Jonas Rebmann says:
====================
improve multicast join group performance
This series seeks to improve performance on updating igmp group
memberships such as with IP_ADD_MEMBERSHIP or MCAST_JOIN_SOURCE_GROUP.
Our use case was to add 2000 multicast memberships on a TQMLS1046A which
took about 3.6 seconds for the membership additions alone. Our userspace
reproducer tool was instrumented to log runtimes of the individual
setsockopt invocations which clearly indicated quadratic complexity of
setting up the membership with regard to the total number of multicast
groups to be joined. We used perf to locate the hotspots and
subsequently optimized the most costly sections of code.
This series includes a patch to Linux igmp handling as well as a patch
to the DPAA/Freescale driver. With both patches applied, our memberships can
be set up in only about 87 miliseconds, which corresponds to a speedup
of around 40.
While we have acheived practically linear run-time complexity on the
kernel side, a small quadratic factor remains in parts of the freescale
driver code which we haven't yet optimized. We have by now payed little
attention to the optimization potential in dropping group memberships,
yet the dpaa patch applies to joining and leaving groups alike.
Overall, this patch series brings great improvements in use cases
involving large numbers of multicast groups, particularly when using the
fsl_dpa driver, without noteworthy drawbacks in other scenarios.
====================
Signed-off-by: Jonas Rebmann <jre@pengutronix.de>
Signed-off-by: David S. Miller <davem@davemloft.net>
The original driver first unregisters then re-registers all multicast
addresses in the struct net_device_ops::ndo_set_rx_mode() callback.
As the networking stack calls ndo_set_rx_mode() if a single multicast
address change occurs, a significant amount of time may be used to first
unregister and then re-register unchanged multicast addresses. This
leads to performance issues when tracking large numbers of multicast
addresses.
Replace the unregister and register loop and the hand crafted
mc_addr_list list handling with __dev_mc_sync(), to only update entries
which have changed.
On profiling with an fsl_dpa NIC, this patch presented a speedup of
around 40 when successively setting up 2000 multicast groups using
setsockopt(), without drawbacks on smaller numbers of multicast groups.
Signed-off-by: Jonas Rebmann <jre@pengutronix.de>
Reviewed-by: Sean Anderson <sean.anderson@seco.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
The runtime cost of joining a single multicast group in the current
implementation of ____ip_mc_inc_group grows linearly with the number of
existing memberships. This is caused by the linear search for an
existing group record in the multicast address list.
This linear complexity results in quadratic complexity when successively
adding memberships, which becomes a performance bottleneck when setting
up large numbers of multicast memberships.
If available, use the existing multicast hash map mc_hash to quickly
search for an existing group membership record. This leads to
near-constant complexity on the addition of a new multicast record,
significantly improving performance for workloads involving many
multicast memberships.
On profiling with a loopback device, this patch presented a speedup of
around 6 when successively setting up 2000 multicast groups using
setsockopt without measurable drawbacks on smaller numbers of
multicast groups.
Signed-off-by: Jonas Rebmann <jre@pengutronix.de>
Reviewed-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
The vmclock device addresses the problem of live migration with
precision clocks. The tolerances of a hardware counter (e.g. TSC) are
typically around ±50PPM. A guest will use NTP/PTP/PPS to discipline that
counter against an external source of 'real' time, and track the precise
frequency of the counter as it changes with environmental conditions.
When a guest is live migrated, anything it knows about the frequency of
the underlying counter becomes invalid. It may move from a host where
the counter running at -50PPM of its nominal frequency, to a host where
it runs at +50PPM. There will also be a step change in the value of the
counter, as the correctness of its absolute value at migration is
limited by the accuracy of the source and destination host's time
synchronization.
In its simplest form, the device merely advertises a 'disruption_marker'
which indicates that the guest should throw away any NTP synchronization
it thinks it has, and start again.
Because the shared memory region can be exposed all the way to userspace
through the /dev/vmclock0 node, applications can still use time from a
fast vDSO 'system call', and check the disruption marker to be sure that
their timestamp is indeed truthful.
The structure also allows for the precise time, as known by the host, to
be exposed directly to guests so that they don't have to wait for NTP to
resync from scratch. The PTP driver consumes this information if present.
Like the KVM PTP clock, this PTP driver can convert TSC-based cross
timestamps into KVM clock values. Unlike the KVM PTP clock, it does so
only when such is actually helpful.
The values and fields are based on the nascent virtio-rtc specification,
and the intent is that a version (hopefully precisely this version) of
this structure will be included as an optional part of that spec. In the
meantime, this driver supports the simple ACPI form of the device which
is being shipped in certain commercial hypervisors (and submitted for
inclusion in QEMU).
Signed-off-by: David Woodhouse <dwmw@amazon.co.uk>
Acked-by: Richard Cochran <richardcochran@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Russell King says:
====================
net: pcs: xpcs: cleanups batch 2
This is the second cleanup series for XPCS.
