Attached is revision of this patch that now treats the latch in
PGPROC, waitLatch, as the generic "process latch", rather than just
using it for sync rep; It is initialised appropriately as a shared
latch generically, within InitProcGlobal(), and ownership is
subsequently set within InitProcess(). We were doing so before, though
only for the benefit of sync rep.
The idea here is to have a per-process latch to guard against time-out
invalidation issues from within each generic signal handler, by
calling SetLatch() on our generic latch there, much as we already do
from within non-generic archiver process signal handlers on the
archiver's static, non-generic latch (the archiver has its MyProc
pointer set to NULL, and we allow for the possibility that generic
handlers may be registered within processes that have a NULL MyProc -
though latch timeout invalidation issues then become the
responsibility of the process exclusively, just as is currently the
case with the archiver. In other words, they better not register a
generic signal handler).
It doesn't really matter that the SetLatch() call will usually be
unnecessary, because, as Heikki once pointed out, redundantly setting
a latch that is already set is very cheap. We don't check if it's set
directly in advance of setting the latch, because the Latch struct is
"logically opaque" and there is no "public function" to check if it's
set, nor should there be; the checking simply happens in SetLatch().
On 18 July 2011 20:06, Heikki Linnakangas
<heikki.linnakangas@enterprisedb.com> wrote:
> Right, we can easily change the timeout argument to be in milliseconds
> instead of microseconds.
I've done so in this latest revision as a precautionary measure. I
don't see much point in sub-millisecond granularity, and besides, the
Windows implementation will not provide that granularity anyway as
things stand.
Thoughts?
--
Peter Geoghegan http://www.2ndQuadrant.com/
PostgreSQL Development, 24x7 Support, Training and Services
