xfs
[Top] [All Lists]

Re: [PATCH 3/5] xfs_repair: fix dir refcount when '.' missing and dir is

To: Brian Foster <bfoster@xxxxxxxxxx>, Eric Sandeen <sandeen@xxxxxxxxxx>
Subject: Re: [PATCH 3/5] xfs_repair: fix dir refcount when '.' missing and dir is rebuilt
From: Eric Sandeen <sandeen@xxxxxxxxxxx>
Date: Mon, 08 Sep 2014 09:44:10 -0500
Cc: xfs@xxxxxxxxxxx
Delivered-to: xfs@xxxxxxxxxxx
In-reply-to: <20140908142529.GD52419@xxxxxxxxxxxxxxx>
References: <1410108065-18156-1-git-send-email-sandeen@xxxxxxxxxx> <1410108065-18156-4-git-send-email-sandeen@xxxxxxxxxx> <20140908134524.GC52419@xxxxxxxxxxxxxxx> <20140908142529.GD52419@xxxxxxxxxxxxxxx>
On 9/8/14 9:25 AM, Brian Foster wrote:
On Mon, Sep 08, 2014 at 09:45:25AM -0400, Brian Foster wrote:
On Sun, Sep 07, 2014 at 11:41:03AM -0500, Eric Sandeen wrote:
In phase 6's longform_dir2_entry_check, if we never
find a '.' entry we never add a reference to that entry;
if we subsequently rebuild it, '.' gets added, but
no ref to it is ever made.  This leads to Phase 7 doing
i.e.:

   Phase 7 - verify and correct link counts...
   resetting inode 5184 nlinks from 2 to 1

and the next run will do:

   Phase 7 - verify and correct link counts...
   resetting inode 5184 nlinks from 1 to 2

So if '.' was never found, but the directory got
rebuilt, manually add the ref for it.

Signed-off-by: Eric Sandeen <sandeen@xxxxxxxxxx>
---
  repair/phase6.c |    6 ++++++
  1 files changed, 6 insertions(+), 0 deletions(-)

diff --git a/repair/phase6.c b/repair/phase6.c
index f13069f..cc36a9c 100644
--- a/repair/phase6.c
+++ b/repair/phase6.c
@@ -2288,6 +2288,12 @@ out_fix:
                        if (bplist[i])
                                libxfs_putbuf(bplist[i]);
                longform_dir2_rebuild(mp, ino, ip, irec, ino_offset, hashtab);
+               /*
+                * If we didn't find a dot, we never added a ref for it;
+                * it's there now after the rebuild, so mark it as reached.
+                */
+               if (*need_dot)
+                       add_inode_ref(irec, ino_offset);

So if I follow this correctly, we iterate through the dir, add each name
to the hashtable and handle the inode reference count in the first
longform_dir2_entry_check() loop. If something is wrong, we call
longform_dir2_rebuild() to rebuild the dir from the hashtable of
names/inodes. We may or may not have added a reference for dot at that
point, and need_dot is set appropriately.

This seems Ok, but where is the dot entry actually added? Hmm, I see
that we handle dot in the longform_dir2_rebuild() loop by just skipping
over it...


It looks like this happens in process_dir_inode() after this whole
check/rebuild sequence, directory format permitting. There's also an
add_inode_ref() there. Perhaps the bug here is that we clear need_dot
when we shouldn't..?

If we do that, the first run says:

bad hash table for directory inode 5184 (no data entry): rebuilding
rebuilding directory inode 5184
creating missing "." entry in dir ino 5184

and then the 2nd run says:

multiple . entries in directory inode 5184: clearing entry

so, no.  ;)

The issue is that add_inode_ref() is keeping track (in repair)
of reached paths to the inode, in counted_nlinks.

If we didn't find '.' originally, we didn't add that ref.

When we do:


longform_dir2_rebuild
        xfs_dir_init()  // creates shortform
        <loop over names>
                xfs_dir_createname
                        xfs_dir2_sf_to_block when it's big enough
                                add '.' entry

and then we've added the '.' but haven't added the reference repair needs
internally.

-Eric

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