Recent VFS locking vs kqueue

[Suleiman Souhlal]

Hi

On Jun 16, 2005, at 5:04 AM, Xin LI wrote:

> Hi, Jeff and -current@,
>
> With a fresh -CURRENT kernel (today's source) I got the following
> messages from console and /var/log/messages on my laptop, with a  
> "cvs -R
> up -PdA 2>/dev/null" running on one console (which results in disk/fs
> activities) and "tail -F /var/log/messages" (which uses kqueue  
> events).
>
> The kernel is built with ULE scheduler, PREEMPTION+FULL_PREEMPTION
> enabled, and all debugging options, like INVARIANTS and WITESS turned
> on.
>
> BTW.  I have found something that can lead to panic: Negative nice
> count, which indicates that there are some races in our code, but it
> seems that the bug is not easy to exercise.  Will report more
> information/patches once I got some findings.
>
> Thanks in advance!
>
> Jun 16 16:44:09 spirit kernel: Acquiring lockmgr lock "ufs" with the
> following non-sleepable locks held:
> Jun 16 16:44:09 spirit kernel: exclusive sleep mutex kqueue r = 0
> (0xc1b60a00) locked @ /usr/src/sys/kern/kern_event.c:1524
> Jun 16 16:44:09 spirit kernel: exclusive sleep mutex vnode pollinfo  
> r =
> 0 (0xc2074000) locked @ /usr/src/sys/kern/kern_event.c:1512

This has been reported before by Jun Kuriyama.. Unfortunately, I'm a  
little busy this week, so I've not had time to fix it yet, but  
hopefully, I'll have something this weekend.
In the meantime, you can patch -R this: http://people.freebsd.org/ 
~ssouhlal/testing/kqueue-hooks-20050603.diff .

Bye.
--
Suleiman Souhlal     | ssouhlal at vt.edu
The FreeBSD Project  | ssouhlal at FreeBSD.org