Warner Losh wrote this message on Thu, May 25, 2006 at 22:06 -0600:
> > In the past, I've been against mandating that callouts/timeouts/generic
> > taskqueues should not be allowed to sleep. However, after looking over
> > the history of this problem as well as others, it seems that it's just
> > too easy for driver authors to make bad assumptions and wind up with a
> > priority inversion/deadlock like this. It would be relatively trivial
> > to mark these contexts as being non-sleepable and have the msleep code
> > enforce it, like is done with ithreads. What do you think? Anyways,
> > thanks for looking at this and fixing it.
>
> At the very least, we should mandate that timeouts are a non-sleepable
> event. Sleeping just doesn't work there. taskqueues, I'm less sure
> of, since short sleeps there work, but do degrade performance. I like
> this idea.
People worried about things like this should create their own thread
for their taskqueue.. It's quite easy (simple macro declaration), and
I did that for handling kq in kq...
--
John-Mark Gurney Voice: +1 415 225 5579
"All that I will do, has been done, All that I have, has not."