- Linux-kselftest-mirror - lists.linaro.org

[PATCH net-next v5 0/7] Suspend IRQs during application busy periods

by Joe Damato

Greetings: Welcome to v5, see changelog below. Note that our performance tests were not re-run for this revision as we updated a commit message, fixed typos, removed a short unnecessary paragraph from the documentation and added a very minor functional change to return without suspended IRQs in certain error cases. 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. We've added an additional test case, 'fullbusy', achieved by modifying libevent for comparison. See below for a detailed description, link to the patch, and test results. 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. ~ Design rationale The implementation of the IRQ suspension mechanism very nicely dovetails with the existing mechanism for IRQ deferral when preferred busy poll is enabled (introduced in commit 7fd3253a7de6 ("net: Introduce preferred busy-polling"), see that commit message for more details). While it would be possible to inject the suspend timeout via the existing epoll ioctl, it is more natural to avoid this path for one main reason: An epoll context is linked to NAPI IDs as file descriptors are added; this means any epoll context might suddenly be associated with a different net_device if the application were to replace all existing fds with fds from a different device. In this case, the scope of the suspend timeout becomes unclear and many edge cases for both the user application and the kernel are introduced Only a single iteration through napi busy polling is needed for this mechanism to work effectively. Since an important objective for this mechanism is preserving cpu cycles, exactly one iteration of the napi busy loop is invoked when busy_poll_usecs is set to 0. ~ 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 - 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 at least 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 are higher than our previous submission, but the relative differences between variants are equivalent. Because the patches have been rebased on 6.12, several factors have likely influenced the overall performance. Most importantly, we had to switch to a new set of basic kernel options, which has likely altered the baseline performance. Because the overall comparison of variants still holds, we have not attempted to recreate the exact set of kernel options from the previous submission. 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. 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 200K 200024 127 254 458 25 12748 11289 defer10 200K 199991 64 128 166 27 18763 16574 defer20 200K 199986 72 135 178 25 15405 14173 defer50 200K 200025 91 149 198 23 12275 12203 defer200 200K 199996 182 266 326 18 8595 9183 fullbusy 200K 200040 58 123 167 100 43641 23145 napibusy 200K 200009 115 244 299 56 24797 24693 suspend10 200K 200005 63 128 167 32 19559 17240 suspend20 200K 199952 69 132 170 29 16324 14838 suspend50 200K 200019 84 144 189 26 13106 12516 suspend200 200K 199978 168 264 326 20 9331 9643 testcase load qps avglat 95%lat 99%lat cpu cpq ipq base 400K 400017 157 292 762 39 9287 9325 defer10 400K 400033 71 141 204 53 13950 12943 defer20 400K 399935 79 150 212 47 12027 11673 defer50 400K 399888 101 171 231 39 9556 9921 defer200 400K 399993 200 287 358 32 7428 8576 fullbusy 400K 400018 63 132 203 100 21827 16062 napibusy 400K 399970 89 230 292 83 18156 16508 suspend10 400K 400061 69 139 202 54 13576 13057 suspend20 400K 399988 73 144 206 49 11930 11773 suspend50 400K 399975 88 161 218 42 9996 10270 suspend200 400K 399954 172 276 353 34 7847 8713 testcase load qps avglat 95%lat 99%lat cpu cpq ipq base 600K 600031 166 289 631 61 9188 8787 defer10 600K 599967 85 167 262 75 11833 10947 defer20 600K 599888 89 165 243 66 10513 10362 defer50 600K 600072 109 185 253 55 8664 9190 defer200 600K 599951 222 315 393 45 6892 8213 fullbusy 600K 600041 69 145 227 100 14549 13936 napibusy 600K 599980 79 188 280 96 13927 14155 suspend10 600K 600028 78 159 267 69 10877 11032 suspend20 600K 600026 81 159 254 64 9922 10320 suspend50 600K 600007 96 178 258 57 8681 9331 suspend200 600K 599964 177 295 369 47 7115 8366 testcase load qps avglat 95%lat 99%lat cpu cpq ipq base 800K 800034 198 329 698 84 9366 8338 defer10 800K 799718 243 642 1457 95 10532 9007 defer20 800K 800009 132 245 399 89 9956 8979 defer50 800K 800024 136 228 378 80 9002 8598 defer200 800K 799965 255 362 473 66 7481 8147 fullbusy 800K 799927 78 157 253 100 10915 12533 napibusy 800K 799870 81 173 273 99 10826 12532 suspend10 800K 799991 84 167 269 83 9380 9802 suspend20 800K 799979 90 172 290 78 8765 9404 suspend50 800K 800031 106 191 307 71 7945 8805 suspend200 800K 799905 182 307 411 62 6985 8242 testcase load qps avglat 95%lat 99%lat cpu cpq ipq base 1000K 919543 3805 6390 14229 98 9324 7978 defer10 1000K 850751 4574 7382 15370 99 10218 8470 defer20 1000K 890296 4736 6862 14858 99 9708 8277 defer50 1000K 932694 3463 6180 13251 97 9148 8053 defer200 1000K 951311 3524 6052 13599 96 8875 7845 fullbusy 1000K 1000011 90 181 278 100 8731 10686 napibusy 1000K 1000050 93 184 280 100 8721 10547 suspend10 1000K 999962 101 193 306 92 8138 8980 suspend20 1000K 1000030 103 191 324 88 7844 8763 suspend50 1000K 1000001 114 202 320 83 7396 8431 suspend200 1000K 999965 185 314 428 76 6733 8072 testcase load qps avglat 95%lat 99%lat cpu cpq ipq base MAX 1005592 4651 6594 14979 100 8679 7918 defer10 MAX 928204 5106 7286 15199 100 9398 8380 defer20 MAX 984663 4774 6518 14920 100 8861 8063 defer50 MAX 1044099 4431 6368 14652 100 8350 7948 defer200 MAX 1040451 4423 6610 14674 100 8380 7931 fullbusy MAX 1236608 3715 3987 12805 100 7051 7936 napibusy MAX 1077516 4345 10155 15957 100 8080 7842 suspend10 MAX 1218344 3760 3990 12585 100 7150 7935 suspend20 MAX 1220056 3752 4053 12602 100 7150 7961 suspend50 MAX 1213666 3791 4103 12919 100 7183 7959 suspend200 MAX 1217411 3768 3988 12863 100 7161 7954 ~ 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. We ran experiments with these parameters set to zero and the results are as expected and essentially the same as the base case. - Can the new timeout value be threaded through the new epoll ioctl ? Only with difficulty. The epoll ioctl sets options on an epoll context and the NAPI ID associated with an epoll context can change based on what file descriptors a user app adds to the epoll context. This would introduce complexity in the API from the user perspective and also complexity in the kernel. - 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. Thanks, Martin and Joe [1]: https://lore.kernel.org/netdev/20240812125717.413108-1-jdamato@fastly.com/ [2]: https://doi.org/10.1145/3626780 [3]: https://github.com/memcached/memcached/blob/master/doc/napi_ids.txt [4]: https://github.com/leverich/mutilate [5]: https://raw.githubusercontent.com/martinkarsten/irqsuspend/main/patches/mem… [6]: https://raw.githubusercontent.com/martinkarsten/irqsuspend/main/patches/lib… [7]: https://github.com/martinkarsten/irqsuspend [8]: https://github.com/martinkarsten/irqsuspend/tree/main/results v5: - Adjusted patch 5 to only suspend IRQs when ep_send_events returns a positive return value. This issue was pointed out by Hillf Danton. - Updated the commit message of patch 6 which still mentioned netcat, despite the code being updated in v4 to replace it with socat and fixed misspelling of netdevsim. - Fixed a minor typo in patch 7 and removed an unnecessary paragraph. - Added Sridhar Samudrala's Reviewed-by to patch 1-5 and 7. v4: https://lore.kernel.org/netdev/20241102005214.32443-1-jdamato@fastly.com/ - Added a new FAQ item to cover letter. - Updated patch 6 to use socat instead of nc in busy_poll_test.sh and updated busy_poller.c to use netlink directly to configure napi params. - Updated the kernel documentation in patch 7 to include more details. - Dropped Stanislav's Acked-by and Bagas' Reviewed-by from patch 7 since the documentation was updated. v3: https://lore.kernel.org/netdev/20241101004846.32532-1-jdamato@fastly.com/ - Added Stanislav Fomichev's Acked-by to every patch except the newly added selftest. - Added Bagas Sanjaya's Reviewed-by to the documentation patch. - Fixed the commit message of patch 2 to remove a reference to the now non-existent sysfs setting. - Added a self test which tests both "regular" busy poll and busy poll with suspend enabled. This was added as patch 6 as requested by Paolo. netdevsim was chosen instead of veth due to netdevsim's pre-existing support for netdev-genl. See the commit message of patch 6 for more details. v2: https://lore.kernel.org/bpf/20241021015311.95468-1-jdamato@fastly.com/ - Cover letter updated, including a re-run of test data. - Patch 1 rewritten to use netdev-genl instead of sysfs. - Patch 3 updated with a comment added to napi_resume_irqs. - Patch 4 rebased to apply now that commit b9ca079dd6b0 ("eventpoll: Annotate data-race of busy_poll_usecs") has been picked up from VFS. - Patch 6 updated the kernel documentation. rfc -> v1: - Cover letter updated to include more details. - Patch 1 updated to remove the documentation added. This was moved to patch 6 with the rest of the docs (see below). - Patch 5 updated to fix an error uncovered by the kernel build robot. See patch 5's changelog for more details. - Patch 6 added which updates kernel documentation. Joe Damato (2): selftests: net: Add busy_poll_test docs: networking: Describe irq suspension Martin Karsten (5): net: Add napi_struct parameter irq_suspend_timeout net: Suspend softirq when prefer_busy_poll is set net: Add control functions for irq suspension eventpoll: Trigger napi_busy_loop, if prefer_busy_poll is set eventpoll: Control irq suspension for prefer_busy_poll Documentation/netlink/specs/netdev.yaml | 7 + Documentation/networking/napi.rst | 172 ++++++++- fs/eventpoll.c | 36 +- include/linux/netdevice.h | 2 + include/net/busy_poll.h | 3 + include/uapi/linux/netdev.h | 1 + net/core/dev.c | 58 +++- net/core/dev.h | 25 ++ net/core/netdev-genl-gen.c | 5 +- net/core/netdev-genl.c | 12 + tools/include/uapi/linux/netdev.h | 1 + tools/testing/selftests/net/.gitignore | 1 + tools/testing/selftests/net/Makefile | 3 +- tools/testing/selftests/net/busy_poll_test.sh | 164 +++++++++ tools/testing/selftests/net/busy_poller.c | 328 ++++++++++++++++++ 15 files changed, 807 insertions(+), 11 deletions(-) create mode 100755 tools/testing/selftests/net/busy_poll_test.sh create mode 100644 tools/testing/selftests/net/busy_poller.c base-commit: dbb9a7ef347828870df3e5e6ddf19469a3277fc9 -- 2.25.1

