Dragos Tatulea [Thu, 7 Nov 2024 19:43:56 +0000 (21:43 +0200)]
net/mlx5e: SHAMPO, Drop info array
The info array is used to store a pointer to the
dma address of the header and to the frag page. However,
this array is not really required:
- The frag page can be calculated from the header index
frag page index = header index / headers per page.
- The dma address can be calculated through a formula:
dma page address + header offset.
This series gets rid of the info array and uses the above
formulas instead.
The current_page_index was used in conjunction with the info array to
store page fragment indices. This variable is dropped as well.
Dragos Tatulea [Thu, 7 Nov 2024 19:43:55 +0000 (21:43 +0200)]
net/mlx5e: SHAMPO, Change frag page setup order during allocation
Now that the UMR allocation has been simplified, it is no longer
possible to have a leftover page from a previous call to
mlx5e_build_shampo_hd_umr().
This patch simplifies the code by switching the order of operations:
first take the frag page and then increment the index. This is more
straightforward and it also paves the way for dropping the info
array.
When calculating the index for the next frag page slot, the divisor is
incorrect: it should be the number of pages per queue not the number of
headers per queue. This is currently harmless because frag pages are not
used directly, but they are intermediated through the info array. But it
needs to be fixed as an upcoming patch will get rid of the info array.
This patch introduces a new pages per queue variable and plugs it in the
formula.
Now that this variable exists, additional code can be simplified in the
SHAMPO initialization code.
Dragos Tatulea [Thu, 7 Nov 2024 19:43:53 +0000 (21:43 +0200)]
net/mlx5e: SHAMPO, Simplify UMR allocation for headers
Allocating page fragments for header data split is currently
more complicated than it should be. That's because the number
of KSM entries allocated is not aligned to the number of headers
per page. This leads to having leftovers in the next allocation
which require additional accounting and needlessly complicated
code.
This patch aligns (down) the number of KSM entries in the
UMR WQE to the number of headers per page by:
1) Aligning the max number of entries allocated per UMR WQE
(max_ksm_entries) to MLX5E_SHAMPO_WQ_HEADER_PER_PAGE.
2) Aligning the total number of free headers to
MLX5E_SHAMPO_WQ_HEADER_PER_PAGE.
... and then it drops the extra accounting code from
mlx5e_build_shampo_hd_umr().
Although the number of entries allocated per UMR WQE is slightly
smaller due to aligning down, no performance impact was observed.
Carolina Jubran [Thu, 7 Nov 2024 19:43:52 +0000 (21:43 +0200)]
net/mlx5: Make vport QoS enablement more flexible for future extensions
Refactor esw_qos_vport_enable to support more generic configurations,
allowing it to be reused for new vport node types in future patches.
This refactor includes a new way to change the vport parent node by
disabling the current setup and re-enabling it with the new parent.
This change sets the foundation for adapting configuration based on the
parent type in future patches.
Carolina Jubran [Thu, 7 Nov 2024 19:43:51 +0000 (21:43 +0200)]
net/mlx5: Integrate esw_qos_vport_enable logic into rate operations
Fold the esw_qos_vport_enable function into operations for configuring
maximum and minimum rates, simplifying QoS logic. This change
consolidates enabling and updating the scheduling element
configuration, streamlining how vport QoS is initialized and adjusted.
Carolina Jubran [Thu, 7 Nov 2024 19:43:50 +0000 (21:43 +0200)]
net/mlx5: Generalize scheduling element operations
Introduce helper functions to create and destroy scheduling elements,
allowing flexible configuration for different scheduling element types.
The new helper functions streamline the process by centralizing error
handling and logging through esw_qos_sched_elem_op_warn, which now
accepts the operation type (create, destroy, or modify).
The changes also adjust the esw_qos_vport_enable and
mlx5_esw_qos_vport_disable functions to leverage the new generalized
create/destroy helpers.
The destroy functions now log errors with esw_warn without returning
them. This prevents unnecessary error handling since the node was
already destroyed and no further action is required from callers.
Carolina Jubran [Thu, 7 Nov 2024 19:43:49 +0000 (21:43 +0200)]
net/mlx5: Refactor scheduling element configuration bitmasks
Refactor esw_qos_sched_elem_config to set bitmasks only when max_rate
or bw_share values change, allowing the function to configure nodes
with only one of these parameters.
This enables more flexible usage for nodes where only one parameter
requires configuration.
Remove scattered assignments and checks to centralize them within this
function, removing the now redundant esw_qos_set_node_max_rate
entirely.
With this refactor, also remove the assignment of the vport scheduling
node max rate to the parent max rate for unlimited vports
(where max rate is set to zero), as firmware already handles this
behavior.
Carolina Jubran [Thu, 7 Nov 2024 19:43:47 +0000 (21:43 +0200)]
net/mlx5: Simplify QoS normalization by removing error handling
This change updates esw_qos_normalize_min_rate to not return errors,
significantly simplifying the code.
Normalization failures are software bugs, and it's unnecessary to
handle them with rollback mechanisms. Instead,
`esw_qos_update_sched_node_bw_share` and `esw_qos_normalize_min_rate`
now return void, with any errors logged as warnings to indicate
potential software issues.
This approach avoids compensating for hidden bugs and removes error
handling from all places that perform normalization, streamlining
future patches.
The E-switch mode was previously updated before removing and re-adding the
IB device, which could cause a temporary mismatch between the E-switch mode
and the IB device configuration.
To prevent this discrepancy, the IB device is now removed first, then
the E-switch mode is updated, and finally, the IB device is re-added.
This sequence ensures consistent alignment between the E-switch mode and
the IB device whenever the mode changes, regardless of the new mode value.
Vladimir Vdovin [Fri, 8 Nov 2024 09:34:24 +0000 (09:34 +0000)]
net: ipv4: Cache pmtu for all packet paths if multipath enabled
Check number of paths by fib_info_num_path(),
and update_or_create_fnhe() for every path.
Problem is that pmtu is cached only for the oif
that has received icmp message "need to frag",
other oifs will still try to use "default" iface mtu.
host1 have enabled multipath and
sysctl net.ipv4.fib_multipath_hash_policy = 1:
default proto static src 10.179.20.18
nexthop via 10.179.2.12 dev ens17f1 weight 1
nexthop via 10.179.2.140 dev ens17f0 weight 1
When host1 tries to do pmtud from 10.179.20.18/32 to host2,
host1 receives at ens17f1 iface an icmp packet from ro3 that ro3 mtu=1500.
