On 18.03.21 07:34, Craig Ringer wrote:
The main reason I didn't want to add more tracepoints than were strictly
necessary is that Pg doesn't enable the systemtap semaphores feature, so
right now we do a T_NAME(lock) evaluation each time we pass a tracepoint
if --enable-dtrace is compiled in, whether or not anything is tracing.
This was fine on pg11 where it was just:
#define T_NAME(lock) \
(LWLockTrancheArray[(lock)->tranche])
but since pg13 it instead expands to
GetLWTrancheName((lock)->tranche)
where GetLWTrancheName isn't especially trivial. We'll run that function
every single time we pass any of these tracepoints and then discard the
result, which is ... not ideal. That applies so long as Pg is compiled
with --enable-dtrace. I've been meaning to look at enabling the
systemtap semaphores feature in our build so these can be wrapped in
unlikely(TRACE_POSTGRESQL_LWLOCK_RELEASE_ENABLED()) guards, but I wanted
to wrap this patch set up first as there are some complexities around
enabling the semaphores feature.
There is already support for that. See the documentation at the end of
this page:
https://www.postgresql.org/docs/devel/dynamic-trace.html#DEFINING-TRACE-POINTS
