Testing xmit_more support with netperf and connected UDP sockets,
I found strange dst refcount false sharing.
Current handling of IFF_XMIT_DST_RELEASE is not optimal.
Dropping dst in validate_xmit_skb() is certainly too late in case
packet was queued by cpu X but dequeued by cpu Y
The logical point to take care of drop/force is in __dev_queue_xmit()
before even taking qdisc lock.
As Julian Anastasov pointed out, need for skb_dst() might come from some
packet schedulers or classifiers.
This patch adds new helper to cleanly express needs of various drivers
or qdiscs/classifiers.
Drivers that need skb_dst() in their ndo_start_xmit() should call
following helper in their setup instead of the prior :
dev->priv_flags &= ~IFF_XMIT_DST_RELEASE;
->
netif_keep_dst(dev);
Instead of using a single bit, we use two bits, one being
eventually rebuilt in bonding/team drivers.
The other one, is permanent and blocks IFF_XMIT_DST_RELEASE being
rebuilt in bonding/team. Eventually, we could add something
smarter later.
Signed-off-by: Eric Dumazet <edumazet@google.com>
Cc: Julian Anastasov <ja@ssi.bg>
Signed-off-by: David S. Miller <davem@davemloft.net>
Probably not a big deal, but we'd better just use the
one we get in retry loop.
Fixes: commit 22e0f8b9322cb1a48b1357e8 ("net: sched: make bstats per cpu and estimator RCU safe")
Reported-by: Joe Perches <joe@perches.com>
Cc: John Fastabend <john.fastabend@gmail.com>
Signed-off-by: Cong Wang <xiyou.wangcong@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Fix a openvswitch compilation error when CONFIG_INET is not set:
=====================================================
In file included from include/net/geneve.h:4:0,
from net/openvswitch/flow_netlink.c:45:
include/net/udp_tunnel.h: In function 'udp_tunnel_handle_offloads':
>> include/net/udp_tunnel.h💯2: error: implicit declaration of function 'iptunnel_handle_offloads' [-Werror=implicit-function-declaration]
>> return iptunnel_handle_offloads(skb, udp_csum, type);
>> ^
>> >> include/net/udp_tunnel.h💯2: warning: return makes pointer from integer without a cast
>> >> cc1: some warnings being treated as errors
=====================================================
Reported-by: kbuild test robot <fengguang.wu@intel.com>
Signed-off-by: Andy Zhou <azhou@nicira.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Fix a sparse warning introduced by commit:
f5796684069e0c71c65bce6a6d4766114aec1396 (openvswitch: Add support for
Geneve tunneling.) caught by kbuild test robot:
reproduce:
# apt-get install sparse
# git checkout f5796684069e0c71c65bce6a6d4766114aec1396
# make ARCH=x86_64 allmodconfig
# make C=1 CF=-D__CHECK_ENDIAN__
#
#
# sparse warnings: (new ones prefixed by >>)
#
# >> net/openvswitch/vport-geneve.c:109:15: sparse: incorrect type in assignment (different base types)
# net/openvswitch/vport-geneve.c:109:15: expected restricted __be16 [usertype] sport
# net/openvswitch/vport-geneve.c:109:15: got int
# >> net/openvswitch/vport-geneve.c:110:56: sparse: incorrect type in argument 3 (different base types)
# net/openvswitch/vport-geneve.c:110:56: expected unsigned short [unsigned] [usertype] value
# net/openvswitch/vport-geneve.c:110:56: got restricted __be16 [usertype] sport
Reported-by: kbuild test robot <fengguang.wu@intel.com>
Signed-off-by: Andy Zhou <azhou@nicira.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Cc: YOSHIFUJI Hideaki <hideaki@yoshifuji.org>
Cc: Martin Lau <kafai@fb.com>
Signed-off-by: Hannes Frederic Sowa <hannes@stressinduktion.org>
Signed-off-by: David S. Miller <davem@davemloft.net>
Try to reduce number of possible fn_sernum mutation by constraining them
to their namespace.