And host1 caches it in nexthop exceptions cache.
Problem is that it is cached only for the iface that has received icmp,
and there is no way that ro3 will send icmp msg to host1 via another path.
Host1 now have this routes to host2:
ip r g 10.10.30.30 sport 30000 dport 443
10.10.30.30 via 10.179.2.12 dev ens17f1 src 10.179.20.18 uid 0
cache expires 521sec mtu 1500
ip r g 10.10.30.30 sport 30033 dport 443
10.10.30.30 via 10.179.2.140 dev ens17f0 src 10.179.20.18 uid 0
cache
So when host1 tries again to reach host2 with mtu>1500,
if packet flow is lucky enough to be hashed with oif=ens17f1 its ok,
if oif=ens17f0 it blackholes and still gets icmp msgs from ro3 to ens17f1,
until lucky day when ro3 will send it through another flow to ens17f0.
Breno Leitao [Fri, 8 Nov 2024 14:59:25 +0000 (06:59 -0800)]
net: netconsole: selftests: Check if netdevsim is available
The netconsole selftest relies on the availability of the netdevsim module.
To ensure the test can run correctly, we need to check if the netdevsim
module is either loaded or built-in before proceeding.
Update the netconsole selftest to check for the existence of
the /sys/bus/netdevsim/new_device file before running the test. If the
file is not found, the test is skipped with an explanation that the
CONFIG_NETDEVSIM kernel config option may not be enabled.
net: phylink: remove switch() statement in resolve handling
The switch() statement doesn't sit very well with the preceeding if()
statements, so let's just convert everything to if()s. As a result of
the two preceding commits, there is now only one case in the switch()
statement. Remove the switch statement and reduce the code indentation.
Code reformatting will be in the following commit.
net: phylink: move MLO_AN_PHY resolve handling to if() statement
The switch() statement doesn't sit very well with the preceeding if()
statements, and results in excessive indentation that spoils code
readability. Continue cleaning this up by converting the MLO_AN_PHY
case to use an if() statmeent.
net: phylink: move MLO_AN_FIXED resolve handling to if() statement
The switch() statement doesn't sit very well with the preceeding if()
statements, and results in excessive indentation that spoils code
readability. Begin cleaning this up by converting the MLO_AN_FIXED case
to an if() statement.
====================
Suspend IRQs during application busy periods
This series introduces a new mechanism, IRQ suspension, which allows
network applications using epoll to mask IRQs during periods of high
traffic while also reducing tail latency (compared to existing
mechanisms, see below) during periods of low traffic. In doing so, this
balances CPU consumption with network processing efficiency.
Martin Karsten (CC'd) and I have been collaborating on this series for
several months and have appreciated the feedback from the community on
our RFC [1]. We've updated the cover letter and kernel documentation in
an attempt to more clearly explain how this mechanism works, how
applications can use it, and how it compares to existing mechanisms in
the kernel.
I briefly mentioned this idea at netdev conf 2024 (for those who were
there) and Martin described this idea in an earlier paper presented at
Sigmetrics 2024 [2].
~ The short explanation (TL;DR)
We propose adding a new napi config parameter: irq_suspend_timeout to
help balance CPU usage and network processing efficiency when using IRQ
deferral and napi busy poll.
If this parameter is set to a non-zero value *and* a user application
has enabled preferred busy poll on a busy poll context (via the
EPIOCSPARAMS ioctl introduced in commit 18e2bf0edf4d ("eventpoll: Add
epoll ioctl for epoll_params")), then application calls to epoll_wait
for that context will cause device IRQs and softirq processing to be
suspended as long as epoll_wait successfully retrieves data from the
NAPI. Each time data is retrieved, the irq_suspend_timeout is deferred.
If/when network traffic subsides and epoll_wait returns no data, IRQ
suspension is immediately reverted back to the existing
napi_defer_hard_irqs and gro_flush_timeout mechanism which was
introduced in commit 6f8b12d661d0 ("net: napi: add hard irqs deferral
feature")).
The irq_suspend_timeout serves as a safety mechanism. If userland takes
a long time processing data, irq_suspend_timeout will fire and restart
normal NAPI processing.
For a more in depth explanation, please continue reading.
~ Comparison with existing mechanisms
Interrupt mitigation can be accomplished in napi software, by setting
napi_defer_hard_irqs and gro_flush_timeout, or via interrupt coalescing
in the NIC. This can be quite efficient, but in both cases, a fixed
timeout (or packet count) needs to be configured. However, a fixed
timeout cannot effectively support both low- and high-load situations:
At low load, an application typically processes a few requests and then
waits to receive more input data. In this scenario, a large timeout will
cause unnecessary latency.
At high load, an application typically processes many requests before
being ready to receive more input data. In this case, a small timeout
will likely fire prematurely and trigger irq/softirq processing, which
interferes with the application's execution. This causes overhead, most
likely due to cache contention.
While NICs attempt to provide adaptive interrupt coalescing schemes,
these cannot properly take into account application-level processing.
An alternative packet delivery mechanism is busy-polling, which results
in perfect alignment of application processing and network polling. It
delivers optimal performance (throughput and latency), but results in
100% cpu utilization and is thus inefficient for below-capacity
workloads.
We propose to add a new packet delivery mode that properly alternates
between busy polling and interrupt-based delivery depending on busy and
idle periods of the application. During a busy period, the system
operates in busy-polling mode, which avoids interference. During an idle
period, the system falls back to interrupt deferral, but with a small
timeout to avoid excessive latencies. This delivery mode can also be
viewed as an extension of basic interrupt deferral, but alternating
between a small and a very large timeout.
This delivery mode is efficient, because it avoids softirq execution
interfering with application processing during busy periods. It can be
used with blocking epoll_wait to conserve cpu cycles during idle
periods. The effect of alternating between busy and idle periods is that
performance (throughput and latency) is very close to full busy polling,
while cpu utilization is lower and very close to interrupt mitigation.
~ Usage details
IRQ suspension is introduced via a per-NAPI configuration parameter that
controls the maximum time that IRQs can be suspended.
Here's how it is intended to work:
- The user application (or system administrator) uses the netdev-genl
netlink interface to set the pre-existing napi_defer_hard_irqs and
gro_flush_timeout NAPI config parameters to enable IRQ deferral.
