On 04/28/2015 09:19 AM, Jeff Trawick wrote:
On 04/28/2015 09:12 AM, Jim Jagielski wrote:
Yeah, that's it, at least from what I can see. If I add a sleep(1)
right after the send_msg, we pass.
To cover the potential regression concern, I'll backport the new test
to 1.5.1 and confirm that it fails for me on FreeBSD 9 and Lion (even
if intermittently).
I copied 1.5.2/test/testpoll.c to 1.5.1/test/ and now see the same issue
on FreeBSD 9, which passed the 1.5.1 tests just before. For Lion, 1.5.2
testpoll.c on top of 1.5.1 fails intermittently (15% of the time?).
FWIW, the kqueue changes in 1.5.2 are all of this form,
@@ -115,12 +115,20 @@
{
int flags;
- if ((flags = fcntl(pollset->p->kqueue_fd, F_GETFD)) == -1)
- return errno;
+ if ((flags = fcntl(pollset->p->kqueue_fd, F_GETFD)) == -1) {
+ rv = errno;
+ close(pollset->p->kqueue_fd);
+ pollset->p = NULL;
+ return rv;
+ }
On Apr 28, 2015, at 9:10 AM, Jim Jagielski <[email protected]> wrote:
Could it be that we are seeing a race condition... that the
data isn't available by the time we wakeup and poll?
On Apr 28, 2015, at 9:02 AM, Jeff Trawick <[email protected]> wrote:
On 04/28/2015 08:52 AM, Jim Jagielski wrote:
Under FreeBSD 10, I see no issue w/ testpoll.
So it's looking like an OSX issue and not kqueue per-se.
Fails on FreeBSD 9 (also kqueue)...
Note that this is a new test I added in 1.5.2-dev. The history of
that is something like this:
* I had an uncommitted change to poll on my OS X box for a long
time, to report APR_INTR (4) when both data and wakeup occurred,
along with configure bits to use poll instead of kqueue.
* I committed the change to poll recently but luckily Yann didn't
let me get away with it.
* As penance, I added a test to confirm that when both data and
wakeup occur the caller sees APR_SUCCESS instead of APR_EINTR.
But with N apr_pollset_poll() implementations they may not be
acting consistently. We should see if the kqueue implementation
can change behaviors intermittently???
On Apr 28, 2015, at 8:40 AM, Jeff Trawick <[email protected]> wrote:
On 04/28/2015 08:38 AM, Jim Jagielski wrote:
Hmmm... seeing a regression on OSX 10.10.3 w/ Xcode 6.3.1:
% ./testall -v testpoll
testpoll : /Line 816: expected <0>, but saw <4>
FAILED 1 of 23
Failed Tests Total Fail Failed %
===================================================
testpoll 23 1 4.35%
Investigating.
PS: OSX sets HAVE_KQUEUE 1
"Cool" :) I'll fire up Lion and FreeBSD...
On Apr 25, 2015, at 9:39 AM, Jeff Trawick <[email protected]>
wrote:
Tarballs/zipfiles are at http://apr.apache.org/dev/dist/
Shortcut to CHANGES:
http://apr.apache.org/dev/dist/CHANGES-APR-1.5.2
autoconf version: 2.69 (same as apr 1.5.1)
libtool version: 2.4.2 (same as apr 1.5.1)
+/-1
[ ] Release APR 1.5.2 as GA
I'll hold the vote open for 72 hours unless something out of
the ordinary occurs.
Thanks in advance for testing!