Also remove rt_genid which I forgot to remove in 705f1c869d577c ("ipv6:
remove rt6i_genid").
Cc: YOSHIFUJI Hideaki <hideaki@yoshifuji.org>
Cc: Martin Lau <kafai@fb.com>
Signed-off-by: Hannes Frederic Sowa <hannes@stressinduktion.org>
Signed-off-by: David S. Miller <davem@davemloft.net>
Cc: YOSHIFUJI Hideaki <hideaki@yoshifuji.org>
Cc: Martin Lau <kafai@fb.com>
Signed-off-by: Hannes Frederic Sowa <hannes@stressinduktion.org>
Signed-off-by: David S. Miller <davem@davemloft.net>
Cc: YOSHIFUJI Hideaki <hideaki@yoshifuji.org>
Cc: Martin Lau <kafai@fb.com>
Signed-off-by: Hannes Frederic Sowa <hannes@stressinduktion.org>
Signed-off-by: David S. Miller <davem@davemloft.net>
Also renamed struct fib6_walker_t to fib6_walker and enum fib_walk_state_t
to fib6_walk_state as recommended by Cong Wang.
Cc: Cong Wang <cwang@twopensource.com>
Cc: YOSHIFUJI Hideaki <hideaki@yoshifuji.org>
Cc: Martin Lau <kafai@fb.com>
Signed-off-by: Hannes Frederic Sowa <hannes@stressinduktion.org>
Signed-off-by: David S. Miller <davem@davemloft.net>
Marking this as static allows compiler to inline it.
Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Using the tcf_proto pointer 'tp' from inside the classifiers callback
is not valid because it may have been cleaned up by another call_rcu
occuring on another CPU.
'tp' is currently being used by tcf_unbind_filter() in this patch we
move instances of tcf_unbind_filter outside of the call_rcu() context.
This is safe to do because any running schedulers will either read the
valid class field or it will be zeroed.
And all schedulers today when the class is 0 do a lookup using the
same call used by the tcf_exts_bind(). So even if we have a running
classifier hit the null class pointer it will do a lookup and get
to the same result. This is particularly fragile at the moment because
the only way to verify this is to audit the schedulers call sites.
Reported-by: Cong Wang <xiyou.wangconf@gmail.com>
Signed-off-by: John Fastabend <john.r.fastabend@intel.com>
Acked-by: Cong Wang <cwang@twopensource.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
It is not RCU safe to destroy the action chain while there
is a possibility of readers accessing it. Move this code
into the rcu callback using the same rcu callback used in the
code patch to make a change to head.
Signed-off-by: John Fastabend <john.r.fastabend@intel.com>
Acked-by: Cong Wang <cwang@twopensource.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
This removes the tcf_proto argument from the ematch code paths that
only need it to reference the net namespace. This allows simplifying
qdisc code paths especially when we need to tear down the ematch
from an RCU callback. In this case we can not guarentee that the
tcf_proto structure is still valid.
Signed-off-by: John Fastabend <john.r.fastabend@intel.com>
Acked-by: Cong Wang <cwang@twopensource.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Some TSO engines might have a too heavy setup cost, that impacts
performance on hosts sending small bursts (2 MSS per packet).
This patch adds a device gso_min_segs, allowing drivers to set
a minimum segment size for TSO packets, according to the NIC
performance.
Tested on a mlx4 NIC, this allows to get a ~110% increase of
throughput when sending 2 MSS per packet.
Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
In case we find a general query with non-zero number of sources, we
are dropping the skb as it's malformed.
RFC3376, section 4.1.8. Number of Sources (N):
This number is zero in a General Query or a Group-Specific Query,
and non-zero in a Group-and-Source-Specific Query.
Therefore, reflect that by using kfree_skb() instead of consume_skb().
Fixes: d679c5324d9a ("igmp: avoid drop_monitor false positives")
Signed-off-by: Daniel Borkmann <dborkman@redhat.com>
Acked-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Use new ethtool [sg]et_tunable() to set tx_copybread (inline threshold)
Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: Amir Vadai <amirv@mellanox.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Standard qdisc API to setup a timer implies an atomic operation on every
packet dequeue : qdisc_unthrottled()
It turns out this is not really needed for FQ, as FQ has no concept of
global qdisc throttling, being a qdisc handling many different flows,
some of them can be throttled, while others are not.
Fix is straightforward : add a 'bool throttle' to
qdisc_watchdog_schedule_ns(), and remove calls to qdisc_unthrottled()
in sch_fq.
Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Its unfortunate we have to walk again skb list to find the tail
after segmentation, even if data is probably hot in cpu caches.
skb_segment() can store the tail of the list into segs->prev,
and validate_xmit_skb_list() can immediately get the tail.
Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
The Openvswitch implementation is completely agnostic to the options
that are in use and can handle newly defined options without
further work. It does this by simply matching on a byte array
of options and allowing userspace to setup flows on this array.
Signed-off-by: Jesse Gross <jesse@nicira.com>
Singed-off-by: Ansis Atteka <aatteka@nicira.com>
Signed-off-by: Andy Zhou <azhou@nicira.com>
Acked-by: Thomas Graf <tgraf@noironetworks.com>
Acked-by: Pravin B Shelar <pshelar@nicira.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
As the size of the flow key grows, it can put some pressure on the
stack. This is particularly true in ovs_flow_cmd_set(), which needs several
copies of the key on the stack. One of those uses is logically separate,
so this factors it out to reduce stack pressure and improve readibility.
Signed-off-by: Jesse Gross <jesse@nicira.com>
Signed-off-by: Andy Zhou <azhou@nicira.com>
Acked-by: Pravin B Shelar <pshelar@nicira.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Currently, the flow information that is matched for tunnels and
the tunnel data passed around with packets is the same. However,
as additional information is added this is not necessarily desirable,
as in the case of pointers.
This adds a new structure for tunnel metadata which currently contains
only the existing struct. This change is purely internal to the kernel
since the current OVS_KEY_ATTR_IPV4_TUNNEL is simply a compressed version
of OVS_KEY_ATTR_TUNNEL that is translated at flow setup.
Signed-off-by: Jesse Gross <jesse@nicira.com>
Signed-off-by: Andy Zhou <azhou@nicira.com>
Acked-by: Pravin B Shelar <pshelar@nicira.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Some tunnel formats have mechanisms for indicating that packets are
OAM frames that should be handled specially (either as high priority or
not forwarded beyond an endpoint). This provides support for allowing
those types of packets to be matched.
Signed-off-by: Jesse Gross <jesse@nicira.com>
Signed-off-by: Andy Zhou <azhou@nicira.com>
Acked-by: Pravin B Shelar <pshelar@nicira.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
As new protocols are added, the size of the flow key tends to
increase although few protocols care about all of the fields. In
order to optimize this for hashing and matching, OVS uses a variable
length portion of the key. However, when fields are extracted from
the packet we must still zero out the entire key.
This is no longer necessary now that OVS implements masking. Any
fields (or holes in the structure) which are not part of a given
protocol will be by definition not part of the mask and zeroed out
during lookup. Furthermore, since masking already uses variable
length keys this zeroing operation automatically benefits as well.
In principle, the only thing that needs to be done at this point
is remove the memset() at the beginning of flow. However, some
fields assume that they are initialized to zero, which now must be
done explicitly. In addition, in the event of an error we must also
zero out corresponding fields to signal that there is no valid data
present. These increase the total amount of code but very little of
it is executed in non-error situations.
Removing the memset() reduces the profile of ovs_flow_extract()
from 0.64% to 0.56% when tested with large packets on a 10G link.
