Several perf subcommands (sched stats, lock contention) use the pattern
of forking a workload child, calling evlist__start_workload() to uncork
it, and then calling pause() to wait for a signal (typically SIGCHLD
when the child exits, or SIGINT/SIGTERM from the user).

This pattern has a race condition: if the workload is very short-lived,
the child can exit and deliver SIGCHLD in the window between
evlist__start_workload() and pause(). Since pause() only returns when a
signal is received *while the process is suspended*, and SIGCHLD has
already been delivered and handled by the empty sighandler(), pause()
blocks indefinitely.

The fix replaces pause() with a simpler approach:

 - The signal handler now sets a 'volatile sig_atomic_t done' flag
   to record delivery.  The flag is reset before handler registration
   so that a signal arriving during the setup phase is not discarded.

 - Replace pause() with a unified loop:

       while (!done) {
           if (argc && waitpid(evlist->workload.pid, NULL, WNOHANG) > 0)
               break;
           sleep(1);
       }

   This handles both workload mode (child exit detected via waitpid)
   and system-wide mode (user sends SIGINT/SIGTERM setting done).
   Using WNOHANG avoids the SA_RESTART problem where a blocking
   waitpid() would auto-restart and ignore the done flag if the child
   doesn't exit on signal.

Three call sites are affected across two files:
  - perf_sched__schedstat_record() in builtin-sched.c
  - perf_sched__schedstat_live()   in builtin-sched.c
  - __cmd_contention()             in builtin-lock.c

The two pause() sites in builtin-kwork.c are NOT affected because they
do not register SIGCHLD or fork workload children; they only wait for
user-initiated SIGINT/SIGTERM.

v3: https://lore.kernel.org/all/20260422064700.12345-1-swapnil.sapkal@amd.com/
Changes since v3:
  - Replaced blocking waitpid() with waitpid(WNOHANG) in a unified
    loop that also checks 'done'.  Because perf is compiled with
    _GNU_SOURCE, glibc's signal() sets SA_RESTART, causing a blocking
    waitpid() to auto-restart after the handler returns.  If the
    workload ignores SIGINT, the tool would hang.  (Sashiko review)
  - Moved 'done = 0' before signal handler registration so that an
    early signal during the setup phase is not discarded by a later
    reset.  (Sashiko review)
  - Used single unified loop for both workload and system-wide cases
    as suggested by Namhyung.

v2: https://lore.kernel.org/all/20260409162249.25581-1-swapnil.sapkal@amd.com/
Changes since v2:
  - The signal handler now sets a 'volatile sig_atomic_t done' flag
    instead of being an empty function.  pause()/sigsuspend() alone
    cannot cover the window between signal() handler registration and
    sigprocmask(): a signal consumed during that unblocked window was
    silently lost.  The while(!done) loop detects such early delivery.
    (Sashiko review)
  - Replaced sigprocmask()/sigsuspend() approach with the much simpler
    waitpid() + while(!done) sleep(1) pattern as suggested by the
    maintainer.

v1: https://lore.kernel.org/all/20260401064114.141066-1-swapnil.sapkal@amd.com/
Changes since v1:
  - Moved sigprocmask() to after evlist__prepare_workload() so the
    forked child does not inherit a blocked SIGCHLD mask, which would
    break workloads relying on SIGCHLD (e.g., Node.js, Python asyncio).
    (Sashiko review)
  - Block SIGINT and SIGTERM in addition to SIGCHLD to prevent an
    early Ctrl+C during setup from being consumed before sigsuspend().
  - Added signal mask restoration on error paths.
    (Sashiko review)

Swapnil Sapkal (3):
  perf sched stats: Fix SIGCHLD vs pause() race in schedstat_record()
  perf sched stats: Fix SIGCHLD vs pause() race in schedstat_live()
  perf lock contention: Fix SIGCHLD vs pause() race in __cmd_contention()

 tools/perf/builtin-lock.c  | 12 ++++++++++--
 tools/perf/builtin-sched.c | 20 ++++++++++++++++----
 2 files changed, 26 insertions(+), 6 deletions(-)

-- 
2.43.0

On Wed, May 20, 2026 at 10:20:14AM +0000, Swapnil Sapkal wrote:
> Several perf subcommands (sched stats, lock contention) use the pattern
> of forking a workload child, calling evlist__start_workload() to uncork
> it, and then calling pause() to wait for a signal (typically SIGCHLD
> when the child exits, or SIGINT/SIGTERM from the user).
> 
> This pattern has a race condition: if the workload is very short-lived,
> the child can exit and deliver SIGCHLD in the window between
> evlist__start_workload() and pause(). Since pause() only returns when a
> signal is received *while the process is suspended*, and SIGCHLD has
> already been delivered and handled by the empty sighandler(), pause()
> blocks indefinitely.
> 
> The fix replaces pause() with a simpler approach:
> 
>  - The signal handler now sets a 'volatile sig_atomic_t done' flag
>    to record delivery.  The flag is reset before handler registration
>    so that a signal arriving during the setup phase is not discarded.
> 
>  - Replace pause() with a unified loop:
> 
>        while (!done) {
>            if (argc && waitpid(evlist->workload.pid, NULL, WNOHANG) > 0)
>                break;
>            sleep(1);
>        }
> 
>    This handles both workload mode (child exit detected via waitpid)
>    and system-wide mode (user sends SIGINT/SIGTERM setting done).
>    Using WNOHANG avoids the SA_RESTART problem where a blocking
>    waitpid() would auto-restart and ignore the done flag if the child
>    doesn't exit on signal.
> 
> Three call sites are affected across two files:
>   - perf_sched__schedstat_record() in builtin-sched.c
>   - perf_sched__schedstat_live()   in builtin-sched.c
>   - __cmd_contention()             in builtin-lock.c
> 
> The two pause() sites in builtin-kwork.c are NOT affected because they
> do not register SIGCHLD or fork workload children; they only wait for
> user-initiated SIGINT/SIGTERM.

Thanks, applied to perf-tools-next, for v7.2.

- Arnaldo