Re: [patch] zfs livelock and thread priorities

From: John Baldwin <jhb_at_freebsd.org>
Date: Mon, 18 May 2009 13:31:10 -0400
On Monday 18 May 2009 1:12:59 pm Attilio Rao wrote:
> 2009/5/18 John Baldwin <jhb_at_freebsd.org>:
> > On Saturday 16 May 2009 12:40:44 pm Ben Kelly wrote:
> >>    1) It changes the kproc(9) API by adding a kproc_create_priority()
> >> function that allows you to set the priority of the newly created
> >> thread.  I'm not sure how people feel about this.
> >
> > Actually, I almost think we should just add a priority argument to each of 
the
> > routines that creates a new kthread/kproc.  Perhaps allow a priority of 0 
to
> > let the thread run with the default priority.  Hmm, it looks like kthreads
> > default to running with whatever thread0 runs at (PVM) which is probably 
not
> > really ideal.  Having an explicit priority for every kthread would 
probably
> > be best.  Most kthreads should probably be at PZERO by default I think.
> 
> I'm not sure I agree here.
> 1) Maybe I missed it (so please point me to the right one) but I
> didn't see a deep analysis of what messed up with the priorities there

Solaris makes certain assumptions about the relative priorities of ZFS threads 
and our ZFS doesn't set the priorities the same.  I think specifically there 
are "cleaner" threads which have a higher priority on Solaris than other ZFS 
threads and Solaris depends on that to avoid deadlocks.

> 2) I think this KPI can be dangerous and lead to problems. Priority is
> something highly fragile.

All the more reason to make developers _think_ about the priority of each 
kthread they create.  Right now all these threads start out with a priority 
of PVM since that is what thread0 runs at.  Does that sound right to you?  Do 
you think many folks realize that?  It sounds very bogus to me.  I think 
forcing people to pick a sensible priority for each thread is far better than 
the complete lack of thought that often happens now.

-- 
John Baldwin
Received on Mon May 18 2009 - 15:31:19 UTC

This archive was generated by hypermail 2.4.0 : Wed May 19 2021 - 11:39:47 UTC