Обсуждение: myProcLocks initialization
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
Вложения
Robert Haas <robertmhaas@gmail.com> writes: > 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. That's not really enough to excite me, and the prospect of problems in one session corrupting an unrelated later one is pretty scary from a debugging standpoint. How about at least an Assert that the lock is in a clean state? regards, tom lane
On Sun, Oct 30, 2011 at 11:13 PM, Tom Lane <tgl@sss.pgh.pa.us> wrote: > Robert Haas <robertmhaas@gmail.com> writes: >> 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. > > That's not really enough to excite me, and the prospect of problems in > one session corrupting an unrelated later one is pretty scary from a > debugging standpoint. How about at least an Assert that the lock is in > a clean state? I can go for that. -- Robert Haas EnterpriseDB: http://www.enterprisedb.com The Enterprise PostgreSQL Company
On Sun, Oct 30, 2011 at 11:26 PM, Robert Haas <robertmhaas@gmail.com> wrote: > On Sun, Oct 30, 2011 at 11:13 PM, Tom Lane <tgl@sss.pgh.pa.us> wrote: >> Robert Haas <robertmhaas@gmail.com> writes: >>> 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. >> >> That's not really enough to excite me, and the prospect of problems in >> one session corrupting an unrelated later one is pretty scary from a >> debugging standpoint. How about at least an Assert that the lock is in >> a clean state? > > I can go for that. Revised patch attached. I think it would be useful to assert this both at process startup time and at process shutdown, since it would really be much nicer to have the process that didn't clean up fail the assertion, rather than the new one that innocently inherited its slot; so the attached patch takes that approach. -- Robert Haas EnterpriseDB: http://www.enterprisedb.com The Enterprise PostgreSQL Company
Вложения
Robert Haas <robertmhaas@gmail.com> writes: > Revised patch attached. I think it would be useful to assert this > both at process startup time and at process shutdown, since it would > really be much nicer to have the process that didn't clean up fail the > assertion, rather than the new one that innocently inherited its slot; > so the attached patch takes that approach. +1 regards, tom lane
Simon Riggs <simon@2ndQuadrant.com> writes: > On Mon, Oct 31, 2011 at 7:54 PM, Robert Haas <robertmhaas@gmail.com> wrote: >> Revised patch attached. �I think it would be useful to assert this >> both at process startup time and at process shutdown, since it would >> really be much nicer to have the process that didn't clean up fail the >> assertion, rather than the new one that innocently inherited its slot; >> so the attached patch takes that approach. > Something stronger than an assertion at shutdown? Run-time test? There's currently no evidence to suggest this will ever fire at all, especially not in non-development builds, so an assert seems enough to me. regards, tom lane
On Mon, Oct 31, 2011 at 7:54 PM, Robert Haas <robertmhaas@gmail.com> wrote: > Revised patch attached. I think it would be useful to assert this > both at process startup time and at process shutdown, since it would > really be much nicer to have the process that didn't clean up fail the > assertion, rather than the new one that innocently inherited its slot; > so the attached patch takes that approach. Something stronger than an assertion at shutdown? Run-time test? -- Simon Riggs http://www.2ndQuadrant.com/ PostgreSQL Development, 24x7 Support, Training & Services