Re: svn commit: r746597 - /apr/apr/trunk/poll/unix/epoll.c

[Jeff Trawick <tr...@gmail.com>]

On Sat, Feb 21, 2009 at 5:22 PM, <tr...@apache.org> wrote:

> Author: trawick
> Date: Sat Feb 21 22:22:34 2009
> New Revision: 746597
>
> URL: http://svn.apache.org/viewvc?rev=746597&view=rev
> Log:
> simplifications and minor improvements to error reporting and
> pathlength


I think there may still be feedback issues in impl_pollset_poll() with rv,
*num, and *descriptors, possibly specific to  APR_POLLSET_WAKEABLE.  I don't
see a testcase for that; maybe I'll get a chance to play with that soon.

Meanwhile, testpoll is crashing on Leopard, and I didn't think I changed any
code that affects Leopard (yet) ;)

Re: Re: svn commit: r746597 - /apr/apr/trunk/poll/unix/epoll.c

[tr...@gmail.com]

On Feb 22, 2009 3:53am, Mladen Turk <mt...@apache.org> wrote:
> Jeff Trawick wrote:

> On Sat, Feb 21, 2009 at 5:22 PM, trawick@apache.org

> I think there may still be feedback issues in impl_pollset_poll() with  
> rv, *num, and *descriptors, possibly specific to APR_POLLSET_WAKEABLE. I  
> don't see a testcase for that; maybe I'll get a chance to play with that  
> soon.


> The only difference is that if only the wakeup socked was signaled

> the *num is set to 0, descriptors are untouched and rv is APR_EINTR.

> Test case would be cool. I have one, but it needs a

> little bit polishing.




> Meanwhile, testpoll is crashing on Leopard, and I didn't think I changed  
> any code that affects Leopard (yet) ;)


> It passes on Darwin. Does Leopard uses kqueue as well?

yes

As it turns out, I was lucky to see it crash a couple of times in a row, as  
it is reproduced infrequently.

These hacks work around most intermittent testpoll failures I see on on  
leopard (Darwin Kernel Version 9.6.0: Mon Nov 24 17:37:00 PST 2008 ...  
i386), where kevent returns unexpectedly:

Index: testpoll.c
===================================================================
--- testpoll.c (revision 746597)
+++ testpoll.c (working copy)
@@ -382,6 +382,11 @@

send_msg(s, sa, 0, tc);
rv = apr_pollset_poll(pollset, 0, &num, &descs);
+ if (rv != APR_SUCCESS) {
+ /* rare glitch seen on Leopard; wait just a little longer */
+ printf(" \n DELAY %d\n ", __LINE__);
+ rv = apr_pollset_poll(pollset, apr_time_from_sec(1), &num, &descs);
+ }
ABTS_INT_EQUAL(tc, APR_SUCCESS, rv);
ABTS_INT_EQUAL(tc, 1, num);
ABTS_PTR_NOTNULL(tc, descs);
@@ -444,6 +449,11 @@

send_msg(s, sa, LARGE_NUM_SOCKETS - 1, tc);
rv = apr_pollset_poll(pollset, 0, &num, &descs);
+ if (rv != APR_SUCCESS) {
+ /* rare glitch seen on Leopard; wait just a little longer */
+ printf(" \n DELAY %d\n ", __LINE__);
+ rv = apr_pollset_poll(pollset, apr_time_from_sec(1), &num, &descs);
+ }
ABTS_INT_EQUAL(tc, APR_SUCCESS, rv);
ABTS_INT_EQUAL(tc, 1, num);
ABTS_PTR_NOTNULL(tc, descs);
@@ -607,7 +617,13 @@
send_msg(s, sa, 0, tc);
pcb.tc = tc;
pcb.count = 0;
- rv = apr_pollcb_poll(pollcb, 0, trigger_pollcb_cb, &pcb);
+ rv = apr_pollcb_poll(pollcb, 0, trigger_pollcb_cb, &pcb);
+ if (APR_STATUS_IS_TIMEUP(rv)) {
+ /* rare glitch seen on Leopard; wait just a little longer */
+ printf(" \n DELAY %d\n ", __LINE__);
+ rv = apr_pollcb_poll(pollcb, apr_time_from_sec(1), trigger_pollcb_cb,
+ &pcb);
+ }
ABTS_INT_EQUAL(tc, 0, APR_STATUS_IS_TIMEUP(rv));
ABTS_INT_EQUAL(tc, 1, pcb.count);

Re: svn commit: r746597 - /apr/apr/trunk/poll/unix/epoll.c

[Mladen Turk <mt...@apache.org>]

Jeff Trawick wrote:
> On Sat, Feb 21, 2009 at 5:22 PM, <trawick@apache.org 
> 
> I think there may still be feedback issues in impl_pollset_poll() with 
> rv, *num, and *descriptors, possibly specific to  APR_POLLSET_WAKEABLE. 
>  I don't see a testcase for that; maybe I'll get a chance to play with 
> that soon.
> 

The only difference is that if only the wakeup socked was signaled
the *num is set to 0, descriptors are untouched and rv is APR_EINTR.
Test case would be cool. I have one, but it needs a
little bit polishing.

> Meanwhile, testpoll is crashing on Leopard, and I didn't think I changed 
> any code that affects Leopard (yet) ;)
> 

It passes on Darwin. Does Leopard uses kqueue as well?

Regards
-- 
^(TM)