Re: 2.6.27-rc4: lots of 'in_atomic():1, irqs_disabled():0' with software-raid1

[Jens Axboe <jens.axboe@oracle.com>]

On Thu, Aug 28 2008, Neil Brown wrote:
> On Thursday August 28, jens.axboe@oracle.com wrote:
> > On Thu, Aug 28 2008, Andrew Morton wrote:
> > > 
> > > Cant sleep inside rcu_read_lock(), with CONFIG_PREEMPT_RCU=n, at least.
> > > 
> > > Dunno if it's legal if CONFIG_PREEMPT_RCU=y.  Hopefully not - that
> > > would be insane.  But I've failed to keep up with rcu goings-on
> > > recently.
> > 
> > Doh right, we of course can't block inside a RCU section. Then
> > bitmap.c:write_sb_page() wants fixing:
> > 
> >         rcu_read_lock();
> >         rdev_for_each_rcu(...)
> >                 md_super_write(...)
> >                         bio_alloc(GFP_NOIO, 1);
> > 
> > Neil?
> 
> 
> Yes......
> 
> And not only can't I call bio_alloc inside the rcu_read_lock, I also
> cannot call submit_bio, as that can do a mempool alloc for a request
> structure.

Right, it's not the only broken part there.

> I can get around that by putting the bios on the ->biolist that
> md_super_wait will resubmit requests from.  But I still need to
> allocate those bios.
> Maybe I can count how many there need to be, then allocate them and
> make a list, then pass them down into md_super_write.
> It's a bit ugly but it should work.
> 
> I think I'll have to think about it a bit more.

That was my initial thought as well, but the problem there is that you
cannot prealloc > 1 bio without matching it with a submit_bio() as well,
or you'd violate the bio_alloc() restriction on having the previous in
flight before allocating a new one.

-- 
Jens Axboe