FreeBSD 6.3 deadlock (vm_map?) with DDB output

[John Baldwin]

On Sunday 15 June 2008 07:23:19 am Stef Walter wrote:
> I've been trying to track down a deadlock on some newish production
> servers running FreeBSD 6.3-RELEASE-p2. The deadlock occurs on a
> specific (although mundane) hardware configuration, and each of several
> servers running this hardware deadlock about once per week.
> 
> Although I suspect that this is not hardware related, from a (naive)
> perusal of the attached stack traces.
> 
> Forgive me if my interpretation of this is all wrong, but I'm pretty
> desperate for help. So here's my basic understanding of the deadlock:
> 
> These processes seem to be waiting on the page queue mutex:
>  sendmail (in vm_mmap > vm_map_find > vm_map_insert > vm_map_pmap_enter)
>  bsnmpd (in malloc, uma_large_malloc > page_alloc > kmem_malloc)
>  httpd (in trap > trap_pfault > vm_fault)
>  [g_up] (in g_vfs_done > bufdone)
> 
> The page queue mutex is held by rsync process:
>  rsync (in trap > trap_pfault > vm_fault > pmap_enter)
> 
> Rsync kernel process (in pmap_enter) was interrupted while holding the
> page queue lock?
> 
> 
> Giant is enabled in loader.conf due to the needs of the pf firewall when
> dealing with user credentials lookups. I do not believe that Giant plays
> into this deadlock. Kernel config attached.
> 
> Any and all help or info is welcome. Thanks in advance.

Try this change:

jhb         2007-10-27 22:07:40 UTC

  FreeBSD src repository

  Modified files:
    sys/kern             sched_4bsd.c
  Log:
  Change the roundrobin implementation in the 4BSD scheduler to trigger a
  userland preemption directly from hardclock() via sched_clock() when a
  thread uses up a full quantum instead of using a periodic timeout to cause
  a userland preemption every so often.  This fixes a potential deadlock
  when IPI_PREEMPTION isn't enabled where softclock blocks on a lock held
  by a thread pinned or bound to another CPU.  The current thread on that
  CPU will never be preempted while softclock is blocked.

  Note that ULE already drives its round-robin userland preemption from
  sched_clock() as well and always enables IPI_PREEMPT.

  MFC after:      1 week

  Revision  Changes    Path
  1.108     +8 -29     src/sys/kern/sched_4bsd.c

We use it at work on 6.x.  W/o this fix, round-robin stops working on 4BSD 
when softclock() (swi4: clock) blocks on a lock like Giant.

-- 
John Baldwin