to retrieve these details from the
sk_buff content when needed.
(6) recvmsg() now has to handle jumbo subpackets, where each subpacket is
individually secured and needs to be individually decrypted. The code
to do this is broken out into rxrpc_recvmsg_data() and shared with the
kernel API. It now iterates over the call's receive buffer rather
than walking the socket receive queue.
Additional changes:
(1) The timers are condensed to a single timer that is set for the soonest
of three timeouts (delayed ACK generation, DATA retransmission and
call lifespan).
(2) Transmission of ACK and ABORT packets is effected immediately from
process-context socket ops/kernel API calls that cause them instead of
them being punted off to a background work item. The data_ready
handler still has to defer to the background, though.
(3) A shutdown op is added to the AF_RXRPC socket so that the AFS
filesystem can shut down the socket and flush its own work items
before closing the socket to deal with any in-progress service calls.
Future additional changes that will need to be considered:
(1) Make sure that a call doesn't hog the front of the queue by receiving
data from the network as fast as userspace is consuming it to the
exclusion of other calls.
(2) Transmit delayed ACKs from within recvmsg() when we've consumed
sufficiently more packets to avoid the background work item needing to
run.
Signed-off-by: David Howells <dhowells@redhat.com>
2016-09-08 10:10:12 +00:00
|
|
|
*/
|
2022-10-06 14:43:51 +00:00
|
|
|
static void rxrpc_input_data_one(struct rxrpc_call *call, struct sk_buff *skb,
|
2024-05-03 15:07:40 +00:00
|
|
|
bool *_notify, rxrpc_serial_t *_ack_serial, int *_ack_reason)
|
rxrpc: Rewrite the data and ack handling code
Rewrite the data and ack handling code such that:
(1) Parsing of received ACK and ABORT packets and the distribution and the
filing of DATA packets happens entirely within the data_ready context
called from the UDP socket. This allows us to process and discard ACK
and ABORT packets much more quickly (they're no longer stashed on a
queue for a background thread to process).
(2) We avoid calling skb_clone(), pskb_pull() and pskb_trim(). We instead
keep track of the offset and length of the content of each packet in
the sk_buff metadata. This means we don't do any allocation in the
receive path.
(3) Jumbo DATA packet parsing is now done in data_ready context. Rather
than cloning the packet once for each subpacket and pulling/trimming
it, we file the packet multiple times with an annotation for each
indicating which subpacket is there. From that we can directly
calculate the offset and length.
(4) A call's receive queue can be accessed without taking locks (memory
barriers do have to be used, though).
(5) Incoming calls are set up from preallocated resources and immediately
made live. They can than have packets queued upon them and ACKs
generated. If insufficient resources exist, DATA packet #1 is given a
BUSY reply and other DATA packets are discarded).
(6) sk_buffs no longer take a ref on their parent call.
To make this work, the following changes are made:
(1) Each call's receive buffer is now a circular buffer of sk_buff
pointers (rxtx_buffer) rather than a number of sk_buff_heads spread
between the call and the socket. This permits each sk_buff to be in
the buffer multiple times. The receive buffer is reused for the
transmit buffer.
(2) A circular buffer of annotations (rxtx_annotations) is kept parallel
to the data buffer. Transmission phase annotations indicate whether a
buffered packet has been ACK'd or not and whether it needs
retransmission.
Receive phase annotations indicate whether a slot holds a whole packet
or a jumbo subpacket and, if the latter, which subpacket. They also
note whether the packet has been decrypted in place.
(3) DATA packet window tracking is much simplified. Each phase has just
two numbers representing the window (rx_hard_ack/rx_top and
tx_hard_ack/tx_top).
The hard_ack number is the sequence number before base of the window,
representing the last packet the other side says it has consumed.
hard_ack starts from 0 and the first packet is sequence number 1.
The top number is the sequence number of the highest-numbered packet
residing in the buffer. Packets between hard_ack+1 and top are
soft-ACK'd to indicate they've been received, but not yet consumed.
Four macros, before(), before_eq(), after() and after_eq() are added
to compare sequence numbers within the window. This allows for the
top of the window to wrap when the hard-ack sequence number gets close
to the limit.
Two flags, RXRPC_CALL_RX_LAST and RXRPC_CALL_TX_LAST, are added also
to indicate when rx_top and tx_top point at the packets with the
LAST_PACKET bit set, indicating the end of the phase.
(4) Calls are queued on the socket 'receive queue' rather than packets.
This means that we don't need have to invent dummy packets to queue to
indicate abnormal/terminal states and we don't have to keep metadata
packets (such as ABORTs) around
(5) The offset and length of a (sub)packet's content are now passed to
the verify_packet security op. This is currently expected to decrypt
the packet in place and validate it.
However, there's now nowhere to store the revised offset and length of
the actual data within the decrypted blob (there may be a header and
padding to skip) because an sk_buff may represent multiple packets, so
a locate_data security op is added to retrieve these details from the
sk_buff content when needed.
(6) recvmsg() now has to handle jumbo subpackets, where each subpacket is
individually secured and needs to be individually decrypted. The code
to do this is broken out into rxrpc_recvmsg_data() and shared with the
kernel API. It now iterates over the call's receive buffer rather
than walking the socket receive queue.
Additional changes:
(1) The timers are condensed to a single timer that is set for the soonest
of three timeouts (delayed ACK generation, DATA retransmission and
call lifespan).
(2) Transmission of ACK and ABORT packets is effected immediately from
process-context socket ops/kernel API calls that cause them instead of
them being punted off to a background work item. The data_ready
handler still has to defer to the background, though.
(3) A shutdown op is added to the AF_RXRPC socket so that the AFS
filesystem can shut down the socket and flush its own work items
before closing the socket to deal with any in-progress service calls.
Future additional changes that will need to be considered:
(1) Make sure that a call doesn't hog the front of the queue by receiving
data from the network as fast as userspace is consuming it to the
exclusion of other calls.
(2) Transmit delayed ACKs from within recvmsg() when we've consumed
sufficiently more packets to avoid the background work item needing to
run.
Signed-off-by: David Howells <dhowells@redhat.com>
2016-09-08 10:10:12 +00:00
|
|
|
{
|
|
|
|
|
struct rxrpc_skb_priv *sp = rxrpc_skb(skb);
|
2022-08-27 13:27:56 +00:00
|
|
|
struct sk_buff *oos;
|
2022-10-07 16:44:39 +00:00
|
|
|
rxrpc_serial_t serial = sp->hdr.serial;
|
2022-10-16 07:01:32 +00:00
|
|
|
unsigned int sack = call->ackr_sack_base;
|
2022-10-17 10:44:22 +00:00
|
|
|
rxrpc_seq_t window = call->ackr_window;
|
|
|
|
|
rxrpc_seq_t wtop = call->ackr_wtop;
|
2022-08-27 13:27:56 +00:00
|
|
|
rxrpc_seq_t wlimit = window + call->rx_winsize - 1;
|
|
|
|
|
rxrpc_seq_t seq = sp->hdr.seq;
|
2022-10-07 16:44:39 +00:00
|
|
|
bool last = sp->hdr.flags & RXRPC_LAST_PACKET;
|
2022-08-27 13:27:56 +00:00
|
|
|
int ack_reason = -1;
|
rxrpc: Rewrite the data and ack handling code
Rewrite the data and ack handling code such that:
(1) Parsing of received ACK and ABORT packets and the distribution and the
filing of DATA packets happens entirely within the data_ready context
called from the UDP socket. This allows us to process and discard ACK
and ABORT packets much more quickly (they're no longer stashed on a
queue for a background thread to process).
(2) We avoid calling skb_clone(), pskb_pull() and pskb_trim(). We instead
keep track of the offset and length of the content of each packet in
the sk_buff metadata. This means we don't do any allocation in the
receive path.
(3) Jumbo DATA packet parsing is now done in data_ready context. Rather
than cloning the packet once for each subpacket and pulling/trimming
it, we file the packet multiple times with an annotation for each
indicating which subpacket is there. From that we can directly
calculate the offset and length.
(4) A call's receive queue can be accessed without taking locks (memory
barriers do have to be used, though).
(5) Incoming calls are set up from preallocated resources and immediately
made live. They can than have packets queued upon them and ACKs
generated. If insufficient resources exist, DATA packet #1 is given a
BUSY reply and other DATA packets are discarded).
(6) sk_buffs no longer take a ref on their parent call.
To make this work, the following changes are made:
(1) Each call's receive buffer is now a circular buffer of sk_buff
pointers (rxtx_buffer) rather than a number of sk_buff_heads spread
between the call and the socket. This permits each sk_buff to be in
the buffer multiple times. The receive buffer is reused for the
transmit buffer.
(2) A circular buffer of annotations (rxtx_annotations) is kept parallel
to the data buffer. Transmission phase annotations indicate whether a
buffered packet has been ACK'd or not and whether it needs
retransmission.
Receive phase annotations indicate whether a slot holds a whole packet
or a jumbo subpacket and, if the latter, which subpacket. They also
note whether the packet has been decrypted in place.
(3) DATA packet window tracking is much simplified. Each phase has just
two numbers representing the window (rx_hard_ack/rx_top and
tx_hard_ack/tx_top).
The hard_ack number is the sequence number before base of the window,
representing the last packet the other side says it has consumed.
hard_ack starts from 0 and the first packet is sequence number 1.
The top number is the sequence number of the highest-numbered packet
residing in the buffer. Packets between hard_ack+1 and top are
soft-ACK'd to indicate they've been received, but not yet consumed.
Four macros, before(), before_eq(), after() and after_eq() are added
to compare sequence numbers within the window. This allows for the
top of the window to wrap when the hard-ack sequence number gets close
to the limit.
Two flags, RXRPC_CALL_RX_LAST and RXRPC_CALL_TX_LAST, are added also
to indicate when rx_top and tx_top point at the packets with the
LAST_PACKET bit set, indicating the end of the phase.
(4) Calls are queued on the socket 'receive queue' rather than packets.
This means that we don't need have to invent dummy packets to queue to
indicate abnormal/terminal states and we don't have to keep metadata
packets (such as ABORTs) around
(5) The offset and length of a (sub)packet's content are now passed to
the verify_packet security op. This is currently expected to decrypt
the packet in place and validate it.
However, there's now nowhere to store the revised offset and length of
the actual data within the decrypted blob (there may be a header and
padding to skip) because an sk_buff may represent multiple packets, so
a locate_data security op is added to retrieve these details from the
sk_buff content when needed.
(6) recvmsg() now has to handle jumbo subpackets, where each subpacket is
individually secured and needs to be individually decrypted. The code
to do this is broken out into rxrpc_recvmsg_data() and shared with the
kernel API. It now iterates over the call's receive buffer rather
than walking the socket receive queue.
Additional changes:
(1) The timers are condensed to a single timer that is set for the soonest
of three timeouts (delayed ACK generation, DATA retransmission and
call lifespan).
(2) Transmission of ACK and ABORT packets is effected immediately from
process-context socket ops/kernel API calls that cause them instead of
them being punted off to a background work item. The data_ready
handler still has to defer to the background, though.
(3) A shutdown op is added to the AF_RXRPC socket so that the AFS
filesystem can shut down the socket and flush its own work items
before closing the socket to deal with any in-progress service calls.
Future additional changes that will need to be considered:
(1) Make sure that a call doesn't hog the front of the queue by receiving
data from the network as fast as userspace is consuming it to the
exclusion of other calls.
(2) Transmit delayed ACKs from within recvmsg() when we've consumed
sufficiently more packets to avoid the background work item needing to
run.
Signed-off-by: David Howells <dhowells@redhat.com>
2016-09-08 10:10:12 +00:00
|
|
|
|
2022-10-07 16:44:39 +00:00
|
|
|
rxrpc_inc_stat(call->rxnet, stat_rx_data);
|
|
|
|
|
if (sp->hdr.flags & RXRPC_REQUEST_ACK)
|
|
|
|
|
rxrpc_inc_stat(call->rxnet, stat_rx_data_reqack);
|
|
|
|
|
if (sp->hdr.flags & RXRPC_JUMBO_PACKET)
|
|
|
|
|
rxrpc_inc_stat(call->rxnet, stat_rx_data_jumbo);
|
2019-08-19 08:25:37 +00:00
|
|
|
|
2022-10-07 16:44:39 +00:00
|
|
|
if (last) {
|
2022-08-27 13:27:56 +00:00
|
|
|
if (test_and_set_bit(RXRPC_CALL_RX_LAST, &call->flags) &&
|
2022-10-06 20:45:42 +00:00
|
|
|
seq + 1 != wtop)
|
|
|
|
|
return rxrpc_proto_abort(call, seq, rxrpc_eproto_different_last);
|
2022-10-07 16:44:39 +00:00
|
|
|
} else {
|
|
|
|
|
if (test_bit(RXRPC_CALL_RX_LAST, &call->flags) &&
|
2022-08-27 13:27:56 +00:00
|
|
|
after_eq(seq, wtop)) {
|
|
|
|
|
pr_warn("Packet beyond last: c=%x q=%x window=%x-%x wlimit=%x\n",
|
|
|
|
|
call->debug_id, seq, window, wtop, wlimit);
|
2022-10-06 20:45:42 +00:00
|
|
|
return rxrpc_proto_abort(call, seq, rxrpc_eproto_data_after_last);
|
2022-10-07 16:44:39 +00:00
|
|
|
}
|
2019-08-19 08:25:37 +00:00
|
|
|
}
|
rxrpc: Rewrite the data and ack handling code
Rewrite the data and ack handling code such that:
(1) Parsing of received ACK and ABORT packets and the distribution and the
filing of DATA packets happens entirely within the data_ready context
called from the UDP socket. This allows us to process and discard ACK
and ABORT packets much more quickly (they're no longer stashed on a
queue for a background thread to process).
(2) We avoid calling skb_clone(), pskb_pull() and pskb_trim(). We instead
keep track of the offset and length of the content of each packet in
the sk_buff metadata. This means we don't do any allocation in the
receive path.
(3) Jumbo DATA packet parsing is now done in data_ready context. Rather
than cloning the packet once for each subpacket and pulling/trimming
it, we file the packet multiple times with an annotation for each
indicating which subpacket is there. From that we can directly
calculate the offset and length.
(4) A call's receive queue can be accessed without taking locks (memory
barriers do have to be used, though).
(5) Incoming calls are set up from preallocated resources and immediately
made live. They can than have packets queued upon them and ACKs
generated. If insufficient resources exist, DATA packet #1 is given a
BUSY reply and other DATA packets are discarded).
(6) sk_buffs no longer take a ref on their parent call.
To make this work, the following changes are made:
(1) Each call's receive buffer is now a circular buffer of sk_buff
pointers (rxtx_buffer) rather than a number of sk_buff_heads spread
between the call and the socket. This permits each sk_buff to be in
the buffer multiple times. The receive buffer is reused for the
transmit buffer.
(2) A circular buffer of annotations (rxtx_annotations) is kept parallel
to the data buffer. Transmission phase annotations indicate whether a
buffered packet has been ACK'd or not and whether it needs
retransmission.
Receive phase annotations indicate whether a slot holds a whole packet
or a jumbo subpacket and, if the latter, which subpacket. They also
note whether the packet has been decrypted in place.
(3) DATA packet window tracking is much simplified. Each phase has just
two numbers representing the window (rx_hard_ack/rx_top and
tx_hard_ack/tx_top).
The hard_ack number is the sequence number before base of the window,
representing the last packet the other side says it has consumed.
hard_ack starts from 0 and the first packet is sequence number 1.
The top number is the sequence number of the highest-numbered packet
residing in the buffer. Packets between hard_ack+1 and top are
soft-ACK'd to indicate they've been received, but not yet consumed.
Four macros, before(), before_eq(), after() and after_eq() are added
to compare sequence numbers within the window. This allows for the
top of the window to wrap when the hard-ack sequence number gets close
to the limit.
Two flags, RXRPC_CALL_RX_LAST and RXRPC_CALL_TX_LAST, are added also
to indicate when rx_top and tx_top point at the packets with the
LAST_PACKET bit set, indicating the end of the phase.
(4) Calls are queued on the socket 'receive queue' rather than packets.
This means that we don't need have to invent dummy packets to queue to
indicate abnormal/terminal states and we don't have to keep metadata
packets (such as ABORTs) around
(5) The offset and length of a (sub)packet's content are now passed to
the verify_packet security op. This is currently expected to decrypt
the packet in place and validate it.
However, there's now nowhere to store the revised offset and length of
the actual data within the decrypted blob (there may be a header and
padding to skip) because an sk_buff may represent multiple packets, so
a locate_data security op is added to retrieve these details from the
sk_buff content when needed.
(6) recvmsg() now has to handle jumbo subpackets, where each subpacket is
individually secured and needs to be individually decrypted. The code
to do this is broken out into rxrpc_recvmsg_data() and shared with the
kernel API. It now iterates over the call's receive buffer rather
than walking the socket receive queue.
Additional changes:
(1) The timers are condensed to a single timer that is set for the soonest
of three timeouts (delayed ACK generation, DATA retransmission and
call lifespan).
(2) Transmission of ACK and ABORT packets is effected immediately from
process-context socket ops/kernel API calls that cause them instead of
them being punted off to a background work item. The data_ready
handler still has to defer to the background, though.
(3) A shutdown op is added to the AF_RXRPC socket so that the AFS
filesystem can shut down the socket and flush its own work items
before closing the socket to deal with any in-progress service calls.
Future additional changes that will need to be considered:
(1) Make sure that a call doesn't hog the front of the queue by receiving
data from the network as fast as userspace is consuming it to the
exclusion of other calls.
(2) Transmit delayed ACKs from within recvmsg() when we've consumed
sufficiently more packets to avoid the background work item needing to
run.
Signed-off-by: David Howells <dhowells@redhat.com>
2016-09-08 10:10:12 +00:00
|
|
|
|
2022-08-27 13:27:56 +00:00
|
|
|
if (after(seq, call->rx_highest_seq))
|
|
|
|
|
call->rx_highest_seq = seq;
|
|
|
|
|
|
2022-10-07 16:44:39 +00:00
|
|
|
trace_rxrpc_rx_data(call->debug_id, seq, serial, sp->hdr.flags);
|
2007-04-26 22:48:28 +00:00
|
|
|
|
2022-08-27 13:27:56 +00:00
|
|
|
if (before(seq, window)) {
|
|
|
|
|
ack_reason = RXRPC_ACK_DUPLICATE;
|
|
|
|
|
goto send_ack;
|
2022-10-07 16:44:39 +00:00
|
|
|
}
|
2022-08-27 13:27:56 +00:00
|
|
|
if (after(seq, wlimit)) {
|
|
|
|
|
ack_reason = RXRPC_ACK_EXCEEDS_WINDOW;
|
|
|
|
|
goto send_ack;
|
2022-10-07 16:44:39 +00:00
|
|
|
}
|
|
|
|
|
|
2022-08-27 13:27:56 +00:00
|
|
|
/* Queue the packet. */
|
|
|
|
|
if (seq == window) {
|
|
|
|
|
if (sp->hdr.flags & RXRPC_REQUEST_ACK)
|
|
|
|
|
ack_reason = RXRPC_ACK_REQUESTED;
|
|
|
|
|
/* Send an immediate ACK if we fill in a hole */
|
|
|
|
|
else if (!skb_queue_empty(&call->rx_oos_queue))
|
|
|
|
|
ack_reason = RXRPC_ACK_DELAY;
|
2022-10-07 16:44:39 +00:00
|
|
|
|
2022-08-27 13:27:56 +00:00
|
|
|
window++;
|
2022-10-16 07:01:32 +00:00
|
|
|
if (after(window, wtop)) {
|
|
|
|
|
trace_rxrpc_sack(call, seq, sack, rxrpc_sack_none);
|
2022-08-27 13:27:56 +00:00
|
|
|
wtop = window;
|
2022-10-16 07:01:32 +00:00
|
|
|
} else {
|
|
|
|
|
trace_rxrpc_sack(call, seq, sack, rxrpc_sack_advance);
|
|
|
|
|
sack = (sack + 1) % RXRPC_SACK_SIZE;
|
|
|
|
|
}
|
|
|
|
|
|
2022-10-07 16:44:39 +00:00
|
|
|
|
2022-10-06 14:43:51 +00:00
|
|
|
rxrpc_get_skb(skb, rxrpc_skb_get_to_recvmsg);
|
|
|
|
|
|
2022-08-27 13:27:56 +00:00
|
|
|
rxrpc_input_queue_data(call, skb, window, wtop, rxrpc_receive_queue);
|
2022-10-06 14:43:51 +00:00
|
|
|
*_notify = true;
|
2022-08-27 13:27:56 +00:00
|
|
|
|
|
|
|
|
while ((oos = skb_peek(&call->rx_oos_queue))) {
|
|
|
|
|
struct rxrpc_skb_priv *osp = rxrpc_skb(oos);
|
|
|
|
|
|
|
|
|
|
if (after(osp->hdr.seq, window))
|
|
|
|
|
break;
|
|
|
|
|
|
|
|
|
|
__skb_unlink(oos, &call->rx_oos_queue);
|
|
|
|
|
last = osp->hdr.flags & RXRPC_LAST_PACKET;
|
|
|
|
|
seq = osp->hdr.seq;
|
2022-10-16 07:01:32 +00:00
|
|
|
call->ackr_sack_table[sack] = 0;
|
|
|
|
|
trace_rxrpc_sack(call, seq, sack, rxrpc_sack_fill);
|
|
|
|
|
sack = (sack + 1) % RXRPC_SACK_SIZE;
|
2022-08-27 13:27:56 +00:00
|
|
|
|
|
|
|
|
window++;
|
|
|
|
|
rxrpc_input_queue_data(call, oos, window, wtop,
|
2022-10-16 07:01:32 +00:00
|
|
|
rxrpc_receive_queue_oos);
|
2022-10-07 16:44:39 +00:00
|
|
|
}
|
|
|
|
|
|
2022-10-16 07:01:32 +00:00
|
|
|
call->ackr_sack_base = sack;
|
2022-10-07 16:44:39 +00:00
|
|
|
} else {
|
2022-10-16 07:01:32 +00:00
|
|
|
unsigned int slot;
|
2022-10-07 16:44:39 +00:00
|
|
|
|
2022-08-27 13:27:56 +00:00
|
|
|
ack_reason = RXRPC_ACK_OUT_OF_SEQUENCE;
|
|
|
|
|
|
2022-10-16 07:01:32 +00:00
|
|
|
slot = seq - window;
|
|
|
|
|
sack = (sack + slot) % RXRPC_SACK_SIZE;
|
|
|
|
|
|
|
|
|
|
if (call->ackr_sack_table[sack % RXRPC_SACK_SIZE]) {
|
|
|
|
|
ack_reason = RXRPC_ACK_DUPLICATE;
|
|
|
|
|
goto send_ack;
|
2022-08-27 13:27:56 +00:00
|
|
|
}
|
|
|
|
|
|
2022-10-16 07:01:32 +00:00
|
|
|
call->ackr_sack_table[sack % RXRPC_SACK_SIZE] |= 1;
|
|
|
|
|
trace_rxrpc_sack(call, seq, sack, rxrpc_sack_oos);
|
|
|
|
|
|
2022-08-27 13:27:56 +00:00
|
|
|
if (after(seq + 1, wtop)) {
|
|
|
|
|
wtop = seq + 1;
|
|
|
|
|
rxrpc_input_update_ack_window(call, window, wtop);
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
skb_queue_walk(&call->rx_oos_queue, oos) {
|
|
|
|
|
struct rxrpc_skb_priv *osp = rxrpc_skb(oos);
|
|
|
|
|
|
|
|
|
|
if (after(osp->hdr.seq, seq)) {
|
2022-10-06 14:43:51 +00:00
|
|
|
rxrpc_get_skb(skb, rxrpc_skb_get_to_recvmsg_oos);
|
2022-08-27 13:27:56 +00:00
|
|
|
__skb_queue_before(&call->rx_oos_queue, oos, skb);
|
|
|
|
|
goto oos_queued;
|
|
|
|
|
}
|
2022-10-07 16:44:39 +00:00
|
|
|
}
|
2022-08-27 13:27:56 +00:00
|
|
|
|
2022-10-06 14:43:51 +00:00
|
|
|
rxrpc_get_skb(skb, rxrpc_skb_get_to_recvmsg_oos);
|
2022-08-27 13:27:56 +00:00
|
|
|
__skb_queue_tail(&call->rx_oos_queue, skb);
|
|
|
|
|
oos_queued:
|
|
|
|
|
trace_rxrpc_receive(call, last ? rxrpc_receive_oos_last : rxrpc_receive_oos,
|
|
|
|
|
sp->hdr.serial, sp->hdr.seq);
|
2022-10-07 16:44:39 +00:00
|
|
|
}
|
|
|
|
|
|
2022-08-27 13:27:56 +00:00
|
|
|
send_ack:
|
2024-05-03 15:07:40 +00:00
|
|
|
if (ack_reason >= 0) {
|
|
|
|
|
if (rxrpc_ack_priority[ack_reason] > rxrpc_ack_priority[*_ack_reason]) {
|
|
|
|
|
*_ack_serial = serial;
|
|
|
|
|
*_ack_reason = ack_reason;
|
|
|
|
|
} else if (rxrpc_ack_priority[ack_reason] == rxrpc_ack_priority[*_ack_reason] &&
|
|
|
|
|
ack_reason == RXRPC_ACK_REQUESTED) {
|
|
|
|
|
*_ack_serial = serial;
|
|
|
|
|
*_ack_reason = ack_reason;
|
|
|
|
|
}
|
|
|
|
|
}
|
2007-04-26 22:48:28 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
|
|
/*
|
2022-10-07 16:44:39 +00:00
|
|
|
* Split a jumbo packet and file the bits separately.
|
2007-04-26 22:48:28 +00:00
|
|
|
*/
|
2022-10-07 16:44:39 +00:00
|
|
|
static bool rxrpc_input_split_jumbo(struct rxrpc_call *call, struct sk_buff *skb)
|
2007-04-26 22:48:28 +00:00
|
|
|
{
|
2022-10-07 16:44:39 +00:00
|
|
|
struct rxrpc_jumbo_header jhdr;
|
|
|
|
|
struct rxrpc_skb_priv *sp = rxrpc_skb(skb), *jsp;
|
|
|
|
|
struct sk_buff *jskb;
|
2024-05-03 15:07:40 +00:00
|
|
|
rxrpc_serial_t ack_serial = 0;
|
2022-10-07 16:44:39 +00:00
|
|
|
unsigned int offset = sizeof(struct rxrpc_wire_header);
|
|
|
|
|
unsigned int len = skb->len - offset;
|
2022-10-06 14:43:51 +00:00
|
|
|
bool notify = false;
|
2024-12-04 07:46:49 +00:00
|
|
|
int ack_reason = 0, count = 1, stat_ix;
|
2007-04-26 22:48:28 +00:00
|
|
|
|
2022-10-07 16:44:39 +00:00
|
|
|
while (sp->hdr.flags & RXRPC_JUMBO_PACKET) {
|
|
|
|
|
if (len < RXRPC_JUMBO_SUBPKTLEN)
|
|
|
|
|
goto protocol_error;
|
|
|
|
|
if (sp->hdr.flags & RXRPC_LAST_PACKET)
|
|
|
|
|
goto protocol_error;
|
|
|
|
|
if (skb_copy_bits(skb, offset + RXRPC_JUMBO_DATALEN,
|
|
|
|
|
&jhdr, sizeof(jhdr)) < 0)
|
|
|
|
|
goto protocol_error;
|
|
|
|
|
|
2020-01-23 13:13:41 +00:00
|
|
|
jskb = skb_clone(skb, GFP_NOFS);
|
2022-10-07 16:44:39 +00:00
|
|
|
if (!jskb) {
|
|
|
|
|
kdebug("couldn't clone");
|
|
|
|
|
return false;
|
|
|
|
|
}
|
2022-10-21 14:31:21 +00:00
|
|
|
rxrpc_new_skb(jskb, rxrpc_skb_new_jumbo_subpacket);
|
2022-10-07 16:44:39 +00:00
|
|
|
jsp = rxrpc_skb(jskb);
|
|
|
|
|
jsp->offset = offset;
|
|
|
|
|
jsp->len = RXRPC_JUMBO_DATALEN;
|
2024-05-03 15:07:40 +00:00
|
|
|
rxrpc_input_data_one(call, jskb, ¬ify, &ack_serial, &ack_reason);
|
2022-10-06 14:43:51 +00:00
|
|
|
rxrpc_free_skb(jskb, rxrpc_skb_put_jumbo_subpacket);
|
2022-10-07 16:44:39 +00:00
|
|
|
|
|
|
|
|
sp->hdr.flags = jhdr.flags;
|
|
|
|
|
sp->hdr._rsvd = ntohs(jhdr._rsvd);
|
|
|
|
|
sp->hdr.seq++;
|
|
|
|
|
sp->hdr.serial++;
|
|
|
|
|
offset += RXRPC_JUMBO_SUBPKTLEN;
|
|
|
|
|
len -= RXRPC_JUMBO_SUBPKTLEN;
|
2024-12-04 07:46:49 +00:00
|
|
|
count++;
|
rxrpc: Rewrite the data and ack handling code
Rewrite the data and ack handling code such that:
(1) Parsing of received ACK and ABORT packets and the distribution and the
filing of DATA packets happens entirely within the data_ready context
called from the UDP socket. This allows us to process and discard ACK
and ABORT packets much more quickly (they're no longer stashed on a
queue for a background thread to process).
(2) We avoid calling skb_clone(), pskb_pull() and pskb_trim(). We instead
keep track of the offset and length of the content of each packet in
the sk_buff metadata. This means we don't do any allocation in the
receive path.
(3) Jumbo DATA packet parsing is now done in data_ready context. Rather
than cloning the packet once for each subpacket and pulling/trimming
it, we file the packet multiple times with an annotation for each
indicating which subpacket is there. From that we can directly
calculate the offset and length.
(4) A call's receive queue can be accessed without taking locks (memory
barriers do have to be used, though).
(5) Incoming calls are set up from preallocated resources and immediately
made live. They can than have packets queued upon them and ACKs
generated. If insufficient resources exist, DATA packet #1 is given a
BUSY reply and other DATA packets are discarded).
(6) sk_buffs no longer take a ref on their parent call.
To make this work, the following changes are made:
(1) Each call's receive buffer is now a circular buffer of sk_buff
pointers (rxtx_buffer) rather than a number of sk_buff_heads spread
between the call and the socket. This permits each sk_buff to be in
the buffer multiple times. The receive buffer is reused for the
transmit buffer.
(2) A circular buffer of annotations (rxtx_annotations) is kept parallel
to the data buffer. Transmission phase annotations indicate whether a
buffered packet has been ACK'd or not and whether it needs
retransmission.
Receive phase annotations indicate whether a slot holds a whole packet
or a jumbo subpacket and, if the latter, which subpacket. They also
note whether the packet has been decrypted in place.
(3) DATA packet window tracking is much simplified. Each phase has just
two numbers representing the window (rx_hard_ack/rx_top and
tx_hard_ack/tx_top).
The hard_ack number is the sequence number before base of the window,
representing the last packet the other side says it has consumed.
hard_ack starts from 0 and the first packet is sequence number 1.
The top number is the sequence number of the highest-numbered packet
residing in the buffer. Packets between hard_ack+1 and top are
soft-ACK'd to indicate they've been received, but not yet consumed.
Four macros, before(), before_eq(), after() and after_eq() are added
to compare sequence numbers within the window. This allows for the
top of the window to wrap when the hard-ack sequence number gets close
to the limit.
Two flags, RXRPC_CALL_RX_LAST and RXRPC_CALL_TX_LAST, are added also
to indicate when rx_top and tx_top point at the packets with the
LAST_PACKET bit set, indicating the end of the phase.
(4) Calls are queued on the socket 'receive queue' rather than packets.
This means that we don't need have to invent dummy packets to queue to
indicate abnormal/terminal states and we don't have to keep metadata
packets (such as ABORTs) around
(5) The offset and length of a (sub)packet's content are now passed to
the verify_packet security op. This is currently expected to decrypt
the packet in place and validate it.
However, there's now nowhere to store the revised offset and length of
the actual data within the decrypted blob (there may be a header and
padding to skip) because an sk_buff may represent multiple packets, so
a locate_data security op is added to retrieve these details from the
sk_buff content when needed.
(6) recvmsg() now has to handle jumbo subpackets, where each subpacket is
individually secured and needs to be individually decrypted. The code
to do this is broken out into rxrpc_recvmsg_data() and shared with the
kernel API. It now iterates over the call's receive buffer rather
than walking the socket receive queue.
Additional changes:
(1) The timers are condensed to a single timer that is set for the soonest
of three timeouts (delayed ACK generation, DATA retransmission and
call lifespan).
(2) Transmission of ACK and ABORT packets is effected immediately from
process-context socket ops/kernel API calls that cause them instead of
them being punted off to a background work item. The data_ready
handler still has to defer to the background, though.
(3) A shutdown op is added to the AF_RXRPC socket so that the AFS
filesystem can shut down the socket and flush its own work items
before closing the socket to deal with any in-progress service calls.
Future additional changes that will need to be considered:
(1) Make sure that a call doesn't hog the front of the queue by receiving
data from the network as fast as userspace is consuming it to the
exclusion of other calls.
(2) Transmit delayed ACKs from within recvmsg() when we've consumed
sufficiently more packets to avoid the background work item needing to
run.
Signed-off-by: David Howells <dhowells@redhat.com>
2016-09-08 10:10:12 +00:00
|
|
|
}
|
2022-10-07 16:44:39 +00:00
|
|
|
|
|
|
|
|
sp->offset = offset;
|
|
|
|
|
sp->len = len;
|
2024-05-03 15:07:40 +00:00
|
|
|
rxrpc_input_data_one(call, skb, ¬ify, &ack_serial, &ack_reason);
|
|
|
|
|
|
2024-12-04 07:46:49 +00:00
|
|
|
stat_ix = umin(count, ARRAY_SIZE(call->rxnet->stat_rx_jumbo)) - 1;
|
|
|
|
|
atomic_inc(&call->rxnet->stat_rx_jumbo[stat_ix]);
|
|
|
|
|
|
2024-05-03 15:07:40 +00:00
|
|
|
if (ack_reason > 0) {
|
|
|
|
|
rxrpc_send_ACK(call, ack_reason, ack_serial,
|
|
|
|
|
rxrpc_propose_ack_input_data);
|
|
|
|
|
} else {
|
|
|
|
|
call->ackr_nr_unacked++;
|
|
|
|
|
rxrpc_propose_delay_ACK(call, sp->hdr.serial,
|
|
|
|
|
rxrpc_propose_ack_input_data);
|
|
|
|
|
}
|
2022-10-06 14:43:51 +00:00
|
|
|
if (notify) {
|
|
|
|
|
trace_rxrpc_notify_socket(call->debug_id, sp->hdr.serial);
|
|
|
|
|
rxrpc_notify_socket(call);
|
|
|
|
|
}
|
2022-10-07 16:44:39 +00:00
|
|
|
return true;
|
|
|
|
|
|
|
|
|
|
protocol_error:
|
|
|
|
|
return false;
|
rxrpc: Rewrite the data and ack handling code
Rewrite the data and ack handling code such that:
(1) Parsing of received ACK and ABORT packets and the distribution and the
filing of DATA packets happens entirely within the data_ready context
called from the UDP socket. This allows us to process and discard ACK
and ABORT packets much more quickly (they're no longer stashed on a
queue for a background thread to process).
(2) We avoid calling skb_clone(), pskb_pull() and pskb_trim(). We instead
keep track of the offset and length of the content of each packet in
the sk_buff metadata. This means we don't do any allocation in the
receive path.
(3) Jumbo DATA packet parsing is now done in data_ready context. Rather
than cloning the packet once for each subpacket and pulling/trimming
it, we file the packet multiple times with an annotation for each
indicating which subpacket is there. From that we can directly
calculate the offset and length.
(4) A call's receive queue can be accessed without taking locks (memory
barriers do have to be used, though).
(5) Incoming calls are set up from preallocated resources and immediately
made live. They can than have packets queued upon them and ACKs
generated. If insufficient resources exist, DATA packet #1 is given a
BUSY reply and other DATA packets are discarded).
(6) sk_buffs no longer take a ref on their parent call.
To make this work, the following changes are made:
(1) Each call's receive buffer is now a circular buffer of sk_buff
pointers (rxtx_buffer) rather than a number of sk_buff_heads spread
between the call and the socket. This permits each sk_buff to be in
the buffer multiple times. The receive buffer is reused for the
transmit buffer.
(2) A circular buffer of annotations (rxtx_annotations) is kept parallel
to the data buffer. Transmission phase annotations indicate whether a
buffered packet has been ACK'd or not and whether it needs
retransmission.
Receive phase annotations indicate whether a slot holds a whole packet
or a jumbo subpacket and, if the latter, which subpacket. They also
note whether the packet has been decrypted in place.
(3) DATA packet window tracking is much simplified. Each phase has just
two numbers representing the window (rx_hard_ack/rx_top and
tx_hard_ack/tx_top).
The hard_ack number is the sequence number before base of the window,
representing the last packet the other side says it has consumed.
hard_ack starts from 0 and the first packet is sequence number 1.
The top number is the sequence number of the highest-numbered packet
residing in the buffer. Packets between hard_ack+1 and top are
soft-ACK'd to indicate they've been received, but not yet consumed.
Four macros, before(), before_eq(), after() and after_eq() are added
to compare sequence numbers within the window. This allows for the
top of the window to wrap when the hard-ack sequence number gets close
to the limit.
Two flags, RXRPC_CALL_RX_LAST and RXRPC_CALL_TX_LAST, are added also
to indicate when rx_top and tx_top point at the packets with the
LAST_PACKET bit set, indicating the end of the phase.
(4) Calls are queued on the socket 'receive queue' rather than packets.
This means that we don't need have to invent dummy packets to queue to
indicate abnormal/terminal states and we don't have to keep metadata
packets (such as ABORTs) around
(5) The offset and length of a (sub)packet's content are now passed to
the verify_packet security op. This is currently expected to decrypt
the packet in place and validate it.
However, there's now nowhere to store the revised offset and length of
the actual data within the decrypted blob (there may be a header and
padding to skip) because an sk_buff may represent multiple packets, so
a locate_data security op is added to retrieve these details from the
sk_buff content when needed.
(6) recvmsg() now has to handle jumbo subpackets, where each subpacket is
individually secured and needs to be individually decrypted. The code
to do this is broken out into rxrpc_recvmsg_data() and shared with the
kernel API. It now iterates over the call's receive buffer rather
than walking the socket receive queue.
Additional changes:
(1) The timers are condensed to a single timer that is set for the soonest
of three timeouts (delayed ACK generation, DATA retransmission and
call lifespan).
(2) Transmission of ACK and ABORT packets is effected immediately from
process-context socket ops/kernel API calls that cause them instead of
them being punted off to a background work item. The data_ready
handler still has to defer to the background, though.
(3) A shutdown op is added to the AF_RXRPC socket so that the AFS
filesystem can shut down the socket and flush its own work items
before closing the socket to deal with any in-progress service calls.
Future additional changes that will need to be considered:
(1) Make sure that a call doesn't hog the front of the queue by receiving
data from the network as fast as userspace is consuming it to the
exclusion of other calls.
(2) Transmit delayed ACKs from within recvmsg() when we've consumed
sufficiently more packets to avoid the background work item needing to
run.
Signed-off-by: David Howells <dhowells@redhat.com>
2016-09-08 10:10:12 +00:00
|
|
|
}
|
2007-04-26 22:48:28 +00:00
|
|
|
|
rxrpc: Rewrite the data and ack handling code
Rewrite the data and ack handling code such that:
(1) Parsing of received ACK and ABORT packets and the distribution and the
filing of DATA packets happens entirely within the data_ready context
called from the UDP socket. This allows us to process and discard ACK
and ABORT packets much more quickly (they're no longer stashed on a
queue for a background thread to process).
(2) We avoid calling skb_clone(), pskb_pull() and pskb_trim(). We instead
keep track of the offset and length of the content of each packet in
the sk_buff metadata. This means we don't do any allocation in the
receive path.
(3) Jumbo DATA packet parsing is now done in data_ready context. Rather
than cloning the packet once for each subpacket and pulling/trimming
it, we file the packet multiple times with an annotation for each
indicating which subpacket is there. From that we can directly
calculate the offset and length.
(4) A call's receive queue can be accessed without taking locks (memory
barriers do have to be used, though).
(5) Incoming calls are set up from preallocated resources and immediately
made live. They can than have packets queued upon them and ACKs
generated. If insufficient resources exist, DATA packet #1 is given a
BUSY reply and other DATA packets are discarded).
(6) sk_buffs no longer take a ref on their parent call.
To make this work, the following changes are made:
(1) Each call's receive buffer is now a circular buffer of sk_buff
pointers (rxtx_buffer) rather than a number of sk_buff_heads spread
between the call and the socket. This permits each sk_buff to be in
the buffer multiple times. The receive buffer is reused for the
transmit buffer.
(2) A circular buffer of annotations (rxtx_annotations) is kept parallel
to the data buffer. Transmission phase annotations indicate whether a
buffered packet has been ACK'd or not and whether it needs
retransmission.
Receive phase annotations indicate whether a slot holds a whole packet
or a jumbo subpacket and, if the latter, which subpacket. They also
note whether the packet has been decrypted in place.
(3) DATA packet window tracking is much simplified. Each phase has just
two numbers representing the window (rx_hard_ack/rx_top and
tx_hard_ack/tx_top).
The hard_ack number is the sequence number before base of the window,
representing the last packet the other side says it has consumed.
hard_ack starts from 0 and the first packet is sequence number 1.
The top number is the sequence number of the highest-numbered packet
residing in the buffer. Packets between hard_ack+1 and top are
soft-ACK'd to indicate they've been received, but not yet consumed.
Four macros, before(), before_eq(), after() and after_eq() are added
to compare sequence numbers within the window. This allows for the
top of the window to wrap when the hard-ack sequence number gets close
to the limit.
Two flags, RXRPC_CALL_RX_LAST and RXRPC_CALL_TX_LAST, are added also
to indicate when rx_top and tx_top point at the packets with the
LAST_PACKET bit set, indicating the end of the phase.
(4) Calls are queued on the socket 'receive queue' rather than packets.
This means that we don't need have to invent dummy packets to queue to
indicate abnormal/terminal states and we don't have to keep metadata
packets (such as ABORTs) around
(5) The offset and length of a (sub)packet's content are now passed to
the verify_packet security op. This is currently expected to decrypt
the packet in place and validate it.
However, there's now nowhere to store the revised offset and length of
the actual data within the decrypted blob (there may be a header and
padding to skip) because an sk_buff may represent multiple packets, so
a locate_data security op is added to retrieve these details from the
sk_buff content when needed.
(6) recvmsg() now has to handle jumbo subpackets, where each subpacket is
individually secured and needs to be individually decrypted. The code
to do this is broken out into rxrpc_recvmsg_data() and shared with the
kernel API. It now iterates over the call's receive buffer rather
than walking the socket receive queue.
Additional changes:
(1) The timers are condensed to a single timer that is set for the soonest
of three timeouts (delayed ACK generation, DATA retransmission and
call lifespan).
(2) Transmission of ACK and ABORT packets is effected immediately from
process-context socket ops/kernel API calls that cause them instead of
them being punted off to a background work item. The data_ready
handler still has to defer to the background, though.
(3) A shutdown op is added to the AF_RXRPC socket so that the AFS
filesystem can shut down the socket and flush its own work items
before closing the socket to deal with any in-progress service calls.
Future additional changes that will need to be considered:
(1) Make sure that a call doesn't hog the front of the queue by receiving
data from the network as fast as userspace is consuming it to the
exclusion of other calls.
(2) Transmit delayed ACKs from within recvmsg() when we've consumed
sufficiently more packets to avoid the background work item needing to
run.
Signed-off-by: David Howells <dhowells@redhat.com>
2016-09-08 10:10:12 +00:00
|
|
|
/*
|
2019-08-19 08:25:36 +00:00
|
|
|
* Process a DATA packet, adding the packet to the Rx ring. The caller's
|
|
|
|
|
* packet ref must be passed on or discarded.
|
rxrpc: Rewrite the data and ack handling code
Rewrite the data and ack handling code such that:
(1) Parsing of received ACK and ABORT packets and the distribution and the
filing of DATA packets happens entirely within the data_ready context
called from the UDP socket. This allows us to process and discard ACK
and ABORT packets much more quickly (they're no longer stashed on a
queue for a background thread to process).
(2) We avoid calling skb_clone(), pskb_pull() and pskb_trim(). We instead
keep track of the offset and length of the content of each packet in
the sk_buff metadata. This means we don't do any allocation in the
receive path.
(3) Jumbo DATA packet parsing is now done in data_ready context. Rather
than cloning the packet once for each subpacket and pulling/trimming
it, we file the packet multiple times with an annotation for each
indicating which subpacket is there. From that we can directly
calculate the offset and length.
(4) A call's receive queue can be accessed without taking locks (memory
barriers do have to be used, though).
(5) Incoming calls are set up from preallocated resources and immediately
made live. They can than have packets queued upon them and ACKs
generated. If insufficient resources exist, DATA packet #1 is given a
BUSY reply and other DATA packets are discarded).
(6) sk_buffs no longer take a ref on their parent call.
To make this work, the following changes are made:
(1) Each call's receive buffer is now a circular buffer of sk_buff
pointers (rxtx_buffer) rather than a number of sk_buff_heads spread
between the call and the socket. This permits each sk_buff to be in
the buffer multiple times. The receive buffer is reused for the
transmit buffer.
(2) A circular buffer of annotations (rxtx_annotations) is kept parallel
to the data buffer. Transmission phase annotations indicate whether a
buffered packet has been ACK'd or not and whether it needs
retransmission.
Receive phase annotations indicate whether a slot holds a whole packet
or a jumbo subpacket and, if the latter, which subpacket. They also
note whether the packet has been decrypted in place.
(3) DATA packet window tracking is much simplified. Each phase has just
two numbers representing the window (rx_hard_ack/rx_top and
tx_hard_ack/tx_top).
The hard_ack number is the sequence number before base of the window,
representing the last packet the other side says it has consumed.
hard_ack starts from 0 and the first packet is sequence number 1.
The top number is the sequence number of the highest-numbered packet
residing in the buffer. Packets between hard_ack+1 and top are
soft-ACK'd to indicate they've been received, but not yet consumed.
Four macros, before(), before_eq(), after() and after_eq() are added
to compare sequence numbers within the window. This allows for the
top of the window to wrap when the hard-ack sequence number gets close
to the limit.
Two flags, RXRPC_CALL_RX_LAST and RXRPC_CALL_TX_LAST, are added also
to indicate when rx_top and tx_top point at the packets with the
LAST_PACKET bit set, indicating the end of the phase.
(4) Calls are queued on the socket 'receive queue' rather than packets.
This means that we don't need have to invent dummy packets to queue to
indicate abnormal/terminal states and we don't have to keep metadata
packets (such as ABORTs) around
(5) The offset and length of a (sub)packet's content are now passed to
the verify_packet security op. This is currently expected to decrypt
the packet in place and validate it.
However, there's now nowhere to store the revised offset and length of
the actual data within the decrypted blob (there may be a header and
padding to skip) because an sk_buff may represent multiple packets, so
a locate_data security op is added to retrieve these details from the
sk_buff content when needed.
(6) recvmsg() now has to handle jumbo subpackets, where each subpacket is
individually secured and needs to be individually decrypted. The code
to do this is broken out into rxrpc_recvmsg_data() and shared with the
kernel API. It now iterates over the call's receive buffer rather
than walking the socket receive queue.
Additional changes:
(1) The timers are condensed to a single timer that is set for the soonest
of three timeouts (delayed ACK generation, DATA retransmission and
call lifespan).
(2) Transmission of ACK and ABORT packets is effected immediately from
process-context socket ops/kernel API calls that cause them instead of
them being punted off to a background work item. The data_ready
handler still has to defer to the background, though.
(3) A shutdown op is added to the AF_RXRPC socket so that the AFS
filesystem can shut down the socket and flush its own work items
before closing the socket to deal with any in-progress service calls.
Future additional changes that will need to be considered:
(1) Make sure that a call doesn't hog the front of the queue by receiving
data from the network as fast as userspace is consuming it to the
exclusion of other calls.
(2) Transmit delayed ACKs from within recvmsg() when we've consumed
sufficiently more packets to avoid the background work item needing to
run.
Signed-off-by: David Howells <dhowells@redhat.com>
2016-09-08 10:10:12 +00:00
|
|
|
*/
|
2019-08-09 14:20:41 +00:00
|
|
|
static void rxrpc_input_data(struct rxrpc_call *call, struct sk_buff *skb)
|
rxrpc: Rewrite the data and ack handling code
Rewrite the data and ack handling code such that:
(1) Parsing of received ACK and ABORT packets and the distribution and the
filing of DATA packets happens entirely within the data_ready context
called from the UDP socket. This allows us to process and discard ACK
and ABORT packets much more quickly (they're no longer stashed on a
queue for a background thread to process).
(2) We avoid calling skb_clone(), pskb_pull() and pskb_trim(). We instead
keep track of the offset and length of the content of each packet in
the sk_buff metadata. This means we don't do any allocation in the
receive path.
(3) Jumbo DATA packet parsing is now done in data_ready context. Rather
than cloning the packet once for each subpacket and pulling/trimming
it, we file the packet multiple times with an annotation for each
indicating which subpacket is there. From that we can directly
calculate the offset and length.
(4) A call's receive queue can be accessed without taking locks (memory
barriers do have to be used, though).
(5) Incoming calls are set up from preallocated resources and immediately
made live. They can than have packets queued upon them and ACKs
generated. If insufficient resources exist, DATA packet #1 is given a
BUSY reply and other DATA packets are discarded).
(6) sk_buffs no longer take a ref on their parent call.
To make this work, the following changes are made:
(1) Each call's receive buffer is now a circular buffer of sk_buff
pointers (rxtx_buffer) rather than a number of sk_buff_heads spread
between the call and the socket. This permits each sk_buff to be in
the buffer multiple times. The receive buffer is reused for the
transmit buffer.
(2) A circular buffer of annotations (rxtx_annotations) is kept parallel
to the data buffer. Transmission phase annotations indicate whether a
buffered packet has been ACK'd or not and whether it needs
retransmission.
Receive phase annotations indicate whether a slot holds a whole packet
or a jumbo subpacket and, if the latter, which subpacket. They also
note whether the packet has been decrypted in place.
(3) DATA packet window tracking is much simplified. Each phase has just
two numbers representing the window (rx_hard_ack/rx_top and
tx_hard_ack/tx_top).
The hard_ack number is the sequence number before base of the window,
representing the last packet the other side says it has consumed.
hard_ack starts from 0 and the first packet is sequence number 1.
The top number is the sequence number of the highest-numbered packet
residing in the buffer. Packets between hard_ack+1 and top are
soft-ACK'd to indicate they've been received, but not yet consumed.
Four macros, before(), before_eq(), after() and after_eq() are added
to compare sequence numbers within the window. This allows for the
top of the window to wrap when the hard-ack sequence number gets close
to the limit.
Two flags, RXRPC_CALL_RX_LAST and RXRPC_CALL_TX_LAST, are added also
to indicate when rx_top and tx_top point at the packets with the
LAST_PACKET bit set, indicating the end of the phase.
(4) Calls are queued on the socket 'receive queue' rather than packets.
This means that we don't need have to invent dummy packets to queue to
indicate abnormal/terminal states and we don't have to keep metadata
packets (such as ABORTs) around
(5) The offset and length of a (sub)packet's content are now passed to
the verify_packet security op. This is currently expected to decrypt
the packet in place and validate it.
However, there's now nowhere to store the revised offset and length of
the actual data within the decrypted blob (there may be a header and
padding to skip) because an sk_buff may represent multiple packets, so
a locate_data security op is added to retrieve these details from the
sk_buff content when needed.
(6) recvmsg() now has to handle jumbo subpackets, where each subpacket is
individually secured and needs to be individually decrypted. The code
to do this is broken out into rxrpc_recvmsg_data() and shared with the
kernel API. It now iterates over the call's receive buffer rather
than walking the socket receive queue.
Additional changes:
(1) The timers are condensed to a single timer that is set for the soonest
of three timeouts (delayed ACK generation, DATA retransmission and
call lifespan).
(2) Transmission of ACK and ABORT packets is effected immediately from
process-context socket ops/kernel API calls that cause them instead of
them being punted off to a background work item. The data_ready
handler still has to defer to the background, though.
(3) A shutdown op is added to the AF_RXRPC socket so that the AFS
filesystem can shut down the socket and flush its own work items
before closing the socket to deal with any in-progress service calls.
Future additional changes that will need to be considered:
(1) Make sure that a call doesn't hog the front of the queue by receiving
data from the network as fast as userspace is consuming it to the
exclusion of other calls.
(2) Transmit delayed ACKs from within recvmsg() when we've consumed
sufficiently more packets to avoid the background work item needing to
run.
Signed-off-by: David Howells <dhowells@redhat.com>
2016-09-08 10:10:12 +00:00
|
|
|
{
|
|
|
|
|
struct rxrpc_skb_priv *sp = rxrpc_skb(skb);
|
2022-10-07 16:44:39 +00:00
|
|
|
rxrpc_serial_t serial = sp->hdr.serial;
|
|
|
|
|
rxrpc_seq_t seq0 = sp->hdr.seq;
|
2007-04-26 22:48:28 +00:00
|
|
|
|
2022-10-17 10:44:22 +00:00
|
|
|
_enter("{%x,%x,%x},{%u,%x}",
|
|
|
|
|
call->ackr_window, call->ackr_wtop, call->rx_highest_seq,
|
2022-08-27 13:27:56 +00:00
|
|
|
skb->len, seq0);
|
2007-04-26 22:48:28 +00:00
|
|
|
|
2022-10-27 10:25:55 +00:00
|
|
|
if (__rxrpc_call_is_complete(call))
|
rxrpc: Rewrite the data and ack handling code
Rewrite the data and ack handling code such that:
(1) Parsing of received ACK and ABORT packets and the distribution and the
filing of DATA packets happens entirely within the data_ready context
called from the UDP socket. This allows us to process and discard ACK
and ABORT packets much more quickly (they're no longer stashed on a
queue for a background thread to process).
(2) We avoid calling skb_clone(), pskb_pull() and pskb_trim(). We instead
keep track of the offset and length of the content of each packet in
the sk_buff metadata. This means we don't do any allocation in the
receive path.
(3) Jumbo DATA packet parsing is now done in data_ready context. Rather
than cloning the packet once for each subpacket and pulling/trimming
it, we file the packet multiple times with an annotation for each
indicating which subpacket is there. From that we can directly
calculate the offset and length.
(4) A call's receive queue can be accessed without taking locks (memory
barriers do have to be used, though).
(5) Incoming calls are set up from preallocated resources and immediately
made live. They can than have packets queued upon them and ACKs
generated. If insufficient resources exist, DATA packet #1 is given a
BUSY reply and other DATA packets are discarded).
(6) sk_buffs no longer take a ref on their parent call.
To make this work, the following changes are made:
(1) Each call's receive buffer is now a circular buffer of sk_buff
pointers (rxtx_buffer) rather than a number of sk_buff_heads spread
between the call and the socket. This permits each sk_buff to be in
the buffer multiple times. The receive buffer is reused for the
transmit buffer.
(2) A circular buffer of annotations (rxtx_annotations) is kept parallel
to the data buffer. Transmission phase annotations indicate whether a
buffered packet has been ACK'd or not and whether it needs
retransmission.
Receive phase annotations indicate whether a slot holds a whole packet
or a jumbo subpacket and, if the latter, which subpacket. They also
note whether the packet has been decrypted in place.
(3) DATA packet window tracking is much simplified. Each phase has just
two numbers representing the window (rx_hard_ack/rx_top and
tx_hard_ack/tx_top).
The hard_ack number is the sequence number before base of the window,
representing the last packet the other side says it has consumed.
hard_ack starts from 0 and the first packet is sequence number 1.
The top number is the sequence number of the highest-numbered packet
residing in the buffer. Packets between hard_ack+1 and top are
soft-ACK'd to indicate they've been received, but not yet consumed.
Four macros, before(), before_eq(), after() and after_eq() are added
to compare sequence numbers within the window. This allows for the
top of the window to wrap when the hard-ack sequence number gets close
to the limit.
Two flags, RXRPC_CALL_RX_LAST and RXRPC_CALL_TX_LAST, are added also
to indicate when rx_top and tx_top point at the packets with the
LAST_PACKET bit set, indicating the end of the phase.
(4) Calls are queued on the socket 'receive queue' rather than packets.
This means that we don't need have to invent dummy packets to queue to
indicate abnormal/terminal states and we don't have to keep metadata
packets (such as ABORTs) around
(5) The offset and length of a (sub)packet's content are now passed to
the verify_packet security op. This is currently expected to decrypt
the packet in place and validate it.
However, there's now nowhere to store the revised offset and length of
the actual data within the decrypted blob (there may be a header and
padding to skip) because an sk_buff may represent multiple packets, so
a locate_data security op is added to retrieve these details from the
sk_buff content when needed.
(6) recvmsg() now has to handle jumbo subpackets, where each subpacket is
individually secured and needs to be individually decrypted. The code
to do this is broken out into rxrpc_recvmsg_data() and shared with the
kernel API. It now iterates over the call's receive buffer rather
than walking the socket receive queue.
Additional changes:
(1) The timers are condensed to a single timer that is set for the soonest
of three timeouts (delayed ACK generation, DATA retransmission and
call lifespan).
(2) Transmission of ACK and ABORT packets is effected immediately from
process-context socket ops/kernel API calls that cause them instead of
them being punted off to a background work item. The data_ready
handler still has to defer to the background, though.
(3) A shutdown op is added to the AF_RXRPC socket so that the AFS
filesystem can shut down the socket and flush its own work items
before closing the socket to deal with any in-progress service calls.
Future additional changes that will need to be considered:
(1) Make sure that a call doesn't hog the front of the queue by receiving
data from the network as fast as userspace is consuming it to the
exclusion of other calls.
(2) Transmit delayed ACKs from within recvmsg() when we've consumed
sufficiently more packets to avoid the background work item needing to
run.
Signed-off-by: David Howells <dhowells@redhat.com>
2016-09-08 10:10:12 +00:00
|
|
|
return;
|
2007-04-26 22:48:28 +00:00
|
|
|
|
2022-10-27 10:25:55 +00:00
|
|
|
switch (__rxrpc_call_state(call)) {
|
|
|
|
|
case RXRPC_CALL_CLIENT_SEND_REQUEST:
|
|
|
|
|
case RXRPC_CALL_CLIENT_AWAIT_REPLY:
|
|
|
|
|
/* Received data implicitly ACKs all of the request
|
|
|
|
|
* packets we sent when we're acting as a client.
|
|
|
|
|
*/
|
|
|
|
|
if (!rxrpc_receiving_reply(call))
|
|
|
|
|
goto out_notify;
|
|
|
|
|
break;
|
|
|
|
|
|
|
|
|
|
case RXRPC_CALL_SERVER_RECV_REQUEST: {
|
2017-11-24 10:18:41 +00:00
|
|
|
unsigned long timo = READ_ONCE(call->next_req_timo);
|
|
|
|
|
|
|
|
|
|
if (timo) {
|
2024-01-30 21:37:16 +00:00
|
|
|
ktime_t delay = ms_to_ktime(timo);
|
|
|
|
|
|
|
|
|
|
call->expect_req_by = ktime_add(ktime_get_real(), delay);
|
|
|
|
|
trace_rxrpc_timer_set(call, delay, rxrpc_timer_trace_idle);
|
2017-11-24 10:18:41 +00:00
|
|
|
}
|
2022-10-27 10:25:55 +00:00
|
|
|
break;
|
2017-11-24 10:18:41 +00:00
|
|
|
}
|
|
|
|
|
|
2022-10-27 10:25:55 +00:00
|
|
|
default:
|
|
|
|
|
break;
|
|
|
|
|
}
|
2020-01-30 21:48:13 +00:00
|
|
|
|
2022-10-07 16:44:39 +00:00
|
|
|
if (!rxrpc_input_split_jumbo(call, skb)) {
|
2022-10-06 20:45:42 +00:00
|
|
|
rxrpc_proto_abort(call, sp->hdr.seq, rxrpc_badmsg_bad_jumbo);
|
2020-01-23 13:13:41 +00:00
|
|
|
goto out_notify;
|
rxrpc: Rewrite the data and ack handling code
Rewrite the data and ack handling code such that:
(1) Parsing of received ACK and ABORT packets and the distribution and the
filing of DATA packets happens entirely within the data_ready context
called from the UDP socket. This allows us to process and discard ACK
and ABORT packets much more quickly (they're no longer stashed on a
queue for a background thread to process).
(2) We avoid calling skb_clone(), pskb_pull() and pskb_trim(). We instead
keep track of the offset and length of the content of each packet in
the sk_buff metadata. This means we don't do any allocation in the
receive path.
(3) Jumbo DATA packet parsing is now done in data_ready context. Rather
than cloning the packet once for each subpacket and pulling/trimming
it, we file the packet multiple times with an annotation for each
indicating which subpacket is there. From that we can directly
calculate the offset and length.
(4) A call's receive queue can be accessed without taking locks (memory
barriers do have to be used, though).
(5) Incoming calls are set up from preallocated resources and immediately
made live. They can than have packets queued upon them and ACKs
generated. If insufficient resources exist, DATA packet #1 is given a
BUSY reply and other DATA packets are discarded).
(6) sk_buffs no longer take a ref on their parent call.
To make this work, the following changes are made:
(1) Each call's receive buffer is now a circular buffer of sk_buff
pointers (rxtx_buffer) rather than a number of sk_buff_heads spread
between the call and the socket. This permits each sk_buff to be in
the buffer multiple times. The receive buffer is reused for the
transmit buffer.
(2) A circular buffer of annotations (rxtx_annotations) is kept parallel
to the data buffer. Transmission phase annotations indicate whether a
buffered packet has been ACK'd or not and whether it needs
retransmission.
Receive phase annotations indicate whether a slot holds a whole packet
or a jumbo subpacket and, if the latter, which subpacket. They also
note whether the packet has been decrypted in place.
(3) DATA packet window tracking is much simplified. Each phase has just
two numbers representing the window (rx_hard_ack/rx_top and
tx_hard_ack/tx_top).
The hard_ack number is the sequence number before base of the window,
representing the last packet the other side says it has consumed.
hard_ack starts from 0 and the first packet is sequence number 1.
The top number is the sequence number of the highest-numbered packet
residing in the buffer. Packets between hard_ack+1 and top are
soft-ACK'd to indicate they've been received, but not yet consumed.
Four macros, before(), before_eq(), after() and after_eq() are added
to compare sequence numbers within the window. This allows for the
top of the window to wrap when the hard-ack sequence number gets close
to the limit.
Two flags, RXRPC_CALL_RX_LAST and RXRPC_CALL_TX_LAST, are added also
to indicate when rx_top and tx_top point at the packets with the
LAST_PACKET bit set, indicating the end of the phase.
(4) Calls are queued on the socket 'receive queue' rather than packets.
This means that we don't need have to invent dummy packets to queue to
indicate abnormal/terminal states and we don't have to keep metadata
packets (such as ABORTs) around
(5) The offset and length of a (sub)packet's content are now passed to
the verify_packet security op. This is currently expected to decrypt
the packet in place and validate it.
However, there's now nowhere to store the revised offset and length of
the actual data within the decrypted blob (there may be a header and
padding to skip) because an sk_buff may represent multiple packets, so
a locate_data security op is added to retrieve these details from the
sk_buff content when needed.
(6) recvmsg() now has to handle jumbo subpackets, where each subpacket is
individually secured and needs to be individually decrypted. The code
to do this is broken out into rxrpc_recvmsg_data() and shared with the
kernel API. It now iterates over the call's receive buffer rather
than walking the socket receive queue.
Additional changes:
(1) The timers are condensed to a single timer that is set for the soonest
of three timeouts (delayed ACK generation, DATA retransmission and
call lifespan).
(2) Transmission of ACK and ABORT packets is effected immediately from
process-context socket ops/kernel API calls that cause them instead of
them being punted off to a background work item. The data_ready
handler still has to defer to the background, though.
(3) A shutdown op is added to the AF_RXRPC socket so that the AFS
filesystem can shut down the socket and flush its own work items
before closing the socket to deal with any in-progress service calls.
Future additional changes that will need to be considered:
(1) Make sure that a call doesn't hog the front of the queue by receiving
data from the network as fast as userspace is consuming it to the
exclusion of other calls.
(2) Transmit delayed ACKs from within recvmsg() when we've consumed
sufficiently more packets to avoid the background work item needing to
run.
Signed-off-by: David Howells <dhowells@redhat.com>
2016-09-08 10:10:12 +00:00
|
|
|
}
|
2023-02-15 21:48:05 +00:00
|
|
|
return;
|
2007-04-26 22:48:28 +00:00
|
|
|
|
2020-01-23 13:13:41 +00:00
|
|
|
out_notify:
|
2020-01-30 21:50:36 +00:00
|
|
|
trace_rxrpc_notify_socket(call->debug_id, serial);
|
|
|
|
|
rxrpc_notify_socket(call);
|
rxrpc: Rewrite the data and ack handling code
Rewrite the data and ack handling code such that:
(1) Parsing of received ACK and ABORT packets and the distribution and the
filing of DATA packets happens entirely within the data_ready context
called from the UDP socket. This allows us to process and discard ACK
and ABORT packets much more quickly (they're no longer stashed on a
queue for a background thread to process).
(2) We avoid calling skb_clone(), pskb_pull() and pskb_trim(). We instead
keep track of the offset and length of the content of each packet in
the sk_buff metadata. This means we don't do any allocation in the
receive path.
(3) Jumbo DATA packet parsing is now done in data_ready context. Rather
than cloning the packet once for each subpacket and pulling/trimming
it, we file the packet multiple times with an annotation for each
indicating which subpacket is there. From that we can directly
calculate the offset and length.
(4) A call's receive queue can be accessed without taking locks (memory
barriers do have to be used, though).
(5) Incoming calls are set up from preallocated resources and immediately
made live. They can than have packets queued upon them and ACKs
generated. If insufficient resources exist, DATA packet #1 is given a
BUSY reply and other DATA packets are discarded).
(6) sk_buffs no longer take a ref on their parent call.
To make this work, the following changes are made:
(1) Each call's receive buffer is now a circular buffer of sk_buff
pointers (rxtx_buffer) rather than a number of sk_buff_heads spread
between the call and the socket. This permits each sk_buff to be in
the buffer multiple times. The receive buffer is reused for the
transmit buffer.
(2) A circular buffer of annotations (rxtx_annotations) is kept parallel
to the data buffer. Transmission phase annotations indicate whether a
buffered packet has been ACK'd or not and whether it needs
retransmission.
Receive phase annotations indicate whether a slot holds a whole packet
or a jumbo subpacket and, if the latter, which subpacket. They also
note whether the packet has been decrypted in place.
(3) DATA packet window tracking is much simplified. Each phase has just
two numbers representing the window (rx_hard_ack/rx_top and
tx_hard_ack/tx_top).
The hard_ack number is the sequence number before base of the window,
representing the last packet the other side says it has consumed.
hard_ack starts from 0 and the first packet is sequence number 1.
The top number is the sequence number of the highest-numbered packet
residing in the buffer. Packets between hard_ack+1 and top are
soft-ACK'd to indicate they've been received, but not yet consumed.
Four macros, before(), before_eq(), after() and after_eq() are added
to compare sequence numbers within the window. This allows for the
top of the window to wrap when the hard-ack sequence number gets close
to the limit.
Two flags, RXRPC_CALL_RX_LAST and RXRPC_CALL_TX_LAST, are added also
to indicate when rx_top and tx_top point at the packets with the
LAST_PACKET bit set, indicating the end of the phase.
(4) Calls are queued on the socket 'receive queue' rather than packets.
This means that we don't need have to invent dummy packets to queue to
indicate abnormal/terminal states and we don't have to keep metadata
packets (such as ABORTs) around
(5) The offset and length of a (sub)packet's content are now passed to
the verify_packet security op. This is currently expected to decrypt
the packet in place and validate it.
However, there's now nowhere to store the revised offset and length of
the actual data within the decrypted blob (there may be a header and
padding to skip) because an sk_buff may represent multiple packets, so
a locate_data security op is added to retrieve these details from the
sk_buff content when needed.
(6) recvmsg() now has to handle jumbo subpackets, where each subpacket is
individually secured and needs to be individually decrypted. The code
to do this is broken out into rxrpc_recvmsg_data() and shared with the
kernel API. It now iterates over the call's receive buffer rather
than walking the socket receive queue.
Additional changes:
(1) The timers are condensed to a single timer that is set for the soonest
of three timeouts (delayed ACK generation, DATA retransmission and
call lifespan).
(2) Transmission of ACK and ABORT packets is effected immediately from
process-context socket ops/kernel API calls that cause them instead of
them being punted off to a background work item. The data_ready
handler still has to defer to the background, though.
(3) A shutdown op is added to the AF_RXRPC socket so that the AFS
filesystem can shut down the socket and flush its own work items
before closing the socket to deal with any in-progress service calls.
Future additional changes that will need to be considered:
(1) Make sure that a call doesn't hog the front of the queue by receiving
data from the network as fast as userspace is consuming it to the
exclusion of other calls.
(2) Transmit delayed ACKs from within recvmsg() when we've consumed
sufficiently more packets to avoid the background work item needing to
run.
Signed-off-by: David Howells <dhowells@redhat.com>
2016-09-08 10:10:12 +00:00
|
|
|
_leave(" [queued]");
|
2007-04-26 22:48:28 +00:00
|
|
|
}
|
|
|
|
|
|
2016-09-21 23:29:31 +00:00
|
|
|
/*
|
rxrpc: Fix loss of RTT samples due to interposed ACK
The Rx protocol has a mechanism to help generate RTT samples that works by
a client transmitting a REQUESTED-type ACK when it receives a DATA packet
that has the REQUEST_ACK flag set.
The peer, however, may interpose other ACKs before transmitting the
REQUESTED-ACK, as can be seen in the following trace excerpt:
rxrpc_tx_data: c=00000044 DATA d0b5ece8:00000001 00000001 q=00000001 fl=07
rxrpc_rx_ack: c=00000044 00000001 PNG r=00000000 f=00000002 p=00000000 n=0
rxrpc_rx_ack: c=00000044 00000002 REQ r=00000001 f=00000002 p=00000001 n=0
...
DATA packet 1 (q=xx) has REQUEST_ACK set (bit 1 of fl=xx). The incoming
ping (labelled PNG) hard-acks the request DATA packet (f=xx exceeds the
sequence number of the DATA packet), causing it to be discarded from the Tx
ring. The ACK that was requested (labelled REQ, r=xx references the serial
of the DATA packet) comes after the ping, but the sk_buff holding the
timestamp has gone and the RTT sample is lost.
This is particularly noticeable on RPC calls used to probe the service
offered by the peer. A lot of peers end up with an unknown RTT because we
only ever sent a single RPC. This confuses the server rotation algorithm.
Fix this by caching the information about the outgoing packet in RTT
calculations in the rxrpc_call struct rather than looking in the Tx ring.
A four-deep buffer is maintained and both REQUEST_ACK-flagged DATA and
PING-ACK transmissions are recorded in there. When the appropriate
response ACK is received, the buffer is checked for a match and, if found,
an RTT sample is recorded.
If a received ACK refers to a packet with a later serial number than an
entry in the cache, that entry is presumed lost and the entry is made
available to record a new transmission.
ACKs types other than REQUESTED-type and PING-type cause any matching
sample to be cancelled as they don't necessarily represent a useful
measurement.
If there's no space in the buffer on ping/data transmission, the sample
base is discarded.
Fixes: 50235c4b5a2f ("rxrpc: Obtain RTT data by requesting ACKs on DATA packets")
Signed-off-by: David Howells <dhowells@redhat.com>
2020-08-19 22:29:16 +00:00
|
|
|
* See if there's a cached RTT probe to complete.
|
2016-09-21 23:29:31 +00:00
|
|
|
*/
|
rxrpc: Fix loss of RTT samples due to interposed ACK
The Rx protocol has a mechanism to help generate RTT samples that works by
a client transmitting a REQUESTED-type ACK when it receives a DATA packet
that has the REQUEST_ACK flag set.
The peer, however, may interpose other ACKs before transmitting the
REQUESTED-ACK, as can be seen in the following trace excerpt:
rxrpc_tx_data: c=00000044 DATA d0b5ece8:00000001 00000001 q=00000001 fl=07
rxrpc_rx_ack: c=00000044 00000001 PNG r=00000000 f=00000002 p=00000000 n=0
rxrpc_rx_ack: c=00000044 00000002 REQ r=00000001 f=00000002 p=00000001 n=0
...
DATA packet 1 (q=xx) has REQUEST_ACK set (bit 1 of fl=xx). The incoming
ping (labelled PNG) hard-acks the request DATA packet (f=xx exceeds the
sequence number of the DATA packet), causing it to be discarded from the Tx
ring. The ACK that was requested (labelled REQ, r=xx references the serial
of the DATA packet) comes after the ping, but the sk_buff holding the
timestamp has gone and the RTT sample is lost.
This is particularly noticeable on RPC calls used to probe the service
offered by the peer. A lot of peers end up with an unknown RTT because we
only ever sent a single RPC. This confuses the server rotation algorithm.
Fix this by caching the information about the outgoing packet in RTT
calculations in the rxrpc_call struct rather than looking in the Tx ring.
A four-deep buffer is maintained and both REQUEST_ACK-flagged DATA and
PING-ACK transmissions are recorded in there. When the appropriate
response ACK is received, the buffer is checked for a match and, if found,
an RTT sample is recorded.
If a received ACK refers to a packet with a later serial number than an
entry in the cache, that entry is presumed lost and the entry is made
available to record a new transmission.
ACKs types other than REQUESTED-type and PING-type cause any matching
sample to be cancelled as they don't necessarily represent a useful
measurement.
If there's no space in the buffer on ping/data transmission, the sample
base is discarded.
Fixes: 50235c4b5a2f ("rxrpc: Obtain RTT data by requesting ACKs on DATA packets")
Signed-off-by: David Howells <dhowells@redhat.com>
2020-08-19 22:29:16 +00:00
|
|
|
static void rxrpc_complete_rtt_probe(struct rxrpc_call *call,
|
|
|
|
|
ktime_t resp_time,
|
|
|
|
|
rxrpc_serial_t acked_serial,
|
|
|
|
|
rxrpc_serial_t ack_serial,
|
|
|
|
|
enum rxrpc_rtt_rx_trace type)
|
2016-09-21 23:29:31 +00:00
|
|
|
{
|
rxrpc: Fix loss of RTT samples due to interposed ACK
The Rx protocol has a mechanism to help generate RTT samples that works by
a client transmitting a REQUESTED-type ACK when it receives a DATA packet
that has the REQUEST_ACK flag set.
The peer, however, may interpose other ACKs before transmitting the
REQUESTED-ACK, as can be seen in the following trace excerpt:
rxrpc_tx_data: c=00000044 DATA d0b5ece8:00000001 00000001 q=00000001 fl=07
rxrpc_rx_ack: c=00000044 00000001 PNG r=00000000 f=00000002 p=00000000 n=0
rxrpc_rx_ack: c=00000044 00000002 REQ r=00000001 f=00000002 p=00000001 n=0
...
DATA packet 1 (q=xx) has REQUEST_ACK set (bit 1 of fl=xx). The incoming
ping (labelled PNG) hard-acks the request DATA packet (f=xx exceeds the
sequence number of the DATA packet), causing it to be discarded from the Tx
ring. The ACK that was requested (labelled REQ, r=xx references the serial
of the DATA packet) comes after the ping, but the sk_buff holding the
timestamp has gone and the RTT sample is lost.
This is particularly noticeable on RPC calls used to probe the service
offered by the peer. A lot of peers end up with an unknown RTT because we
only ever sent a single RPC. This confuses the server rotation algorithm.
Fix this by caching the information about the outgoing packet in RTT
calculations in the rxrpc_call struct rather than looking in the Tx ring.
A four-deep buffer is maintained and both REQUEST_ACK-flagged DATA and
PING-ACK transmissions are recorded in there. When the appropriate
response ACK is received, the buffer is checked for a match and, if found,
an RTT sample is recorded.
If a received ACK refers to a packet with a later serial number than an
entry in the cache, that entry is presumed lost and the entry is made
available to record a new transmission.
ACKs types other than REQUESTED-type and PING-type cause any matching
sample to be cancelled as they don't necessarily represent a useful
measurement.
If there's no space in the buffer on ping/data transmission, the sample
base is discarded.
Fixes: 50235c4b5a2f ("rxrpc: Obtain RTT data by requesting ACKs on DATA packets")
Signed-off-by: David Howells <dhowells@redhat.com>
2020-08-19 22:29:16 +00:00
|
|
|
rxrpc_serial_t orig_serial;
|
|
|
|
|
unsigned long avail;
|
2016-09-21 23:29:31 +00:00
|
|
|
ktime_t sent_at;
|
rxrpc: Fix loss of RTT samples due to interposed ACK
The Rx protocol has a mechanism to help generate RTT samples that works by
a client transmitting a REQUESTED-type ACK when it receives a DATA packet
that has the REQUEST_ACK flag set.
The peer, however, may interpose other ACKs before transmitting the
REQUESTED-ACK, as can be seen in the following trace excerpt:
rxrpc_tx_data: c=00000044 DATA d0b5ece8:00000001 00000001 q=00000001 fl=07
rxrpc_rx_ack: c=00000044 00000001 PNG r=00000000 f=00000002 p=00000000 n=0
rxrpc_rx_ack: c=00000044 00000002 REQ r=00000001 f=00000002 p=00000001 n=0
...
DATA packet 1 (q=xx) has REQUEST_ACK set (bit 1 of fl=xx). The incoming
ping (labelled PNG) hard-acks the request DATA packet (f=xx exceeds the
sequence number of the DATA packet), causing it to be discarded from the Tx
ring. The ACK that was requested (labelled REQ, r=xx references the serial
of the DATA packet) comes after the ping, but the sk_buff holding the
timestamp has gone and the RTT sample is lost.
This is particularly noticeable on RPC calls used to probe the service
offered by the peer. A lot of peers end up with an unknown RTT because we
only ever sent a single RPC. This confuses the server rotation algorithm.
Fix this by caching the information about the outgoing packet in RTT
calculations in the rxrpc_call struct rather than looking in the Tx ring.
A four-deep buffer is maintained and both REQUEST_ACK-flagged DATA and
PING-ACK transmissions are recorded in there. When the appropriate
response ACK is received, the buffer is checked for a match and, if found,
an RTT sample is recorded.
If a received ACK refers to a packet with a later serial number than an
entry in the cache, that entry is presumed lost and the entry is made
available to record a new transmission.
ACKs types other than REQUESTED-type and PING-type cause any matching
sample to be cancelled as they don't necessarily represent a useful
measurement.
If there's no space in the buffer on ping/data transmission, the sample
base is discarded.
Fixes: 50235c4b5a2f ("rxrpc: Obtain RTT data by requesting ACKs on DATA packets")
Signed-off-by: David Howells <dhowells@redhat.com>
2020-08-19 22:29:16 +00:00
|
|
|
bool matched = false;
|
|
|
|
|
int i;
|
2016-09-21 23:29:31 +00:00
|
|
|
|
rxrpc: Fix loss of RTT samples due to interposed ACK
The Rx protocol has a mechanism to help generate RTT samples that works by
a client transmitting a REQUESTED-type ACK when it receives a DATA packet
that has the REQUEST_ACK flag set.
The peer, however, may interpose other ACKs before transmitting the
REQUESTED-ACK, as can be seen in the following trace excerpt:
rxrpc_tx_data: c=00000044 DATA d0b5ece8:00000001 00000001 q=00000001 fl=07
rxrpc_rx_ack: c=00000044 00000001 PNG r=00000000 f=00000002 p=00000000 n=0
rxrpc_rx_ack: c=00000044 00000002 REQ r=00000001 f=00000002 p=00000001 n=0
...
DATA packet 1 (q=xx) has REQUEST_ACK set (bit 1 of fl=xx). The incoming
ping (labelled PNG) hard-acks the request DATA packet (f=xx exceeds the
sequence number of the DATA packet), causing it to be discarded from the Tx
ring. The ACK that was requested (labelled REQ, r=xx references the serial
of the DATA packet) comes after the ping, but the sk_buff holding the
timestamp has gone and the RTT sample is lost.
This is particularly noticeable on RPC calls used to probe the service
offered by the peer. A lot of peers end up with an unknown RTT because we
only ever sent a single RPC. This confuses the server rotation algorithm.
Fix this by caching the information about the outgoing packet in RTT
calculations in the rxrpc_call struct rather than looking in the Tx ring.
A four-deep buffer is maintained and both REQUEST_ACK-flagged DATA and
PING-ACK transmissions are recorded in there. When the appropriate
response ACK is received, the buffer is checked for a match and, if found,
an RTT sample is recorded.
If a received ACK refers to a packet with a later serial number than an
entry in the cache, that entry is presumed lost and the entry is made
available to record a new transmission.
ACKs types other than REQUESTED-type and PING-type cause any matching
sample to be cancelled as they don't necessarily represent a useful
measurement.
If there's no space in the buffer on ping/data transmission, the sample
base is discarded.
Fixes: 50235c4b5a2f ("rxrpc: Obtain RTT data by requesting ACKs on DATA packets")
Signed-off-by: David Howells <dhowells@redhat.com>
2020-08-19 22:29:16 +00:00
|
|
|
avail = READ_ONCE(call->rtt_avail);
|
|
|
|
|
smp_rmb(); /* Read avail bits before accessing data. */
|
2016-09-21 23:29:31 +00:00
|
|
|
|
rxrpc: Fix loss of RTT samples due to interposed ACK
The Rx protocol has a mechanism to help generate RTT samples that works by
a client transmitting a REQUESTED-type ACK when it receives a DATA packet
that has the REQUEST_ACK flag set.
