Skip to content

Fix GH-25515: FPM ondemand counts a child being idle-killed as idle - #25517

Open
lazerg wants to merge 1 commit into
php:PHP-8.4from
lazerg:fix/gh-25515-ondemand-idle-kill
Open

lazerg wants to merge 1 commit into
php:PHP-8.4from
lazerg:fix/gh-25515-ondemand-idle-kill

Conversation

@lazerg

@lazerg lazerg commented Nov 17, 2026

Copy link
Copy Markdown
Contributor

When the master has already sent SIGQUIT to the last idle child for pm.process_idle_timeout, fpm_pctl_on_socket_accept() still counts that child as idle and doesn't fork. The child exits without accepting, and since the ondemand event is edge-triggered the connection sits in the backlog until another client connects. Skipping children with idle_kill set makes the master fork a new child for it.

I reproduced it with the reporter's script on Linux (2 of 10 pools hung within 5 minutes, none after the fix). There's no test because the window is too small to hit reliably in CI.

Fixes #25515

@lazerg
lazerg requested a review from bukka as a code owner November 17, 2026 12:43
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant