dax_clear_blocks() needs a valid struct block_device and previously it
was using inode->i_sb->s_bdev in all cases.  This is correct for normal
inodes on mounted ext2, ext4 and XFS filesystems, but is incorrect for
DAX raw block devices and for XFS real-time devices.

Instead, rename dax_clear_blocks() to dax_clear_sectors(), and change
its arguments to take a bdev and a sector instead of an inode and a
block.  This better reflects what the function does, and it allows the
filesystem and raw block device code to pass in an appropriate struct
block_device.
Signed-off-by: NRoss Zwisler <ross.zwisler@linux.intel.com>
Suggested-by: NDan Williams <dan.j.williams@intel.com>
Reviewed-by: NJan Kara <jack@suse.cz>
Cc: Theodore Ts'o <tytso@mit.edu>
Cc: Al Viro <viro@ftp.linux.org.uk>
Cc: Dave Chinner <david@fromorbit.com>
Cc: Jens Axboe <axboe@fb.com>
Cc: Matthew Wilcox <matthew.r.wilcox@intel.com>
Cc: Ross Zwisler <ross.zwisler@linux.intel.com>
Cc: Theodore Ts'o <tytso@mit.edu>
Signed-off-by: NAndrew Morton <akpm@linux-foundation.org>
Signed-off-by: NLinus Torvalds <torvalds@linux-foundation.org>

Calls to xfs_bmap_finish() and xfs_trans_ijoin(), and the
associated comments were replicated several times across
the attribute code, all dealing with what to do if the
transaction was or wasn't committed.

And in that replicated code, an ASSERT() test of an
uninitialized variable occurs in several locations:

	error = xfs_attr_thing(&args);
	if (!error) {
		error = xfs_bmap_finish(&args.trans, args.flist,
					&committed);
	}
	if (error) {
		ASSERT(committed);

If the first xfs_attr_thing() failed, we'd skip the xfs_bmap_finish,
never set "committed", and then test it in the ASSERT.

Fix this up by moving the committed state internal to xfs_bmap_finish,
and add a new inode argument.  If an inode is passed in, it is passed
through to __xfs_trans_roll() and joined to the transaction there if
the transaction was committed.

xfs_qm_dqalloc() was a little unique in that it called bjoin rather
than ijoin, but as Dave points out we can detect the committed state
but checking whether (*tpp != tp).

Addresses-Coverity-Id: 102360
Addresses-Coverity-Id: 102361
Addresses-Coverity-Id: 102363
Addresses-Coverity-Id: 102364
Signed-off-by: NEric Sandeen <sandeen@redhat.com>
Reviewed-by: NChristoph Hellwig <hch@lst.de>
Signed-off-by: NDave Chinner <david@fromorbit.com>

To enable DAX to do atomic allocation of zeroed extents, we need to
drive the block zeroing deep into the allocator. Because
xfs_bmapi_write() can return merged extents on allocation that were
only partially allocated (i.e. requested range spans allocated and
hole regions, allocation into the hole was contiguous), we cannot
zero the extent returned from xfs_bmapi_write() as that can
overwrite existing data with zeros.

Hence we have to drive the extent zeroing into the allocation code,
prior to where we merge the extents into the BMBT and return the
resultant map. This means we need to propagate this need down to
the xfs_alloc_vextent() and issue the block zeroing at this point.

While this functionality is being introduced for DAX, there is no
reason why it is specific to DAX - we can per-zero blocks during the
allocation transaction on any type of device. It's just slow (and
usually slower than unwritten allocation and conversion) on
traditional block devices so doesn't tend to get used. We can,
however, hook hardware zeroing optimisations via sb_issue_zeroout()
to this operation, so it may be useful in future and hence the
"allocate zeroed blocks" API needs to be implementation neutral.
Signed-off-by: NDave Chinner <dchinner@redhat.com>
Reviewed-by: NBrian Foster <bfoster@redhat.com>
Signed-off-by: NDave Chinner <david@fromorbit.com>

The total field from struct xfs_alloc_arg is a bit of an unknown
commodity. It is documented as the total block requirement for the
transaction and is used in this manner from most call sites by virtue of
passing the total block reservation of the transaction associated with
an allocation. Several xfs_bmapi_write() callers pass hardcoded values
of 0 or 1 for the total block requirement, which is a historical oddity
without any clear reasoning.

The xfs_iomap_write_direct() caller, for example, passes 0 for the total
block requirement. This has been determined to cause problems in the
form of ABBA deadlocks of AGF buffers due to incorrect AG selection in
the block allocator. Specifically, the xfs_alloc_space_available()
function incorrectly selects an AG that doesn't actually have sufficient
space for the allocation. This occurs because the args.total field is 0
and thus the remaining free space check on the AG doesn't actually
consider the size of the allocation request. This locks the AGF buffer,
the allocation attempt proceeds and ultimately fails (in
xfs_alloc_fix_minleft()), and xfs_alloc_vexent() moves on to the next
AG. In turn, this can lead to incorrect AG locking order (if the
allocator wraps around, attempting to lock AG 0 after acquiring AG N)
and thus deadlock if racing with another operation. This problem has
been reproduced via generic/299 on smallish (1GB) ramdisk test devices.

To avoid this problem, replace the undocumented hardcoded total
parameters from the iomap and utility callers to pass the block
reservation used for the associated transaction. This is consistent with
other xfs_bmapi_write() callers throughout XFS. The assumption is that
the total field allows the selection of an AG that can handle the entire
operation rather than simply the allocation/range being requested (e.g.,
resulting btree splits, etc.). This addresses the aforementioned
generic/299 hang by ensuring AG selection only occurs when the
allocation can be satisfied by the AG.
Reported-by: NRoss Zwisler <ross.zwisler@linux.intel.com>
Signed-off-by: NBrian Foster <bfoster@redhat.com>
Reviewed-by: NDave Chinner <dchinner@redhat.com>
Signed-off-by: NDave Chinner <david@fromorbit.com>

If a failure occurs after the bmap free list is populated and before
xfs_bmap_finish() completes successfully (which returns a partial
list on failure), the bmap free list must be cancelled. Otherwise,
the extent items on the list are never freed and a memory leak
occurs.

Several random error paths throughout the code suffer this problem.
Fix these up such that xfs_bmap_cancel() is always called on error.
Signed-off-by: NBrian Foster <bfoster@redhat.com>
Reviewed-by: NDave Chinner <dchinner@redhat.com>
Signed-off-by: NDave Chinner <david@fromorbit.com>

Log recovery attempts to free extents with leftover EFIs in the AIL
after initial processing. If the extent free fails (e.g., due to
unrelated fs corruption), the transaction is cancelled, though it
might not be dirtied at the time. If this is the case, the EFD does
not abort and thus does not release the EFI. This can lead to hangs
as the EFI pins the AIL.

Update xlog_recover_process_efi() to log the EFD in the transaction
before xfs_free_extent() errors are handled to ensure the
transaction is dirty, aborts the EFD and releases the EFI on error.
Since this is a requirement for EFD processing (and consistent with
xfs_bmap_finish()), update the EFD logging helper to do the extent
free and unconditionally log the EFD. This encodes the required EFD
logging behavior into the helper and reduces the likelihood of
errors down the road.

[dchinner: re-add xfs_alloc.h to xfs_log_recover.c to fix build
 failure.]
Signed-off-by: NBrian Foster <bfoster@redhat.com>
Reviewed-by: NDave Chinner <dchinner@redhat.com>
Signed-off-by: NDave Chinner <david@fromorbit.com>

Freeing an extent in XFS involves logging an EFI (extent free
intention), freeing the actual extent, and logging an EFD (extent
free done). The EFI object is created with a reference count of 2:
one for the current transaction and one for the subsequently created
EFD. Under normal circumstances, the first reference is dropped when
the EFI is unpinned and the second reference is dropped when the EFD
is committed to the on-disk log.

In event of errors or filesystem shutdown, there are various
potential cleanup scenarios depending on the state of the EFI/EFD.
The cleanup scenarios are confusing and racy, as demonstrated by the
following test sequence:

	# mount $dev $mnt
	# fsstress -d $mnt -n 99999 -p 16 -z -f fallocate=1 \
		-f punch=1 -f creat=1 -f unlink=1 &
	# sleep 5
	# killall -9 fsstress; wait
	# godown -f $mnt
	# umount

... in which the final umount can hang due to the AIL being pinned
indefinitely by one or more EFI items. This can occur due to several
conditions. For example, if the shutdown occurs after the EFI is
committed to the on-disk log and the EFD committed to the CIL, but
before the EFD committed to the log, the EFD iop_committed() abort
handler does not drop its reference to the EFI. Alternatively,
manual error injection in the xfs_bmap_finish() codepath shows that
if an error occurs after the EFI transaction is committed but before
the EFD is constructed and logged, the EFI is never released from
the AIL.

Update the EFI/EFD item handling code to use a more straightforward
and reliable approach to error handling. If an error occurs after
the EFI transaction is committed and before the EFD is constructed,
release the EFI explicitly from xfs_bmap_finish(). If the EFI
transaction is cancelled, release the EFI in the unlock handler.

Once the EFD is constructed, it is responsible for releasing the EFI
under any circumstances (including whether the EFI item aborts due
to log I/O error). Update the EFD item handlers to release the EFI
if the transaction is cancelled or aborts due to log I/O error.
Finally, update xfs_bmap_finish() to log at least one EFD extent to
the transaction before xfs_free_extent() errors are handled to
ensure the transaction is dirty and EFD item error handling is
triggered.
Signed-off-by: NBrian Foster <bfoster@redhat.com>
Reviewed-by: NDave Chinner <dchinner@redhat.com>
Signed-off-by: NDave Chinner <david@fromorbit.com>

The flags argument to xfs_trans_commit is not useful for most callers, as
a commit of a transaction without a permanent log reservation must pass
0 here, and all callers for a transaction with a permanent log reservation
except for xfs_trans_roll must pass XFS_TRANS_RELEASE_LOG_RES.  So remove
the flags argument from the public xfs_trans_commit interfaces, and
introduce low-level __xfs_trans_commit variant just for xfs_trans_roll
that regrants a log reservation instead of releasing it.
Signed-off-by: NChristoph Hellwig <hch@lst.de>
Reviewed-by: NDave Chinner <dchinner@redhat.com>
Signed-off-by: NDave Chinner <david@fromorbit.com>

xfs_trans_cancel takes two flags arguments: XFS_TRANS_RELEASE_LOG_RES and
XFS_TRANS_ABORT.  Both of them are a direct product of the transaction
state, and can be deducted:

 - any dirty transaction needs XFS_TRANS_ABORT to be properly canceled,
   and XFS_TRANS_ABORT is a noop for a transaction that is not dirty.
 - any transaction with a permanent log reservation needs
   XFS_TRANS_RELEASE_LOG_RES to be properly canceled, and passing
   XFS_TRANS_RELEASE_LOG_RES for a transaction without a permanent
   log reservation is invalid.

So just remove the flags argument and do the right thing.
Signed-off-by: NChristoph Hellwig <hch@lst.de>
Reviewed-by: NDave Chinner <dchinner@redhat.com>
Signed-off-by: NDave Chinner <david@fromorbit.com>

We have three remaining callers of xfs_trans_dup:

 - xfs_itruncate_extents which open codes xfs_trans_roll
 - xfs_bmap_finish doesn't have an xfs_inode argument and thus leaves
   attaching them to it's callers, but otherwise is identical to
   xfs_trans_roll
 - xfs_dir_ialloc looks at the log reservations in the old xfs_trans
   structure instead of the log reservation parameters, but otherwise
   is identical to xfs_trans_roll.

By allowing a NULL xfs_inode argument to xfs_trans_roll we can switch
these three remaining users over to xfs_trans_roll and mark xfs_trans_dup
static.
Signed-off-by: NChristoph Hellwig <hch@lst.de>
Reviewed-by: NDave Chinner <dchinner@redhat.com>
Signed-off-by: NDave Chinner <david@fromorbit.com>

Add initial support for DAX block zeroing operations to XFS. DAX
cannot use buffered IO through the page cache for zeroing, nor do we
need to issue IO for uncached block zeroing. In both cases, we can
simply call out to the dax block zeroing function.
Signed-off-by: NDave Chinner <dchinner@redhat.com>
Reviewed-by: NBrian Foster <bfoster@redhat.com>
Signed-off-by: NDave Chinner <david@fromorbit.com>

Signed-off-by: NFengguang Wu <fengguang.wu@intel.com>
Reviewed-by: NDave Chinner <dchinner@redhat.com>
Signed-off-by: NDave Chinner <david@fromorbit.com>

This patch implements fallocate's FALLOC_FL_INSERT_RANGE for XFS.

1) Make sure that both offset and len are block size aligned.
2) Update the i_size of inode by len bytes.
3) Compute the file's logical block number against offset. If the computed
   block number is not the starting block of the extent, split the extent
   such that the block number is the starting block of the extent.
4) Shift all the extents which are lying bewteen [offset, last allocated extent]
   towards right by len bytes. This step will make a hole of len bytes
   at offset.
Signed-off-by: NNamjae Jeon <namjae.jeon@samsung.com>
Signed-off-by: NAshish Sangwan <a.sangwan@samsung.com>
Reviewed-by: NBrian Foster <bfoster@redhat.com>
Reviewed-by: NDave Chinner <dchinner@redhat.com>
Signed-off-by: NDave Chinner <david@fromorbit.com>

Extent swap operations are another extent manipulation operation
that we need to ensure does not race against mmap page faults. The
current code returns if the file is mapped prior to the swap being
done, but it could potentially race against new page faults while
the swap is in progress. Hence we should use the XFS_MMAPLOCK_EXCL
for this operation, too.

