Re: scheduling with io_lock held in 2.4.6

["Peter T. Breuer" <ptb@it.uc3m.es>]

"A month of sundays ago Andrew Morton wrote:"
> "Peter T. Breuer" wrote:
> > I've been plagued for a month by smp lockups in my block driver
> > that I eventually deduced were due to somebody else scheduling while
> > holding the io_request_lock spinlock.

> >   Aug 17 01:41:00 xilofon kernel: Scheduling with io lock held in process 0
> >   Aug 17 01:41:01 xilofon last message repeated 87 times
> >   Aug 17 01:41:01 xilofon kernel: Scheduling with io lock held in process 1141

> Replace the printk with a BUG(), feed the result into ksymooops.
> Or use show_trace(0).
> 
> But if you're running SMP, scheduling with a lock held
> is quite legal - it'll be held by another CPU.  In that case

Err, yes, I had initially made that mistake, but was fortunately running
on a single cpu machine. I fixed the test to check that the spinlock
was taken on the same cpu as we are now scheduling on and the test still
triggers.

> you'll need to record which CPU holds the lock.

My initial conclusion, based on recording file and line numbers every
time the spinlock is taken, is that end_that_request_last from ll_rw_blk.c
sometimes schedules under the io_request_lock.

I am still investigating, in the hope of pinning it down more exactly. If
anyone recognizes what goes on, please tell me.

Peter