cascardo/linux.git
7 years agoxfs: try other AGs to allocate a BMBT block
Darrick J. Wong [Mon, 3 Oct 2016 16:11:45 +0000 (09:11 -0700)]
xfs: try other AGs to allocate a BMBT block

Prior to the introduction of reflink, allocating a block and mapping
it into a file was performed in a single transaction with a single
block reservation, and the allocator was supposed to find enough
blocks to allocate the extent and any BMBT blocks that might be
necessary (unless we're low on space).

However, due to the way copy on write works, allocation and mapping
have been split into two transactions, which means that we must be
able to handle the case where we allocate an extent for CoW but that
AG runs out of free space before the blocks can be mapped into a file,
and the mapping requires a new BMBT block.  When this happens, look in
one of the other AGs for a BMBT block instead of taking the FS down.

The same applies to the functions that convert a data fork to extents
and later btree format.

Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
7 years agoxfs: don't allow reflink when the AG is low on space
Darrick J. Wong [Mon, 3 Oct 2016 16:11:45 +0000 (09:11 -0700)]
xfs: don't allow reflink when the AG is low on space

If the AG free space is down to the reserves, refuse to reflink our
way out of space.  Hopefully userspace will make a real copy and/or go
elsewhere.

Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
7 years agoxfs: preallocate blocks for worst-case btree expansion
Darrick J. Wong [Mon, 3 Oct 2016 16:11:44 +0000 (09:11 -0700)]
xfs: preallocate blocks for worst-case btree expansion

To gracefully handle the situation where a CoW operation turns a
single refcount extent into a lot of tiny ones and then run out of
space when a tree split has to happen, use the per-AG reserved block
pool to pre-allocate all the space we'll ever need for a maximal
btree.  For a 4K block size, this only costs an overhead of 0.3% of
available disk space.

When reflink is enabled, we have an unfortunate problem with rmap --
since we can share a block billions of times, this means that the
reverse mapping btree can expand basically infinitely.  When an AG is
so full that there are no free blocks with which to expand the rmapbt,
the filesystem will shut down hard.

This is rather annoying to the user, so use the AG reservation code to
reserve a "reasonable" amount of space for rmap.  We'll prevent
reflinks and CoW operations if we think we're getting close to
exhausting an AG's free space rather than shutting down, but this
permanent reservation should be enough for "most" users.  Hopefully.

Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
[hch@lst.de: ensure that we invalidate the freed btree buffer]
Signed-off-by: Christoph Hellwig <hch@lst.de>
7 years agoxfs: create a separate cow extent size hint for the allocator
Darrick J. Wong [Mon, 3 Oct 2016 16:11:43 +0000 (09:11 -0700)]
xfs: create a separate cow extent size hint for the allocator

Create a per-inode extent size allocator hint for copy-on-write.  This
hint is separate from the existing extent size hint so that CoW can
take advantage of the fragmentation-reducing properties of extent size
hints without disabling delalloc for regular writes.

The extent size hint that's fed to the allocator during a copy on
write operation is the greater of the cowextsize and regular extsize
hint.

During reflink, if we're sharing the entire source file to the entire
destination file and the destination file doesn't already have a
cowextsize hint, propagate the source file's cowextsize hint to the
destination file.

Furthermore, zero the bulkstat buffer prior to setting the fields
so that we don't copy kernel memory contents into userspace.

Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
7 years agoxfs: unshare a range of blocks via fallocate
Darrick J. Wong [Mon, 3 Oct 2016 16:11:43 +0000 (09:11 -0700)]
xfs: unshare a range of blocks via fallocate

Unshare all shared extents if the user calls fallocate with the new
unshare mode flag set, so that we can guarantee that a subsequent
write will not ENOSPC.

Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
[hch: pass inode instead of file to xfs_reflink_dirty_range,
      use iomap infrastructure for copy up]
Signed-off-by: Christoph Hellwig <hch@lst.de>
7 years agoxfs: swap inode reflink flags when swapping inode extents
Darrick J. Wong [Mon, 3 Oct 2016 16:11:42 +0000 (09:11 -0700)]
xfs: swap inode reflink flags when swapping inode extents

When we're swapping the extents of two inodes, be sure to swap the
reflink inode flag too.

Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
7 years agoxfs: teach get_bmapx about shared extents and the CoW fork
Darrick J. Wong [Mon, 3 Oct 2016 16:11:41 +0000 (09:11 -0700)]
xfs: teach get_bmapx about shared extents and the CoW fork

Teach xfs_getbmapx how to report shared extents and CoW fork contents
accurately in the bmap output by querying the refcount btree
appropriately.

Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
7 years agoxfs: add dedupe range vfs function
Darrick J. Wong [Mon, 3 Oct 2016 16:11:41 +0000 (09:11 -0700)]
xfs: add dedupe range vfs function

Define a VFS function which allows userspace to request that the
kernel reflink a range of blocks between two files if the ranges'
contents match.  The function fits the new VFS ioctl that standardizes
the checking for the btrfs EXTENT SAME ioctl.

Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
7 years agoxfs: add clone file and clone range vfs functions
Darrick J. Wong [Mon, 3 Oct 2016 16:11:40 +0000 (09:11 -0700)]
xfs: add clone file and clone range vfs functions

Define two VFS functions which allow userspace to reflink a range of
blocks between two files or to reflink one file's contents to another.
These functions fit the new VFS ioctls that standardize the checking
for the btrfs CLONE and CLONE RANGE ioctls.

Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
7 years agoxfs: reflink extents from one file to another
Darrick J. Wong [Mon, 3 Oct 2016 16:11:40 +0000 (09:11 -0700)]
xfs: reflink extents from one file to another

Reflink extents from one file to another; that is to say, iteratively
remove the mappings from the destination file, copy the mappings from
the source file to the destination file, and increment the reference
count of all the blocks that got remapped.

Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
7 years agoxfs: store in-progress CoW allocations in the refcount btree
Darrick J. Wong [Mon, 3 Oct 2016 16:11:39 +0000 (09:11 -0700)]
xfs: store in-progress CoW allocations in the refcount btree

Due to the way the CoW algorithm in XFS works, there's an interval
during which blocks allocated to handle a CoW can be lost -- if the FS
goes down after the blocks are allocated but before the block
remapping takes place.  This is exacerbated by the cowextsz hint --
allocated reservations can sit around for a while, waiting to get
used.

Since the refcount btree doesn't normally store records with refcount
of 1, we can use it to record these in-progress extents.  In-progress
blocks cannot be shared because they're not user-visible, so there
shouldn't be any conflicts with other programs.  This is a better
solution than holding EFIs during writeback because (a) EFIs can't be
relogged currently, (b) even if they could, EFIs are bound by
available log space, which puts an unnecessary upper bound on how much
CoW we can have in flight, and (c) we already have a mechanism to
track blocks.

At mount time, read the refcount records and free anything we find
with a refcount of 1 because those were in-progress when the FS went
down.

Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
7 years agoxfs: cancel pending CoW reservations when destroying inodes
Darrick J. Wong [Mon, 3 Oct 2016 16:11:38 +0000 (09:11 -0700)]
xfs: cancel pending CoW reservations when destroying inodes

When destroying the inode, cancel all pending reservations in the CoW
fork so that all the reserved blocks go back to the free pile.  In
theory this sort of cleanup is only needed to clean up after write
errors.

Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
7 years agoxfs: cancel CoW reservations and clear inode reflink flag when freeing blocks
Darrick J. Wong [Mon, 3 Oct 2016 16:11:38 +0000 (09:11 -0700)]
xfs: cancel CoW reservations and clear inode reflink flag when freeing blocks

When we're freeing blocks (truncate, punch, etc.), clear all CoW
reservations in the range being freed.  If the file block count
drops to zero, also clear the inode reflink flag.

Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
7 years agoxfs: implement CoW for directio writes
Darrick J. Wong [Mon, 3 Oct 2016 16:11:37 +0000 (09:11 -0700)]
xfs: implement CoW for directio writes

For O_DIRECT writes to shared blocks, we have to CoW them just like
we would with buffered writes.  For writes that are not block-aligned,
just bounce them to the page cache.

For block-aligned writes, however, we can do better than that.  Use
the same mechanisms that we employ for buffered CoW to set up a
delalloc reservation, allocate all the blocks at once, issue the
writes against the new blocks and use the same ioend functions to
remap the blocks after the write.  This should be fairly performant.

Christoph discovered that xfs_reflink_allocate_cow_range may stumble
over invalid entries in the extent array given that it drops the ilock
but still expects the index to be stable.  Simple fixing it to a new
lookup for every iteration still isn't correct given that
xfs_bmapi_allocate will trigger a BUG_ON() if hitting a hole, and
there is nothing preventing a xfs_bunmapi_cow call removing extents
once we dropped the ilock either.

This patch duplicates the inner loop of xfs_bmapi_allocate into a
helper for xfs_reflink_allocate_cow_range so that it can be done under
the same ilock critical section as our CoW fork delayed allocation.
The directio CoW warts will be revisited in a later patch.

Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
Signed-off-by: Christoph Hellwig <hch@lst.de>
7 years agoxfs: report shared extent mappings to userspace correctly
Darrick J. Wong [Mon, 3 Oct 2016 16:11:36 +0000 (09:11 -0700)]
xfs: report shared extent mappings to userspace correctly

Report shared extents through the iomap interface so that FIEMAP flags
shared blocks accurately.  Have xfs_vm_bmap return zero for reflinked
files because the bmap-based swap code requires static block mappings,
which is incompatible with copy on write.

NOTE: Existing userspace bmap users such as lilo will have the same
problem with reflink files.

Signed-off-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
7 years agoxfs: move mappings from cow fork to data fork after copy-write
Darrick J. Wong [Mon, 3 Oct 2016 16:11:35 +0000 (09:11 -0700)]
xfs: move mappings from cow fork to data fork after copy-write

After the write component of a copy-write operation finishes, clean up
the bookkeeping left behind.  On error, we simply free the new blocks
and pass the error up.  If we succeed, however, then we must remove
the old data fork mapping and move the cow fork mapping to the data
fork.

Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
[hch: Call the CoW failure function during xfs_cancel_ioend]
Signed-off-by: Christoph Hellwig <hch@lst.de>
7 years agoxfs: support removing extents from CoW fork
Darrick J. Wong [Mon, 3 Oct 2016 16:11:35 +0000 (09:11 -0700)]
xfs: support removing extents from CoW fork

Create a helper method to remove extents from the CoW fork without
any of the side effects (rmapbt/bmbt updates) of the regular extent
deletion routine.  We'll eventually use this to clear out the CoW fork
during ioend processing.

Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
7 years agoxfs: allocate delayed extents in CoW fork
Darrick J. Wong [Mon, 3 Oct 2016 16:11:34 +0000 (09:11 -0700)]
xfs: allocate delayed extents in CoW fork

Modify the writepage handler to find and convert pending delalloc
extents to real allocations.  Furthermore, when we're doing non-cow
writes to a part of a file that already has a CoW reservation (the
cowextsz hint that we set up in a subsequent patch facilitates this),
promote the write to copy-on-write so that the entire extent can get
written out as a single extent on disk, thereby reducing post-CoW
fragmentation.

Christoph moved the CoW support code in _map_blocks to a separate helper
function, refactored other functions, and reduced the number of CoW fork
lookups, so I merged those changes here to reduce churn.

Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
Signed-off-by: Christoph Hellwig <hch@lst.de>
7 years agoxfs: support allocating delayed extents in CoW fork
Darrick J. Wong [Mon, 3 Oct 2016 16:11:34 +0000 (09:11 -0700)]
xfs: support allocating delayed extents in CoW fork

Modify xfs_bmap_add_extent_delay_real() so that we can convert delayed
allocation extents in the CoW fork to real allocations, and wire this
up all the way back to xfs_iomap_write_allocate().  In a subsequent
patch, we'll modify the writepage handler to call this.

Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
7 years agoxfs: create delalloc extents in CoW fork
Darrick J. Wong [Mon, 3 Oct 2016 16:11:33 +0000 (09:11 -0700)]
xfs: create delalloc extents in CoW fork

Wire up iomap_begin to detect shared extents and create delayed allocation
extents in the CoW fork:

 1) Check if we already have an extent in the COW fork for the area.
    If so nothing to do, we can move along.
 2) Look up block number for the current extent, and if there is none
    it's not shared move along.
 3) Unshare the current extent as far as we are going to write into it.
    For this we avoid an additional COW fork lookup and use the
    information we set aside in step 1) above.
 4) Goto 1) unless we've covered the whole range.

Last but not least, this updates the xfs_reflink_reserve_cow_range calling
convention to pass a byte offset and length, as that is what both callers
expect anyway.  This patch has been refactored considerably as part of the
iomap transition.

Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
Signed-off-by: Christoph Hellwig <hch@lst.de>
7 years agoxfs: support bmapping delalloc extents in the CoW fork
Darrick J. Wong [Mon, 3 Oct 2016 16:11:32 +0000 (09:11 -0700)]
xfs: support bmapping delalloc extents in the CoW fork

Allow the creation of delayed allocation extents in the CoW fork.  In
a subsequent patch we'll wire up iomap_begin to actually do this via
reflink helper functions.

Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
7 years agoxfs: introduce the CoW fork
Darrick J. Wong [Mon, 3 Oct 2016 16:11:32 +0000 (09:11 -0700)]
xfs: introduce the CoW fork

Introduce a new in-core fork for storing copy-on-write delalloc
reservations and allocated extents that are in the process of being
written out.

Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
7 years agoxfs: don't allow reflinked dir/dev/fifo/socket/pipe files
Darrick J. Wong [Mon, 3 Oct 2016 16:11:31 +0000 (09:11 -0700)]
xfs: don't allow reflinked dir/dev/fifo/socket/pipe files

Only non-rt files can be reflinked, so check that when we load an
inode.  Also, don't leak the attr fork if there's a failure.

Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
7 years agoxfs: add reflink feature flag to geometry
Darrick J. Wong [Mon, 3 Oct 2016 16:11:30 +0000 (09:11 -0700)]
xfs: add reflink feature flag to geometry

Report the reflink feature in the XFS geometry so that xfs_info and
friends know the filesystem has this feature.

Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
7 years agoxfs: define tracepoints for reflink activities
Darrick J. Wong [Mon, 3 Oct 2016 16:11:30 +0000 (09:11 -0700)]
xfs: define tracepoints for reflink activities

Define all the tracepoints we need to inspect the runtime operation
of reflink/dedupe/copy-on-write.

Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
7 years agoxfs: return work remaining at the end of a bunmapi operation
Darrick J. Wong [Mon, 3 Oct 2016 16:11:29 +0000 (09:11 -0700)]
xfs: return work remaining at the end of a bunmapi operation

Return the range of file blocks that bunmapi didn't free.  This hint
is used by CoW and reflink to figure out what part of an extent
actually got freed so that it can set up the appropriate atomic
remapping of just the freed range.

Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
7 years agoxfs: when replaying bmap operations, don't let unlinked inodes get reaped
Darrick J. Wong [Mon, 3 Oct 2016 16:11:29 +0000 (09:11 -0700)]
xfs: when replaying bmap operations, don't let unlinked inodes get reaped

Log recovery will iget an inode to replay BUI items and iput the inode
when it's done.  Unfortunately, if the inode was unlinked, the iput
will see that i_nlink == 0 and decide to truncate & free the inode,
which prevents us from replaying subsequent BUIs.  We can't skip the
BUIs because we have to replay all the redo items to ensure that
atomic operations complete.

Since unlinked inode recovery will reap the inode anyway, we can
safely introduce a new inode flag to indicate that an inode is in this
'unlinked recovery' state and should not be auto-reaped in the
drop_inode path.

Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
7 years agoxfs: implement deferred bmbt map/unmap operations
Darrick J. Wong [Mon, 3 Oct 2016 16:11:28 +0000 (09:11 -0700)]
xfs: implement deferred bmbt map/unmap operations

Implement deferred versions of the inode block map/unmap functions.
These will be used in subsequent patches to make reflink operations
atomic.

Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
7 years agoxfs: pass bmapi flags through to bmap_del_extent
Darrick J. Wong [Mon, 3 Oct 2016 16:11:27 +0000 (09:11 -0700)]
xfs: pass bmapi flags through to bmap_del_extent

Pass BMAPI_ flags from bunmapi into bmap_del_extent and extend
BMAPI_REMAP (which means "don't touch the allocator or the quota
accounting") to apply to bunmapi as well.  This will be used to
implement the unmap operation, which will be used by swapext.

Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
7 years agoxfs: map an inode's offset to an exact physical block
Darrick J. Wong [Mon, 3 Oct 2016 16:11:27 +0000 (09:11 -0700)]
xfs: map an inode's offset to an exact physical block

Teach the bmap routine to know how to map a range of file blocks to a
specific range of physical blocks, instead of simply allocating fresh
blocks.  This enables reflink to map a file to blocks that are already
in use.

Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
7 years agoxfs: log bmap intent items
Darrick J. Wong [Mon, 3 Oct 2016 16:11:26 +0000 (09:11 -0700)]
xfs: log bmap intent items

Provide a mechanism for higher levels to create BUI/BUD items, submit
them to the log, and a stub function to deal with recovered BUI items.
These parts will be connected to the rmapbt in a later patch.

Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
7 years agoxfs: create bmbt update intent log items
Darrick J. Wong [Mon, 3 Oct 2016 16:11:25 +0000 (09:11 -0700)]
xfs: create bmbt update intent log items

Create bmbt update intent/done log items to record redo information in
the log.  Because we roll transactions multiple times for reflink
operations, we also have to track the status of the metadata updates
that will be recorded in the post-roll transactions in case we crash
before committing the final transaction.  This mechanism enables log
recovery to finish what was already started.

Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
7 years agoxfs: introduce reflink utility functions
Darrick J. Wong [Mon, 3 Oct 2016 16:11:25 +0000 (09:11 -0700)]
xfs: introduce reflink utility functions

These functions will be used by the other reflink functions to find
the maximum length of a range of shared blocks.