Patch 1 removes the enum indexing the dw_xpcs_compat array. The index is
never used except to place entries in the array and to size the array.
Patch 2 removes the interface arrays - each of which only contain one
interface.
Patch 3 makes xpcs_find_compat() take the xpcs structure rather than the
ID - the previous series removed the reason for xpcs_find_compat needing
to take the ID.
Patch 4 provides a helper to convert xpcs structure to a regular
phylink_pcs structure, which leads to patch 5.
Patch 5 moves the definition of struct dw_xpcs to the private xpcs
header - with patch 4 in place, nothing outside of the xpcs driver
accesses the contents of the dw_xpcs structure.
Patch 6 renames xpcs_get_id() to xpcs_read_id() since it's reading the
ID, rather than doing anything further with it. (Prior versions of this
series renamed it to xpcs_read_phys_id() since that more accurately
described that it was reading the physical ID registers.)
Patch 7 moves the searching of the ID list out of line as this is a
separate functional block.
Patch 8 converts xpcs to use the bitmap macros, which eliminates the
need for _SHIFT definitions.
Patch 9 adds and uses _modify() accessors as there are a large amount
of read-modify-write operations in this driver. This conversion found
a bug in xpcs-wx code that has been reported and already fixed.
Patch 10 converts xpcs to use read_poll_timeout() rather than open
coding that.
Patch 11 converts all printed messages to use the dev_*() functions so
the driver and devie name are always printed.
Patch 12 moves DW_VR_MII_DIG_CTRL1_2G5_EN to the correct place in the
header file, rather than amongst another register's definitions.
Patch 13 moves the Wangxun workaround to a common location rather than
duplicating it in two places. We also reformat this to fit within
80 columns.
====================
Tested-by: Serge Semin <fancer.lancer@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
According to commits 2a22b7ae2fa3 ("net: pcs: xpcs: adapt Wangxun NICs
for SGMII mode") and 2deea43f386d ("net: pcs: xpcs: add 1000BASE-X AN
interrupt support"), Wangxun devices need special VR_XS_PCS_DIG_CTRL1
settings for SGMII and 1000BASE-X. Both SGMII and 1000BASE-X use the
same settings.
Rather than placing these in the individual xpcs_config_*() functions,
move it to where we already test for the Wangxun devices in
xpcs_do_config().
Signed-off-by: Russell King (Oracle) <rmk+kernel@armlinux.org.uk>
Signed-off-by: David S. Miller <davem@davemloft.net>
Place DW_VR_MII_DIG_CTRL1_2G5_EN with the other DW_VR_MII_DIG_CTRL1
definitions rather than in the middle of a register list.
Signed-off-by: Russell King (Oracle) <rmk+kernel@armlinux.org.uk>
Signed-off-by: David S. Miller <davem@davemloft.net>
Use the dev_*() family of functions to print all messages from the XPCS
driver so we know which instance issues the messages.
Signed-off-by: Russell King (Oracle) <rmk+kernel@armlinux.org.uk>
Signed-off-by: David S. Miller <davem@davemloft.net>
Convert the xpcs driver to use read_poll_timeout() when waiting for
reset to complete, rather than open-coding this.
Signed-off-by: Russell King (Oracle) <rmk+kernel@armlinux.org.uk>
Signed-off-by: David S. Miller <davem@davemloft.net>
The xpcs driver does a lot of read-modify-write operations on
registers, which leads to long-winded code to read the register, check
whether the read was successful, modify the value in some way, and then
write it back.
We have a mdiodev _modify() accessor that encapsulates this, and does
the register modification under the MDIO bus lock ensuring that the
modification is atomic with respect to other bus operations. Convert
the xpcs driver to use this accessor.
Signed-off-by: Russell King (Oracle) <rmk+kernel@armlinux.org.uk>
Signed-off-by: David S. Miller <davem@davemloft.net>
Convert xpcs to use the bitfield macros rather than definining the
bitfield shifts and open-coding the insertion and extraction of these
bitfields.
Signed-off-by: Russell King (Oracle) <rmk+kernel@armlinux.org.uk>
Signed-off-by: David S. Miller <davem@davemloft.net>
Move the searching of the physical ID out of xpcs_create() and into
its own xpcs_identify() function, which makes it self contained.
This reduces the complexity in xpcs_craete(), making it easier to
follow, rather than having a lot of once-run code in the big for()
loop.
Signed-off-by: Russell King (Oracle) <rmk+kernel@armlinux.org.uk>
Signed-off-by: David S. Miller <davem@davemloft.net>
Rename xpcs_get_id() to xpcs_read_id() which more closely reflects
the purpose of this function.
Signed-off-by: Russell King (Oracle) <rmk+kernel@armlinux.org.uk>
Signed-off-by: David S. Miller <davem@davemloft.net>
There should be no reason for anything outside the XPCS code to know
the contents of struct dw_xpcs - this is a private structure to XPCS.
