Invalidate the cached dentries that point to the file that we're moving
to lost+found before we actually move it.
Signed-off-by: Darrick J. Wong <djwong@kernel.org>
Reviewed-by: Christoph Hellwig <hch@lst.de>

The atomic file exchange-range functionality is now a permanent
filesystem feature instead of a dynamic log-incompat feature.  It cannot
be turned on at runtime, so we no longer need the XCHK_FSGATES flags and
whatnot that supported it.  Remove the flag and the enable function, and
move the xfs_has_exchange_range checks to the start of the repair
functions.
Signed-off-by: Darrick J. Wong <djwong@kernel.org>
Reviewed-by: Christoph Hellwig <hch@lst.de>

If the transaction allocation in xrep_adoption_trans_alloc fails, we
should drop only the locks that we took.  In this case this is
ILOCK_EXCL of both the orphanage and the file being repaired.  Dropping
any IOLOCK here is incorrect.

Found by fuzzing u3.sfdir3.list[1].name = zeroes in xfs/1546.
Signed-off-by: Darrick J. Wong <djwong@kernel.org>
Reviewed-by: Christoph Hellwig <hch@lst.de>

If the transaction allocation in the !orphanage_available case of
xrep_nlinks_repair_inode fails, we need to drop the IOLOCK of the file
being scrubbed before exiting.

Found by fuzzing u3.sfdir3.list[1].name = zeroes in xfs/1546.
Signed-off-by: Darrick J. Wong <djwong@kernel.org>
Reviewed-by: Christoph Hellwig <hch@lst.de>

If a program wants us to perform a scrub on a file handle and the fd
passed to ioctl() is not the file referenced in the handle, iget the
file once and pass it into the scrub code.  This amortizes the untrusted
iget lookup over /all/ the scrubbers mentioned in the scrubv call.

When running fstests in "rebuild all metadata after each test" mode, I
observed a 10% reduction in runtime on account of avoiding repeated
inobt lookups.
Signed-off-by: Darrick J. Wong <djwong@kernel.org>
Reviewed-by: Christoph Hellwig <hch@lst.de>

Back when I wrote commit a03297a0, I had thought that we'd be doing
users a favor by only marking inodes dontcache at the end of a scrub
operation, and only if there's only one reference to that inode.  This
was more or less true back when I_DONTCACHE was an XFS iflag and the
only thing it did was change the outcome of xfs_fs_drop_inode to 1.

Note: If there are dentries pointing to the inode when scrub finishes,
the inode will have positive i_count and stay around in cache until
dentry reclaim.

But now we have d_mark_dontcache, which cause the inode *and* the
dentries attached to it all to be marked I_DONTCACHE, which means that
we drop the dentries ASAP, which drops the inode ASAP.

This is bad if scrub found problems with the inode, because now they can
be scheduled for inactivation, which can cause inodegc to trip on it and
shut down the filesystem.

Even if the inode isn't bad, this is still suboptimal because phases 3-7
each initiate inode scans.  Dropping the inode immediately during phase
3 is silly because phase 5 will reload it and drop it immediately, etc.
It's fine to mark the inodes dontcache, but if there have been accesses
to the file that set up dentries, we should keep them.

I validated this by setting up ftrace to capture xfs_iget_recycle*
tracepoints and ran xfs/285 for 30 seconds.  With current djwong-wtf I
saw ~30,000 recycle events.  I then dropped the d_mark_dontcache calls
and set XFS_IGET_DONTCACHE, and the recycle events dropped to ~5,000 per
30 seconds.

Therefore, grab the inode with XFS_IGET_DONTCACHE, which only has the
effect of setting I_DONTCACHE for cache misses.  Remove the
d_mark_dontcache call that can happen in xchk_irele.

Fixes: a03297a0 ("xfs: manage inode DONTCACHE status at irele time")
Signed-off-by: Darrick J. Wong <djwong@kernel.org>
Reviewed-by: Christoph Hellwig <hch@lst.de>

Introduce a variant on XFS_SCRUB_METADATA that allows for a vectored
mode.  The caller specifies the principal metadata object that they want
to scrub (allocation group, inode, etc.) once, followed by an array of
scrub types they want called on that object.  The kernel runs the scrub
operations and writes the output flags and errno code to the
corresponding array element.

A new pseudo scrub type BARRIER is introduced to force the kernel to
return to userspace if any corruptions have been found when scrubbing
the previous scrub types in the array.  This enables userspace to
schedule, for example, the sequence:

 1. data fork
 2. barrier
 3. directory

If the data fork scrub is clean, then the kernel will perform the
directory scrub.  If not, the barrier in 2 will exit back to userspace.

