Fix GH-9067: random extension is not thread safe

[cmb69]

For thread-safety, we need to initialize global variables in GINIT (or
RINIT), but not in MINIT.

[TimWolla]

Thank you for tackling the issue. As this part of the engine is not something I have interacted with before: Do you have a reference of when example the various "callbacks" in question are executed?

Is MINIT once and GINIT once per thread?

[cmb69]

Is MINIT once and GINIT once per thread?

Indeed. See https://www.phpinternalsbook.com/php7/extensions_design/php_lifecycle.html?highlight=lifecycle#globals-initialization-ginit for details.

[TimWolla]

Thanks, I have read through that and also through the explanation for globals.

I'm a little confused, because according to the link the GINIT runs before MINIT. I would have expected the reverse order (first MINIT then GINIT). But if I understand the issue and inner workings correctly:

• random_globals is a thread-local variable
• Because MINIT() only runs once (likely for the very first thread?), the state in random_globals will not be correctly set up for any additional thread?

[TimWolla]

Based on my understanding of the PHP engine this is the correct fix. Each thread should have its own copy of the state.

[cmb69]

random_globals is a thread-local variable

Basically, yes. (Actually, it is a parameter which is an alias to the thread locals; elsewhere they're access via RANDOM_G.)

• Because MINIT() only runs once (likely for the very first thread?), the state in random_globals will not be correctly set up for any additional thread?

Right. MINIT only runs for the very first thread, so the thread locals of other threads need to be initialized elsewhere (GINIT or RINIT).