Hi,
On 2023-01-10 22:11:16 -0800, Nathan Bossart wrote:
> The attached patch fixes this by always calling WaitLatch(), even if
> wal_retrieve_retry_interval milliseconds have already elapsed and the
> timeout is 0.
It doesn't seem right to call WaitLatch() just for that purpose - nor
necessarily a complete fix.
Given that we check for interrupts in other parts of recovery with
HandleStartupProcInterrupt(), which doesn't interact with latches, isn't the
actual bug that HandleStartupProcInterrupt() doesn't contain the same black
magic that CHECK_FOR_INTERRUPTS() contains on windows? Namely this stuff:
#ifndef WIN32
...
#else
#define INTERRUPTS_PENDING_CONDITION() \
(unlikely(UNBLOCKED_SIGNAL_QUEUE()) ? pgwin32_dispatch_queued_signals()
: 0, \
unlikely(InterruptPending))
#endif
/* Service interrupt, if one is pending and it's safe to service it now */
#define CHECK_FOR_INTERRUPTS() \
do { \
if (INTERRUPTS_PENDING_CONDITION()) \
ProcessInterrupts(); \
} while(0)
Looks like we have that bug in quite a few places... Some are "protected" by
unconditional WaitLatch() calls, but at least pgarch.c, checkpointer.c via
CheckpointWriteDelay() seem borked.
Greetings,
Andres Freund
