Skip to contentSkip to conversation

perf record: fix poll storm when monitored threads exit

PATCH Applied Jiawei Sun 3 participants 3 messages Jul 7 – Jul 8 perf events

lore
Share & tools
lore

Jiawei Sun

author
#0 lore raw

PATCH 1/1 perf record: fix poll storm when monitored threads exit

When perf record samples a multi-threaded process and one of the target threads exits during the session, perf itself may start burning 100% CPU (up to 200% across two cores) until the session ends. A single dead fd is sufficient to trigger this; it can be reproduced with 15 pthreads in a compute loop where one thread exits halfway through.

The root cause is two independent instances of the same defect: dead perf_event ring-buffer fds are left in a pollfd array. When a monitored thread exits, the kernel closes its ring-buffer fd, which then returns POLLHUP. POSIX specifies that poll() always reports POLLHUP and POLLERR regardless of the events mask, so any dead fd left in the array makes poll() return immediately every time, spinning in a tight loop:

3 seconds: 256,600 poll() calls, 0 context switches, only 21 write() Woken up count goes from ~0 to 1,300,000+

There are two affected poll paths, fixed together here:

1. Record main loop, via fdarray__filter() (tools/lib/api/fd/array.c). Since commit 59b4412f27f1 ("libperf: Avoid internal moving of fdarray fds") it only zeroes events/revents without setting fd to -1, so poll() keeps reporting POLLHUP for the entry. Setting fd = -1 makes poll() skip it, matching the pattern already used in the control-fd path at tools/perf/builtin-record.c:1673.

2. BPF sideband thread, perf_evlist__poll_thread() (tools/perf/util/sideband_evlist.c). This thread polls for PERF_RECORD_BPF_EVENT but, unlike the main record loop, never calls fdarray__filter() at all, so dead fds accumulate forever and it spins at 100% CPU:

Before fix: dJiffies=101, wchan=0 (running)
After fix: dJiffies=0, wchan=do_sys_poll (blocking)

Fixed by calling the existing evlist__filter_pollfd() helper after evlist__poll(), mirroring the main record loop. <poll.h> is included for the POLLERR/POLLHUP macros (previously unused there).

The two fixes compose: fix 1 makes poll() ignore dead fds (fd=-1); fix 2 ensures the sideband thread actually performs the filtering. Both paths are affected in all kernels from v5.1/v5.9 to the current master (7.2-rc1); the source of both functions is byte-identical across them.

BPF event recording is preserved: after the fix, perf.data still contains PERF_RECORD_BPF_EVENT records and bpf_prog_info entries.

Verified on perf 6.1.76, 6.6.143 and 7.2-rc1 with a minimal reproducer (Woken up 1,300,000 -> 3, CPU 100% -> 0%) and an A/B orthogonal test:
keeping the unpatched binary but preventing the target thread from exiting also makes the storm disappear, confirming the trigger.

Fixes 59b4412f27f1 ("libperf: Avoid internal moving of fdarray fds")
Fixes 657ee5531903 ("perf evlist: Introduce side band thread")
---
 tools/lib/api/fd/array.c          |  6 ++++++
 tools/perf/util/sideband_evlist.c | 14 ++++++++++++++
 2 files changed, 20 insertions(+)
tools/lib/api/fd/array.c +6 -0
index f0f195207..ffe8272af 100644
@@ -122,6 +122,12 @@ int fdarray__filter(struct fdarray *fda, short revents,
 			if (entry_destructor)
 				entry_destructor(fda, fd, arg);
 
+			/*
+			 * Set fd to -1 so poll() ignores this entry; otherwise
+			 * POLLHUP/POLLERR are still reported for events=0 fds
+			 * (POSIX: always checked), causing a poll storm.
+			 */
+			fda->entries[fd].fd = -1;
 			fda->entries[fd].revents = fda->entries[fd].events = 0;
 			continue;
 		}
tools/perf/util/sideband_evlist.c +14 -0
index c07dacf3c..ba043db6c 100644
@@ -8,6 +8,7 @@
 #include <perf/mmap.h>
 #include <linux/perf_event.h>
 #include <limits.h>
+#include <poll.h>
 #include <pthread.h>
 #include <sched.h>
 #include <stdbool.h>
@@ -55,6 +56,19 @@ static void *perf_evlist__poll_thread(void *arg)
 		if (!draining)
 			evlist__poll(evlist, 1000);
 
+		/*
+		 * When a thread of the monitored target exits, its per-cpu
+		 * ring-buffer fd is closed and starts returning POLLHUP. Such
+		 * dead fds are never requested for POLLIN, but poll() reports
+		 * POLLHUP/POLLERR unconditionally, so leaving them in the
+		 * pollfd array makes the following evlist__poll() return
+		 * immediately forever, spinning this thread at 100% CPU.
+		 *
+		 * Filter them out here, mirroring what the 'perf record' main
+		 * loop does after fdarray__poll().
+		 */
+		evlist__filter_pollfd(evlist, POLLERR | POLLHUP);
+
 		for (i = 0; i < evlist__core(evlist)->nr_mmaps; i++) {
 			struct mmap *map = &evlist__mmap(evlist)[i];
 			union perf_event *event;
Signature · 1 line
2.39.5
reviewer Reviewed-by on v1 1/1
#1 lore raw
Quoted 47 lines · Jiawei Sun

When perf record samples a multi-threaded process and one of the target threads exits during the session, perf itself may start burning 100% CPU (up to 200% across two cores) until the session ends. A single dead fd is sufficient to trigger this; it can be reproduced with 15 pthreads in a compute loop where one thread exits halfway through.

The root cause is two independent instances of the same defect: dead perf_event ring-buffer fds are left in a pollfd array. When a monitored thread exits, the kernel closes its ring-buffer fd, which then returns POLLHUP. POSIX specifies that poll() always reports POLLHUP and POLLERR regardless of the events mask, so any dead fd left in the array makes poll() return immediately every time, spinning in a tight loop:

3 seconds: 256,600 poll() calls, 0 context switches, only 21 write() Woken up count goes from ~0 to 1,300,000+

There are two affected poll paths, fixed together here:

1. Record main loop, via fdarray__filter() (tools/lib/api/fd/array.c). Since commit 59b4412f27f1 ("libperf: Avoid internal moving of fdarray fds") it only zeroes events/revents without setting fd to -1, so poll() keeps reporting POLLHUP for the entry. Setting fd = -1 makes poll() skip it, matching the pattern already used in the control-fd path at tools/perf/builtin-record.c:1673.

2. BPF sideband thread, perf_evlist__poll_thread() (tools/perf/util/sideband_evlist.c). This thread polls for PERF_RECORD_BPF_EVENT but, unlike the main record loop, never calls fdarray__filter() at all, so dead fds accumulate forever and it spins at 100% CPU:

Before fix: dJiffies=101, wchan=0 (running)
After fix: dJiffies=0, wchan=do_sys_poll (blocking)

Fixed by calling the existing evlist__filter_pollfd() helper after evlist__poll(), mirroring the main record loop. <poll.h> is included for the POLLERR/POLLHUP macros (previously unused there).

The two fixes compose: fix 1 makes poll() ignore dead fds (fd=-1); fix 2 ensures the sideband thread actually performs the filtering. Both paths are affected in all kernels from v5.1/v5.9 to the current master (7.2-rc1); the source of both functions is byte-identical across them.

BPF event recording is preserved: after the fix, perf.data still contains PERF_RECORD_BPF_EVENT records and bpf_prog_info entries.

Quoted 8 lines

Verified on perf 6.1.76, 6.6.143 and 7.2-rc1 with a minimal reproducer (Woken up 1,300,000 -> 3, CPU 100% -> 0%) and an A/B orthogonal test:
keeping the unpatched binary but preventing the target thread from exiting also makes the storm disappear, confirming the trigger.

Fixes 59b4412f27f1 ("libperf: Avoid internal moving of fdarray fds")
Fixes 657ee5531903 ("perf evlist: Introduce side band thread")

This all makes sense and reviewing the surrounding code I don't see issues. Sashiko is also happy:
https://sashiko.dev/#/patchset/20260707153801.315352-1-abyssmystery%40gmail.com

Thanks,
Ian

Quoted 57 lines
---
 tools/lib/api/fd/array.c          |  6 ++++++
 tools/perf/util/sideband_evlist.c | 14 ++++++++++++++
 2 files changed, 20 insertions(+)
tools/lib/api/fd/array.c +6 -0
index f0f195207..ffe8272af 100644
@@ -122,6 +122,12 @@ int fdarray__filter(struct fdarray *fda, short revents,
                        if (entry_destructor)
                                entry_destructor(fda, fd, arg);
 
+                       /*
+                        * Set fd to -1 so poll() ignores this entry; otherwise
+                        * POLLHUP/POLLERR are still reported for events=0 fds
+                        * (POSIX: always checked), causing a poll storm.
+                        */
+                       fda->entries[fd].fd = -1;
                        fda->entries[fd].revents = fda->entries[fd].events = 0;
                        continue;
                }
tools/perf/util/sideband_evlist.c +14 -1
index c07dacf3c..ba043db6c 100644
@@ -8,6 +8,7 @@
 #include <perf/mmap.h>
 #include <linux/perf_event.h>
 #include <limits.h>
+#include <poll.h>
 #include <pthread.h>
 #include <sched.h>
 #include <stdbool.h>
@@ -55,6 +56,19 @@ static void *perf_evlist__poll_thread(void *arg)
                if (!draining)
                        evlist__poll(evlist, 1000);
 
+               /*
+                * When a thread of the monitored target exits, its per-cpu
+                * ring-buffer fd is closed and starts returning POLLHUP. Such
+                * dead fds are never requested for POLLIN, but poll() reports
+                * POLLHUP/POLLERR unconditionally, so leaving them in the
+                * pollfd array makes the following evlist__poll() return
+                * immediately forever, spinning this thread at 100% CPU.
+                *
+                * Filter them out here, mirroring what the 'perf record' main
+                * loop does after fdarray__poll().
+                */
+               evlist__filter_pollfd(evlist, POLLERR | POLLHUP);
+
                for (i = 0; i < evlist__core(evlist)->nr_mmaps; i++) {
                        struct mmap *map = &evlist__mmap(evlist)[i];
                        union perf_event *event;
--

2.39.5

maintainer Applied
#2 lore raw
In reply to Jiawei Sun
Quoted 6 lines · Jiawei Sun

When perf record samples a multi-threaded process and one of the target threads exits during the session, perf itself may start burning 100% CPU (up to 200% across two cores) until the session ends. A single dead fd is sufficient to trigger this; it can be reproduced with 15 pthreads in a compute loop where one thread exits halfway through.

Quoted 8 lines

The root cause is two independent instances of the same defect: dead perf_event ring-buffer fds are left in a pollfd array. When a monitored thread exits, the kernel closes its ring-buffer fd, which then returns POLLHUP. POSIX specifies that poll() always reports POLLHUP and POLLERR regardless of the events mask, so any dead fd left in the array makes poll() return immediately every time, spinning in a tight loop:

[...]

Applied to perf-tools-next, thanks!

Best regards,
Namhyung