While there, fix the error path handling that can result in double
unlocks of the inodes when cancelling the swapext transaction.
Signed-off-by: NDave Chinner <dchinner@redhat.com>
Reviewed-by: NBrian Foster <bfoster@redhat.com>
Signed-off-by: NDave Chinner <david@fromorbit.com>

More on-disk format consolidation.
Signed-off-by: NChristoph Hellwig <hch@lst.de>
Reviewed-by: NDave Chinner <dchinner@redhat.com>
Signed-off-by: NDave Chinner <david@fromorbit.com>

More on-disk format consolidation.  A few declarations that weren't on-disk
format related move into better suitable spots.
Signed-off-by: NChristoph Hellwig <hch@lst.de>
Reviewed-by: NDave Chinner <dchinner@redhat.com>
Signed-off-by: NDave Chinner <david@fromorbit.com>

More consolidatation for the on-disk format defintions.  Note that the
XFS_IS_REALTIME_INODE moves to xfs_linux.h instead as it is not related
to the on disk format, but depends on a CONFIG_ option.
Signed-off-by: NChristoph Hellwig <hch@lst.de>
Reviewed-by: NDave Chinner <dchinner@redhat.com>
Signed-off-by: NDave Chinner <david@fromorbit.com>

The zero range operation is analogous to fallocate with the exception of
converting the range to zeroes. E.g., it attempts to allocate zeroed
blocks over the range specified by the caller. The XFS implementation
kills all delalloc blocks currently over the aligned range, converts the
range to allocated zero blocks (unwritten extents) and handles the
partial pages at the ends of the range by sending writes through the
pagecache.

The current implementation suffers from several problems associated with
inode size. If the aligned range covers an extending I/O, said I/O is
discarded and an inode size update from a previous write never makes it
to disk. Further, if an unaligned zero range extends beyond eof, the
page write induced for the partial end page can itself increase the
inode size, even if the zero range request is not supposed to update
i_size (via KEEP_SIZE, similar to an fallocate beyond EOF).

The latter behavior not only incorrectly increases the inode size, but
can lead to stray delalloc blocks on the inode. Typically, post-eof
preallocation blocks are either truncated on release or inode eviction
or explicitly written to by xfs_zero_eof() on natural file size
extension. If the inode size increases due to zero range, however,
associated blocks leak into the address space having never been
converted or mapped to pagecache pages. A direct I/O to such an
uncovered range cannot convert the extent via writeback and will BUG().
For example:

$ xfs_io -fc "pwrite 0 128k" -c "fzero -k 1m 54321" <file>
...
$ xfs_io -d -c "pread 128k 128k" <file>
<BUG>

If the entire delalloc extent happens to not have page coverage
whatsoever (e.g., delalloc conversion couldn't find a large enough free
space extent), even a full file writeback won't convert what's left of
the extent and we'll assert on inode eviction.

Rework xfs_zero_file_space() to avoid buffered I/O for partial pages.
Use the existing hole punch and prealloc mechanisms as primitives for
zero range. This implementation is not efficient nor ideal as we
writeback dirty data over the range and remove existing extents rather
than convert to unwrittern. The former writeback, however, is currently
the only mechanism available to ensure consistency between pagecache and
extent state. Even a pagecache truncate/delalloc punch prior to hole
punch has lead to inconsistencies due to racing with writeback.

This provides a consistent, correct implementation of zero range that
survives fsstress/fsx testing without assert failures. The
implementation can be optimized from this point forward once the
fundamental issue of pagecache and delalloc extent state consistency is
addressed.
Signed-off-by: NBrian Foster <bfoster@redhat.com>
Reviewed-by: NDave Chinner <dchinner@redhat.com>
Signed-off-by: NDave Chinner <david@fromorbit.com>

XFS currently discards delalloc blocks within the target range of a
zero range request. Unaligned start and end offsets are zeroed
through the page cache and the internal, aligned blocks are
converted to unwritten extents.

If EOF is page aligned and covered by a delayed allocation extent.
The inode size is not updated until I/O completion. If a zero range
request discards a delalloc range that covers page aligned EOF as
such, the inode size update never occurs. For example:

$ rm -f /mnt/file
$ xfs_io -fc "pwrite 0 64k" -c "zero 60k 4k" /mnt/file
$ stat -c "%s" /mnt/file
65536
$ umount /mnt
$ mount <dev> /mnt
$ stat -c "%s" /mnt/file
61440

Update xfs_zero_file_space() to flush the range rather than discard
delalloc blocks to ensure that inode size updates occur
appropriately.

