[Top] [All Lists]

Re: [PATCH] XFS: Let the broken fiemap work in query mode.

To: Dave Chinner <david@xxxxxxxxxxxxx>
Subject: Re: [PATCH] XFS: Let the broken fiemap work in query mode.
From: Tao Ma <tao.ma@xxxxxxxxxx>
Date: Wed, 28 Apr 2010 10:00:01 +0800
Cc: linux-kernel@xxxxxxxxxxxxxxx, xfs@xxxxxxxxxxx, Eric Sandeen <sandeen@xxxxxxxxxx>, Christoph Hellwig <hch@xxxxxx>, Alex Elder <aelder@xxxxxxx>
In-reply-to: <20100428015035.GC9783@dastard>
References: <1272349065-16383-1-git-send-email-tao.ma@xxxxxxxxxx> <20100428015035.GC9783@dastard>
User-agent: Thunderbird (X11/20100411)
Hi Dave,

Dave Chinner wrote:
On Tue, Apr 27, 2010 at 02:17:45PM +0800, Tao Ma wrote:
According to Documentation/filesystems/fiemap.txt, If fm_extent_count
is zero, then the fm_extents[] array is ignored (no extents will be
returned), and the fm_mapped_extents count will hold the number of
extents needed.

But as the commit 97db39a1f6f69e906e98118392400de5217aa33a has changed
bmv_count to the caller's input buffer, this number query function can't
work any more. As this commit is written to change bmv_count from
MAXEXTNUM because of ENOMEM, we can't find a really suitable number to
set bmv_count now in xfs_vn_fiemap. Since we really have no idea of how
much extents the file has, a big number may cause ENOMEM, while a small
one will mask the real extent no.

So this patch try to resolve this problem by adding a temporary getbmapx
in xfs_getbmap. If the caller didn't give bmv_count, we don't allocate
the "out" either. Instead, every time we want to use 'out', use '&tmp'

I know this solution is a bit ugly, but I can't find a way to resolve
this issue while not changing the codes too much. So any good suggestion
is welcomed.

I don't see a need to change xfs_getbmap() to fix this. We can limit
the maximum allocation size to something realistic just by setting
bm.bmv.count to something sane. e.g, in xfs_vn_fiemap:

-       bm.bmv_count = fieinfo->fi_extents_max + 1;
+       bm.bmv.count = !fieinfo->fi_extents_max ? MAXEXTNUM :
+                                       fieinfo->fi_extents_max - 1;
+       bm.bmv_count = MIN(bm.bmv_count,
                                (PAGE_SIZE * 16 / sizeof(struct getbmapx)));

Unless I'm missing something, that should also prevent the case of
an application providing a really large fi_extents_max from
triggering ENOMEM in most cases as well.
I just worry about one thing: What if the real extent number is larger than the PAGE_SIZE * 16 / sizeof(struct getbmapx)? In this case, we will
give up the wrong extent number to the user space.

FWIW, how did you find this? Is it possible for you to add a test
for this regression into xfstests so that we don't break it again
in future?
Sure, I will check and see whether I can add it in xfstests.


<Prev in Thread] Current Thread [Next in Thread>