Check if restart is pending before trying to lock SHM

[mikhainin]

This reduces lock contention when Opcache restart is scheduled but not yet started.

[mikhainin]

Our application consists of 300K of PHP files (~3Gb of code) and the server has 220 FPM workers.

When we tried to enable file_cache, we got an extremely poor performance of the system, each request took over 30-60s when usually it finishes within 0.5s

perf top showed that the system spends most of the time inside the file lock:

  84.28%  [k] __pv_queued_spin_lock_slowpath                                                                                                                                                                         
   1.99%  [k] _raw_spin_unlock_irqrestore                                                                                                                                                                            
   0.92%  [k] finish_task_switch                                                                                                                                                                                     
   0.67%  [k] posix_locks_conflict                                                                                                                                                                                   
   0.24%  [k] posix_lock_inode                                                                                                                                                                                       
   0.21%  [k] copy_user_enhanced_fast_string                                                                                                                                                                         
   0.15%  [k] __raw_callee_save___pv_queued_spin_unlock                                                                                                                                                              
   0.14%  [k] _raw_spin_lock                                                                                                                                                                                         
   0.11%  [k] __do_page_fault                                                                                                                                                                                        
   0.09%  [k] clear_page_erms                                                                                                                                                                                        
   0.08%  [k] __d_lookup_rcu                                                                                                                                                                                         

PHP code wasn't even in the top 30.

The typical stack trace (we're using PHP8.0) for the FPM worker:

#0  0x00007f3778335f07 in fcntl64 () from /lib64/libc.so.6
#1  0x00007f37776d7922 in zend_shared_alloc_lock () at /build/php.git/ext/opcache/zend_shared_alloc.c:481
#2  0x00007f37776d5b7d in zend_file_cache_script_load (file_handle=<optimized out>) at /build/php.git/ext/opcache/zend_file_cache.c:1753
#3  0x00007fffffffe000 in ?? ()
#4  0x00007f3679ba83c8 in ?? ()
#5  0x0000000000000000 in ?? ()

zend_file_cache.c:1753 is zend_shared_alloc_lock(); under this condition.

After we tried this patch, performance was significantly improved (still worse than with shared memory but way better), and p99 of requests fit into 8 seconds.

perf top looks now healthier (when in-memory opcache is inactive):

  18.65%  [kernel]                                                        [k] __pv_queued_spin_lock_slowpath
   8.91%  php-fpm                                                         [.] execute_ex
   4.57%  php-fpm                                                         [.] zend_hash_find
   3.26%  php-fpm                                                         [.] _zend_hash_find_known_hash
   2.11%  [kernel]                                                        [k] _raw_spin_unlock_irqrestore
   1.79%  php-fpm                                                         [.] ZEND_DO_FCALL_SPEC_OBSERVER_HANDLER
   1.57%  opcache.so                                                      [.] zend_accel_load_script
   1.57%  php-fpm                                                         [.] _zend_observe_fcall_begin

I assume that the lock is being held by one of the FPM workers in accel_activate() or it's just at typical contention.

[mikhainin]

I can collect more data if you're interested or amend the patch if you feel like something is wrong

[dstogov]

This looks fine.
Could you please re-target the pull request to PHP-8.1 branch.

[mikhainin]

Done, it's now based on PHP-8.1