I'd like to propose the attached patch, which initializes each
PGPROC's myProcLocks just once at postmaster startup, rather than
every time the PGPROC is handed out to a backend. These lists should
always be emptied before a backend shuts down, so a newly initialized
backend will find the lists empty anyway. Not reinitializing them
shaves a few cycles. In my testing, it saves about 1% of the cost of
setting up and tearing down a connection, which is not a ton, but a
cycle saved is a cycle earned.
Of course, we have a few outstanding reports, like this one from Dave
Gould, indicating that maybe we have a bug in there somewhere:
http://archives.postgresql.org/message-id/20110822073131.GC3363@sonic.net
...but in that case it seems to me that this doesn't make anything
worse than it already is. If the myProcLocks pointers are pointing to
random garbage, we're just kidding ourselves whatever we do; the
system is screwed, and we ought to PANIC, and anything we do here is
laughably inadequate. OTOH, if it just so happens that a backend
found a sneaky way through the exit path that doesn't involve calling
LockReleaseAll(), then overwriting the shm queue pointers removes our
last hope that the next backend can clean up the mess. I'm not
putting a lot of faith in that actually working, just saying that the
current code doesn't seem to be accomplishing anything in the
robustness department.
--
Robert Haas
EnterpriseDB: http://www.enterprisedb.com
The Enterprise PostgreSQL Company