The peer, however, may interpose other ACKs before transmitting the
REQUESTED-ACK, as can be seen in the following trace excerpt:
rxrpc_tx_data: c=00000044 DATA d0b5ece8:00000001 00000001 q=00000001 fl=07
rxrpc_rx_ack: c=00000044 00000001 PNG r=00000000 f=00000002 p=00000000 n=0
rxrpc_rx_ack: c=00000044 00000002 REQ r=00000001 f=00000002 p=00000001 n=0
...
DATA packet 1 (q=xx) has REQUEST_ACK set (bit 1 of fl=xx). The incoming
ping (labelled PNG) hard-acks the request DATA packet (f=xx exceeds the
sequence number of the DATA packet), causing it to be discarded from the Tx
ring. The ACK that was requested (labelled REQ, r=xx references the serial
of the DATA packet) comes after the ping, but the sk_buff holding the
timestamp has gone and the RTT sample is lost.
This is particularly noticeable on RPC calls used to probe the service
offered by the peer. A lot of peers end up with an unknown RTT because we
only ever sent a single RPC. This confuses the server rotation algorithm.
Fix this by caching the information about the outgoing packet in RTT
calculations in the rxrpc_call struct rather than looking in the Tx ring.
A four-deep buffer is maintained and both REQUEST_ACK-flagged DATA and
PING-ACK transmissions are recorded in there. When the appropriate
response ACK is received, the buffer is checked for a match and, if found,
an RTT sample is recorded.
If a received ACK refers to a packet with a later serial number than an
entry in the cache, that entry is presumed lost and the entry is made
available to record a new transmission.
ACKs types other than REQUESTED-type and PING-type cause any matching
sample to be cancelled as they don't necessarily represent a useful
measurement.
If there's no space in the buffer on ping/data transmission, the sample
base is discarded.
Fixes: 50235c4b5a2f ("rxrpc: Obtain RTT data by requesting ACKs on DATA packets")
Signed-off-by: David Howells <dhowells@redhat.com>
2020-08-19 22:29:16 +00:00
|
|
|
for (i = 0; i < ARRAY_SIZE(call->rtt_serial); i++) {
|
|
|
|
|
if (!test_bit(i + RXRPC_CALL_RTT_PEND_SHIFT, &avail))
|
2016-09-21 23:29:31 +00:00
|
|
|
continue;
|
2018-09-27 14:13:08 +00:00
|
|
|
|
rxrpc: Fix loss of RTT samples due to interposed ACK
The Rx protocol has a mechanism to help generate RTT samples that works by
a client transmitting a REQUESTED-type ACK when it receives a DATA packet
that has the REQUEST_ACK flag set.
The peer, however, may interpose other ACKs before transmitting the
REQUESTED-ACK, as can be seen in the following trace excerpt:
rxrpc_tx_data: c=00000044 DATA d0b5ece8:00000001 00000001 q=00000001 fl=07
rxrpc_rx_ack: c=00000044 00000001 PNG r=00000000 f=00000002 p=00000000 n=0
rxrpc_rx_ack: c=00000044 00000002 REQ r=00000001 f=00000002 p=00000001 n=0
...
DATA packet 1 (q=xx) has REQUEST_ACK set (bit 1 of fl=xx). The incoming
ping (labelled PNG) hard-acks the request DATA packet (f=xx exceeds the
sequence number of the DATA packet), causing it to be discarded from the Tx
ring. The ACK that was requested (labelled REQ, r=xx references the serial
of the DATA packet) comes after the ping, but the sk_buff holding the
timestamp has gone and the RTT sample is lost.
This is particularly noticeable on RPC calls used to probe the service
offered by the peer. A lot of peers end up with an unknown RTT because we
only ever sent a single RPC. This confuses the server rotation algorithm.
Fix this by caching the information about the outgoing packet in RTT
calculations in the rxrpc_call struct rather than looking in the Tx ring.
A four-deep buffer is maintained and both REQUEST_ACK-flagged DATA and
PING-ACK transmissions are recorded in there. When the appropriate
response ACK is received, the buffer is checked for a match and, if found,
an RTT sample is recorded.
If a received ACK refers to a packet with a later serial number than an
entry in the cache, that entry is presumed lost and the entry is made
available to record a new transmission.
ACKs types other than REQUESTED-type and PING-type cause any matching
sample to be cancelled as they don't necessarily represent a useful
measurement.
If there's no space in the buffer on ping/data transmission, the sample
base is discarded.
Fixes: 50235c4b5a2f ("rxrpc: Obtain RTT data by requesting ACKs on DATA packets")
Signed-off-by: David Howells <dhowells@redhat.com>
2020-08-19 22:29:16 +00:00
|
|
|
sent_at = call->rtt_sent_at[i];
|
|
|
|
|
orig_serial = call->rtt_serial[i];
|
|
|
|
|
|
|
|
|
|
if (orig_serial == acked_serial) {
|
|
|
|
|
clear_bit(i + RXRPC_CALL_RTT_PEND_SHIFT, &call->rtt_avail);
|
|
|
|
|
smp_mb(); /* Read data before setting avail bit */
|
|
|
|
|
set_bit(i, &call->rtt_avail);
|
2024-12-04 07:47:05 +00:00
|
|
|
rxrpc_call_add_rtt(call, type, i, acked_serial, ack_serial,
|
2023-11-16 13:12:58 +00:00
|
|
|
sent_at, resp_time);
|
rxrpc: Fix loss of RTT samples due to interposed ACK
The Rx protocol has a mechanism to help generate RTT samples that works by
a client transmitting a REQUESTED-type ACK when it receives a DATA packet
that has the REQUEST_ACK flag set.
The peer, however, may interpose other ACKs before transmitting the
REQUESTED-ACK, as can be seen in the following trace excerpt:
rxrpc_tx_data: c=00000044 DATA d0b5ece8:00000001 00000001 q=00000001 fl=07
rxrpc_rx_ack: c=00000044 00000001 PNG r=00000000 f=00000002 p=00000000 n=0
rxrpc_rx_ack: c=00000044 00000002 REQ r=00000001 f=00000002 p=00000001 n=0
...
DATA packet 1 (q=xx) has REQUEST_ACK set (bit 1 of fl=xx). The incoming
ping (labelled PNG) hard-acks the request DATA packet (f=xx exceeds the
sequence number of the DATA packet), causing it to be discarded from the Tx
ring. The ACK that was requested (labelled REQ, r=xx references the serial
of the DATA packet) comes after the ping, but the sk_buff holding the
timestamp has gone and the RTT sample is lost.
This is particularly noticeable on RPC calls used to probe the service
offered by the peer. A lot of peers end up with an unknown RTT because we
only ever sent a single RPC. This confuses the server rotation algorithm.
Fix this by caching the information about the outgoing packet in RTT
calculations in the rxrpc_call struct rather than looking in the Tx ring.
A four-deep buffer is maintained and both REQUEST_ACK-flagged DATA and
PING-ACK transmissions are recorded in there. When the appropriate
response ACK is received, the buffer is checked for a match and, if found,
an RTT sample is recorded.
If a received ACK refers to a packet with a later serial number than an
entry in the cache, that entry is presumed lost and the entry is made
available to record a new transmission.
ACKs types other than REQUESTED-type and PING-type cause any matching
sample to be cancelled as they don't necessarily represent a useful
measurement.
If there's no space in the buffer on ping/data transmission, the sample
base is discarded.
Fixes: 50235c4b5a2f ("rxrpc: Obtain RTT data by requesting ACKs on DATA packets")
Signed-off-by: David Howells <dhowells@redhat.com>
2020-08-19 22:29:16 +00:00
|
|
|
matched = true;
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
/* If a later serial is being acked, then mark this slot as
|
|
|
|
|
* being available.
|
|
|
|
|
*/
|
|
|
|
|
if (after(acked_serial, orig_serial)) {
|
|
|
|
|
trace_rxrpc_rtt_rx(call, rxrpc_rtt_rx_obsolete, i,
|
2024-12-04 07:46:55 +00:00
|
|
|
orig_serial, acked_serial, 0, 0, 0);
|
rxrpc: Fix loss of RTT samples due to interposed ACK
The Rx protocol has a mechanism to help generate RTT samples that works by
a client transmitting a REQUESTED-type ACK when it receives a DATA packet
that has the REQUEST_ACK flag set.
The peer, however, may interpose other ACKs before transmitting the
REQUESTED-ACK, as can be seen in the following trace excerpt:
rxrpc_tx_data: c=00000044 DATA d0b5ece8:00000001 00000001 q=00000001 fl=07
rxrpc_rx_ack: c=00000044 00000001 PNG r=00000000 f=00000002 p=00000000 n=0
rxrpc_rx_ack: c=00000044 00000002 REQ r=00000001 f=00000002 p=00000001 n=0
...
DATA packet 1 (q=xx) has REQUEST_ACK set (bit 1 of fl=xx). The incoming
ping (labelled PNG) hard-acks the request DATA packet (f=xx exceeds the
sequence number of the DATA packet), causing it to be discarded from the Tx
ring. The ACK that was requested (labelled REQ, r=xx references the serial
of the DATA packet) comes after the ping, but the sk_buff holding the
timestamp has gone and the RTT sample is lost.
This is particularly noticeable on RPC calls used to probe the service
offered by the peer. A lot of peers end up with an unknown RTT because we
only ever sent a single RPC. This confuses the server rotation algorithm.
Fix this by caching the information about the outgoing packet in RTT
calculations in the rxrpc_call struct rather than looking in the Tx ring.
A four-deep buffer is maintained and both REQUEST_ACK-flagged DATA and
PING-ACK transmissions are recorded in there. When the appropriate
response ACK is received, the buffer is checked for a match and, if found,
an RTT sample is recorded.
If a received ACK refers to a packet with a later serial number than an
entry in the cache, that entry is presumed lost and the entry is made
available to record a new transmission.
ACKs types other than REQUESTED-type and PING-type cause any matching
sample to be cancelled as they don't necessarily represent a useful
measurement.
If there's no space in the buffer on ping/data transmission, the sample
base is discarded.
Fixes: 50235c4b5a2f ("rxrpc: Obtain RTT data by requesting ACKs on DATA packets")
Signed-off-by: David Howells <dhowells@redhat.com>
2020-08-19 22:29:16 +00:00
|
|
|
clear_bit(i + RXRPC_CALL_RTT_PEND_SHIFT, &call->rtt_avail);
|
|
|
|
|
smp_wmb();
|
|
|
|
|
set_bit(i, &call->rtt_avail);
|
|
|
|
|
}
|
|
|
|
|
}
|
2016-09-21 23:29:31 +00:00
|
|
|
|
rxrpc: Fix loss of RTT samples due to interposed ACK
The Rx protocol has a mechanism to help generate RTT samples that works by
a client transmitting a REQUESTED-type ACK when it receives a DATA packet
that has the REQUEST_ACK flag set.
The peer, however, may interpose other ACKs before transmitting the
REQUESTED-ACK, as can be seen in the following trace excerpt:
rxrpc_tx_data: c=00000044 DATA d0b5ece8:00000001 00000001 q=00000001 fl=07
rxrpc_rx_ack: c=00000044 00000001 PNG r=00000000 f=00000002 p=00000000 n=0
rxrpc_rx_ack: c=00000044 00000002 REQ r=00000001 f=00000002 p=00000001 n=0
...
DATA packet 1 (q=xx) has REQUEST_ACK set (bit 1 of fl=xx). The incoming
ping (labelled PNG) hard-acks the request DATA packet (f=xx exceeds the
sequence number of the DATA packet), causing it to be discarded from the Tx
ring. The ACK that was requested (labelled REQ, r=xx references the serial
of the DATA packet) comes after the ping, but the sk_buff holding the
timestamp has gone and the RTT sample is lost.
This is particularly noticeable on RPC calls used to probe the service
offered by the peer. A lot of peers end up with an unknown RTT because we
only ever sent a single RPC. This confuses the server rotation algorithm.
Fix this by caching the information about the outgoing packet in RTT
calculations in the rxrpc_call struct rather than looking in the Tx ring.
A four-deep buffer is maintained and both REQUEST_ACK-flagged DATA and
PING-ACK transmissions are recorded in there. When the appropriate
response ACK is received, the buffer is checked for a match and, if found,
an RTT sample is recorded.
If a received ACK refers to a packet with a later serial number than an
entry in the cache, that entry is presumed lost and the entry is made
available to record a new transmission.
ACKs types other than REQUESTED-type and PING-type cause any matching
sample to be cancelled as they don't necessarily represent a useful
measurement.
If there's no space in the buffer on ping/data transmission, the sample
base is discarded.
Fixes: 50235c4b5a2f ("rxrpc: Obtain RTT data by requesting ACKs on DATA packets")
Signed-off-by: David Howells <dhowells@redhat.com>
2020-08-19 22:29:16 +00:00
|
|
|
if (!matched)
|
2024-12-04 07:46:55 +00:00
|
|
|
trace_rxrpc_rtt_rx(call, rxrpc_rtt_rx_lost, 9, 0, acked_serial, 0, 0, 0);
|
2016-09-21 23:29:31 +00:00
|
|
|
}
|
|
|
|
|
|
2007-04-26 22:48:28 +00:00
|
|
|
/*
|
rxrpc: Rewrite the data and ack handling code
Rewrite the data and ack handling code such that:
(1) Parsing of received ACK and ABORT packets and the distribution and the
filing of DATA packets happens entirely within the data_ready context
called from the UDP socket. This allows us to process and discard ACK
and ABORT packets much more quickly (they're no longer stashed on a
queue for a background thread to process).
(2) We avoid calling skb_clone(), pskb_pull() and pskb_trim(). We instead
keep track of the offset and length of the content of each packet in
the sk_buff metadata. This means we don't do any allocation in the
receive path.
(3) Jumbo DATA packet parsing is now done in data_ready context. Rather
than cloning the packet once for each subpacket and pulling/trimming
it, we file the packet multiple times with an annotation for each
indicating which subpacket is there. From that we can directly
calculate the offset and length.
(4) A call's receive queue can be accessed without taking locks (memory
barriers do have to be used, though).
(5) Incoming calls are set up from preallocated resources and immediately
made live. They can than have packets queued upon them and ACKs
generated. If insufficient resources exist, DATA packet #1 is given a
BUSY reply and other DATA packets are discarded).
(6) sk_buffs no longer take a ref on their parent call.
To make this work, the following changes are made:
(1) Each call's receive buffer is now a circular buffer of sk_buff
pointers (rxtx_buffer) rather than a number of sk_buff_heads spread
between the call and the socket. This permits each sk_buff to be in
the buffer multiple times. The receive buffer is reused for the
transmit buffer.
(2) A circular buffer of annotations (rxtx_annotations) is kept parallel
to the data buffer. Transmission phase annotations indicate whether a
buffered packet has been ACK'd or not and whether it needs
retransmission.
Receive phase annotations indicate whether a slot holds a whole packet
or a jumbo subpacket and, if the latter, which subpacket. They also
note whether the packet has been decrypted in place.
(3) DATA packet window tracking is much simplified. Each phase has just
two numbers representing the window (rx_hard_ack/rx_top and
tx_hard_ack/tx_top).
The hard_ack number is the sequence number before base of the window,
representing the last packet the other side says it has consumed.
hard_ack starts from 0 and the first packet is sequence number 1.
The top number is the sequence number of the highest-numbered packet
residing in the buffer. Packets between hard_ack+1 and top are
soft-ACK'd to indicate they've been received, but not yet consumed.
Four macros, before(), before_eq(), after() and after_eq() are added
to compare sequence numbers within the window. This allows for the
top of the window to wrap when the hard-ack sequence number gets close
to the limit.
Two flags, RXRPC_CALL_RX_LAST and RXRPC_CALL_TX_LAST, are added also
to indicate when rx_top and tx_top point at the packets with the
LAST_PACKET bit set, indicating the end of the phase.
(4) Calls are queued on the socket 'receive queue' rather than packets.
This means that we don't need have to invent dummy packets to queue to
indicate abnormal/terminal states and we don't have to keep metadata
packets (such as ABORTs) around
(5) The offset and length of a (sub)packet's content are now passed to
the verify_packet security op. This is currently expected to decrypt
the packet in place and validate it.
However, there's now nowhere to store the revised offset and length of
the actual data within the decrypted blob (there may be a header and
padding to skip) because an sk_buff may represent multiple packets, so
a locate_data security op is added to retrieve these details from the
sk_buff content when needed.
(6) recvmsg() now has to handle jumbo subpackets, where each subpacket is
individually secured and needs to be individually decrypted. The code
to do this is broken out into rxrpc_recvmsg_data() and shared with the
kernel API. It now iterates over the call's receive buffer rather
than walking the socket receive queue.
Additional changes:
(1) The timers are condensed to a single timer that is set for the soonest
of three timeouts (delayed ACK generation, DATA retransmission and
call lifespan).
(2) Transmission of ACK and ABORT packets is effected immediately from
process-context socket ops/kernel API calls that cause them instead of
them being punted off to a background work item. The data_ready
handler still has to defer to the background, though.
(3) A shutdown op is added to the AF_RXRPC socket so that the AFS
filesystem can shut down the socket and flush its own work items
before closing the socket to deal with any in-progress service calls.
Future additional changes that will need to be considered:
(1) Make sure that a call doesn't hog the front of the queue by receiving
data from the network as fast as userspace is consuming it to the
exclusion of other calls.
(2) Transmit delayed ACKs from within recvmsg() when we've consumed
sufficiently more packets to avoid the background work item needing to
run.
Signed-off-by: David Howells <dhowells@redhat.com>
2016-09-08 10:10:12 +00:00
|
|
|
* Process the extra information that may be appended to an ACK packet
|
2007-04-26 22:48:28 +00:00
|
|
|
*/
|
2024-01-26 16:17:03 +00:00
|
|
|
static void rxrpc_input_ack_trailer(struct rxrpc_call *call, struct sk_buff *skb,
|
|
|
|
|
struct rxrpc_acktrailer *trailer)
|
2007-04-26 22:48:28 +00:00
|
|
|
{
|
rxrpc: Rewrite the data and ack handling code
Rewrite the data and ack handling code such that:
(1) Parsing of received ACK and ABORT packets and the distribution and the
filing of DATA packets happens entirely within the data_ready context
called from the UDP socket. This allows us to process and discard ACK
and ABORT packets much more quickly (they're no longer stashed on a
queue for a background thread to process).
(2) We avoid calling skb_clone(), pskb_pull() and pskb_trim(). We instead
keep track of the offset and length of the content of each packet in
the sk_buff metadata. This means we don't do any allocation in the
receive path.
(3) Jumbo DATA packet parsing is now done in data_ready context. Rather
than cloning the packet once for each subpacket and pulling/trimming
it, we file the packet multiple times with an annotation for each
indicating which subpacket is there. From that we can directly
calculate the offset and length.
(4) A call's receive queue can be accessed without taking locks (memory
barriers do have to be used, though).
(5) Incoming calls are set up from preallocated resources and immediately
made live. They can than have packets queued upon them and ACKs
generated. If insufficient resources exist, DATA packet #1 is given a
BUSY reply and other DATA packets are discarded).
(6) sk_buffs no longer take a ref on their parent call.
To make this work, the following changes are made:
(1) Each call's receive buffer is now a circular buffer of sk_buff
pointers (rxtx_buffer) rather than a number of sk_buff_heads spread
between the call and the socket. This permits each sk_buff to be in
the buffer multiple times. The receive buffer is reused for the
transmit buffer.
(2) A circular buffer of annotations (rxtx_annotations) is kept parallel
to the data buffer. Transmission phase annotations indicate whether a
buffered packet has been ACK'd or not and whether it needs
retransmission.
Receive phase annotations indicate whether a slot holds a whole packet
or a jumbo subpacket and, if the latter, which subpacket. They also
note whether the packet has been decrypted in place.
(3) DATA packet window tracking is much simplified. Each phase has just
two numbers representing the window (rx_hard_ack/rx_top and
tx_hard_ack/tx_top).
The hard_ack number is the sequence number before base of the window,
representing the last packet the other side says it has consumed.
hard_ack starts from 0 and the first packet is sequence number 1.
The top number is the sequence number of the highest-numbered packet
residing in the buffer. Packets between hard_ack+1 and top are
soft-ACK'd to indicate they've been received, but not yet consumed.
Four macros, before(), before_eq(), after() and after_eq() are added
to compare sequence numbers within the window. This allows for the
top of the window to wrap when the hard-ack sequence number gets close
to the limit.
Two flags, RXRPC_CALL_RX_LAST and RXRPC_CALL_TX_LAST, are added also
to indicate when rx_top and tx_top point at the packets with the
LAST_PACKET bit set, indicating the end of the phase.
(4) Calls are queued on the socket 'receive queue' rather than packets.
This means that we don't need have to invent dummy packets to queue to
indicate abnormal/terminal states and we don't have to keep metadata
packets (such as ABORTs) around
(5) The offset and length of a (sub)packet's content are now passed to
the verify_packet security op. This is currently expected to decrypt
the packet in place and validate it.
However, there's now nowhere to store the revised offset and length of
the actual data within the decrypted blob (there may be a header and
padding to skip) because an sk_buff may represent multiple packets, so
a locate_data security op is added to retrieve these details from the
sk_buff content when needed.
(6) recvmsg() now has to handle jumbo subpackets, where each subpacket is
individually secured and needs to be individually decrypted. The code
to do this is broken out into rxrpc_recvmsg_data() and shared with the
kernel API. It now iterates over the call's receive buffer rather
than walking the socket receive queue.
Additional changes:
(1) The timers are condensed to a single timer that is set for the soonest
of three timeouts (delayed ACK generation, DATA retransmission and
call lifespan).
(2) Transmission of ACK and ABORT packets is effected immediately from
process-context socket ops/kernel API calls that cause them instead of
them being punted off to a background work item. The data_ready
handler still has to defer to the background, though.
(3) A shutdown op is added to the AF_RXRPC socket so that the AFS
filesystem can shut down the socket and flush its own work items
before closing the socket to deal with any in-progress service calls.
Future additional changes that will need to be considered:
(1) Make sure that a call doesn't hog the front of the queue by receiving
data from the network as fast as userspace is consuming it to the
exclusion of other calls.
(2) Transmit delayed ACKs from within recvmsg() when we've consumed
sufficiently more packets to avoid the background work item needing to
run.
Signed-off-by: David Howells <dhowells@redhat.com>
2016-09-08 10:10:12 +00:00
|
|
|
struct rxrpc_skb_priv *sp = rxrpc_skb(skb);
|
2024-12-04 07:46:37 +00:00
|
|
|
struct rxrpc_peer *peer = call->peer;
|
2024-12-04 07:46:39 +00:00
|
|
|
unsigned int max_data, capacity;
|
2017-03-10 07:48:49 +00:00
|
|
|
bool wake = false;
|
2024-12-04 07:46:39 +00:00
|
|
|
u32 max_mtu = ntohl(trailer->maxMTU);
|
|
|
|
|
//u32 if_mtu = ntohl(trailer->ifMTU);
|
|
|
|
|
u32 rwind = ntohl(trailer->rwind);
|
|
|
|
|
u32 jumbo_max = ntohl(trailer->jumbo_max);
|
rxrpc: Rewrite the data and ack handling code
Rewrite the data and ack handling code such that:
(1) Parsing of received ACK and ABORT packets and the distribution and the
filing of DATA packets happens entirely within the data_ready context
called from the UDP socket. This allows us to process and discard ACK
and ABORT packets much more quickly (they're no longer stashed on a
queue for a background thread to process).
(2) We avoid calling skb_clone(), pskb_pull() and pskb_trim(). We instead
keep track of the offset and length of the content of each packet in
the sk_buff metadata. This means we don't do any allocation in the
receive path.
(3) Jumbo DATA packet parsing is now done in data_ready context. Rather
than cloning the packet once for each subpacket and pulling/trimming
it, we file the packet multiple times with an annotation for each
indicating which subpacket is there. From that we can directly
calculate the offset and length.
(4) A call's receive queue can be accessed without taking locks (memory
barriers do have to be used, though).
(5) Incoming calls are set up from preallocated resources and immediately
made live. They can than have packets queued upon them and ACKs
generated. If insufficient resources exist, DATA packet #1 is given a
BUSY reply and other DATA packets are discarded).
(6) sk_buffs no longer take a ref on their parent call.
To make this work, the following changes are made:
(1) Each call's receive buffer is now a circular buffer of sk_buff
pointers (rxtx_buffer) rather than a number of sk_buff_heads spread
between the call and the socket. This permits each sk_buff to be in
the buffer multiple times. The receive buffer is reused for the
transmit buffer.
(2) A circular buffer of annotations (rxtx_annotations) is kept parallel
to the data buffer. Transmission phase annotations indicate whether a
buffered packet has been ACK'd or not and whether it needs
retransmission.
Receive phase annotations indicate whether a slot holds a whole packet
or a jumbo subpacket and, if the latter, which subpacket. They also
note whether the packet has been decrypted in place.
(3) DATA packet window tracking is much simplified. Each phase has just
two numbers representing the window (rx_hard_ack/rx_top and
tx_hard_ack/tx_top).
The hard_ack number is the sequence number before base of the window,
representing the last packet the other side says it has consumed.
hard_ack starts from 0 and the first packet is sequence number 1.
The top number is the sequence number of the highest-numbered packet
residing in the buffer. Packets between hard_ack+1 and top are
soft-ACK'd to indicate they've been received, but not yet consumed.
Four macros, before(), before_eq(), after() and after_eq() are added
to compare sequence numbers within the window. This allows for the
top of the window to wrap when the hard-ack sequence number gets close
to the limit.
Two flags, RXRPC_CALL_RX_LAST and RXRPC_CALL_TX_LAST, are added also
to indicate when rx_top and tx_top point at the packets with the
LAST_PACKET bit set, indicating the end of the phase.
(4) Calls are queued on the socket 'receive queue' rather than packets.
This means that we don't need have to invent dummy packets to queue to
indicate abnormal/terminal states and we don't have to keep metadata
packets (such as ABORTs) around
(5) The offset and length of a (sub)packet's content are now passed to
the verify_packet security op. This is currently expected to decrypt
the packet in place and validate it.
However, there's now nowhere to store the revised offset and length of
the actual data within the decrypted blob (there may be a header and
padding to skip) because an sk_buff may represent multiple packets, so
a locate_data security op is added to retrieve these details from the
sk_buff content when needed.
(6) recvmsg() now has to handle jumbo subpackets, where each subpacket is
individually secured and needs to be individually decrypted. The code
to do this is broken out into rxrpc_recvmsg_data() and shared with the
kernel API. It now iterates over the call's receive buffer rather
than walking the socket receive queue.
Additional changes:
(1) The timers are condensed to a single timer that is set for the soonest
of three timeouts (delayed ACK generation, DATA retransmission and
call lifespan).
(2) Transmission of ACK and ABORT packets is effected immediately from
process-context socket ops/kernel API calls that cause them instead of
them being punted off to a background work item. The data_ready
handler still has to defer to the background, though.
(3) A shutdown op is added to the AF_RXRPC socket so that the AFS
filesystem can shut down the socket and flush its own work items
before closing the socket to deal with any in-progress service calls.
Future additional changes that will need to be considered:
(1) Make sure that a call doesn't hog the front of the queue by receiving
data from the network as fast as userspace is consuming it to the
exclusion of other calls.
(2) Transmit delayed ACKs from within recvmsg() when we've consumed
sufficiently more packets to avoid the background work item needing to
run.
Signed-off-by: David Howells <dhowells@redhat.com>
2016-09-08 10:10:12 +00:00
|
|
|
|
2022-03-31 22:55:08 +00:00
|
|
|
if (rwind > RXRPC_TX_MAX_WINDOW)
|
|
|
|
|
rwind = RXRPC_TX_MAX_WINDOW;
|
2017-03-10 07:48:49 +00:00
|
|
|
if (call->tx_winsize != rwind) {
|
|
|
|
|
if (rwind > call->tx_winsize)
|
|
|
|
|
wake = true;
|
2020-06-17 22:01:23 +00:00
|
|
|
trace_rxrpc_rx_rwind_change(call, sp->hdr.serial, rwind, wake);
|
2017-03-10 07:48:49 +00:00
|
|
|
call->tx_winsize = rwind;
|
|
|
|
|
}
|
|
|
|
|
|
2024-12-04 07:46:39 +00:00
|
|
|
max_mtu = clamp(max_mtu, 500, 65535);
|
|
|
|
|
peer->ackr_max_data = max_mtu;
|
rxrpc: Rewrite the data and ack handling code
Rewrite the data and ack handling code such that:
(1) Parsing of received ACK and ABORT packets and the distribution and the
filing of DATA packets happens entirely within the data_ready context
called from the UDP socket. This allows us to process and discard ACK
and ABORT packets much more quickly (they're no longer stashed on a
queue for a background thread to process).
(2) We avoid calling skb_clone(), pskb_pull() and pskb_trim(). We instead
keep track of the offset and length of the content of each packet in
the sk_buff metadata. This means we don't do any allocation in the
receive path.
(3) Jumbo DATA packet parsing is now done in data_ready context. Rather
than cloning the packet once for each subpacket and pulling/trimming
it, we file the packet multiple times with an annotation for each
indicating which subpacket is there. From that we can directly
calculate the offset and length.
(4) A call's receive queue can be accessed without taking locks (memory
barriers do have to be used, though).
(5) Incoming calls are set up from preallocated resources and immediately
made live. They can than have packets queued upon them and ACKs
generated. If insufficient resources exist, DATA packet #1 is given a
BUSY reply and other DATA packets are discarded).
(6) sk_buffs no longer take a ref on their parent call.
To make this work, the following changes are made:
(1) Each call's receive buffer is now a circular buffer of sk_buff
pointers (rxtx_buffer) rather than a number of sk_buff_heads spread
between the call and the socket. This permits each sk_buff to be in
the buffer multiple times. The receive buffer is reused for the
transmit buffer.
(2) A circular buffer of annotations (rxtx_annotations) is kept parallel
to the data buffer. Transmission phase annotations indicate whether a
buffered packet has been ACK'd or not and whether it needs
retransmission.
Receive phase annotations indicate whether a slot holds a whole packet
or a jumbo subpacket and, if the latter, which subpacket. They also
note whether the packet has been decrypted in place.
(3) DATA packet window tracking is much simplified. Each phase has just
two numbers representing the window (rx_hard_ack/rx_top and
tx_hard_ack/tx_top).
The hard_ack number is the sequence number before base of the window,
representing the last packet the other side says it has consumed.
hard_ack starts from 0 and the first packet is sequence number 1.
The top number is the sequence number of the highest-numbered packet
residing in the buffer. Packets between hard_ack+1 and top are
soft-ACK'd to indicate they've been received, but not yet consumed.
Four macros, before(), before_eq(), after() and after_eq() are added
to compare sequence numbers within the window. This allows for the
top of the window to wrap when the hard-ack sequence number gets close
to the limit.
Two flags, RXRPC_CALL_RX_LAST and RXRPC_CALL_TX_LAST, are added also
to indicate when rx_top and tx_top point at the packets with the
LAST_PACKET bit set, indicating the end of the phase.
(4) Calls are queued on the socket 'receive queue' rather than packets.
This means that we don't need have to invent dummy packets to queue to
indicate abnormal/terminal states and we don't have to keep metadata
packets (such as ABORTs) around
(5) The offset and length of a (sub)packet's content are now passed to
the verify_packet security op. This is currently expected to decrypt
the packet in place and validate it.
However, there's now nowhere to store the revised offset and length of
the actual data within the decrypted blob (there may be a header and
padding to skip) because an sk_buff may represent multiple packets, so
a locate_data security op is added to retrieve these details from the
sk_buff content when needed.
(6) recvmsg() now has to handle jumbo subpackets, where each subpacket is
individually secured and needs to be individually decrypted. The code
to do this is broken out into rxrpc_recvmsg_data() and shared with the
kernel API. It now iterates over the call's receive buffer rather
than walking the socket receive queue.
Additional changes:
(1) The timers are condensed to a single timer that is set for the soonest
of three timeouts (delayed ACK generation, DATA retransmission and
call lifespan).
(2) Transmission of ACK and ABORT packets is effected immediately from
process-context socket ops/kernel API calls that cause them instead of
them being punted off to a background work item. The data_ready
handler still has to defer to the background, though.
(3) A shutdown op is added to the AF_RXRPC socket so that the AFS
filesystem can shut down the socket and flush its own work items
before closing the socket to deal with any in-progress service calls.
Future additional changes that will need to be considered:
(1) Make sure that a call doesn't hog the front of the queue by receiving
data from the network as fast as userspace is consuming it to the
exclusion of other calls.
(2) Transmit delayed ACKs from within recvmsg() when we've consumed
sufficiently more packets to avoid the background work item needing to
run.
Signed-off-by: David Howells <dhowells@redhat.com>
2016-09-08 10:10:12 +00:00
|
|
|
|
2024-12-04 07:46:39 +00:00
|
|
|
if (max_mtu < peer->max_data) {
|
|
|
|
|
trace_rxrpc_pmtud_reduce(peer, sp->hdr.serial, max_mtu,
|
2024-12-04 07:46:37 +00:00
|
|
|
rxrpc_pmtud_reduce_ack);
|
|
|
|
|
write_seqcount_begin(&peer->mtu_lock);
|
2024-12-04 07:46:39 +00:00
|
|
|
peer->max_data = max_mtu;
|
2024-12-04 07:46:37 +00:00
|
|
|
write_seqcount_end(&peer->mtu_lock);
|
rxrpc: Rewrite the data and ack handling code
Rewrite the data and ack handling code such that:
(1) Parsing of received ACK and ABORT packets and the distribution and the
filing of DATA packets happens entirely within the data_ready context
called from the UDP socket. This allows us to process and discard ACK
and ABORT packets much more quickly (they're no longer stashed on a
queue for a background thread to process).
(2) We avoid calling skb_clone(), pskb_pull() and pskb_trim(). We instead
keep track of the offset and length of the content of each packet in
the sk_buff metadata. This means we don't do any allocation in the
receive path.
(3) Jumbo DATA packet parsing is now done in data_ready context. Rather
than cloning the packet once for each subpacket and pulling/trimming
it, we file the packet multiple times with an annotation for each
indicating which subpacket is there. From that we can directly
calculate the offset and length.
(4) A call's receive queue can be accessed without taking locks (memory
barriers do have to be used, though).
(5) Incoming calls are set up from preallocated resources and immediately
made live. They can than have packets queued upon them and ACKs
generated. If insufficient resources exist, DATA packet #1 is given a
BUSY reply and other DATA packets are discarded).
(6) sk_buffs no longer take a ref on their parent call.
To make this work, the following changes are made:
(1) Each call's receive buffer is now a circular buffer of sk_buff
pointers (rxtx_buffer) rather than a number of sk_buff_heads spread
between the call and the socket. This permits each sk_buff to be in
the buffer multiple times. The receive buffer is reused for the
transmit buffer.
(2) A circular buffer of annotations (rxtx_annotations) is kept parallel
to the data buffer. Transmission phase annotations indicate whether a
buffered packet has been ACK'd or not and whether it needs
retransmission.
Receive phase annotations indicate whether a slot holds a whole packet
or a jumbo subpacket and, if the latter, which subpacket. They also
note whether the packet has been decrypted in place.
(3) DATA packet window tracking is much simplified. Each phase has just
two numbers representing the window (rx_hard_ack/rx_top and
tx_hard_ack/tx_top).
The hard_ack number is the sequence number before base of the window,
representing the last packet the other side says it has consumed.
hard_ack starts from 0 and the first packet is sequence number 1.
The top number is the sequence number of the highest-numbered packet
residing in the buffer. Packets between hard_ack+1 and top are
soft-ACK'd to indicate they've been received, but not yet consumed.
Four macros, before(), before_eq(), after() and after_eq() are added
to compare sequence numbers within the window. This allows for the
top of the window to wrap when the hard-ack sequence number gets close
to the limit.
Two flags, RXRPC_CALL_RX_LAST and RXRPC_CALL_TX_LAST, are added also
to indicate when rx_top and tx_top point at the packets with the
LAST_PACKET bit set, indicating the end of the phase.
(4) Calls are queued on the socket 'receive queue' rather than packets.
This means that we don't need have to invent dummy packets to queue to
indicate abnormal/terminal states and we don't have to keep metadata
packets (such as ABORTs) around
(5) The offset and length of a (sub)packet's content are now passed to
the verify_packet security op. This is currently expected to decrypt
the packet in place and validate it.
However, there's now nowhere to store the revised offset and length of
the actual data within the decrypted blob (there may be a header and
padding to skip) because an sk_buff may represent multiple packets, so
a locate_data security op is added to retrieve these details from the
sk_buff content when needed.
(6) recvmsg() now has to handle jumbo subpackets, where each subpacket is
individually secured and needs to be individually decrypted. The code
to do this is broken out into rxrpc_recvmsg_data() and shared with the
kernel API. It now iterates over the call's receive buffer rather
than walking the socket receive queue.
Additional changes:
(1) The timers are condensed to a single timer that is set for the soonest
of three timeouts (delayed ACK generation, DATA retransmission and
call lifespan).
(2) Transmission of ACK and ABORT packets is effected immediately from
process-context socket ops/kernel API calls that cause them instead of
them being punted off to a background work item. The data_ready
handler still has to defer to the background, though.
(3) A shutdown op is added to the AF_RXRPC socket so that the AFS
filesystem can shut down the socket and flush its own work items
before closing the socket to deal with any in-progress service calls.
Future additional changes that will need to be considered:
(1) Make sure that a call doesn't hog the front of the queue by receiving
data from the network as fast as userspace is consuming it to the
exclusion of other calls.
(2) Transmit delayed ACKs from within recvmsg() when we've consumed
sufficiently more packets to avoid the background work item needing to
run.
Signed-off-by: David Howells <dhowells@redhat.com>
2016-09-08 10:10:12 +00:00
|
|
|
}
|
2017-03-10 07:48:49 +00:00
|
|
|
|
2024-12-04 07:46:39 +00:00
|
|
|
max_data = umin(max_mtu, peer->max_data);
|
|
|
|
|
capacity = max_data;
|
|
|
|
|
capacity += sizeof(struct rxrpc_jumbo_header); /* First subpacket has main hdr, not jumbo */
|
|
|
|
|
capacity /= sizeof(struct rxrpc_jumbo_header) + RXRPC_JUMBO_DATALEN;
|
|
|
|
|
|
|
|
|
|
if (jumbo_max == 0) {
|
|
|
|
|
/* The peer says it supports pmtu discovery */
|
|
|
|
|
peer->ackr_adv_pmtud = true;
|
|
|
|
|
} else {
|
|
|
|
|
peer->ackr_adv_pmtud = false;
|
2024-12-04 07:46:59 +00:00
|
|
|
capacity = clamp(capacity, 1, jumbo_max);
|
2024-12-04 07:46:39 +00:00
|
|
|
}
|
|
|
|
|
|
2024-12-04 07:46:59 +00:00
|
|
|
call->tx_jumbo_max = capacity;
|
|
|
|
|
|
2017-03-10 07:48:49 +00:00
|
|
|
if (wake)
|
|
|
|
|
wake_up(&call->waitq);
|
rxrpc: Rewrite the data and ack handling code
Rewrite the data and ack handling code such that:
(1) Parsing of received ACK and ABORT packets and the distribution and the
filing of DATA packets happens entirely within the data_ready context
called from the UDP socket. This allows us to process and discard ACK
and ABORT packets much more quickly (they're no longer stashed on a
queue for a background thread to process).
(2) We avoid calling skb_clone(), pskb_pull() and pskb_trim(). We instead
keep track of the offset and length of the content of each packet in
the sk_buff metadata. This means we don't do any allocation in the
receive path.
(3) Jumbo DATA packet parsing is now done in data_ready context. Rather
than cloning the packet once for each subpacket and pulling/trimming
it, we file the packet multiple times with an annotation for each
indicating which subpacket is there. From that we can directly
calculate the offset and length.
(4) A call's receive queue can be accessed without taking locks (memory
barriers do have to be used, though).
(5) Incoming calls are set up from preallocated resources and immediately
made live. They can than have packets queued upon them and ACKs
generated. If insufficient resources exist, DATA packet #1 is given a
BUSY reply and other DATA packets are discarded).
(6) sk_buffs no longer take a ref on their parent call.
To make this work, the following changes are made:
(1) Each call's receive buffer is now a circular buffer of sk_buff
pointers (rxtx_buffer) rather than a number of sk_buff_heads spread
between the call and the socket. This permits each sk_buff to be in
the buffer multiple times. The receive buffer is reused for the
transmit buffer.
(2) A circular buffer of annotations (rxtx_annotations) is kept parallel
to the data buffer. Transmission phase annotations indicate whether a
buffered packet has been ACK'd or not and whether it needs
retransmission.
Receive phase annotations indicate whether a slot holds a whole packet
or a jumbo subpacket and, if the latter, which subpacket. They also
note whether the packet has been decrypted in place.
