https://bz.apache.org/bugzilla/show_bug.cgi?id=69592

--- Comment #2 from abernal <[email protected]> ---
Update: an independent review caught a real concurrency bug in the first
version of the fix, specific to LuaScope server. ap_lua_release_state() for
that scope hands the parent VM back to its apr_reslist pool, making it
available for another thread to acquire immediately; the original patch called
lua_pop(parent, 1) after that release, so a second thread could acquire the
same parent and push its own coroutine onto it before the first thread's pop
ran, causing the pop to remove the second thread's coroutine instead of its
own.

Reproduced this for real with temporary logging and an injected delay between
release and pop, 2 concurrent requests against LuaScope server 1 1 (forcing
both threads to share one VM): saw the exact predicted cross-removal every
time. Moving lua_pop() before ap_lua_release_state(), now in PR
https://github.com/apache/httpd/pull/1044, closes it: the same forced-delay
scenario against the corrected ordering shows clean, fully serialized
acquire/push/pop/release per request. The existing regression test and the rest
of the modules/lua suite (61 tests) still pass with the corrected ordering.

-- 
You are receiving this mail because:
You are the assignee for the bug.
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]

Reply via email to