PLEASE TEST: IPI deadlock avoidance patch
Willem Jan Withagen
wjw at withagen.nl
Thu Aug 26 10:55:00 PDT 2004
Doug White wrote:
>On Mon, 23 Aug 2004, Garance A Drosihn wrote:
>>At 12:05 PM -0700 8/22/04, Doug White wrote:
>>>Hey folks, [...] please try this patch:
>>>This patch avoids a deadlock between the smp_rendezvous()
>>>mechanism and TLB shootdowns via pmap by forcing them to
>>>share a mutex.
>>I have completed a series of buildworlds with -j3 to -j10 with
>>no problem. I then started up a "folding at home" client, and
>>repeated the buildworlds. No panics, but in the second set I
>>did have one buildworld (-j9) which failed with four processes
>>apparently getting a "*** Signal 6". I am not sure what that
>>was about. Note that all these builds were done with a `make'
>>that was compiled to USE_KQUEUE , so maybe that's where the
>>Signal's came from. In any case, the machine is still running
>>fine, even after that heavy pounding.
>Signal 6 is SIGABRT, which is usually intentional. You'd have to check the
>output for a specific process that abended. I'd also have to scan the
>make code for any abort() calls.
I have not given it much attention yet, since I'm bussy doing other
things right now.
But on my dual AMD64 box compiling the current 5.3 Beta generated also
'Abort Trap' while doing a
buildworld -j 32. Compiling without -j worked fine, and now I'm at 5.3
B1, so I can test again.
It ws running a 5.3B1 kernel.
I'll let you know if the problem persists
More information about the freebsd-current