(3) DATA packet window tracking is much simplified. Each phase has just
two numbers representing the window (rx_hard_ack/rx_top and
tx_hard_ack/tx_top).
The hard_ack number is the sequence number before base of the window,
representing the last packet the other side says it has consumed.
hard_ack starts from 0 and the first packet is sequence number 1.
The top number is the sequence number of the highest-numbered packet
residing in the buffer. Packets between hard_ack+1 and top are
soft-ACK'd to indicate they've been received, but not yet consumed.
Four macros, before(), before_eq(), after() and after_eq() are added
to compare sequence numbers within the window. This allows for the
top of the window to wrap when the hard-ack sequence number gets close
to the limit.
Two flags, RXRPC_CALL_RX_LAST and RXRPC_CALL_TX_LAST, are added also
to indicate when rx_top and tx_top point at the packets with the
LAST_PACKET bit set, indicating the end of the phase.
(4) Calls are queued on the socket 'receive queue' rather than packets.
This means that we don't need have to invent dummy packets to queue to
indicate abnormal/terminal states and we don't have to keep metadata
packets (such as ABORTs) around
(5) The offset and length of a (sub)packet's content are now passed to
the verify_packet security op. This is currently expected to decrypt
the packet in place and validate it.
However, there's now nowhere to store the revised offset and length of
the actual data within the decrypted blob (there may be a header and
padding to skip) because an sk_buff may represent multiple packets, so
a locate_data security op is added to retrieve these details from the
sk_buff content when needed.
(6) recvmsg() now has to handle jumbo subpackets, where each subpacket is
individually secured and needs to be individually decrypted. The code
to do this is broken out into rxrpc_recvmsg_data() and shared with the
kernel API. It now iterates over the call's receive buffer rather
than walking the socket receive queue.
Additional changes:
(1) The timers are condensed to a single timer that is set for the soonest
of three timeouts (delayed ACK generation, DATA retransmission and
call lifespan).
(2) Transmission of ACK and ABORT packets is effected immediately from
process-context socket ops/kernel API calls that cause them instead of
them being punted off to a background work item. The data_ready
handler still has to defer to the background, though.
(3) A shutdown op is added to the AF_RXRPC socket so that the AFS
filesystem can shut down the socket and flush its own work items
before closing the socket to deal with any in-progress service calls.
Future additional changes that will need to be considered:
(1) Make sure that a call doesn't hog the front of the queue by receiving
data from the network as fast as userspace is consuming it to the
exclusion of other calls.
(2) Transmit delayed ACKs from within recvmsg() when we've consumed
sufficiently more packets to avoid the background work item needing to
run.
Signed-off-by: David Howells <dhowells@redhat.com>
2016-09-08 10:10:12 +00:00
|
|
|
}
|
2007-04-26 22:48:28 +00:00
|
|
|
|
2024-12-04 07:46:51 +00:00
|
|
|
#if defined(CONFIG_X86) && __GNUC__ && !defined(__clang__)
|
|
|
|
|
/* Clang doesn't support the %z constraint modifier */
|
|
|
|
|
#define shiftr_adv_rotr(shift_from, rotate_into) ({ \
|
|
|
|
|
asm(" shr%z1 %1\n" \
|
|
|
|
|
" inc %0\n" \
|
|
|
|
|
" rcr%z2 %2\n" \
|
|
|
|
|
: "+d"(shift_from), "+m"(*(shift_from)), "+rm"(rotate_into) \
|
|
|
|
|
); \
|
|
|
|
|
})
|
|
|
|
|
#else
|
|
|
|
|
#define shiftr_adv_rotr(shift_from, rotate_into) ({ \
|
|
|
|
|
typeof(rotate_into) __bit0 = *(shift_from) & 1; \
|
|
|
|
|
*(shift_from) >>= 1; \
|
|
|
|
|
shift_from++; \
|
|
|
|
|
rotate_into >>= 1; \
|
|
|
|
|
rotate_into |= __bit0 << (sizeof(rotate_into) * 8 - 1); \
|
|
|
|
|
})
|
|
|
|
|
#endif
|
|
|
|
|
|
2024-12-04 07:46:52 +00:00
|
|
|
/*
|
|
|
|
|
* Deal with RTT samples from soft ACKs.
|
|
|
|
|
*/
|
|
|
|
|
static void rxrpc_input_soft_rtt(struct rxrpc_call *call,
|
|
|
|
|
struct rxrpc_ack_summary *summary,
|
2024-12-04 07:47:02 +00:00
|
|
|
struct rxrpc_txqueue *tq)
|
2024-12-04 07:46:52 +00:00
|
|
|
{
|
|
|
|
|
for (int ix = 0; ix < RXRPC_NR_TXQUEUE; ix++)
|
|
|
|
|
if (summary->acked_serial == tq->segment_serial[ix])
|
2024-12-04 07:47:02 +00:00
|
|
|
return rxrpc_add_data_rtt_sample(call, summary, tq, ix);
|
2024-12-04 07:46:52 +00:00
|
|
|
}
|
|
|
|
|
|
2024-02-02 15:19:16 +00:00
|
|
|
/*
|
2024-12-04 07:46:51 +00:00
|
|
|
* Process a batch of soft ACKs specific to a transmission queue segment.
|
2024-02-02 15:19:16 +00:00
|
|
|
*/
|
2024-12-04 07:46:51 +00:00
|
|
|
static void rxrpc_input_soft_ack_tq(struct rxrpc_call *call,
|
|
|
|
|
struct rxrpc_ack_summary *summary,
|
|
|
|
|
struct rxrpc_txqueue *tq,
|
|
|
|
|
unsigned long extracted_acks,
|
|
|
|
|
int nr_reported,
|
|
|
|
|
rxrpc_seq_t seq,
|
|
|
|
|
rxrpc_seq_t *lowest_nak)
|
2024-02-02 15:19:16 +00:00
|
|
|
{
|
2024-12-04 07:47:07 +00:00
|
|
|
unsigned long old_reported = 0, flipped, new_acks = 0;
|
|
|
|
|
unsigned long a_to_n, n_to_a = 0;
|
2024-12-04 07:46:51 +00:00
|
|
|
int new, a, n;
|
|
|
|
|
|
2024-12-04 07:47:07 +00:00
|
|
|
if (tq->nr_reported_acks > 0)
|
|
|
|
|
old_reported = ~0UL >> (RXRPC_NR_TXQUEUE - tq->nr_reported_acks);
|
|
|
|
|
|
2024-12-04 07:46:51 +00:00
|
|
|
_enter("{%x,%lx,%d},%lx,%d,%x",
|
|
|
|
|
tq->qbase, tq->segment_acked, tq->nr_reported_acks,
|
|
|
|
|
extracted_acks, nr_reported, seq);
|
|
|
|
|
|
|
|
|
|
_debug("[%x]", tq->qbase);
|
|
|
|
|
_debug("tq %16lx %u", tq->segment_acked, tq->nr_reported_acks);
|
|
|
|
|
_debug("sack %16lx %u", extracted_acks, nr_reported);
|
|
|
|
|
|
|
|
|
|
/* See how many previously logged ACKs/NAKs have flipped. */
|
|
|
|
|
flipped = (tq->segment_acked ^ extracted_acks) & old_reported;
|
|
|
|
|
if (flipped) {
|
|
|
|
|
n_to_a = ~tq->segment_acked & flipped; /* Old NAK -> ACK */
|
|
|
|
|
a_to_n = tq->segment_acked & flipped; /* Old ACK -> NAK */
|
|
|
|
|
a = hweight_long(n_to_a);
|
|
|
|
|
n = hweight_long(a_to_n);
|
|
|
|
|
_debug("flip %16lx", flipped);
|
|
|
|
|
_debug("ntoa %16lx %d", n_to_a, a);
|
|
|
|
|
_debug("aton %16lx %d", a_to_n, n);
|
|
|
|
|
call->acks_nr_sacks += a - n;
|
|
|
|
|
call->acks_nr_snacks += n - a;
|
|
|
|
|
summary->nr_new_sacks += a;
|
|
|
|
|
summary->nr_new_snacks += n;
|
|
|
|
|
}
|
2024-02-02 15:19:16 +00:00
|
|
|
|
2024-12-04 07:46:51 +00:00
|
|
|
/* See how many new ACKs/NAKs have been acquired. */
|
|
|
|
|
new = nr_reported - tq->nr_reported_acks;
|
|
|
|
|
if (new > 0) {
|
|
|
|
|
new_acks = extracted_acks & ~old_reported;
|
|
|
|
|
if (new_acks) {
|
|
|
|
|
a = hweight_long(new_acks);
|
|
|
|
|
n = new - a;
|
|
|
|
|
_debug("new_a %16lx new=%d a=%d n=%d", new_acks, new, a, n);
|
|
|
|
|
call->acks_nr_sacks += a;
|
|
|
|
|
call->acks_nr_snacks += n;
|
|
|
|
|
summary->nr_new_sacks += a;
|
|
|
|
|
summary->nr_new_snacks += n;
|
|
|
|
|
} else {
|
|
|
|
|
call->acks_nr_snacks += new;
|
|
|
|
|
summary->nr_new_snacks += new;
|
2024-02-02 15:19:16 +00:00
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
|
2024-12-04 07:46:51 +00:00
|
|
|
tq->nr_reported_acks = nr_reported;
|
|
|
|
|
tq->segment_acked = extracted_acks;
|
|
|
|
|
trace_rxrpc_apply_acks(call, tq);
|
|
|
|
|
|
|
|
|
|
if (extracted_acks != ~0UL) {
|
|
|
|
|
rxrpc_seq_t lowest = seq + ffz(extracted_acks);
|
|
|
|
|
|
|
|
|
|
if (before(lowest, *lowest_nak))
|
|
|
|
|
*lowest_nak = lowest;
|
|
|
|
|
}
|
2024-12-04 07:47:07 +00:00
|
|
|
|
|
|
|
|
if (summary->acked_serial)
|
|
|
|
|
rxrpc_input_soft_rtt(call, summary, tq);
|
|
|
|
|
|
|
|
|
|
new_acks |= n_to_a;
|
|
|
|
|
if (new_acks)
|
|
|
|
|
rxrpc_input_rack(call, summary, tq, new_acks);
|
|
|
|
|
|
|
|
|
|
if (call->tlp_serial &&
|
|
|
|
|
rxrpc_seq_in_txq(tq, call->tlp_seq) &&
|
|
|
|
|
test_bit(call->tlp_seq - tq->qbase, &new_acks))
|
|
|
|
|
summary->tlp_probe_acked = true;
|
2024-02-02 15:19:16 +00:00
|
|
|
}
|
|
|
|
|
|
rxrpc: Rewrite the data and ack handling code
Rewrite the data and ack handling code such that:
(1) Parsing of received ACK and ABORT packets and the distribution and the
filing of DATA packets happens entirely within the data_ready context
called from the UDP socket. This allows us to process and discard ACK
and ABORT packets much more quickly (they're no longer stashed on a
queue for a background thread to process).
(2) We avoid calling skb_clone(), pskb_pull() and pskb_trim(). We instead
keep track of the offset and length of the content of each packet in
the sk_buff metadata. This means we don't do any allocation in the
receive path.
(3) Jumbo DATA packet parsing is now done in data_ready context. Rather
than cloning the packet once for each subpacket and pulling/trimming
it, we file the packet multiple times with an annotation for each
indicating which subpacket is there. From that we can directly
calculate the offset and length.
(4) A call's receive queue can be accessed without taking locks (memory
barriers do have to be used, though).
(5) Incoming calls are set up from preallocated resources and immediately
made live. They can than have packets queued upon them and ACKs
generated. If insufficient resources exist, DATA packet #1 is given a
BUSY reply and other DATA packets are discarded).
(6) sk_buffs no longer take a ref on their parent call.
To make this work, the following changes are made:
(1) Each call's receive buffer is now a circular buffer of sk_buff
pointers (rxtx_buffer) rather than a number of sk_buff_heads spread
between the call and the socket. This permits each sk_buff to be in
the buffer multiple times. The receive buffer is reused for the
transmit buffer.
(2) A circular buffer of annotations (rxtx_annotations) is kept parallel
to the data buffer. Transmission phase annotations indicate whether a
buffered packet has been ACK'd or not and whether it needs
retransmission.
Receive phase annotations indicate whether a slot holds a whole packet
or a jumbo subpacket and, if the latter, which subpacket. They also
note whether the packet has been decrypted in place.
(3) DATA packet window tracking is much simplified. Each phase has just
two numbers representing the window (rx_hard_ack/rx_top and
tx_hard_ack/tx_top).
The hard_ack number is the sequence number before base of the window,
representing the last packet the other side says it has consumed.
hard_ack starts from 0 and the first packet is sequence number 1.
The top number is the sequence number of the highest-numbered packet
residing in the buffer. Packets between hard_ack+1 and top are
soft-ACK'd to indicate they've been received, but not yet consumed.
Four macros, before(), before_eq(), after() and after_eq() are added
to compare sequence numbers within the window. This allows for the
top of the window to wrap when the hard-ack sequence number gets close
to the limit.
Two flags, RXRPC_CALL_RX_LAST and RXRPC_CALL_TX_LAST, are added also
to indicate when rx_top and tx_top point at the packets with the
LAST_PACKET bit set, indicating the end of the phase.
(4) Calls are queued on the socket 'receive queue' rather than packets.
This means that we don't need have to invent dummy packets to queue to
indicate abnormal/terminal states and we don't have to keep metadata
packets (such as ABORTs) around
(5) The offset and length of a (sub)packet's content are now passed to
the verify_packet security op. This is currently expected to decrypt
the packet in place and validate it.
However, there's now nowhere to store the revised offset and length of
the actual data within the decrypted blob (there may be a header and
padding to skip) because an sk_buff may represent multiple packets, so
a locate_data security op is added to retrieve these details from the
sk_buff content when needed.
(6) recvmsg() now has to handle jumbo subpackets, where each subpacket is
individually secured and needs to be individually decrypted. The code
to do this is broken out into rxrpc_recvmsg_data() and shared with the
kernel API. It now iterates over the call's receive buffer rather
than walking the socket receive queue.
Additional changes:
(1) The timers are condensed to a single timer that is set for the soonest
of three timeouts (delayed ACK generation, DATA retransmission and
call lifespan).
(2) Transmission of ACK and ABORT packets is effected immediately from
process-context socket ops/kernel API calls that cause them instead of
them being punted off to a background work item. The data_ready
handler still has to defer to the background, though.
(3) A shutdown op is added to the AF_RXRPC socket so that the AFS
filesystem can shut down the socket and flush its own work items
before closing the socket to deal with any in-progress service calls.
Future additional changes that will need to be considered:
(1) Make sure that a call doesn't hog the front of the queue by receiving
data from the network as fast as userspace is consuming it to the
exclusion of other calls.
(2) Transmit delayed ACKs from within recvmsg() when we've consumed
sufficiently more packets to avoid the background work item needing to
run.
Signed-off-by: David Howells <dhowells@redhat.com>
2016-09-08 10:10:12 +00:00
|
|
|
/*
|
|
|
|
|
* Process individual soft ACKs.
|
|
|
|
|
*
|
|
|
|
|
* Each ACK in the array corresponds to one packet and can be either an ACK or
|
|
|
|
|
* a NAK. If we get find an explicitly NAK'd packet we resend immediately;
|
|
|
|
|
* packets that lie beyond the end of the ACK list are scheduled for resend by
|
|
|
|
|
* the timer on the basis that the peer might just not have processed them at
|
|
|
|
|
* the time the ACK was sent.
|
|
|
|
|
*/
|
2024-02-02 15:19:16 +00:00
|
|
|
static void rxrpc_input_soft_acks(struct rxrpc_call *call,
|
|
|
|
|
struct rxrpc_ack_summary *summary,
|
2024-12-04 07:46:51 +00:00
|
|
|
struct sk_buff *skb)
|
rxrpc: Rewrite the data and ack handling code
Rewrite the data and ack handling code such that:
(1) Parsing of received ACK and ABORT packets and the distribution and the
filing of DATA packets happens entirely within the data_ready context
called from the UDP socket. This allows us to process and discard ACK
and ABORT packets much more quickly (they're no longer stashed on a
queue for a background thread to process).
(2) We avoid calling skb_clone(), pskb_pull() and pskb_trim(). We instead
keep track of the offset and length of the content of each packet in
the sk_buff metadata. This means we don't do any allocation in the
receive path.
(3) Jumbo DATA packet parsing is now done in data_ready context. Rather
than cloning the packet once for each subpacket and pulling/trimming
it, we file the packet multiple times with an annotation for each
indicating which subpacket is there. From that we can directly
calculate the offset and length.
(4) A call's receive queue can be accessed without taking locks (memory
barriers do have to be used, though).
(5) Incoming calls are set up from preallocated resources and immediately
made live. They can than have packets queued upon them and ACKs
generated. If insufficient resources exist, DATA packet #1 is given a
BUSY reply and other DATA packets are discarded).
(6) sk_buffs no longer take a ref on their parent call.
To make this work, the following changes are made:
(1) Each call's receive buffer is now a circular buffer of sk_buff
pointers (rxtx_buffer) rather than a number of sk_buff_heads spread
between the call and the socket. This permits each sk_buff to be in
the buffer multiple times. The receive buffer is reused for the
transmit buffer.
(2) A circular buffer of annotations (rxtx_annotations) is kept parallel
to the data buffer. Transmission phase annotations indicate whether a
buffered packet has been ACK'd or not and whether it needs
retransmission.
Receive phase annotations indicate whether a slot holds a whole packet
or a jumbo subpacket and, if the latter, which subpacket. They also
note whether the packet has been decrypted in place.
(3) DATA packet window tracking is much simplified. Each phase has just
two numbers representing the window (rx_hard_ack/rx_top and
tx_hard_ack/tx_top).
The hard_ack number is the sequence number before base of the window,
representing the last packet the other side says it has consumed.
hard_ack starts from 0 and the first packet is sequence number 1.
The top number is the sequence number of the highest-numbered packet
residing in the buffer. Packets between hard_ack+1 and top are
soft-ACK'd to indicate they've been received, but not yet consumed.
Four macros, before(), before_eq(), after() and after_eq() are added
to compare sequence numbers within the window. This allows for the
top of the window to wrap when the hard-ack sequence number gets close
to the limit.
Two flags, RXRPC_CALL_RX_LAST and RXRPC_CALL_TX_LAST, are added also
to indicate when rx_top and tx_top point at the packets with the
LAST_PACKET bit set, indicating the end of the phase.
(4) Calls are queued on the socket 'receive queue' rather than packets.
This means that we don't need have to invent dummy packets to queue to
indicate abnormal/terminal states and we don't have to keep metadata
packets (such as ABORTs) around
(5) The offset and length of a (sub)packet's content are now passed to
the verify_packet security op. This is currently expected to decrypt
the packet in place and validate it.
However, there's now nowhere to store the revised offset and length of
the actual data within the decrypted blob (there may be a header and
padding to skip) because an sk_buff may represent multiple packets, so
a locate_data security op is added to retrieve these details from the
sk_buff content when needed.
(6) recvmsg() now has to handle jumbo subpackets, where each subpacket is
individually secured and needs to be individually decrypted. The code
to do this is broken out into rxrpc_recvmsg_data() and shared with the
kernel API. It now iterates over the call's receive buffer rather
than walking the socket receive queue.
Additional changes:
(1) The timers are condensed to a single timer that is set for the soonest
of three timeouts (delayed ACK generation, DATA retransmission and
call lifespan).
(2) Transmission of ACK and ABORT packets is effected immediately from
process-context socket ops/kernel API calls that cause them instead of
them being punted off to a background work item. The data_ready
handler still has to defer to the background, though.
(3) A shutdown op is added to the AF_RXRPC socket so that the AFS
filesystem can shut down the socket and flush its own work items
before closing the socket to deal with any in-progress service calls.
Future additional changes that will need to be considered:
(1) Make sure that a call doesn't hog the front of the queue by receiving
data from the network as fast as userspace is consuming it to the
exclusion of other calls.
(2) Transmit delayed ACKs from within recvmsg() when we've consumed
sufficiently more packets to avoid the background work item needing to
run.
Signed-off-by: David Howells <dhowells@redhat.com>
2016-09-08 10:10:12 +00:00
|
|
|
{
|
2024-02-02 15:19:16 +00:00
|
|
|
struct rxrpc_skb_priv *sp = rxrpc_skb(skb);
|
2024-12-04 07:46:51 +00:00
|
|
|
struct rxrpc_txqueue *tq = call->tx_queue;
|
|
|
|
|
unsigned long extracted = ~0UL;
|
|
|
|
|
unsigned int nr = 0;
|
|
|
|
|
rxrpc_seq_t seq = call->acks_hard_ack + 1;
|
|
|
|
|
rxrpc_seq_t lowest_nak = seq + sp->ack.nr_acks;
|
2024-02-02 15:19:16 +00:00
|
|
|
u8 *acks = skb->data + sizeof(struct rxrpc_wire_header) + sizeof(struct rxrpc_ackpacket);
|
2022-03-31 22:55:08 +00:00
|
|
|
|
2024-12-04 07:46:51 +00:00
|
|
|
_enter("%x,%x,%u", tq->qbase, seq, sp->ack.nr_acks);
|
|
|
|
|
|
|
|
|
|
while (after(seq, tq->qbase + RXRPC_NR_TXQUEUE - 1))
|
|
|
|
|
tq = tq->next;
|
2024-02-02 15:19:16 +00:00
|
|
|
|
2024-12-04 07:46:51 +00:00
|
|
|
for (unsigned int i = 0; i < sp->ack.nr_acks; i++) {
|
|
|
|
|
/* Decant ACKs until we hit a txqueue boundary. */
|
|
|
|
|
shiftr_adv_rotr(acks, extracted);
|
|
|
|
|
if (i == 256) {
|
|
|
|
|
acks -= i;
|
|
|
|
|
i = 0;
|
2007-04-26 22:48:28 +00:00
|
|
|
}
|
2024-12-04 07:46:51 +00:00
|
|
|
seq++;
|
|
|
|
|
nr++;
|
|
|
|
|
if ((seq & RXRPC_TXQ_MASK) != 0)
|
|
|
|
|
continue;
|
|
|
|
|
|
|
|
|
|
_debug("bound %16lx %u", extracted, nr);
|
|
|
|
|
|
|
|
|
|
rxrpc_input_soft_ack_tq(call, summary, tq, extracted, RXRPC_NR_TXQUEUE,
|
|
|
|
|
seq - RXRPC_NR_TXQUEUE, &lowest_nak);
|
|
|
|
|
extracted = ~0UL;
|
|
|
|
|
nr = 0;
|
|
|
|
|
tq = tq->next;
|
|
|
|
|
prefetch(tq);
|
2024-02-02 15:19:16 +00:00
|
|
|
}
|
|
|
|
|
|
2024-12-04 07:46:51 +00:00
|
|
|
if (nr) {
|
|
|
|
|
unsigned int nr_reported = seq & RXRPC_TXQ_MASK;
|
|
|
|
|
|
|
|
|
|
extracted >>= RXRPC_NR_TXQUEUE - nr_reported;
|
|
|
|
|
_debug("tail %16lx %u", extracted, nr_reported);
|
|
|
|
|
rxrpc_input_soft_ack_tq(call, summary, tq, extracted, nr_reported,
|
|
|
|
|
seq & ~RXRPC_TXQ_MASK, &lowest_nak);
|
2007-04-26 22:48:28 +00:00
|
|
|
}
|
2024-02-02 15:19:16 +00:00
|
|
|
|
|
|
|
|
/* We *can* have more nacks than we did - the peer is permitted to drop
|
|
|
|
|
* packets it has soft-acked and re-request them. Further, it is
|
|
|
|
|
* possible for the nack distribution to change whilst the number of
|
|
|
|
|
* nacks stays the same or goes down.
|
|
|
|
|
*/
|
2024-12-04 07:46:51 +00:00
|
|
|
if (lowest_nak != call->acks_lowest_nak) {
|
|
|
|
|
call->acks_lowest_nak = lowest_nak;
|
|
|
|
|
summary->new_low_snack = true;
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
_debug("summary A=%d+%d N=%d+%d",
|
|
|
|
|
call->acks_nr_sacks, summary->nr_new_sacks,
|
|
|
|
|
call->acks_nr_snacks, summary->nr_new_snacks);
|
2007-04-26 22:48:28 +00:00
|
|
|
}
|
|
|
|
|
|
rxrpc: Fix ack discard
The Rx protocol has a "previousPacket" field in it that is not handled in
the same way by all protocol implementations. Sometimes it contains the
serial number of the last DATA packet received, sometimes the sequence
number of the last DATA packet received and sometimes the highest sequence
number so far received.
AF_RXRPC is using this to weed out ACKs that are out of date (it's possible
for ACK packets to get reordered on the wire), but this does not work with
OpenAFS which will just stick the sequence number of the last packet seen
into previousPacket.
The issue being seen is that big AFS FS.StoreData RPC (eg. of ~256MiB) are
timing out when partly sent. A trace was captured, with an additional
tracepoint to show ACKs being discarded in rxrpc_input_ack(). Here's an
excerpt showing the problem.
52873.203230: rxrpc_tx_data: c=000004ae DATA ed1a3584:00000002 0002449c q=00024499 fl=09
A DATA packet with sequence number 00024499 has been transmitted (the "q="
field).
...
52873.243296: rxrpc_rx_ack: c=000004ae 00012a2b DLY r=00024499 f=00024497 p=00024496 n=0
52873.243376: rxrpc_rx_ack: c=000004ae 00012a2c IDL r=0002449b f=00024499 p=00024498 n=0
52873.243383: rxrpc_rx_ack: c=000004ae 00012a2d OOS r=0002449d f=00024499 p=0002449a n=2
The Out-Of-Sequence ACK indicates that the server didn't see DATA sequence
number 00024499, but did see seq 0002449a (previousPacket, shown as "p=",
skipped the number, but firstPacket, "f=", which shows the bottom of the
window is set at that point).
52873.252663: rxrpc_retransmit: c=000004ae q=24499 a=02 xp=14581537
52873.252664: rxrpc_tx_data: c=000004ae DATA ed1a3584:00000002 000244bc q=00024499 fl=0b *RETRANS*
The packet has been retransmitted. Retransmission recurs until the peer
says it got the packet.
52873.271013: rxrpc_rx_ack: c=000004ae 00012a31 OOS r=000244a1 f=00024499 p=0002449e n=6
More OOS ACKs indicate that the other packets that are already in the
transmission pipeline are being received. The specific-ACK list is up to 6
ACKs and NAKs.
...
52873.284792: rxrpc_rx_ack: c=000004ae 00012a49 OOS r=000244b9 f=00024499 p=000244b6 n=30
52873.284802: rxrpc_retransmit: c=000004ae q=24499 a=0a xp=63505500
52873.284804: rxrpc_tx_data: c=000004ae DATA ed1a3584:00000002 000244c2 q=00024499 fl=0b *RETRANS*
52873.287468: rxrpc_rx_ack: c=000004ae 00012a4a OOS r=000244ba f=00024499 p=000244b7 n=31
52873.287478: rxrpc_rx_ack: c=000004ae 00012a4b OOS r=000244bb f=00024499 p=000244b8 n=32
At this point, the server's receive window is full (n=32) with presumably 1
NAK'd packet and 31 ACK'd packets. We can't transmit any more packets.
52873.287488: rxrpc_retransmit: c=000004ae q=24499 a=0a xp=61327980
52873.287489: rxrpc_tx_data: c=000004ae DATA ed1a3584:00000002 000244c3 q=00024499 fl=0b *RETRANS*
52873.293850: rxrpc_rx_ack: c=000004ae 00012a4c DLY r=000244bc f=000244a0 p=00024499 n=25
And now we've received an ACK indicating that a DATA retransmission was
received. 7 packets have been processed (the occupied part of the window
moved, as indicated by f= and n=).
52873.293853: rxrpc_rx_discard_ack: c=000004ae r=00012a4c 000244a0<00024499 00024499<000244b8
However, the DLY ACK gets discarded because its previousPacket has gone
backwards (from p=000244b8, in the ACK at 52873.287478 to p=00024499 in the
ACK at 52873.293850).
We then end up in a continuous cycle of retransmit/discard. kafs fails to
update its window because it's discarding the ACKs and can't transmit an
extra packet that would clear the issue because the window is full.
OpenAFS doesn't change the previousPacket value in the ACKs because no new
DATA packets are received with a different previousPacket number.
Fix this by altering the discard check to only discard an ACK based on
previousPacket if there was no advance in the firstPacket. This allows us
to transmit a new packet which will cause previousPacket to advance in the
next ACK.
The check, however, needs to allow for the possibility that previousPacket
may actually have had the serial number placed in it instead - in which
case it will go outside the window and we should ignore it.
Fixes: 1a2391c30c0b ("rxrpc: Fix detection of out of order acks")
Reported-by: Dave Botsch <botsch@cnf.cornell.edu>
Signed-off-by: David Howells <dhowells@redhat.com>
2020-04-29 22:48:43 +00:00
|
|
|
/*
|
|
|
|
|
* Return true if the ACK is valid - ie. it doesn't appear to have regressed
|
|
|
|
|
* with respect to the ack state conveyed by preceding ACKs.
|
|
|
|
|
*/
|
|
|
|
|
static bool rxrpc_is_ack_valid(struct rxrpc_call *call,
|
2024-12-04 07:46:48 +00:00
|
|
|
rxrpc_seq_t hard_ack, rxrpc_seq_t prev_pkt)
|
rxrpc: Fix ack discard
The Rx protocol has a "previousPacket" field in it that is not handled in
the same way by all protocol implementations. Sometimes it contains the
serial number of the last DATA packet received, sometimes the sequence
number of the last DATA packet received and sometimes the highest sequence
number so far received.
AF_RXRPC is using this to weed out ACKs that are out of date (it's possible
for ACK packets to get reordered on the wire), but this does not work with
OpenAFS which will just stick the sequence number of the last packet seen
into previousPacket.
The issue being seen is that big AFS FS.StoreData RPC (eg. of ~256MiB) are
timing out when partly sent. A trace was captured, with an additional
tracepoint to show ACKs being discarded in rxrpc_input_ack(). Here's an
excerpt showing the problem.
52873.203230: rxrpc_tx_data: c=000004ae DATA ed1a3584:00000002 0002449c q=00024499 fl=09
A DATA packet with sequence number 00024499 has been transmitted (the "q="
field).
...
52873.243296: rxrpc_rx_ack: c=000004ae 00012a2b DLY r=00024499 f=00024497 p=00024496 n=0
52873.243376: rxrpc_rx_ack: c=000004ae 00012a2c IDL r=0002449b f=00024499 p=00024498 n=0
52873.243383: rxrpc_rx_ack: c=000004ae 00012a2d OOS r=0002449d f=00024499 p=0002449a n=2
The Out-Of-Sequence ACK indicates that the server didn't see DATA sequence
number 00024499, but did see seq 0002449a (previousPacket, shown as "p=",
skipped the number, but firstPacket, "f=", which shows the bottom of the
window is set at that point).
52873.252663: rxrpc_retransmit: c=000004ae q=24499 a=02 xp=14581537
52873.252664: rxrpc_tx_data: c=000004ae DATA ed1a3584:00000002 000244bc q=00024499 fl=0b *RETRANS*
The packet has been retransmitted. Retransmission recurs until the peer
says it got the packet.
52873.271013: rxrpc_rx_ack: c=000004ae 00012a31 OOS r=000244a1 f=00024499 p=0002449e n=6
More OOS ACKs indicate that the other packets that are already in the
transmission pipeline are being received. The specific-ACK list is up to 6
ACKs and NAKs.
...
52873.284792: rxrpc_rx_ack: c=000004ae 00012a49 OOS r=000244b9 f=00024499 p=000244b6 n=30
52873.284802: rxrpc_retransmit: c=000004ae q=24499 a=0a xp=63505500
52873.284804: rxrpc_tx_data: c=000004ae DATA ed1a3584:00000002 000244c2 q=00024499 fl=0b *RETRANS*
52873.287468: rxrpc_rx_ack: c=000004ae 00012a4a OOS r=000244ba f=00024499 p=000244b7 n=31
52873.287478: rxrpc_rx_ack: c=000004ae 00012a4b OOS r=000244bb f=00024499 p=000244b8 n=32
At this point, the server's receive window is full (n=32) with presumably 1
NAK'd packet and 31 ACK'd packets. We can't transmit any more packets.
52873.287488: rxrpc_retransmit: c=000004ae q=24499 a=0a xp=61327980
52873.287489: rxrpc_tx_data: c=000004ae DATA ed1a3584:00000002 000244c3 q=00024499 fl=0b *RETRANS*
52873.293850: rxrpc_rx_ack: c=000004ae 00012a4c DLY r=000244bc f=000244a0 p=00024499 n=25
And now we've received an ACK indicating that a DATA retransmission was
received. 7 packets have been processed (the occupied part of the window
moved, as indicated by f= and n=).
52873.293853: rxrpc_rx_discard_ack: c=000004ae r=00012a4c 000244a0<00024499 00024499<000244b8
However, the DLY ACK gets discarded because its previousPacket has gone
backwards (from p=000244b8, in the ACK at 52873.287478 to p=00024499 in the
ACK at 52873.293850).
We then end up in a continuous cycle of retransmit/discard. kafs fails to
update its window because it's discarding the ACKs and can't transmit an
extra packet that would clear the issue because the window is full.
OpenAFS doesn't change the previousPacket value in the ACKs because no new
DATA packets are received with a different previousPacket number.
Fix this by altering the discard check to only discard an ACK based on
previousPacket if there was no advance in the firstPacket. This allows us
to transmit a new packet which will cause previousPacket to advance in the
next ACK.
The check, however, needs to allow for the possibility that previousPacket
may actually have had the serial number placed in it instead - in which
case it will go outside the window and we should ignore it.
Fixes: 1a2391c30c0b ("rxrpc: Fix detection of out of order acks")
Reported-by: Dave Botsch <botsch@cnf.cornell.edu>
Signed-off-by: David Howells <dhowells@redhat.com>
2020-04-29 22:48:43 +00:00
|
|
|
{
|
2024-12-04 07:46:48 +00:00
|
|
|
rxrpc_seq_t base = READ_ONCE(call->acks_hard_ack);
|
rxrpc: Fix ack discard
The Rx protocol has a "previousPacket" field in it that is not handled in
the same way by all protocol implementations. Sometimes it contains the
serial number of the last DATA packet received, sometimes the sequence
number of the last DATA packet received and sometimes the highest sequence
number so far received.
AF_RXRPC is using this to weed out ACKs that are out of date (it's possible
for ACK packets to get reordered on the wire), but this does not work with
OpenAFS which will just stick the sequence number of the last packet seen
into previousPacket.
The issue being seen is that big AFS FS.StoreData RPC (eg. of ~256MiB) are
timing out when partly sent. A trace was captured, with an additional
tracepoint to show ACKs being discarded in rxrpc_input_ack(). Here's an
excerpt showing the problem.
52873.203230: rxrpc_tx_data: c=000004ae DATA ed1a3584:00000002 0002449c q=00024499 fl=09
A DATA packet with sequence number 00024499 has been transmitted (the "q="
field).
...
52873.243296: rxrpc_rx_ack: c=000004ae 00012a2b DLY r=00024499 f=00024497 p=00024496 n=0
52873.243376: rxrpc_rx_ack: c=000004ae 00012a2c IDL r=0002449b f=00024499 p=00024498 n=0
52873.243383: rxrpc_rx_ack: c=000004ae 00012a2d OOS r=0002449d f=00024499 p=0002449a n=2
The Out-Of-Sequence ACK indicates that the server didn't see DATA sequence
number 00024499, but did see seq 0002449a (previousPacket, shown as "p=",
skipped the number, but firstPacket, "f=", which shows the bottom of the
window is set at that point).
52873.252663: rxrpc_retransmit: c=000004ae q=24499 a=02 xp=14581537
52873.252664: rxrpc_tx_data: c=000004ae DATA ed1a3584:00000002 000244bc q=00024499 fl=0b *RETRANS*
The packet has been retransmitted. Retransmission recurs until the peer
says it got the packet.
52873.271013: rxrpc_rx_ack: c=000004ae 00012a31 OOS r=000244a1 f=00024499 p=0002449e n=6
More OOS ACKs indicate that the other packets that are already in the
transmission pipeline are being received. The specific-ACK list is up to 6
ACKs and NAKs.
...
52873.284792: rxrpc_rx_ack: c=000004ae 00012a49 OOS r=000244b9 f=00024499 p=000244b6 n=30
52873.284802: rxrpc_retransmit: c=000004ae q=24499 a=0a xp=63505500
52873.284804: rxrpc_tx_data: c=000004ae DATA ed1a3584:00000002 000244c2 q=00024499 fl=0b *RETRANS*
52873.287468: rxrpc_rx_ack: c=000004ae 00012a4a OOS r=000244ba f=00024499 p=000244b7 n=31
52873.287478: rxrpc_rx_ack: c=000004ae 00012a4b OOS r=000244bb f=00024499 p=000244b8 n=32
At this point, the server's receive window is full (n=32) with presumably 1
NAK'd packet and 31 ACK'd packets. We can't transmit any more packets.
52873.287488: rxrpc_retransmit: c=000004ae q=24499 a=0a xp=61327980
52873.287489: rxrpc_tx_data: c=000004ae DATA ed1a3584:00000002 000244c3 q=00024499 fl=0b *RETRANS*
52873.293850: rxrpc_rx_ack: c=000004ae 00012a4c DLY r=000244bc f=000244a0 p=00024499 n=25
And now we've received an ACK indicating that a DATA retransmission was
received. 7 packets have been processed (the occupied part of the window
moved, as indicated by f= and n=).
52873.293853: rxrpc_rx_discard_ack: c=000004ae r=00012a4c 000244a0<00024499 00024499<000244b8
However, the DLY ACK gets discarded because its previousPacket has gone
backwards (from p=000244b8, in the ACK at 52873.287478 to p=00024499 in the
ACK at 52873.293850).
We then end up in a continuous cycle of retransmit/discard. kafs fails to
update its window because it's discarding the ACKs and can't transmit an
extra packet that would clear the issue because the window is full.
OpenAFS doesn't change the previousPacket value in the ACKs because no new
DATA packets are received with a different previousPacket number.
Fix this by altering the discard check to only discard an ACK based on
previousPacket if there was no advance in the firstPacket. This allows us
to transmit a new packet which will cause previousPacket to advance in the
next ACK.
The check, however, needs to allow for the possibility that previousPacket
may actually have had the serial number placed in it instead - in which
case it will go outside the window and we should ignore it.
Fixes: 1a2391c30c0b ("rxrpc: Fix detection of out of order acks")
Reported-by: Dave Botsch <botsch@cnf.cornell.edu>
Signed-off-by: David Howells <dhowells@redhat.com>
2020-04-29 22:48:43 +00:00
|
|
|
|
2024-12-04 07:46:48 +00:00
|
|
|
if (after(hard_ack, base))
|
rxrpc: Fix ack discard
The Rx protocol has a "previousPacket" field in it that is not handled in
the same way by all protocol implementations. Sometimes it contains the
serial number of the last DATA packet received, sometimes the sequence
number of the last DATA packet received and sometimes the highest sequence
number so far received.
AF_RXRPC is using this to weed out ACKs that are out of date (it's possible
for ACK packets to get reordered on the wire), but this does not work with
OpenAFS which will just stick the sequence number of the last packet seen
into previousPacket.
The issue being seen is that big AFS FS.StoreData RPC (eg. of ~256MiB) are
timing out when partly sent. A trace was captured, with an additional
tracepoint to show ACKs being discarded in rxrpc_input_ack(). Here's an
excerpt showing the problem.
52873.203230: rxrpc_tx_data: c=000004ae DATA ed1a3584:00000002 0002449c q=00024499 fl=09
A DATA packet with sequence number 00024499 has been transmitted (the "q="
field).
...
52873.243296: rxrpc_rx_ack: c=000004ae 00012a2b DLY r=00024499 f=00024497 p=00024496 n=0
52873.243376: rxrpc_rx_ack: c=000004ae 00012a2c IDL r=0002449b f=00024499 p=00024498 n=0
52873.243383: rxrpc_rx_ack: c=000004ae 00012a2d OOS r=0002449d f=00024499 p=0002449a n=2
The Out-Of-Sequence ACK indicates that the server didn't see DATA sequence
number 00024499, but did see seq 0002449a (previousPacket, shown as "p=",
skipped the number, but firstPacket, "f=", which shows the bottom of the
window is set at that point).
52873.252663: rxrpc_retransmit: c=000004ae q=24499 a=02 xp=14581537
52873.252664: rxrpc_tx_data: c=000004ae DATA ed1a3584:00000002 000244bc q=00024499 fl=0b *RETRANS*
The packet has been retransmitted. Retransmission recurs until the peer
says it got the packet.
52873.271013: rxrpc_rx_ack: c=000004ae 00012a31 OOS r=000244a1 f=00024499 p=0002449e n=6
More OOS ACKs indicate that the other packets that are already in the
transmission pipeline are being received. The specific-ACK list is up to 6
ACKs and NAKs.
...
