KCSAN: data-race in event_sched_out / perf_poll

Status: moderation: reported on 2026/11/20 00:06
Subsystems: perf
Labels: race:benign prio:low
[Documentation on labels]
Reported-by: syzbot+d2f5271536cfba9f371d@syzkaller.appspotmail.com
First crash: 2d22h, last: 2d06h
✨ AI-generated root cause summary:

perf_poll() in kernel/events/core.c performs unmarked reads of event->state (at lines 6322 and 6327, as well as in is_event_hup() at line 6260) without holding the event context lock ctx->lock. Concurrently—or when preempted during perf_poll(), as shown in the KCSAN report—the scheduler's perf_event_task_sched_out() -> event_sched_out() path calls perf_event_set_state(), which updates event->state using WRITE_ONCE(event->state, state) under ctx->lock. This mismatch between the marked write in perf_event_set_state() and the unmarked reads in perf_poll() triggers the KCSAN data-race report, whereas line 6333 of perf_poll() already uses READ_ONCE(event->state).

▶ ▼ ✨ Full AI root cause analysis
Based on kernel commit 666539eecdf4:

Root Cause Analysis

The KCSAN data-race occurs in kernel/events/core.c between lockless reads of struct perf_event::state (enum perf_event_state) in perf_poll() and marked writes to the same field in perf_event_set_state().

Involved Functions and Fields

Mechanism

  1. perf_poll() is invoked on a perf event file descriptor without holding ctx->lock (here via io_uring -> io_file_supports_nowait() -> vfs_poll()).
  2. At kernel/events/core.c:6322 (and again at line 6327, as well as in is_event_hup() at line 6260), perf_poll() performs plain, unmarked reads of event->state:
    if (event->state <= PERF_EVENT_STATE_REVOKED)
        return EPOLLERR;
    
    poll_wait(file, &event->waitq, wait);
    
    if (event->state <= PERF_EVENT_STATE_REVOKED)
        return EPOLLERR;
    
    By contrast, line 6333 in the same function already uses READ_ONCE(event->state).
  3. Concurrently—or in this report, when task 5259 was preempted inside KCSAN's watchpoint delay in perf_poll()—the context-switch path (prepare_task_switch() -> perf_event_task_sched_out() -> event_sched_out()) updates event->state under ctx->lock via perf_event_set_state():
    WRITE_ONCE(event->state, state); /* kernel/events/core.c:790 */
    
  4. Because perf_poll() reads event->state without READ_ONCE() while perf_event_set_state() writes to it concurrently under ctx->lock, KCSAN reports a data race between the marked write in perf_event_set_state() and the unmarked read in perf_poll().

Fix Direction

Use READ_ONCE(event->state) consistently for the lockless reads of event->state in perf_poll() (kernel/events/core.c:6322, 6327) and is_event_hup() (kernel/events/core.c:6260), matching line 6333 (and perf_event_read() / perf_sigtrap()).

▶ ▼ ✨ AI Jobs (2)
ID Workflow Result Correct Bug Created Started Finished Revision
28452c90 assessment-security DenialOfService: ❌ Exploitable: ❌ FilesystemTrigger: ❌ FuzzingArtifact: ❌ MemoryCorruption: ❌ NetworkNeedsSetup: ❌ NetworkTrigger: ❌ PeripheralTrigger: ❌ RemoteTrigger: ❌ Unprivileged: ✅ UserNamespace: ✅ VMGuestTrigger: ❌ VMHostTrigger: ❌ ❓ KCSAN: data-race in event_sched_out / perf_poll 2026/11/19 21:17 2026/11/19 21:17 2026/11/19 21:19 229865fb
8f09d514 assessment-kcsan Benign: ✅ ❓ KCSAN: data-race in event_sched_out / perf_poll 2026/11/19 21:15 2026/11/19 21:15 2026/11/19 21:17 229865fb

Sample crash report:
==================================================================
BUG: KCSAN: data-race in event_sched_out / perf_poll

