On Fri, 27 Jul 2001, Greg Ames wrote:
> whoops...there's a bug with the patch I posted earlier. If multiple
> threads from a dying process are allowed to set "quiescing", then under
> worst case conditions, the same number of new processes are allowed to
> use the same process slot in the scoreboard. The fix is that only the
> first thread that exits is allowed to set "quiescing".
>
> Please use this patch instead.
It fixes a problem, so +1 on concept. For 2.0.22 I'd much rather see
us do something small and digestible like this than a really big
change-the-methodology patch. But there's still a problem.
I was pounding on it by doing ab -n 20000 -c 50 ... and doing a graceful
restart mid-bench. At the same time, I had a browser window open doing a
/server-status?refresh=2 . Because the browser was keepaliving my
server-status requests, it turns out that the whole server came to a
screeching halt when I did the graceful. The new generation didn't take
its workers (unsure if there were any yet) to 'S' state and then start
accepting new requests until after I killed my keptalive connection to the
old generation. That's bad. It doesn't do that without the patch. Fix
that problem, and I'm +1 for committing the patch to 2.0.22.
Thanks,
Cliff
--------------------------------------------------------------
Cliff Woolley
[EMAIL PROTECTED]
Charlottesville, VA