2 months, 1 week

1
1
0 0

[PATCH v1 1/2] selftests/x86/syscall: fix coccinelle WARNING recommending the use of ARRAY_SIZE()

by Mirsad Todorovac

Coccinelle gives WARNING recommending the use of ARRAY_SIZE() macro definition to improve the code readability: ./tools/testing/selftests/x86/syscall_numbering.c:316:35-36: WARNING: Use ARRAY_SIZE Fixes: 15c82d98a0f78 ("selftests/x86/syscall: Update and extend syscall_numbering_64") Cc: Thomas Gleixner <tglx(a)linutronix.de> Cc: Ingo Molnar <mingo(a)redhat.com> Cc: Borislav Petkov <bp(a)alien8.de> Cc: Dave Hansen <dave.hansen(a)linux.intel.com> Cc: x86(a)kernel.org Cc: "H. Peter Anvin" <hpa(a)zytor.com> Cc: Shuah Khan <shuah(a)kernel.org> Cc: linux-kernel(a)vger.kernel.org Cc: linux-kselftest(a)vger.kernel.org Signed-off-by: Mirsad Todorovac <mtodorovac69(a)gmail.com> --- v1: initial version. tools/testing/selftests/x86/syscall_numbering.c | 3 ++- 1 file changed, 2 insertions(+), 1 deletion(-) diff --git a/tools/testing/selftests/x86/syscall_numbering.c b/tools/testing/selftests/x86/syscall_numbering.c index 991591718bb0..41c42b7b54a6 100644 --- a/tools/testing/selftests/x86/syscall_numbering.c +++ b/tools/testing/selftests/x86/syscall_numbering.c @@ -25,6 +25,7 @@ #include <sys/mman.h> #include <linux/ptrace.h> +#include "../kselftest.h" /* Common system call numbers */ #define SYS_READ 0 @@ -313,7 +314,7 @@ static void test_syscall_numbering(void) * The MSB is supposed to be ignored, so we loop over a few * to test that out. */ - for (size_t i = 0; i < sizeof(msbs)/sizeof(msbs[0]); i++) { + for (size_t i = 0; i < ARRAY_SIZE(msbs); i++) { int msb = msbs[i]; run("Checking system calls with msb = %d (0x%x)\n", msb, msb); -- 2.43.0