The alternative would have been an interface where userspace passes a
pointer to an empty buffer, and the kernel formats that with
xfs_scrub_vecs that tell userspace what it scrubbed and what the outcome
was.  With that the kernel would have to communicate that the buffer
needed to have been at least X size, even though for our cases
XFS_SCRUB_TYPE_NR + 2 would always be enough.

Compared to that, this design keeps all the dependency policy and
ordering logic in userspace where it already resides instead of
duplicating it in the kernel. The downside of that is that it needs the
barrier logic.

When running fstests in "rebuild all metadata after each test" mode, I
observed a 10% reduction in runtime due to fewer transitions across the
system call boundary.
Signed-off-by: Darrick J. Wong <djwong@kernel.org>
Reviewed-by: Christoph Hellwig <hch@lst.de>

Move the scrub ioctl handler to scrub.c to keep the code together and to
reduce unnecessary code when CONFIG_XFS_ONLINE_SCRUB=n.
Signed-off-by: Darrick J. Wong <djwong@kernel.org>
Reviewed-by: Christoph Hellwig <hch@lst.de>

We really don't want to call cond_resched every single time we go
through a loop in scrub -- there may be billions of records, and probing
into the scheduler itself has overhead.  Reduce this overhead by only
calling cond_resched 10x per second; and add a counter so that we only
check jiffies once every 1000 records or so.

Surprisingly, this reduces scrub-only fstests runtime by about 2%.  I
used the bmapinflate xfs_db command to produce a billion-extent file and
this stupid gadget reduced the scrub runtime by about 4%.

From a stupid microbenchmark of calling these things 1 billion times, I
estimate that cond_resched costs about 5.5ns per call; jiffes costs
about 0.3ns per read; and fatal_signal_pending costs about 0.4ns per
call.
Signed-off-by: Darrick J. Wong <djwong@kernel.org>
Reviewed-by: Christoph Hellwig <hch@lst.de>

Repair corruptions in the directory tree itself.  Cycles are broken by
removing an incoming parent->child link.  Multiply-owned directories are
fixed by pruning the extra parent -> child links  Disconnected subtrees
are reconnected to the lost and found.
Signed-off-by: Darrick J. Wong <djwong@kernel.org>
Reviewed-by: Christoph Hellwig <hch@lst.de>

Report directories that are the source of corruption in the directory
tree.
Signed-off-by: Darrick J. Wong <djwong@kernel.org>
Reviewed-by: Christoph Hellwig <hch@lst.de>

Add a dirent update hook so that we can detect directory tree updates
that affect any of the paths found by this scrubber and force it to
rescan.
Signed-off-by: Darrick J. Wong <djwong@kernel.org>
Reviewed-by: Christoph Hellwig <hch@lst.de>

Create a new scrubber that detects corruptions within the directory tree
structure itself.  It can detect directories with multiple parents;
loops within the directory tree; and directory loops not accessible from
the root.
Signed-off-by: Darrick J. Wong <djwong@kernel.org>
Reviewed-by: Christoph Hellwig <hch@lst.de>

The runtime parent pointer update code expects that any file being moved
around the directory tree already has an attr fork.  However, if we had
to rebuild an inode core record, there's a chance that we zeroed forkoff
as part of the inode to pass the iget verifiers.

Therefore, if we performed any repairs on an inode core, ensure that the
inode has a nonzero forkoff before unlocking the inode.
Signed-off-by: Darrick J. Wong <djwong@kernel.org>
Reviewed-by: Christoph Hellwig <hch@lst.de>

Since the parent pointer scrubber does not exhaustively search the
filesystem for missing parent pointers, it doesn't have a good way to
determine that there are pointers missing from an otherwise uncorrupt
xattr structure.  Instead, for nondirectories it employs a heuristic of
comparing the file link count to the number of parent pointers found.

However, we don't want this heuristic flagging a false corruption after
a repair has actually scanned the entire filesystem to rebuild the
parent pointers.  Therefore, reset the file link count in this one case
because we actually know the correct link count.
Signed-off-by: Darrick J. Wong <djwong@kernel.org>
Reviewed-by: Christoph Hellwig <hch@lst.de>

Adapt the orphanage's adoption code to update the child file's parent
pointers as part of the reparenting process.  Also ensure that the child
has an attr fork to receive the parent pointer update, since the runtime
code assumes one exists.
Signed-off-by: Darrick J. Wong <djwong@kernel.org>
Reviewed-by: Christoph Hellwig <hch@lst.de>

Once we've assembled all the parent pointers for a file, we need to
commit the new dataset atomically to that file.  Parent pointer records
are embedded in the xattr structure, which means that we must write a
new extended attribute structure, again, atomically.  Therefore, we must
copy the non-parent-pointer attributes from the file being repaired into
the temporary file's extended attributes and then call the atomic extent
swap mechanism to exchange the blocks.
Signed-off-by: Darrick J. Wong <djwong@kernel.org>
Reviewed-by: Christoph Hellwig <hch@lst.de>

Add a second callback function to xchk_xattr_walk so that we can do
something in between attr leaf blocks.  This will be used by the next
patch to see if we should flush cached parent pointer updates to
constrain memory usage.
Signed-off-by: Darrick J. Wong <djwong@kernel.org>
Reviewed-by: Christoph Hellwig <hch@lst.de>

Split this function into two pieces -- one to make the actual changes to
the inode core to add the attr fork, and another one to deal with
getting the transaction and locking the inodes.