Suggested-by: Pravin Shelar <pshelar@nicira.com>
Signed-off-by: Jesse Gross <jesse@nicira.com>
Signed-off-by: Andy Zhou <azhou@nicira.com>
Acked-by: Pravin B Shelar <pshelar@nicira.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
This adds a device level support for Geneve -- Generic Network
Virtualization Encapsulation. The protocol is documented at
http://tools.ietf.org/html/draft-gross-geneve-01
Only protocol layer Geneve support is provided by this driver.
Openvswitch can be used for configuring, set up and tear down
functional Geneve tunnels.
Signed-off-by: Jesse Gross <jesse@nicira.com>
Signed-off-by: Andy Zhou <azhou@nicira.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Currently association restarts do not take into consideration the
state of the socket. When a restart happens, the current assocation
simply transitions into established state. This creates a condition
where a remote system, through a the restart procedure, may create a
local association that is no way reachable by user. The conditions
to trigger this are as follows:
1) Remote does not acknoledge some data causing data to remain
outstanding.
2) Local application calls close() on the socket. Since data
is still outstanding, the association is placed in SHUTDOWN_PENDING
state. However, the socket is closed.
3) The remote tries to create a new association, triggering a restart
on the local system. The association moves from SHUTDOWN_PENDING
to ESTABLISHED. At this point, it is no longer reachable by
any socket on the local system.
This patch addresses the above situation by moving the newly ESTABLISHED
association into SHUTDOWN-SENT state and bundling a SHUTDOWN after
the COOKIE-ACK chunk. This way, the restarted associate immidiately
enters the shutdown procedure and forces the termination of the
unreachable association.
Reported-by: David Laight <David.Laight@aculab.com>
Signed-off-by: Vlad Yasevich <vyasevich@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
John W. Linville says:
====================
pull request: wireless-next 2014-10-03
Please pull tihs batch of updates intended for the 3.18 stream!
For the iwlwifi bits, Emmanuel says:
"I have here a few things that depend on the latest mac80211's changes:
RRM, TPC, Quiet Period etc... Eyal keeps improving our rate control
and we have a new device ID. This last patch should probably have
gone to wireless.git, but at that stage, I preferred to send it to
-next and CC stable."
For (most of) the Atheros bits, Kalle says:
"The only new feature is testmode support from me. Ben added a new method
to crash the firmware with an assert for debug purposes. As usual, we
have lots of smaller fixes from Michal. Matteo fixed a Kconfig
dependency with debugfs. I fixed some warnings recently added to
checkpatch."
For the NFC bits, Samuel says:
"We've had major updates for TI and ST Microelectronics drivers, and a
few NCI related changes.
For TI's trf7970a driver:
- Target mode support for trf7970a
- Suspend/resume support for trf7970a
- DT properties additions to handle different quirks
- A bunch of fixes for smartphone IOP related issues
For ST Microelectronics' ST21NFCA and ST21NFCB drivers:
- ISO15693 support for st21nfcb
- checkpatch and sparse related warning fixes
- Code cleanups and a few minor fixes
Finally, Marvell added ISO15693 support to the NCI stack, together with a
couple of NCI fixes."
For the Bluetooth bits, Johan says:
"This 3.18 pull request replaces the one I did on Monday ("bluetooth-next
2014-09-22", which hasn't been pulled yet). The additions since the last
request are:
- SCO connection fix for devices not supporting eSCO
- Cleanups regarding the SCO establishment logic
- Remove unnecessary return value from logging functions
- Header compression fix for 6lowpan
- Cleanups to the ieee802154/mrf24j40 driver
Here's a copy from previous request that this one replaces:
'
Here are some more patches for 3.18. They include various fixes to the
btusb HCI driver, a fix for LE SMP, as well as adding Jukka to the
MAINTAINERS file for generic 6LoWPAN (as requested by Alexander Aring).
I've held on to this pull request a bit since we were waiting for a SCO
related fix to get sorted out first. However, since the merge window is
getting closer I decided not to wait for it. If we do get the fix sorted
out there'll probably be a second small pull request later this week.
