Re: Sleeping function called from illegal context...

[John Levon]

On Sat, Sep 28, 2002 at 10:05:17AM -0700, Andrew Morton wrote:

> We need a standalone CONFIG_MIGHT_SLEEP.  I sinfully hooked it
> to CONFIG_DEBUG_KERNEL (it's not obvious why CONFIG_DEBUG_KERNEL
> exists actually).

Ah, OK.

> > I have a bit of a problem with __might_sleep because I call sleepable
> > stuff holding a spinlock (yes, it is justified, and yes, it is safe
> > afaics, at least with PREEMPT=n)
> 
> I'm looking at you suspiciously.  How come?

NMI interrupt handler cannot block so it trylocks against a spinlock
instead. Buffer processing code needs to block against concurrent NMI
interrupts so takes the spinlock for them. All actual blocks on the
spinlock are beneath a down() on another semaphore, so a sleep whilst
holding the spinlock won't actually cause deadlock.

I don't know a way out of this that can safely ensure we've finished
processing an NMI on the remote CPU the buffer processing code is about
to look at.

[I'll post a new patch against 2.5.39 in a bit so you can see in
context]

regards
john

-- 
"When your name is Winner, that's it. You don't need a nickname."
	- Loser Lane