I started fuzz-testing the realtime rmap feature with a very large
number of realtime allocation groups. There were so many rt groups that
repair had to rebuild /realtime in the metadata directory tree, and that
directory was big enough to spur the creation of a block format
directory.
Unfortunately, repair then walks both directory trees to look for
unconnceted files. This part of phase 6 emits CRC errors on the newly
created buffers for the /realtime directory, declares the directory to
be garbage, and moves all the rt rmap inodes to /lost+found, resulting
in a corrupt fs.
Poking around in gdb, I noticed that the buffer contents were indeed
zero, and that UPTODATE was not set. This was very strange, until I
added a watch on bp->b_flags to watch for accesses. It turns out that
xfs_repair's prefetch code will _get a buffer and zero the contents if
UPTODATE is not set.
The directory tree code in libxfs will also _get a buffer, initialize
it, and log it to the coordinating transaction, which in this case is
the transactions used to reconnect the rmap btree inodes to /realtime.
At no point does any of that code ever set UPTODATE on the buffer, which
is why prefetch zaps the contents.
Hence change both buffer dirtying functions to set UPTODATE, since a
dirty buffer is by definition at least as recent as whatever's on disk.
Signed-off-by: Darrick J. Wong <djwong@kernel.org>
*/
bp->b_error = 0;
bp->b_flags &= ~LIBXFS_B_STALE;
- bp->b_flags |= LIBXFS_B_DIRTY;
+ bp->b_flags |= LIBXFS_B_DIRTY | LIBXFS_B_UPTODATE;
}
/* Prepare a buffer to be sent to the MRU list. */
ASSERT(bp->b_transp == tp);
ASSERT(bip != NULL);
+ bp->b_flags |= LIBXFS_B_UPTODATE;
tp->t_flags |= XFS_TRANS_DIRTY;
set_bit(XFS_LI_DIRTY, &bip->bli_item.li_flags);
}