Date: Wed, 11 Mar 2009 15:03:02 -0700
From: Andrew Morton <akpm@linux-foundation.org>
To: David Howells <dhowells@redhat.com>
Cc: torvalds@linux-foundation.org, peterz@infradead.org, Enrik.Berkhan@ge.com,
       dhowells@redhat.com, uclinux-dev@uclinux.org,
       linux-kernel@vger.kernel.org, linux-mm@kvack.org,
       Johannes Weiner <hannes@cmpxchg.org>
Subject: Re: [PATCH] NOMMU: Pages allocated to a ramfs inode's pagecache may
 get wrongly discarded
Message-Id: <20090311150302.0ae76cf1.akpm@linux-foundation.org>
In-Reply-To: <20090311153034.9389.19938.stgit@warthog.procyon.org.uk>
References: <20090311153034.9389.19938.stgit@warthog.procyon.org.uk>
Mime-Version: 1.0
Content-Type: text/plain; charset=US-ASCII
Content-Transfer-Encoding: 7bit
Sender: linux-kernel-owner@vger.kernel.org
Content-Length: 1860
Lines: 47

On Wed, 11 Mar 2009 15:30:35 +0000
David Howells <dhowells@redhat.com> wrote:

> From: Enrik Berkhan <Enrik.Berkhan@ge.com>
> 
> The pages attached to a ramfs inode's pagecache by truncation from nothing - as
> done by SYSV SHM for example - may get discarded under memory pressure.

Something has gone wrong in core VM.

> The problem is that the pages are not marked dirty.  Anything that creates data
> in an MMU-based ramfs will cause the pages holding that data will cause the
> set_page_dirty() aop to be called.
> 
> For the NOMMU-based mmap, set_page_dirty() may be called by write(), but it
> won't be called by page-writing faults on writable mmaps, and it isn't called
> by ramfs_nommu_expand_for_mapping() when a file is being truncated from nothing
> to allocate a contiguous run.
> 
> The solution is to mark the pages dirty at the point of allocation by
> the truncation code.

Page reclaim shouldn't be even attempting to reclaim or write back
ramfs pagecache pages - reclaim can't possibly do anything with these
pages!

Arguably those pages shouldn't be on the LRU at all, but we haven't
done that yet.

Now, my problem is that I can't 100% be sure that we _ever_ implemented
this properly.  I _think_ we did, in which case we later broke it.  If
we've always been (stupidly) trying to pageout these pages then OK, I
guess your patch is a suitable 2.6.29 stopgap.

If, however, we broke it then we've probably broken other filesystems
and we should fix the regression instead.

Running bdi_cap_writeback_dirty() in may_write_to_queue() might be the
way to fix all this.

Peter touched it last :)

--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/