2 months, 1 week

3
5
0 0

[PATCH v5] net: ipv4: Cache pmtu for all packet paths if multipath enabled

by Vladimir Vdovin

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. V5: - make self test cleaner V4: - fix selftest, do route lookup before checking cached exceptions V3: - added selftest - fixed compile error V2: - fix fib_info_num_path parameter pass An example topology showing the problem: | host1 +---------+ | dummy0 | 10.179.20.18/32 mtu9000 +---------+ +-----------+----------------+ +---------+ +---------+ | ens17f0 | 10.179.2.141/31 | ens17f1 | 10.179.2.13/31 +---------+ +---------+ | (all here have mtu 9000) | +------+ +------+ | ro1 | 10.179.2.140/31 | ro2 | 10.179.2.12/31 +------+ +------+ | | ---------+------------+-------------------+------ | +-----+ | ro3 | 10.10.10.10 mtu1500 +-----+ | ======================================== some networks ======================================== | +-----+ | eth0| 10.10.30.30 mtu9000 +-----+ | host2 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. Signed-off-by: Vladimir Vdovin <deliran(a)verdict.gg> --- net/ipv4/route.c | 13 +++++ tools/testing/selftests/net/pmtu.sh | 78 ++++++++++++++++++++++++++++- 2 files changed, 90 insertions(+), 1 deletion(-) diff --git a/net/ipv4/route.c b/net/ipv4/route.c index 723ac9181558..41162b5cc4cb 100644 --- a/net/ipv4/route.c +++ b/net/ipv4/route.c @@ -1027,6 +1027,19 @@ static void __ip_rt_update_pmtu(struct rtable *rt, struct flowi4 *fl4, u32 mtu) struct fib_nh_common *nhc; fib_select_path(net, &res, fl4, NULL); +#ifdef CONFIG_IP_ROUTE_MULTIPATH + if (fib_info_num_path(res.fi) > 1) { + int nhsel; + + for (nhsel = 0; nhsel < fib_info_num_path(res.fi); nhsel++) { + nhc = fib_info_nhc(res.fi, nhsel); + update_or_create_fnhe(nhc, fl4->daddr, 0, mtu, lock, + jiffies + net->ipv4.ip_rt_mtu_expires); + } + rcu_read_unlock(); + return; + } +#endif /* CONFIG_IP_ROUTE_MULTIPATH */ nhc = FIB_RES_NHC(res); update_or_create_fnhe(nhc, fl4->daddr, 0, mtu, lock, jiffies + net->ipv4.ip_rt_mtu_expires); diff --git a/tools/testing/selftests/net/pmtu.sh b/tools/testing/selftests/net/pmtu.sh index 569bce8b6383..a0159340fe84 100755 --- a/tools/testing/selftests/net/pmtu.sh +++ b/tools/testing/selftests/net/pmtu.sh @@ -266,7 +266,8 @@ tests=" list_flush_ipv4_exception ipv4: list and flush cached exceptions 1 list_flush_ipv6_exception ipv6: list and flush cached exceptions 1 pmtu_ipv4_route_change ipv4: PMTU exception w/route replace 1 - pmtu_ipv6_route_change ipv6: PMTU exception w/route replace 1" + pmtu_ipv6_route_change ipv6: PMTU exception w/route replace 1 + pmtu_ipv4_mp_exceptions ipv4: PMTU multipath nh exceptions 0" # Addressing and routing for tests with routers: four network segments, with # index SEGMENT between 1 and 4, a common prefix (PREFIX4 or PREFIX6) and an @@ -2329,6 +2330,81 @@ test_pmtu_ipv6_route_change() { test_pmtu_ipvX_route_change 6 } +test_pmtu_ipv4_mp_exceptions() { + setup namespaces routing || return $ksft_skip + + ip nexthop ls >/dev/null 2>&1 + if [ $? -ne 0 ]; then + echo "Nexthop objects not supported; skipping tests" + exit $ksft_skip + fi + + trace "${ns_a}" veth_A-R1 "${ns_r1}" veth_R1-A \ + "${ns_r1}" veth_R1-B "${ns_b}" veth_B-R1 \ + "${ns_a}" veth_A-R2 "${ns_r2}" veth_R2-A \ + "${ns_r2}" veth_R2-B "${ns_b}" veth_B-R2 + + dummy0_a="192.168.99.99" + dummy0_b="192.168.88.88" + + # Set up initial MTU values + mtu "${ns_a}" veth_A-R1 2000 + mtu "${ns_r1}" veth_R1-A 2000 + mtu "${ns_r1}" veth_R1-B 1500 + mtu "${ns_b}" veth_B-R1 1500 + + mtu "${ns_a}" veth_A-R2 2000 + mtu "${ns_r2}" veth_R2-A 2000 + mtu "${ns_r2}" veth_R2-B 1500 + mtu "${ns_b}" veth_B-R2 1500 + + fail=0 + + #Set up host A with multipath routes to host B dummy0_b + run_cmd ${ns_a} sysctl -q net.ipv4.fib_multipath_hash_policy=1 + run_cmd ${ns_a} sysctl -q net.ipv4.ip_forward=1 + run_cmd ${ns_a} ip link add dummy0 mtu 2000 type dummy + run_cmd ${ns_a} ip link set dummy0 up + run_cmd ${ns_a} ip addr add ${dummy0_a} dev dummy0 + run_cmd ${ns_a} ip nexthop add id 201 via ${prefix4}.${a_r1}.2 dev veth_A-R1 + run_cmd ${ns_a} ip nexthop add id 202 via ${prefix4}.${a_r2}.2 dev veth_A-R2 + run_cmd ${ns_a} ip nexthop add id 203 group 201/202 + run_cmd ${ns_a} ip route add ${dummy0_b} nhid 203 + + #Set up host B with multipath routes to host A dummy0_a + run_cmd ${ns_b} sysctl -q net.ipv4.fib_multipath_hash_policy=1 + run_cmd ${ns_b} sysctl -q net.ipv4.ip_forward=1 + run_cmd ${ns_b} ip link add dummy0 mtu 2000 type dummy + run_cmd ${ns_b} ip link set dummy0 up + run_cmd ${ns_b} ip addr add ${dummy0_b} dev dummy0 + run_cmd ${ns_b} ip nexthop add id 201 via ${prefix4}.${b_r1}.2 dev veth_A-R1 + run_cmd ${ns_b} ip nexthop add id 202 via ${prefix4}.${b_r2}.2 dev veth_A-R2 + run_cmd ${ns_b} ip nexthop add id 203 group 201/202 + run_cmd ${ns_b} ip route add ${dummy0_a} nhid 203 + + #Set up routers with routes to dummies + run_cmd ${ns_r1} ip route add ${dummy0_a} via ${prefix4}.${a_r1}.1 + run_cmd ${ns_r2} ip route add ${dummy0_a} via ${prefix4}.${a_r2}.1 + run_cmd ${ns_r1} ip route add ${dummy0_b} via ${prefix4}.${b_r1}.1 + run_cmd ${ns_r2} ip route add ${dummy0_b} via ${prefix4}.${b_r2}.1 + + + #Ping and expect two nexthop exceptions for two routes in nh group + run_cmd ${ns_a} ping -q -M want -i 0.1 -c 2 -s 1800 "${dummy0_b}" + + #Do route lookup before checking cached exceptions + run_cmd ${ns_a} ip route get ${dummy0_b} oif veth_A-R1 + run_cmd ${ns_a} ip route get ${dummy0_b} oif veth_A-R2 + + #Check cached exceptions + if [ "$(${ns_a} ip -oneline route list cache| grep mtu | wc -l)" -ne 2 ]; then + err " there are not enough cached exceptions" + fail=1 + fi + + return ${fail} +} + usage() { echo echo "$0 [OPTIONS] [TEST]..." base-commit: 66600fac7a984dea4ae095411f644770b2561ede -- 2.43.5

