Re: Block device throttling [Re: Distributed storage.]

[Daniel Phillips <phillips@phunq.net>]

On Monday 13 August 2007 05:04, Evgeniy Polyakov wrote:
> On Mon, Aug 13, 2007 at 04:04:26AM -0700, Daniel Phillips 
(phillips@phunq.net) wrote:
> > On Monday 13 August 2007 01:14, Evgeniy Polyakov wrote:
> > > > Oops, and there is also:
> > > >
> > > > 3) The bio throttle, which is supposed to prevent deadlock, can
> > > > itself deadlock.  Let me see if I can remember how it goes.
> > > >
> > > >   * generic_make_request puts a bio in flight
> > > >   * the bio gets past the throttle and initiates network IO
> > > >   * net calls sk_alloc->alloc_pages->shrink_caches
> > > >   * shrink_caches submits a bio recursively to our block device
> > > >   * this bio blocks on the throttle
> > > >   * net may never get the memory it needs, and we are wedged
> > >
> > > If system is in such condition, it is already broken - throttle
> > > limit must be lowered (next time) not to allow such situation.
> >
> > Agreed that the system is broken, however lowering the throttle
> > limit gives no improvement in this case.
>
> How is it ever possible? The whole idea of throttling is to remove
> such situation, and now you say it can not be solved.

It was solved, by not throttling writeout that comes from shrink_caches.
Ugly.

> If limit is for 
> 1gb of pending block io, and system has for example 2gbs of ram (or
> any other resonable parameters), then there is no way we can deadlock
> in allocation, since it will not force page reclaim mechanism.

The problem is that sk_alloc (called from our block driver via 
socket->write) would recurse into shrink_pages, which recursively 
submits IO to our block driver and blocks on the throttle.  Subtle 
indeed, and yet another demonstration of why vm recursion is a Bad 
Thing.

I will find a traceback for you tomorrow, which makes this deadlock much 
clearer.

Regards