Move the definition to the private pcs-xpcs.h header, leaving a
declaration in the global pcs/pcs-xpcs.h
Signed-off-by: Russell King (Oracle) <rmk+kernel@armlinux.org.uk>
Signed-off-by: David S. Miller <davem@davemloft.net>
Provide a helper to provide the pointer to the phylink_pcs struct
given a valid xpcs pointer. This will be necessary when we make
struct dw_xpcs private to pcs-xpcs.c
Signed-off-by: Russell King (Oracle) <rmk+kernel@armlinux.org.uk>
Signed-off-by: David S. Miller <davem@davemloft.net>
xpcs_find_compat() is now always passed xpcs->id. Rather than always
dereferencing this in the caller, move it into xpcs_find_compat(),
thus making this function consistent with most of the other xpcs
functions in taking an xpcs pointer.
Signed-off-by: Russell King (Oracle) <rmk+kernel@armlinux.org.uk>
Signed-off-by: David S. Miller <davem@davemloft.net>
Currently, xpcs uses an array of interfaces that each "compat" entry
supports. When looking up the compat entry for an interface, we
iterate over the compat entries and then over each interface.
Since each compat entry only has a single interface in its interfaces
array, replace the array with a single member in the compat structure.
Signed-off-by: Russell King (Oracle) <rmk+kernel@armlinux.org.uk>
Signed-off-by: David S. Miller <davem@davemloft.net>
There is no reason for the struct dw_xpcs_compat arrays to be a fixed
size other than the way we iterate over them. The index into the array
isn't used for anything, and having them fixed size needlessly wastes
space.
Remove the enum that defines their size, and instead use an empty
array entry (with NULL ->supported) to mark the end of the array.
Signed-off-by: Russell King (Oracle) <rmk+kernel@armlinux.org.uk>
Signed-off-by: David S. Miller <davem@davemloft.net>
Add support for measuring Signal Quality Index for LAN887x T1 PHY.
Signal Quality Index (SQI) is measure of Link Channel Quality from
0 to 7, with 7 as the best. By default, a link loss event shall
indicate an SQI of 0.
Signed-off-by: Tarun Alle <Tarun.Alle@microchip.com>
Reviewed-by: Andrew Lunn <andrew@lunn.ch>
Link: https://patch.msgid.link/20241007063943.3233-1-tarun.alle@microchip.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Niklas Söderlund says:
====================
net: phy: marvell-88q2xxx: Enable auto negotiation for mv88q2110
This series enables auto negotiation for the mv88q2110 device.
Previously this feature have been disabled for mv88q2110, while enabled
for other devices supported by this driver.
The initial driver implementation states this is due to the
configuration sequence provided by the vendor did not work. By comparing
the initialization sequence of other devices this driver supports and
the out-of-tree PHY driver for mv88q2110 found in the Renesas BSP [1]
I was able to figure out a working configuration.
As I have no access to the datasheets of either of these devices it
would be super if someone who has could sanity check the initialization
sequence.
With this series I'm able to auto negotiate both 1000Mbps and 100Mbps
links without issue.
# ethtool eth0
Settings for eth0:
Supported ports: [ ]
Supported link modes: 100baseT1/Full
1000baseT1/Full
Supported pause frame use: Symmetric Receive-only
Supports auto-negotiation: Yes
Supported FEC modes: Not reported
Advertised link modes: 100baseT1/Full
1000baseT1/Full
Advertised pause frame use: No
Advertised auto-negotiation: Yes
Advertised FEC modes: Not reported
Link partner advertised link modes: 100baseT1/Full
1000baseT1/Full
Link partner advertised pause frame use: No
Link partner advertised auto-negotiation: Yes
Link partner advertised FEC modes: Not reported
Speed: 1000Mb/s
Duplex: Full
Auto-negotiation: on
master-slave cfg: preferred master
master-slave status: slave
Port: Twisted Pair
PHYAD: 0
Transceiver: external
MDI-X: Unknown
Link detected: yes
SQI: 15/15
And the performance is good too. Without this change I was not able to
manually configure a 1000Mbps link, only 100Mbps ones. So this gives a
huge performance boost for my use-case.
[ 5] local 10.1.0.2 port 5201 connected to 10.1.0.1 port 38346
[ ID] Interval Transfer Bitrate Retr Cwnd
[ 5] 0.00-1.00 sec 96.8 MBytes 812 Mbits/sec 0 469 KBytes
[ 5] 1.00-2.00 sec 94.3 MBytes 791 Mbits/sec 0 469 KBytes
[ 5] 2.00-3.00 sec 96.1 MBytes 806 Mbits/sec 0 469 KBytes
[ 5] 3.00-4.00 sec 98.3 MBytes 825 Mbits/sec 0 469 KBytes
[ 5] 4.00-5.00 sec 98.4 MBytes 825 Mbits/sec 0 469 KBytes
[ 5] 5.00-6.00 sec 98.4 MBytes 826 Mbits/sec 0 469 KBytes
[ 5] 6.00-7.00 sec 98.9 MBytes 830 Mbits/sec 0 469 KBytes
[ 5] 7.00-8.00 sec 91.7 MBytes 769 Mbits/sec 0 469 KBytes
[ 5] 8.00-9.00 sec 99.4 MBytes 834 Mbits/sec 0 747 KBytes
[ 5] 9.00-10.00 sec 101 MBytes 851 Mbits/sec 0 747 KBytes
Patch 1/3 and 2/3 are preparation patches that align and move functions
around as the mv88q2110 code paths can now reuses much of what is done
for mv88q2220. While patch 3/3 adds the new initialization sequence and
removes the auto negotiation limit for mv88q2110.
