Queue upcall locking (was: [dm-devel] [RFC][PATCH] fix dm_any_congested() to properly sync up with suspend code path)

[Mikulas Patocka]

On Mon, 10 Nov 2008, Peter Zijlstra wrote:

> On Mon, 2008-11-10 at 09:19 -0500, Mikulas Patocka wrote:
> > On Mon, 10 Nov 2008, Christoph Hellwig wrote:
> > 
> > > On Mon, Nov 10, 2008 at 08:11:51AM -0500, Mikulas Patocka wrote:
> > > > For upstream Linux developers: you are holding a spinlock and calling 
> > > > bdi*_congested functions that can take indefinite amount of time (there 
> > > > are even users reporting having 50 disks in one logical volume or so). I 
> > > > think it would be good to move these calls out of spinlocks.
> > > 
> > > Umm, they shouldn't block that long, as that completely defeats their
> > > purpose.  These functions are mostly used to avoid throwing more I/O at
> > > a congested device if pdflush could do more useful things instead.  But
> > > if it blocks in those functions anyway we wouldn't have to bother using
> > > them.  Do you have more details about the uses cases when this happens
> > > and where the routines spend so much time?
> > 
> > For device mapper, congested_fn asks every device in the tree and make OR 
> > of their bits --- so if the user has 50 devices, it asks them all.
> > 
> > For md-linear, md-raid0, md-raid1, md-raid10 and md-multipath it does the 
> > same --- asking every device.
> > 
> > If you have a better idea how to implement congested_fn, say it.
> 
> Fix the infrastructure by adding a function call so that you can have
> the individual devices report their congestion state to the aggregate.
> 
> Then congestion_fn can return a valid state in O(1) because the state is
> keps up-to-date by the individual state changes.
> 
> IOW, add a set_congested_fn() and clear_congested_fn().

If you have a physical disk that has many LVM volumes on it, you end up in 
a situation when disk congestion state change is reported to all the 
volumes. So it will create O(n) problem at the other side.

Mikulas