The next couple of patches will need this to be split into two.  One
patch implements committing new parent pointer recordsets to damaged
files.  If one file has an attr fork and the other does not, we have to
create the missing attr fork before the atomic swap transaction, and can
use the behavior encoded in the current xfs_bmap_add_attrfork.

The second patch adapts /lost+found adoptions to handle parent pointers
correctly.  The adoption process will add a parent pointer to a child
that is being moved to /lost+found, but this requires that the attr fork
already exists.  We don't know if we're actually going to commit the
adoption until we've already reserved a transaction and taken the
ILOCKs, which means that we must have a way to bypass the start of the
current xfs_bmap_add_attrfork.

Therefore, create xfs_attr_add_fork as the helper that creates a
transaction and takes locks; and make xfs_bmap_add_attrfork the function
that updates the inode core and allocates the incore attr fork.
Signed-off-by: Darrick J. Wong <djwong@kernel.org>
Reviewed-by: Christoph Hellwig <hch@lst.de>

Remove this assertion about the inode not having an attr fork from
xfs_bmap_add_attrfork because the function handles that case just fine.
Weirder still, the function actually /requires/ the caller not to hold
the ILOCK, which means that its accesses are not stabilized.
Signed-off-by: Darrick J. Wong <djwong@kernel.org>
Reviewed-by: Christoph Hellwig <hch@lst.de>

While we're scanning the filesystem for dirents that we can turn into
parent pointers, we cannot hold the IOLOCK or ILOCK of the file being
repaired.  Therefore, we need to set up a dirent hook so that we can
keep the temporary file's parent pionters up to date with the rest of
the filesystem.  Hence we add the ability to *remove* pptrs from the
temporary file.
Signed-off-by: Darrick J. Wong <djwong@kernel.org>
Reviewed-by: Christoph Hellwig <hch@lst.de>

If parent pointers are enabled on the filesystem, we can repair the
entire dataset by walking the directories of the filesystem looking for
dirents that we can turn into parent pointers.  Once we have a full
incore dataset, we'll figure out what to do with it, but that's for a
subsequent patch.
Signed-off-by: Darrick J. Wong <djwong@kernel.org>
Reviewed-by: Christoph Hellwig <hch@lst.de>

There are a few places where the extended attribute repair code drops
the ILOCK to apply stashed xattrs to the temporary file.  Although
setxattr and removexattr are still locked out because we retain our hold
on the IOLOCK, this doesn't prevent renames from updating parent
pointers, because the VFS doesn't take i_rwsem on children that are
being moved.

Therefore, set up a dirent hook to capture parent pointer updates for
this file, and replay(?) the updates.
Signed-off-by: Darrick J. Wong <djwong@kernel.org>
Reviewed-by: Christoph Hellwig <hch@lst.de>

While we're scanning the filesystem for parent pointers that we can turn
into dirents, we cannot hold the IOLOCK or ILOCK of the directory being
repaired.  Therefore, we need to set up a dirent hook so that we can
keep the temporary directory up to date with the rest of the filesystem.
Hence we add the ability to *remove* entries from the temporary dir.
Signed-off-by: Darrick J. Wong <djwong@kernel.org>
Reviewed-by: Christoph Hellwig <hch@lst.de>

For filesystems with parent pointers, scan the entire filesystem looking
for parent pointers that target the directory we're rebuilding instead
of trying to salvage whatever we can from the directory data blocks.
This will be more robust than salvaging, but there's more code to come.
Signed-off-by: Darrick J. Wong <djwong@kernel.org>
Reviewed-by: Christoph Hellwig <hch@lst.de>

Add a couple of utility functions to set or remove parent pointers from
a file.  These functions will be used by repair code, hence they skip
the xattr logging that regular parent pointer updates use.
Signed-off-by: Darrick J. Wong <djwong@kernel.org>
Reviewed-by: Christoph Hellwig <hch@lst.de>

When we're salvaging extended attributes, make sure we validate the ones
that claim to be parent pointers before adding them to the salvage pile.
Signed-off-by: Darrick J. Wong <djwong@kernel.org>
Reviewed-by: Christoph Hellwig <hch@lst.de>

