Skip to content

ZTS: pcntl signal handled by one PHP thread ends the process or is lost when delivered elsewhere #25549

Open

Description

@phcorp

Description

Signal dispositions are process-wide, but under ZTS the PHP handler for a signal belongs to the thread that registered it. With several threads running PHP requests (an embedding application, for instance), a signal handled by one thread can end the process or be lost. GH-25542 fixes delivery to threads that never ran PHP, but these two cases remain.

1. A signal lands on a PHP thread that has no handler for it.

  1. Thread A runs a request and calls pcntl_signal(SIGUSR1, $handler).
  2. Thread B runs a request and has no handler for SIGUSR1.
  3. SIGUSR1 is delivered to thread B.

2. Two threads handle the same signal and one finishes first.

  1. Threads A and B each run a request and call pcntl_signal(SIGUSR1, $handler).
  2. Thread B finishes its request.
  3. SIGUSR1 is delivered to a thread that never ran PHP, while thread A is still running.

I reproduced both with a small embed SAPI program that runs each request on its own pthread and calls raise() on the target thread. Each result below is 10 of 10 runs, on a debug ZTS build:

Case With zend signals With --disable-zend-signals
1. PHP thread without a handler Process terminated by SIGUSR1 Signal lost, A's handler never runs
2. Same signal, one thread finished Process terminated by SIGUSR1 Process terminated by SIGUSR1

I expected thread A's handler to run in every case.

The results are the same with and without GH-25542. In case 2 without zend signals, thread B's RSHUTDOWN resets the signal to SIG_DFL for the whole process while thread A still relies on it.

This looks related to the discussion in GH-10193 about making the zend signal state process-wide rather than per thread.

PHP Version

PHP-8.4 branch (8.4.29-dev)

Operating System

macOS (arm64)

Activity

  1. phcorp commented on Nov 18, 2026

    @phcorp
    Author

    Re-tested with the current version of GH-25542 (f582dba): both cases behave as described above.

    One related detail for whoever looks at this: on macOS, sigprocmask() changes the signal mask of every thread in the process, which I confirmed with a small C program. pcntl_signal_dispatch() uses it, so while one thread dispatches, signals are blocked in the other threads as well.

  2. added a commit that references this issue on Nov 18, 2026
    198e0d1
  3. phcorp commented on Nov 18, 2026

    @phcorp
    Author

    GH-25542 now includes a fix for this issue, with a regression test.

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Type

    No type

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions