On Mon, Feb 22, 2021 at 10:34:42AM -0500, Brian Foster wrote:

Freed extents are marked busy from the point the freeing transaction
commits until the associated CIL context is checkpointed to the log.
This prevents reuse and overwrite of recently freed blocks before
the changes are committed to disk, which can lead to corruption
after a crash. The exception to this rule is that metadata
allocation is allowed to reuse busy extents because metadata changes
are also logged.

As of commit 97d3ac75e5e0 ("xfs: exact busy extent tracking"), XFS
has allowed modification or complete invalidation of outstanding
busy extents for metadata allocations. This implementation assumes
that use of the associated extent is imminent, which is not always
the case. For example, the trimmed extent might not satisfy the
minimum length of the allocation request, or the allocation
algorithm might be involved in a search for the optimal result based
on locality.

generic/019 reproduces a corruption caused by this scenario. First,
a metadata block (usually a bmbt or symlink block) is freed from an
inode. A subsequent bmbt split on an unrelated inode attempts a near
mode allocation request that invalidates the busy block during the
search, but does not ultimately allocate it. Due to the busy state
invalidation, the block is no longer considered busy to subsequent
allocation. A direct I/O write request immediately allocates the
block and writes to it.

I really hope there's a fstest case coming for this... :)

Finally, the filesystem crashes while in a
state where the initial metadata block free had not committed to the
on-disk log. After recovery, the original metadata block is in its
original location as expected, but has been corrupted by the
aforementioned dio.

Wheee!

Looking at xfs_alloc_ag_vextent_exact, I guess the allocator will go
find a freespace record, call xfs_extent_busy_trim (which could erase
the busy extent entry), decide that it's not interested after all, and
bail out without restoring the busy entry.

Similarly, xfs_alloc_cur_check calls _busy_trim (same side effects) as
we wander around the free space btrees looking for a good chunk of
space... and doesn't restore the busy record if it decides to consider a
different extent.

So I guess this "speculatively remove busy records and forget to restore
them" behavior opens the door to the write allocating blocks that aren't
yet free and nonbusy, right?  And the solution presented here is to
avoid letting go of the busy record for the bmbt allocation, and if the
btree split caller decides it really /must/ have that block for the bmbt
it can force the log and try again, just like we do for a file data
allocation?

Another solution could have been to restore the record if we decide not
to go ahead with the allocation, but as we haven't yet committed to
using the space, there's no sense in thrashing the busy records?

Assuming I got all that right,
Reviewed-by: Darrick J. Wong <djwong@kernel.org>

--D