2 months, 1 week

3
4
0 0

[PATCH 0/2] kunit: enable hardware virtualization

by Tamir Duberstein

This series implements feature detection of hardware virtualization on Linux and macOS; the latter being my primary use case. This yields approximately a 6x improvement using HVF on M3 Pro. Signed-off-by: Tamir Duberstein <tamird(a)gmail.com> --- Tamir Duberstein (2): kunit: add fallback for os.sched_getaffinity kunit: enable hardware acceleration when available tools/testing/kunit/kunit.py | 11 ++++++++++- tools/testing/kunit/kunit_kernel.py | 26 +++++++++++++++++++++++++- tools/testing/kunit/qemu_configs/arm64.py | 2 +- 3 files changed, 36 insertions(+), 3 deletions(-) --- base-commit: ae90f6a6170d7a7a1aa4fddf664fbd093e3023bc change-id: 20241025-kunit-qemu-accel-macos-2840e4c2def5 Best regards, -- Tamir Duberstein <tamird(a)gmail.com>

2 months, 1 week

3
6
0 0

[GIT PULL] Kselftest fixes update for Linux 6.12-rc6

by Shuah Khan

Hi Linus, Please pull the following kselftest fixes update for Linux 6.12-rc6. Kselftest fixes for Linux 6.12-rc6 - fix syntax error in frequency calculation arithmetic expression in intel_pstate run.sh - add missing cpupower dependency check intel_pstate run.sh - fix idmap_mount_tree_invalid test failure due to incorrect argument - fix watchdog-test run leaving the watchdog timer enabled causing system reboot. With this fix, the test disables the watchdog timer when it gets terminated with SIGTERM, SIGKILL, and SIGQUIT in addition to SIGINT diff is attached. thanks, -- Shuah ---------------------------------------------------------------- The following changes since commit fe05c40ca9c18cfdb003f639a30fc78a7ab49519: selftest: hid: add the missing tests directory (2024-10-16 15:55:14 -0600) are available in the Git repository at: git://git.kernel.org/pub/scm/linux/kernel/git/shuah/linux-kselftest tags/linux_kselftest-fixes-6.12-rc6 for you to fetch changes up to dc1308bee1ed03b4d698d77c8bd670d399dcd04d: selftests/watchdog-test: Fix system accidentally reset after watchdog-test (2024-10-28 21:34:43 -0600) ---------------------------------------------------------------- linux_kselftest-fixes-6.12-rc6 Kselftest fixes for Linux 6.12-rc6 - fix syntax error in frequency calculation arithmetic expression in intel_pstate run.sh - add missing cpupower dependency check intel_pstate run.sh - fix idmap_mount_tree_invalid test failure due to incorrect argument - fix watchdog-test run leaving the watchdog timer enabled causing system reboot. With this fix, the test disables the watchdog timer when it gets terminated with SIGTERM, SIGKILL, and SIGQUIT in addition to SIGINT ---------------------------------------------------------------- Alessandro Zanni (2): selftests/intel_pstate: fix operand expected error selftests/intel_pstate: check if cpupower is installed Li Zhijian (1): selftests/watchdog-test: Fix system accidentally reset after watchdog-test zhouyuhang (1): selftests/mount_setattr: fix idmap_mount_tree_invalid failed to run tools/testing/selftests/intel_pstate/run.sh | 9 +++++++-- tools/testing/selftests/mount_setattr/mount_setattr_test.c | 9 +++++++++ tools/testing/selftests/watchdog/watchdog-test.c | 6 ++++++ 3 files changed, 22 insertions(+), 2 deletions(-) ---------------------------------------------------------------