1. 2a1f07d0e7
====================
Link: https://patch.msgid.link/20241005112412.544360-1-niklas.soderlund+renesas@ragnatech.se
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
The initial marvell-88q2xxx driver only supported the Marvell 88Q2110
PHY without auto negotiation support. The reason documented states that
the provided initialization sequence did not to work. Now a method to
enable auto negotiation have been found by comparing the initialization
of other supported devices and an out-of-tree PHY driver.
Perform the minimal needed initialization of the PHY to get auto
negotiation working and remove the limitation that disables the auto
negotiation feature for the mv88q2110 device.
With this change a 1000Mbps full duplex link is able to be negotiated
between two mv88q2110 and the link works perfectly. The other side also
reflects the manually configure settings of the master device.
# ethtool eth0
Settings for eth0:
Supported ports: [ ]
Supported link modes: 100baseT1/Full
1000baseT1/Full
Supported pause frame use: Symmetric Receive-only
Supports auto-negotiation: Yes
Supported FEC modes: Not reported
Advertised link modes: 100baseT1/Full
1000baseT1/Full
Advertised pause frame use: No
Advertised auto-negotiation: Yes
Advertised FEC modes: Not reported
Link partner advertised link modes: 100baseT1/Full
1000baseT1/Full
Link partner advertised pause frame use: No
Link partner advertised auto-negotiation: Yes
Link partner advertised FEC modes: Not reported
Speed: 1000Mb/s
Duplex: Full
Auto-negotiation: on
master-slave cfg: preferred master
master-slave status: slave
Port: Twisted Pair
PHYAD: 0
Transceiver: external
MDI-X: Unknown
Link detected: yes
SQI: 15/15
Before this change I was not able to manually configure 1000Mbps link,
only a 100Mpps link so this change providers an improvement in
performance for this device.
[ 5] local 10.1.0.2 port 5201 connected to 10.1.0.1 port 38346
[ ID] Interval Transfer Bitrate Retr Cwnd
[ 5] 0.00-1.00 sec 96.8 MBytes 812 Mbits/sec 0 469 KBytes
[ 5] 1.00-2.00 sec 94.3 MBytes 791 Mbits/sec 0 469 KBytes
[ 5] 2.00-3.00 sec 96.1 MBytes 806 Mbits/sec 0 469 KBytes
[ 5] 3.00-4.00 sec 98.3 MBytes 825 Mbits/sec 0 469 KBytes
[ 5] 4.00-5.00 sec 98.4 MBytes 825 Mbits/sec 0 469 KBytes
[ 5] 5.00-6.00 sec 98.4 MBytes 826 Mbits/sec 0 469 KBytes
[ 5] 6.00-7.00 sec 98.9 MBytes 830 Mbits/sec 0 469 KBytes
[ 5] 7.00-8.00 sec 91.7 MBytes 769 Mbits/sec 0 469 KBytes
[ 5] 8.00-9.00 sec 99.4 MBytes 834 Mbits/sec 0 747 KBytes
[ 5] 9.00-10.00 sec 101 MBytes 851 Mbits/sec 0 747 KBytes
Signed-off-by: Niklas Söderlund <niklas.soderlund+renesas@ragnatech.se>
Tested-by: Stefan Eichenberger <eichest@gmail.com>
Reviewed-by: Andrew Lunn <andrew@lunn.ch>
Link: https://patch.msgid.link/20241005112412.544360-4-niklas.soderlund+renesas@ragnatech.se
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
In preparation to adding auto negotiation support to mv88q2110 move and
rename the helper function used to write an array of register values to
the PHY.
Just as for mv88q2220 devices this helper will be needed to for the
initial configuration of the mv88q2110 to support auto negotiation.
The function is moved verbatim, there is no change in behavior.
Signed-off-by: Niklas Söderlund <niklas.soderlund+renesas@ragnatech.se>
Reviewed-by: Andrew Lunn <andrew@lunn.ch>
Tested-by: Dimitri Fedrau <dima.fedrau@gmail.com>
Tested-by: Stefan Eichenberger <eichest@gmail.com>
Link: https://patch.msgid.link/20241005112412.544360-3-niklas.soderlund+renesas@ragnatech.se
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
The soft reset implementations for mv88q2110 and mv88q2220 differ as the
later need to consider that auto negation is supported on mv88q2220
devices. In preparation of enabling auto negotiation on mv88q2110 merge
the two rest functions into a device generic one.
