Optimize memory allocation when receiving packets #16
Assignees
Comments
zanebeckwith
pushed a commit
that referenced
this issue
Aug 5, 2020
When experimenting with bpf_send_signal() helper in our production environment (5.2 based), we experienced a deadlock in NMI mode: #5 [ffffc9002219f770] queued_spin_lock_slowpath at ffffffff8110be24 #6 [ffffc9002219f770] _raw_spin_lock_irqsave at ffffffff81a43012 #7 [ffffc9002219f780] try_to_wake_up at ffffffff810e7ecd #8 [ffffc9002219f7e0] signal_wake_up_state at ffffffff810c7b55 #9 [ffffc9002219f7f0] __send_signal at ffffffff810c8602 #10 [ffffc9002219f830] do_send_sig_info at ffffffff810ca31a #11 [ffffc9002219f868] bpf_send_signal at ffffffff8119d227 #12 [ffffc9002219f988] bpf_overflow_handler at ffffffff811d4140 #13 [ffffc9002219f9e0] __perf_event_overflow at ffffffff811d68cf #14 [ffffc9002219fa10] perf_swevent_overflow at ffffffff811d6a09 #15 [ffffc9002219fa38] ___perf_sw_event at ffffffff811e0f47 #16 [ffffc9002219fc30] __schedule at ffffffff81a3e04d #17 [ffffc9002219fc90] schedule at ffffffff81a3e219 #18 [ffffc9002219fca0] futex_wait_queue_me at ffffffff8113d1b9 #19 [ffffc9002219fcd8] futex_wait at ffffffff8113e529 #20 [ffffc9002219fdf0] do_futex at ffffffff8113ffbc #21 [ffffc9002219fec0] __x64_sys_futex at ffffffff81140d1c #22 [ffffc9002219ff38] do_syscall_64 at ffffffff81002602 torvalds#23 [ffffc9002219ff50] entry_SYSCALL_64_after_hwframe at ffffffff81c00068 The above call stack is actually very similar to an issue reported by Commit eac9153 ("bpf/stackmap: Fix deadlock with rq_lock in bpf_get_stack()") by Song Liu. The only difference is bpf_send_signal() helper instead of bpf_get_stack() helper. The above deadlock is triggered with a perf_sw_event. Similar to Commit eac9153, the below almost identical reproducer used tracepoint point sched/sched_switch so the issue can be easily caught. /* stress_test.c */ #include <stdio.h> #include <stdlib.h> #include <sys/mman.h> #include <pthread.h> #include <sys/types.h> #include <sys/stat.h> #include <fcntl.h> #define THREAD_COUNT 1000 char *filename; void *worker(void *p) { void *ptr; int fd; char *pptr; fd = open(filename, O_RDONLY); if (fd < 0) return NULL; while (1) { struct timespec ts = {0, 1000 + rand() % 2000}; ptr = mmap(NULL, 4096 * 64, PROT_READ, MAP_PRIVATE, fd, 0); usleep(1); if (ptr == MAP_FAILED) { printf("failed to mmap\n"); break; } munmap(ptr, 4096 * 64); usleep(1); pptr = malloc(1); usleep(1); pptr[0] = 1; usleep(1); free(pptr); usleep(1); nanosleep(&ts, NULL); } close(fd); return NULL; } int main(int argc, char *argv[]) { void *ptr; int i; pthread_t threads[THREAD_COUNT]; if (argc < 2) return 0; filename = argv[1]; for (i = 0; i < THREAD_COUNT; i++) { if (pthread_create(threads + i, NULL, worker, NULL)) { fprintf(stderr, "Error creating thread\n"); return 0; } } for (i = 0; i < THREAD_COUNT; i++) pthread_join(threads[i], NULL); return 0; } and the following command: 1. run `stress_test /bin/ls` in one windown 2. hack bcc trace.py with the following change: --- a/tools/trace.py +++ b/tools/trace.py @@ -513,6 +513,7 @@ BPF_PERF_OUTPUT(%s); __data.tgid = __tgid; __data.pid = __pid; bpf_get_current_comm(&__data.comm, sizeof(__data.comm)); + bpf_send_signal(10); %s %s %s.perf_submit(%s, &__data, sizeof(__data)); 3. in a different window run ./trace.py -p $(pidof stress_test) t:sched:sched_switch The deadlock can be reproduced in our production system. Similar to Song's fix, the fix is to delay sending signal if irqs is disabled to avoid deadlocks involving with rq_lock. With this change, my above stress-test in our production system won't cause deadlock any more. I also implemented a scale-down version of reproducer in the selftest (a subsequent commit). With latest bpf-next, it complains for the following potential deadlock. [ 32.832450] -> #1 (&p->pi_lock){-.-.}: [ 32.833100] _raw_spin_lock_irqsave+0x44/0x80 [ 32.833696] task_rq_lock+0x2c/0xa0 [ 32.834182] task_sched_runtime+0x59/0xd0 [ 32.834721] thread_group_cputime+0x250/0x270 [ 32.835304] thread_group_cputime_adjusted+0x2e/0x70 [ 32.835959] do_task_stat+0x8a7/0xb80 [ 32.836461] proc_single_show+0x51/0xb0 ... [ 32.839512] -> #0 (&(&sighand->siglock)->rlock){....}: [ 32.840275] __lock_acquire+0x1358/0x1a20 [ 32.840826] lock_acquire+0xc7/0x1d0 [ 32.841309] _raw_spin_lock_irqsave+0x44/0x80 [ 32.841916] __lock_task_sighand+0x79/0x160 [ 32.842465] do_send_sig_info+0x35/0x90 [ 32.842977] bpf_send_signal+0xa/0x10 [ 32.843464] bpf_prog_bc13ed9e4d3163e3_send_signal_tp_sched+0x465/0x1000 [ 32.844301] trace_call_bpf+0x115/0x270 [ 32.844809] perf_trace_run_bpf_submit+0x4a/0xc0 [ 32.845411] perf_trace_sched_switch+0x10f/0x180 [ 32.846014] __schedule+0x45d/0x880 [ 32.846483] schedule+0x5f/0xd0 ... [ 32.853148] Chain exists of: [ 32.853148] &(&sighand->siglock)->rlock --> &p->pi_lock --> &rq->lock [ 32.853148] [ 32.854451] Possible unsafe locking scenario: [ 32.854451] [ 32.855173] CPU0 CPU1 [ 32.855745] ---- ---- [ 32.856278] lock(&rq->lock); [ 32.856671] lock(&p->pi_lock); [ 32.857332] lock(&rq->lock); [ 32.857999] lock(&(&sighand->siglock)->rlock); Deadlock happens on CPU0 when it tries to acquire &sighand->siglock but it has been held by CPU1 and CPU1 tries to grab &rq->lock and cannot get it. This is not exactly the callstack in our production environment, but sympotom is similar and both locks are using spin_lock_irqsave() to acquire the lock, and both involves rq_lock. The fix to delay sending signal when irq is disabled also fixed this issue. Signed-off-by: Yonghong Song <[email protected]> Signed-off-by: Alexei Starovoitov <[email protected]> Cc: Song Liu <[email protected]> Link: https://lore.kernel.org/bpf/[email protected]
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Currently whenever a command comes in we allocate a work item and packet, one is free'd in the work function and the other in the proxy call (functions are just expected to know this). This is inflexible and won't be scalable to high volume applications.
In general
So rather than perpetual workqueue iterations for tiny ops, it may make more sense to have a single thread handle this and work with persistent buffers rather than many fragmented allocs.
The text was updated successfully, but these errors were encountered: