On Wed, Jul 4, 2018 at 2:30 PM, Andres Freund <andres@anarazel.de> wrote:
Hi,
On 2018-07-04 08:50:12 -0400, Jeff Janes wrote: > Reading through the comments touched by the commit, it seems obvious what > the bug is. It says "cause the processing to occur just before we next go > idle", but also says "This is called just *after* waiting for a frontend > command", which is too late to be "before we next go idle"
I've not looked at this issue in depth yet. So I might be completely off base. But I'm confused by your comment - we're doing it *after*, because we do a non-blocking read. And the latch will notify us (event.events & WL_LATCH_SET) if there was a read.
We get a signal while we are busy. We set the flag and the latch. The latch
wakes us up, but since we are busy (in a transaction, specifically) we don't
do anything. Later, the transaction ends and we are about to go idle, but
no one checks the flag again. We start a read using the latch mechanism, but
the flag notifyInterruptPending being set true from a now-long-gone signal is not on
the list of things the latch wakes us for. It is technically a non-blocking read but from
the perspective if the pending notify message it is a blocking read, unless another