The mv88q2220 behavior is kept as is but extended to wait for the reset
bit to be clears before continuing, as was done previously on mv88q2220.
Signed-off-by: Niklas Söderlund <niklas.soderlund+renesas@ragnatech.se>
Reviewed-by: Andrew Lunn <andrew@lunn.ch>
Tested-by: Dimitri Fedrau <dima.fedrau@gmail.com>
Tested-by: Stefan Eichenberger <eichest@gmail.com>
Link: https://patch.msgid.link/20241005112412.544360-2-niklas.soderlund+renesas@ragnatech.se
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Fix a typo in comments: bellow -> below.
Reported-by: Matthew Wilcox <willy@infradead.org>
Signed-off-by: Andrew Kreimer <algonell@gmail.com>
Reviewed-by: Simon Horman <horms@kernel.org>
Link: https://patch.msgid.link/20241006130829.13967-1-algonell@gmail.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Despite supporting Auto MDI-X, it looks like Aquantia only supports
swapping pair (1,2) with pair (3,6) like it used to be for MDI-X on
100MBit/s networks.
When all 4 pairs are in use (for 1000MBit/s or faster) the link does not
come up with pair order is not configured correctly, either using
MDI_CFG pin or using the "PMA Receive Reserved Vendor Provisioning 1"
register.
Normally, the order of MDI pairs being either ABCD or DCBA is configured
by pulling the MDI_CFG pin.
However, some hardware designs require overriding the value configured
by that bootstrap pin. The PHY allows doing that by setting a bit in
"PMA Receive Reserved Vendor Provisioning 1" register which allows
ignoring the state of the MDI_CFG pin and another bit configuring
whether the order of MDI pairs should be normal (ABCD) or reverse
(DCBA). Pair polarity is not affected and remains identical in both
settings.
Introduce property "marvell,mdi-cfg-order" which allows forcing either
normal or reverse order of the MDI pairs from DT.
If the property isn't present, the behavior is unchanged and MDI pair
order configuration is untouched (ie. either the result of MDI_CFG pin
pull-up/pull-down, or pair order override already configured by the
bootloader before Linux is started).
Forcing normal pair order is required on the Adtran SDG-8733A Wi-Fi 7
residential gateway.
Signed-off-by: Daniel Golle <daniel@makrotopia.org>
Reviewed-by: Andrew Lunn <andrew@lunn.ch>
Link: https://patch.msgid.link/9ed760ff87d5fc456f31e407ead548bbb754497d.1728058550.git.daniel@makrotopia.org
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Usually the MDI pair order reversal configuration is defined by
bootstrap pin MDI_CFG. Some designs, however, require overriding the MDI
pair order and force either normal or reverse order.
Add property 'marvell,mdi-cfg-order' to allow forcing either normal or
reverse order of the MDI pairs.
Signed-off-by: Daniel Golle <daniel@makrotopia.org>
Reviewed-by: Rob Herring (Arm) <robh@kernel.org>
Link: https://patch.msgid.link/7ccf25d6d7859f1ce9983c81a2051cfdfb0e0a99.1728058550.git.daniel@makrotopia.org
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Petr Machata says:
====================
selftests: mlxsw: Stabilize RED tests
Tweak the mlxsw-specific RED selftests to increase stability on
Spectrum-3 and Spectrum-4 machines.
====================
Link: https://patch.msgid.link/cover.1728316370.git.petrm@nvidia.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
The RED test uses a pair of TBF shapers. The first to get predictably-sized
stream of traffic, and second to get a 100% saturated chokepoint. To this
chokepoint it injects individual packets. Because the chokepoint is
saturated, these additional packets go straight to the backlog. This allows
the test to check RED behavior across various queue sizes.
The shapers are rated at 1Gbps, for historical reasons (before mlxsw
supported TBF offload, the test used port speed to create the chokepoints).
Machines with a low-power CPU may have trouble consistently generating
1Gbps of traffic, and the test then spuriously fails.
Instead, drop the rate to 200Mbps (Spectrum has a guaranteed shaper rate
granularity of 200Mbps, so anything lower is not guaranteed to work well).
Because that means fewer packets will be mirrored in the ECN-mark test,
adjust the passing condition accordingly.
Signed-off-by: Petr Machata <petrm@nvidia.com>
Link: https://patch.msgid.link/c6712f9c5de75ae0bc2ab3d8ea7d92aaaf93af95.1728316370.git.petrm@nvidia.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
This test works by injecting into a port with a maxed-out queue a couple
packets and checks if a corresponding number of packets were dropped. This
has worked well on Spectrum<4, but on Spectrum-4 it has been noisy. This
is in line with the observation that on Spectrum-4, queue size tends to
fluctuate more. A handful of packets could then still be accepted to the
queue even though it was nominally full just recently.
In order to accommodate this behavior, send many more packets. The buffer
can fit N extra packets, but not N% packets. This therefore allows us to
set wider absolute margins, while actually narrowing them relatively.
Signed-off-by: Petr Machata <petrm@nvidia.com>
Reviewed-by: Ido Schimmel <idosch@nvidia.com>
Link: https://patch.msgid.link/abc869b9f6003d400d6293ddd5edb2f4517f44d5.1728316370.git.petrm@nvidia.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
The qdisc stats are taken from the port's periodic HW stats, which are
updated once a second. We try to accommodate the latency by using busywait
in build_backlog().
The issue in that seems to be that when do_mark_test() builds the backlog,
it makes the decision whether to send more packets based on the first
instance of the queue depth stat exceeding the current value, when in fact
more traffic is on the way and the queue depth would increase further. This
leads to failures in TC 1 of mark-mirror test, where we see the following
failure:
TEST: TC 0: marked packets mirror'd [ OK ]
TEST: TC 1: marked packets mirror'd [FAIL]
Spurious packets (1680 -> 2290) observed without buffer pressure
Fix by waiting for the full second before reading the queue depth for the
first time, to make sure it reflects all in-flight traffic.
Signed-off-by: Petr Machata <petrm@nvidia.com>
Reviewed-by: Amit Cohen <amcohen@nvidia.com>
Link: https://patch.msgid.link/321dcf8b3e9a1f0766429c8cf3e3f1746f1bc375.1728316370.git.petrm@nvidia.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Backlog fluctuates on Spectrum-4 much more than on <4. In practice we can
sample queue depth values going from about -12% to about +7% of the
configured RED limit. The test which checks the queue size has a limit of
+-10%, and as a result often fails. We attempted to fix the issue by
busywaiting for several seconds hoping to get within the bounds, but that
still proved to be too noisy (or the wait time would be impractically
long). Unfortunately we have to bump the value tolerance from 10% to 15%,
which in this patch do.
Signed-off-by: Petr Machata <petrm@nvidia.com>
Reviewed-by: Amit Cohen <amcohen@nvidia.com>
Link: https://patch.msgid.link/f54950df2a8fcba46c3ddc1053376352fa2e592b.1728316370.git.petrm@nvidia.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Backlog fluctuates on Spectrum-4 much more than on <4. Increasing the
desired backlog seems to help, as the constant fluctuations do not overlap
into the territory where packets are marked.
Signed-off-by: Petr Machata <petrm@nvidia.com>
Reviewed-by: Amit Cohen <amcohen@nvidia.com>
Link: https://patch.msgid.link/0821fb3aa8bb6a6c0d3000baab04995517c9a0cc.1728316370.git.petrm@nvidia.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Fold the separate call to clk_set_rate() into the clock getter.
Signed-off-by: Bartosz Golaszewski <bartosz.golaszewski@linaro.org>
Reviewed-by: Andrew Lunn <andrew@lunn.ch>
Link: https://patch.msgid.link/20241007134100.107921-1-brgl@bgdev.pl
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
chtls_set_tcb_tflag() has been unused since 2021's commit
827d329105bf ("chtls: Remove invalid set_tcb call")
Remove it.
Signed-off-by: Dr. David Alan Gilbert <linux@treblig.org>
Reviewed-by: Simon Horman <horms@kernel.org>
Link: https://patch.msgid.link/20241007004652.150065-1-linux@treblig.org
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
cfsrvl_getphyid() has been unused since 2011's commit
f36214408470 ("caif: Use RCU and lists in cfcnfg.c for managing caif link layers")
Remove it.
Signed-off-by: Dr. David Alan Gilbert <linux@treblig.org>
Reviewed-by: Simon Horman <horms@kernel.org>
Link: https://patch.msgid.link/20241007004456.149899-1-linux@treblig.org
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Let it be tuned in per netns by admins.
Signed-off-by: Jason Xing <kernelxing@tencent.com>
Reviewed-by: Kuniyuki Iwashima <kuniyu@amazon.com>
Reviewed-by: Eric Dumazet <edumazet@google.com>
Reviewed-by: Willem de Bruijn <willemb@google.com>
Link: https://patch.msgid.link/20241005222609.94980-1-kerneljasonxing@gmail.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Add a cached FID bitmap. This mitigates the need to walk all VTU entries
to find the next free FID.
When flushing the VTU (during init), zero the FID bitmap. Use and
manipulate this bitmap from now on, instead of reading HW for the FID
map.
The repeated VTU walks are costly and can take ~40 mins if ~4000 vlans
are added. Caching the FID map reduces this time to <2 mins.
Signed-off-by: Aryan Srivastava <aryan.srivastava@alliedtelesis.co.nz>
Reviewed-by: Andrew Lunn <andrew@lunn.ch>
Link: https://patch.msgid.link/20241006212905.3142976-1-aryan.srivastava@alliedtelesis.co.nz
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
The sizeof(struct napi_struct) can change. Don't hardcode the size to
400 bytes and instead use "sizeof(struct napi_struct)".
