xfs
[Top] [All Lists]

Re: XFS corruption

To: Eric Sandeen <sandeen@xxxxxxxxxxx>
Subject: Re: XFS corruption
From: Dave Chinner <david@xxxxxxxxxxxxx>
Date: Mon, 22 Dec 2014 10:08:18 +1100
Cc: Alex Lyakas <alex@xxxxxxxxxxxxxxxxx>, xfs@xxxxxxxxxxx
Delivered-to: xfs@xxxxxxxxxxx
In-reply-to: <54970DD9.6080707@xxxxxxxxxxx>
References: <B97466D39E354AFB808620584A4ABE5C@alyakaslap> <54970DD9.6080707@xxxxxxxxxxx>
User-agent: Mutt/1.5.21 (2010-09-15)
On Sun, Dec 21, 2014 at 12:13:45PM -0600, Eric Sandeen wrote:
> On 12/21/14 5:42 AM, Alex Lyakas wrote:
> > Greetings,
> > we encountered XFS corruption:
> 
> > kernel: [774772.852316] ffff8801018c5000: 05 d1 fd 01 fd ff 2f ec 2f 8d 82 
> > 6a 81 fe c2 0f  .....././..j....    
> 
> There should have been 64 bytes of hexdump, not just the single line above, 
> no?

Yeah, really need the whole dmesg, because we've got readahead in
the picture here so the number of times the corruption error is seen
is actually important....

> 
> > [813114.622928] IP: [<ffffffffa077bad9>] xfs_bmbt_get_all+0x9/0x20 [xfs]
> > [813114.622928] PGD 0
> > [813114.622928] Oops: 0000 [#1] SMP
> > [813114.622928] CPU 2
> > [813114.622928] Pid: 31120, comm: smbd Tainted: GF       W  O 
> > 3.8.13-030813-generic #201305111843 Bochs Bochs
> > [813114.622928] RIP: 0010:[<ffffffffa077bad9>]  [<ffffffffa077bad9>] 
> > xfs_bmbt_get_all+0x9/0x20 [xfs]
> > [813114.622928] RSP: 0018:ffff88010a193798  EFLAGS: 00010297
> > [813114.622928] RAX: 0000000000000964 RBX: ffff880180fa9c38 RCX: 
> > ffffa5a5a5a5a5a5

RCX implies gotp->br_startblock was not overwritten by the
extent search. i.e. we've called xfs_bmap_search_multi_extents()
but no extent was actually found.

> > We analyzed several suspects, but all of them fall on disk addresses
> > not near the corrupted disk address. I realize that running somewhat
> > outdated kernel + our changes within XFSs, points back at us, but
> > this is first time we see XFS corruption after about a year of this
> > code being exercised. So posting here, just in case this is a known
> > issue.
> 
> well, xfs should _never_ oops, even if it encounters corruption.  So hopefully
> we can work backwards from the trace above to what went wrong here.
> 
> offhand, in xfs_bmap_search_multi_extents():
> 
>         ep = xfs_iext_bno_to_ext(ifp, bno, &lastx);
>         if (lastx > 0) {
>                 xfs_bmbt_get_all(xfs_iext_get_ext(ifp, lastx - 1), prevp);
>         }
>         if (lastx < (ifp->if_bytes / (uint)sizeof(xfs_bmbt_rec_t))) {
>                 xfs_bmbt_get_all(ep, gotp);
>                 *eofp = 0;
> 
> xfs_iext_bno_to_ext() can return NULL with lastx set to 0:
> 
>         nextents = ifp->if_bytes / (uint)sizeof(xfs_bmbt_rec_t);
>         if (nextents == 0) {
>                 *idxp = 0;
>                 return NULL;
>         }
> 
> (where idxp is the &lastx we sent in)

> and if we do that, it sure seems like the "if lastx < ...." test will wind up
> sending a null ep into xfs_bmbt_get_all, which would do a null ptr deref.

No, it shouldn't because lastx = 0 to get it set that way
ifp->if_bytes / (uint)sizeof(xfs_bmbt_rec_t) must be zero.
Therefore, this:

        if (lastx < (ifp->if_bytes / (uint)sizeof(xfs_bmbt_rec_t)))

evaulates as:

        if (0 < 0)

which is not true, so we fall into the else case:

        } else {
                if (lastx > 0) {
                        *gotp = *prevp;
                }
                *eofp = 1;
                ep = NULL;
        }
        *lastxp = lastx;
        return ep;

Which basically overwrites *eofp and *lastxp, neither of which are
NULL.

However, the stack trace clearly shows we've just called
xfs_bmap_search_multi_extents() - the "?" before the function name
means it found the symbol in the stack, but not in the direct line
of the frame pointers the current function stack points to.

That makes me doubt the accuracy of the stack trace, because the
only caller of xfs_bmap_search_multi_extents() is
xfs_bmap_search_extents() and xfs_bmap_search_extents does not call
xfs_bmbt_get_all() directly like the stack trace would lead us to
beleive. Hence I don't think we can trust the stack trace to be
pointing use at the correct caller of xfs_bmbt_get_all(), which
makes it real hard to isolate the cause...

Cheers,

Dave.
-- 
Dave Chinner
david@xxxxxxxxxxxxx

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