Fix #81206: Multiple PHP processes crash with JIT enabled

[cmb69]

The fix for bug 80175 (PR #6268) broke most SAPIs wrt. JIT, most
notably cgi-fcgi, because for these SAPIs SUCCESSFULLY_REATTACHED
actually has to set the reattached flag.

---

I'm not sure whether some other APIs might need to be handled as well:

php-src/ext/opcache/ZendAccelerator.c

Lines 2601 to 2612 in 65bd8d2

	static const char *supported_sapis[] = {
	"apache",
	"fastcgi",
	"cli-server",
	"cgi-fcgi",
	"fpm-fcgi",
	"fpmi-fcgi",
	"apache2handler",
	"litespeed",
	"uwsgi",
	NULL
	};

What is "uwsgi"?

[nikic]

What's the relevant difference between apache2handler and other SAPIs here? Can you add a code comment for why it has special handling?

[cmb69]

Actually, that may only be relevant for mpm_winnt, where the control process attaches to SHM, and afterwards the handler process reattaches, although it needs to do the JIT initialization. I wonder if the proper solution would be to suppress OPcache initialization (or at least JIT initialization) for the control process in the first place. I have no idea how to do that, and what would need to be done for Linux, especially for worker and event (although these are not recommended according to the PHP manual).

[cmb69]

Well, after more debugging I found that there is no relevant difference between the SAPIs here at all. The problem is just that we need to initialize the tracing JIT handlers for tracing even when reattaching on Windows. The new patch does this in the most simple way, namely by ignoring ZEND_VM_KIND_HYBRID which is not available on Windows anyway, and by doing the initialization on Windows only (I don't think this initialization is ever needed on other OSs).

Note that it was sufficient for the OP to just revert 2a545ba, because they had opcache.jit=1205 (= "function").

[nikic]

Ah yes, this makes sense. The stubs themselves are in SHM to which we reattach, but we need to reload the pointers to those stubs in per-process memory.

[dktapps]

I don't think this initialization is ever needed on other OSs

Why is this? The handlers are defaulted to NULL, and only inited currently in zend_jit_make_stubs() - I don't see any other place where they are initialized, nor any special behaviour for *nix - am I missing something here?

[cmb69]

@dktapps, I think that re-attaching is a Windowes only thing.

[dktapps]

I eventually managed to answer my own question - fork() is used everywhere except Windows, and SHM reattach is only done from forked subprocess on non-Windows. In a forked subprocess these handlers would already be initialized anyway because of the forking.