52873.284792: rxrpc_rx_ack: c=000004ae 00012a49 OOS r=000244b9 f=00024499 p=000244b6 n=30
52873.284802: rxrpc_retransmit: c=000004ae q=24499 a=0a xp=63505500
52873.284804: rxrpc_tx_data: c=000004ae DATA ed1a3584:00000002 000244c2 q=00024499 fl=0b *RETRANS*
52873.287468: rxrpc_rx_ack: c=000004ae 00012a4a OOS r=000244ba f=00024499 p=000244b7 n=31
52873.287478: rxrpc_rx_ack: c=000004ae 00012a4b OOS r=000244bb f=00024499 p=000244b8 n=32
At this point, the server's receive window is full (n=32) with presumably 1
NAK'd packet and 31 ACK'd packets. We can't transmit any more packets.
52873.287488: rxrpc_retransmit: c=000004ae q=24499 a=0a xp=61327980
52873.287489: rxrpc_tx_data: c=000004ae DATA ed1a3584:00000002 000244c3 q=00024499 fl=0b *RETRANS*
52873.293850: rxrpc_rx_ack: c=000004ae 00012a4c DLY r=000244bc f=000244a0 p=00024499 n=25
And now we've received an ACK indicating that a DATA retransmission was
received. 7 packets have been processed (the occupied part of the window
moved, as indicated by f= and n=).
52873.293853: rxrpc_rx_discard_ack: c=000004ae r=00012a4c 000244a0<00024499 00024499<000244b8
However, the DLY ACK gets discarded because its previousPacket has gone
backwards (from p=000244b8, in the ACK at 52873.287478 to p=00024499 in the
ACK at 52873.293850).
We then end up in a continuous cycle of retransmit/discard. kafs fails to
update its window because it's discarding the ACKs and can't transmit an
extra packet that would clear the issue because the window is full.
OpenAFS doesn't change the previousPacket value in the ACKs because no new
DATA packets are received with a different previousPacket number.
Fix this by altering the discard check to only discard an ACK based on
previousPacket if there was no advance in the firstPacket. This allows us
to transmit a new packet which will cause previousPacket to advance in the
next ACK.
The check, however, needs to allow for the possibility that previousPacket
may actually have had the serial number placed in it instead - in which
case it will go outside the window and we should ignore it.
Fixes: 1a2391c30c0b ("rxrpc: Fix detection of out of order acks")
Reported-by: Dave Botsch <botsch@cnf.cornell.edu>
Signed-off-by: David Howells <dhowells@redhat.com>
2020-04-29 22:48:43 +00:00
|
|
|
return true; /* The window advanced */
|
|
|
|
|
|
2024-12-04 07:46:48 +00:00
|
|
|
if (before(hard_ack, base))
|
rxrpc: Fix ack discard
The Rx protocol has a "previousPacket" field in it that is not handled in
the same way by all protocol implementations. Sometimes it contains the
serial number of the last DATA packet received, sometimes the sequence
number of the last DATA packet received and sometimes the highest sequence
number so far received.
AF_RXRPC is using this to weed out ACKs that are out of date (it's possible
for ACK packets to get reordered on the wire), but this does not work with
OpenAFS which will just stick the sequence number of the last packet seen
into previousPacket.
The issue being seen is that big AFS FS.StoreData RPC (eg. of ~256MiB) are
timing out when partly sent. A trace was captured, with an additional
tracepoint to show ACKs being discarded in rxrpc_input_ack(). Here's an
excerpt showing the problem.
52873.203230: rxrpc_tx_data: c=000004ae DATA ed1a3584:00000002 0002449c q=00024499 fl=09
A DATA packet with sequence number 00024499 has been transmitted (the "q="
field).
...
52873.243296: rxrpc_rx_ack: c=000004ae 00012a2b DLY r=00024499 f=00024497 p=00024496 n=0
52873.243376: rxrpc_rx_ack: c=000004ae 00012a2c IDL r=0002449b f=00024499 p=00024498 n=0
52873.243383: rxrpc_rx_ack: c=000004ae 00012a2d OOS r=0002449d f=00024499 p=0002449a n=2
The Out-Of-Sequence ACK indicates that the server didn't see DATA sequence
number 00024499, but did see seq 0002449a (previousPacket, shown as "p=",
skipped the number, but firstPacket, "f=", which shows the bottom of the
window is set at that point).
52873.252663: rxrpc_retransmit: c=000004ae q=24499 a=02 xp=14581537
52873.252664: rxrpc_tx_data: c=000004ae DATA ed1a3584:00000002 000244bc q=00024499 fl=0b *RETRANS*
The packet has been retransmitted. Retransmission recurs until the peer
says it got the packet.
52873.271013: rxrpc_rx_ack: c=000004ae 00012a31 OOS r=000244a1 f=00024499 p=0002449e n=6
More OOS ACKs indicate that the other packets that are already in the
transmission pipeline are being received. The specific-ACK list is up to 6
ACKs and NAKs.
...
52873.284792: rxrpc_rx_ack: c=000004ae 00012a49 OOS r=000244b9 f=00024499 p=000244b6 n=30
52873.284802: rxrpc_retransmit: c=000004ae q=24499 a=0a xp=63505500
52873.284804: rxrpc_tx_data: c=000004ae DATA ed1a3584:00000002 000244c2 q=00024499 fl=0b *RETRANS*
52873.287468: rxrpc_rx_ack: c=000004ae 00012a4a OOS r=000244ba f=00024499 p=000244b7 n=31
52873.287478: rxrpc_rx_ack: c=000004ae 00012a4b OOS r=000244bb f=00024499 p=000244b8 n=32
At this point, the server's receive window is full (n=32) with presumably 1
NAK'd packet and 31 ACK'd packets. We can't transmit any more packets.
52873.287488: rxrpc_retransmit: c=000004ae q=24499 a=0a xp=61327980
52873.287489: rxrpc_tx_data: c=000004ae DATA ed1a3584:00000002 000244c3 q=00024499 fl=0b *RETRANS*
52873.293850: rxrpc_rx_ack: c=000004ae 00012a4c DLY r=000244bc f=000244a0 p=00024499 n=25
And now we've received an ACK indicating that a DATA retransmission was
received. 7 packets have been processed (the occupied part of the window
moved, as indicated by f= and n=).
52873.293853: rxrpc_rx_discard_ack: c=000004ae r=00012a4c 000244a0<00024499 00024499<000244b8
However, the DLY ACK gets discarded because its previousPacket has gone
backwards (from p=000244b8, in the ACK at 52873.287478 to p=00024499 in the
ACK at 52873.293850).
We then end up in a continuous cycle of retransmit/discard. kafs fails to
update its window because it's discarding the ACKs and can't transmit an
extra packet that would clear the issue because the window is full.
OpenAFS doesn't change the previousPacket value in the ACKs because no new
DATA packets are received with a different previousPacket number.
Fix this by altering the discard check to only discard an ACK based on
previousPacket if there was no advance in the firstPacket. This allows us
to transmit a new packet which will cause previousPacket to advance in the
next ACK.
The check, however, needs to allow for the possibility that previousPacket
may actually have had the serial number placed in it instead - in which
case it will go outside the window and we should ignore it.
Fixes: 1a2391c30c0b ("rxrpc: Fix detection of out of order acks")
Reported-by: Dave Botsch <botsch@cnf.cornell.edu>
Signed-off-by: David Howells <dhowells@redhat.com>
2020-04-29 22:48:43 +00:00
|
|
|
return false; /* firstPacket regressed */
|
|
|
|
|
|
2022-05-21 08:03:18 +00:00
|
|
|
if (after_eq(prev_pkt, call->acks_prev_seq))
|
rxrpc: Fix ack discard
The Rx protocol has a "previousPacket" field in it that is not handled in
the same way by all protocol implementations. Sometimes it contains the
serial number of the last DATA packet received, sometimes the sequence
number of the last DATA packet received and sometimes the highest sequence
number so far received.
AF_RXRPC is using this to weed out ACKs that are out of date (it's possible
for ACK packets to get reordered on the wire), but this does not work with
OpenAFS which will just stick the sequence number of the last packet seen
into previousPacket.
The issue being seen is that big AFS FS.StoreData RPC (eg. of ~256MiB) are
timing out when partly sent. A trace was captured, with an additional
tracepoint to show ACKs being discarded in rxrpc_input_ack(). Here's an
excerpt showing the problem.
52873.203230: rxrpc_tx_data: c=000004ae DATA ed1a3584:00000002 0002449c q=00024499 fl=09
A DATA packet with sequence number 00024499 has been transmitted (the "q="
field).
...
52873.243296: rxrpc_rx_ack: c=000004ae 00012a2b DLY r=00024499 f=00024497 p=00024496 n=0
52873.243376: rxrpc_rx_ack: c=000004ae 00012a2c IDL r=0002449b f=00024499 p=00024498 n=0
52873.243383: rxrpc_rx_ack: c=000004ae 00012a2d OOS r=0002449d f=00024499 p=0002449a n=2
The Out-Of-Sequence ACK indicates that the server didn't see DATA sequence
number 00024499, but did see seq 0002449a (previousPacket, shown as "p=",
skipped the number, but firstPacket, "f=", which shows the bottom of the
window is set at that point).
52873.252663: rxrpc_retransmit: c=000004ae q=24499 a=02 xp=14581537
52873.252664: rxrpc_tx_data: c=000004ae DATA ed1a3584:00000002 000244bc q=00024499 fl=0b *RETRANS*
The packet has been retransmitted. Retransmission recurs until the peer
says it got the packet.
52873.271013: rxrpc_rx_ack: c=000004ae 00012a31 OOS r=000244a1 f=00024499 p=0002449e n=6
More OOS ACKs indicate that the other packets that are already in the
transmission pipeline are being received. The specific-ACK list is up to 6
ACKs and NAKs.
...
52873.284792: rxrpc_rx_ack: c=000004ae 00012a49 OOS r=000244b9 f=00024499 p=000244b6 n=30
52873.284802: rxrpc_retransmit: c=000004ae q=24499 a=0a xp=63505500
52873.284804: rxrpc_tx_data: c=000004ae DATA ed1a3584:00000002 000244c2 q=00024499 fl=0b *RETRANS*
52873.287468: rxrpc_rx_ack: c=000004ae 00012a4a OOS r=000244ba f=00024499 p=000244b7 n=31
52873.287478: rxrpc_rx_ack: c=000004ae 00012a4b OOS r=000244bb f=00024499 p=000244b8 n=32
At this point, the server's receive window is full (n=32) with presumably 1
NAK'd packet and 31 ACK'd packets. We can't transmit any more packets.
52873.287488: rxrpc_retransmit: c=000004ae q=24499 a=0a xp=61327980
52873.287489: rxrpc_tx_data: c=000004ae DATA ed1a3584:00000002 000244c3 q=00024499 fl=0b *RETRANS*
52873.293850: rxrpc_rx_ack: c=000004ae 00012a4c DLY r=000244bc f=000244a0 p=00024499 n=25
And now we've received an ACK indicating that a DATA retransmission was
received. 7 packets have been processed (the occupied part of the window
moved, as indicated by f= and n=).
52873.293853: rxrpc_rx_discard_ack: c=000004ae r=00012a4c 000244a0<00024499 00024499<000244b8
However, the DLY ACK gets discarded because its previousPacket has gone
backwards (from p=000244b8, in the ACK at 52873.287478 to p=00024499 in the
ACK at 52873.293850).
We then end up in a continuous cycle of retransmit/discard. kafs fails to
update its window because it's discarding the ACKs and can't transmit an
extra packet that would clear the issue because the window is full.
OpenAFS doesn't change the previousPacket value in the ACKs because no new
DATA packets are received with a different previousPacket number.
Fix this by altering the discard check to only discard an ACK based on
previousPacket if there was no advance in the firstPacket. This allows us
to transmit a new packet which will cause previousPacket to advance in the
next ACK.
The check, however, needs to allow for the possibility that previousPacket
may actually have had the serial number placed in it instead - in which
case it will go outside the window and we should ignore it.
Fixes: 1a2391c30c0b ("rxrpc: Fix detection of out of order acks")
Reported-by: Dave Botsch <botsch@cnf.cornell.edu>
Signed-off-by: David Howells <dhowells@redhat.com>
2020-04-29 22:48:43 +00:00
|
|
|
return true; /* previousPacket hasn't regressed. */
|
|
|
|
|
|
|
|
|
|
/* Some rx implementations put a serial number in previousPacket. */
|
2024-12-04 07:46:48 +00:00
|
|
|
if (after(prev_pkt, base + call->tx_winsize))
|
rxrpc: Fix ack discard
The Rx protocol has a "previousPacket" field in it that is not handled in
the same way by all protocol implementations. Sometimes it contains the
serial number of the last DATA packet received, sometimes the sequence
number of the last DATA packet received and sometimes the highest sequence
number so far received.
AF_RXRPC is using this to weed out ACKs that are out of date (it's possible
for ACK packets to get reordered on the wire), but this does not work with
OpenAFS which will just stick the sequence number of the last packet seen
into previousPacket.
The issue being seen is that big AFS FS.StoreData RPC (eg. of ~256MiB) are
timing out when partly sent. A trace was captured, with an additional
tracepoint to show ACKs being discarded in rxrpc_input_ack(). Here's an
excerpt showing the problem.
52873.203230: rxrpc_tx_data: c=000004ae DATA ed1a3584:00000002 0002449c q=00024499 fl=09
A DATA packet with sequence number 00024499 has been transmitted (the "q="
field).
...
52873.243296: rxrpc_rx_ack: c=000004ae 00012a2b DLY r=00024499 f=00024497 p=00024496 n=0
52873.243376: rxrpc_rx_ack: c=000004ae 00012a2c IDL r=0002449b f=00024499 p=00024498 n=0
52873.243383: rxrpc_rx_ack: c=000004ae 00012a2d OOS r=0002449d f=00024499 p=0002449a n=2
The Out-Of-Sequence ACK indicates that the server didn't see DATA sequence
number 00024499, but did see seq 0002449a (previousPacket, shown as "p=",
skipped the number, but firstPacket, "f=", which shows the bottom of the
window is set at that point).
52873.252663: rxrpc_retransmit: c=000004ae q=24499 a=02 xp=14581537
52873.252664: rxrpc_tx_data: c=000004ae DATA ed1a3584:00000002 000244bc q=00024499 fl=0b *RETRANS*
The packet has been retransmitted. Retransmission recurs until the peer
says it got the packet.
52873.271013: rxrpc_rx_ack: c=000004ae 00012a31 OOS r=000244a1 f=00024499 p=0002449e n=6
More OOS ACKs indicate that the other packets that are already in the
transmission pipeline are being received. The specific-ACK list is up to 6
ACKs and NAKs.
...
52873.284792: rxrpc_rx_ack: c=000004ae 00012a49 OOS r=000244b9 f=00024499 p=000244b6 n=30
52873.284802: rxrpc_retransmit: c=000004ae q=24499 a=0a xp=63505500
52873.284804: rxrpc_tx_data: c=000004ae DATA ed1a3584:00000002 000244c2 q=00024499 fl=0b *RETRANS*
52873.287468: rxrpc_rx_ack: c=000004ae 00012a4a OOS r=000244ba f=00024499 p=000244b7 n=31
52873.287478: rxrpc_rx_ack: c=000004ae 00012a4b OOS r=000244bb f=00024499 p=000244b8 n=32
At this point, the server's receive window is full (n=32) with presumably 1
NAK'd packet and 31 ACK'd packets. We can't transmit any more packets.
52873.287488: rxrpc_retransmit: c=000004ae q=24499 a=0a xp=61327980
52873.287489: rxrpc_tx_data: c=000004ae DATA ed1a3584:00000002 000244c3 q=00024499 fl=0b *RETRANS*
52873.293850: rxrpc_rx_ack: c=000004ae 00012a4c DLY r=000244bc f=000244a0 p=00024499 n=25
And now we've received an ACK indicating that a DATA retransmission was
received. 7 packets have been processed (the occupied part of the window
moved, as indicated by f= and n=).
52873.293853: rxrpc_rx_discard_ack: c=000004ae r=00012a4c 000244a0<00024499 00024499<000244b8
However, the DLY ACK gets discarded because its previousPacket has gone
backwards (from p=000244b8, in the ACK at 52873.287478 to p=00024499 in the
ACK at 52873.293850).
We then end up in a continuous cycle of retransmit/discard. kafs fails to
update its window because it's discarding the ACKs and can't transmit an
extra packet that would clear the issue because the window is full.
OpenAFS doesn't change the previousPacket value in the ACKs because no new
DATA packets are received with a different previousPacket number.
Fix this by altering the discard check to only discard an ACK based on
previousPacket if there was no advance in the firstPacket. This allows us
to transmit a new packet which will cause previousPacket to advance in the
next ACK.
The check, however, needs to allow for the possibility that previousPacket
may actually have had the serial number placed in it instead - in which
case it will go outside the window and we should ignore it.
Fixes: 1a2391c30c0b ("rxrpc: Fix detection of out of order acks")
Reported-by: Dave Botsch <botsch@cnf.cornell.edu>
Signed-off-by: David Howells <dhowells@redhat.com>
2020-04-29 22:48:43 +00:00
|
|
|
return false;
|
|
|
|
|
return true;
|
|
|
|
|
}
|
|
|
|
|
|
2007-04-26 22:48:28 +00:00
|
|
|
/*
|
rxrpc: Rewrite the data and ack handling code
Rewrite the data and ack handling code such that:
(1) Parsing of received ACK and ABORT packets and the distribution and the
filing of DATA packets happens entirely within the data_ready context
called from the UDP socket. This allows us to process and discard ACK
and ABORT packets much more quickly (they're no longer stashed on a
queue for a background thread to process).
(2) We avoid calling skb_clone(), pskb_pull() and pskb_trim(). We instead
keep track of the offset and length of the content of each packet in
the sk_buff metadata. This means we don't do any allocation in the
receive path.
(3) Jumbo DATA packet parsing is now done in data_ready context. Rather
than cloning the packet once for each subpacket and pulling/trimming
it, we file the packet multiple times with an annotation for each
indicating which subpacket is there. From that we can directly
calculate the offset and length.
(4) A call's receive queue can be accessed without taking locks (memory
barriers do have to be used, though).
(5) Incoming calls are set up from preallocated resources and immediately
made live. They can than have packets queued upon them and ACKs
generated. If insufficient resources exist, DATA packet #1 is given a
BUSY reply and other DATA packets are discarded).
(6) sk_buffs no longer take a ref on their parent call.
To make this work, the following changes are made:
(1) Each call's receive buffer is now a circular buffer of sk_buff
pointers (rxtx_buffer) rather than a number of sk_buff_heads spread
between the call and the socket. This permits each sk_buff to be in
the buffer multiple times. The receive buffer is reused for the
transmit buffer.
(2) A circular buffer of annotations (rxtx_annotations) is kept parallel
to the data buffer. Transmission phase annotations indicate whether a
buffered packet has been ACK'd or not and whether it needs
retransmission.
Receive phase annotations indicate whether a slot holds a whole packet
or a jumbo subpacket and, if the latter, which subpacket. They also
note whether the packet has been decrypted in place.
(3) DATA packet window tracking is much simplified. Each phase has just
two numbers representing the window (rx_hard_ack/rx_top and
tx_hard_ack/tx_top).
The hard_ack number is the sequence number before base of the window,
representing the last packet the other side says it has consumed.
hard_ack starts from 0 and the first packet is sequence number 1.
The top number is the sequence number of the highest-numbered packet
residing in the buffer. Packets between hard_ack+1 and top are
soft-ACK'd to indicate they've been received, but not yet consumed.
Four macros, before(), before_eq(), after() and after_eq() are added
to compare sequence numbers within the window. This allows for the
top of the window to wrap when the hard-ack sequence number gets close
to the limit.
Two flags, RXRPC_CALL_RX_LAST and RXRPC_CALL_TX_LAST, are added also
to indicate when rx_top and tx_top point at the packets with the
LAST_PACKET bit set, indicating the end of the phase.
(4) Calls are queued on the socket 'receive queue' rather than packets.
This means that we don't need have to invent dummy packets to queue to
indicate abnormal/terminal states and we don't have to keep metadata
packets (such as ABORTs) around
(5) The offset and length of a (sub)packet's content are now passed to
the verify_packet security op. This is currently expected to decrypt
the packet in place and validate it.
However, there's now nowhere to store the revised offset and length of
the actual data within the decrypted blob (there may be a header and
padding to skip) because an sk_buff may represent multiple packets, so
a locate_data security op is added to retrieve these details from the
sk_buff content when needed.
(6) recvmsg() now has to handle jumbo subpackets, where each subpacket is
individually secured and needs to be individually decrypted. The code
to do this is broken out into rxrpc_recvmsg_data() and shared with the
kernel API. It now iterates over the call's receive buffer rather
than walking the socket receive queue.
Additional changes:
(1) The timers are condensed to a single timer that is set for the soonest
of three timeouts (delayed ACK generation, DATA retransmission and
call lifespan).
(2) Transmission of ACK and ABORT packets is effected immediately from
process-context socket ops/kernel API calls that cause them instead of
them being punted off to a background work item. The data_ready
handler still has to defer to the background, though.
(3) A shutdown op is added to the AF_RXRPC socket so that the AFS
filesystem can shut down the socket and flush its own work items
before closing the socket to deal with any in-progress service calls.
Future additional changes that will need to be considered:
(1) Make sure that a call doesn't hog the front of the queue by receiving
data from the network as fast as userspace is consuming it to the
exclusion of other calls.
(2) Transmit delayed ACKs from within recvmsg() when we've consumed
sufficiently more packets to avoid the background work item needing to
run.
Signed-off-by: David Howells <dhowells@redhat.com>
2016-09-08 10:10:12 +00:00
|
|
|
* Process an ACK packet.
|
|
|
|
|
*
|
|
|
|
|
* ack.firstPacket is the sequence number of the first soft-ACK'd/NAK'd packet
|
|
|
|
|
* in the ACK array. Anything before that is hard-ACK'd and may be discarded.
|
|
|
|
|
*
|
|
|
|
|
* A hard-ACK means that a packet has been processed and may be discarded; a
|
|
|
|
|
* soft-ACK means that the packet may be discarded and retransmission
|
|
|
|
|
* requested. A phase is complete when all packets are hard-ACK'd.
|
2007-04-26 22:48:28 +00:00
|
|
|
*/
|
2019-08-09 14:20:41 +00:00
|
|
|
static void rxrpc_input_ack(struct rxrpc_call *call, struct sk_buff *skb)
|
2007-04-26 22:48:28 +00:00
|
|
|
{
|
2016-09-24 17:05:26 +00:00
|
|
|
struct rxrpc_ack_summary summary = { 0 };
|
2024-01-26 16:17:03 +00:00
|
|
|
struct rxrpc_acktrailer trailer;
|
2024-12-04 07:46:48 +00:00
|
|
|
struct rxrpc_skb_priv *sp = rxrpc_skb(skb);
|
2024-12-04 07:46:51 +00:00
|
|
|
rxrpc_seq_t first_soft_ack, hard_ack, prev_pkt;
|
2016-09-30 12:26:03 +00:00
|
|
|
int nr_acks, offset, ioffset;
|
rxrpc: Rewrite the data and ack handling code
Rewrite the data and ack handling code such that:
(1) Parsing of received ACK and ABORT packets and the distribution and the
filing of DATA packets happens entirely within the data_ready context
called from the UDP socket. This allows us to process and discard ACK
and ABORT packets much more quickly (they're no longer stashed on a
queue for a background thread to process).
(2) We avoid calling skb_clone(), pskb_pull() and pskb_trim(). We instead
keep track of the offset and length of the content of each packet in
the sk_buff metadata. This means we don't do any allocation in the
receive path.
(3) Jumbo DATA packet parsing is now done in data_ready context. Rather
than cloning the packet once for each subpacket and pulling/trimming
it, we file the packet multiple times with an annotation for each
indicating which subpacket is there. From that we can directly
calculate the offset and length.
(4) A call's receive queue can be accessed without taking locks (memory
barriers do have to be used, though).
(5) Incoming calls are set up from preallocated resources and immediately
made live. They can than have packets queued upon them and ACKs
generated. If insufficient resources exist, DATA packet #1 is given a
BUSY reply and other DATA packets are discarded).
(6) sk_buffs no longer take a ref on their parent call.
To make this work, the following changes are made:
(1) Each call's receive buffer is now a circular buffer of sk_buff
pointers (rxtx_buffer) rather than a number of sk_buff_heads spread
between the call and the socket. This permits each sk_buff to be in
the buffer multiple times. The receive buffer is reused for the
transmit buffer.
(2) A circular buffer of annotations (rxtx_annotations) is kept parallel
to the data buffer. Transmission phase annotations indicate whether a
buffered packet has been ACK'd or not and whether it needs
retransmission.
Receive phase annotations indicate whether a slot holds a whole packet
or a jumbo subpacket and, if the latter, which subpacket. They also
note whether the packet has been decrypted in place.
(3) DATA packet window tracking is much simplified. Each phase has just
two numbers representing the window (rx_hard_ack/rx_top and
tx_hard_ack/tx_top).
The hard_ack number is the sequence number before base of the window,
representing the last packet the other side says it has consumed.
hard_ack starts from 0 and the first packet is sequence number 1.
The top number is the sequence number of the highest-numbered packet
residing in the buffer. Packets between hard_ack+1 and top are
soft-ACK'd to indicate they've been received, but not yet consumed.
Four macros, before(), before_eq(), after() and after_eq() are added
to compare sequence numbers within the window. This allows for the
top of the window to wrap when the hard-ack sequence number gets close
to the limit.
Two flags, RXRPC_CALL_RX_LAST and RXRPC_CALL_TX_LAST, are added also
to indicate when rx_top and tx_top point at the packets with the
LAST_PACKET bit set, indicating the end of the phase.
(4) Calls are queued on the socket 'receive queue' rather than packets.
This means that we don't need have to invent dummy packets to queue to
indicate abnormal/terminal states and we don't have to keep metadata
packets (such as ABORTs) around
(5) The offset and length of a (sub)packet's content are now passed to
the verify_packet security op. This is currently expected to decrypt
the packet in place and validate it.
However, there's now nowhere to store the revised offset and length of
the actual data within the decrypted blob (there may be a header and
padding to skip) because an sk_buff may represent multiple packets, so
a locate_data security op is added to retrieve these details from the
sk_buff content when needed.
(6) recvmsg() now has to handle jumbo subpackets, where each subpacket is
individually secured and needs to be individually decrypted. The code
to do this is broken out into rxrpc_recvmsg_data() and shared with the
kernel API. It now iterates over the call's receive buffer rather
than walking the socket receive queue.
Additional changes:
(1) The timers are condensed to a single timer that is set for the soonest
of three timeouts (delayed ACK generation, DATA retransmission and
call lifespan).
(2) Transmission of ACK and ABORT packets is effected immediately from
process-context socket ops/kernel API calls that cause them instead of
them being punted off to a background work item. The data_ready
handler still has to defer to the background, though.
(3) A shutdown op is added to the AF_RXRPC socket so that the AFS
filesystem can shut down the socket and flush its own work items
before closing the socket to deal with any in-progress service calls.
Future additional changes that will need to be considered:
(1) Make sure that a call doesn't hog the front of the queue by receiving
data from the network as fast as userspace is consuming it to the
exclusion of other calls.
(2) Transmit delayed ACKs from within recvmsg() when we've consumed
sufficiently more packets to avoid the background work item needing to
run.
Signed-off-by: David Howells <dhowells@redhat.com>
2016-09-08 10:10:12 +00:00
|
|
|
|
|
|
|
|
_enter("");
|
|
|
|
|
|
2024-02-01 10:29:36 +00:00
|
|
|
offset = sizeof(struct rxrpc_wire_header) + sizeof(struct rxrpc_ackpacket);
|
|
|
|
|
|
2024-12-04 07:47:02 +00:00
|
|
|
summary.ack_serial = sp->hdr.serial;
|
|
|
|
|
first_soft_ack = sp->ack.first_ack;
|
|
|
|
|
prev_pkt = sp->ack.prev_ack;
|
|
|
|
|
nr_acks = sp->ack.nr_acks;
|
|
|
|
|
hard_ack = first_soft_ack - 1;
|
|
|
|
|
summary.acked_serial = sp->ack.acked_serial;
|
|
|
|
|
summary.ack_reason = (sp->ack.reason < RXRPC_ACK__INVALID ?
|
|
|
|
|
sp->ack.reason : RXRPC_ACK__INVALID);
|
rxrpc: Rewrite the data and ack handling code
Rewrite the data and ack handling code such that:
(1) Parsing of received ACK and ABORT packets and the distribution and the
filing of DATA packets happens entirely within the data_ready context
called from the UDP socket. This allows us to process and discard ACK
and ABORT packets much more quickly (they're no longer stashed on a
queue for a background thread to process).
(2) We avoid calling skb_clone(), pskb_pull() and pskb_trim(). We instead
keep track of the offset and length of the content of each packet in
the sk_buff metadata. This means we don't do any allocation in the
receive path.
(3) Jumbo DATA packet parsing is now done in data_ready context. Rather
than cloning the packet once for each subpacket and pulling/trimming
it, we file the packet multiple times with an annotation for each
indicating which subpacket is there. From that we can directly
calculate the offset and length.
(4) A call's receive queue can be accessed without taking locks (memory
barriers do have to be used, though).
(5) Incoming calls are set up from preallocated resources and immediately
made live. They can than have packets queued upon them and ACKs
generated. If insufficient resources exist, DATA packet #1 is given a
BUSY reply and other DATA packets are discarded).
(6) sk_buffs no longer take a ref on their parent call.
To make this work, the following changes are made:
(1) Each call's receive buffer is now a circular buffer of sk_buff
pointers (rxtx_buffer) rather than a number of sk_buff_heads spread
between the call and the socket. This permits each sk_buff to be in
the buffer multiple times. The receive buffer is reused for the
transmit buffer.
(2) A circular buffer of annotations (rxtx_annotations) is kept parallel
to the data buffer. Transmission phase annotations indicate whether a
buffered packet has been ACK'd or not and whether it needs
retransmission.
Receive phase annotations indicate whether a slot holds a whole packet
or a jumbo subpacket and, if the latter, which subpacket. They also
note whether the packet has been decrypted in place.
(3) DATA packet window tracking is much simplified. Each phase has just
two numbers representing the window (rx_hard_ack/rx_top and
tx_hard_ack/tx_top).
The hard_ack number is the sequence number before base of the window,
representing the last packet the other side says it has consumed.
hard_ack starts from 0 and the first packet is sequence number 1.
The top number is the sequence number of the highest-numbered packet
residing in the buffer. Packets between hard_ack+1 and top are
soft-ACK'd to indicate they've been received, but not yet consumed.
Four macros, before(), before_eq(), after() and after_eq() are added
to compare sequence numbers within the window. This allows for the
top of the window to wrap when the hard-ack sequence number gets close
to the limit.
Two flags, RXRPC_CALL_RX_LAST and RXRPC_CALL_TX_LAST, are added also
to indicate when rx_top and tx_top point at the packets with the
LAST_PACKET bit set, indicating the end of the phase.
(4) Calls are queued on the socket 'receive queue' rather than packets.
This means that we don't need have to invent dummy packets to queue to
indicate abnormal/terminal states and we don't have to keep metadata
packets (such as ABORTs) around
(5) The offset and length of a (sub)packet's content are now passed to
the verify_packet security op. This is currently expected to decrypt
the packet in place and validate it.
However, there's now nowhere to store the revised offset and length of
the actual data within the decrypted blob (there may be a header and
padding to skip) because an sk_buff may represent multiple packets, so
a locate_data security op is added to retrieve these details from the
sk_buff content when needed.
(6) recvmsg() now has to handle jumbo subpackets, where each subpacket is
individually secured and needs to be individually decrypted. The code
to do this is broken out into rxrpc_recvmsg_data() and shared with the
kernel API. It now iterates over the call's receive buffer rather
than walking the socket receive queue.
Additional changes:
(1) The timers are condensed to a single timer that is set for the soonest
of three timeouts (delayed ACK generation, DATA retransmission and
call lifespan).
(2) Transmission of ACK and ABORT packets is effected immediately from
process-context socket ops/kernel API calls that cause them instead of
them being punted off to a background work item. The data_ready
handler still has to defer to the background, though.
(3) A shutdown op is added to the AF_RXRPC socket so that the AFS
filesystem can shut down the socket and flush its own work items
before closing the socket to deal with any in-progress service calls.
Future additional changes that will need to be considered:
(1) Make sure that a call doesn't hog the front of the queue by receiving
data from the network as fast as userspace is consuming it to the
exclusion of other calls.
(2) Transmit delayed ACKs from within recvmsg() when we've consumed
sufficiently more packets to avoid the background work item needing to
run.
Signed-off-by: David Howells <dhowells@redhat.com>
2016-09-08 10:10:12 +00:00
|
|
|
|
2024-12-04 07:46:48 +00:00
|
|
|
trace_rxrpc_rx_ack(call, sp);
|
2024-02-01 10:29:36 +00:00
|
|
|
rxrpc_inc_stat(call->rxnet, stat_rx_acks[summary.ack_reason]);
|
2024-12-04 07:46:51 +00:00
|
|
|
prefetch(call->tx_queue);
|
2016-09-17 09:49:13 +00:00
|
|
|
|
afs: Adjust ACK interpretation to try and cope with NAT
If a client's address changes, say if it is NAT'd, this can disrupt an in
progress operation. For most operations, this is not much of a problem,
but StoreData can be different as some servers modify the target file as
the data comes in, so if a store request is disrupted, the file can get
corrupted on the server.
The problem is that the server doesn't recognise packets that come after
the change of address as belonging to the original client and will bounce
them, either by sending an OUT_OF_SEQUENCE ACK to the apparent new call if
the packet number falls within the initial sequence number window of a call
or by sending an EXCEEDS_WINDOW ACK if it falls outside and then aborting
it. In both cases, firstPacket will be 1 and previousPacket will be 0 in
the ACK information.
Fix this by the following means:
(1) If a client call receives an EXCEEDS_WINDOW ACK with firstPacket as 1
and previousPacket as 0, assume this indicates that the server saw the
incoming packets from a different peer and thus as a different call.
Fail the call with error -ENETRESET.
(2) Also fail the call if a similar OUT_OF_SEQUENCE ACK occurs if the
first packet has been hard-ACK'd. If it hasn't been hard-ACK'd, the
ACK packet will cause it to get retransmitted, so the call will just
be repeated.
(3) Make afs_select_fileserver() treat -ENETRESET as a straight fail of
the operation.
(4) Prioritise the error code over things like -ECONNRESET as the server
did actually respond.
(5) Make writeback treat -ENETRESET as a retryable error and make it
redirty all the pages involved in a write so that the VM will retry.
Note that there is still a circumstance that I can't easily deal with: if
the operation is fully received and processed by the server, but the reply
is lost due to address change. There's no way to know if the op happened.
We can examine the server, but a conflicting change could have been made by
a third party - and we can't tell the difference. In such a case, a
message like:
kAFS: vnode modified {100058:146266} b7->b8 YFS.StoreData64 (op=2646a)
will be logged to dmesg on the next op to touch the file and the client
will reset the inode state, including invalidating clean parts of the
pagecache.
Reported-by: Marc Dionne <marc.dionne@auristor.com>
Signed-off-by: David Howells <dhowells@redhat.com>
cc: linux-afs@lists.infradead.org
Link: http://lists.infradead.org/pipermail/linux-afs/2021-December/004811.html # v1
Signed-off-by: David S. Miller <davem@davemloft.net>
2022-05-21 07:45:55 +00:00
|
|
|
/* If we get an EXCEEDS_WINDOW ACK from the server, it probably
|
|
|
|
|
* indicates that the client address changed due to NAT. The server
|
|
|
|
|
* lost the call because it switched to a different peer.
|
|
|
|
|
*/
|
2024-02-01 10:29:36 +00:00
|
|
|
if (unlikely(summary.ack_reason == RXRPC_ACK_EXCEEDS_WINDOW) &&
|
2024-12-04 07:46:48 +00:00
|
|
|
hard_ack == 0 &&
|
afs: Adjust ACK interpretation to try and cope with NAT
If a client's address changes, say if it is NAT'd, this can disrupt an in
progress operation. For most operations, this is not much of a problem,
but StoreData can be different as some servers modify the target file as
the data comes in, so if a store request is disrupted, the file can get
corrupted on the server.
The problem is that the server doesn't recognise packets that come after
the change of address as belonging to the original client and will bounce
them, either by sending an OUT_OF_SEQUENCE ACK to the apparent new call if
the packet number falls within the initial sequence number window of a call
or by sending an EXCEEDS_WINDOW ACK if it falls outside and then aborting
it. In both cases, firstPacket will be 1 and previousPacket will be 0 in
the ACK information.
Fix this by the following means:
(1) If a client call receives an EXCEEDS_WINDOW ACK with firstPacket as 1
and previousPacket as 0, assume this indicates that the server saw the
incoming packets from a different peer and thus as a different call.
Fail the call with error -ENETRESET.
(2) Also fail the call if a similar OUT_OF_SEQUENCE ACK occurs if the
first packet has been hard-ACK'd. If it hasn't been hard-ACK'd, the
ACK packet will cause it to get retransmitted, so the call will just
be repeated.
(3) Make afs_select_fileserver() treat -ENETRESET as a straight fail of
the operation.
(4) Prioritise the error code over things like -ECONNRESET as the server
did actually respond.
(5) Make writeback treat -ENETRESET as a retryable error and make it
redirty all the pages involved in a write so that the VM will retry.
Note that there is still a circumstance that I can't easily deal with: if
the operation is fully received and processed by the server, but the reply
is lost due to address change. There's no way to know if the op happened.
We can examine the server, but a conflicting change could have been made by
a third party - and we can't tell the difference. In such a case, a
message like:
kAFS: vnode modified {100058:146266} b7->b8 YFS.StoreData64 (op=2646a)
will be logged to dmesg on the next op to touch the file and the client
will reset the inode state, including invalidating clean parts of the
pagecache.
Reported-by: Marc Dionne <marc.dionne@auristor.com>
Signed-off-by: David Howells <dhowells@redhat.com>
cc: linux-afs@lists.infradead.org
Link: http://lists.infradead.org/pipermail/linux-afs/2021-December/004811.html # v1
Signed-off-by: David S. Miller <davem@davemloft.net>
2022-05-21 07:45:55 +00:00
|
|
|
prev_pkt == 0 &&
|
|
|
|
|
rxrpc_is_client_call(call)) {
|
|
|
|
|
rxrpc_set_call_completion(call, RXRPC_CALL_REMOTELY_ABORTED,
|
|
|
|
|
0, -ENETRESET);
|
2023-11-16 13:12:59 +00:00
|
|
|
goto send_response;
|
afs: Adjust ACK interpretation to try and cope with NAT
If a client's address changes, say if it is NAT'd, this can disrupt an in
progress operation. For most operations, this is not much of a problem,
but StoreData can be different as some servers modify the target file as
the data comes in, so if a store request is disrupted, the file can get
corrupted on the server.
The problem is that the server doesn't recognise packets that come after
the change of address as belonging to the original client and will bounce
them, either by sending an OUT_OF_SEQUENCE ACK to the apparent new call if
the packet number falls within the initial sequence number window of a call
or by sending an EXCEEDS_WINDOW ACK if it falls outside and then aborting
it. In both cases, firstPacket will be 1 and previousPacket will be 0 in
the ACK information.
Fix this by the following means:
(1) If a client call receives an EXCEEDS_WINDOW ACK with firstPacket as 1
and previousPacket as 0, assume this indicates that the server saw the
incoming packets from a different peer and thus as a different call.
Fail the call with error -ENETRESET.
(2) Also fail the call if a similar OUT_OF_SEQUENCE ACK occurs if the
first packet has been hard-ACK'd. If it hasn't been hard-ACK'd, the
ACK packet will cause it to get retransmitted, so the call will just
be repeated.
(3) Make afs_select_fileserver() treat -ENETRESET as a straight fail of
the operation.
(4) Prioritise the error code over things like -ECONNRESET as the server
did actually respond.
(5) Make writeback treat -ENETRESET as a retryable error and make it
redirty all the pages involved in a write so that the VM will retry.
Note that there is still a circumstance that I can't easily deal with: if
the operation is fully received and processed by the server, but the reply
is lost due to address change. There's no way to know if the op happened.
We can examine the server, but a conflicting change could have been made by
a third party - and we can't tell the difference. In such a case, a
message like:
kAFS: vnode modified {100058:146266} b7->b8 YFS.StoreData64 (op=2646a)
will be logged to dmesg on the next op to touch the file and the client
will reset the inode state, including invalidating clean parts of the
pagecache.
Reported-by: Marc Dionne <marc.dionne@auristor.com>
Signed-off-by: David Howells <dhowells@redhat.com>
cc: linux-afs@lists.infradead.org
Link: http://lists.infradead.org/pipermail/linux-afs/2021-December/004811.html # v1
Signed-off-by: David S. Miller <davem@davemloft.net>
2022-05-21 07:45:55 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
|
|
/* If we get an OUT_OF_SEQUENCE ACK from the server, that can also
|
|
|
|
|
* indicate a change of address. However, we can retransmit the call
|
|
|
|
|
* if we still have it buffered to the beginning.
|
|
|
|
|
*/
|
2024-02-01 10:29:36 +00:00
|
|
|
if (unlikely(summary.ack_reason == RXRPC_ACK_OUT_OF_SEQUENCE) &&
|
2024-12-04 07:46:48 +00:00
|
|
|
hard_ack == 0 &&
|
afs: Adjust ACK interpretation to try and cope with NAT
If a client's address changes, say if it is NAT'd, this can disrupt an in
progress operation. For most operations, this is not much of a problem,
but StoreData can be different as some servers modify the target file as
the data comes in, so if a store request is disrupted, the file can get
corrupted on the server.
The problem is that the server doesn't recognise packets that come after
the change of address as belonging to the original client and will bounce
them, either by sending an OUT_OF_SEQUENCE ACK to the apparent new call if
the packet number falls within the initial sequence number window of a call
or by sending an EXCEEDS_WINDOW ACK if it falls outside and then aborting
it. In both cases, firstPacket will be 1 and previousPacket will be 0 in
the ACK information.
Fix this by the following means:
(1) If a client call receives an EXCEEDS_WINDOW ACK with firstPacket as 1
and previousPacket as 0, assume this indicates that the server saw the
incoming packets from a different peer and thus as a different call.
Fail the call with error -ENETRESET.
(2) Also fail the call if a similar OUT_OF_SEQUENCE ACK occurs if the
first packet has been hard-ACK'd. If it hasn't been hard-ACK'd, the
ACK packet will cause it to get retransmitted, so the call will just
be repeated.
(3) Make afs_select_fileserver() treat -ENETRESET as a straight fail of
the operation.
(4) Prioritise the error code over things like -ECONNRESET as the server
did actually respond.
(5) Make writeback treat -ENETRESET as a retryable error and make it
redirty all the pages involved in a write so that the VM will retry.
Note that there is still a circumstance that I can't easily deal with: if
the operation is fully received and processed by the server, but the reply
is lost due to address change. There's no way to know if the op happened.
We can examine the server, but a conflicting change could have been made by
a third party - and we can't tell the difference. In such a case, a
message like:
kAFS: vnode modified {100058:146266} b7->b8 YFS.StoreData64 (op=2646a)
will be logged to dmesg on the next op to touch the file and the client
will reset the inode state, including invalidating clean parts of the
pagecache.
Reported-by: Marc Dionne <marc.dionne@auristor.com>
Signed-off-by: David Howells <dhowells@redhat.com>
cc: linux-afs@lists.infradead.org
Link: http://lists.infradead.org/pipermail/linux-afs/2021-December/004811.html # v1
Signed-off-by: David S. Miller <davem@davemloft.net>
2022-05-21 07:45:55 +00:00
|
|
|
prev_pkt == 0 &&
|
2024-12-04 07:46:47 +00:00
|
|
|
call->tx_bottom == 0 &&
|
afs: Adjust ACK interpretation to try and cope with NAT
If a client's address changes, say if it is NAT'd, this can disrupt an in
progress operation. For most operations, this is not much of a problem,
but StoreData can be different as some servers modify the target file as
the data comes in, so if a store request is disrupted, the file can get
corrupted on the server.
The problem is that the server doesn't recognise packets that come after
the change of address as belonging to the original client and will bounce
them, either by sending an OUT_OF_SEQUENCE ACK to the apparent new call if
the packet number falls within the initial sequence number window of a call
or by sending an EXCEEDS_WINDOW ACK if it falls outside and then aborting
it. In both cases, firstPacket will be 1 and previousPacket will be 0 in
the ACK information.
Fix this by the following means:
(1) If a client call receives an EXCEEDS_WINDOW ACK with firstPacket as 1
and previousPacket as 0, assume this indicates that the server saw the
incoming packets from a different peer and thus as a different call.
Fail the call with error -ENETRESET.
(2) Also fail the call if a similar OUT_OF_SEQUENCE ACK occurs if the
first packet has been hard-ACK'd. If it hasn't been hard-ACK'd, the
ACK packet will cause it to get retransmitted, so the call will just
be repeated.
(3) Make afs_select_fileserver() treat -ENETRESET as a straight fail of
the operation.
(4) Prioritise the error code over things like -ECONNRESET as the server
did actually respond.
(5) Make writeback treat -ENETRESET as a retryable error and make it
redirty all the pages involved in a write so that the VM will retry.
Note that there is still a circumstance that I can't easily deal with: if
the operation is fully received and processed by the server, but the reply
is lost due to address change. There's no way to know if the op happened.
We can examine the server, but a conflicting change could have been made by
a third party - and we can't tell the difference. In such a case, a
message like:
kAFS: vnode modified {100058:146266} b7->b8 YFS.StoreData64 (op=2646a)
will be logged to dmesg on the next op to touch the file and the client
will reset the inode state, including invalidating clean parts of the
pagecache.
Reported-by: Marc Dionne <marc.dionne@auristor.com>
Signed-off-by: David Howells <dhowells@redhat.com>
cc: linux-afs@lists.infradead.org
Link: http://lists.infradead.org/pipermail/linux-afs/2021-December/004811.html # v1
Signed-off-by: David S. Miller <davem@davemloft.net>
2022-05-21 07:45:55 +00:00
|
|
|
rxrpc_is_client_call(call)) {
|
|
|
|
|
rxrpc_set_call_completion(call, RXRPC_CALL_REMOTELY_ABORTED,
|
|
|
|
|
0, -ENETRESET);
|
2023-11-16 13:12:59 +00:00
|
|
|
goto send_response;
|
afs: Adjust ACK interpretation to try and cope with NAT
If a client's address changes, say if it is NAT'd, this can disrupt an in
progress operation. For most operations, this is not much of a problem,
but StoreData can be different as some servers modify the target file as
the data comes in, so if a store request is disrupted, the file can get
corrupted on the server.
The problem is that the server doesn't recognise packets that come after
the change of address as belonging to the original client and will bounce
them, either by sending an OUT_OF_SEQUENCE ACK to the apparent new call if
the packet number falls within the initial sequence number window of a call
or by sending an EXCEEDS_WINDOW ACK if it falls outside and then aborting
it. In both cases, firstPacket will be 1 and previousPacket will be 0 in
the ACK information.
Fix this by the following means:
(1) If a client call receives an EXCEEDS_WINDOW ACK with firstPacket as 1
and previousPacket as 0, assume this indicates that the server saw the
incoming packets from a different peer and thus as a different call.
Fail the call with error -ENETRESET.
(2) Also fail the call if a similar OUT_OF_SEQUENCE ACK occurs if the
first packet has been hard-ACK'd. If it hasn't been hard-ACK'd, the
ACK packet will cause it to get retransmitted, so the call will just
be repeated.
(3) Make afs_select_fileserver() treat -ENETRESET as a straight fail of
the operation.
(4) Prioritise the error code over things like -ECONNRESET as the server
did actually respond.
(5) Make writeback treat -ENETRESET as a retryable error and make it
redirty all the pages involved in a write so that the VM will retry.
Note that there is still a circumstance that I can't easily deal with: if
the operation is fully received and processed by the server, but the reply
is lost due to address change. There's no way to know if the op happened.
We can examine the server, but a conflicting change could have been made by
a third party - and we can't tell the difference. In such a case, a
message like:
kAFS: vnode modified {100058:146266} b7->b8 YFS.StoreData64 (op=2646a)
will be logged to dmesg on the next op to touch the file and the client
will reset the inode state, including invalidating clean parts of the
pagecache.
Reported-by: Marc Dionne <marc.dionne@auristor.com>
Signed-off-by: David Howells <dhowells@redhat.com>
cc: linux-afs@lists.infradead.org
Link: http://lists.infradead.org/pipermail/linux-afs/2021-December/004811.html # v1
Signed-off-by: David S. Miller <davem@davemloft.net>
2022-05-21 07:45:55 +00:00
|
|
|
}
|
|
|
|
|
|
2019-04-12 15:34:16 +00:00
|
|
|
/* Discard any out-of-order or duplicate ACKs (outside lock). */
|
2024-12-04 07:46:48 +00:00
|
|
|
if (!rxrpc_is_ack_valid(call, hard_ack, prev_pkt)) {
|
2024-12-04 07:47:02 +00:00
|
|
|
trace_rxrpc_rx_discard_ack(call, summary.ack_serial, hard_ack, prev_pkt);
|
2024-12-04 07:47:07 +00:00
|
|
|
goto send_response; /* Still respond if requested. */
|
2020-04-28 21:06:54 +00:00
|
|
|
}
|
rxrpc: Fix the packet reception routine
The rxrpc_input_packet() function and its call tree was built around the
assumption that data_ready() handler called from UDP to inform a kernel
service that there is data to be had was non-reentrant. This means that
certain locking could be dispensed with.
This, however, turns out not to be the case with a multi-queue network card
that can deliver packets to multiple cpus simultaneously. Each of those
cpus can be in the rxrpc_input_packet() function at the same time.
Fix by adding or changing some structure members:
(1) Add peer->rtt_input_lock to serialise access to the RTT buffer.
(2) Make conn->service_id into a 32-bit variable so that it can be
cmpxchg'd on all arches.
(3) Add call->input_lock to serialise access to the Rx/Tx state. Note
that although the Rx and Tx states are (almost) entirely separate,
there's no point completing the separation and having separate locks
since it's a bi-phasal RPC protocol rather than a bi-direction
streaming protocol. Data transmission and data reception do not take
place simultaneously on any particular call.
and making the following functional changes:
(1) In rxrpc_input_data(), hold call->input_lock around the core to
prevent simultaneous producing of packets into the Rx ring and
updating of tracking state for a particular call.
(2) In rxrpc_input_ping_response(), only read call->ping_serial once, and
check it before checking RXRPC_CALL_PINGING as that's a cheaper test.
The bit test and bit clear can then be combined. No further locking
is needed here.
(3) In rxrpc_input_ack(), take call->input_lock after we've parsed much of
the ACK packet. The superseded ACK check is then done both before and
after the lock is taken.
The handing of ackinfo data is split, parsing before the lock is taken
and processing with it held. This is keyed on rxMTU being non-zero.
Congestion management is also done within the locked section.
(4) In rxrpc_input_ackall(), take call->input_lock around the Tx window
rotation. The ACKALL packet carries no information and is only really
useful after all packets have been transmitted since it's imprecise.
(5) In rxrpc_input_implicit_end_call(), we use rx->incoming_lock to
prevent calls being simultaneously implicitly ended on two cpus and
also to prevent any races with incoming call setup.
(6) In rxrpc_input_packet(), use cmpxchg() to effect the service upgrade
on a connection. It is only permitted to happen once for a
connection.
(7) In rxrpc_new_incoming_call(), we have to recheck the routing inside
rx->incoming_lock to see if someone else set up the call, connection
or peer whilst we were getting there. We can't trust the values from
the earlier routing check unless we pin refs on them - which we want
to avoid.
Further, we need to allow for an incoming call to have its state
changed on another CPU between us making it live and us adjusting it
because the conn is now in the RXRPC_CONN_SERVICE state.
(8) In rxrpc_peer_add_rtt(), take peer->rtt_input_lock around the access
to the RTT buffer. Don't need to lock around setting peer->rtt.
For reference, the inventory of state-accessing or state-altering functions
used by the packet input procedure is:
> rxrpc_input_packet()
* PACKET CHECKING
* ROUTING
> rxrpc_post_packet_to_local()
> rxrpc_find_connection_rcu() - uses RCU
> rxrpc_lookup_peer_rcu() - uses RCU
> rxrpc_find_service_conn_rcu() - uses RCU
> idr_find() - uses RCU
* CONNECTION-LEVEL PROCESSING
- Service upgrade
- Can only happen once per conn
! Changed to use cmpxchg
> rxrpc_post_packet_to_conn()
- Setting conn->hi_serial
- Probably safe not using locks
- Maybe use cmpxchg
* CALL-LEVEL PROCESSING
> Old-call checking
> rxrpc_input_implicit_end_call()
> rxrpc_call_completed()
> rxrpc_queue_call()
! Need to take rx->incoming_lock
> __rxrpc_disconnect_call()
> rxrpc_notify_socket()
> rxrpc_new_incoming_call()
- Uses rx->incoming_lock for the entire process
- Might be able to drop this earlier in favour of the call lock
> rxrpc_incoming_call()
! Conflicts with rxrpc_input_implicit_end_call()
> rxrpc_send_ping()
- Don't need locks to check rtt state
> rxrpc_propose_ACK
* PACKET DISTRIBUTION
> rxrpc_input_call_packet()
> rxrpc_input_data()
* QUEUE DATA PACKET ON CALL
> rxrpc_reduce_call_timer()
- Uses timer_reduce()
! Needs call->input_lock()
> rxrpc_receiving_reply()
! Needs locking around ack state
> rxrpc_rotate_tx_window()
> rxrpc_end_tx_phase()
> rxrpc_proto_abort()
> rxrpc_input_dup_data()
- Fills the Rx buffer
- rxrpc_propose_ACK()
- rxrpc_notify_socket()
> rxrpc_input_ack()
* APPLY ACK PACKET TO CALL AND DISCARD PACKET
> rxrpc_input_ping_response()
- Probably doesn't need any extra locking
! Need READ_ONCE() on call->ping_serial
> rxrpc_input_check_for_lost_ack()
- Takes call->lock to consult Tx buffer
> rxrpc_peer_add_rtt()
! Needs to take a lock (peer->rtt_input_lock)
! Could perhaps manage with cmpxchg() and xadd() instead
> rxrpc_input_requested_ack
- Consults Tx buffer
! Probably needs a lock
> rxrpc_peer_add_rtt()
> rxrpc_propose_ack()
> rxrpc_input_ackinfo()
- Changes call->tx_winsize
! Use cmpxchg to handle change
! Should perhaps track serial number
- Uses peer->lock to record MTU specification changes
> rxrpc_proto_abort()
! Need to take call->input_lock
> rxrpc_rotate_tx_window()
> rxrpc_end_tx_phase()
> rxrpc_input_soft_acks()
- Consults the Tx buffer
> rxrpc_congestion_management()
- Modifies the Tx annotations
! Needs call->input_lock()
> rxrpc_queue_call()
> rxrpc_input_abort()
* APPLY ABORT PACKET TO CALL AND DISCARD PACKET
> rxrpc_set_call_completion()
> rxrpc_notify_socket()
> rxrpc_input_ackall()
* APPLY ACKALL PACKET TO CALL AND DISCARD PACKET
! Need to take call->input_lock
> rxrpc_rotate_tx_window()
> rxrpc_end_tx_phase()
> rxrpc_reject_packet()
There are some functions used by the above that queue the packet, after
which the procedure is terminated:
- rxrpc_post_packet_to_local()
- local->event_queue is an sk_buff_head
- local->processor is a work_struct
- rxrpc_post_packet_to_conn()
- conn->rx_queue is an sk_buff_head
- conn->processor is a work_struct
- rxrpc_reject_packet()
- local->reject_queue is an sk_buff_head
- local->processor is a work_struct
And some that offload processing to process context:
- rxrpc_notify_socket()
- Uses RCU lock
- Uses call->notify_lock to call call->notify_rx
- Uses call->recvmsg_lock to queue recvmsg side
- rxrpc_queue_call()
- call->processor is a work_struct
- rxrpc_propose_ACK()
- Uses call->lock to wrap __rxrpc_propose_ACK()
And a bunch that complete a call, all of which use call->state_lock to
protect the call state:
- rxrpc_call_completed()
- rxrpc_set_call_completion()
- rxrpc_abort_call()
- rxrpc_proto_abort()
- Also uses rxrpc_queue_call()
Fixes: 17926a79320a ("[AF_RXRPC]: Provide secure RxRPC sockets for use by userspace and kernel both")
Signed-off-by: David Howells <dhowells@redhat.com>
2018-10-08 14:46:25 +00:00
|
|
|
|
2024-01-26 16:17:03 +00:00
|
|
|
trailer.maxMTU = 0;
|
2016-09-30 12:26:03 +00:00
|
|
|
ioffset = offset + nr_acks + 3;
|
2024-01-26 16:17:03 +00:00
|
|
|
if (skb->len >= ioffset + sizeof(trailer) &&
|
|
|
|
|
skb_copy_bits(skb, ioffset, &trailer, sizeof(trailer)) < 0)
|
|
|
|
|
return rxrpc_proto_abort(call, 0, rxrpc_badmsg_short_ack_trailer);
|
2022-05-07 09:06:13 +00:00
|
|
|
|
|
|
|
|
if (nr_acks > 0)
|
|
|
|
|
skb_condense(skb);
|
rxrpc: Fix the packet reception routine
The rxrpc_input_packet() function and its call tree was built around the
assumption that data_ready() handler called from UDP to inform a kernel
service that there is data to be had was non-reentrant. This means that
certain locking could be dispensed with.
This, however, turns out not to be the case with a multi-queue network card
that can deliver packets to multiple cpus simultaneously. Each of those
cpus can be in the rxrpc_input_packet() function at the same time.
Fix by adding or changing some structure members:
(1) Add peer->rtt_input_lock to serialise access to the RTT buffer.
(2) Make conn->service_id into a 32-bit variable so that it can be
cmpxchg'd on all arches.
(3) Add call->input_lock to serialise access to the Rx/Tx state. Note
that although the Rx and Tx states are (almost) entirely separate,
there's no point completing the separation and having separate locks
since it's a bi-phasal RPC protocol rather than a bi-direction
streaming protocol. Data transmission and data reception do not take
place simultaneously on any particular call.
and making the following functional changes:
(1) In rxrpc_input_data(), hold call->input_lock around the core to
prevent simultaneous producing of packets into the Rx ring and
updating of tracking state for a particular call.
(2) In rxrpc_input_ping_response(), only read call->ping_serial once, and
check it before checking RXRPC_CALL_PINGING as that's a cheaper test.
The bit test and bit clear can then be combined. No further locking
is needed here.
(3) In rxrpc_input_ack(), take call->input_lock after we've parsed much of
the ACK packet. The superseded ACK check is then done both before and
after the lock is taken.
The handing of ackinfo data is split, parsing before the lock is taken
and processing with it held. This is keyed on rxMTU being non-zero.
Congestion management is also done within the locked section.
(4) In rxrpc_input_ackall(), take call->input_lock around the Tx window
rotation. The ACKALL packet carries no information and is only really
useful after all packets have been transmitted since it's imprecise.
(5) In rxrpc_input_implicit_end_call(), we use rx->incoming_lock to
prevent calls being simultaneously implicitly ended on two cpus and
also to prevent any races with incoming call setup.
(6) In rxrpc_input_packet(), use cmpxchg() to effect the service upgrade
on a connection. It is only permitted to happen once for a
connection.
(7) In rxrpc_new_incoming_call(), we have to recheck the routing inside
rx->incoming_lock to see if someone else set up the call, connection
or peer whilst we were getting there. We can't trust the values from
the earlier routing check unless we pin refs on them - which we want
to avoid.
Further, we need to allow for an incoming call to have its state
changed on another CPU between us making it live and us adjusting it
because the conn is now in the RXRPC_CONN_SERVICE state.
(8) In rxrpc_peer_add_rtt(), take peer->rtt_input_lock around the access
to the RTT buffer. Don't need to lock around setting peer->rtt.
For reference, the inventory of state-accessing or state-altering functions
used by the packet input procedure is:
> rxrpc_input_packet()
* PACKET CHECKING
* ROUTING
> rxrpc_post_packet_to_local()
> rxrpc_find_connection_rcu() - uses RCU
> rxrpc_lookup_peer_rcu() - uses RCU
> rxrpc_find_service_conn_rcu() - uses RCU
> idr_find() - uses RCU
* CONNECTION-LEVEL PROCESSING
- Service upgrade
- Can only happen once per conn
! Changed to use cmpxchg
> rxrpc_post_packet_to_conn()
- Setting conn->hi_serial
- Probably safe not using locks
- Maybe use cmpxchg
* CALL-LEVEL PROCESSING
> Old-call checking
> rxrpc_input_implicit_end_call()
> rxrpc_call_completed()
> rxrpc_queue_call()
! Need to take rx->incoming_lock
> __rxrpc_disconnect_call()
> rxrpc_notify_socket()
> rxrpc_new_incoming_call()
- Uses rx->incoming_lock for the entire process
- Might be able to drop this earlier in favour of the call lock
> rxrpc_incoming_call()
! Conflicts with rxrpc_input_implicit_end_call()
> rxrpc_send_ping()
- Don't need locks to check rtt state
> rxrpc_propose_ACK
* PACKET DISTRIBUTION
> rxrpc_input_call_packet()
> rxrpc_input_data()
* QUEUE DATA PACKET ON CALL
> rxrpc_reduce_call_timer()
- Uses timer_reduce()
! Needs call->input_lock()
> rxrpc_receiving_reply()
! Needs locking around ack state
> rxrpc_rotate_tx_window()
> rxrpc_end_tx_phase()
> rxrpc_proto_abort()
> rxrpc_input_dup_data()
- Fills the Rx buffer
- rxrpc_propose_ACK()
- rxrpc_notify_socket()
> rxrpc_input_ack()
* APPLY ACK PACKET TO CALL AND DISCARD PACKET
> rxrpc_input_ping_response()
- Probably doesn't need any extra locking
! Need READ_ONCE() on call->ping_serial
> rxrpc_input_check_for_lost_ack()
- Takes call->lock to consult Tx buffer
> rxrpc_peer_add_rtt()
! Needs to take a lock (peer->rtt_input_lock)
! Could perhaps manage with cmpxchg() and xadd() instead
> rxrpc_input_requested_ack
- Consults Tx buffer
! Probably needs a lock
> rxrpc_peer_add_rtt()
> rxrpc_propose_ack()
> rxrpc_input_ackinfo()
- Changes call->tx_winsize
! Use cmpxchg to handle change
! Should perhaps track serial number
- Uses peer->lock to record MTU specification changes
> rxrpc_proto_abort()
! Need to take call->input_lock
> rxrpc_rotate_tx_window()
> rxrpc_end_tx_phase()
> rxrpc_input_soft_acks()
- Consults the Tx buffer
> rxrpc_congestion_management()
- Modifies the Tx annotations
! Needs call->input_lock()
> rxrpc_queue_call()
> rxrpc_input_abort()
* APPLY ABORT PACKET TO CALL AND DISCARD PACKET
> rxrpc_set_call_completion()
> rxrpc_notify_socket()
> rxrpc_input_ackall()
* APPLY ACKALL PACKET TO CALL AND DISCARD PACKET
! Need to take call->input_lock
> rxrpc_rotate_tx_window()
> rxrpc_end_tx_phase()
> rxrpc_reject_packet()
There are some functions used by the above that queue the packet, after
which the procedure is terminated:
- rxrpc_post_packet_to_local()
- local->event_queue is an sk_buff_head
- local->processor is a work_struct
- rxrpc_post_packet_to_conn()
- conn->rx_queue is an sk_buff_head
- conn->processor is a work_struct
- rxrpc_reject_packet()
- local->reject_queue is an sk_buff_head
- local->processor is a work_struct
And some that offload processing to process context:
- rxrpc_notify_socket()
- Uses RCU lock
- Uses call->notify_lock to call call->notify_rx
- Uses call->recvmsg_lock to queue recvmsg side
- rxrpc_queue_call()
- call->processor is a work_struct
- rxrpc_propose_ACK()
- Uses call->lock to wrap __rxrpc_propose_ACK()
And a bunch that complete a call, all of which use call->state_lock to
protect the call state:
- rxrpc_call_completed()
- rxrpc_set_call_completion()
- rxrpc_abort_call()
- rxrpc_proto_abort()
- Also uses rxrpc_queue_call()
Fixes: 17926a79320a ("[AF_RXRPC]: Provide secure RxRPC sockets for use by userspace and kernel both")
Signed-off-by: David Howells <dhowells@redhat.com>
2018-10-08 14:46:25 +00:00
|
|
|
|
2024-12-04 07:46:53 +00:00
|
|
|
call->acks_latest_ts = ktime_get_real();
|
2024-12-04 07:46:48 +00:00
|
|
|
call->acks_hard_ack = hard_ack;
|
2022-05-21 08:03:18 +00:00
|
|
|
call->acks_prev_seq = prev_pkt;
|
2019-04-12 15:34:16 +00:00
|
|
|
|
2024-12-04 07:47:07 +00:00
|
|
|
if (summary.acked_serial) {
|
|
|
|
|
switch (summary.ack_reason) {
|
|
|
|
|
case RXRPC_ACK_PING_RESPONSE:
|
|
|
|
|
rxrpc_complete_rtt_probe(call, call->acks_latest_ts,
|
|
|
|
|
summary.acked_serial, summary.ack_serial,
|
|
|
|
|
rxrpc_rtt_rx_ping_response);
|
|
|
|
|
break;
|
|
|
|
|
default:
|
|
|
|
|
if (after(summary.acked_serial, call->acks_highest_serial))
|
|
|
|
|
call->acks_highest_serial = summary.acked_serial;
|
|
|
|
|
summary.rtt_sample_avail = true;
|
|
|
|
|
break;
|
|
|
|
|
}
|
2022-05-07 09:06:13 +00:00
|
|
|
}
|
2022-04-28 07:30:47 +00:00
|
|
|
|
2018-10-08 14:46:11 +00:00
|
|
|
/* Parse rwind and mtu sizes if provided. */
|
2024-01-26 16:17:03 +00:00
|
|
|
if (trailer.maxMTU)
|
|
|
|
|
rxrpc_input_ack_trailer(call, skb, &trailer);
|
2007-04-26 22:48:28 +00:00
|
|
|
|
2024-12-04 07:46:48 +00:00
|
|
|
if (hard_ack + 1 == 0)
|
2022-10-06 20:45:42 +00:00
|
|
|
return rxrpc_proto_abort(call, 0, rxrpc_eproto_ackr_zero);
|
2007-04-26 22:48:28 +00:00
|
|
|
|
rxrpc: Rewrite the data and ack handling code
Rewrite the data and ack handling code such that:
(1) Parsing of received ACK and ABORT packets and the distribution and the
filing of DATA packets happens entirely within the data_ready context
called from the UDP socket. This allows us to process and discard ACK
and ABORT packets much more quickly (they're no longer stashed on a
queue for a background thread to process).
(2) We avoid calling skb_clone(), pskb_pull() and pskb_trim(). We instead
keep track of the offset and length of the content of each packet in
the sk_buff metadata. This means we don't do any allocation in the
receive path.
(3) Jumbo DATA packet parsing is now done in data_ready context. Rather
than cloning the packet once for each subpacket and pulling/trimming
it, we file the packet multiple times with an annotation for each
indicating which subpacket is there. From that we can directly
calculate the offset and length.
(4) A call's receive queue can be accessed without taking locks (memory
barriers do have to be used, though).
(5) Incoming calls are set up from preallocated resources and immediately
made live. They can than have packets queued upon them and ACKs
generated. If insufficient resources exist, DATA packet #1 is given a
BUSY reply and other DATA packets are discarded).
(6) sk_buffs no longer take a ref on their parent call.
To make this work, the following changes are made:
(1) Each call's receive buffer is now a circular buffer of sk_buff
pointers (rxtx_buffer) rather than a number of sk_buff_heads spread
between the call and the socket. This permits each sk_buff to be in
the buffer multiple times. The receive buffer is reused for the
transmit buffer.
(2) A circular buffer of annotations (rxtx_annotations) is kept parallel
to the data buffer. Transmission phase annotations indicate whether a
buffered packet has been ACK'd or not and whether it needs
retransmission.
Receive phase annotations indicate whether a slot holds a whole packet
or a jumbo subpacket and, if the latter, which subpacket. They also
note whether the packet has been decrypted in place.
(3) DATA packet window tracking is much simplified. Each phase has just
two numbers representing the window (rx_hard_ack/rx_top and
tx_hard_ack/tx_top).
The hard_ack number is the sequence number before base of the window,
representing the last packet the other side says it has consumed.
hard_ack starts from 0 and the first packet is sequence number 1.
The top number is the sequence number of the highest-numbered packet
residing in the buffer. Packets between hard_ack+1 and top are
soft-ACK'd to indicate they've been received, but not yet consumed.
Four macros, before(), before_eq(), after() and after_eq() are added
to compare sequence numbers within the window. This allows for the
top of the window to wrap when the hard-ack sequence number gets close
to the limit.
Two flags, RXRPC_CALL_RX_LAST and RXRPC_CALL_TX_LAST, are added also
to indicate when rx_top and tx_top point at the packets with the
LAST_PACKET bit set, indicating the end of the phase.
(4) Calls are queued on the socket 'receive queue' rather than packets.
This means that we don't need have to invent dummy packets to queue to
indicate abnormal/terminal states and we don't have to keep metadata
packets (such as ABORTs) around
(5) The offset and length of a (sub)packet's content are now passed to
the verify_packet security op. This is currently expected to decrypt
the packet in place and validate it.
However, there's now nowhere to store the revised offset and length of
the actual data within the decrypted blob (there may be a header and
padding to skip) because an sk_buff may represent multiple packets, so
a locate_data security op is added to retrieve these details from the
sk_buff content when needed.
(6) recvmsg() now has to handle jumbo subpackets, where each subpacket is
individually secured and needs to be individually decrypted. The code
to do this is broken out into rxrpc_recvmsg_data() and shared with the
kernel API. It now iterates over the call's receive buffer rather
than walking the socket receive queue.
Additional changes:
(1) The timers are condensed to a single timer that is set for the soonest
of three timeouts (delayed ACK generation, DATA retransmission and
call lifespan).
(2) Transmission of ACK and ABORT packets is effected immediately from
process-context socket ops/kernel API calls that cause them instead of
them being punted off to a background work item. The data_ready
handler still has to defer to the background, though.
(3) A shutdown op is added to the AF_RXRPC socket so that the AFS
filesystem can shut down the socket and flush its own work items
before closing the socket to deal with any in-progress service calls.
Future additional changes that will need to be considered:
(1) Make sure that a call doesn't hog the front of the queue by receiving
data from the network as fast as userspace is consuming it to the
exclusion of other calls.
(2) Transmit delayed ACKs from within recvmsg() when we've consumed
sufficiently more packets to avoid the background work item needing to
run.
Signed-off-by: David Howells <dhowells@redhat.com>
2016-09-08 10:10:12 +00:00
|
|
|
/* Ignore ACKs unless we are or have just been transmitting. */
|
2022-10-27 10:25:55 +00:00
|
|
|
switch (__rxrpc_call_state(call)) {
|
rxrpc: Rewrite the data and ack handling code
Rewrite the data and ack handling code such that:
(1) Parsing of received ACK and ABORT packets and the distribution and the
filing of DATA packets happens entirely within the data_ready context
called from the UDP socket. This allows us to process and discard ACK
and ABORT packets much more quickly (they're no longer stashed on a
queue for a background thread to process).
(2) We avoid calling skb_clone(), pskb_pull() and pskb_trim(). We instead
keep track of the offset and length of the content of each packet in
the sk_buff metadata. This means we don't do any allocation in the
receive path.
(3) Jumbo DATA packet parsing is now done in data_ready context. Rather
than cloning the packet once for each subpacket and pulling/trimming
it, we file the packet multiple times with an annotation for each
indicating which subpacket is there. From that we can directly
calculate the offset and length.
(4) A call's receive queue can be accessed without taking locks (memory
barriers do have to be used, though).
(5) Incoming calls are set up from preallocated resources and immediately
made live. They can than have packets queued upon them and ACKs
generated. If insufficient resources exist, DATA packet #1 is given a
BUSY reply and other DATA packets are discarded).
(6) sk_buffs no longer take a ref on their parent call.
To make this work, the following changes are made:
(1) Each call's receive buffer is now a circular buffer of sk_buff
pointers (rxtx_buffer) rather than a number of sk_buff_heads spread
between the call and the socket. This permits each sk_buff to be in
the buffer multiple times. The receive buffer is reused for the
transmit buffer.
(2) A circular buffer of annotations (rxtx_annotations) is kept parallel
to the data buffer. Transmission phase annotations indicate whether a
buffered packet has been ACK'd or not and whether it needs
retransmission.
Receive phase annotations indicate whether a slot holds a whole packet
or a jumbo subpacket and, if the latter, which subpacket. They also
note whether the packet has been decrypted in place.
(3) DATA packet window tracking is much simplified. Each phase has just
two numbers representing the window (rx_hard_ack/rx_top and
tx_hard_ack/tx_top).
The hard_ack number is the sequence number before base of the window,
representing the last packet the other side says it has consumed.
hard_ack starts from 0 and the first packet is sequence number 1.
The top number is the sequence number of the highest-numbered packet
residing in the buffer. Packets between hard_ack+1 and top are
soft-ACK'd to indicate they've been received, but not yet consumed.
Four macros, before(), before_eq(), after() and after_eq() are added
to compare sequence numbers within the window. This allows for the
top of the window to wrap when the hard-ack sequence number gets close
to the limit.
Two flags, RXRPC_CALL_RX_LAST and RXRPC_CALL_TX_LAST, are added also
to indicate when rx_top and tx_top point at the packets with the
LAST_PACKET bit set, indicating the end of the phase.
(4) Calls are queued on the socket 'receive queue' rather than packets.
This means that we don't need have to invent dummy packets to queue to
indicate abnormal/terminal states and we don't have to keep metadata
packets (such as ABORTs) around
(5) The offset and length of a (sub)packet's content are now passed to
the verify_packet security op. This is currently expected to decrypt
the packet in place and validate it.
However, there's now nowhere to store the revised offset and length of
the actual data within the decrypted blob (there may be a header and
padding to skip) because an sk_buff may represent multiple packets, so
a locate_data security op is added to retrieve these details from the
sk_buff content when needed.
(6) recvmsg() now has to handle jumbo subpackets, where each subpacket is
individually secured and needs to be individually decrypted. The code
to do this is broken out into rxrpc_recvmsg_data() and shared with the
kernel API. It now iterates over the call's receive buffer rather
than walking the socket receive queue.
Additional changes:
(1) The timers are condensed to a single timer that is set for the soonest
of three timeouts (delayed ACK generation, DATA retransmission and
call lifespan).
(2) Transmission of ACK and ABORT packets is effected immediately from
process-context socket ops/kernel API calls that cause them instead of
them being punted off to a background work item. The data_ready
handler still has to defer to the background, though.
(3) A shutdown op is added to the AF_RXRPC socket so that the AFS
filesystem can shut down the socket and flush its own work items
before closing the socket to deal with any in-progress service calls.
Future additional changes that will need to be considered:
(1) Make sure that a call doesn't hog the front of the queue by receiving
data from the network as fast as userspace is consuming it to the
exclusion of other calls.
(2) Transmit delayed ACKs from within recvmsg() when we've consumed
sufficiently more packets to avoid the background work item needing to
run.
Signed-off-by: David Howells <dhowells@redhat.com>
2016-09-08 10:10:12 +00:00
|
|
|
case RXRPC_CALL_CLIENT_SEND_REQUEST:
|
|
|
|
|
case RXRPC_CALL_CLIENT_AWAIT_REPLY:
|
|
|
|
|
case RXRPC_CALL_SERVER_SEND_REPLY:
|
|
|
|
|
case RXRPC_CALL_SERVER_AWAIT_ACK:
|
|
|
|
|
break;
|
2007-04-26 22:48:28 +00:00
|
|
|
default:
|
2023-11-16 13:12:59 +00:00
|
|
|
goto send_response;
|
rxrpc: Rewrite the data and ack handling code
Rewrite the data and ack handling code such that:
(1) Parsing of received ACK and ABORT packets and the distribution and the
filing of DATA packets happens entirely within the data_ready context
called from the UDP socket. This allows us to process and discard ACK
and ABORT packets much more quickly (they're no longer stashed on a
queue for a background thread to process).
(2) We avoid calling skb_clone(), pskb_pull() and pskb_trim(). We instead
keep track of the offset and length of the content of each packet in
the sk_buff metadata. This means we don't do any allocation in the
receive path.
(3) Jumbo DATA packet parsing is now done in data_ready context. Rather
than cloning the packet once for each subpacket and pulling/trimming
it, we file the packet multiple times with an annotation for each
indicating which subpacket is there. From that we can directly
calculate the offset and length.
(4) A call's receive queue can be accessed without taking locks (memory
barriers do have to be used, though).
(5) Incoming calls are set up from preallocated resources and immediately
made live. They can than have packets queued upon them and ACKs
generated. If insufficient resources exist, DATA packet #1 is given a
BUSY reply and other DATA packets are discarded).
(6) sk_buffs no longer take a ref on their parent call.
To make this work, the following changes are made:
(1) Each call's receive buffer is now a circular buffer of sk_buff
pointers (rxtx_buffer) rather than a number of sk_buff_heads spread
between the call and the socket. This permits each sk_buff to be in
the buffer multiple times. The receive buffer is reused for the
transmit buffer.
(2) A circular buffer of annotations (rxtx_annotations) is kept parallel
to the data buffer. Transmission phase annotations indicate whether a
buffered packet has been ACK'd or not and whether it needs
retransmission.
Receive phase annotations indicate whether a slot holds a whole packet
or a jumbo subpacket and, if the latter, which subpacket. They also
note whether the packet has been decrypted in place.
(3) DATA packet window tracking is much simplified. Each phase has just
two numbers representing the window (rx_hard_ack/rx_top and
tx_hard_ack/tx_top).
The hard_ack number is the sequence number before base of the window,
representing the last packet the other side says it has consumed.
hard_ack starts from 0 and the first packet is sequence number 1.
The top number is the sequence number of the highest-numbered packet
residing in the buffer. Packets between hard_ack+1 and top are
soft-ACK'd to indicate they've been received, but not yet consumed.
Four macros, before(), before_eq(), after() and after_eq() are added
to compare sequence numbers within the window. This allows for the
top of the window to wrap when the hard-ack sequence number gets close
to the limit.
Two flags, RXRPC_CALL_RX_LAST and RXRPC_CALL_TX_LAST, are added also
to indicate when rx_top and tx_top point at the packets with the
LAST_PACKET bit set, indicating the end of the phase.
(4) Calls are queued on the socket 'receive queue' rather than packets.
This means that we don't need have to invent dummy packets to queue to
indicate abnormal/terminal states and we don't have to keep metadata
packets (such as ABORTs) around
(5) The offset and length of a (sub)packet's content are now passed to
the verify_packet security op. This is currently expected to decrypt
the packet in place and validate it.
However, there's now nowhere to store the revised offset and length of
the actual data within the decrypted blob (there may be a header and
padding to skip) because an sk_buff may represent multiple packets, so
a locate_data security op is added to retrieve these details from the
sk_buff content when needed.
(6) recvmsg() now has to handle jumbo subpackets, where each subpacket is
individually secured and needs to be individually decrypted. The code
to do this is broken out into rxrpc_recvmsg_data() and shared with the
kernel API. It now iterates over the call's receive buffer rather
than walking the socket receive queue.
Additional changes:
(1) The timers are condensed to a single timer that is set for the soonest
of three timeouts (delayed ACK generation, DATA retransmission and
call lifespan).
(2) Transmission of ACK and ABORT packets is effected immediately from
process-context socket ops/kernel API calls that cause them instead of
them being punted off to a background work item. The data_ready
handler still has to defer to the background, though.
(3) A shutdown op is added to the AF_RXRPC socket so that the AFS
filesystem can shut down the socket and flush its own work items
before closing the socket to deal with any in-progress service calls.
Future additional changes that will need to be considered:
(1) Make sure that a call doesn't hog the front of the queue by receiving
data from the network as fast as userspace is consuming it to the
exclusion of other calls.
(2) Transmit delayed ACKs from within recvmsg() when we've consumed
sufficiently more packets to avoid the background work item needing to
run.