[dchinner: Note that this is really a workaround to avoid the
underlying problems. More work is needed (and ongoing) to fix those
issues so this fix is being added as a temporary stop-gap measure. ]
Signed-off-by: NBrian Foster <bfoster@redhat.com>
Reviewed-by: NDave Chinner <dchinner@redhat.com>
Signed-off-by: NDave Chinner <david@fromorbit.com>

xfs_zero_remaining_bytes() open codes a log of buffer manupulations
to do a read forllowed by a write. It can simply be replaced by an
uncached read followed by a xfs_bwrite() call.
Signed-off-by: NChristoph Hellwig <hch@lst.de>
Reviewed-by: NDave Chinner <dchinner@redhat.com>
Signed-off-by: NDave Chinner <david@fromorbit.com>

There is a lot of cookie-cutter code that looks like:

	if (shutdown)
		handle buffer error
	xfs_buf_iorequest(bp)
	error = xfs_buf_iowait(bp)
	if (error)
		handle buffer error

spread through XFS. There's significant complexity now in
xfs_buf_iorequest() to specifically handle this sort of synchronous
IO pattern, but there's all sorts of nasty surprises in different
error handling code dependent on who owns the buffer references and
the locks.

Pull this pattern into a single helper, where we can hide all the
synchronous IO warts and hence make the error handling for all the
callers much saner. This removes the need for a special extra
reference to protect IO completion processing, as we can now hold a
single reference across dispatch and waiting, simplifying the sync
IO smeantics and error handling.

In doing this, also rename xfs_buf_iorequest to xfs_buf_submit and
make it explicitly handle on asynchronous IO. This forces all users
to be switched specifically to one interface or the other and
removes any ambiguity between how the interfaces are to be used. It
also means that xfs_buf_iowait() goes away.
Signed-off-by: NDave Chinner <dchinner@redhat.com>
Reviewed-by: NChristoph Hellwig <hch@lst.de>
Signed-off-by: NDave Chinner <david@fromorbit.com>

Fix sparse warning introduced by commit 4ef897a2 ("xfs: flush both
inodes in xfs_swap_extents").
Signed-off-by: NFengguang Wu <fengguang.wu@intel.com>
Reviewed-by: NDave Chinner <dchinner@redhat.com>
Signed-off-by: NDave Chinner <david@fromorbit.com>

xfs_free_file_space() only affects the range of the file for which space
is being freed. It currently writes and truncates the page cache from
the start offset of the free to EOF.

Modify xfs_free_file_space() to write back and truncate page cache of
just the range being freed.
Signed-off-by: NBrian Foster <bfoster@redhat.com>
Reviewed-by: NDave Chinner <dchinner@redhat.com>
Signed-off-by: NDave Chinner <david@fromorbit.com>

The collapse range operation currently writes the entire file before
starting the collapse to avoid changes in the in-core extent list due to
writeback causing the extent count to change. Now that collapse range is
fsb based rather than extent index based it can sustain changes in the
extent list during the shift sequence without disruption.

Modify xfs_collapse_file_space() to writeback and invalidate pages
associated with the range of the file to be shifted.
xfs_free_file_space() currently has similar behavior, but the space free
need only affect the region of the file that is freed and this could
change in the future.

Also update the comments to reflect the current implementation. We
retain the eofblocks trim permanently as a best option for dealing with
delalloc extents. We don't shift delalloc extents because this scenario
only occurs with post-eof preallocation (since data must be flushed such
that the cache can be invalidated and data can be shifted). That means
said space must also be initialized before being shifted into the
accessible region of the file only to be immediately truncated off as
the last part of the collapse. In other words, the eofblocks trim will
happen anyways, we just run it first to ensure the file remains in a
consistent state throughout the collapse.

Finally, detect and fail explicitly in the event of a delalloc extent
during the extent shift. The implementation does not support delalloc
extents and the caller is expected to prevent this scenario in advance
as is done by collapse.
Signed-off-by: NBrian Foster <bfoster@redhat.com>
Reviewed-by: NDave Chinner <dchinner@redhat.com>
Signed-off-by: NDave Chinner <david@fromorbit.com>

The collapse range implementation uses a transaction per extent shift.
The progress of the overall operation is tracked via the current extent
index of the in-core extent list. This is racy because the ilock must be
dropped and reacquired for each transaction according to locking and log
reservation rules. Therefore, writeback to prior regions of the file is
possible and can change the extent count. This changes the extent to
which the current index refers and causes the collapse to fail mid
operation. To avoid this problem, the entire file is currently written
back before the collapse operation starts.

