[morrisonlevi]

If an extension re-arms EG(vm_interrupt) inside of zend_interrupt_function (as-in, unconditionally, every time), then on some VM kinds there will be infinite recursion. For the HYBRID VM this will crash with stack overflow and for TAILCALL this is just regular infinite recursion (on 8.6, no problem on 8.5).

The VM should always make progress. This is a useful testing property (does my interrupt handler work properly under ASAN for every possible vm interrupt point that was provided?) but it's also a nice property to protect production against a misbehaving extension.

I've discussed this with @arnaud-lb previously (see his draft #24472) but I wanted to fix it more generically, backport it to PHP 8.4, and add a test to prevent regressions. The core of the fix is this:

-		ZEND_VM_ENTER();
+		execute_data = EG(current_execute_data);
+		LOAD_OPLINE();
+		ZEND_VM_LEAVE();

[morrisonlevi]

Here's how ZEND_VM_ENTER and ZEND_VM_LEAVE expand across VM kinds (courtesy of claude) so you can see how the change from ZEND_VM_ENTER(); to execute_data = EG(current_execute_data); LOAD_OPLINE(); ZEND_VM_LEAVE(); changes things:

Kind	ZEND_VM_ENTER()	ZEND_VM_LEAVE()
HYBRID (8.4–8.6)	execute_data = EG(current_execute_data); LOAD_OPLINE(); then ZEND_VM_INTERRUPT_CHECK(); then ZEND_VM_CONTINUE()	ZEND_VM_CONTINUE()
CALL, 8.4	return 1	return 2
CALL, 8.5–8.6 (and 8.5 TAILCALL)	execute_data = EG(current_execute_data); LOAD_OPLINE(); return opline | ZEND_VM_ENTER_BIT	return opline | ZEND_VM_ENTER_BIT
TAILCALL, 8.6	execute_data = EG(current_execute_data); LOAD_OPLINE(); then ZEND_VM_INTERRUPT_CHECK(); then ZEND_VM_CONTINUE() (musttail call to the next handler)	ZEND_VM_CONTINUE()

I don't understand return 1 and return 2 on PHP 8.4's CALL VM but the rest seem correct to me.

[arnaud-lb]

Thank you!