Signed-off-by: David Howells <dhowells@redhat.com>
2016-09-08 10:10:12 +00:00
|
|
|
}
|
2007-04-26 22:48:28 +00:00
|
|
|
|
2024-12-04 07:46:47 +00:00
|
|
|
if (before(hard_ack, call->tx_bottom) ||
|
2020-01-23 13:13:41 +00:00
|
|
|
after(hard_ack, call->tx_top))
|
2022-10-06 20:45:42 +00:00
|
|
|
return rxrpc_proto_abort(call, 0, rxrpc_eproto_ackr_outside_window);
|
2020-01-23 13:13:41 +00:00
|
|
|
if (nr_acks > call->tx_top - hard_ack)
|
2022-10-06 20:45:42 +00:00
|
|
|
return rxrpc_proto_abort(call, 0, rxrpc_eproto_ackr_sack_overflow);
|
2007-04-26 22:48:28 +00:00
|
|
|
|
2024-12-04 07:46:47 +00:00
|
|
|
if (after(hard_ack, call->tx_bottom)) {
|
2024-12-04 07:47:02 +00:00
|
|
|
if (rxrpc_rotate_tx_window(call, hard_ack, &summary)) {
|
2022-10-06 20:45:42 +00:00
|
|
|
rxrpc_end_tx_phase(call, false, rxrpc_eproto_unexpected_ack);
|
2023-11-16 13:12:59 +00:00
|
|
|
goto send_response;
|
2018-10-08 14:46:01 +00:00
|
|
|
}
|
|
|
|
|
}
|
2007-04-26 22:48:28 +00:00
|
|
|
|
2016-09-23 11:39:22 +00:00
|
|
|
if (nr_acks > 0) {
|
2020-01-23 13:13:41 +00:00
|
|
|
if (offset > (int)skb->len - nr_acks)
|
2022-10-06 20:45:42 +00:00
|
|
|
return rxrpc_proto_abort(call, 0, rxrpc_eproto_ackr_short_sack);
|
2024-12-04 07:46:51 +00:00
|
|
|
rxrpc_input_soft_acks(call, &summary, skb);
|
2016-09-23 11:39:22 +00:00
|
|
|
}
|
|
|
|
|
|
2022-03-31 22:55:08 +00:00
|
|
|
if (test_bit(RXRPC_CALL_TX_LAST, &call->flags) &&
|
2024-12-04 07:46:50 +00:00
|
|
|
call->acks_nr_sacks == call->tx_top - hard_ack &&
|
2016-10-06 07:11:49 +00:00
|
|
|
rxrpc_is_client_call(call))
|
2024-12-04 07:47:02 +00:00
|
|
|
rxrpc_propose_ping(call, summary.ack_serial,
|
2020-01-30 21:48:13 +00:00
|
|
|
rxrpc_propose_ack_ping_for_lost_reply);
|
2016-09-24 17:05:27 +00:00
|
|
|
|
2024-12-04 07:47:07 +00:00
|
|
|
/* Drive the congestion management algorithm first and then RACK-TLP as
|
|
|
|
|
* the latter depends on the state/change in state in the former.
|
|
|
|
|
*/
|
2024-12-04 07:46:52 +00:00
|
|
|
rxrpc_congestion_management(call, &summary);
|
2024-12-04 07:47:07 +00:00
|
|
|
rxrpc_rack_detect_loss_and_arm_timer(call, &summary);
|
|
|
|
|
rxrpc_tlp_process_ack(call, &summary);
|
|
|
|
|
if (call->tlp_serial && after_eq(summary.acked_serial, call->tlp_serial))
|
|
|
|
|
call->tlp_serial = 0;
|
2023-11-16 13:12:59 +00:00
|
|
|
|
|
|
|
|
send_response:
|
2024-02-01 10:29:36 +00:00
|
|
|
if (summary.ack_reason == RXRPC_ACK_PING)
|
2024-12-04 07:47:02 +00:00
|
|
|
rxrpc_send_ACK(call, RXRPC_ACK_PING_RESPONSE, summary.ack_serial,
|
2023-11-16 13:12:59 +00:00
|
|
|
rxrpc_propose_ack_respond_to_ping);
|
|
|
|
|
else if (sp->hdr.flags & RXRPC_REQUEST_ACK)
|
2024-12-04 07:47:02 +00:00
|
|
|
rxrpc_send_ACK(call, RXRPC_ACK_REQUESTED, summary.ack_serial,
|
2023-11-16 13:12:59 +00:00
|
|
|
rxrpc_propose_ack_respond_to_ack);
|
2007-04-26 22:48:28 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
|
|
/*
|
rxrpc: Rewrite the data and ack handling code
Rewrite the data and ack handling code such that:
(1) Parsing of received ACK and ABORT packets and the distribution and the
filing of DATA packets happens entirely within the data_ready context
called from the UDP socket. This allows us to process and discard ACK
and ABORT packets much more quickly (they're no longer stashed on a
queue for a background thread to process).
(2) We avoid calling skb_clone(), pskb_pull() and pskb_trim(). We instead
keep track of the offset and length of the content of each packet in
the sk_buff metadata. This means we don't do any allocation in the
receive path.
(3) Jumbo DATA packet parsing is now done in data_ready context. Rather
than cloning the packet once for each subpacket and pulling/trimming
it, we file the packet multiple times with an annotation for each
indicating which subpacket is there. From that we can directly
calculate the offset and length.
(4) A call's receive queue can be accessed without taking locks (memory
barriers do have to be used, though).
(5) Incoming calls are set up from preallocated resources and immediately
made live. They can than have packets queued upon them and ACKs
generated. If insufficient resources exist, DATA packet #1 is given a
BUSY reply and other DATA packets are discarded).
(6) sk_buffs no longer take a ref on their parent call.
To make this work, the following changes are made:
(1) Each call's receive buffer is now a circular buffer of sk_buff
pointers (rxtx_buffer) rather than a number of sk_buff_heads spread
between the call and the socket. This permits each sk_buff to be in
the buffer multiple times. The receive buffer is reused for the
transmit buffer.
(2) A circular buffer of annotations (rxtx_annotations) is kept parallel
to the data buffer. Transmission phase annotations indicate whether a
buffered packet has been ACK'd or not and whether it needs
retransmission.
Receive phase annotations indicate whether a slot holds a whole packet
or a jumbo subpacket and, if the latter, which subpacket. They also
note whether the packet has been decrypted in place.
(3) DATA packet window tracking is much simplified. Each phase has just
two numbers representing the window (rx_hard_ack/rx_top and
tx_hard_ack/tx_top).
The hard_ack number is the sequence number before base of the window,
representing the last packet the other side says it has consumed.
hard_ack starts from 0 and the first packet is sequence number 1.
The top number is the sequence number of the highest-numbered packet
residing in the buffer. Packets between hard_ack+1 and top are
soft-ACK'd to indicate they've been received, but not yet consumed.
Four macros, before(), before_eq(), after() and after_eq() are added
to compare sequence numbers within the window. This allows for the
top of the window to wrap when the hard-ack sequence number gets close
to the limit.
Two flags, RXRPC_CALL_RX_LAST and RXRPC_CALL_TX_LAST, are added also
to indicate when rx_top and tx_top point at the packets with the
LAST_PACKET bit set, indicating the end of the phase.
(4) Calls are queued on the socket 'receive queue' rather than packets.
This means that we don't need have to invent dummy packets to queue to
indicate abnormal/terminal states and we don't have to keep metadata
packets (such as ABORTs) around
(5) The offset and length of a (sub)packet's content are now passed to
the verify_packet security op. This is currently expected to decrypt
the packet in place and validate it.
However, there's now nowhere to store the revised offset and length of
the actual data within the decrypted blob (there may be a header and
padding to skip) because an sk_buff may represent multiple packets, so
a locate_data security op is added to retrieve these details from the
sk_buff content when needed.
(6) recvmsg() now has to handle jumbo subpackets, where each subpacket is
individually secured and needs to be individually decrypted. The code
to do this is broken out into rxrpc_recvmsg_data() and shared with the
kernel API. It now iterates over the call's receive buffer rather
than walking the socket receive queue.
Additional changes:
(1) The timers are condensed to a single timer that is set for the soonest
of three timeouts (delayed ACK generation, DATA retransmission and
call lifespan).
(2) Transmission of ACK and ABORT packets is effected immediately from
process-context socket ops/kernel API calls that cause them instead of
them being punted off to a background work item. The data_ready
handler still has to defer to the background, though.
(3) A shutdown op is added to the AF_RXRPC socket so that the AFS
filesystem can shut down the socket and flush its own work items
before closing the socket to deal with any in-progress service calls.
Future additional changes that will need to be considered:
(1) Make sure that a call doesn't hog the front of the queue by receiving
data from the network as fast as userspace is consuming it to the
exclusion of other calls.
(2) Transmit delayed ACKs from within recvmsg() when we've consumed
sufficiently more packets to avoid the background work item needing to
run.
Signed-off-by: David Howells <dhowells@redhat.com>
2016-09-08 10:10:12 +00:00
|
|
|
* Process an ACKALL packet.
|
2007-04-26 22:48:28 +00:00
|
|
|
*/
|
rxrpc: Rewrite the data and ack handling code
Rewrite the data and ack handling code such that:
(1) Parsing of received ACK and ABORT packets and the distribution and the
filing of DATA packets happens entirely within the data_ready context
called from the UDP socket. This allows us to process and discard ACK
and ABORT packets much more quickly (they're no longer stashed on a
queue for a background thread to process).
(2) We avoid calling skb_clone(), pskb_pull() and pskb_trim(). We instead
keep track of the offset and length of the content of each packet in
the sk_buff metadata. This means we don't do any allocation in the
receive path.
(3) Jumbo DATA packet parsing is now done in data_ready context. Rather
than cloning the packet once for each subpacket and pulling/trimming
it, we file the packet multiple times with an annotation for each
indicating which subpacket is there. From that we can directly
calculate the offset and length.
(4) A call's receive queue can be accessed without taking locks (memory
barriers do have to be used, though).
(5) Incoming calls are set up from preallocated resources and immediately
made live. They can than have packets queued upon them and ACKs
generated. If insufficient resources exist, DATA packet #1 is given a
BUSY reply and other DATA packets are discarded).
(6) sk_buffs no longer take a ref on their parent call.
To make this work, the following changes are made:
(1) Each call's receive buffer is now a circular buffer of sk_buff
pointers (rxtx_buffer) rather than a number of sk_buff_heads spread
between the call and the socket. This permits each sk_buff to be in
the buffer multiple times. The receive buffer is reused for the
transmit buffer.
(2) A circular buffer of annotations (rxtx_annotations) is kept parallel
to the data buffer. Transmission phase annotations indicate whether a
buffered packet has been ACK'd or not and whether it needs
retransmission.
Receive phase annotations indicate whether a slot holds a whole packet
or a jumbo subpacket and, if the latter, which subpacket. They also
note whether the packet has been decrypted in place.
(3) DATA packet window tracking is much simplified. Each phase has just
two numbers representing the window (rx_hard_ack/rx_top and
tx_hard_ack/tx_top).
The hard_ack number is the sequence number before base of the window,
representing the last packet the other side says it has consumed.
hard_ack starts from 0 and the first packet is sequence number 1.
The top number is the sequence number of the highest-numbered packet
residing in the buffer. Packets between hard_ack+1 and top are
soft-ACK'd to indicate they've been received, but not yet consumed.
Four macros, before(), before_eq(), after() and after_eq() are added
to compare sequence numbers within the window. This allows for the
top of the window to wrap when the hard-ack sequence number gets close
to the limit.
Two flags, RXRPC_CALL_RX_LAST and RXRPC_CALL_TX_LAST, are added also
to indicate when rx_top and tx_top point at the packets with the
LAST_PACKET bit set, indicating the end of the phase.
(4) Calls are queued on the socket 'receive queue' rather than packets.
This means that we don't need have to invent dummy packets to queue to
indicate abnormal/terminal states and we don't have to keep metadata
packets (such as ABORTs) around
(5) The offset and length of a (sub)packet's content are now passed to
the verify_packet security op. This is currently expected to decrypt
the packet in place and validate it.
However, there's now nowhere to store the revised offset and length of
the actual data within the decrypted blob (there may be a header and
padding to skip) because an sk_buff may represent multiple packets, so
a locate_data security op is added to retrieve these details from the
sk_buff content when needed.
(6) recvmsg() now has to handle jumbo subpackets, where each subpacket is
individually secured and needs to be individually decrypted. The code
to do this is broken out into rxrpc_recvmsg_data() and shared with the
kernel API. It now iterates over the call's receive buffer rather
than walking the socket receive queue.
Additional changes:
(1) The timers are condensed to a single timer that is set for the soonest
of three timeouts (delayed ACK generation, DATA retransmission and
call lifespan).
(2) Transmission of ACK and ABORT packets is effected immediately from
process-context socket ops/kernel API calls that cause them instead of
them being punted off to a background work item. The data_ready
handler still has to defer to the background, though.
(3) A shutdown op is added to the AF_RXRPC socket so that the AFS
filesystem can shut down the socket and flush its own work items
before closing the socket to deal with any in-progress service calls.
Future additional changes that will need to be considered:
(1) Make sure that a call doesn't hog the front of the queue by receiving
data from the network as fast as userspace is consuming it to the
exclusion of other calls.
(2) Transmit delayed ACKs from within recvmsg() when we've consumed
sufficiently more packets to avoid the background work item needing to
run.
Signed-off-by: David Howells <dhowells@redhat.com>
2016-09-08 10:10:12 +00:00
|
|
|
static void rxrpc_input_ackall(struct rxrpc_call *call, struct sk_buff *skb)
|
2007-04-26 22:48:28 +00:00
|
|
|
{
|
2016-09-24 17:05:26 +00:00
|
|
|
struct rxrpc_ack_summary summary = { 0 };
|
2007-04-26 22:48:28 +00:00
|
|
|
|
2024-12-04 07:47:02 +00:00
|
|
|
if (rxrpc_rotate_tx_window(call, call->tx_top, &summary))
|
2022-10-06 20:45:42 +00:00
|
|
|
rxrpc_end_tx_phase(call, false, rxrpc_eproto_unexpected_ackall);
|
rxrpc: Rewrite the data and ack handling code
Rewrite the data and ack handling code such that:
(1) Parsing of received ACK and ABORT packets and the distribution and the
filing of DATA packets happens entirely within the data_ready context
called from the UDP socket. This allows us to process and discard ACK
and ABORT packets much more quickly (they're no longer stashed on a
queue for a background thread to process).
(2) We avoid calling skb_clone(), pskb_pull() and pskb_trim(). We instead
keep track of the offset and length of the content of each packet in
the sk_buff metadata. This means we don't do any allocation in the
receive path.
(3) Jumbo DATA packet parsing is now done in data_ready context. Rather
than cloning the packet once for each subpacket and pulling/trimming
it, we file the packet multiple times with an annotation for each
indicating which subpacket is there. From that we can directly
calculate the offset and length.
(4) A call's receive queue can be accessed without taking locks (memory
barriers do have to be used, though).
(5) Incoming calls are set up from preallocated resources and immediately
made live. They can than have packets queued upon them and ACKs
generated. If insufficient resources exist, DATA packet #1 is given a
BUSY reply and other DATA packets are discarded).
(6) sk_buffs no longer take a ref on their parent call.
To make this work, the following changes are made:
(1) Each call's receive buffer is now a circular buffer of sk_buff
pointers (rxtx_buffer) rather than a number of sk_buff_heads spread
between the call and the socket. This permits each sk_buff to be in
the buffer multiple times. The receive buffer is reused for the
transmit buffer.
(2) A circular buffer of annotations (rxtx_annotations) is kept parallel
to the data buffer. Transmission phase annotations indicate whether a
buffered packet has been ACK'd or not and whether it needs
retransmission.
Receive phase annotations indicate whether a slot holds a whole packet
or a jumbo subpacket and, if the latter, which subpacket. They also
note whether the packet has been decrypted in place.
(3) DATA packet window tracking is much simplified. Each phase has just
two numbers representing the window (rx_hard_ack/rx_top and
tx_hard_ack/tx_top).
The hard_ack number is the sequence number before base of the window,
representing the last packet the other side says it has consumed.
hard_ack starts from 0 and the first packet is sequence number 1.
The top number is the sequence number of the highest-numbered packet
residing in the buffer. Packets between hard_ack+1 and top are
soft-ACK'd to indicate they've been received, but not yet consumed.
Four macros, before(), before_eq(), after() and after_eq() are added
to compare sequence numbers within the window. This allows for the
top of the window to wrap when the hard-ack sequence number gets close
to the limit.
Two flags, RXRPC_CALL_RX_LAST and RXRPC_CALL_TX_LAST, are added also
to indicate when rx_top and tx_top point at the packets with the
LAST_PACKET bit set, indicating the end of the phase.
(4) Calls are queued on the socket 'receive queue' rather than packets.
This means that we don't need have to invent dummy packets to queue to
indicate abnormal/terminal states and we don't have to keep metadata
packets (such as ABORTs) around
(5) The offset and length of a (sub)packet's content are now passed to
the verify_packet security op. This is currently expected to decrypt
the packet in place and validate it.
However, there's now nowhere to store the revised offset and length of
the actual data within the decrypted blob (there may be a header and
padding to skip) because an sk_buff may represent multiple packets, so
a locate_data security op is added to retrieve these details from the
sk_buff content when needed.
(6) recvmsg() now has to handle jumbo subpackets, where each subpacket is
individually secured and needs to be individually decrypted. The code
to do this is broken out into rxrpc_recvmsg_data() and shared with the
kernel API. It now iterates over the call's receive buffer rather
than walking the socket receive queue.
Additional changes:
(1) The timers are condensed to a single timer that is set for the soonest
of three timeouts (delayed ACK generation, DATA retransmission and
call lifespan).
(2) Transmission of ACK and ABORT packets is effected immediately from
process-context socket ops/kernel API calls that cause them instead of
them being punted off to a background work item. The data_ready
handler still has to defer to the background, though.
(3) A shutdown op is added to the AF_RXRPC socket so that the AFS
filesystem can shut down the socket and flush its own work items
before closing the socket to deal with any in-progress service calls.
Future additional changes that will need to be considered:
(1) Make sure that a call doesn't hog the front of the queue by receiving
data from the network as fast as userspace is consuming it to the
exclusion of other calls.
(2) Transmit delayed ACKs from within recvmsg() when we've consumed
sufficiently more packets to avoid the background work item needing to
run.
Signed-off-by: David Howells <dhowells@redhat.com>
2016-09-08 10:10:12 +00:00
|
|
|
}
|
2007-04-26 22:48:28 +00:00
|
|
|
|
rxrpc: Rewrite the data and ack handling code
Rewrite the data and ack handling code such that:
(1) Parsing of received ACK and ABORT packets and the distribution and the
filing of DATA packets happens entirely within the data_ready context
called from the UDP socket. This allows us to process and discard ACK
and ABORT packets much more quickly (they're no longer stashed on a
queue for a background thread to process).
(2) We avoid calling skb_clone(), pskb_pull() and pskb_trim(). We instead
keep track of the offset and length of the content of each packet in
the sk_buff metadata. This means we don't do any allocation in the
receive path.
(3) Jumbo DATA packet parsing is now done in data_ready context. Rather
than cloning the packet once for each subpacket and pulling/trimming
it, we file the packet multiple times with an annotation for each
indicating which subpacket is there. From that we can directly
calculate the offset and length.
(4) A call's receive queue can be accessed without taking locks (memory
barriers do have to be used, though).
(5) Incoming calls are set up from preallocated resources and immediately
made live. They can than have packets queued upon them and ACKs
generated. If insufficient resources exist, DATA packet #1 is given a
BUSY reply and other DATA packets are discarded).
(6) sk_buffs no longer take a ref on their parent call.
To make this work, the following changes are made:
(1) Each call's receive buffer is now a circular buffer of sk_buff
pointers (rxtx_buffer) rather than a number of sk_buff_heads spread
between the call and the socket. This permits each sk_buff to be in
the buffer multiple times. The receive buffer is reused for the
transmit buffer.
(2) A circular buffer of annotations (rxtx_annotations) is kept parallel
to the data buffer. Transmission phase annotations indicate whether a
buffered packet has been ACK'd or not and whether it needs
retransmission.
Receive phase annotations indicate whether a slot holds a whole packet
or a jumbo subpacket and, if the latter, which subpacket. They also
note whether the packet has been decrypted in place.
(3) DATA packet window tracking is much simplified. Each phase has just
two numbers representing the window (rx_hard_ack/rx_top and
tx_hard_ack/tx_top).
The hard_ack number is the sequence number before base of the window,
representing the last packet the other side says it has consumed.
hard_ack starts from 0 and the first packet is sequence number 1.
The top number is the sequence number of the highest-numbered packet
residing in the buffer. Packets between hard_ack+1 and top are
soft-ACK'd to indicate they've been received, but not yet consumed.
Four macros, before(), before_eq(), after() and after_eq() are added
to compare sequence numbers within the window. This allows for the
top of the window to wrap when the hard-ack sequence number gets close
to the limit.
Two flags, RXRPC_CALL_RX_LAST and RXRPC_CALL_TX_LAST, are added also
to indicate when rx_top and tx_top point at the packets with the
LAST_PACKET bit set, indicating the end of the phase.
(4) Calls are queued on the socket 'receive queue' rather than packets.
This means that we don't need have to invent dummy packets to queue to
indicate abnormal/terminal states and we don't have to keep metadata
packets (such as ABORTs) around
(5) The offset and length of a (sub)packet's content are now passed to
the verify_packet security op. This is currently expected to decrypt
the packet in place and validate it.
However, there's now nowhere to store the revised offset and length of
the actual data within the decrypted blob (there may be a header and
padding to skip) because an sk_buff may represent multiple packets, so
a locate_data security op is added to retrieve these details from the
sk_buff content when needed.
(6) recvmsg() now has to handle jumbo subpackets, where each subpacket is
individually secured and needs to be individually decrypted. The code
to do this is broken out into rxrpc_recvmsg_data() and shared with the
kernel API. It now iterates over the call's receive buffer rather
than walking the socket receive queue.
Additional changes:
(1) The timers are condensed to a single timer that is set for the soonest
of three timeouts (delayed ACK generation, DATA retransmission and
call lifespan).
(2) Transmission of ACK and ABORT packets is effected immediately from
process-context socket ops/kernel API calls that cause them instead of
them being punted off to a background work item. The data_ready
handler still has to defer to the background, though.
(3) A shutdown op is added to the AF_RXRPC socket so that the AFS
filesystem can shut down the socket and flush its own work items
before closing the socket to deal with any in-progress service calls.
Future additional changes that will need to be considered:
(1) Make sure that a call doesn't hog the front of the queue by receiving
data from the network as fast as userspace is consuming it to the
exclusion of other calls.
(2) Transmit delayed ACKs from within recvmsg() when we've consumed
sufficiently more packets to avoid the background work item needing to
run.
Signed-off-by: David Howells <dhowells@redhat.com>
2016-09-08 10:10:12 +00:00
|
|
|
/*
|
2017-04-06 09:12:00 +00:00
|
|
|
* Process an ABORT packet directed at a call.
|
rxrpc: Rewrite the data and ack handling code
Rewrite the data and ack handling code such that:
(1) Parsing of received ACK and ABORT packets and the distribution and the
filing of DATA packets happens entirely within the data_ready context
called from the UDP socket. This allows us to process and discard ACK
and ABORT packets much more quickly (they're no longer stashed on a
queue for a background thread to process).
(2) We avoid calling skb_clone(), pskb_pull() and pskb_trim(). We instead
keep track of the offset and length of the content of each packet in
the sk_buff metadata. This means we don't do any allocation in the
receive path.
(3) Jumbo DATA packet parsing is now done in data_ready context. Rather
than cloning the packet once for each subpacket and pulling/trimming
it, we file the packet multiple times with an annotation for each
indicating which subpacket is there. From that we can directly
calculate the offset and length.
(4) A call's receive queue can be accessed without taking locks (memory
barriers do have to be used, though).
(5) Incoming calls are set up from preallocated resources and immediately
made live. They can than have packets queued upon them and ACKs
generated. If insufficient resources exist, DATA packet #1 is given a
BUSY reply and other DATA packets are discarded).
(6) sk_buffs no longer take a ref on their parent call.
To make this work, the following changes are made:
(1) Each call's receive buffer is now a circular buffer of sk_buff
pointers (rxtx_buffer) rather than a number of sk_buff_heads spread
between the call and the socket. This permits each sk_buff to be in
the buffer multiple times. The receive buffer is reused for the
transmit buffer.
(2) A circular buffer of annotations (rxtx_annotations) is kept parallel
to the data buffer. Transmission phase annotations indicate whether a
buffered packet has been ACK'd or not and whether it needs
retransmission.
Receive phase annotations indicate whether a slot holds a whole packet
or a jumbo subpacket and, if the latter, which subpacket. They also
note whether the packet has been decrypted in place.
(3) DATA packet window tracking is much simplified. Each phase has just
two numbers representing the window (rx_hard_ack/rx_top and
tx_hard_ack/tx_top).
The hard_ack number is the sequence number before base of the window,
representing the last packet the other side says it has consumed.
hard_ack starts from 0 and the first packet is sequence number 1.
The top number is the sequence number of the highest-numbered packet
residing in the buffer. Packets between hard_ack+1 and top are
soft-ACK'd to indicate they've been received, but not yet consumed.
Four macros, before(), before_eq(), after() and after_eq() are added
to compare sequence numbers within the window. This allows for the
top of the window to wrap when the hard-ack sequence number gets close
to the limit.
Two flags, RXRPC_CALL_RX_LAST and RXRPC_CALL_TX_LAST, are added also
to indicate when rx_top and tx_top point at the packets with the
LAST_PACKET bit set, indicating the end of the phase.
(4) Calls are queued on the socket 'receive queue' rather than packets.
This means that we don't need have to invent dummy packets to queue to
indicate abnormal/terminal states and we don't have to keep metadata
packets (such as ABORTs) around
(5) The offset and length of a (sub)packet's content are now passed to
the verify_packet security op. This is currently expected to decrypt
the packet in place and validate it.
However, there's now nowhere to store the revised offset and length of
the actual data within the decrypted blob (there may be a header and
padding to skip) because an sk_buff may represent multiple packets, so
a locate_data security op is added to retrieve these details from the
sk_buff content when needed.
(6) recvmsg() now has to handle jumbo subpackets, where each subpacket is
individually secured and needs to be individually decrypted. The code
to do this is broken out into rxrpc_recvmsg_data() and shared with the
kernel API. It now iterates over the call's receive buffer rather
than walking the socket receive queue.
Additional changes:
(1) The timers are condensed to a single timer that is set for the soonest
of three timeouts (delayed ACK generation, DATA retransmission and
call lifespan).
(2) Transmission of ACK and ABORT packets is effected immediately from
process-context socket ops/kernel API calls that cause them instead of
them being punted off to a background work item. The data_ready
handler still has to defer to the background, though.
(3) A shutdown op is added to the AF_RXRPC socket so that the AFS
filesystem can shut down the socket and flush its own work items
before closing the socket to deal with any in-progress service calls.
Future additional changes that will need to be considered:
(1) Make sure that a call doesn't hog the front of the queue by receiving
data from the network as fast as userspace is consuming it to the
exclusion of other calls.
(2) Transmit delayed ACKs from within recvmsg() when we've consumed
sufficiently more packets to avoid the background work item needing to
run.
Signed-off-by: David Howells <dhowells@redhat.com>
2016-09-08 10:10:12 +00:00
|
|
|
*/
|
|
|
|
|
static void rxrpc_input_abort(struct rxrpc_call *call, struct sk_buff *skb)
|
|
|
|
|
{
|
|
|
|
|
struct rxrpc_skb_priv *sp = rxrpc_skb(skb);
|
2007-04-26 22:48:28 +00:00
|
|
|
|
2022-11-14 12:21:32 +00:00
|
|
|
trace_rxrpc_rx_abort(call, sp->hdr.serial, skb->priority);
|
2017-04-06 09:12:00 +00:00
|
|
|
|
2020-06-03 21:21:16 +00:00
|
|
|
rxrpc_set_call_completion(call, RXRPC_CALL_REMOTELY_ABORTED,
|
2022-11-14 12:21:32 +00:00
|
|
|
skb->priority, -ECONNABORTED);
|
2007-04-26 22:48:28 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
|
|
/*
|
rxrpc: Rewrite the data and ack handling code
Rewrite the data and ack handling code such that:
(1) Parsing of received ACK and ABORT packets and the distribution and the
filing of DATA packets happens entirely within the data_ready context
called from the UDP socket. This allows us to process and discard ACK
and ABORT packets much more quickly (they're no longer stashed on a
queue for a background thread to process).
(2) We avoid calling skb_clone(), pskb_pull() and pskb_trim(). We instead
keep track of the offset and length of the content of each packet in
the sk_buff metadata. This means we don't do any allocation in the
receive path.
(3) Jumbo DATA packet parsing is now done in data_ready context. Rather
than cloning the packet once for each subpacket and pulling/trimming
it, we file the packet multiple times with an annotation for each
indicating which subpacket is there. From that we can directly
calculate the offset and length.
(4) A call's receive queue can be accessed without taking locks (memory
barriers do have to be used, though).
(5) Incoming calls are set up from preallocated resources and immediately
made live. They can than have packets queued upon them and ACKs
generated. If insufficient resources exist, DATA packet #1 is given a
BUSY reply and other DATA packets are discarded).
(6) sk_buffs no longer take a ref on their parent call.
To make this work, the following changes are made:
(1) Each call's receive buffer is now a circular buffer of sk_buff
pointers (rxtx_buffer) rather than a number of sk_buff_heads spread
between the call and the socket. This permits each sk_buff to be in
the buffer multiple times. The receive buffer is reused for the
transmit buffer.
(2) A circular buffer of annotations (rxtx_annotations) is kept parallel
to the data buffer. Transmission phase annotations indicate whether a
buffered packet has been ACK'd or not and whether it needs
retransmission.
Receive phase annotations indicate whether a slot holds a whole packet
or a jumbo subpacket and, if the latter, which subpacket. They also
note whether the packet has been decrypted in place.
(3) DATA packet window tracking is much simplified. Each phase has just
two numbers representing the window (rx_hard_ack/rx_top and
tx_hard_ack/tx_top).
The hard_ack number is the sequence number before base of the window,
representing the last packet the other side says it has consumed.
hard_ack starts from 0 and the first packet is sequence number 1.
The top number is the sequence number of the highest-numbered packet
residing in the buffer. Packets between hard_ack+1 and top are
soft-ACK'd to indicate they've been received, but not yet consumed.
Four macros, before(), before_eq(), after() and after_eq() are added
to compare sequence numbers within the window. This allows for the
top of the window to wrap when the hard-ack sequence number gets close
to the limit.
Two flags, RXRPC_CALL_RX_LAST and RXRPC_CALL_TX_LAST, are added also
to indicate when rx_top and tx_top point at the packets with the
LAST_PACKET bit set, indicating the end of the phase.
(4) Calls are queued on the socket 'receive queue' rather than packets.
This means that we don't need have to invent dummy packets to queue to
indicate abnormal/terminal states and we don't have to keep metadata
packets (such as ABORTs) around
(5) The offset and length of a (sub)packet's content are now passed to
the verify_packet security op. This is currently expected to decrypt
the packet in place and validate it.
However, there's now nowhere to store the revised offset and length of
the actual data within the decrypted blob (there may be a header and
padding to skip) because an sk_buff may represent multiple packets, so
a locate_data security op is added to retrieve these details from the
sk_buff content when needed.
(6) recvmsg() now has to handle jumbo subpackets, where each subpacket is
individually secured and needs to be individually decrypted. The code
to do this is broken out into rxrpc_recvmsg_data() and shared with the
kernel API. It now iterates over the call's receive buffer rather
than walking the socket receive queue.
Additional changes:
(1) The timers are condensed to a single timer that is set for the soonest
of three timeouts (delayed ACK generation, DATA retransmission and
call lifespan).
(2) Transmission of ACK and ABORT packets is effected immediately from
process-context socket ops/kernel API calls that cause them instead of
them being punted off to a background work item. The data_ready
handler still has to defer to the background, though.
(3) A shutdown op is added to the AF_RXRPC socket so that the AFS
filesystem can shut down the socket and flush its own work items
before closing the socket to deal with any in-progress service calls.
Future additional changes that will need to be considered:
(1) Make sure that a call doesn't hog the front of the queue by receiving
data from the network as fast as userspace is consuming it to the
exclusion of other calls.
(2) Transmit delayed ACKs from within recvmsg() when we've consumed
sufficiently more packets to avoid the background work item needing to
run.
Signed-off-by: David Howells <dhowells@redhat.com>
2016-09-08 10:10:12 +00:00
|
|
|
* Process an incoming call packet.
|
2007-04-26 22:48:28 +00:00
|
|
|
*/
|
2020-01-23 13:13:41 +00:00
|
|
|
void rxrpc_input_call_packet(struct rxrpc_call *call, struct sk_buff *skb)
|
2007-04-26 22:48:28 +00:00
|
|
|
{
|
rxrpc: Rewrite the data and ack handling code
Rewrite the data and ack handling code such that:
(1) Parsing of received ACK and ABORT packets and the distribution and the
filing of DATA packets happens entirely within the data_ready context
called from the UDP socket. This allows us to process and discard ACK
and ABORT packets much more quickly (they're no longer stashed on a
queue for a background thread to process).
(2) We avoid calling skb_clone(), pskb_pull() and pskb_trim(). We instead
keep track of the offset and length of the content of each packet in
the sk_buff metadata. This means we don't do any allocation in the
receive path.
(3) Jumbo DATA packet parsing is now done in data_ready context. Rather
than cloning the packet once for each subpacket and pulling/trimming
it, we file the packet multiple times with an annotation for each
indicating which subpacket is there. From that we can directly
calculate the offset and length.
(4) A call's receive queue can be accessed without taking locks (memory
barriers do have to be used, though).
(5) Incoming calls are set up from preallocated resources and immediately
made live. They can than have packets queued upon them and ACKs
generated. If insufficient resources exist, DATA packet #1 is given a
BUSY reply and other DATA packets are discarded).
(6) sk_buffs no longer take a ref on their parent call.
To make this work, the following changes are made:
(1) Each call's receive buffer is now a circular buffer of sk_buff
pointers (rxtx_buffer) rather than a number of sk_buff_heads spread
between the call and the socket. This permits each sk_buff to be in
the buffer multiple times. The receive buffer is reused for the
transmit buffer.
(2) A circular buffer of annotations (rxtx_annotations) is kept parallel
to the data buffer. Transmission phase annotations indicate whether a
buffered packet has been ACK'd or not and whether it needs
retransmission.
Receive phase annotations indicate whether a slot holds a whole packet
or a jumbo subpacket and, if the latter, which subpacket. They also
note whether the packet has been decrypted in place.
(3) DATA packet window tracking is much simplified. Each phase has just
two numbers representing the window (rx_hard_ack/rx_top and
tx_hard_ack/tx_top).
The hard_ack number is the sequence number before base of the window,
representing the last packet the other side says it has consumed.
hard_ack starts from 0 and the first packet is sequence number 1.
The top number is the sequence number of the highest-numbered packet
residing in the buffer. Packets between hard_ack+1 and top are
soft-ACK'd to indicate they've been received, but not yet consumed.
Four macros, before(), before_eq(), after() and after_eq() are added
to compare sequence numbers within the window. This allows for the
top of the window to wrap when the hard-ack sequence number gets close
to the limit.
Two flags, RXRPC_CALL_RX_LAST and RXRPC_CALL_TX_LAST, are added also
to indicate when rx_top and tx_top point at the packets with the
LAST_PACKET bit set, indicating the end of the phase.
(4) Calls are queued on the socket 'receive queue' rather than packets.
This means that we don't need have to invent dummy packets to queue to
indicate abnormal/terminal states and we don't have to keep metadata
packets (such as ABORTs) around
(5) The offset and length of a (sub)packet's content are now passed to
the verify_packet security op. This is currently expected to decrypt
the packet in place and validate it.
However, there's now nowhere to store the revised offset and length of
the actual data within the decrypted blob (there may be a header and
padding to skip) because an sk_buff may represent multiple packets, so
a locate_data security op is added to retrieve these details from the
sk_buff content when needed.
(6) recvmsg() now has to handle jumbo subpackets, where each subpacket is
individually secured and needs to be individually decrypted. The code
to do this is broken out into rxrpc_recvmsg_data() and shared with the
kernel API. It now iterates over the call's receive buffer rather
than walking the socket receive queue.
Additional changes:
(1) The timers are condensed to a single timer that is set for the soonest
of three timeouts (delayed ACK generation, DATA retransmission and
call lifespan).
(2) Transmission of ACK and ABORT packets is effected immediately from
process-context socket ops/kernel API calls that cause them instead of
them being punted off to a background work item. The data_ready
handler still has to defer to the background, though.
(3) A shutdown op is added to the AF_RXRPC socket so that the AFS
filesystem can shut down the socket and flush its own work items
before closing the socket to deal with any in-progress service calls.
Future additional changes that will need to be considered:
(1) Make sure that a call doesn't hog the front of the queue by receiving
data from the network as fast as userspace is consuming it to the
exclusion of other calls.
(2) Transmit delayed ACKs from within recvmsg() when we've consumed
sufficiently more packets to avoid the background work item needing to
run.
Signed-off-by: David Howells <dhowells@redhat.com>
2016-09-08 10:10:12 +00:00
|
|
|
struct rxrpc_skb_priv *sp = rxrpc_skb(skb);
|
2017-11-24 10:18:41 +00:00
|
|
|
unsigned long timo;
|
2007-04-26 22:48:28 +00:00
|
|
|
|
2014-03-03 23:04:45 +00:00
|
|
|
_enter("%p,%p", call, skb);
|
2007-04-26 22:48:28 +00:00
|
|
|
|
2020-01-23 13:13:41 +00:00
|
|
|
if (sp->hdr.serviceId != call->dest_srx.srx_service)
|
|
|
|
|
call->dest_srx.srx_service = sp->hdr.serviceId;
|
|
|
|
|
if ((int)sp->hdr.serial - (int)call->rx_serial > 0)
|
|
|
|
|
call->rx_serial = sp->hdr.serial;
|
|
|
|
|
if (!test_bit(RXRPC_CALL_RX_HEARD, &call->flags))
|
|
|
|
|
set_bit(RXRPC_CALL_RX_HEARD, &call->flags);
|
|
|
|
|
|
2017-11-24 10:18:41 +00:00
|
|
|
timo = READ_ONCE(call->next_rx_timo);
|
|
|
|
|
if (timo) {
|
2024-01-30 21:37:16 +00:00
|
|
|
ktime_t delay = ms_to_ktime(timo);
|
2017-11-24 10:18:41 +00:00
|
|
|
|
2024-01-30 21:37:16 +00:00
|
|
|
call->expect_rx_by = ktime_add(ktime_get_real(), delay);
|
|
|
|
|
trace_rxrpc_timer_set(call, delay, rxrpc_timer_trace_expect_rx);
|
2017-11-24 10:18:41 +00:00
|
|
|
}
|
2017-11-29 14:25:50 +00:00
|
|
|
|
rxrpc: Rewrite the data and ack handling code
Rewrite the data and ack handling code such that:
(1) Parsing of received ACK and ABORT packets and the distribution and the
filing of DATA packets happens entirely within the data_ready context
called from the UDP socket. This allows us to process and discard ACK
and ABORT packets much more quickly (they're no longer stashed on a
queue for a background thread to process).
(2) We avoid calling skb_clone(), pskb_pull() and pskb_trim(). We instead
keep track of the offset and length of the content of each packet in
the sk_buff metadata. This means we don't do any allocation in the
receive path.
(3) Jumbo DATA packet parsing is now done in data_ready context. Rather
than cloning the packet once for each subpacket and pulling/trimming
it, we file the packet multiple times with an annotation for each
indicating which subpacket is there. From that we can directly
calculate the offset and length.
(4) A call's receive queue can be accessed without taking locks (memory
barriers do have to be used, though).
(5) Incoming calls are set up from preallocated resources and immediately
made live. They can than have packets queued upon them and ACKs
generated. If insufficient resources exist, DATA packet #1 is given a
BUSY reply and other DATA packets are discarded).
(6) sk_buffs no longer take a ref on their parent call.
To make this work, the following changes are made:
(1) Each call's receive buffer is now a circular buffer of sk_buff
pointers (rxtx_buffer) rather than a number of sk_buff_heads spread
between the call and the socket. This permits each sk_buff to be in
the buffer multiple times. The receive buffer is reused for the
transmit buffer.
(2) A circular buffer of annotations (rxtx_annotations) is kept parallel
to the data buffer. Transmission phase annotations indicate whether a
buffered packet has been ACK'd or not and whether it needs
retransmission.
Receive phase annotations indicate whether a slot holds a whole packet
or a jumbo subpacket and, if the latter, which subpacket. They also
note whether the packet has been decrypted in place.
(3) DATA packet window tracking is much simplified. Each phase has just
two numbers representing the window (rx_hard_ack/rx_top and
tx_hard_ack/tx_top).
The hard_ack number is the sequence number before base of the window,
representing the last packet the other side says it has consumed.
hard_ack starts from 0 and the first packet is sequence number 1.
The top number is the sequence number of the highest-numbered packet
residing in the buffer. Packets between hard_ack+1 and top are
soft-ACK'd to indicate they've been received, but not yet consumed.
Four macros, before(), before_eq(), after() and after_eq() are added
to compare sequence numbers within the window. This allows for the
top of the window to wrap when the hard-ack sequence number gets close
to the limit.
Two flags, RXRPC_CALL_RX_LAST and RXRPC_CALL_TX_LAST, are added also
to indicate when rx_top and tx_top point at the packets with the
LAST_PACKET bit set, indicating the end of the phase.
(4) Calls are queued on the socket 'receive queue' rather than packets.
This means that we don't need have to invent dummy packets to queue to
indicate abnormal/terminal states and we don't have to keep metadata
packets (such as ABORTs) around
(5) The offset and length of a (sub)packet's content are now passed to
the verify_packet security op. This is currently expected to decrypt
the packet in place and validate it.
However, there's now nowhere to store the revised offset and length of
the actual data within the decrypted blob (there may be a header and
padding to skip) because an sk_buff may represent multiple packets, so
a locate_data security op is added to retrieve these details from the
sk_buff content when needed.