2 months, 1 week

2
1
0 0

[PATCH net-next v4 0/7] Suspend IRQs during application busy periods

by Joe Damato

Greetings: Welcome to v4, see changelog below. Note that our performance tests were not re-run for this revision as we updated the selftest, FAQ in the cover letter, and kernel documentation. No functional/code changes. 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. We've added an additional test case, 'fullbusy', achieved by modifying libevent for comparison. See below for a detailed description, link to the patch, and test results. 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. ~ Design rationale The implementation of the IRQ suspension mechanism very nicely dovetails with the existing mechanism for IRQ deferral when preferred busy poll is enabled (introduced in commit 7fd3253a7de6 ("net: Introduce preferred busy-polling"), see that commit message for more details). While it would be possible to inject the suspend timeout via the existing epoll ioctl, it is more natural to avoid this path for one main reason: An epoll context is linked to NAPI IDs as file descriptors are added; this means any epoll context might suddenly be associated with a different net_device if the application were to replace all existing fds with fds from a different device. In this case, the scope of the suspend timeout becomes unclear and many edge cases for both the user application and the kernel are introduced Only a single iteration through napi busy polling is needed for this mechanism to work effectively. Since an important objective for this mechanism is preserving cpu cycles, exactly one iteration of the napi busy loop is invoked when busy_poll_usecs is set to 0. ~ Important call outs 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. - Patches apply cleanly on commit 160a810b2a85 ("net: vxlan: update the document for vxlan_snoop()"). ~ 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 - 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 at least 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 are higher than our previous submission, but the relative differences between variants are equivalent. Because the patches have been rebased on 6.12, several factors have likely influenced the overall performance. Most importantly, we had to switch to a new set of basic kernel options, which has likely altered the baseline performance. Because the overall comparison of variants still holds, we have not attempted to recreate the exact set of kernel options from the previous submission. 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. 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 200K 200024 127 254 458 25 12748 11289 defer10 200K 199991 64 128 166 27 18763 16574 defer20 200K 199986 72 135 178 25 15405 14173 defer50 200K 200025 91 149 198 23 12275 12203 defer200 200K 199996 182 266 326 18 8595 9183 fullbusy 200K 200040 58 123 167 100 43641 23145 napibusy 200K 200009 115 244 299 56 24797 24693 suspend10 200K 200005 63 128 167 32 19559 17240 suspend20 200K 199952 69 132 170 29 16324 14838 suspend50 200K 200019 84 144 189 26 13106 12516 suspend200 200K 199978 168 264 326 20 9331 9643 testcase load qps avglat 95%lat 99%lat cpu cpq ipq base 400K 400017 157 292 762 39 9287 9325 defer10 400K 400033 71 141 204 53 13950 12943 defer20 400K 399935 79 150 212 47 12027 11673 defer50 400K 399888 101 171 231 39 9556 9921 defer200 400K 399993 200 287 358 32 7428 8576 fullbusy 400K 400018 63 132 203 100 21827 16062 napibusy 400K 399970 89 230 292 83 18156 16508 suspend10 400K 400061 69 139 202 54 13576 13057 suspend20 400K 399988 73 144 206 49 11930 11773 suspend50 400K 399975 88 161 218 42 9996 10270 suspend200 400K 399954 172 276 353 34 7847 8713 testcase load qps avglat 95%lat 99%lat cpu cpq ipq base 600K 600031 166 289 631 61 9188 8787 defer10 600K 599967 85 167 262 75 11833 10947 defer20 600K 599888 89 165 243 66 10513 10362 defer50 600K 600072 109 185 253 55 8664 9190 defer200 600K 599951 222 315 393 45 6892 8213 fullbusy 600K 600041 69 145 227 100 14549 13936 napibusy 600K 599980 79 188 280 96 13927 14155 suspend10 600K 600028 78 159 267 69 10877 11032 suspend20 600K 600026 81 159 254 64 9922 10320 suspend50 600K 600007 96 178 258 57 8681 9331 suspend200 600K 599964 177 295 369 47 7115 8366 testcase load qps avglat 95%lat 99%lat cpu cpq ipq base 800K 800034 198 329 698 84 9366 8338 defer10 800K 799718 243 642 1457 95 10532 9007 defer20 800K 800009 132 245 399 89 9956 8979 defer50 800K 800024 136 228 378 80 9002 8598 defer200 800K 799965 255 362 473 66 7481 8147 fullbusy 800K 799927 78 157 253 100 10915 12533 napibusy 800K 799870 81 173 273 99 10826 12532 suspend10 800K 799991 84 167 269 83 9380 9802 suspend20 800K 799979 90 172 290 78 8765 9404 suspend50 800K 800031 106 191 307 71 7945 8805 suspend200 800K 799905 182 307 411 62 6985 8242 testcase load qps avglat 95%lat 99%lat cpu cpq ipq base 1000K 919543 3805 6390 14229 98 9324 7978 defer10 1000K 850751 4574 7382 15370 99 10218 8470 defer20 1000K 890296 4736 6862 14858 99 9708 8277 defer50 1000K 932694 3463 6180 13251 97 9148 8053 defer200 1000K 951311 3524 6052 13599 96 8875 7845 fullbusy 1000K 1000011 90 181 278 100 8731 10686 napibusy 1000K 1000050 93 184 280 100 8721 10547 suspend10 1000K 999962 101 193 306 92 8138 8980 suspend20 1000K 1000030 103 191 324 88 7844 8763 suspend50 1000K 1000001 114 202 320 83 7396 8431 suspend200 1000K 999965 185 314 428 76 6733 8072 testcase load qps avglat 95%lat 99%lat cpu cpq ipq base MAX 1005592 4651 6594 14979 100 8679 7918 defer10 MAX 928204 5106 7286 15199 100 9398 8380 defer20 MAX 984663 4774 6518 14920 100 8861 8063 defer50 MAX 1044099 4431 6368 14652 100 8350 7948 defer200 MAX 1040451 4423 6610 14674 100 8380 7931 fullbusy MAX 1236608 3715 3987 12805 100 7051 7936 napibusy MAX 1077516 4345 10155 15957 100 8080 7842 suspend10 MAX 1218344 3760 3990 12585 100 7150 7935 suspend20 MAX 1220056 3752 4053 12602 100 7150 7961 suspend50 MAX 1213666 3791 4103 12919 100 7183 7959 suspend200 MAX 1217411 3768 3988 12863 100 7161 7954 ~ 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. We ran experiments with these parameters set to zero and the results are as expected and essentially the same as the base case. - Can the new timeout value be threaded through the new epoll ioctl ? Only with difficulty. The epoll ioctl sets options on an epoll context and the NAPI ID associated with an epoll context can change based on what file descriptors a user app adds to the epoll context. This would introduce complexity in the API from the user perspective and also complexity in the kernel. - 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 result. 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. Thanks, Martin and Joe [1]: https://lore.kernel.org/netdev/20240812125717.413108-1-jdamato@fastly.com/ [2]: https://doi.org/10.1145/3626780 [3]: https://github.com/memcached/memcached/blob/master/doc/napi_ids.txt [4]: https://github.com/leverich/mutilate [5]: https://raw.githubusercontent.com/martinkarsten/irqsuspend/main/patches/mem… [6]: https://raw.githubusercontent.com/martinkarsten/irqsuspend/main/patches/lib… [7]: https://github.com/martinkarsten/irqsuspend [8]: https://github.com/martinkarsten/irqsuspend/tree/main/results v4: - Added a new FAQ item to cover letter. - Updated patch 6 to use socat instead of nc in busy_poll_test.sh and updated busy_poller.c to use netlink directly to configure napi params. - Updated the kernel documentation in patch 7 to include more details. - Dropped Stanislav's Acked-by and Bagas' Reviewed-by from patch 7 since the documentation was updated. v3: - Added Stanislav Fomichev's Acked-by to every patch except the newly added selftest. - Added Bagas Sanjaya's Reviewed-by to the documentation patch. - Fixed the commit message of patch 2 to remove a reference to the now non-existent sysfs setting. - Added a self test which tests both "regular" busy poll and busy poll with suspend enabled. This was added as patch 6 as requested by Paolo. netdevsim was chosen instead of veth due to netdevsim's pre-existing support for netdev-genl. See the commit message of patch 6 for more details. v2: https://lore.kernel.org/bpf/20241021015311.95468-1-jdamato@fastly.com/ - Cover letter updated, including a re-run of test data. - Patch 1 rewritten to use netdev-genl instead of sysfs. - Patch 3 updated with a comment added to napi_resume_irqs. - Patch 4 rebased to apply now that commit b9ca079dd6b0 ("eventpoll: Annotate data-race of busy_poll_usecs") has been picked up from VFS. - Patch 6 updated the kernel documentation. rfc -> v1: - Cover letter updated to include more details. - Patch 1 updated to remove the documentation added. This was moved to patch 6 with the rest of the docs (see below). - Patch 5 updated to fix an error uncovered by the kernel build robot. See patch 5's changelog for more details. - Patch 6 added which updates kernel documentation. Joe Damato (2): selftests: net: Add busy_poll_test docs: networking: Describe irq suspension Martin Karsten (5): net: Add napi_struct parameter irq_suspend_timeout net: Suspend softirq when prefer_busy_poll is set net: Add control functions for irq suspension eventpoll: Trigger napi_busy_loop, if prefer_busy_poll is set eventpoll: Control irq suspension for prefer_busy_poll Documentation/netlink/specs/netdev.yaml | 7 + Documentation/networking/napi.rst | 176 +++++++++- fs/eventpoll.c | 35 +- include/linux/netdevice.h | 2 + include/net/busy_poll.h | 3 + include/uapi/linux/netdev.h | 1 + net/core/dev.c | 58 +++- net/core/dev.h | 25 ++ net/core/netdev-genl-gen.c | 5 +- net/core/netdev-genl.c | 12 + tools/include/uapi/linux/netdev.h | 1 + tools/testing/selftests/net/.gitignore | 1 + tools/testing/selftests/net/Makefile | 3 +- tools/testing/selftests/net/busy_poll_test.sh | 164 +++++++++ tools/testing/selftests/net/busy_poller.c | 328 ++++++++++++++++++ 15 files changed, 810 insertions(+), 11 deletions(-) create mode 100755 tools/testing/selftests/net/busy_poll_test.sh create mode 100644 tools/testing/selftests/net/busy_poller.c base-commit: dbb9a7ef347828870df3e5e6ddf19469a3277fc9 -- 2.25.1

