selftests/bpf: perf_link: avoid failures in concurrent mode
Subsystems (3)
PATCH 1/1 selftests/bpf: perf_link: avoid failures in concurrent mode
perf_link creates a system-wide perf event pinned to CPU 0 (pid=-1, cpu=0) and also pins the test thread to CPU 0. Under concurrent selftests this can lead to cross-test interference and CPU 0 contention, making the test flaky.
Create a per-task perf event instead (pid=0, cpu=-1) and drop CPU pinning from burn_cpu(). Use barrier() to prevent the burn loop from being optimized away. Drop the serial_ prefix so the test can run in parallel. Also remove the stale TODO comment.
Tested:
./test_progs -t perf_link -vv
./test_progs -j$(nproc) -t perf_link -vv
for i in $(seq 1 50); do ./test_progs -j$(nproc) -t perf_link; done
--- .../selftests/bpf/prog_tests/perf_link.c | 20 +++++-------------- 1 file changed, 5 insertions(+), 15 deletions(-)
tools/testing/selftests/bpf/prog_tests/perf_link.c +5 -15
@@ -1,8 +1,7 @@// SPDX-License-Identifier: GPL-2.0 /* Copyright (c) 2021 Facebook */ #define _GNU_SOURCE -#include <pthread.h> -#include <sched.h> +#include <linux/compiler.h> #include <test_progs.h> #include "testing_helpers.h" #include "test_perf_link.skel.h"
@@ -12,23 +11,14 @@static void burn_cpu (void ) { -volatile int j =0 ; -cpu_set_t cpu_set; -int i, err; - -/* generate some branches on cpu 0 */ -CPU_ZERO (&cpu_set); -CPU_SET (0 , &cpu_set); - err =pthread_setaffinity_np (pthread_self (),sizeof (cpu_set), &cpu_set); -ASSERT_OK (err,"set_thread_affinity" ); +int i;/* spin the loop for a while (random high number) */ for (i =0 ; i <1000000 ; ++i) - ++j; +barrier (); } -/* TODO: often fails in concurrent mode */ -void serial_test_perf_link (void ) +void test_perf_link (void ) {struct test_perf_link *skel =NULL ;struct perf_event_attr attr;
@@ -45,7 +35,7 @@ void serial_test_perf_link(void) attr.config =PERF_COUNT_SW_CPU_CLOCK ; attr.freq =1 ; attr.sample_freq =1000 ; - pfd =syscall (__NR_perf_event_open, &attr,-1 ,0 ,-1 ,PERF_FLAG_FD_CLOEXEC ); + pfd =syscall (__NR_perf_event_open, &attr,0 ,-1 ,-1 ,PERF_FLAG_FD_CLOEXEC );if (!ASSERT_GE (pfd,0 ,"perf_fd" )) goto cleanup;
base-commit: 0b3bb205808195159be633a8cefb602670e856fb
Signature · 1 line
2.43.0
Quoted 11 lines · Sun Jian
perf_linkcreates a system-wide perf event pinned to CPU 0 (pid=-1, cpu=0) and also pins the test thread to CPU 0. Under concurrent selftests this can lead to cross-test interference and CPU 0 contention, making the test flaky.Create a per-task perf event instead (pid=0, cpu=-1) and drop CPU pinning from
burn_cpu(). Usebarrier()to prevent the burn loop from being optimized away. Drop theserial_prefix so the test can run in parallel. Also remove the stale TODO comment.
Quoted 7 lines
Tested:
./test_progs -tperf_link-vv
./test_progs -j$(nproc) -tperf_link-vv
for i in $(seq 1 50); do ./test_progs -j$(nproc) -tperf_link; doneSigned-off-by Sun Jian <[email protected]>---
I tried reproducing the flakiness when removing serial_ prefix without this patch for a while, but wasn't successful.
That said, the logic makes sense to me, the bit about burn_cpu() being optimized away can be true, and I think it is an improvement.
Quoted 1 line
[...]
This patch was applied to bpf/bpf-next.git (master) by Alexei Starovoitov <[email protected]>:
Hello:
This patch was applied to bpf/bpf-next.git (master) by Alexei Starovoitov <[email protected]>:
Quoted 11 lines · On Thu, 5 Mar 2026 16:43:05 +0800 you wrote:
perf_linkcreates a system-wide perf event pinned to CPU 0 (pid=-1, cpu=0) and also pins the test thread to CPU 0. Under concurrent selftests this can lead to cross-test interference and CPU 0 contention, making the test flaky.Create a per-task perf event instead (pid=0, cpu=-1) and drop CPU pinning from
burn_cpu(). Usebarrier()to prevent the burn loop from being optimized away. Drop theserial_prefix so the test can run in parallel. Also remove the stale TODO comment.[...]
Here is the summary with links:
- [bpf-next] selftests/bpf: perf_link: avoid failures in concurrent mode
https://git.kernel.org/bpf/bpf-next/c/888329ba6c8b
You are awesome, thank you!
Signature · 2 lines
Deet-doot-dot, I am a bot. https://korg.docs.kernel.org/patchwork/pwbot.html