(6) recvmsg() now has to handle jumbo subpackets, where each subpacket is
individually secured and needs to be individually decrypted. The code
to do this is broken out into rxrpc_recvmsg_data() and shared with the
kernel API. It now iterates over the call's receive buffer rather
than walking the socket receive queue.
Additional changes:
(1) The timers are condensed to a single timer that is set for the soonest
of three timeouts (delayed ACK generation, DATA retransmission and
call lifespan).
(2) Transmission of ACK and ABORT packets is effected immediately from
process-context socket ops/kernel API calls that cause them instead of
them being punted off to a background work item. The data_ready
handler still has to defer to the background, though.
(3) A shutdown op is added to the AF_RXRPC socket so that the AFS
filesystem can shut down the socket and flush its own work items
before closing the socket to deal with any in-progress service calls.
Future additional changes that will need to be considered:
(1) Make sure that a call doesn't hog the front of the queue by receiving
data from the network as fast as userspace is consuming it to the
exclusion of other calls.
(2) Transmit delayed ACKs from within recvmsg() when we've consumed
sufficiently more packets to avoid the background work item needing to
run.
Signed-off-by: David Howells <dhowells@redhat.com>
2016-09-08 10:10:12 +00:00
|
|
|
switch (sp->hdr.type) {
|
|
|
|
|
case RXRPC_PACKET_TYPE_DATA:
|
2022-10-06 20:45:42 +00:00
|
|
|
return rxrpc_input_data(call, skb);
|
2016-08-30 08:49:28 +00:00
|
|
|
|
rxrpc: Rewrite the data and ack handling code
Rewrite the data and ack handling code such that:
(1) Parsing of received ACK and ABORT packets and the distribution and the
filing of DATA packets happens entirely within the data_ready context
called from the UDP socket. This allows us to process and discard ACK
and ABORT packets much more quickly (they're no longer stashed on a
queue for a background thread to process).
(2) We avoid calling skb_clone(), pskb_pull() and pskb_trim(). We instead
keep track of the offset and length of the content of each packet in
the sk_buff metadata. This means we don't do any allocation in the
receive path.
(3) Jumbo DATA packet parsing is now done in data_ready context. Rather
than cloning the packet once for each subpacket and pulling/trimming
it, we file the packet multiple times with an annotation for each
indicating which subpacket is there. From that we can directly
calculate the offset and length.
(4) A call's receive queue can be accessed without taking locks (memory
barriers do have to be used, though).
(5) Incoming calls are set up from preallocated resources and immediately
made live. They can than have packets queued upon them and ACKs
generated. If insufficient resources exist, DATA packet #1 is given a
BUSY reply and other DATA packets are discarded).
(6) sk_buffs no longer take a ref on their parent call.
To make this work, the following changes are made:
(1) Each call's receive buffer is now a circular buffer of sk_buff
pointers (rxtx_buffer) rather than a number of sk_buff_heads spread
between the call and the socket. This permits each sk_buff to be in
the buffer multiple times. The receive buffer is reused for the
transmit buffer.
(2) A circular buffer of annotations (rxtx_annotations) is kept parallel
to the data buffer. Transmission phase annotations indicate whether a
buffered packet has been ACK'd or not and whether it needs
retransmission.
Receive phase annotations indicate whether a slot holds a whole packet
or a jumbo subpacket and, if the latter, which subpacket. They also
note whether the packet has been decrypted in place.
(3) DATA packet window tracking is much simplified. Each phase has just
two numbers representing the window (rx_hard_ack/rx_top and
tx_hard_ack/tx_top).
The hard_ack number is the sequence number before base of the window,
representing the last packet the other side says it has consumed.
hard_ack starts from 0 and the first packet is sequence number 1.
The top number is the sequence number of the highest-numbered packet
residing in the buffer. Packets between hard_ack+1 and top are
soft-ACK'd to indicate they've been received, but not yet consumed.
Four macros, before(), before_eq(), after() and after_eq() are added
to compare sequence numbers within the window. This allows for the
top of the window to wrap when the hard-ack sequence number gets close
to the limit.
Two flags, RXRPC_CALL_RX_LAST and RXRPC_CALL_TX_LAST, are added also
to indicate when rx_top and tx_top point at the packets with the
LAST_PACKET bit set, indicating the end of the phase.
(4) Calls are queued on the socket 'receive queue' rather than packets.
This means that we don't need have to invent dummy packets to queue to
indicate abnormal/terminal states and we don't have to keep metadata
packets (such as ABORTs) around
(5) The offset and length of a (sub)packet's content are now passed to
the verify_packet security op. This is currently expected to decrypt
the packet in place and validate it.
However, there's now nowhere to store the revised offset and length of
the actual data within the decrypted blob (there may be a header and
padding to skip) because an sk_buff may represent multiple packets, so
a locate_data security op is added to retrieve these details from the
sk_buff content when needed.
(6) recvmsg() now has to handle jumbo subpackets, where each subpacket is
individually secured and needs to be individually decrypted. The code
to do this is broken out into rxrpc_recvmsg_data() and shared with the
kernel API. It now iterates over the call's receive buffer rather
than walking the socket receive queue.
Additional changes:
(1) The timers are condensed to a single timer that is set for the soonest
of three timeouts (delayed ACK generation, DATA retransmission and
call lifespan).
(2) Transmission of ACK and ABORT packets is effected immediately from
process-context socket ops/kernel API calls that cause them instead of
them being punted off to a background work item. The data_ready
handler still has to defer to the background, though.
(3) A shutdown op is added to the AF_RXRPC socket so that the AFS
filesystem can shut down the socket and flush its own work items
before closing the socket to deal with any in-progress service calls.
Future additional changes that will need to be considered:
(1) Make sure that a call doesn't hog the front of the queue by receiving
data from the network as fast as userspace is consuming it to the
exclusion of other calls.
(2) Transmit delayed ACKs from within recvmsg() when we've consumed
sufficiently more packets to avoid the background work item needing to
run.
Signed-off-by: David Howells <dhowells@redhat.com>
2016-09-08 10:10:12 +00:00
|
|
|
case RXRPC_PACKET_TYPE_ACK:
|
2022-10-06 20:45:42 +00:00
|
|
|
return rxrpc_input_ack(call, skb);
|
2007-04-26 22:48:28 +00:00
|
|
|
|
rxrpc: Rewrite the data and ack handling code
Rewrite the data and ack handling code such that:
(1) Parsing of received ACK and ABORT packets and the distribution and the
filing of DATA packets happens entirely within the data_ready context
called from the UDP socket. This allows us to process and discard ACK
and ABORT packets much more quickly (they're no longer stashed on a
queue for a background thread to process).
(2) We avoid calling skb_clone(), pskb_pull() and pskb_trim(). We instead
keep track of the offset and length of the content of each packet in
the sk_buff metadata. This means we don't do any allocation in the
receive path.
(3) Jumbo DATA packet parsing is now done in data_ready context. Rather
than cloning the packet once for each subpacket and pulling/trimming
it, we file the packet multiple times with an annotation for each
indicating which subpacket is there. From that we can directly
calculate the offset and length.
(4) A call's receive queue can be accessed without taking locks (memory
barriers do have to be used, though).
(5) Incoming calls are set up from preallocated resources and immediately
made live. They can than have packets queued upon them and ACKs
generated. If insufficient resources exist, DATA packet #1 is given a
BUSY reply and other DATA packets are discarded).
(6) sk_buffs no longer take a ref on their parent call.
To make this work, the following changes are made:
(1) Each call's receive buffer is now a circular buffer of sk_buff
pointers (rxtx_buffer) rather than a number of sk_buff_heads spread
between the call and the socket. This permits each sk_buff to be in
the buffer multiple times. The receive buffer is reused for the
transmit buffer.
(2) A circular buffer of annotations (rxtx_annotations) is kept parallel
to the data buffer. Transmission phase annotations indicate whether a
buffered packet has been ACK'd or not and whether it needs
retransmission.
Receive phase annotations indicate whether a slot holds a whole packet
or a jumbo subpacket and, if the latter, which subpacket. They also
note whether the packet has been decrypted in place.
(3) DATA packet window tracking is much simplified. Each phase has just
two numbers representing the window (rx_hard_ack/rx_top and
tx_hard_ack/tx_top).
The hard_ack number is the sequence number before base of the window,
representing the last packet the other side says it has consumed.
hard_ack starts from 0 and the first packet is sequence number 1.
The top number is the sequence number of the highest-numbered packet
residing in the buffer. Packets between hard_ack+1 and top are
soft-ACK'd to indicate they've been received, but not yet consumed.
Four macros, before(), before_eq(), after() and after_eq() are added
to compare sequence numbers within the window. This allows for the
top of the window to wrap when the hard-ack sequence number gets close
to the limit.
Two flags, RXRPC_CALL_RX_LAST and RXRPC_CALL_TX_LAST, are added also
to indicate when rx_top and tx_top point at the packets with the
LAST_PACKET bit set, indicating the end of the phase.
(4) Calls are queued on the socket 'receive queue' rather than packets.
This means that we don't need have to invent dummy packets to queue to
indicate abnormal/terminal states and we don't have to keep metadata
packets (such as ABORTs) around
(5) The offset and length of a (sub)packet's content are now passed to
the verify_packet security op. This is currently expected to decrypt
the packet in place and validate it.
However, there's now nowhere to store the revised offset and length of
the actual data within the decrypted blob (there may be a header and
padding to skip) because an sk_buff may represent multiple packets, so
a locate_data security op is added to retrieve these details from the
sk_buff content when needed.
(6) recvmsg() now has to handle jumbo subpackets, where each subpacket is
individually secured and needs to be individually decrypted. The code
to do this is broken out into rxrpc_recvmsg_data() and shared with the
kernel API. It now iterates over the call's receive buffer rather
than walking the socket receive queue.
Additional changes:
(1) The timers are condensed to a single timer that is set for the soonest
of three timeouts (delayed ACK generation, DATA retransmission and
call lifespan).
(2) Transmission of ACK and ABORT packets is effected immediately from
process-context socket ops/kernel API calls that cause them instead of
them being punted off to a background work item. The data_ready
handler still has to defer to the background, though.
(3) A shutdown op is added to the AF_RXRPC socket so that the AFS
filesystem can shut down the socket and flush its own work items
before closing the socket to deal with any in-progress service calls.
Future additional changes that will need to be considered:
(1) Make sure that a call doesn't hog the front of the queue by receiving
data from the network as fast as userspace is consuming it to the
exclusion of other calls.
(2) Transmit delayed ACKs from within recvmsg() when we've consumed
sufficiently more packets to avoid the background work item needing to
run.
Signed-off-by: David Howells <dhowells@redhat.com>
2016-09-08 10:10:12 +00:00
|
|
|
case RXRPC_PACKET_TYPE_BUSY:
|
|
|
|
|
/* Just ignore BUSY packets from the server; the retry and
|
|
|
|
|
* lifespan timers will take care of business. BUSY packets
|
|
|
|
|
* from the client don't make sense.
|
|
|
|
|
*/
|
2022-10-06 20:45:42 +00:00
|
|
|
return;
|
2007-04-26 22:48:28 +00:00
|
|
|
|
rxrpc: Rewrite the data and ack handling code
Rewrite the data and ack handling code such that:
(1) Parsing of received ACK and ABORT packets and the distribution and the
filing of DATA packets happens entirely within the data_ready context
called from the UDP socket. This allows us to process and discard ACK
and ABORT packets much more quickly (they're no longer stashed on a
queue for a background thread to process).
(2) We avoid calling skb_clone(), pskb_pull() and pskb_trim(). We instead
keep track of the offset and length of the content of each packet in
the sk_buff metadata. This means we don't do any allocation in the
receive path.
(3) Jumbo DATA packet parsing is now done in data_ready context. Rather
than cloning the packet once for each subpacket and pulling/trimming
it, we file the packet multiple times with an annotation for each
indicating which subpacket is there. From that we can directly
calculate the offset and length.
(4) A call's receive queue can be accessed without taking locks (memory
barriers do have to be used, though).
(5) Incoming calls are set up from preallocated resources and immediately
made live. They can than have packets queued upon them and ACKs
generated. If insufficient resources exist, DATA packet #1 is given a
BUSY reply and other DATA packets are discarded).
(6) sk_buffs no longer take a ref on their parent call.
To make this work, the following changes are made:
(1) Each call's receive buffer is now a circular buffer of sk_buff
pointers (rxtx_buffer) rather than a number of sk_buff_heads spread
between the call and the socket. This permits each sk_buff to be in
the buffer multiple times. The receive buffer is reused for the
transmit buffer.
(2) A circular buffer of annotations (rxtx_annotations) is kept parallel
to the data buffer. Transmission phase annotations indicate whether a
buffered packet has been ACK'd or not and whether it needs
retransmission.
Receive phase annotations indicate whether a slot holds a whole packet
or a jumbo subpacket and, if the latter, which subpacket. They also
note whether the packet has been decrypted in place.
(3) DATA packet window tracking is much simplified. Each phase has just
two numbers representing the window (rx_hard_ack/rx_top and
tx_hard_ack/tx_top).
The hard_ack number is the sequence number before base of the window,
representing the last packet the other side says it has consumed.
hard_ack starts from 0 and the first packet is sequence number 1.
The top number is the sequence number of the highest-numbered packet
residing in the buffer. Packets between hard_ack+1 and top are
soft-ACK'd to indicate they've been received, but not yet consumed.
Four macros, before(), before_eq(), after() and after_eq() are added
to compare sequence numbers within the window. This allows for the
top of the window to wrap when the hard-ack sequence number gets close
to the limit.
Two flags, RXRPC_CALL_RX_LAST and RXRPC_CALL_TX_LAST, are added also
to indicate when rx_top and tx_top point at the packets with the
LAST_PACKET bit set, indicating the end of the phase.
(4) Calls are queued on the socket 'receive queue' rather than packets.
This means that we don't need have to invent dummy packets to queue to
indicate abnormal/terminal states and we don't have to keep metadata
packets (such as ABORTs) around
(5) The offset and length of a (sub)packet's content are now passed to
the verify_packet security op. This is currently expected to decrypt
the packet in place and validate it.
However, there's now nowhere to store the revised offset and length of
the actual data within the decrypted blob (there may be a header and
padding to skip) because an sk_buff may represent multiple packets, so
a locate_data security op is added to retrieve these details from the
sk_buff content when needed.
(6) recvmsg() now has to handle jumbo subpackets, where each subpacket is
individually secured and needs to be individually decrypted. The code
to do this is broken out into rxrpc_recvmsg_data() and shared with the
kernel API. It now iterates over the call's receive buffer rather
than walking the socket receive queue.
Additional changes:
(1) The timers are condensed to a single timer that is set for the soonest
of three timeouts (delayed ACK generation, DATA retransmission and
call lifespan).
(2) Transmission of ACK and ABORT packets is effected immediately from
process-context socket ops/kernel API calls that cause them instead of
them being punted off to a background work item. The data_ready
handler still has to defer to the background, though.
(3) A shutdown op is added to the AF_RXRPC socket so that the AFS
filesystem can shut down the socket and flush its own work items
before closing the socket to deal with any in-progress service calls.
Future additional changes that will need to be considered:
(1) Make sure that a call doesn't hog the front of the queue by receiving
data from the network as fast as userspace is consuming it to the
exclusion of other calls.
(2) Transmit delayed ACKs from within recvmsg() when we've consumed
sufficiently more packets to avoid the background work item needing to
run.
Signed-off-by: David Howells <dhowells@redhat.com>
2016-09-08 10:10:12 +00:00
|
|
|
case RXRPC_PACKET_TYPE_ABORT:
|
2022-10-06 20:45:42 +00:00
|
|
|
return rxrpc_input_abort(call, skb);
|
2007-04-26 22:48:28 +00:00
|
|
|
|
rxrpc: Rewrite the data and ack handling code
Rewrite the data and ack handling code such that:
(1) Parsing of received ACK and ABORT packets and the distribution and the
filing of DATA packets happens entirely within the data_ready context
called from the UDP socket. This allows us to process and discard ACK
and ABORT packets much more quickly (they're no longer stashed on a
queue for a background thread to process).
(2) We avoid calling skb_clone(), pskb_pull() and pskb_trim(). We instead
keep track of the offset and length of the content of each packet in
the sk_buff metadata. This means we don't do any allocation in the
receive path.
(3) Jumbo DATA packet parsing is now done in data_ready context. Rather
than cloning the packet once for each subpacket and pulling/trimming
it, we file the packet multiple times with an annotation for each
indicating which subpacket is there. From that we can directly
calculate the offset and length.
(4) A call's receive queue can be accessed without taking locks (memory
barriers do have to be used, though).
(5) Incoming calls are set up from preallocated resources and immediately
made live. They can than have packets queued upon them and ACKs
generated. If insufficient resources exist, DATA packet #1 is given a
BUSY reply and other DATA packets are discarded).
(6) sk_buffs no longer take a ref on their parent call.
To make this work, the following changes are made:
(1) Each call's receive buffer is now a circular buffer of sk_buff
pointers (rxtx_buffer) rather than a number of sk_buff_heads spread
between the call and the socket. This permits each sk_buff to be in
the buffer multiple times. The receive buffer is reused for the
transmit buffer.
(2) A circular buffer of annotations (rxtx_annotations) is kept parallel
to the data buffer. Transmission phase annotations indicate whether a
buffered packet has been ACK'd or not and whether it needs
retransmission.
Receive phase annotations indicate whether a slot holds a whole packet
or a jumbo subpacket and, if the latter, which subpacket. They also
note whether the packet has been decrypted in place.
(3) DATA packet window tracking is much simplified. Each phase has just
two numbers representing the window (rx_hard_ack/rx_top and
tx_hard_ack/tx_top).
The hard_ack number is the sequence number before base of the window,
representing the last packet the other side says it has consumed.
hard_ack starts from 0 and the first packet is sequence number 1.
The top number is the sequence number of the highest-numbered packet
residing in the buffer. Packets between hard_ack+1 and top are
soft-ACK'd to indicate they've been received, but not yet consumed.
Four macros, before(), before_eq(), after() and after_eq() are added
to compare sequence numbers within the window. This allows for the
top of the window to wrap when the hard-ack sequence number gets close
to the limit.
Two flags, RXRPC_CALL_RX_LAST and RXRPC_CALL_TX_LAST, are added also
to indicate when rx_top and tx_top point at the packets with the
LAST_PACKET bit set, indicating the end of the phase.
(4) Calls are queued on the socket 'receive queue' rather than packets.
This means that we don't need have to invent dummy packets to queue to
indicate abnormal/terminal states and we don't have to keep metadata
packets (such as ABORTs) around
(5) The offset and length of a (sub)packet's content are now passed to
the verify_packet security op. This is currently expected to decrypt
the packet in place and validate it.
However, there's now nowhere to store the revised offset and length of
the actual data within the decrypted blob (there may be a header and
padding to skip) because an sk_buff may represent multiple packets, so
a locate_data security op is added to retrieve these details from the
sk_buff content when needed.
(6) recvmsg() now has to handle jumbo subpackets, where each subpacket is
individually secured and needs to be individually decrypted. The code
to do this is broken out into rxrpc_recvmsg_data() and shared with the
kernel API. It now iterates over the call's receive buffer rather
than walking the socket receive queue.
Additional changes:
(1) The timers are condensed to a single timer that is set for the soonest
of three timeouts (delayed ACK generation, DATA retransmission and
call lifespan).
(2) Transmission of ACK and ABORT packets is effected immediately from
process-context socket ops/kernel API calls that cause them instead of
them being punted off to a background work item. The data_ready
handler still has to defer to the background, though.
(3) A shutdown op is added to the AF_RXRPC socket so that the AFS
filesystem can shut down the socket and flush its own work items
before closing the socket to deal with any in-progress service calls.
Future additional changes that will need to be considered:
(1) Make sure that a call doesn't hog the front of the queue by receiving
data from the network as fast as userspace is consuming it to the
exclusion of other calls.
(2) Transmit delayed ACKs from within recvmsg() when we've consumed
sufficiently more packets to avoid the background work item needing to
run.
Signed-off-by: David Howells <dhowells@redhat.com>
2016-09-08 10:10:12 +00:00
|
|
|
case RXRPC_PACKET_TYPE_ACKALL:
|
2022-10-06 20:45:42 +00:00
|
|
|
return rxrpc_input_ackall(call, skb);
|
2016-08-30 08:49:28 +00:00
|
|
|
|
rxrpc: Rewrite the data and ack handling code
Rewrite the data and ack handling code such that:
(1) Parsing of received ACK and ABORT packets and the distribution and the
filing of DATA packets happens entirely within the data_ready context
called from the UDP socket. This allows us to process and discard ACK
and ABORT packets much more quickly (they're no longer stashed on a
queue for a background thread to process).
(2) We avoid calling skb_clone(), pskb_pull() and pskb_trim(). We instead
keep track of the offset and length of the content of each packet in
the sk_buff metadata. This means we don't do any allocation in the
receive path.
(3) Jumbo DATA packet parsing is now done in data_ready context. Rather
than cloning the packet once for each subpacket and pulling/trimming
it, we file the packet multiple times with an annotation for each
indicating which subpacket is there. From that we can directly
calculate the offset and length.
(4) A call's receive queue can be accessed without taking locks (memory
barriers do have to be used, though).
(5) Incoming calls are set up from preallocated resources and immediately
made live. They can than have packets queued upon them and ACKs
generated. If insufficient resources exist, DATA packet #1 is given a
BUSY reply and other DATA packets are discarded).
(6) sk_buffs no longer take a ref on their parent call.
To make this work, the following changes are made:
(1) Each call's receive buffer is now a circular buffer of sk_buff
pointers (rxtx_buffer) rather than a number of sk_buff_heads spread
between the call and the socket. This permits each sk_buff to be in
the buffer multiple times. The receive buffer is reused for the
transmit buffer.
(2) A circular buffer of annotations (rxtx_annotations) is kept parallel
to the data buffer. Transmission phase annotations indicate whether a
buffered packet has been ACK'd or not and whether it needs
retransmission.
Receive phase annotations indicate whether a slot holds a whole packet
or a jumbo subpacket and, if the latter, which subpacket. They also
note whether the packet has been decrypted in place.
(3) DATA packet window tracking is much simplified. Each phase has just
two numbers representing the window (rx_hard_ack/rx_top and
tx_hard_ack/tx_top).
The hard_ack number is the sequence number before base of the window,
representing the last packet the other side says it has consumed.
hard_ack starts from 0 and the first packet is sequence number 1.
The top number is the sequence number of the highest-numbered packet
residing in the buffer. Packets between hard_ack+1 and top are
soft-ACK'd to indicate they've been received, but not yet consumed.
Four macros, before(), before_eq(), after() and after_eq() are added
to compare sequence numbers within the window. This allows for the
top of the window to wrap when the hard-ack sequence number gets close
to the limit.
Two flags, RXRPC_CALL_RX_LAST and RXRPC_CALL_TX_LAST, are added also
to indicate when rx_top and tx_top point at the packets with the
LAST_PACKET bit set, indicating the end of the phase.
(4) Calls are queued on the socket 'receive queue' rather than packets.
This means that we don't need have to invent dummy packets to queue to
indicate abnormal/terminal states and we don't have to keep metadata
packets (such as ABORTs) around
(5) The offset and length of a (sub)packet's content are now passed to
the verify_packet security op. This is currently expected to decrypt
the packet in place and validate it.
However, there's now nowhere to store the revised offset and length of
the actual data within the decrypted blob (there may be a header and
padding to skip) because an sk_buff may represent multiple packets, so
a locate_data security op is added to retrieve these details from the
sk_buff content when needed.
(6) recvmsg() now has to handle jumbo subpackets, where each subpacket is
individually secured and needs to be individually decrypted. The code
to do this is broken out into rxrpc_recvmsg_data() and shared with the
kernel API. It now iterates over the call's receive buffer rather
than walking the socket receive queue.
Additional changes:
(1) The timers are condensed to a single timer that is set for the soonest
of three timeouts (delayed ACK generation, DATA retransmission and
call lifespan).
(2) Transmission of ACK and ABORT packets is effected immediately from
process-context socket ops/kernel API calls that cause them instead of
them being punted off to a background work item. The data_ready
handler still has to defer to the background, though.
(3) A shutdown op is added to the AF_RXRPC socket so that the AFS
filesystem can shut down the socket and flush its own work items
before closing the socket to deal with any in-progress service calls.
Future additional changes that will need to be considered:
(1) Make sure that a call doesn't hog the front of the queue by receiving
data from the network as fast as userspace is consuming it to the
exclusion of other calls.
(2) Transmit delayed ACKs from within recvmsg() when we've consumed
sufficiently more packets to avoid the background work item needing to
run.
Signed-off-by: David Howells <dhowells@redhat.com>
2016-09-08 10:10:12 +00:00
|
|
|
default:
|
|
|
|
|
break;
|
2007-04-26 22:48:28 +00:00
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
|
2016-10-06 07:11:49 +00:00
|
|
|
/*
|
rxrpc: Fix the packet reception routine
The rxrpc_input_packet() function and its call tree was built around the
assumption that data_ready() handler called from UDP to inform a kernel
service that there is data to be had was non-reentrant. This means that
certain locking could be dispensed with.
This, however, turns out not to be the case with a multi-queue network card
that can deliver packets to multiple cpus simultaneously. Each of those
cpus can be in the rxrpc_input_packet() function at the same time.
Fix by adding or changing some structure members:
(1) Add peer->rtt_input_lock to serialise access to the RTT buffer.
(2) Make conn->service_id into a 32-bit variable so that it can be
cmpxchg'd on all arches.
(3) Add call->input_lock to serialise access to the Rx/Tx state. Note
that although the Rx and Tx states are (almost) entirely separate,
there's no point completing the separation and having separate locks
since it's a bi-phasal RPC protocol rather than a bi-direction
streaming protocol. Data transmission and data reception do not take
place simultaneously on any particular call.
and making the following functional changes:
(1) In rxrpc_input_data(), hold call->input_lock around the core to
prevent simultaneous producing of packets into the Rx ring and
updating of tracking state for a particular call.
(2) In rxrpc_input_ping_response(), only read call->ping_serial once, and
check it before checking RXRPC_CALL_PINGING as that's a cheaper test.
The bit test and bit clear can then be combined. No further locking
is needed here.
(3) In rxrpc_input_ack(), take call->input_lock after we've parsed much of
the ACK packet. The superseded ACK check is then done both before and
after the lock is taken.
The handing of ackinfo data is split, parsing before the lock is taken
and processing with it held. This is keyed on rxMTU being non-zero.
Congestion management is also done within the locked section.
(4) In rxrpc_input_ackall(), take call->input_lock around the Tx window
rotation. The ACKALL packet carries no information and is only really
useful after all packets have been transmitted since it's imprecise.
(5) In rxrpc_input_implicit_end_call(), we use rx->incoming_lock to
prevent calls being simultaneously implicitly ended on two cpus and
also to prevent any races with incoming call setup.
(6) In rxrpc_input_packet(), use cmpxchg() to effect the service upgrade
on a connection. It is only permitted to happen once for a
connection.
(7) In rxrpc_new_incoming_call(), we have to recheck the routing inside
rx->incoming_lock to see if someone else set up the call, connection
or peer whilst we were getting there. We can't trust the values from
the earlier routing check unless we pin refs on them - which we want
to avoid.
Further, we need to allow for an incoming call to have its state
changed on another CPU between us making it live and us adjusting it
because the conn is now in the RXRPC_CONN_SERVICE state.
(8) In rxrpc_peer_add_rtt(), take peer->rtt_input_lock around the access
to the RTT buffer. Don't need to lock around setting peer->rtt.
For reference, the inventory of state-accessing or state-altering functions
used by the packet input procedure is:
> rxrpc_input_packet()
* PACKET CHECKING
* ROUTING
> rxrpc_post_packet_to_local()
> rxrpc_find_connection_rcu() - uses RCU
> rxrpc_lookup_peer_rcu() - uses RCU
> rxrpc_find_service_conn_rcu() - uses RCU
> idr_find() - uses RCU
* CONNECTION-LEVEL PROCESSING
- Service upgrade
- Can only happen once per conn
! Changed to use cmpxchg
> rxrpc_post_packet_to_conn()
- Setting conn->hi_serial
- Probably safe not using locks
- Maybe use cmpxchg
* CALL-LEVEL PROCESSING
> Old-call checking
> rxrpc_input_implicit_end_call()
> rxrpc_call_completed()
> rxrpc_queue_call()
! Need to take rx->incoming_lock
> __rxrpc_disconnect_call()
> rxrpc_notify_socket()
> rxrpc_new_incoming_call()
- Uses rx->incoming_lock for the entire process
- Might be able to drop this earlier in favour of the call lock
> rxrpc_incoming_call()
! Conflicts with rxrpc_input_implicit_end_call()
> rxrpc_send_ping()
- Don't need locks to check rtt state
> rxrpc_propose_ACK
* PACKET DISTRIBUTION
> rxrpc_input_call_packet()
> rxrpc_input_data()
* QUEUE DATA PACKET ON CALL
> rxrpc_reduce_call_timer()
- Uses timer_reduce()
! Needs call->input_lock()
> rxrpc_receiving_reply()
! Needs locking around ack state
> rxrpc_rotate_tx_window()
> rxrpc_end_tx_phase()
> rxrpc_proto_abort()
> rxrpc_input_dup_data()
- Fills the Rx buffer
- rxrpc_propose_ACK()
- rxrpc_notify_socket()
> rxrpc_input_ack()
* APPLY ACK PACKET TO CALL AND DISCARD PACKET
> rxrpc_input_ping_response()
- Probably doesn't need any extra locking
! Need READ_ONCE() on call->ping_serial
> rxrpc_input_check_for_lost_ack()
- Takes call->lock to consult Tx buffer
> rxrpc_peer_add_rtt()
! Needs to take a lock (peer->rtt_input_lock)
! Could perhaps manage with cmpxchg() and xadd() instead
> rxrpc_input_requested_ack
- Consults Tx buffer
! Probably needs a lock
> rxrpc_peer_add_rtt()
> rxrpc_propose_ack()
> rxrpc_input_ackinfo()
- Changes call->tx_winsize
! Use cmpxchg to handle change
! Should perhaps track serial number
- Uses peer->lock to record MTU specification changes
> rxrpc_proto_abort()
! Need to take call->input_lock
> rxrpc_rotate_tx_window()
> rxrpc_end_tx_phase()
> rxrpc_input_soft_acks()
- Consults the Tx buffer
> rxrpc_congestion_management()
- Modifies the Tx annotations
! Needs call->input_lock()
> rxrpc_queue_call()
> rxrpc_input_abort()
* APPLY ABORT PACKET TO CALL AND DISCARD PACKET
> rxrpc_set_call_completion()
> rxrpc_notify_socket()
> rxrpc_input_ackall()
* APPLY ACKALL PACKET TO CALL AND DISCARD PACKET
! Need to take call->input_lock
> rxrpc_rotate_tx_window()
> rxrpc_end_tx_phase()
> rxrpc_reject_packet()
There are some functions used by the above that queue the packet, after
which the procedure is terminated:
- rxrpc_post_packet_to_local()
- local->event_queue is an sk_buff_head
- local->processor is a work_struct
- rxrpc_post_packet_to_conn()
- conn->rx_queue is an sk_buff_head
- conn->processor is a work_struct
- rxrpc_reject_packet()
- local->reject_queue is an sk_buff_head
- local->processor is a work_struct
And some that offload processing to process context:
- rxrpc_notify_socket()
- Uses RCU lock
- Uses call->notify_lock to call call->notify_rx
- Uses call->recvmsg_lock to queue recvmsg side
- rxrpc_queue_call()
- call->processor is a work_struct
- rxrpc_propose_ACK()
- Uses call->lock to wrap __rxrpc_propose_ACK()
And a bunch that complete a call, all of which use call->state_lock to
protect the call state:
- rxrpc_call_completed()
- rxrpc_set_call_completion()
- rxrpc_abort_call()
- rxrpc_proto_abort()
- Also uses rxrpc_queue_call()
Fixes: 17926a79320a ("[AF_RXRPC]: Provide secure RxRPC sockets for use by userspace and kernel both")
Signed-off-by: David Howells <dhowells@redhat.com>
2018-10-08 14:46:25 +00:00
|
|
|
* Handle a new service call on a channel implicitly completing the preceding
|
|
|
|
|
* call on that channel. This does not apply to client conns.
|
2016-10-06 07:11:49 +00:00
|
|
|
*
|
|
|
|
|
* TODO: If callNumber > call_id + 1, renegotiate security.
|
|
|
|
|
*/
|
2020-01-23 13:13:41 +00:00
|
|
|
void rxrpc_implicit_end_call(struct rxrpc_call *call, struct sk_buff *skb)
|
2016-10-06 07:11:49 +00:00
|
|
|
{
|
2022-10-27 10:25:55 +00:00
|
|
|
switch (__rxrpc_call_state(call)) {
|
2016-10-06 07:11:49 +00:00
|
|
|
case RXRPC_CALL_SERVER_AWAIT_ACK:
|
|
|
|
|
rxrpc_call_completed(call);
|
2020-08-23 22:36:59 +00:00
|
|
|
fallthrough;
|
2016-10-06 07:11:49 +00:00
|
|
|
case RXRPC_CALL_COMPLETE:
|
|
|
|
|
break;
|
|
|
|
|
default:
|
2022-10-06 20:45:42 +00:00
|
|
|
rxrpc_abort_call(call, 0, RX_CALL_DEAD, -ESHUTDOWN,
|
|
|
|
|
rxrpc_eproto_improper_term);
|
rxrpc: Fix the packet reception routine
The rxrpc_input_packet() function and its call tree was built around the
assumption that data_ready() handler called from UDP to inform a kernel
service that there is data to be had was non-reentrant. This means that
certain locking could be dispensed with.
This, however, turns out not to be the case with a multi-queue network card
that can deliver packets to multiple cpus simultaneously. Each of those
cpus can be in the rxrpc_input_packet() function at the same time.
Fix by adding or changing some structure members:
(1) Add peer->rtt_input_lock to serialise access to the RTT buffer.
(2) Make conn->service_id into a 32-bit variable so that it can be
cmpxchg'd on all arches.
(3) Add call->input_lock to serialise access to the Rx/Tx state. Note
that although the Rx and Tx states are (almost) entirely separate,
there's no point completing the separation and having separate locks
since it's a bi-phasal RPC protocol rather than a bi-direction
streaming protocol. Data transmission and data reception do not take
place simultaneously on any particular call.
and making the following functional changes:
(1) In rxrpc_input_data(), hold call->input_lock around the core to
prevent simultaneous producing of packets into the Rx ring and
updating of tracking state for a particular call.
(2) In rxrpc_input_ping_response(), only read call->ping_serial once, and
check it before checking RXRPC_CALL_PINGING as that's a cheaper test.
The bit test and bit clear can then be combined. No further locking
is needed here.
(3) In rxrpc_input_ack(), take call->input_lock after we've parsed much of
the ACK packet. The superseded ACK check is then done both before and
after the lock is taken.
The handing of ackinfo data is split, parsing before the lock is taken
and processing with it held. This is keyed on rxMTU being non-zero.
Congestion management is also done within the locked section.
(4) In rxrpc_input_ackall(), take call->input_lock around the Tx window
rotation. The ACKALL packet carries no information and is only really
useful after all packets have been transmitted since it's imprecise.
(5) In rxrpc_input_implicit_end_call(), we use rx->incoming_lock to
prevent calls being simultaneously implicitly ended on two cpus and
also to prevent any races with incoming call setup.
(6) In rxrpc_input_packet(), use cmpxchg() to effect the service upgrade
on a connection. It is only permitted to happen once for a
connection.
(7) In rxrpc_new_incoming_call(), we have to recheck the routing inside
rx->incoming_lock to see if someone else set up the call, connection
or peer whilst we were getting there. We can't trust the values from
the earlier routing check unless we pin refs on them - which we want
to avoid.
Further, we need to allow for an incoming call to have its state
changed on another CPU between us making it live and us adjusting it
because the conn is now in the RXRPC_CONN_SERVICE state.
(8) In rxrpc_peer_add_rtt(), take peer->rtt_input_lock around the access
to the RTT buffer. Don't need to lock around setting peer->rtt.
For reference, the inventory of state-accessing or state-altering functions
used by the packet input procedure is:
> rxrpc_input_packet()
* PACKET CHECKING
* ROUTING
> rxrpc_post_packet_to_local()
> rxrpc_find_connection_rcu() - uses RCU
> rxrpc_lookup_peer_rcu() - uses RCU
> rxrpc_find_service_conn_rcu() - uses RCU
> idr_find() - uses RCU
* CONNECTION-LEVEL PROCESSING
- Service upgrade
- Can only happen once per conn
! Changed to use cmpxchg
> rxrpc_post_packet_to_conn()
- Setting conn->hi_serial
- Probably safe not using locks
- Maybe use cmpxchg
* CALL-LEVEL PROCESSING
> Old-call checking
> rxrpc_input_implicit_end_call()
> rxrpc_call_completed()
> rxrpc_queue_call()
! Need to take rx->incoming_lock
> __rxrpc_disconnect_call()
> rxrpc_notify_socket()
> rxrpc_new_incoming_call()
- Uses rx->incoming_lock for the entire process
- Might be able to drop this earlier in favour of the call lock
> rxrpc_incoming_call()
! Conflicts with rxrpc_input_implicit_end_call()
> rxrpc_send_ping()
- Don't need locks to check rtt state
> rxrpc_propose_ACK
* PACKET DISTRIBUTION
> rxrpc_input_call_packet()
> rxrpc_input_data()
* QUEUE DATA PACKET ON CALL
> rxrpc_reduce_call_timer()
- Uses timer_reduce()
! Needs call->input_lock()
> rxrpc_receiving_reply()
! Needs locking around ack state
> rxrpc_rotate_tx_window()
> rxrpc_end_tx_phase()
> rxrpc_proto_abort()
> rxrpc_input_dup_data()
- Fills the Rx buffer
- rxrpc_propose_ACK()
- rxrpc_notify_socket()
> rxrpc_input_ack()
* APPLY ACK PACKET TO CALL AND DISCARD PACKET
> rxrpc_input_ping_response()
- Probably doesn't need any extra locking
! Need READ_ONCE() on call->ping_serial
> rxrpc_input_check_for_lost_ack()
- Takes call->lock to consult Tx buffer
> rxrpc_peer_add_rtt()
! Needs to take a lock (peer->rtt_input_lock)
! Could perhaps manage with cmpxchg() and xadd() instead
> rxrpc_input_requested_ack
- Consults Tx buffer
! Probably needs a lock
> rxrpc_peer_add_rtt()
> rxrpc_propose_ack()
> rxrpc_input_ackinfo()
- Changes call->tx_winsize
! Use cmpxchg to handle change
! Should perhaps track serial number
- Uses peer->lock to record MTU specification changes
> rxrpc_proto_abort()
! Need to take call->input_lock
> rxrpc_rotate_tx_window()
> rxrpc_end_tx_phase()
> rxrpc_input_soft_acks()
- Consults the Tx buffer
> rxrpc_congestion_management()
- Modifies the Tx annotations
! Needs call->input_lock()
> rxrpc_queue_call()
> rxrpc_input_abort()
* APPLY ABORT PACKET TO CALL AND DISCARD PACKET
> rxrpc_set_call_completion()
> rxrpc_notify_socket()
> rxrpc_input_ackall()
* APPLY ACKALL PACKET TO CALL AND DISCARD PACKET
! Need to take call->input_lock
> rxrpc_rotate_tx_window()
> rxrpc_end_tx_phase()
> rxrpc_reject_packet()
There are some functions used by the above that queue the packet, after
which the procedure is terminated:
- rxrpc_post_packet_to_local()
- local->event_queue is an sk_buff_head
- local->processor is a work_struct
- rxrpc_post_packet_to_conn()
- conn->rx_queue is an sk_buff_head
- conn->processor is a work_struct
- rxrpc_reject_packet()
- local->reject_queue is an sk_buff_head
- local->processor is a work_struct
And some that offload processing to process context:
- rxrpc_notify_socket()
- Uses RCU lock
- Uses call->notify_lock to call call->notify_rx
- Uses call->recvmsg_lock to queue recvmsg side
- rxrpc_queue_call()
- call->processor is a work_struct
- rxrpc_propose_ACK()
- Uses call->lock to wrap __rxrpc_propose_ACK()
And a bunch that complete a call, all of which use call->state_lock to
protect the call state:
- rxrpc_call_completed()
- rxrpc_set_call_completion()
- rxrpc_abort_call()
- rxrpc_proto_abort()
- Also uses rxrpc_queue_call()
Fixes: 17926a79320a ("[AF_RXRPC]: Provide secure RxRPC sockets for use by userspace and kernel both")
Signed-off-by: David Howells <dhowells@redhat.com>
2018-10-08 14:46:25 +00:00
|
|
|
trace_rxrpc_improper_term(call);
|
2016-10-06 07:11:49 +00:00
|
|
|
break;
|
|
|
|
|
}
|
|
|
|
|
|
2024-12-04 07:46:41 +00:00
|
|
|
rxrpc_input_call_event(call);
|
2016-10-06 07:11:49 +00:00
|
|
|
}
|