To eliminate the need to flush the entire file, use the file offset
(fsb) to track the progress of the overall extent shift operation rather
than the extent index. Modify xfs_bmap_shift_extents() to
unconditionally convert the start_fsb parameter to an extent index and
return the file offset of the extent where the shift left off, if
further extents exist. The bulk of ths function can remain based on
extent index as ilock is held by the caller. xfs_collapse_file_space()
now uses the fsb output as the starting point for the subsequent shift.
Signed-off-by: NBrian Foster <bfoster@redhat.com>
Reviewed-by: NDave Chinner <dchinner@redhat.com>
Signed-off-by: NDave Chinner <david@fromorbit.com>

xfs_collapse_file_space() currently writes back the entire file
undergoing collapse range to settle things down for the extent shift
algorithm. While this prevents changes to the extent list during the
collapse operation, the writeback itself is not enough to prevent
unnecessary collapse failures.

The current shift algorithm uses the extent index to iterate the in-core
extent list. If a post-eof delalloc extent persists after the writeback
(e.g., a prior zero range op where the end of the range aligns with eof
can separate the post-eof blocks such that they are not written back and
converted), xfs_bmap_shift_extents() becomes confused over the encoded
br_startblock value and fails the collapse.

As with the full writeback, this is a temporary fix until the algorithm
is improved to cope with a volatile extent list and avoid attempts to
shift post-eof extents.
Signed-off-by: NBrian Foster <bfoster@redhat.com>
Reviewed-by: NDave Chinner <dchinner@redhat.com>
Reviewed-by: NChristoph Hellwig <hch@lst.de>
Signed-off-by: NDave Chinner <david@fromorbit.com>

If we have delalloc extents on a file before we run a collapse range
opertaion, we sync the range that we are going to collapse to
convert delalloc extents in that region to real extents to simplify
the shift operation.

However, the shift operation then assumes that the extent list is
not going to change as it iterates over the extent list moving
things about. Unfortunately, this isn't true because we can't hold
the ILOCK over all the operations. We can prevent new IO from
modifying the extent list by holding the IOLOCK, but that doesn't
prevent writeback from running....

And when writeback runs, it can convert delalloc extents is the
range of the file prior to the region being collapsed, and this
changes the indexes of all the extents in the file. That causes the
collapse range operation to Go Bad.

The right fix is to rewrite the extent shift operation not to be
dependent on the extent list not changing across the entire
operation, but this is a fairly significant piece of work to do.
Hence, as a short-term workaround for the problem, sync the entire
file before starting a collapse operation to remove all delalloc
ranges from the file and so avoid the problem of concurrent
writeback changing the extent list.
Diagnosed-and-Reported-by: NBrian Foster <bfoster@redhat.com>
Signed-off-by: NDave Chinner <dchinner@redhat.com>
Reviewed-by: NBrian Foster <bfoster@redhat.com>
Reviewed-by: NChristoph Hellwig <hch@lst.de>
Signed-off-by: NDave Chinner <david@fromorbit.com>

We need to treat both inodes identically from a page cache point of
view when prepareing them for extent swapping. We don't do this
right now - we assume that one of the inodes empty, because that's
what xfs_fsr currently does. Remove this assumption from the code.

While factoring out the flushing and related checks, move the
transactions reservation to immeidately after the flushes so that we
don't need to pick up and then drop the ilock to do the transaction
reservation. There are no issues with aborting the transaction it if
the checks fail before we join the inodes to the transaction and
dirty them, so this is a safe change to make.
Signed-off-by: NDave Chinner <dchinner@redhat.com>
Reviewed-by: NChristoph Hellwig <hch@lst.de>
Signed-off-by: NDave Chinner <david@fromorbit.com>

xfs_swap_extents() holds the ilock over a call to
filemap_write_and_wait(), which can then try to write data and take
the ilock. That causes a self-deadlock.

Fix the deadlock and clean up the code by separating the locking
appropriately. Add a lockflags variable to track what locks we are
holding as we gain and drop them and cleanup the error handling to
always use "out_unlock" with the lockflags variable.
Signed-off-by: NDave Chinner <dchinner@redhat.com>
Reviewed-by: NChristoph Hellwig <hch@lst.de>
Signed-off-by: NDave Chinner <david@fromorbit.com>

Only one user, no longer needed.
Signed-off-by: NDave Chinner <dchinner@redhat.com>
Reviewed-by: NChristoph Hellwig <hch@lst.de>
Signed-off-by: NDave Chinner <david@fromorbit.com>

Only has 2 users, has outlived it's usefulness.
Signed-off-by: NDave Chinner <dchinner@redhat.com>
Reviewed-by: NChristoph Hellwig <hch@lst.de>
Signed-off-by: NDave Chinner <david@fromorbit.com>

Trying to support tiny disks only and saving a bit memory might have
made sense on an SGI O2 15 years ago, but is pretty pointless today.

