On Sat, 23 Feb 2008 04:41:55 +1100, Eric Sandeen <sandeen@xxxxxxxxxxx>
wrote:
Barry Naujok wrote:
The attached patch fixes mkfs.xfs writing the bad features2 in the first
place (the change to xfs_sb.h does this).
Next xfs_db support printing of this superblock field and xfs_check can
report the bad_features2 field is set.
xfs_repair can correct the error in the same fashion that David
Chinner's
mount code does it.
This patch applies to the lazy-count xfs_repair conversion code that I
posted a short time before this patch.
Barry.
+ * Check bad_features2, if set and features2 is zero, copy
+ * bad_features2 to features2 and zero bad_features2.
+ */
+ if (sb->sb_bad_features2 != 0) {
+ if (sb->sb_features2 == 0)
+ sb->sb_features2 = sb->sb_bad_features2;
+ sb->sb_bad_features2 = 0;
+ primary_sb_modified = 1;
+ do_warn(_("superblock's features2 field is in the wrong "
+ "location, correcting\n"));
+ }
My only thought here is that if you repair it, then use an older kernel
w/o the fix, suddenly your fs behavior changes, whereas before you often
got lucky, and both userspace & kernelspace swapped the same way, and
you found the bits you were looking for out of luck :) (same goes for
the recent kernel fix too, I guess)
I believe the kernel code never tried to access "bad_features2" part
of the superblock, it always did the correct thing (correct me if I'm
wrong of course :).
But, as Dave pointed out, I should be OR'ing the two together before
zeroing the bad one.
Should we flag the bad_features2 as "already copied to features2" but
otherwise leave it? Hmm I guess that would look like an unsupported
feature flag to old kernels, wouldn't it. Urgh.
-Eric
|