Hi,
I found that in the PG source code function BackendInitialize(), handler for
SIGTERM was set to be startup_die(). And in startup_die(), we simply call
proc_exit to exit the process.
What's more, early in BackendInitialize() function, we called pq_init to
setup socket_close() as a process exit callback.
The problem is: in socket_close, we have some none-reentrant calls like
free(). It may cause deadlock of this backend if we are excueting
malloc/free right before we step into the signal hander startup_die(). And I
experienced that on my local server :)
Similar to the problem in this thread:
https://www.postgresql-archive.org/PG-signal-handler-and-non-reentrant-malloc-free-calls-td3403162.html.
Is there a way to avoid this deadlock in startup_die()? Thanks..
--
Sent from: http://www.postgresql-archive.org/PostgreSQL-hackers-f1928748.html
