[aguinaldotupy]

Prepared with an AI assistant (Claude), which wrote the patch and ran the tests below; I asked for it.

Fixes #2572.

register_internal_extensions() freed the module list after the first php_module_startup(), so every later startup registered no custom extension. Their functions, classes and constants disappeared after a main thread reboot (file watcher, RestartWorkers() and the admin restart endpoint, opcache_reset()) and after Shutdown() and Init(), which Caddy runs on config reloads.

This picks up #2579 by @ousamabenyounes, rebased on main, and follows the review it got there.

Changes

• register_internal_extensions() keeps the C copy of the module list for the process lifetime, so each startup registers the modules again.
• register_extensions() returns early once the hook is installed, or when the list is empty. A second call can no longer make the hook its own predecessor.
• No Go or public API change.

This matches what PHP does for its own static modules: zend_startup() creates module_registry on each startup, and module_destructor() runs MSHUTDOWN, frees the ZTS globals id and resets module_started on each shutdown. Re-registering the same static zend_module_entry is therefore a fresh registration.

Tests

internal/testext now restarts PHP twice, first with RestartWorkers() and then with Shutdown() and Init(), and checks after each restart that both test modules are loaded. ext2 gets a function, a MINIT and module globals (GINIT), so the test covers the ZTS globals id being freed and allocated again. MINIT increments a process-wide counter and exposes it as EXT2_MINIT_CALLS, which must read 1, 2 and then 3. If RestartWorkers() returns without rebooting, the count stays at 1 and the test fails, so no API change is needed to observe the reboot.

RED on main (dev image, PHP 8.5 ZTS debug, go test -race ./internal/testext):

exttest.go:67:
    expected: "ext1: loaded\next2: loaded\next2_answer: 42\nEXT2_MINIT_CALLS: 2\n"
    actual  : "ext1: missing\next2: missing\next2_answer: missing\nEXT2_MINIT_CALLS: missing\n"
exttest.go:72:
    expected: "ext1: loaded\next2: loaded\next2_answer: 42\nEXT2_MINIT_CALLS: 3\n"
    actual  : "ext1: missing\next2: missing\next2_answer: missing\nEXT2_MINIT_CALLS: missing\n"
--- FAIL: TestRegisterExtension (0.08s)

GREEN with the fix:

--- PASS: TestRegisterExtension (0.08s)
ok  	github.com/dunglas/frankenphp/internal/testext	1.099s

Also run in the dev image: gotestsum -- -race ./... for the library (1192 tests, 8 skipped because the image has no pcntl) and for caddy/, the internal/testext test 20 times in fresh processes, golangci-lint (0 issues), go mod tidy -diff, and super-linter slim v9.0.0 with the CI settings (clang-format 22.1.3 included).

Not in this PR

• A late RegisterExtension() after the first Init() is still ignored silently. Reporting it would need an API change, so it belongs in a separate PR.
• A registration failure on reboot is no more observable than one at first boot: PHP prints Unable to start builtin modules and php_main() ignores the startup result. Every startup registers the same list into a new registry, so a reboot should not fail where the first boot succeeded. I can handle the startup result in a follow-up if you want it.

[ousamabenyounes]

Hi @aguinaldotupy,

Thanks for taking over #2579 and bringing the fix forward! Glad to see this work moving ahead.

Since your PR builds on our original contribution, would you be open to adding me as a co-author?

Thanks again for your work on this!

[aguinaldotupy]

Hi @ousamabenyounes, of course — the fix builds directly on your work in #2579. I've added you as co-author on the commit (Co-authored-by trailer). Thanks for the original patch and the investigation!