Re: [patch 1/5] mm: add nofail variants of kmalloc kcalloc and kzalloc

[Dave Chinner <david@fromorbit.com>]

On Wed, Aug 25, 2010 at 02:48:36PM +0200, Peter Zijlstra wrote:
> On Wed, 2010-08-25 at 07:57 -0400, Ted Ts'o wrote:
> > On Wed, Aug 25, 2010 at 01:35:32PM +0200, Peter Zijlstra wrote:
> > > On Wed, 2010-08-25 at 07:24 -0400, Ted Ts'o wrote:
> > > > Part of the problem is that we have a few places in the kernel where
> > > > failure is really not an option --- or rather, if we're going to fail
> > > > while we're in the middle of doing a commit, our choices really are
> > > > (a) retry the loop in the jbd layer (which Andrew really doesn't
> > > > like), (b) keep our own private cache of free memory so we don't fail
> > > > and/or loop, (c) fail the file system and mark it read-only, or (d)
> > > > panic. 
> > > 
> > > d) do the allocation before you're committed to going fwd and can still
> > > fail and back out.
> > 
> > Sure in some cases that can be done, but the commit has to happen at
> > some point, or we run out of journal space, at which point we're back
> > to (c) or (d).
> 
> Well (b) sounds a lot saner than either of those. Simply revert to a
> state that is sub-optimal but has bounded memory use and reserve that
> memory up-front. That way you can always get out of a tight memory spot.
> 
> Its what the block layer has always done to avoid the memory deadlock
> situation, it has a private stash of BIOs that is big enough to always
> service some IO, and as long as IO is happening stuff keeps moving fwd
> and we don't deadlock.
> 
> Filesystems might have a slightly harder time creating such a bounded
> state because there might be more involved like journals and the like,
> but still it should be possible to create something like that (my swap
> over nfs patches created such a state for the network rx side of
> things).

Filesystems are way more complex than the block layer - the block
layer simply doesn't have to handle situations were thread X is
holding A, B and C, while thread Y needs C to complete the
transaction. thread Y is the user of the low memory pool, but has
almost depleted it and so even if we swith to thread X, the pool doe
snot have enouhg memory for X to complete and allow us to switch
back to Y and have it complete, freeing the memory from the pool
that it holds.

That is, the guarantee that we will always make progress simply does
not exist in filesystems, so a mempool-like concept seems to me to
be doomed from the start....

Cheers,

Dave.
-- 
Dave Chinner
david@fromorbit.com