'"
And,
"Unless 3.17 gets delayed this will probably be our last -next pull request for
3.18. We've got:
- New Marvell hardware supportr
- Multicast support for 6lowpan
- Several of 6lowpan fixes & cleanups
- Fix for a (false-positive) lockdep warning in L2CAP
- Minor btusb cleanup"
On top of all that comes the usual sort of updates to ath5k, ath9k,
ath10k, brcmfmac, mwifiex, and wil6210. This time around there are
also a number of rtlwifi updates to enable some new hardware and
to reconcile the in-kernel drivers with some newer releases of the
Realtek vendor drivers. Also of note is some device tree work for
the bcma bus.
Please let me know if there are problems!
====================
Signed-off-by: David S. Miller <davem@davemloft.net>
Pablo Neira Ayuso says:
====================
Netfilter/IPVS updates for net-next
The following patchset contains another batch with Netfilter/IPVS updates
for net-next, they are:
1) Add abstracted ICMP codes to the nf_tables reject expression. We
introduce four reasons to reject using ICMP that overlap in IPv4
and IPv6 from the semantic point of view. This should simplify the
maintainance of dual stack rule-sets through the inet table.
2) Move nf_send_reset() functions from header files to per-family
nf_reject modules, suggested by Patrick McHardy.
3) We have to use IS_ENABLED(CONFIG_BRIDGE_NETFILTER) everywhere in the
code now that br_netfilter can be modularized. Convert remaining spots
in the network stack code.
4) Use rcu_barrier() in the nf_tables module removal path to ensure that
we don't leave object that are still pending to be released via
call_rcu (that may likely result in a crash).
5) Remove incomplete arch 32/64 compat from nft_compat. The original (bad)
idea was to probe the word size based on the xtables match/target info
size, but this assumption is wrong when you have to dump the information
back to userspace.
6) Allow to filter from prerouting and postrouting in the nf_tables bridge.
In order to emulate the ebtables NAT chains (which are actually simple
filter chains with no special semantics), we have support filtering from
this hooks too.
7) Add explicit module dependency between xt_physdev and br_netfilter.
This provides a way to detect if the user needs br_netfilter from
the configuration path. This should reduce the breakage of the
br_netfilter modularization.
8) Cleanup coding style in ip_vs.h, from Simon Horman.
9) Fix crash in the recently added nf_tables masq expression. We have
to register/unregister the notifiers to clean up the conntrack table
entries from the module init/exit path, not from the rule addition /
deletion path. From Arturo Borrero.
====================
Signed-off-by: David S. Miller <davem@davemloft.net>
Currently when vlan filtering is turned on on the bridge, the bridge
will drop all traffic untill the user configures the filter. This
isn't very nice for ports that don't care about vlans and just
want untagged traffic.
A concept of a default_pvid was recently introduced. This patch
adds filtering support for default_pvid. Now, ports that don't
care about vlans and don't define there own filter will belong
to the VLAN of the default_pvid and continue to receive untagged
traffic.
This filtering can be disabled by setting default_pvid to 0.
Signed-off-by: Vladislav Yasevich <vyasevic@redhat.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Currently, if the pvid is not set, we return an illegal vlan value
even though the pvid value is set to 0. Since pvid of 0 is currently
invalid, just return 0 instead. This makes the current and future
checks simpler.
Signed-off-by: Vladislav Yasevich <vyasevic@redhat.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
This patch allows the user to set and retrieve default_pvid
value. A new value can only be stored when vlan filtering
is disabled.
Signed-off-by: Vladislav Yasevich <vyasevic@redhat.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
The result of a negated container has to be inverted before checking for
early ending.
This fixes my previous attempt (17c9c8232663a47f074b7452b9b034efda868ca7) to
make inverted containers work correctly.
Signed-off-by: Ignacy Gawędzki <ignacy.gawedzki@green-communications.fr>
Signed-off-by: David S. Miller <davem@davemloft.net>
Commit f7f1de51edbd ("net: dsa: start and stop the PHY state machine")
add calls to phy_start() in dsa_slave_open() respectively phy_stop() in
dsa_slave_close().
We also call phy_start_aneg() in dsa_slave_create(), and this call is
messing up with the PHY state machine, since we basically start the
auto-negotiation, and later on restart it when calling phy_start().
phy_start() does not currently handle the PHY_FORCING or PHY_AN states
properly, but such a fix would be too invasive for this window.
Fixes: f7f1de51edbd ("net: dsa: start and stop the PHY state machine")
Signed-off-by: Florian Fainelli <f.fainelli@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
SKB_FCLONE_UNAVAILABLE has overloaded meaning depending on type of skb.
1: If skb is allocated from head_cache, it indicates fclone is not available.
2: If skb is a companion fclone skb (allocated from fclone_cache), it indicates
it is available to be used.
To avoid confusion for case 2 above, this patch replaces
SKB_FCLONE_UNAVAILABLE with SKB_FCLONE_FREE where appropriate. For fclone
companion skbs, this indicates it is free for use.
SKB_FCLONE_UNAVAILABLE will now simply indicate skb is from head_cache and
cannot / will not have a companion fclone.
Signed-off-by: Vijay Subramanian <subramanian.vijay@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
In xmit path, we build a flowi6 which will be used for the output route lookup.
We are sending a GRE packet, neither IPv4 nor IPv6 encapsulated packet, thus the
protocol should be IPPROTO_GRE.
Fixes: c12b395a4664 ("gre: Support GRE over IPv6")
Reported-by: Matthieu Ternisien d'Ouville <matthieu.tdo@6wind.com>
Signed-off-by: Nicolas Dichtel <nicolas.dichtel@6wind.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
This patch allows configuring IPIP, sit, and GRE tunnels to use GUE.
This is very similar to fou excpet that we need to insert the GUE header
in addition to the UDP header on transmit.
Signed-off-by: Tom Herbert <therbert@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
This patch adds support receiving for GUE packets in the fou module. The
fou module now supports direct foo-over-udp (no encapsulation header)
and GUE. To support this a type parameter is added to the fou netlink
parameters.
For a GUE socket we define gue_udp_recv, gue_gro_receive, and
gue_gro_complete to handle the specifics of the GUE protocol. Most
of the code to manage and configure sockets is common with the fou.
Signed-off-by: Tom Herbert <therbert@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
This patch removes fou[46]_gro_receive and fou[46]_gro_complete
functions. The v4 or v6 variants were chosen for the UDP offloads
based on the address family of the socket this is not necessary
or correct. Alternatively, this patch adds is_ipv6 to napi_gro_skb.
This is set in udp6_gro_receive and unset in udp4_gro_receive. In
fou_gro_receive the value is used to select the correct inet_offloads
for the protocol of the outer IP header.
Signed-off-by: Tom Herbert <therbert@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
When adjusting max_header for the tunnel interface based on egress
device we need to account for any extra bytes in secondary encapsulation
(e.g. FOU).
Signed-off-by: Tom Herbert <therbert@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
skb_gro_receive() is only called from tcp_gro_receive() which is
not in a module.
Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Validation of skb can be pretty expensive :
GSO segmentation and/or checksum computations.
We can do this without holding qdisc lock, so that other cpus
can queue additional packets.
Trick is that requeued packets were already validated, so we carry
a boolean so that sch_direct_xmit() can validate a fresh skb list,
or directly use an old one.
Tested on 40Gb NIC (8 TX queues) and 200 concurrent flows, 48 threads
host.
Turning TSO on or off had no effect on throughput, only few more cpu
cycles. Lock contention on qdisc lock disappeared.
Same if disabling TX checksum offload.
Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
I got a report of a double free happening at RDS slab cache. One
suspicion was that may be somewhere we were doing a sock_hold/sock_put
on an already freed sock. Thus after providing a kernel with the
following change:
static inline void sock_hold(struct sock *sk)
{
- atomic_inc(&sk->sk_refcnt);
+ if (!atomic_inc_not_zero(&sk->sk_refcnt))
+ WARN(1, "Trying to hold sock already gone: %p (family: %hd)\n",
+ sk, sk->sk_family);
}
The warning successfuly triggered:
Trying to hold sock already gone: ffff81f6dda61280 (family: 21)
WARNING: at include/net/sock.h:350 sock_hold()
Call Trace:
<IRQ> [<ffffffff8adac135>] :rds:rds_send_remove_from_sock+0xf0/0x21b
[<ffffffff8adad35c>] :rds:rds_send_drop_acked+0xbf/0xcf
[<ffffffff8addf546>] :rds_rdma:rds_ib_recv_tasklet_fn+0x256/0x2dc
[<ffffffff8009899a>] tasklet_action+0x8f/0x12b
[<ffffffff800125a2>] __do_softirq+0x89/0x133
[<ffffffff8005f30c>] call_softirq+0x1c/0x28
[<ffffffff8006e644>] do_softirq+0x2c/0x7d
[<ffffffff8006e4d4>] do_IRQ+0xee/0xf7
[<ffffffff8005e625>] ret_from_intr+0x0/0xa
<EOI>
Looking at the call chain above, the only way I think this would be
possible is if somewhere we already released the same socket->sock which
is assigned to the rds_message at rds_send_remove_from_sock. Which seems
only possible to happen after the tear down done on rds_release.
rds_release properly calls rds_send_drop_to to drop the socket from any
rds_message, and some proper synchronization is in place to avoid race
with rds_send_drop_acked/rds_send_remove_from_sock. However, I still see
a very narrow window where it may be possible we touch a sock already
released: when rds_release races with rds_send_drop_acked, we check
RDS_MSG_ON_CONN to avoid cleanup on the same rds_message, but in this
specific case we don't clear rm->m_rs. In this case, it seems we could
then go on at rds_send_drop_to and after it returns, the sock is freed
by last sock_put on rds_release, with concurrently we being at
rds_send_remove_from_sock; then at some point in the loop at
rds_send_remove_from_sock we process an rds_message which didn't have
rm->m_rs unset for a freed sock, and a possible sock_hold on an sock
already gone at rds_release happens.
This hopefully address the described condition above and avoids a double
free on "second last" sock_put. In addition, I removed the comment about
socket destruction on top of rds_send_drop_acked: we call rds_send_drop_to
in rds_release and we should have things properly serialized there, thus
I can't see the comment being accurate there.
Signed-off-by: Herton R. Krzesinski <herton@redhat.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
I see two problems if we consider the sock->ops->connect attempt to fail in
rds_tcp_conn_connect. The first issue is that for example we don't remove the
previously added rds_tcp_connection item to rds_tcp_tc_list at
rds_tcp_set_callbacks, which means that on a next reconnect attempt for the
same rds_connection, when rds_tcp_conn_connect is called we can again call
rds_tcp_set_callbacks, resulting in duplicated items on rds_tcp_tc_list,
leading to list corruption: to avoid this just make sure we call
properly rds_tcp_restore_callbacks before we exit. The second issue
is that we should also release the sock properly, by setting sock = NULL
only if we are returning without error.
Signed-off-by: Herton R. Krzesinski <herton@redhat.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
The TSO and GSO segmented packets already benefit from bulking
on their own.
The TSO packets have always taken advantage of the only updating
the tailptr once for a large packet.
The GSO segmented packets have recently taken advantage of
bulking xmit_more API, via merge commit 53fda7f7f9e8 ("Merge
branch 'xmit_list'"), specifically via commit 7f2e870f2a4 ("net:
Move main gso loop out of dev_hard_start_xmit() into helper.")
allowing qdisc requeue of remaining list. And via commit
ce93718fb7cd ("net: Don't keep around original SKB when we
software segment GSO frames.").
This patch allow further bulking of TSO/GSO packets together,
when dequeueing from the qdisc.
Testing:
Measuring HoL (Head-of-Line) blocking for TSO and GSO, with
netperf-wrapper. Bulking several TSO show no performance regressions
(requeues were in the area 32 requeues/sec).
Bulking several GSOs does show small regression or very small
improvement (requeues were in the area 8000 requeues/sec).
Using ixgbe 10Gbit/s with GSO bulking, we can measure some additional
latency. Base-case, which is "normal" GSO bulking, sees varying
high-prio queue delay between 0.38ms to 0.47ms. Bulking several GSOs
together, result in a stable high-prio queue delay of 0.50ms.
Using igb at 100Mbit/s with GSO bulking, shows an improvement.
Base-case sees varying high-prio queue delay between 2.23ms to 2.35ms
Signed-off-by: David S. Miller <davem@davemloft.net>
Based on DaveM's recent API work on dev_hard_start_xmit(), that allows
sending/processing an entire skb list.
This patch implements qdisc bulk dequeue, by allowing multiple packets
to be dequeued in dequeue_skb().
The optimization principle for this is two fold, (1) to amortize
locking cost and (2) avoid expensive tailptr update for notifying HW.
(1) Several packets are dequeued while holding the qdisc root_lock,
amortizing locking cost over several packet. The dequeued SKB list is
processed under the TXQ lock in dev_hard_start_xmit(), thus also
amortizing the cost of the TXQ lock.
(2) Further more, dev_hard_start_xmit() will utilize the skb->xmit_more
API to delay HW tailptr update, which also reduces the cost per
packet.
One restriction of the new API is that every SKB must belong to the
same TXQ. This patch takes the easy way out, by restricting bulk
dequeue to qdisc's with the TCQ_F_ONETXQUEUE flag, that specifies the
qdisc only have attached a single TXQ.
Some detail about the flow; dev_hard_start_xmit() will process the skb
list, and transmit packets individually towards the driver (see
xmit_one()). In case the driver stops midway in the list, the
remaining skb list is returned by dev_hard_start_xmit(). In
sch_direct_xmit() this returned list is requeued by dev_requeue_skb().
To avoid overshooting the HW limits, which results in requeuing, the
patch limits the amount of bytes dequeued, based on the drivers BQL
limits. In-effect bulking will only happen for BQL enabled drivers.
Small amounts for extra HoL blocking (2x MTU/0.24ms) were
measured at 100Mbit/s, with bulking 8 packets, but the
oscillating nature of the measurement indicate something, like
sched latency might be causing this effect. More comparisons
show, that this oscillation goes away occationally. Thus, we
disregard this artifact completely and remove any "magic" bulking
limit.
For now, as a conservative approach, stop bulking when seeing TSO and
segmented GSO packets. They already benefit from bulking on their own.
A followup patch add this, to allow easier bisect-ability for finding
regressions.
Jointed work with Hannes, Daniel and Florian.
Signed-off-by: Jesper Dangaard Brouer <brouer@redhat.com>
Signed-off-by: Hannes Frederic Sowa <hannes@stressinduktion.org>
Signed-off-by: Daniel Borkmann <dborkman@redhat.com>
Signed-off-by: Florian Westphal <fw@strlen.de>
Signed-off-by: David S. Miller <davem@davemloft.net>
We have to register the notifiers in the masquerade expression from
the the module _init and _exit path.
This fixes crashes when removing the masquerade rule with no
ipt_MASQUERADE support in place (which was masking the problem).
Fixes: 9ba1f72 ("netfilter: nf_tables: add new nft_masq expression")
Signed-off-by: Arturo Borrero Gonzalez <arturo.borrero.glez@gmail.com>
Signed-off-by: Pablo Neira Ayuso <pablo@netfilter.org>
Conflicts:
drivers/net/usb/r8152.c
net/netfilter/nfnetlink.c
Both r8152 and nfnetlink conflicts were simple overlapping changes.
Signed-off-by: David S. Miller <davem@davemloft.net>