Make the use of reserved blocks an explicit parameter to xfs_attr_set.
Userspace setting XFS_ATTR_ROOT attrs should continue to be able to use
it, but for online repairs we can back out and therefore do not care.
Signed-off-by: Darrick J. Wong <djwong@kernel.org>
Reviewed-by: Christoph Hellwig <hch@lst.de>

In preparation for online/offline repair wanting to use xfs_attr_set,
move some of the boilerplate out of this function into the callers.
Repair can initialize the da_args completely, and the userspace flag
handling/twisting goes away once we move it to xfs_attr_change.
Signed-off-by: Darrick J. Wong <djwong@kernel.org>
Reviewed-by: Christoph Hellwig <hch@lst.de>

Check parent pointer xattrs as part of scrubbing xattrs.
Signed-off-by: Darrick J. Wong <djwong@kernel.org>
Reviewed-by: Christoph Hellwig <hch@lst.de>

If the filesystem has parent pointers enabled, walk the parent pointers
of subdirectories to determine the true backref count.  In theory each
subdir should have a single parent reachable via dotdot, but in the case
of (corrupt) subdirs with multiple parents, we need to keep the link
counts high enough that the directory loop detector will be able to
correct the multiple parents problems.
Signed-off-by: Darrick J. Wong <djwong@kernel.org>
Reviewed-by: Christoph Hellwig <hch@lst.de>

If the trylock-based dirent check fails, retain those parent pointers
and check them at the end.  This may involve dropping the locks on the
file being scanned, so yay.
Signed-off-by: Darrick J. Wong <djwong@kernel.org>
Reviewed-by: Christoph Hellwig <hch@lst.de>

Actually check parent pointers now.
Signed-off-by: Darrick J. Wong <djwong@kernel.org>
Reviewed-by: Christoph Hellwig <hch@lst.de>

If the trylock-based parent pointer check fails, retain those dirents
and check them at the end.  This may involve dropping the locks on the
file being scanned, so yay.
Signed-off-by: Darrick J. Wong <djwong@kernel.org>
Reviewed-by: Christoph Hellwig <hch@lst.de>

If the fs has parent pointers, we need to check that each child dirent
points to a file that has a parent pointer pointing back at us.
Signed-off-by: Darrick J. Wong <djwong@kernel.org>
Reviewed-by: Christoph Hellwig <hch@lst.de>

In my haste to fix what I thought was a performance problem in the attr
scrub code, I neglected to notice that the xfs_attr_get_ilocked also had
the effect of checking that attributes can actually be looked up through
the attr dabtree.  Fix this.

Fixes: 44af6c7e ("xfs: don't load local xattr values during scrub")
Signed-off-by: Darrick J. Wong <djwong@kernel.org>
Reviewed-by: Christoph Hellwig <hch@lst.de>

Add parent pointers to the list of supported features.
Signed-off-by: Darrick J. Wong <djwong@kernel.org>
Reviewed-by: Christoph Hellwig <hch@lst.de>

Let's also drop the oversized minimum log computations for reflink and
rmap that were the result of bugs introduced many years ago.
Signed-off-by: Darrick J. Wong <djwong@kernel.org>
Reviewed-by: Allison Henderson <allison.henderson@oracle.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>

Dave and I were discussing some recent test regressions as a result of
me turning on nrext64=1 on realtime filesystems, when we noticed that
the minimum log size of a 32M filesystem jumped from 954 blocks to 4287
blocks.

Digging through xfs_log_calc_max_attrsetm_res, Dave noticed that @size
contains the maximum estimated amount of space needed for a local format
xattr, in bytes, but we feed this quantity to XFS_NEXTENTADD_SPACE_RES,
which requires units of blocks.  This has resulted in an overestimation
of the minimum log size over the years.

We should nominally correct this, but there's a backwards compatibility
problem -- if we enable it now, the minimum log size will decrease.  If
a corrected mkfs formats a filesystem with this new smaller log size, a
user will encounter mount failures on an uncorrected kernel due to the
larger minimum log size computations there.

Therefore, turn this on for parent pointers because it wasn't merged at
all upstream when this issue was discovered.
Signed-off-by: Darrick J. Wong <djwong@kernel.org>
Reviewed-by: Allison Henderson <allison.henderson@oracle.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>

Create an incompat feature bit and a fs geometry flag so that we can
enable the feature in the ondisk superblock and advertise its existence
to userspace.
Signed-off-by: Mark Tinguely <mark.tinguely@oracle.com>
Signed-off-by: Dave Chinner <dchinner@redhat.com>
Signed-off-by: Allison Henderson <allison.henderson@oracle.com>
Reviewed-by: Darrick J. Wong <darrick.wong@oracle.com>
Reviewed-by: Darrick J. Wong <djwong@kernel.org>
Signed-off-by: Darrick J. Wong <djwong@kernel.org>