Remove the rarely tested codepath that uses various smaller in-memory
types to reduce our test matrix and make the codebase a little bit
smaller and less complicated.
Signed-off-by: NChristoph Hellwig <hch@lst.de>
Reviewed-by: NBen Myers <bpm@sgi.com>
Signed-off-by: NDave Chinner <david@fromorbit.com>

The allocation stack switch at xfs_bmapi_allocate() has served it's
purpose, but is no longer a sufficient solution to the stack usage
problem we have in the XFS allocation path.

Whilst the kernel stack size is now 16k, that is not a valid reason
for undoing all our "keep stack usage down" modifications. What it
does allow us to do is have the freedom to refine and perfect the
modifications knowing that if we get it wrong it won't blow up in
our faces - we have a safety net now.

This is important because we still have the issue of older kernels
having smaller stacks and that they are still supported and are
demonstrating a wide range of different stack overflows.  Red Hat
has several open bugs for allocation based stack overflows from
directory modifications and direct IO block allocation and these
problems still need to be solved. If we can solve them upstream,
then distro's won't need to bake their own unique solutions.

To that end, I've observed that every allocation based stack
overflow report has had a specific characteristic - it has happened
during or directly after a bmap btree block split. That event
requires a new block to be allocated to the tree, and so we
effectively stack one allocation stack on top of another, and that's
when we get into trouble.

A further observation is that bmap btree block splits are much rarer
than writeback allocation - over a range of different workloads I've
observed the ratio of bmap btree inserts to splits ranges from 100:1
(xfstests run) to 10000:1 (local VM image server with sparse files
that range in the hundreds of thousands to millions of extents).
Either way, bmap btree split events are much, much rarer than
allocation events.

Finally, we have to move the kswapd state to the allocation workqueue
work when allocation is done on behalf of kswapd. This is proving to
cause significant perturbation in performance under memory pressure
and appears to be generating allocation deadlock warnings under some
workloads, so avoiding the use of a workqueue for the majority of
kswapd writeback allocation will minimise the impact of such
behaviour.

Hence it makes sense to move the stack switch to xfs_btree_split()
and only do it for bmap btree splits. Stack switches during
allocation will be much rarer, so there won't be significant
performacne overhead caused by switching stacks. The worse case
stack from all allocation paths will be split, not just writeback.
And the majority of memory allocations will be done in the correct
context (e.g. kswapd) without causing additional latency, and so we
simplify the memory reclaim interactions between processes,
workqueues and kswapd.

The worst stack I've been able to generate with this patch in place
is 5600 bytes deep. It's very revealing because we exit XFS at:

37)     1768      64   kmem_cache_alloc+0x13b/0x170

about 1800 bytes of stack consumed, and the remaining 3800 bytes
(and 36 functions) is memory reclaim, swap and the IO stack. And
this occurs in the inode allocation from an open(O_CREAT) syscall,
not writeback.

The amount of stack being used is much less than I've previously be
able to generate - fs_mark testing has been able to generate stack
usage of around 7k without too much trouble; with this patch it's
only just getting to 5.5k. This is primarily because the metadata
allocation paths (e.g. directory blocks) are no longer causing
double splits on the same stack, and hence now stack tracing is
showing swapping being the worst stack consumer rather than XFS.

Performance of fs_mark inode create workloads is unchanged.
Performance of fs_mark async fsync workloads is consistently good
with context switches reduced by around 150,000/s (30%).
Performance of dbench, streaming IO and postmark is unchanged.
Allocation deadlock warnings have not been seen on the workloads
that generated them since adding this patch.
Signed-off-by: NDave Chinner <dchinner@redhat.com>
Reviewed-by: NBrian Foster <bfoster@redhat.com>
Signed-off-by: NDave Chinner <david@fromorbit.com>

This reverts commit 1f6d6482.

This commit resulted in regressions in performance in low
memory situations where kswapd was doing writeback of delayed
allocation blocks. It resulted in significant parallelism of the
kswapd work and with the special kswapd flags meant that hundreds of
active allocation could dip into kswapd specific memory reserves and
avoid being throttled. This cause a large amount of performance
variation, as well as random OOM-killer invocations that didn't
previously exist.
Signed-off-by: NDave Chinner <dchinner@redhat.com>
Reviewed-by: NBrian Foster <bfoster@redhat.com>
Signed-off-by: NDave Chinner <david@fromorbit.com>

Convert all the errors the core XFs code to negative error signs
like the rest of the kernel and remove all the sign conversion we
do in the interface layers.

