PHP‑FPM workers exceed pm.max_requests

[qcybb]

Description

PHP‑FPM workers consistently exceed the configured pm.max_requests value under sustained load. Even with pm.max_requests = 5000, individual workers regularly process 8000–10,000+ requests before recycling.

$ sudo php-fpm -tt
[07-Aug-2026 20:20:46] NOTICE: [global]
[07-Aug-2026 20:20:46] NOTICE:  pid = /var/run/php-fpm.pid
[07-Aug-2026 20:20:46] NOTICE:  error_log = /var/log/php-fpm/php-fpm.log
[07-Aug-2026 20:20:46] NOTICE:  syslog.ident = php-fpm
[07-Aug-2026 20:20:46] NOTICE:  syslog.facility = 24
[07-Aug-2026 20:20:46] NOTICE:  log_buffering = yes
[07-Aug-2026 20:20:46] NOTICE:  log_level = unknown value
[07-Aug-2026 20:20:46] NOTICE:  log_limit = 1024
[07-Aug-2026 20:20:46] NOTICE:  emergency_restart_interval = 60s
[07-Aug-2026 20:20:46] NOTICE:  emergency_restart_threshold = 10
[07-Aug-2026 20:20:46] NOTICE:  process_control_timeout = 10s
[07-Aug-2026 20:20:46] NOTICE:  process.max = 0
[07-Aug-2026 20:20:46] NOTICE:  process.priority = undefined
[07-Aug-2026 20:20:46] NOTICE:  daemonize = yes
[07-Aug-2026 20:20:46] NOTICE:  rlimit_files = 0
[07-Aug-2026 20:20:46] NOTICE:  rlimit_core = 0
[07-Aug-2026 20:20:46] NOTICE:  events.mechanism = kqueue
[07-Aug-2026 20:20:46] NOTICE:
[07-Aug-2026 20:20:46] NOTICE: [www]
[07-Aug-2026 20:20:46] NOTICE:  prefix = undefined
[07-Aug-2026 20:20:46] NOTICE:  user = www
[07-Aug-2026 20:20:46] NOTICE:  group = www
[07-Aug-2026 20:20:46] NOTICE:  listen = /var/run/php-fpm.sock
[07-Aug-2026 20:20:46] NOTICE:  listen.backlog = -1
[07-Aug-2026 20:20:46] NOTICE:  listen.owner = nginx
[07-Aug-2026 20:20:46] NOTICE:  listen.group = nginx
[07-Aug-2026 20:20:46] NOTICE:  listen.mode = 0660
[07-Aug-2026 20:20:46] NOTICE:  listen.allowed_clients = 127.0.0.1
[07-Aug-2026 20:20:46] NOTICE:  listen.setfib = -1
[07-Aug-2026 20:20:46] NOTICE:  process.priority = undefined
[07-Aug-2026 20:20:46] NOTICE:  process.dumpable = no
[07-Aug-2026 20:20:46] NOTICE:  pm = dynamic
[07-Aug-2026 20:20:46] NOTICE:  pm.max_children = 50
[07-Aug-2026 20:20:46] NOTICE:  pm.start_servers = 10
[07-Aug-2026 20:20:46] NOTICE:  pm.min_spare_servers = 10
[07-Aug-2026 20:20:46] NOTICE:  pm.max_spare_servers = 25
[07-Aug-2026 20:20:46] NOTICE:  pm.max_spawn_rate = 32
[07-Aug-2026 20:20:46] NOTICE:  pm.process_idle_timeout = 10
[07-Aug-2026 20:20:46] NOTICE:  pm.max_requests = 5000
[07-Aug-2026 20:20:46] NOTICE:  pm.status_path = /php-status
[07-Aug-2026 20:20:46] NOTICE:  pm.status_listen = undefined
[07-Aug-2026 20:20:46] NOTICE:  ping.path = undefined
[07-Aug-2026 20:20:46] NOTICE:  ping.response = undefined
[07-Aug-2026 20:20:46] NOTICE:  access.log = undefined
[07-Aug-2026 20:20:46] NOTICE:  access.format = undefined
[07-Aug-2026 20:20:46] NOTICE:  slowlog = undefined
[07-Aug-2026 20:20:46] NOTICE:  request_slowlog_timeout = 0s
[07-Aug-2026 20:20:46] NOTICE:  request_slowlog_trace_depth = 20
[07-Aug-2026 20:20:46] NOTICE:  request_terminate_timeout = 0s
[07-Aug-2026 20:20:46] NOTICE:  request_terminate_timeout_track_finished = no
[07-Aug-2026 20:20:46] NOTICE:  rlimit_files = 0
[07-Aug-2026 20:20:46] NOTICE:  rlimit_core = 0
[07-Aug-2026 20:20:46] NOTICE:  chroot = undefined
[07-Aug-2026 20:20:46] NOTICE:  chdir = undefined
[07-Aug-2026 20:20:46] NOTICE:  catch_workers_output = no
[07-Aug-2026 20:20:46] NOTICE:  decorate_workers_output = yes
[07-Aug-2026 20:20:46] NOTICE:  clear_env = yes
[07-Aug-2026 20:20:46] NOTICE:  security.limit_extensions = .php .phar
[07-Aug-2026 20:20:46] NOTICE:
[07-Aug-2026 20:20:46] NOTICE: configuration file /usr/local/etc/php-fpm.conf test is successful

Idle Processes status

PHP Version

PHP 8.4.24 (cli) (built: Aug  1 2026 14:53:20) (NTS)
Copyright (c) The PHP Group
Zend Engine v4.4.24, Copyright (c) Zend Technologies
    with Zend OPcache v8.4.24, Copyright (c), by Zend Technologies

Operating System

FreeBSD 14.4-RELEASE-p8

[qcybb]

Additional observation:

After lowering pm.max_requests from 5000 to 2500 and restarting PHP‑FPM, workers are still recycling only after serving roughly double the configured limit.

Example:

• With pm.max_requests = 2500, two workers reached 4996 and 4998 requests.
• Both transitioned from idle -> active and immediately exited.
• New workers spawned in their place.

This matches earlier behavior when the limit was set to 5000, where workers recycled around 10,000 requests.

This suggests that recycling is not triggered at pm.max_requests, but instead at approximately 2 x pm.max_requests.

This doubling behavior is consistent across restarts, reboots, and different threshold values.

[s2x]

This looks like the same root cause as #14212 (active-process counting), where nginx fastcgi_keep_conn on; / keepalive connections desync FPM's per-request state tracking — see PR #19191 for the fix there. @qcybb could you check if you have keepalive enabled toward FPM in nginx? If disabling it fixes the ~2x overshoot, it's the same bug class.

[qcybb]

I can confirm that removing fastcgi_keep_conn on; from my nginx.conf has solved this issue.