Suggested-by: Alexander Lobakin <aleksander.lobakin@intel.com>
Signed-off-by: Joe Damato <jdamato@fastly.com>
Acked-by: Alexander Lobakin <aleksander.lobakin@intel.com>
Link: https://patch.msgid.link/20241004105407.73585-1-jdamato@fastly.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Kuniyuki Iwashima says:
====================
rtnetlink: Per-netns RTNL.
rtnl_lock() is a "Big Kernel Lock" in the networking slow path and
serialised all rtnetlink requests until 4.13.
Since RTNL_FLAG_DOIT_UNLOCKED and RTNL_FLAG_DUMP_UNLOCKED have been
introduced in 4.14 and 6.9, respectively, rtnetlink message handlers
are ready to be converted to RTNL-less/free.
15 out of 44 dumpit()s have been converted to RCU so far, and the
progress is pretty good. We can now dump various major network
resources without RTNL.
12 out of 87 doit()s have been converted, but most of the converted
doit()s are also on the reader side of RTNL; their message types are
RTM_GET*.
So, most of RTM_(NEW|DEL|SET)* operations are still serialised by RTNL.
For example, one of our services creates 2K netns and a small number
of network interfaces in each netns that require too many writer-side
rtnetlink requests, and setting up a single host takes 10+ minutes.
RTNL is still a huge pain for network configuration paths, and we need
more granular locking, given converting all doit()s would be unfeasible.
Actually, most RTNL users do not need to freeze multiple netns, and such
users can be protected by per-netns RTNL mutex. The exceptions would be
RTM_NEWLINK, RTM_DELLINK, and RTM_SETLINK. (See [0] and [1])
This series is the first step of the per-netns RTNL conversion that
gradually replaces rtnl_lock() with rtnl_net_lock(net) under
CONFIG_DEBUG_NET_SMALL_RTNL.
[0]: https://netdev.bots.linux.dev/netconf/2024/index.html
[1]: https://lpc.events/event/18/contributions/1959/
====================
Link: https://patch.msgid.link/20241004221031.77743-1-kuniyu@amazon.com
Signed-off-by: Paolo Abeni <pabeni@redhat.com>
The global and per-netns netdev notifier depend on RTNL, and its
dependency is not so clear due to nested calls.
Let's add a placeholder to place ASSERT_RTNL_NET() for each event.
Signed-off-by: Kuniyuki Iwashima <kuniyu@amazon.com>
Reviewed-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: Paolo Abeni <pabeni@redhat.com>
Once an RTNL scope is converted with rtnl_net_lock(), we will replace
RTNL helper functions inside the scope with the following per-netns
alternatives:
ASSERT_RTNL() -> ASSERT_RTNL_NET(net)
rcu_dereference_rtnl(p) -> rcu_dereference_rtnl_net(net, p)
Note that the per-netns helpers are equivalent to the conventional
helpers unless CONFIG_DEBUG_NET_SMALL_RTNL is enabled.
Signed-off-by: Kuniyuki Iwashima <kuniyu@amazon.com>
Reviewed-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: Paolo Abeni <pabeni@redhat.com>
The goal is to break RTNL down into per-netns mutex.
This patch adds per-netns mutex and its helper functions, rtnl_net_lock()
and rtnl_net_unlock().
rtnl_net_lock() acquires the global RTNL and per-netns RTNL mutex, and
rtnl_net_unlock() releases them.
We will replace 800+ rtnl_lock() with rtnl_net_lock() and finally removes
rtnl_lock() in rtnl_net_lock().
When we need to nest per-netns RTNL mutex, we will use __rtnl_net_lock(),
and its locking order is defined by rtnl_net_lock_cmp_fn() as follows:
1. init_net is first
2. netns address ascending order
Note that the conversion will be done under CONFIG_DEBUG_NET_SMALL_RTNL
with LOCKDEP so that we can carefully add the extra mutex without slowing
down RTNL operations during conversion.
Signed-off-by: Kuniyuki Iwashima <kuniyu@amazon.com>
Reviewed-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: Paolo Abeni <pabeni@redhat.com>
This reverts commit 464eb03c4a7cfb32cb3324249193cf6bb5b35152.
Once we have a per-netns RTNL, we won't use guard(rtnl).
Also, there's no users for now.
$ grep -rnI "guard(rtnl" || true
$
Suggested-by: Eric Dumazet <edumazet@google.com>
Link: https://lore.kernel.org/netdev/CANn89i+KoYzUH+VPLdGmLABYf5y4TW0hrM4UAeQQJ9AREty0iw@mail.gmail.com/
Signed-off-by: Kuniyuki Iwashima <kuniyu@amazon.com>
Reviewed-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: Paolo Abeni <pabeni@redhat.com>
Depending on the SoC where the FEC is integrated into the PPS channel
might be routed to different timer instances. Make this configurable
from the devicetree.
When the related DT property is not present fallback to the previous
default and use channel 0.