- The user application (or system administrator) sets the proposed
irq_suspend_timeout parameter via the netdev-genl netlink interface
to a larger value than gro_flush_timeout to enable IRQ suspension.
- The user application issues the existing epoll ioctl to set the
prefer_busy_poll flag on the epoll context.
- The user application then calls epoll_wait to busy poll for network
events, as it normally would.
- If epoll_wait returns events to userland, IRQs are suspended for the
duration of irq_suspend_timeout.
- If epoll_wait finds no events and the thread is about to go to
sleep, IRQ handling using napi_defer_hard_irqs and gro_flush_timeout
is resumed.
As long as epoll_wait is retrieving events, IRQs (and softirq
processing) for the NAPI being polled remain disabled. When network
traffic reduces, eventually a busy poll loop in the kernel will retrieve
no data. When this occurs, regular IRQ deferral using gro_flush_timeout
for the polled NAPI is re-enabled.
Unless IRQ suspension is continued by subsequent calls to epoll_wait, it
automatically times out after the irq_suspend_timeout timer expires.
Regular deferral is also immediately re-enabled when the epoll context
is destroyed.
~ Usage scenario
The target scenario for IRQ suspension as packet delivery mode is a
system that runs a dominant application with substantial network I/O.
The target application can be configured to receive input data up to a
certain batch size (via epoll_wait maxevents parameter) and this batch
size determines the worst-case latency that application requests might
experience. Because packet delivery is suspended during the target
application's processing, the batch size also determines the worst-case
latency of concurrent applications using the same RX queue(s).
gro_flush_timeout should be set as small as possible, but large enough to
make sure that a single request is likely not being interfered with.
irq_suspend_timeout is largely a safety mechanism against misbehaving
applications. It should be set large enough to cover the processing of an
entire application batch, i.e., the factor between gro_flush_timeout and
irq_suspend_timeout should roughly correspond to the maximum batch size
that the target application would process in one go.
~ Important call out in the implementation
- Enabling per epoll-context preferred busy poll will now effectively
lead to a nonblocking iteration through napi_busy_loop, even when
busy_poll_usecs is 0. See patch 4.
~ Benchmark configs & descriptions
The changes were benchmarked with memcached [3] using the benchmarking
tool mutilate [4].
To facilitate benchmarking, a small patch [5] was applied to memcached
1.6.29 to allow setting per-epoll context preferred busy poll and other
settings via environment variables. Another small patch [6] was applied
to libevent to enable full busy-polling.
Multiple scenarios were benchmarked as described below and the scripts
used for producing these results can be found on github [7] (note: all
scenarios use NAPI-based traffic splitting via SO_INCOMING_ID by passing
-N to memcached):
- base:
- no other options enabled
- deferX:
- set defer_hard_irqs to 100
- set gro_flush_timeout to X,000
- napibusy:
- set defer_hard_irqs to 100
- set gro_flush_timeout to 200,000
- enable busy poll via the existing ioctl (busy_poll_usecs = 64,
busy_poll_budget = 64, prefer_busy_poll = true)
- fullbusy:
- set defer_hard_irqs to 100
- set gro_flush_timeout to 5,000,000
- enable busy poll via the existing ioctl (busy_poll_usecs = 1000,
busy_poll_budget = 64, prefer_busy_poll = true)
- change memcached's nonblocking epoll_wait invocation (via
libevent) to using a 1 ms timeout
- suspend0:
- set defer_hard_irqs to 0
- set gro_flush_timeout to 0
- set irq_suspend_timeout to 20,000,000
- enable busy poll via the existing ioctl (busy_poll_usecs = 0,
busy_poll_budget = 64, prefer_busy_poll = true)
- suspendX:
- set defer_hard_irqs to 100
- set gro_flush_timeout to X,000
- set irq_suspend_timeout to 20,000,000
- enable busy poll via the existing ioctl (busy_poll_usecs = 0,
busy_poll_budget = 64, prefer_busy_poll = true)
~ Benchmark results
Tested on:
Single socket AMD EPYC 7662 64-Core Processor
Hyperthreading disabled
4 NUMA Zones (NPS=4)
16 CPUs per NUMA zone (64 cores total)
2 x Dual port 100gbps Mellanox Technologies ConnectX-5 Ex EN NIC
The test machine is configured such that a single interface has 8 RX
queues. The queues' IRQs and memcached are pinned to CPUs that are
NUMA-local to the interface which is under test. The NIC's interrupt
coalescing configuration is left at boot-time defaults.
Results:
Results are shown below. The mechanism added by this series is
represented by the 'suspend' cases. Data presented shows a summary over
nearly 10 runs of each test case [8] using the scripts on github [7].
For latency, the median is shown. For throughput and CPU utilization,
the average is shown.
The results also include cycles-per-query (cpq) and
instruction-per-query (ipq) metrics, following the methodology proposed
in [2], to augment the CPU utilization numbers, which could be skewed
due to frequency scaling. We find that this does not appear to be the
case as CPU utilization and low-level metrics show similar trends.
These results were captured using the scripts on github [7] to
illustrate how this approach compares with other pre-existing
mechanisms. This data is not to be interpreted as scientific data
captured in a fully isolated lab setting, but instead as best effort,
illustrative information comparing and contrasting tradeoffs.
The absolute QPS results shift between submissions, but the
relative differences are equivalent. As patches are rebased,
several factors likely influence overall performance.
Compare:
- Throughput (MAX) and latencies of base vs suspend.
- CPU usage of napibusy and fullbusy during lower load (200K, 400K for
example) vs suspend.
- Latency of the defer variants vs suspend as timeout and load
increases.
- suspend0, which sets defer_hard_irqs and gro_flush_timeout to 0, has
nearly the same performance as the base case (this is FAQ item #1).
The overall takeaway is that the suspend variants provide a superior
combination of high throughput, low latency, and low cpu utilization
compared to all other variants. Each of the suspend variants works very
well, but some fine-tuning between latency and cpu utilization is still
possible by tuning the small timeout (gro_flush_timeout).
Note: we've reorganized the results to make comparison among testcases
with the same load easier.
testcase load qps avglat 95%lat 99%lat cpu cpq ipq
base MAX 1037654 4184 5453 5810 100 8411 7938
defer10 MAX 905607 4840 6151 6380 100 9639 8431
defer20 MAX 986463 4455 5594 5796 100 8848 8110
defer50 MAX 1077030 4000 5073 5299 100 8104 7920
defer200 MAX 1040728 4152 5385 5765 100 8379 7849
fullbusy MAX 1247536 3518 3935 3984 100 6998 7930
napibusy MAX 1136310 3799 7756 9964 100 7670 7877
suspend0 MAX 1057509 4132 5724 6185 100 8253 7918
suspend10 MAX 1215147 3580 3957 4041 100 7185 7944
suspend20 MAX 1216469 3576 3953 3988 100 7175 7950
suspend50 MAX 1215871 3577 3961 4075 100 7181 7949
suspend200 MAX 1216882 3556 3951 3988 100 7175 7955
~ FAQ
- Why is a new parameter needed? Does irq_suspend_timeout override
gro_flush_timeout?
Using the suspend mechanism causes the system to alternate between
polling mode and irq-driven packet delivery. During busy periods,
irq_suspend_timeout overrides gro_flush_timeout and keeps the system
busy polling, but when epoll finds no events, the setting of
gro_flush_timeout and napi_defer_hard_irqs determine the next step.
There are essentially three possible loops for network processing and
packet delivery:
1) hardirq -> softirq -> napi poll; basic interrupt delivery
2) timer -> softirq -> napi poll; deferred irq processing
3) epoll -> busy-poll -> napi poll; busy looping
Loop 2 can take control from Loop 1, if gro_flush_timeout and
napi_defer_hard_irqs are set.
If gro_flush_timeout and napi_defer_hard_irqs are set, Loops 2 and
3 "wrestle" with each other for control. During busy periods,
irq_suspend_timeout is used as timer in Loop 2, which essentially
tilts this in favour of Loop 3.
If gro_flush_timeout and napi_defer_hard_irqs are not set, Loop 3
cannot take control from Loop 1.
Therefore, setting gro_flush_timeout and napi_defer_hard_irqs is the
recommended usage, because otherwise setting irq_suspend_timeout
might not have any discernible effect.
This is shown in the results above: compare suspend0 with the base
case. Note that the lack of napi_defer_hard_irqs and
gro_flush_timeout produce similar results for both, which encourages
the use of napi_defer_hard_irqs and gro_flush_timeout in addition to
irq_suspend_timeout.
- Can the new timeout value be threaded through the new epoll ioctl ?
It is possible, but presents challenges for userspace. User
applications must ensure that the file descriptors added to epoll
contexts have the same NAPI ID to support busy polling.
An epoll context is not permanently tied to any particular NAPI ID.
So, a user application could decide to clear the file descriptors
from the context and add a new set of file descriptors with a
different NAPI ID to the context. Busy polling would work as
expected, but the meaning of the suspend timeout becomes ambiguous
because IRQs are not inherently associated with epoll contexts, but
rather with the NAPI. The user program would need to reissue the
ioctl to set the irq_suspend_timeout, but the napi_defer_hard_irqs
and gro_flush_timeout settings would come from the NAPI's
napi_config (which are set either by sysfs or by netlink). Such an
interface seems awkard to use from a user perspective.
Further, IRQs are related to NAPIs, which is why they are stored in
the napi_config space. Putting the irq_suspend_timeout in
the epoll context while other IRQ deferral mechanisms remain in the
NAPI's napi_config space seems like an odd design choice.
We've opted to keep all of the IRQ deferral parameters together and
place the irq_suspend_timeout in napi_config. This has nice benefits
for userspace: if a user app were to remove all file descriptors
from an epoll context and add new file descriptors with a new NAPI ID,
the correct suspend timeout for that NAPI ID would be used automatically
without the user application needing to do anything (like re-issuing an
ioctl, for example). All IRQ deferral related parameters are in one
place and can all be set the same way: with netlink.
- Can irq suspend be built by combining NIC coalescing and
gro_flush_timeout ?
No. The problem is that the long timeout must engage if and only if
prefer-busy is active.
When using NIC coalescing for the short timeout (without
napi_defer_hard_irqs/gro_flush_timeout), an interrupt after an idle
period will trigger softirq, which will run napi polling. At this
point, prefer-busy is not active, so NIC interrupts would be
re-enabled. Then it is not possible for the longer timeout to
interject to switch control back to polling. In other words, only by
using the software timer for the short timeout, it is possible to
extend the timeout without having to reprogram the NIC timer or
reach down directly and disable interrupts.
Using gro_flush_timeout for the long timeout also has problems, for
the same underlying reason. In the current napi implementation,
gro_flush_timeout is not tied to prefer-busy. We'd either have to
change that and in the process modify the existing deferral
mechanism, or introduce a state variable to determine whether
gro_flush_timeout is used as long timeout for irq suspend or whether
it is used for its default purpose. In an earlier version, we did
try something similar to the latter and made it work, but it ends up
being a lot more convoluted than our current proposal.
- Isn't it already possible to combine busy looping with irq deferral?
Yes, in fact enabling irq deferral via napi_defer_hard_irqs and
gro_flush_timeout is a precondition for prefer_busy_poll to have an
effect. If the application also uses a tight busy loop with
essentially nonblocking epoll_wait (accomplished with a very short
timeout parameter), this is the fullbusy case shown in the results.
An application using blocking epoll_wait is shown as the napibusy
case in the results. It's a hybrid approach that provides limited
latency benefits compared to the base case and plain irq deferral,
but not as good as fullbusy or suspend.
~ Special thanks
Several people were involved in earlier stages of the development of this
mechanism whom we'd like to thank:
- Peter Cai (CC'd), for the initial kernel patch and his contributions
to the paper.
- Mohammadamin Shafie (CC'd), for testing various versions of the kernel
patch and providing helpful feedback.
Joe Damato [Sat, 9 Nov 2024 05:02:35 +0000 (05:02 +0000)]
selftests: net: Add busy_poll_test
Add an epoll busy poll test using netdevsim.
This test is comprised of:
- busy_poller (via busy_poller.c)
- busy_poll_test.sh which loads netdevsim, sets up network namespaces,
and runs busy_poller to receive data and socat to send data.
The selftest tests two different scenarios:
- busy poll (the pre-existing version in the kernel)
- busy poll with suspend enabled (what this series adds)
The data transmit is a 1MiB temporary file generated from /dev/urandom
and the test is considered passing if the md5sum of the input file to
socat matches the md5sum of the output file from busy_poller.
netdevsim was chosen instead of veth due to netdevsim's support for
netdev-genl.