2 months, 1 week

1
1
0 0

[PATCH net-next v3 0/7] Suspend IRQs during application busy periods

by Joe Damato

Greetings: Welcome to v3, see changelog below. Note that our performance tests were not re-run for this revision as we only updated a commit message and added a selftest. 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. We've added an additional test case, 'fullbusy', achieved by modifying libevent for comparison. See below for a detailed description, link to the patch, and test results. 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. ~ Design rationale The implementation of the IRQ suspension mechanism very nicely dovetails with the existing mechanism for IRQ deferral when preferred busy poll is enabled (introduced in commit 7fd3253a7de6 ("net: Introduce preferred busy-polling"), see that commit message for more details). While it would be possible to inject the suspend timeout via the existing epoll ioctl, it is more natural to avoid this path for one main reason: An epoll context is linked to NAPI IDs as file descriptors are added; this means any epoll context might suddenly be associated with a different net_device if the application were to replace all existing fds with fds from a different device. In this case, the scope of the suspend timeout becomes unclear and many edge cases for both the user application and the kernel are introduced Only a single iteration through napi busy polling is needed for this mechanism to work effectively. Since an important objective for this mechanism is preserving cpu cycles, exactly one iteration of the napi busy loop is invoked when busy_poll_usecs is set to 0. ~ Important call outs 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. - Patches apply cleanly on commit 160a810b2a85 ("net: vxlan: update the document for vxlan_snoop()"). ~ 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 - 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 at least 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 are higher than our previous submission, but the relative differences between variants are equivalent. Because the patches have been rebased on 6.12, several factors have likely influenced the overall performance. Most importantly, we had to switch to a new set of basic kernel options, which has likely altered the baseline performance. Because the overall comparison of variants still holds, we have not attempted to recreate the exact set of kernel options from the previous submission. 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. 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 200K 200024 127 254 458 25 12748 11289 defer10 200K 199991 64 128 166 27 18763 16574 defer20 200K 199986 72 135 178 25 15405 14173 defer50 200K 200025 91 149 198 23 12275 12203 defer200 200K 199996 182 266 326 18 8595 9183 fullbusy 200K 200040 58 123 167 100 43641 23145 napibusy 200K 200009 115 244 299 56 24797 24693 suspend10 200K 200005 63 128 167 32 19559 17240 suspend20 200K 199952 69 132 170 29 16324 14838 suspend50 200K 200019 84 144 189 26 13106 12516 suspend200 200K 199978 168 264 326 20 9331 9643 testcase load qps avglat 95%lat 99%lat cpu cpq ipq base 400K 400017 157 292 762 39 9287 9325 defer10 400K 400033 71 141 204 53 13950 12943 defer20 400K 399935 79 150 212 47 12027 11673 defer50 400K 399888 101 171 231 39 9556 9921 defer200 400K 399993 200 287 358 32 7428 8576 fullbusy 400K 400018 63 132 203 100 21827 16062 napibusy 400K 399970 89 230 292 83 18156 16508 suspend10 400K 400061 69 139 202 54 13576 13057 suspend20 400K 399988 73 144 206 49 11930 11773 suspend50 400K 399975 88 161 218 42 9996 10270 suspend200 400K 399954 172 276 353 34 7847 8713 testcase load qps avglat 95%lat 99%lat cpu cpq ipq base 600K 600031 166 289 631 61 9188 8787 defer10 600K 599967 85 167 262 75 11833 10947 defer20 600K 599888 89 165 243 66 10513 10362 defer50 600K 600072 109 185 253 55 8664 9190 defer200 600K 599951 222 315 393 45 6892 8213 fullbusy 600K 600041 69 145 227 100 14549 13936 napibusy 600K 599980 79 188 280 96 13927 14155 suspend10 600K 600028 78 159 267 69 10877 11032 suspend20 600K 600026 81 159 254 64 9922 10320 suspend50 600K 600007 96 178 258 57 8681 9331 suspend200 600K 599964 177 295 369 47 7115 8366 testcase load qps avglat 95%lat 99%lat cpu cpq ipq base 800K 800034 198 329 698 84 9366 8338 defer10 800K 799718 243 642 1457 95 10532 9007 defer20 800K 800009 132 245 399 89 9956 8979 defer50 800K 800024 136 228 378 80 9002 8598 defer200 800K 799965 255 362 473 66 7481 8147 fullbusy 800K 799927 78 157 253 100 10915 12533 napibusy 800K 799870 81 173 273 99 10826 12532 suspend10 800K 799991 84 167 269 83 9380 9802 suspend20 800K 799979 90 172 290 78 8765 9404 suspend50 800K 800031 106 191 307 71 7945 8805 suspend200 800K 799905 182 307 411 62 6985 8242 testcase load qps avglat 95%lat 99%lat cpu cpq ipq base 1000K 919543 3805 6390 14229 98 9324 7978 defer10 1000K 850751 4574 7382 15370 99 10218 8470 defer20 1000K 890296 4736 6862 14858 99 9708 8277 defer50 1000K 932694 3463 6180 13251 97 9148 8053 defer200 1000K 951311 3524 6052 13599 96 8875 7845 fullbusy 1000K 1000011 90 181 278 100 8731 10686 napibusy 1000K 1000050 93 184 280 100 8721 10547 suspend10 1000K 999962 101 193 306 92 8138 8980 suspend20 1000K 1000030 103 191 324 88 7844 8763 suspend50 1000K 1000001 114 202 320 83 7396 8431 suspend200 1000K 999965 185 314 428 76 6733 8072 testcase load qps avglat 95%lat 99%lat cpu cpq ipq base MAX 1005592 4651 6594 14979 100 8679 7918 defer10 MAX 928204 5106 7286 15199 100 9398 8380 defer20 MAX 984663 4774 6518 14920 100 8861 8063 defer50 MAX 1044099 4431 6368 14652 100 8350 7948 defer200 MAX 1040451 4423 6610 14674 100 8380 7931 fullbusy MAX 1236608 3715 3987 12805 100 7051 7936 napibusy MAX 1077516 4345 10155 15957 100 8080 7842 suspend10 MAX 1218344 3760 3990 12585 100 7150 7935 suspend20 MAX 1220056 3752 4053 12602 100 7150 7961 suspend50 MAX 1213666 3791 4103 12919 100 7183 7959 suspend200 MAX 1217411 3768 3988 12863 100 7161 7954 ~ FAQ - Can the new timeout value be threaded through the new epoll ioctl ? Only with difficulty. The epoll ioctl sets options on an epoll context and the NAPI ID associated with an epoll context can change based on what file descriptors a user app adds to the epoll context. This would introduce complexity in the API from the user perspective and also complexity in the kernel. - 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 result. 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. Thanks, Martin and Joe [1]: https://lore.kernel.org/netdev/20240812125717.413108-1-jdamato@fastly.com/ [2]: https://doi.org/10.1145/3626780 [3]: https://github.com/memcached/memcached/blob/master/doc/napi_ids.txt [4]: https://github.com/leverich/mutilate [5]: https://raw.githubusercontent.com/martinkarsten/irqsuspend/main/patches/mem… [6]: https://raw.githubusercontent.com/martinkarsten/irqsuspend/main/patches/lib… [7]: https://github.com/martinkarsten/irqsuspend [8]: https://github.com/martinkarsten/irqsuspend/tree/main/results v3: - Added Stanislav Fomichev's Acked-by to every patch except the newly added selftest. - Added Bagas Sanjaya's Reviewed-by to the documentation patch. - Fixed the commit message of patch 2 to remove a reference to the now non-existent sysfs setting. - Added a self test which tests both "regular" busy poll and busy poll with suspend enabled. This was added as patch 6 as requested by Paolo. netdevsim was chosen instead of veth due to netdevsim's pre-existing support for netdev-genl. See the commit message of patch 6 for more details. v2: https://lore.kernel.org/bpf/20241021015311.95468-1-jdamato@fastly.com/ - Cover letter updated, including a re-run of test data. - Patch 1 rewritten to use netdev-genl instead of sysfs. - Patch 3 updated with a comment added to napi_resume_irqs. - Patch 4 rebased to apply now that commit b9ca079dd6b0 ("eventpoll: Annotate data-race of busy_poll_usecs") has been picked up from VFS. - Patch 6 updated the kernel documentation. rfc -> v1: - Cover letter updated to include more details. - Patch 1 updated to remove the documentation added. This was moved to patch 6 with the rest of the docs (see below). - Patch 5 updated to fix an error uncovered by the kernel build robot. See patch 5's changelog for more details. - Patch 6 added which updates kernel documentation. Joe Damato (2): selftests: net: Add busy_poll_test docs: networking: Describe irq suspension Martin Karsten (5): net: Add napi_struct parameter irq_suspend_timeout net: Suspend softirq when prefer_busy_poll is set net: Add control functions for irq suspension eventpoll: Trigger napi_busy_loop, if prefer_busy_poll is set eventpoll: Control irq suspension for prefer_busy_poll Documentation/netlink/specs/netdev.yaml | 7 + Documentation/networking/napi.rst | 132 +++++++++- fs/eventpoll.c | 35 ++- include/linux/netdevice.h | 2 + include/net/busy_poll.h | 3 + include/uapi/linux/netdev.h | 1 + net/core/dev.c | 58 ++++- net/core/dev.h | 25 ++ net/core/netdev-genl-gen.c | 5 +- net/core/netdev-genl.c | 12 + tools/include/uapi/linux/netdev.h | 1 + tools/testing/selftests/net/.gitignore | 1 + tools/testing/selftests/net/Makefile | 2 + tools/testing/selftests/net/busy_poll_test.sh | 164 ++++++++++++ tools/testing/selftests/net/busy_poller.c | 245 ++++++++++++++++++ 15 files changed, 683 insertions(+), 10 deletions(-) create mode 100755 tools/testing/selftests/net/busy_poll_test.sh create mode 100644 tools/testing/selftests/net/busy_poller.c base-commit: 9e114ec8084020e10e1cb7b43dbbf6e69940866b -- 2.25.1