Errors for conversion (and comparison) found via searches like:

$ git grep " E" fs/xfs
$ git grep "return E" fs/xfs
$ git grep " E[A-Z].*;$" fs/xfs

Negation points found via searches like:

$ git grep "= -[a-z,A-Z]" fs/xfs
$ git grep "return -[a-z,A-D,F-Z]" fs/xfs
$ git grep " -[a-z].*;" fs/xfs

[ with some bits I missed from Brian Foster ]
Signed-off-by: NDave Chinner <dchinner@redhat.com>
Reviewed-by: NBrian Foster <bfoster@redhat.com>
Signed-off-by: NDave Chinner <david@fromorbit.com>

XFS_ERROR was designed long ago to trap return values, but it's not
runtime configurable, it's not consistently used, and we can do
similar error trapping with ftrace scripts and triggers from
userspace.

Just nuke XFS_ERROR and associated bits.
Signed-off-by: NEric Sandeen <sandeen@redhat.com>
Reviewed-by: NDave Chinner <dchinner@redhat.com>
Signed-off-by: NDave Chinner <david@fromorbit.com>

Upon memory pressure, kswapd calls xfs_vm_writepage() from
shrink_page_list(). This can result in delayed allocation occurring
and that gets deferred to the the allocation workqueue.

The allocation then runs outside kswapd context, which means if it
needs memory (and it does to demand page metadata from disk) it can
block in shrink_inactive_list() waiting for IO congestion. These
blocking waits are normally avoiding in kswapd context, so under
memory pressure writeback from kswapd can be arbitrarily delayed by
memory reclaim.

To avoid this, pass the kswapd context to the allocation being done
by the workqueue, so that memory reclaim understands correctly that
the work is being done for kswapd and therefore it is not blocked
and does not delay memory reclaim.

To avoid issues with int->char conversion of flag fields (as noticed
in v1 of this patch) convert the flag fields in the struct
xfs_bmalloca to bool types. pahole indicates these variables are
still single byte variables, so no extra space is consumed by this
change.

cc: <stable@vger.kernel.org>
Reported-by: NTetsuo Handa <penguin-kernel@I-love.SAKURA.ne.jp>
Signed-off-by: NDave Chinner <dchinner@redhat.com>
Reviewed-by: NChristoph Hellwig <hch@lst.de>
Signed-off-by: NDave Chinner <david@fromorbit.com>

There is no need to dip into reserve pool. Reserve pool is used for much
more important things. And xfs_trans_reserve will never return ENOSPC
because punch hole is already done. If we get ENOSPC, collapse range
will be simply failed.

Cc: Brian Foster <bfoster@redhat.com>
Signed-off-by: NNamjae Jeon <namjae.jeon@samsung.com>
Signed-off-by: NAshish Sangwan <a.sangwan@samsung.com>
Reviewed-by: NBrian Foster <bfoster@redhat.com>
Signed-off-by: NDave Chinner <david@fromorbit.com>

When we are zeroing space andit is covered by a delalloc range, we
need to punch the delalloc range out before we truncate the page
cache. Failing to do so leaves and inconsistency between the page
cache and the extent tree, which we later trip over when doing
direct IO over the same range.
Signed-off-by: NDave Chinner <dchinner@redhat.com>
Tested-by: NBrian Foster <bfoster@redhat.com>
Reviewed-by: NChristoph Hellwig <hch@lst.de>
Signed-off-by: NDave Chinner <david@fromorbit.com>

This patch implements fallocate's FALLOC_FL_COLLAPSE_RANGE for XFS.

The semantics of this flag are following:
1) It collapses the range lying between offset and length by removing any data
   blocks which are present in this range and than updates all the logical
   offsets of extents beyond "offset + len" to nullify the hole created by
   removing blocks. In short, it does not leave a hole.
2) It should be used exclusively. No other fallocate flag in combination.
3) Offset and length supplied to fallocate should be fs block size aligned
   in case of xfs and ext4.
4) Collaspe range does not work beyond i_size.
Signed-off-by: NNamjae Jeon <namjae.jeon@samsung.com>
Signed-off-by: NAshish Sangwan <a.sangwan@samsung.com>
Reviewed-by: NDave Chinner <dchinner@redhat.com>
Signed-off-by: NDave Chinner <david@fromorbit.com>