fix(reactor): no PHP object after the object store is gone (#264)

[EdmondDantes]

Closes#264.

async_new_exception returns NULL while EG(active) is 0, and the four call
sites in libuv_reactor.c that released its result unconditionally take that
NULL: :2204, :2870, :3126, :6286. The other 23 in that file already
carry NULL onward — assigned to req->base.exception and passed to NOTIFY, or
released under a check — and async_rethrow_exception(NULL) is a no-op on both
of its branches.

The window this guards

zend_deactivate runs shutdown_executor() and only then
ZEND_ASYNC_ENGINE_SHUTDOWN(). By that point zend_shutdown_executor_values
has cleared EG(active) ("No PHP callback functions should be called after this
point") and zend_objects_store_destroy has freed the bucket table.
libuv_reactor_shutdown still turns the loop, so a write libuv cancels there
reaches io_pipe_writev_cb with status != 0 and the old code built an
exception object with no store to put it in.

Nothing else runs in that window: every uv callback in this extension is defined
in libuv_reactor.c (scheduler.c mentions libuv only in comments), and the
listeners a NOTIFY reaches consume an exception rather than build one. Callers
outside that file run from PHP-visible code, where EG(active) is 1.

The test

tests/cleanup/005-write_cancelled_at_reactor_shutdown.phpt needs the shape
built on purpose: a queued write on a stream handle comes only from
ZEND_ASYNC_IO_WRITEV_EX, whose one consumer is true-async/server, and
php-src's single caller of the write API — main/streams/plain_wrapper.c:553 —
writes files, whose queued work the drain waits out instead of cancelling. So a
build configured with --enable-async-fuzz queues one such write at reactor
shutdown when ASYNC_FUZZ_CANCELLED_WRITE=1 is set, and the test skips on every
other build. run-tests hands a test's ENV section to its SKIPIF process as
well, so the hook declines to fire when TEST_PHP_EVALUATING_SKIPIF is set;
a putenv() in the SKIPIF cannot do it, because PHP restores the original
environment at request shutdown, before the reactor stops.

Without the guard the test prints its expected output and then segfaults, which
run-tests reports as Termsig=0; with it, 5 of 5 runs pass. The full suite on
this build: 1293 tests, 1192 passed, 98 skipped, 3 failed. All three failures
are include __DIR__ . '/../../../../ext/...' resolving outside php-src,
because this checkout is reached through a symlink; nothing in this change can
reach them.

Debug builds crash. Release builds take the fast-shutdown path
(shutdown_executor, #if ZEND_DEBUG → fast_shutdown = 0), where the bucket
table outlives the teardown and the exception was built into a store nothing
reads again.

Where the symptom came from

true-async/server's h2/060-h2-write-deadline-reclaims failed 4 of 30
standalone runs and 2 of 17 full -j4 sweeps, its own assertions passing every
time. Against a PHP carrying this change: 0 of 30, and the server's phpt suite
is 349 passed, 20 skipped, 0 failed.

A second commit, test-only

pdo_mysql/011 failed on this branch's first CI run, and it is not this change:
four MySQL tests took the default fixture name async_test in one database, so
under parallel workers the DROP TABLE IF EXISTS of whichever started second
removed the table the first had just created. It reproduces on the PHP built on
14 August — 10 failing runs out of 10 over those four tests at -j4 — and 0 out
of 10 once the default becomes async_test_<pid>. It rides here because CI
cannot go green without it; the suite reaches MySQL only intermittently, which
is why the collision surfaced on this branch rather than earlier.

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.

📢 Thoughts on this report? Let us know!