2 months, 1 week

2
3
0 0

[PATCH] kselftest/arm64: Use ksft_perror() to log MTE failures

by Mark Brown

The logging in the allocation helpers variously uses ksft_print_msg() with very intermittent logging of errno and perror() (which won't produce KTAP conformant output) when logging the result of API calls that set errno. Standardise on using the ksft_perror() helper in these cases so that more information is available should the tests fail. Signed-off-by: Mark Brown <broonie(a)kernel.org> --- tools/testing/selftests/arm64/mte/mte_common_util.c | 14 +++++++------- 1 file changed, 7 insertions(+), 7 deletions(-) diff --git a/tools/testing/selftests/arm64/mte/mte_common_util.c b/tools/testing/selftests/arm64/mte/mte_common_util.c index 00ffd34c66d301ee7d5c99e6b8d9d5d944520b7f..46958b58801e90ceb79be76f57c7f72b50d43b3c 100644 --- a/tools/testing/selftests/arm64/mte/mte_common_util.c +++ b/tools/testing/selftests/arm64/mte/mte_common_util.c @@ -150,13 +150,13 @@ static void *__mte_allocate_memory_range(size_t size, int mem_type, int mapping, map_flag |= MAP_PRIVATE; ptr = mmap(NULL, entire_size, prot_flag, map_flag, fd, 0); if (ptr == MAP_FAILED) { - ksft_print_msg("FAIL: mmap allocation\n"); + ksft_perror("mmap()"); return NULL; } if (mem_type == USE_MPROTECT) { if (mprotect(ptr, entire_size, prot_flag | PROT_MTE)) { + ksft_perror("mprotect(PROT_MTE)"); munmap(ptr, size); - ksft_print_msg("FAIL: mprotect PROT_MTE property\n"); return NULL; } } @@ -190,13 +190,13 @@ void *mte_allocate_file_memory(size_t size, int mem_type, int mapping, bool tags lseek(fd, 0, SEEK_SET); for (index = INIT_BUFFER_SIZE; index < size; index += INIT_BUFFER_SIZE) { if (write(fd, buffer, INIT_BUFFER_SIZE) != INIT_BUFFER_SIZE) { - perror("initialising buffer"); + ksft_perror("initialising buffer"); return NULL; } } index -= INIT_BUFFER_SIZE; if (write(fd, buffer, size - index) != size - index) { - perror("initialising buffer"); + ksft_perror("initialising buffer"); return NULL; } return __mte_allocate_memory_range(size, mem_type, mapping, 0, 0, tags, fd); @@ -217,12 +217,12 @@ void *mte_allocate_file_memory_tag_range(size_t size, int mem_type, int mapping, lseek(fd, 0, SEEK_SET); for (index = INIT_BUFFER_SIZE; index < map_size; index += INIT_BUFFER_SIZE) if (write(fd, buffer, INIT_BUFFER_SIZE) != INIT_BUFFER_SIZE) { - perror("initialising buffer"); + ksft_perror("initialising buffer"); return NULL; } index -= INIT_BUFFER_SIZE; if (write(fd, buffer, map_size - index) != map_size - index) { - perror("initialising buffer"); + ksft_perror("initialising buffer"); return NULL; } return __mte_allocate_memory_range(size, mem_type, mapping, range_before, @@ -359,7 +359,7 @@ int create_temp_file(void) /* Create a file in the tmpfs filesystem */ fd = mkstemp(&filename[0]); if (fd == -1) { - perror(filename); + ksft_perror(filename); ksft_print_msg("FAIL: Unable to open temporary file\n"); return 0; } --- base-commit: 8e929cb546ee42c9a61d24fae60605e9e3192354 change-id: 20241028-arm64-mte-test-logging-6c6e737b1c62 Best regards, -- Mark Brown <broonie(a)kernel.org>

