Pierangelo Masarati skrev:
Rein Tollevik wrote:
Howard Chu skrev:
Howard Chu wrote:
There's not much difference between HEAD and RE24 really. And I've seen test050 fail on both, at random times. Still haven't tracked it down.
Hm. I just saw test050 lock up on me, a repeat of the discussion in ITS#5454. Seems like we need to go back to a trylock here.
This reopens the race condition window that ITS closed. If a trylock is required then we must find some other way of asserting that the socket is either not enabled while the lock is held, or that any messages received while it is locked is acted upon.
I have run test050 with this patch, and see the same failures as in http://www.openldap.org/lists/openldap-devel/200809/msg00051.html I thought the errors was due to some recent changes in the code, but it appears to be the populating of consumer2 added in revision 1.11 of test050 itself that triggers it. I.e, the actual bug can have been there for some time :-(
AFAIR, writing to consumer2 was exactly added to trigger an issue that test050 was hiding in its initial form.
Do you remember what issue? After applying the patch to ITS#5719 this test has switch from failing almost half of the times I run it to now managing several hundreds iterations before deadlocking (i.e not the same symptoms as I used to see). I have had two cases where it was hung, and they seem to be related to the pool pausing that takes place when the updates to the config is replicated to the second consumer.
The tests hangs in ldapadd populating consumer2, where at least one thread (on the second consumer) replicating the config is hanging in handle_pause(). Note, this is without the lock -> trylock change in syncrepl.c, as I'm convinced that would reintroduce the syncrepl hangs we have seen earlier. None of the threads seem to be waiting on this lock, but I haven't had time to investigate it enough to conclude anything yet.
Rein