Re: [PATCH] VM: kswapd should not do blocking memory allocations

[Trond Myklebust <Trond.Myklebust@netapp.com>]

On Wed, 2010-08-18 at 15:34 -0400, Chris Mason wrote:
> On Wed, Aug 18, 2010 at 03:04:01PM -0400, Trond Myklebust wrote:
> > From: Trond Myklebust <Trond.Myklebust@netapp.com>
> > 
> > Allowing kswapd to do GFP_KERNEL memory allocations (or any blocking memory
> > allocations) is wrong and can cause deadlocks in try_to_release_page(), as
> > the filesystem believes it is safe to allocate new memory and block,
> > whereas kswapd is there specifically to clear a low-memory situation...
> > 
> > Set the gfp_mask to GFP_IOFS instead.
> 
> I always thought releasepage was supposed to do almost zero work.  It
> could release an instantly freeable page but it wasn't supposed to dive
> in and solve world hunger or anything.
> 
> I thought the VM would be using writepage for that.

writepage isn't sufficient for the NFS case: the page may be in the
'clean but unstable' state, in which case the NFS client needs to send a
COMMIT rpc call before the page can finally be released.

That is why we need the gfp_flag to tell us when it is safe to do this,
and when it is not.
The main case where it is safe and necessary for try_to_release_page()
to initiate a COMMIT call is in the invalidate_inode_pages2(). We might
want to do it in the kswapd case too, but in that case, we definitely
should tell the filesystem that it is unsafe to block.

Trond