write (marked) to 0xffff88811b78a4a8 of 4 bytes by task 5259 on cpu 1:
 perf_event_set_state kernel/events/core.c:790 [inline]
 event_sched_out+0x534/0x740 kernel/events/core.c:2506
 group_sched_out kernel/events/core.c:2530 [inline]
 __pmu_ctx_sched_out+0x3e7/0x530 kernel/events/core.c:3582
 ctx_sched_out+0x273/0x2d0 kernel/events/core.c:3673
 task_ctx_sched_out+0x4d/0x70 kernel/events/core.c:2982
 perf_event_context_sched_out kernel/events/core.c:3883 [inline]
 __perf_event_task_sched_out+0x441/0x710 kernel/events/core.c:3985
 perf_event_task_sched_out include/linux/perf_event.h:1668 [inline]
 prepare_task_switch kernel/sched/core.c:5366 [inline]
 context_switch kernel/sched/core.c:5529 [inline]
 __schedule+0xd04/0xd30 kernel/sched/core.c:7386
 preempt_schedule_common+0x12/0x40 kernel/sched/core.c:7565
 preempt_schedule_thunk+0x16/0x40 arch/x86/entry/thunk.S:12
 delay_tsc+0x64/0xa0 arch/x86/lib/delay.c:75
 udelay include/asm-generic/delay.h:64 [inline]
 delay_access kernel/kcsan/core.c:332 [inline]
 kcsan_setup_watchpoint+0x255/0x3f0 kernel/kcsan/core.c:622
 perf_poll+0x36/0x200 kernel/events/core.c:6322
 vfs_poll include/linux/poll.h:82 [inline]
 io_file_supports_nowait io_uring/rw.c:48 [inline]
 __io_read+0x30b/0xbd0 io_uring/rw.c:924
 io_read+0x4a/0x1b0 io_uring/rw.c:1017
 __io_issue_sqe+0xfd/0x2e0 io_uring/io_uring.c:1514
 io_issue_sqe+0x1e7/0xd20 io_uring/io_uring.c:1537
 io_queue_sqe io_uring/io_uring.c:1794 [inline]
 io_req_task_submit+0x46/0xa0 io_uring/io_uring.c:1143
 tctx_task_work_run+0x3cd/0x550 io_uring/tw.c:96
 tctx_task_work+0x3f/0x60 io_uring/tw.c:133
 task_work_run+0x145/0x1a0 kernel/task_work.c:233
 get_signal+0xc12/0xd30 kernel/signal.c:2885
 arch_do_signal_or_restart+0x96/0x480 arch/x86/kernel/signal.c:337
 __exit_to_user_mode_loop kernel/entry/common.c:66 [inline]
 exit_to_user_mode_loop+0x161/0x940 kernel/entry/common.c:101
 __exit_to_user_mode_prepare include/linux/irq-entry-common.h:207 [inline]
 syscall_exit_to_user_mode_prepare include/linux/irq-entry-common.h:230 [inline]
 syscall_exit_to_user_mode include/linux/entry-common.h:343 [inline]
 do_syscall_64+0x23b/0x3a0 arch/x86/entry/syscall_64.c:107
 entry_SYSCALL_64_after_hwframe+0x77/0x7f

read to 0xffff88811b78a4a8 of 4 bytes by task 5259 on cpu 1:
 perf_poll+0x36/0x200 kernel/events/core.c:6322
 vfs_poll include/linux/poll.h:82 [inline]
 io_file_supports_nowait io_uring/rw.c:48 [inline]
 __io_read+0x30b/0xbd0 io_uring/rw.c:924
 io_read+0x4a/0x1b0 io_uring/rw.c:1017
 __io_issue_sqe+0xfd/0x2e0 io_uring/io_uring.c:1514
 io_issue_sqe+0x1e7/0xd20 io_uring/io_uring.c:1537
 io_queue_sqe io_uring/io_uring.c:1794 [inline]
 io_req_task_submit+0x46/0xa0 io_uring/io_uring.c:1143
 tctx_task_work_run+0x3cd/0x550 io_uring/tw.c:96
 tctx_task_work+0x3f/0x60 io_uring/tw.c:133
 task_work_run+0x145/0x1a0 kernel/task_work.c:233
 get_signal+0xc12/0xd30 kernel/signal.c:2885
 arch_do_signal_or_restart+0x96/0x480 arch/x86/kernel/signal.c:337
 __exit_to_user_mode_loop kernel/entry/common.c:66 [inline]
 exit_to_user_mode_loop+0x161/0x940 kernel/entry/common.c:101
 __exit_to_user_mode_prepare include/linux/irq-entry-common.h:207 [inline]
 syscall_exit_to_user_mode_prepare include/linux/irq-entry-common.h:230 [inline]
 syscall_exit_to_user_mode include/linux/entry-common.h:343 [inline]
 do_syscall_64+0x23b/0x3a0 arch/x86/entry/syscall_64.c:107
 entry_SYSCALL_64_after_hwframe+0x77/0x7f

Reported by Kernel Concurrency Sanitizer on:
CPU: 1 UID: 0 PID: 5259 Comm: syz.6.503 Not tainted syzkaller #0 PREEMPT(lazy) 
Hardware name: Google Google Compute Engine/Google Compute Engine, BIOS Google 11/06/2026
==================================================================

Crashes (4):
Time Kernel Commit Syzkaller Config Log Report Syz repro C repro VM info Assets (help?) Manager Title
2026/11/17 18:35 upstream 666539eecdf4 122f94d5 .config console log report info [disk image] [vmlinux] [kernel image] ci2-upstream-kcsan-gce KCSAN: data-race in event_sched_out / perf_poll
2026/11/17 12:48 upstream 666539eecdf4 43af6cac .config console log report info [disk image] [vmlinux] [kernel image] ci2-upstream-kcsan-gce KCSAN: data-race in event_sched_out / perf_poll
2026/11/17 03:02 upstream 666539eecdf4 a3e64cfc .config console log report info [disk image] [vmlinux] [kernel image] ci2-upstream-kcsan-gce KCSAN: data-race in event_sched_out / perf_poll
2026/11/17 03:02 upstream 666539eecdf4 a3e64cfc .config console log report info [disk image] [vmlinux] [kernel image] ci2-upstream-kcsan-gce KCSAN: data-race in event_sched_out / perf_poll
* Struck through repros no longer work on HEAD.