From bf0e1069cb6898f24a5607bdf1d0291f1f351be1 Mon Sep 17 00:00:00 2001 From: Thomas Munro Date: Fri, 13 Dec 2024 01:11:01 +1300 Subject: [PATCH v2 2/2] Adjust pending_pm_pmsignal processing priority. Commit 239b175 gave high priority to shutdown requests, child exit notifications, reload requests and PM signals received from backends. Undo that last case. While the others have good reasons for queue-jumping (supremacy of administrative commands, ordering requirement of reload-connect sequences, and desire to release child process resources ASAP), PM signals do not. Let's go back to processing PM signals only when a WL_LATCH_SET event is received, and thus only once per server loop. It didn't make much difference before, because we'd very likely run the inner loop that scan events[] just once, because only WL_LATCH_SET was reported in the common case. Now that we've adjusted WaitEventSetWait() to return WL_LATCH_SET and potentially WL_SOCKET_ACCEPT at the same time when the latch was already set, let's reconsider that. We don't want to run BackgroundWorkerStateChange() and similar PM-signal-processing code twice as often as before in high load scenarios. XXX experimental Reported-by: Nathan Bossart Discussion: https://postgr.es/m/CA%2BhUKGLOcxUa6m7UinPN1gZXFyr92L8btG_pGTHPiWY2YbRw2w%40mail.gmail.com --- src/backend/postmaster/postmaster.c | 16 +++++++++++++--- 1 file changed, 13 insertions(+), 3 deletions(-) diff --git a/src/backend/postmaster/postmaster.c b/src/backend/postmaster/postmaster.c index f0f9c66487c..11e7a938f9f 100644 --- a/src/backend/postmaster/postmaster.c +++ b/src/backend/postmaster/postmaster.c @@ -1657,7 +1657,8 @@ ServerLoop(void) * didn't see WL_LATCH_SET. This gives high priority to shutdown * and reload requests where the latch happens to appear later in * events[] or will be reported by a later call to - * WaitEventSetWait(). + * WaitEventSetWait(). We also want to release child resources as + * soon as as possible, before accepting sockets. */ if (pending_pm_shutdown_request) process_pm_shutdown_request(); @@ -1665,8 +1666,17 @@ ServerLoop(void) process_pm_reload_request(); if (pending_pm_child_exit) process_pm_child_exit(); - if (pending_pm_pmsignal) - process_pm_pmsignal(); + + /* + * When receiving high frequency PM signals, we only want to + * process them once per server loop, not once per event of any + * kind. They are neither very high priority, nor causally linked + * to socket events that might arrive out of order, so we defer + * processing until we see WL_LATCH_SET. + */ + if (events[i].events & WL_LATCH_SET) + if (pending_pm_pmsignal) + process_pm_pmsignal(); if (events[i].events & WL_SOCKET_ACCEPT) { -- 2.47.0