On Thu, Oct 27, 2022 at 02:32:42PM -0700, Darrick J. Wong wrote:
> On Fri, Oct 28, 2022 at 07:49:57AM +1100, Dave Chinner wrote:
> > On Thu, Oct 27, 2022 at 10:14:14AM -0700, Darrick J. Wong wrote:
> > > From: Darrick J. Wong <djwong@xxxxxxxxxx>
> > >
> > > If we're in the middle of a deferred refcount operation and decide to
> > > roll the transaction to avoid overflowing the transaction space, we need
> > > to check the new agbno/aglen parameters that we're about to record in
> > > the new intent. Specifically, we need to check that the new extent is
> > > completely within the filesystem, and that continuation does not put us
> > > into a different AG.
> > >
> > > If the keys of a node block are wrong, the lookup to resume an
> > > xfs_refcount_adjust_extents operation can put us into the wrong record
> > > block. If this happens, we might not find that we run out of aglen at
> > > an exact record boundary, which will cause the loop control to do the
> > > wrong thing.
> > >
> > > The previous patch should take care of that problem, but let's add this
> > > extra sanity check to stop corruption problems sooner than later.
> > >
> > > Signed-off-by: Darrick J. Wong <djwong@xxxxxxxxxx>
> > > ---
> > > fs/xfs/libxfs/xfs_refcount.c | 48 ++++++++++++++++++++++++++++++++++++++++--
> > > 1 file changed, 46 insertions(+), 2 deletions(-)
> > >
> > >
> > > diff --git a/fs/xfs/libxfs/xfs_refcount.c b/fs/xfs/libxfs/xfs_refcount.c
> > > index 831353ba96dc..c6aa832a8713 100644
> > > --- a/fs/xfs/libxfs/xfs_refcount.c
> > > +++ b/fs/xfs/libxfs/xfs_refcount.c
> > > @@ -1138,6 +1138,44 @@ xfs_refcount_finish_one_cleanup(
> > > xfs_trans_brelse(tp, agbp);
> > > }
> > >
> > > +/*
> > > + * Set up a continuation a deferred refcount operation by updating the intent.
> > > + * Checks to make sure we're not going to run off the end of the AG.
> > > + */
> > > +static inline int
> > > +xfs_refcount_continue_op(
> > > + struct xfs_btree_cur *cur,
> > > + xfs_fsblock_t startblock,
> > > + xfs_agblock_t new_agbno,
> > > + xfs_extlen_t new_len,
> > > + xfs_fsblock_t *fsbp)
> > > +{
> > > + struct xfs_mount *mp = cur->bc_mp;
> > > + struct xfs_perag *pag = cur->bc_ag.pag;
> > > + xfs_fsblock_t new_fsbno;
> > > + xfs_agnumber_t old_agno;
> > > +
> > > + old_agno = XFS_FSB_TO_AGNO(mp, startblock);
> > > + new_fsbno = XFS_AGB_TO_FSB(mp, pag->pag_agno, new_agbno);
> > > +
> > > + /*
> > > + * If we don't have any work left to do, then there's no need
> > > + * to perform the validation of the new parameters.
> > > + */
> > > + if (!new_len)
> > > + goto done;
> >
> > Shouldn't we be validating new_fsbno rather than just returning
> > whatever we calculated here?
>
> No. Imagine that the deferred work is performed against the last 30
> blocks of the last AG in the filesystem. Let's say that the last AG is
> AG 3 and the AG has 100 blocks. fsblock 3:99 is the last fsblock in the
> filesystem.
>
> Before we start the deferred work, startblock == 3:70 and
> blockcount == 30. We adjust the refcount of those 30 blocks, so we're
> done now. The adjust function passes out new_agbno == 70 + 30 and
> new_len == 30 - 30.
>
> The agbno to fsbno conversion sets new_fsbno to 3:100 and new_len is 0.
> However, fsblock 3/100 is one block past the end of both AG 3 and the
> filesystem, so the check below will fail:
Sure, but my point here is that the function returns this invalid
fsbno in *fsbp and assumes that the caller will handle it correctly.
If the caller knows that we aren't going to continue past the
"new_len == 0" condition, then why is it even calling this function?
i.e. this isn't a "decide if we are going to continue" function,
it's a "calculate and validate next fsbno" function...
i.e. the intent doesn't match the name of the function.
> > > + if (XFS_IS_CORRUPT(mp, !xfs_verify_fsbext(mp, new_fsbno, new_len)))
> > > + return -EFSCORRUPTED;
> > > +
> > > + if (XFS_IS_CORRUPT(mp, old_agno != XFS_FSB_TO_AGNO(mp, new_fsbno)))
> > > + return -EFSCORRUPTED;
> >
> > We already know what agno new_fsbno sits in - we calculated it
> > directly from pag->pag_agno above, so this can jsut check against
> > pag->pag_agno directly, right?
>
> We don't actually know what agno new_fsbno sits in because of the way
> that the agblock -> fsblock conversion works:
>
> #define XFS_AGB_TO_FSB(mp,agno,agbno) \
> (((xfs_fsblock_t)(agno) << (mp)->m_sb.sb_agblklog) | (agbno))
Sure, but FSBs are *sparse* and there is unused, unchecked address
space between the AGs that agbno overruns can fall into. And when we
look at XFS_FSB_TO_AGNO():
#define XFS_FSB_TO_AGNO(mp,fsbno) \
((xfs_agnumber_t)((fsbno) >> (mp)->m_sb.sb_agblklog))
we can see that it simply truncates away the agbno portion to get
back to the agno.
IOWs:
0 sb_agblocks
+--------------------------+------------+
(1 << sb_agblklog)
+------------+
invalid agbnos!
Hence the agbno needs to be checked agains sb_agblocks to capture AG
overruns, not converted to a FSB and back to an AGNO as this will
claim agbnos in the inaccessible address space region between AGs
are valid....
> Notice how we don't mask off the bits of agbno above sb_agblklog? If
> sb_agblklog is (say) 20 but agbno has bit 31 set, that bit 31 will bump
> the AG number by 2^11 AGs.
Yes, but that's only a side effect of the agbno having the high bit
set - it could have many other bits set and still be out of range.
i.e. coverting to fsb and back to agno doesn't actually capture all
cases of the next calculated agbno/fsbno could be invalid.
xfs_verify_fsbext() may capture this by chance because it checks
the entire agbno portion of the fsb (via XFS_FSB_TO_AGBNO) against
xfs_ag_block_count(agno), but it won't capture the overruns that
only bump the AGNO portion of the FSB.
Hence I really think we should be checking new_agbno for validity
here, not relying on side effects of coverting to/from FSBs and
verifying fsb extents to capture ag block count overruns in the
supplied agbno....
Cheers,
Dave.
--
Dave Chinner
david@xxxxxxxxxxxxx