Reviewed-by: Frank Li <Frank.Li@nxp.com>
Tested-by: Rafael Beims <rafael.beims@toradex.com>
Signed-off-by: Francesco Dolcini <francesco.dolcini@toradex.com>
Reviewed-by: Csókás, Bence <csokas.bence@prolan.hu>
Signed-off-by: Paolo Abeni <pabeni@redhat.com>
Preparation patch to allow for PPS channel configuration, no functional
change intended.
Signed-off-by: Francesco Dolcini <francesco.dolcini@toradex.com>
Reviewed-by: Frank Li <Frank.Li@nxp.com>
Reviewed-by: Csókás, Bence <csokas.bence@prolan.hu>
Signed-off-by: Paolo Abeni <pabeni@redhat.com>
Add fsl,pps-channel property to select where to connect the PPS signal.
This depends on the internal SoC routing and on the board, for example
on the i.MX8 SoC it can be connected to an external pin (using channel 1)
or to internal eDMA as DMA request (channel 0).
Signed-off-by: Francesco Dolcini <francesco.dolcini@toradex.com>
Acked-by: Conor Dooley <conor.dooley@microchip.com>
Signed-off-by: Paolo Abeni <pabeni@redhat.com>
Daniel Machon says:
====================
net: sparx5: prepare for lan969x switch driver
== Description:
This series is the first of a multi-part series, that prepares and adds
support for the new lan969x switch driver.
The upstreaming efforts is split into multiple series (might change a
bit as we go along):
1) Prepare the Sparx5 driver for lan969x (this series)
2) Add support lan969x (same basic features as Sparx5 provides +
RGMII, excl. FDMA and VCAP)
3) Add support for lan969x FDMA
4) Add support for lan969x VCAP
== Lan969x in short:
The lan969x Ethernet switch family [1] provides a rich set of
switching features and port configurations (up to 30 ports) from 10Mbps
to 10Gbps, with support for RGMII, SGMII, QSGMII, USGMII, and USXGMII,
ideal for industrial & process automation infrastructure applications,
transport, grid automation, power substation automation, and ring &
intra-ring topologies. The LAN969x family is hardware and software
compatible and scalable supporting 46Gbps to 102Gbps switch bandwidths.
== Preparing Sparx5 for lan969x:
The lan969x switch chip reuses many of the IP's of the Sparx5 switch
chip, therefore it has been decided to add support through the existing
Sparx5 driver, in order to avoid a bunch of duplicate code. However, in
order to reuse the Sparx5 switch driver, we have to introduce some
mechanisms to handle the chip differences that are there. These
mechanisms are:
- Platform match data to contain all the differences that needs to
be handled (constants, ops etc.)
- Register macro indirection layer so that we can reuse the existing
register macros.
- Function for branching out on platform type where required.
In some places we ops out functions and in other places we branch on the
chip type. Exactly when we choose one over the other, is an estimate in
each case.
After this series is applied, the Sparx5 driver will be prepared for
lan969x and still function exactly as before.
== Patch breakdown:
Patch #1 adds private match data
Patch #2 adds register macro indirection layer
Patch #3-#4 does some preparation work
Patch #5-#7 adds chip constants and updates the code to use them
Patch #8-#13 adds and uses ops for handling functions differently on the
two platforms.
Patch #14 adds and uses a macro for branching out on the chip type.
Patch #15 (NEW) redefines macros for internal ports and PGID's.
[1] https://www.microchip.com/en-us/product/lan9698
To: David S. Miller <davem@davemloft.net>
To: Eric Dumazet <edumazet@google.com>
To: Jakub Kicinski <kuba@kernel.org>
To: Paolo Abeni <pabeni@redhat.com>
To: Lars Povlsen <lars.povlsen@microchip.com>
To: Steen Hegelund <Steen.Hegelund@microchip.com>
To: horatiu.vultur@microchip.com
To: jensemil.schulzostergaard@microchip.com
To: UNGLinuxDriver@microchip.com
To: Richard Cochran <richardcochran@gmail.com>
To: horms@kernel.org
To: justinstitt@google.com
To: gal@nvidia.com
To: aakash.r.menon@gmail.com
To: jacob.e.keller@intel.com
To: ast@fiberby.net
Cc: netdev@vger.kernel.org
Cc: linux-arm-kernel@lists.infradead.org
Cc: linux-kernel@vger.kernel.org
Signed-off-by: Daniel Machon <daniel.machon@microchip.com>
====================
Link: https://patch.msgid.link/20241004-b4-sparx5-lan969x-switch-driver-v2-0-d3290f581663@microchip.com
Signed-off-by: Paolo Abeni <pabeni@redhat.com>
Internal ports and PGID's are both defined relative to the number of
front ports on Sparx5. This will not work on lan969x. Instead make them
offsets to the number of front ports and add two helpers to retrieve
them. Use the helpers throughout.
Reviewed-by: Steen Hegelund <Steen.Hegelund@microchip.com>
Signed-off-by: Daniel Machon <daniel.machon@microchip.com>
Signed-off-by: Paolo Abeni <pabeni@redhat.com>