For now, this test uses the functionality that netdevsim provides. In the
future, perhaps netdevsim can be extended to emulate device IRQs to more
thoroughly test all pre-existing kernel options (like defer_hard_irqs)
and suspend.
Add a per-NAPI IRQ suspension parameter, which can be get/set with
netdev-genl.
This patch doesn't change any behavior but prepares the code for other
changes in the following commits which use irq_suspend_timeout as a
timeout for IRQ suspension.
Vadim Fedorenko [Thu, 7 Nov 2024 21:49:17 +0000 (13:49 -0800)]
bnxt_en: add unlocked version of bnxt_refclk_read
Serialization of PHC read with FW reset mechanism uses ptp_lock which
also protects timecounter updates. This means we cannot grab it when
called from bnxt_cc_read(). Let's move locking into different function.
====================
rtnetlink: Convert rtnl_newlink() to per-netns RTNL.
Patch 1 - 3 removes __rtnl_link_unregister and protect link_ops by
its dedicated mutex to move synchronize_srcu() out of RTNL scope.
Patch 4 introduces struct rtnl_nets and helper functions to acquire
multiple per-netns RTNL in rtnl_newlink().
Patch 5 - 8 are to prefetch the peer device's netns in rtnl_newlink().
Patch 9 converts rtnl_newlink() to per-netns RTNL.
Patch 10 pushes RTNL down to rtnl_dellink() and rtnl_setlink(), but
the conversion will not be completed unless we support cases with
peer/upper/lower devices.
I confirmed v3 survived ./rtnetlink.sh; rmmod netdevsim.ko; without
lockdep splat.
rtnetlink: Register rtnl_dellink() and rtnl_setlink() with RTNL_FLAG_DOIT_PERNET_WIP.
Currently, rtnl_setlink() and rtnl_dellink() cannot be fully converted
to per-netns RTNL due to a lack of handling peer/lower/upper devices in
different netns.
For example, when we change a device in rtnl_setlink() and need to
propagate that to its upper devices, we want to avoid acquiring all netns
locks, for which we do not know the upper limit.
The same situation happens when we remove a device.
rtnl_dellink() could be transformed to remove a single device in the
requested netns and delegate other devices to per-netns work, and
rtnl_setlink() might be ?
Until we come up with a better idea, let's use a new flag
RTNL_FLAG_DOIT_PERNET_WIP for rtnl_dellink() and rtnl_setlink().
This will unblock converting RTNL users where such devices are not related.
rtnetlink: Introduce struct rtnl_nets and helpers.
rtnl_newlink() needs to hold 3 per-netns RTNL: 2 for a new device
and 1 for its peer.
We will add rtnl_nets_lock() later, which performs the nested locking
based on struct rtnl_nets, which has an array of struct net pointers.
rtnl_nets_add() adds a net pointer to the array and sorts it so that
rtnl_nets_lock() can simply acquire per-netns RTNL from array[0] to [2].
Before calling rtnl_nets_add(), get_net() must be called for the net,
and rtnl_nets_destroy() will call put_net() for each.
Let's apply the helpers to rtnl_newlink().
When CONFIG_DEBUG_NET_SMALL_RTNL is disabled, we do not call
rtnl_net_lock() thus do not care about the array order, so
rtnl_net_cmp_locks() returns -1 so that the loop in rtnl_nets_add()
can be optimised to NOP.
rtnl_link_unregister() holds RTNL and calls __rtnl_link_unregister(),
where we call synchronize_srcu() to wait inflight RTM_NEWLINK requests
for per-netns RTNL.
We put synchronize_srcu() in __rtnl_link_unregister() due to ifb.ko
and dummy.ko.
However, rtnl_newlink() will acquire SRCU before RTNL later in this
series. Then, lockdep will detect the deadlock:
====================
selftests: ncdevmem: Add ncdevmem to ksft
The goal of the series is to simplify and make it possible to use
ncdevmem in an automated way from the ksft python wrapper.
ncdevmem is slowly mutated into a state where it uses stdout
to print the payload and the python wrapper is added to
make sure the arrived payload matches the expected one.
====================
Only RX side for now and small message to test the setup.
In the future, we can extend it to TX side and to testing
both sides with a couple of megs of data.
make \
-C tools/testing/selftests \
TARGETS="drivers/hw/net" \
install INSTALL_PATH=~/tmp/ksft
selftests: ncdevmem: Run selftest when none of the -s or -c has been provided
This will be used as a 'probe' mode in the selftest to check whether
the device supports the devmem or not. Use hard-coded queue layout
(two last queues) and prevent user from passing custom -q and/or -t.
selftests: ncdevmem: Use YNL to enable TCP header split
In the next patch the hard-coded queue numbers are gonna be removed.
So introduce some initial support for ethtool YNL and use
it to enable header split.
Also, tcp-data-split requires latest ethtool which is unlikely
to be present in the distros right now.
(ideally, we should not shell out to ethtool at all).
====================
net: stmmac: dwmac4: Fixes issues in dwmac4
This patch series fixes issues in the dwmac4 driver. These three patches
don't cause any user-visible issues, so they are targeted for net-next.
Patch #1:
Corrects the masking logic in the MTL Operation Mode RTC mask and shift
macros. The current code lacks the use of the ~ operator, which is
necessary to clear the bits properly.
Patch #2:
Addresses inaccuracies in the MTL_OP_MODE_*_MASK macros. The RTC fields
are located in bits [1:0], and this patch ensures the mask and shift
macros use the appropriate values to reflect this.
Patch #3:
Moves the handling of the Receive Watchdog Timeout (RWT) out of the
Abnormal Interrupt Summary (AIS) condition. According to the databook,
the RWT interrupt is not included in the AIS.
Ley Foon Tan [Thu, 7 Nov 2024 06:36:36 +0000 (14:36 +0800)]
net: stmmac: dwmac4: Receive Watchdog Timeout is not in abnormal interrupt summary
The Receive Watchdog Timeout (RWT, bit[9]) is not part of Abnormal
Interrupt Summary (AIS). Move the RWT handling out of the AIS
condition statement.
