set_timeout() fails to terminate execution when IO Latency is high

Description

I'm not sure if this is a bug or not and I haven't found the proper channel to ask (shall I raise a discussion in Internals?).

We faced a problem in our production cluster. Our setup is following:

SAPI: fpm
max_execution_time: 28
hard_timeout: 2

When Opcache is reloading, we're having excessive IO pressure, which is expected - PHP starts reading everything from disk, etc... However, scripts started executing for 5+ minutes - that wasn't expected.

Further investigation showed that zend_set_timeout_ex uses setitimer(ITIMER_PROF) by default, which means the time the process spends in IOWait doesn't seem to count.

So, basically, even though our gateway cannot wait for such long - it expects that application will finish earlier, PHP still performs the request. One more backside of such behaviour - the restart of Opcache takes longer - it waits until all the workers finish execution (or become killed by opcache when opcache.force_restart_timeout is reached).

This problem doesn't seem to appear when we use ITIMER_REAL. I can provide how we tested this.

I can see a couple of solutions here:

• Switch to ITIMER_REAL (it looks like there are some concerns about the Apache module)
• Choose one over another for some SAPIs when we are sure that there shouldn't be a conflict (e.g. FPM, CLI)
• Make this behaviour configurable (either at runtime or compile time)
• Something else?

Start at zend_set_timeout_ex and compare the reported behavior of ITIMER_PROF with ITIMER_REAL under FPM and high I/O latency. Review the Apache-module concern and the proposed alternatives; done requires a decided timer behavior or configuration approach that prevents requests from running for several minutes.