[lkml]   [2011]   [Dec]   [6]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
Messages in this thread
SubjectRe: [PATCH 07/22] ext4: Create bitmap checksum helper functions
On Tue, Dec 06, 2011 at 10:19:12AM -0700, Andreas Dilger wrote:
> On 2011-12-05, at 9:33, Ted Ts'o <> wrote:
> > Note: it's strictly speaking not necessary to mix in the group and
> > s_csum_seed here. It's useful for the inode table blocks (ITB's)
> > because the checksum for a particular ITB is located *in* the ITB
> > itself. So if an ITB gets written to the wrong place, and in
> > particular, on top of another ITB, we want to be able to know which
> > cloned copy was written to the wrong place on disk.
> >
> > But in the case of the inode and block allocation bitmaps, the
> > checksums are stored in the block group descriptors; so if the bitmap
> > is written to the wrong place (and on top of another bitmap), the
> > checksum will fail to verify, independent of whether we've mixed in
> > the fs-specific csum seed and the group number.
> >
> > So I'd suggest dropping this, which will shave a few cycles off of the
> > checksum calculation, and it will also simplify the code since we
> > won't need this particular function.
> I wouldn't mind keeping the group just to be consistent with all of the other
> checksums that are used in the filesystem, which are largely inside the
> structure being checked.
> The s_uuid is definitely useful to keep as the seed because the block and
> inode bitmaps are not initialized at mke2fs time with uninit_bg, and it is
> possible to read a stale bitmap from disk that might belong to an earlier
> instance of the filesystem in case of a failed or misplaced write of the
> correct bitmap.

Hmm... let's say you have bitmap B before mkfs and bitmap B' after mkfs + some
file writes. B' is lost during write. It would be bad if B != B' and
crc32c(B) == crc32c(B'), in which case you'd use the wrong bitmap.

I suppose having the fsuuid + groupnum would probably help to make the inputs
to crc32c() more distinct, which would be useful since iirc P(collision)
decreases as the Hamming distance increases. I'm running a simulation to check
that claim.


> That isn't important for e2fsck, since it doesn't really use the bitmaps, but
> it is important for the kernel not to use bad bitmaps and corrupt the
> filesystem further.
> Cheers, Andreas--
> To unsubscribe from this list: send the line "unsubscribe linux-fsdevel" in
> the body of a message to
> More majordomo info at

 \ /
  Last update: 2011-12-06 22:03    [W:0.077 / U:3.024 seconds]
©2003-2018 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site