From databook, the AIS is the logical OR of the following interrupt bits:
- Bit 1: Transmit Process Stopped
- Bit 7: Receive Buffer Unavailable
- Bit 8: Receive Process Stopped
- Bit 10: Early Transmit Interrupt
- Bit 12: Fatal Bus Error
- Bit 13: Context Descriptor Error
net: ti: icssg-prueth: Add VLAN support for HSR mode
Add support for VLAN addition/deletion in HSR mode.
In HSR mode, even if the host port is not a member of
the VLAN domain, the slave ports should simply forward the
frames. So allow forwarding of all VLAN frames in HSR mode.
This patch adds support for VLAN ctag based filtering at slave devices.
The slave ethernet device may be capable of filtering ethernet packets
based on VLAN ID. This requires that when the VLAN interface is created
over an HSR/PRP interface, it passes the VID information to the
associated slave ethernet devices so that it updates the hardware
filters to filter ethernet frames based on VID. This patch adds the
required functions to propagate the vid information to the slave
devices.
====================
Side MDIO Support for LAN937x Switches
This patch set introduces support for an internal MDIO bus in LAN937x
switches, enabling the use of a side MDIO channel for PHY management
while keeping SPI as the main interface for switch configuration.
other changelogs are added to separate patches.
====================
Oleksij Rempel [Wed, 6 Nov 2024 07:59:41 +0000 (08:59 +0100)]
net: dsa: microchip: parse PHY config from device tree
Introduce ksz_parse_dt_phy_config() to validate and parse PHY
configuration from the device tree for KSZ switches. This function
ensures proper setup of internal PHYs by checking `phy-handle`
properties, verifying expected PHY IDs, and handling parent node
mismatches. Sets the PHY mask on the MII bus if validation is
successful. Returns -EINVAL on configuration errors.
Oleksij Rempel [Wed, 6 Nov 2024 07:59:40 +0000 (08:59 +0100)]
net: dsa: microchip: add support for side MDIO interface in LAN937x
Implement side MDIO channel support for LAN937x switches, providing an
alternative to SPI for PHY management alongside existing SPI-based
switch configuration. This is needed to reduce SPI load, as SPI can be
relatively expensive for small packets compared to MDIO support.
Also, implemented static mappings for PHY addresses for various LAN937x
models to support different internal PHY configurations. Since the PHY
address mappings are not equal to the port indexes, this patch also
provides PHY address calculation based on hardware strapping
configuration.
Oleksij Rempel [Wed, 6 Nov 2024 07:59:38 +0000 (08:59 +0100)]
net: dsa: microchip: Refactor MDIO handling for side MDIO access
Add support for accessing PHYs via a side MDIO interface in LAN937x
switches. The existing code already supports accessing PHYs via main
management interfaces, which can be SPI, I2C, or MDIO, depending on the
chip variant. This patch enables using a side MDIO bus, where SPI is
used for the main switch configuration and MDIO for managing the
integrated PHYs. On LAN937x, this is optional, allowing them to operate
in both configurations: SPI only, or SPI + MDIO. Typically, the SPI
interface is used for switch configuration, while MDIO handles PHY
management.
Additionally, update interrupt controller code to support non-linear
port to PHY address mapping, enabling correct interrupt handling for
configurations where PHY addresses do not directly correspond to port
indexes. This change ensures that the interrupt mechanism properly
aligns with the new, flexible PHY address mappings introduced by side
MDIO support.
Oleksij Rempel [Wed, 6 Nov 2024 07:59:37 +0000 (08:59 +0100)]
dt-bindings: net: dsa: microchip: add mdio-parent-bus property for internal MDIO
Introduce `mdio-parent-bus` property in the ksz DSA bindings to
reference the parent MDIO bus when the internal MDIO bus is attached to
it, bypassing the main management interface.
Mohammad Heib [Thu, 7 Nov 2024 12:07:39 +0000 (14:07 +0200)]
net: atlantic: use irq_update_affinity_hint()
irq_set_affinity_hint() is deprecated, Use irq_update_affinity_hint()
instead. This removes the side-effect of actually applying the affinity.
The driver does not really need to worry about spreading its IRQs across
CPUs. The core code already takes care of that. when the driver applies the
affinities by itself, it breaks the users' expectations:
1. The user configures irqbalance with IRQBALANCE_BANNED_CPULIST in
order to prevent IRQs from being moved to certain CPUs that run a
real-time workload.
2. atlantic device reopening will resets the affinity
in aq_ndev_open().
3. atlantic has no idea about irqbalance's config, so it may move an IRQ to
a banned CPU. The real-time workload suffers unacceptable latency.
Mohammad Heib [Thu, 7 Nov 2024 11:50:02 +0000 (13:50 +0200)]
nfp: use irq_update_affinity_hint()
irq_set_affinity_hint() is deprecated, Use irq_update_affinity_hint()
instead. This removes the side-effect of actually applying the affinity.
The driver does not really need to worry about spreading its IRQs across
CPUs. The core code already takes care of that. when the driver applies the
affinities by itself, it breaks the users' expectations:
1. The user configures irqbalance with IRQBALANCE_BANNED_CPULIST in
order to prevent IRQs from being moved to certain CPUs that run a
real-time workload.
2. nfp device reopening will resets the affinity
in nfp_net_netdev_open().
3. nfp has no idea about irqbalance's config, so it may move an IRQ to
a banned CPU. The real-time workload suffers unacceptable latency.
Mohammad Heib [Wed, 6 Nov 2024 18:08:11 +0000 (20:08 +0200)]
bnxt_en: use irq_update_affinity_hint()
irq_set_affinity_hint() is deprecated, Use irq_update_affinity_hint()
instead. This removes the side-effect of actually applying the affinity.
The driver does not really need to worry about spreading its IRQs across
CPUs. The core code already takes care of that. when the driver applies the
affinities by itself, it breaks the users' expectations:
1. The user configures irqbalance with IRQBALANCE_BANNED_CPULIST in
order to prevent IRQs from being moved to certain CPUs that run a
real-time workload.
2. bnxt_en device reopening will resets the affinity
in bnxt_open().
3. bnxt_en has no idea about irqbalance's config, so it may move an IRQ to
a banned CPU. The real-time workload suffers unacceptable latency.
ipv6: Fix soft lockups in fib6_select_path under high next hop churn
Soft lockups have been observed on a cluster of Linux-based edge routers
located in a highly dynamic environment. Using the `bird` service, these
routers continuously update BGP-advertised routes due to frequently
changing nexthop destinations, while also managing significant IPv6
traffic. The lockups occur during the traversal of the multipath
circular linked-list in the `fib6_select_path` function, particularly
while iterating through the siblings in the list. The issue typically
arises when the nodes of the linked list are unexpectedly deleted
concurrently on a different core—indicated by their 'next' and
'previous' elements pointing back to the node itself and their reference
count dropping to zero. This results in an infinite loop, leading to a
soft lockup that triggers a system panic via the watchdog timer.
Apply RCU primitives in the problematic code sections to resolve the
issue. Where necessary, update the references to fib6_siblings to
annotate or use the RCU APIs.
Include a test script that reproduces the issue. The script
periodically updates the routing table while generating a heavy load
of outgoing IPv6 traffic through multiple iperf3 clients. It
consistently induces infinite soft lockups within a couple of minutes.
The logic of eq_update_ci() is duplicated in mlx5_eq_update_ci(). The
only additional work done by mlx5_eq_update_ci() is to increment
eq->cons_index. Call eq_update_ci() from mlx5_eq_update_ci() to avoid
the duplication.
The memory barrier in eq_update_ci() after the doorbell write is a
significant hot spot in mlx5_eq_comp_int(). Under heavy TCP load, we see
3% of CPU time spent on the mfence instruction.
98df6d5b877c ("net/mlx5: A write memory barrier is sufficient in EQ ci
update") already relaxed the full memory barrier to just a write barrier
in mlx5_eq_update_ci(), which duplicates eq_update_ci(). So replace mb()
with wmb() in eq_update_ci() too.
On strongly ordered architectures, no barrier is actually needed because
the MMIO writes to the doorbell register are guaranteed to appear to the
device in the order they were made. However, the kernel's ordered MMIO
primitive writel() lacks a convenient big-endian interface.
Therefore, we opt to stick with __raw_writel() + a barrier.
====================
macsec: inherit lower device's features and TSO limits when offloading
When macsec is offloaded to a NIC, we can take advantage of some of
its features, mainly TSO and checksumming. This increases performance
significantly. Some features cannot be inherited, because they require
additional ops that aren't provided by the macsec netdevice.
We also need to inherit TSO limits from the lower device, like
VLAN/macvlan devices do.
This series also moves the existing macsec offload selftest to the
netdevsim selftests before adding tests for the new features. To allow
this new selftest to work, netdevsim's hw_features are expanded.
====================
Sabrina Dubroca [Wed, 6 Nov 2024 23:13:34 +0000 (00:13 +0100)]
selftests: netdevsim: add ethtool features to macsec offload tests
The test verifies that available features aren't changed by toggling
offload on the device. Creating a device with offload off and then
enabling it later should result in the same features as creating the
device with offload enabled directly.
Sabrina Dubroca [Wed, 6 Nov 2024 23:13:29 +0000 (00:13 +0100)]
macsec: add some of the lower device's features when offloading
This commit extends the set of netdevice features supported by macsec
devices when offload is enabled, which increases performance
significantly (for a single TCP stream: 17.5Gbps to 38.5Gbps on my
test machines).
Commit c850240b6c41 ("net: macsec: report real_dev features when HW
offloading is enabled") previously attempted something similar, but
had to be reverted (commit 8bcd560ae878 ("Revert "net: macsec: report
real_dev features when HW offloading is enabled"")) because the set of
features it exposed was too large.
During initialization, all features are set, and they're then removed
via ndo_fix_features (macsec_fix_features). This allows the
offloadable features to be automatically enabled if offloading is
turned on after device creation.
Sabrina Dubroca [Wed, 6 Nov 2024 23:13:27 +0000 (00:13 +0100)]
netdevsim: add more hw_features
netdevsim currently only set HW_TC in its hw_features, but other
features should also be present to better reflect the behavior of real
HW.
In my macsec offload testing, this ends up as HW_CSUM being missing
from hw_features, so it doesn't stick in wanted_features when offload
is turned off. Then HW_CSUM (and thus TSO, thanks to
netdev_fix_features) is not automatically turned back on when offload
is re-enabled.
====================
Replace page_frag with page_frag_cache (Part-1)
This is part 1 of "Replace page_frag with page_frag_cache",
which mainly contain refactoring and optimization for the
implementation of page_frag API before the replacing.
As the discussion in [1], it would be better to target net-next
tree to get more testing as all the callers page_frag API are
in networking, and the chance of conflicting with MM tree seems
low as implementation of page_frag API seems quite self-contained.
After [2], there are still two implementations for page frag:
1. mm/page_alloc.c: net stack seems to be using it in the
rx part with 'struct page_frag_cache' and the main API
being page_frag_alloc_align().
2. net/core/sock.c: net stack seems to be using it in the
tx part with 'struct page_frag' and the main API being
skb_page_frag_refill().
This patchset tries to unfiy the page frag implementation
by replacing page_frag with page_frag_cache for sk_page_frag()
first. net_high_order_alloc_disable_key for the implementation
in net/core/sock.c doesn't seems matter that much now as pcp
is also supported for high-order pages:
commit 44042b449872 ("mm/page_alloc: allow high-order pages to
be stored on the per-cpu lists")
As the related change is mostly related to networking, so
targeting the net-next. And will try to replace the rest
of page_frag in the follow patchset.
After this patchset:
1. Unify the page frag implementation by taking the best out of
two the existing implementations: we are able to save some space
for the 'page_frag_cache' API user, and avoid 'get_page()' for
the old 'page_frag' API user.
2. Future bugfix and performance can be done in one place, hence
improving maintainability of page_frag's implementation.
Kernel Image changing:
Linux Kernel total | text data bss
------------------------------------------------------
after 45250307 | 2727427917209996 766032
before 45254134 | 2727811817209984 766032
delta -3827 | -3839 +12 +0
Yunsheng Lin [Mon, 28 Oct 2024 11:53:42 +0000 (19:53 +0800)]
mm: page_frag: use __alloc_pages() to replace alloc_pages_node()
It seems there is about 24Bytes binary size increase for
__page_frag_cache_refill() after refactoring in arm64 system
with 64K PAGE_SIZE. By doing the gdb disassembling, It seems
we can have more than 100Bytes decrease for the binary size
by using __alloc_pages() to replace alloc_pages_node(), as
there seems to be some unnecessary checking for nid being
NUMA_NO_NODE, especially when page_frag is part of the mm
system.
Yunsheng Lin [Mon, 28 Oct 2024 11:53:41 +0000 (19:53 +0800)]
mm: page_frag: reuse existing space for 'size' and 'pfmemalloc'
Currently there is one 'struct page_frag' for every 'struct
sock' and 'struct task_struct', we are about to replace the
'struct page_frag' with 'struct page_frag_cache' for them.
Before begin the replacing, we need to ensure the size of
'struct page_frag_cache' is not bigger than the size of
'struct page_frag', as there may be tens of thousands of
'struct sock' and 'struct task_struct' instances in the
system.
By or'ing the page order & pfmemalloc with lower bits of
'va' instead of using 'u16' or 'u32' for page size and 'u8'
for pfmemalloc, we are able to avoid 3 or 5 bytes space waste.
And page address & pfmemalloc & order is unchanged for the
same page in the same 'page_frag_cache' instance, it makes
sense to fit them together.
After this patch, the size of 'struct page_frag_cache' should be
the same as the size of 'struct page_frag'.
Yunsheng Lin [Mon, 28 Oct 2024 11:53:40 +0000 (19:53 +0800)]
xtensa: remove the get_order() implementation
As the get_order() implemented by xtensa supporting 'nsau'
instruction seems be the same as the generic implementation
in include/asm-generic/getorder.h when size is not a constant
value as the generic implementation calling the fls*() is also
utilizing the 'nsau' instruction for xtensa.
So remove the get_order() implemented by xtensa, as using the
generic implementation may enable the compiler to do the
computing when size is a constant value instead of runtime
computing and enable the using of get_order() in BUILD_BUG_ON()
macro in next patch.
Yunsheng Lin [Mon, 28 Oct 2024 11:53:38 +0000 (19:53 +0800)]
mm: page_frag: use initial zero offset for page_frag_alloc_align()
We are about to use page_frag_alloc_*() API to not just
allocate memory for skb->data, but also use them to do
the memory allocation for skb frag too. Currently the
implementation of page_frag in mm subsystem is running
the offset as a countdown rather than count-up value,
there may have several advantages to that as mentioned
in [1], but it may have some disadvantages, for example,
it may disable skb frag coalescing and more correct cache
prefetching
We have a trade-off to make in order to have a unified
implementation and API for page_frag, so use a initial zero
offset in this patch, and the following patch will try to
make some optimization to avoid the disadvantages as much
as possible.
Yunsheng Lin [Mon, 28 Oct 2024 11:53:37 +0000 (19:53 +0800)]
mm: move the page fragment allocator from page_alloc into its own file
Inspired by [1], move the page fragment allocator from page_alloc
into its own c file and header file, as we are about to make more
change for it to replace another page_frag implementation in
sock.c
As this patchset is going to replace 'struct page_frag' with
'struct page_frag_cache' in sched.h, including page_frag_cache.h
in sched.h has a compiler error caused by interdependence between
mm_types.h and mm.h for asm-offsets.c, see [2]. So avoid the compiler
error by moving 'struct page_frag_cache' to mm_types_task.h as
suggested by Alexander, see [3].
Yunsheng Lin [Mon, 28 Oct 2024 11:53:36 +0000 (19:53 +0800)]
mm: page_frag: add a test module for page_frag
The testing is done by ensuring that the fragment allocated
from a frag_frag_cache instance is pushed into a ptr_ring
instance in a kthread binded to a specified cpu, and a kthread
binded to a specified cpu will pop the fragment from the
ptr_ring and free the fragment.
Johannes Berg [Fri, 8 Nov 2024 10:41:45 +0000 (11:41 +0100)]
net: convert to nla_get_*_default()
Most of the original conversion is from the spatch below,
but I edited some and left out other instances that were
either buggy after conversion (where default values don't
fit into the type) or just looked strange.
Ido Schimmel [Tue, 5 Nov 2024 13:39:54 +0000 (15:39 +0200)]
bridge: Allow deleting FDB entries with non-existent VLAN
It is currently impossible to delete individual FDB entries (as opposed
to flushing) that were added with a VLAN that no longer exists:
# ip link add name dummy1 up type dummy
# ip link add name br1 up type bridge vlan_filtering 1
# ip link set dev dummy1 master br1
# bridge fdb add 00:11:22:33:44:55 dev dummy1 master static vlan 1
# bridge vlan del vid 1 dev dummy1
# bridge fdb get 00:11:22:33:44:55 br br1 vlan 1
00:11:22:33:44:55 dev dummy1 vlan 1 master br1 static
# bridge fdb del 00:11:22:33:44:55 dev dummy1 master vlan 1
RTNETLINK answers: Invalid argument
# bridge fdb get 00:11:22:33:44:55 br br1 vlan 1
00:11:22:33:44:55 dev dummy1 vlan 1 master br1 static
This is in contrast to MDB entries that can be deleted after the VLAN
was deleted:
# bridge vlan add vid 10 dev dummy1
# bridge mdb add dev br1 port dummy1 grp 239.1.1.1 permanent vid 10
# bridge vlan del vid 10 dev dummy1
# bridge mdb get dev br1 grp 239.1.1.1 vid 10
dev br1 port dummy1 grp 239.1.1.1 permanent vid 10
# bridge mdb del dev br1 port dummy1 grp 239.1.1.1 permanent vid 10
# bridge mdb get dev br1 grp 239.1.1.1 vid 10
Error: bridge: MDB entry not found.
Align the two interfaces and allow user space to delete FDB entries that
were added with a VLAN that no longer exists:
# ip link add name dummy1 up type dummy
# ip link add name br1 up type bridge vlan_filtering 1
# ip link set dev dummy1 master br1
# bridge fdb add 00:11:22:33:44:55 dev dummy1 master static vlan 1
# bridge vlan del vid 1 dev dummy1
# bridge fdb get 00:11:22:33:44:55 br br1 vlan 1
00:11:22:33:44:55 dev dummy1 vlan 1 master br1 static
# bridge fdb del 00:11:22:33:44:55 dev dummy1 master vlan 1
# bridge fdb get 00:11:22:33:44:55 br br1 vlan 1
Error: Fdb entry not found.
Add a selftest to make sure this behavior does not regress:
# ./rtnetlink.sh -t kci_test_fdb_del
PASS: bridge fdb del
projects / linux.git / log