2 months, 1 week

3
2
0 0

[PATCH] kselftest/arm64: Fix encoding for SVE B16B16 test

by Mark Brown

The test for SVE_B16B16 had a cut'n'paste of a SME instruction, fix it with a relevant SVE instruction. Fixes: 44d10c27bd75 ("kselftest/arm64: Add 2023 DPISA hwcap test coverage") Signed-off-by: Mark Brown <broonie(a)kernel.org> --- tools/testing/selftests/arm64/abi/hwcap.c | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-) diff --git a/tools/testing/selftests/arm64/abi/hwcap.c b/tools/testing/selftests/arm64/abi/hwcap.c index f2d6007a2b983eba77a880ec7e614396a6cb1377..1e4ed89802f76c2bcbe4bfd1a9a987970db8b97d 100644 --- a/tools/testing/selftests/arm64/abi/hwcap.c +++ b/tools/testing/selftests/arm64/abi/hwcap.c @@ -361,8 +361,8 @@ static void sveaes_sigill(void) static void sveb16b16_sigill(void) { - /* BFADD ZA.H[W0, 0], {Z0.H-Z1.H} */ - asm volatile(".inst 0xC1E41C00" : : : ); + /* BFADD Z0.H, Z0.H, Z0.H */ + asm volatile(".inst 0x65000000" : : : ); } static void svepmull_sigill(void) --- base-commit: 8e929cb546ee42c9a61d24fae60605e9e3192354 change-id: 20241024-arm64-b16b16-test-a1c7bad2478e Best regards, -- Mark Brown <broonie(a)kernel.org>

2 months, 1 week

2
1
0 0

[PATCH RFC v2 0/5] Verify bias functionality for pinctrl_paris driver through new gpio test

by Nícolas F. R. A. Prado

This series was motivated by the regression fixed by 166bf8af9122 ("pinctrl: mediatek: common-v2: Fix broken bias-disable for PULL_PU_PD_RSEL_TYPE"). A bug was introduced in the pinctrl_paris driver which prevented certain pins from having their bias configured. Running this test on the mt8195-tomato platform with the test plan included below[1] shows the test passing with the fix applied, but failing without the fix: With fix: $ ./gpio-setget-config.py TAP version 13 # Using test plan file: ./google,tomato.yaml 1..3 ok 1 pinctrl_paris.34.pull-up ok 2 pinctrl_paris.34.pull-down ok 3 pinctrl_paris.34.disabled # Totals: pass:3 fail:0 xfail:0 xpass:0 skip:0 error:0 Without fix: $ ./gpio-setget-config.py TAP version 13 # Using test plan file: ./google,tomato.yaml 1..3 # Bias doesn't match: Expected pull-up, read pull-down. not ok 1 pinctrl_paris.34.pull-up ok 2 pinctrl_paris.34.pull-down # Bias doesn't match: Expected disabled, read pull-down. not ok 3 pinctrl_paris.34.disabled # Totals: pass:1 fail:2 xfail:0 xpass:0 skip:0 error:0 In order to achieve this, the first three patches expose bias configuration through the GPIO API in the MediaTek pinctrl drivers, notably, pinctrl_paris, patch 4 extends the gpio-mockup-cdev utility for use by patch 5, and patch 5 introduces a new GPIO kselftest that takes a test plan in YAML, which can be tailored per-platform to specify the configurations to test, and sets and gets back each pin configuration to verify that they match and thus that the driver is behaving as expected. Since the GPIO uAPI only allows setting the pin configuration, getting it back is done through pinconf-pins in the pinctrl debugfs folder. The test currently only verifies bias but it would be easy to extend to verify other pin configurations. The test plan YAML file can be customized for each use-case and is platform-dependant. For that reason, only an example is included in patch 3 and the user is supposed to provide their test plan. That said, the aim is to collect test plans for ease of use at [2]. [1] This is the test plan used for mt8195-tomato: - label: "pinctrl_paris" tests: # Pin 34 has type MTK_PULL_PU_PD_RSEL_TYPE and is unused. # Setting bias to MTK_PULL_PU_PD_RSEL_TYPE pins was fixed by # 166bf8af9122 ("pinctrl: mediatek: common-v2: Fix broken bias-disable for PULL_PU_PD_RSEL_TYPE") - pin: 34 bias: "pull-up" - pin: 34 bias: "pull-down" - pin: 34 bias: "disabled" [2] https://github.com/kernelci/platform-test-parameters Signed-off-by: Nícolas F. R. A. Prado <nfraprado(a)collabora.com> --- Changes in v2: - Added patches 2 and 3 enabling the extra GPIO pin configurations on the other mediatek drivers: pinctrl-moore and pinctrl-mtk-common - Tweaked function name in patch 1: mtk_pinconf_set -> mtk_paris_pin_config_set, to make it clear it is not a pinconf_ops - Adjusted commit message to make it clear the current support is limited to pins supported by the EINT controller - Link to v1: https://lore.kernel.org/r/20240909-kselftest-gpio-set-get-config-v1-0-16a06… --- Nícolas F. R. A. Prado (5): pinctrl: mediatek: paris: Expose more configurations to GPIO set_config pinctrl: mediatek: moore: Expose more configurations to GPIO set_config pinctrl: mediatek: common: Expose more configurations to GPIO set_config selftest: gpio: Add wait flag to gpio-mockup-cdev selftest: gpio: Add a new set-get config test drivers/pinctrl/mediatek/pinctrl-moore.c | 283 +++++++++++---------- drivers/pinctrl/mediatek/pinctrl-mtk-common.c | 48 ++-- drivers/pinctrl/mediatek/pinctrl-paris.c | 26 +- tools/testing/selftests/gpio/Makefile | 2 +- tools/testing/selftests/gpio/gpio-mockup-cdev.c | 14 +- .../gpio-set-get-config-example-test-plan.yaml | 15 ++ .../testing/selftests/gpio/gpio-set-get-config.py | 183 +++++++++++++ 7 files changed, 395 insertions(+), 176 deletions(-) --- base-commit: a39230ecf6b3057f5897bc4744a790070cfbe7a8 change-id: 20240906-kselftest-gpio-set-get-config-6e5bb670c1a5 Best regards, -- Nícolas F. R. A. Prado <nfraprado(a)collabora.com>

2 months, 1 week

3
12
0 0

2025

2024

2023

2022

2021

2020

2019

2018

2017

Linux-kselftest-mirror