[iliaal]

The unserializer now checks each object's closing brace before queuing __unserialize() or __wakeup(), so incomplete objects skip their callback and destructor. Related to #21716, which changes when already-queued callbacks are drained.

[ArtUkrainskiy]

On the PR head (ef25b3a) ASan reports a heap-use-after-free here, while the base (fd3a97d) and current master are fine. The outer object has no closing }, so it goes through the new early zval_ptr_dtor(&ary) after finish_nested_data() fails:

<?php
class T { public function __unserialize(array $data): void {} }
class D { public function __destruct() { var_dump(unserialize('R:2;')); } }
// the closing } of the outer object is missing
var_dump(@unserialize('O:1:"T":2:{i:0;s:5:"hello";i:1;O:1:"D":0:{}'));

Run with USE_ZEND_ALLOC=0 on an ASan build. It could go in as a test, too.

[iliaal]

Ah, indeed. A failed record's data array now stays in the unserialize context until var_destroy(), like a queued one, so the destructor can no longer reach an element it already released. The same UAF was reachable before this PR when an element after the nested object is malformed; both cases are now in __unserialize_malformed_nested_destructor.phpt.

[bukka]

This looks correct and fixes the issues.

[bukka]

I merged it as it conflicts with my other work so want to rebase it on this...