Signed-off-by: Darrick J. Wong <darrick.wong@oracle.coM>
Reviewed-by: Christoph Hellwig <hch@lst.de>
7 years agoxfs: reserve AG space for the refcount btree root
Darrick J. Wong [Mon, 3 Oct 2016 16:11:24 +0000 (09:11 -0700)]
xfs: reserve AG space for the refcount btree root

Reduce the max AG usable space size so that we always have space for
the refcount btree root.

Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
7 years agoxfs: add refcount btree block detection to log recovery
Darrick J. Wong [Mon, 3 Oct 2016 16:11:23 +0000 (09:11 -0700)]
xfs: add refcount btree block detection to log recovery

Identify refcountbt blocks in the log correctly so that we can
validate them during log recovery.

Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
7 years agoxfs: adjust refcount when unmapping file blocks
Darrick J. Wong [Mon, 3 Oct 2016 16:11:23 +0000 (09:11 -0700)]
xfs: adjust refcount when unmapping file blocks

When we're unmapping blocks from a reflinked file, decrease the
refcount of the affected blocks and free the extents that are no
longer in use.

Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
7 years agoxfs: connect refcount adjust functions to upper layers
Darrick J. Wong [Mon, 3 Oct 2016 16:11:22 +0000 (09:11 -0700)]
xfs: connect refcount adjust functions to upper layers

Plumb in the upper level interface to schedule and finish deferred
refcount operations via the deferred ops mechanism.

Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
7 years agoxfs: adjust refcount of an extent of blocks in refcount btree
Darrick J. Wong [Mon, 3 Oct 2016 16:11:21 +0000 (09:11 -0700)]
xfs: adjust refcount of an extent of blocks in refcount btree

Provide functions to adjust the reference counts for an extent of
physical blocks stored in the refcount btree.

Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
Signed-off-by: Christoph Hellwig <hch@lst.de>
7 years agoxfs: log refcount intent items
Darrick J. Wong [Mon, 3 Oct 2016 16:11:21 +0000 (09:11 -0700)]
xfs: log refcount intent items

Provide a mechanism for higher levels to create CUI/CUD items, submit
them to the log, and a stub function to deal with recovered CUI items.
These parts will be connected to the refcountbt in a later patch.

Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
7 years agoxfs: create refcount update intent log items
Darrick J. Wong [Mon, 3 Oct 2016 16:11:20 +0000 (09:11 -0700)]
xfs: create refcount update intent log items

Create refcount update intent/done log items to record redo
information in the log.  Because we need to roll transactions between
updating the bmbt mapping and updating the reverse mapping, we also
have to track the status of the metadata updates that will be recorded
in the post-roll transactions, just in case we crash before committing
the final transaction.  This mechanism enables log recovery to finish
what was already started.

Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
7 years agoxfs: add refcount btree operations
Darrick J. Wong [Mon, 3 Oct 2016 16:11:19 +0000 (09:11 -0700)]
xfs: add refcount btree operations

Implement the generic btree operations required to manipulate refcount
btree blocks.  The implementation is similar to the bmapbt, though it
will only allocate and free blocks from the AG.

Since the refcount root and level fields are separate from the
existing roots and levels array, they need a separate logging flag.

Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
[hch: fix logging of AGF refcount btree fields]
Signed-off-by: Christoph Hellwig <hch@lst.de>
7 years agoxfs: account for the refcount btree in the alloc/free log reservation
Darrick J. Wong [Mon, 3 Oct 2016 16:11:19 +0000 (09:11 -0700)]
xfs: account for the refcount btree in the alloc/free log reservation

Every time we allocate or free a data extent, we might need to split
the refcount btree.  Reserve some blocks in the transaction to handle
this possibility.  Even though the deferred refcount code can roll a
transaction to avoid overloading the transaction, we can still exceed
the reservation.

Certain pathological workloads (1k blocks, no cowextsize hint, random
directio writes), cause a perfect storm wherein a refcount adjustment
of a large range of blocks causes full tree splits in two separate
extents in two separate refcount tree blocks; allocating new refcount
tree blocks causes rmap btree splits; and all the allocation activity
causes the freespace btrees to split, blowing the reservation.

(Reproduced by generic/167 over NFS atop XFS)

Signed-off-by: Christoph Hellwig <hch@lst.de>
[darrick.wong@oracle.com: add commit message]
Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
7 years agoxfs: add refcount btree support to growfs
Darrick J. Wong [Mon, 3 Oct 2016 16:11:18 +0000 (09:11 -0700)]
xfs: add refcount btree support to growfs

Modify the growfs code to initialize new refcount btree blocks.

Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
7 years agoxfs: define the on-disk refcount btree format
Darrick J. Wong [Mon, 3 Oct 2016 16:11:18 +0000 (09:11 -0700)]
xfs: define the on-disk refcount btree format

Start constructing the refcount btree implementation by establishing
the on-disk format and everything needed to read, write, and
manipulate the refcount btree blocks.

Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Christoph Hellwig <hch@lst.de>
7 years agoxfs: refcount btree add more reserved blocks
Darrick J. Wong [Mon, 3 Oct 2016 16:11:17 +0000 (09:11 -0700)]
xfs: refcount btree add more reserved blocks

Since XFS reserves a small amount of space in each AG as the minimum
free space needed for an operation, save some more space in case we
touch the refcount btree.

Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
7 years agoxfs: introduce refcount btree definitions
Darrick J. Wong [Mon, 3 Oct 2016 16:11:16 +0000 (09:11 -0700)]
xfs: introduce refcount btree definitions

Add new per-AG refcount btree definitions to the per-AG structures.

Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Christoph Hellwig <hch@lst.de>
7 years agoxfs: define tracepoints for refcount btree activities
Darrick J. Wong [Mon, 3 Oct 2016 16:11:15 +0000 (09:11 -0700)]
xfs: define tracepoints for refcount btree activities

Define all the tracepoints we need to inspect the refcount btree
runtime operation.

Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
7 years agoxfs: return an error when an inline directory is too small
Darrick J. Wong [Mon, 3 Oct 2016 16:11:15 +0000 (09:11 -0700)]
xfs: return an error when an inline directory is too small

If the size of an inline directory is so small that it doesn't
even cover the required header size, return an error to userspace
instead of ASSERTing and returning 0 like everything's ok.

Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
Reported-by: Jan Kara <jack@suse.cz>
Reviewed-by: Brian Foster <bfoster@redhat.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
7 years agovfs: add a FALLOC_FL_UNSHARE mode to fallocate to unshare a range of blocks
Darrick J. Wong [Mon, 3 Oct 2016 16:11:14 +0000 (09:11 -0700)]
vfs: add a FALLOC_FL_UNSHARE mode to fallocate to unshare a range of blocks

Add a new fallocate mode flag that explicitly unshares blocks on
filesystems that support such features.  The new flag can only
be used with an allocate-mode fallocate call.

Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
7 years agovfs: support FS_XFLAG_COWEXTSIZE and get/set of CoW extent size hint
Darrick J. Wong [Mon, 3 Oct 2016 16:11:13 +0000 (09:11 -0700)]
vfs: support FS_XFLAG_COWEXTSIZE and get/set of CoW extent size hint

Introduce XFLAGs for the new XFS CoW extent size hint, and actually
plumb the CoW extent size hint into the fsxattr structure.

Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
7 years agoMerge branch 'xfs-4.9-log-recovery-fixes' into for-next
Dave Chinner [Sun, 2 Oct 2016 22:56:28 +0000 (09:56 +1100)]
Merge branch 'xfs-4.9-log-recovery-fixes' into for-next

7 years agoMerge branch 'iomap-4.9-dax' into for-next
Dave Chinner [Sun, 2 Oct 2016 22:53:59 +0000 (09:53 +1100)]
Merge branch 'iomap-4.9-dax' into for-next

7 years agoMerge branch 'xfs-4.9-delalloc-rework' into for-next
Dave Chinner [Sun, 2 Oct 2016 22:52:51 +0000 (09:52 +1100)]
Merge branch 'xfs-4.9-delalloc-rework' into for-next

7 years agoMerge branch 'xfs-4.9-reflink-prep' into for-next
Dave Chinner [Sun, 2 Oct 2016 22:52:31 +0000 (09:52 +1100)]
Merge branch 'xfs-4.9-reflink-prep' into for-next

7 years agoMerge branch 'iomap-4.9-misc-fixes-1' into for-next
Dave Chinner [Sun, 2 Oct 2016 22:52:11 +0000 (09:52 +1100)]
Merge branch 'iomap-4.9-misc-fixes-1' into for-next

7 years agofs: update atime before I/O in generic_file_read_iter
Christoph Hellwig [Sun, 2 Oct 2016 22:48:08 +0000 (09:48 +1100)]
fs: update atime before I/O in generic_file_read_iter

After the call to ->direct_IO the final reference to the file might have
been dropped by aio_complete already, and the call to file_accessed might
cause a use after free.

Instead update the access time before the I/O, similar to how we
update the time stamps before writes.

Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Dave Chinner <dchinner@redhat.com>
Signed-off-by: Dave Chinner <david@fromorbit.com>
7 years agoxfs: update atime before I/O in xfs_file_dio_aio_read
Christoph Hellwig [Sun, 2 Oct 2016 22:47:34 +0000 (09:47 +1100)]
xfs: update atime before I/O in xfs_file_dio_aio_read

After the call to __blkdev_direct_IO the final reference to the file
might have been dropped by aio_complete already, and the call to
file_accessed might cause a use after free.

Instead update the access time before the I/O, similar to how we
update the time stamps before writes.

Signed-off-by: Christoph Hellwig <hch@lst.de>
Reported-and-tested-by: Darrick J. Wong <darrick.wong@oracle.com>
Reviewed-by: Darrick J. Wong <darrick.wong@oracle.com>
Signed-off-by: Dave Chinner <david@fromorbit.com>
7 years agoext2: fix possible integer truncation in ext2_iomap_begin
Christoph Hellwig [Sun, 2 Oct 2016 22:46:04 +0000 (09:46 +1100)]
ext2: fix possible integer truncation in ext2_iomap_begin

For 32-bit architectures we need to cast first_block to u64 before
shifting it left.

Signed-off-by: Christoph Hellwig <hch@lst.de>
Reported-by: Jan Kara <jack@suse.cz>
Reviewed-by: Dave Chinner <dchinner@redhat.com>
Signed-off-by: Dave Chinner <david@fromorbit.com>
7 years agoxfs: log recovery tracepoints to track current lsn and buffer submission
Brian Foster [Sun, 25 Sep 2016 22:34:52 +0000 (08:34 +1000)]
xfs: log recovery tracepoints to track current lsn and buffer submission

Log recovery has particular rules around buffer submission along with
tricky corner cases where independent transactions can share an LSN. As
such, it can be difficult to follow when/why buffers are submitted
during recovery.

Add a couple tracepoints to post the current LSN of a record when a new
record is being processed and when a buffer is being skipped due to LSN
ordering. Also, update the recover item class to include the LSN of the
current transaction for the item being processed.

Signed-off-by: Brian Foster <bfoster@redhat.com>
Reviewed-by: Dave Chinner <dchinner@redhat.com>
Signed-off-by: Dave Chinner <david@fromorbit.com>
7 years agoxfs: update metadata LSN in buffers during log recovery
Brian Foster [Sun, 25 Sep 2016 22:34:27 +0000 (08:34 +1000)]
xfs: update metadata LSN in buffers during log recovery

Log recovery is currently broken for v5 superblocks in that it never
updates the metadata LSN of buffers written out during recovery. The
metadata LSN is recorded in various bits of metadata to provide recovery
ordering criteria that prevents transient corruption states reported by
buffer write verifiers. Without such ordering logic, buffer updates can
be replayed out of order and lead to false positive transient corruption
states. This is generally not a corruption vector on its own, but
corruption detection shuts down the filesystem and ultimately prevents a
mount if it occurs during log recovery. This requires an xfs_repair run
that clears the log and potentially loses filesystem updates.

This problem is avoided in most cases as metadata writes during normal
filesystem operation update the metadata LSN appropriately. The problem
with log recovery not updating metadata LSNs manifests if the system
happens to crash shortly after log recovery itself. In this scenario, it
is possible for log recovery to complete all metadata I/O such that the
filesystem is consistent. If a crash occurs after that point but before
the log tail is pushed forward by subsequent operations, however, the
next mount performs the same log recovery over again. If a buffer is
updated multiple times in the dirty range of the log, an earlier update
in the log might not be valid based on the current state of the
associated buffer after all of the updates in the log had been replayed
(before the previous crash). If a verifier happens to detect such a
problem, the filesystem claims corruption and immediately shuts down.

This commonly manifests in practice as directory block verifier failures
such as the following, likely due to directory verifiers being
particularly detailed in their checks as compared to most others:

  ...
  Mounting V5 Filesystem
  XFS (dm-0): Starting recovery (logdev: internal)
  XFS (dm-0): Internal error XFS_WANT_CORRUPTED_RETURN at line ... of \
    file fs/xfs/libxfs/xfs_dir2_data.c.  Caller xfs_dir3_data_verify ...
  ...

Update log recovery to update the metadata LSN of recovered buffers.
Since metadata LSNs are already updated by write verifer functions via
attached log items, attach a dummy log item to the buffer during
validation and explicitly set the LSN of the current transaction. This
ensures that the metadata LSN of a buffer is updated based on whether
the recovery I/O actually completes, and if so, that subsequent recovery
attempts identify that the buffer is already up to date with respect to
the current transaction.

Signed-off-by: Brian Foster <bfoster@redhat.com>
Reviewed-by: Dave Chinner <dchinner@redhat.com>
Signed-off-by: Dave Chinner <david@fromorbit.com>
7 years agoxfs: don't warn on buffers not being recovered due to LSN
Brian Foster [Sun, 25 Sep 2016 22:32:50 +0000 (08:32 +1000)]
xfs: don't warn on buffers not being recovered due to LSN

The log recovery buffer validation function is invoked in cases where a
buffer update may be skipped due to LSN ordering. If the validation
function happens to come across directory conversion situations (e.g., a
dir3 block to data conversion), it may warn about seeing a buffer log
format of one type and a buffer with a magic number of another.

This warning is not valid as the buffer update is ultimately skipped.
This is indicated by a current_lsn of NULLCOMMITLSN provided by the
caller. As such, update xlog_recover_validate_buf_type() to only warn in
such cases when a buffer update is expected.

Signed-off-by: Brian Foster <bfoster@redhat.com>
Reviewed-by: Dave Chinner <dchinner@redhat.com>
Signed-off-by: Dave Chinner <david@fromorbit.com>
7 years agoxfs: pass current lsn to log recovery buffer validation
Brian Foster [Sun, 25 Sep 2016 22:32:07 +0000 (08:32 +1000)]
xfs: pass current lsn to log recovery buffer validation

The current LSN must be available to the buffer validation function to
provide the ability to update the metadata LSN of the buffer. Pass the
current_lsn value down to xlog_recover_validate_buf_type() in
preparation.

Signed-off-by: Brian Foster <bfoster@redhat.com>
Reviewed-by: Dave Chinner <dchinner@redhat.com>
Signed-off-by: Dave Chinner <david@fromorbit.com>
7 years agoxfs: rework log recovery to submit buffers on LSN boundaries
Brian Foster [Sun, 25 Sep 2016 22:22:16 +0000 (08:22 +1000)]
xfs: rework log recovery to submit buffers on LSN boundaries

The fix to log recovery to update the metadata LSN in recovered buffers
introduces the requirement that a buffer is submitted only once per
current LSN. Log recovery currently submits buffers on transaction
boundaries. This is not sufficient as the abstraction between log
records and transactions allows for various scenarios where multiple
transactions can share the same current LSN. If independent transactions
share an LSN and both modify the same buffer, log recovery can
incorrectly skip updates and leave the filesystem in an inconsisent
state.

In preparation for proper metadata LSN updates during log recovery,
update log recovery to submit buffers for write on LSN change boundaries
rather than transaction boundaries. Explicitly track the current LSN in
a new struct xlog field to handle the various corner cases of when the
current LSN may or may not change.

Signed-off-by: Brian Foster <bfoster@redhat.com>
Reviewed-by: Dave Chinner <dchinner@redhat.com>
Signed-off-by: Dave Chinner <david@fromorbit.com>
7 years agoxfs: quiesce the filesystem after recovery on readonly mount
Dave Chinner [Sun, 25 Sep 2016 22:21:44 +0000 (08:21 +1000)]
xfs: quiesce the filesystem after recovery on readonly mount

Recently we've had a number of reports where log recovery on a v5
filesystem has reported corruptions that looked to be caused by
recovery being re-run over the top of an already-recovered
metadata. This has uncovered a bug in recovery (fixed elsewhere)
but the vector that caused this was largely unknown.

A kdump test started tripping over this problem - the system
would be crashed, the kdump kernel and environment would boot and
dump the kernel core image, and then the system would reboot. After
reboot, the root filesystem was triggering log recovery and
corruptions were being detected. The metadumps indicated the above
log recovery issue.

What is happening is that the kdump kernel and environment is
mounting the root device read-only to find the binaries needed to do
it's work. The result of this is that it is running log recovery.
However, because there were unlinked files and EFIs to be processed
by recovery, the completion of phase 1 of log recovery could not
mark the log clean. And because it's a read-only mount, the unmount
process does not write records to the log to mark it clean, either.
Hence on the next mount of the filesystem, log recovery was run
again across all the metadata that had already been recovered and
this is what triggered corruption warnings.

To avoid this problem, we need to ensure that a read-only mount
always updates the log when it completes the second phase of
recovery. We already handle this sort of issue with rw->ro remount
transitions, so the solution is as simple as quiescing the
filesystem at the appropriate time during the mount process. This
results in the log being marked clean so the mount behaviour
recorded in the logs on repeated RO mounts will change (i.e. log
recovery will no longer be run on every mount until a RW mount is
done). This is a user visible change in behaviour, but it is
harmless.

Signed-off-by: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Eric Sandeen <sandeen@redhat.com>
Signed-off-by: Dave Chinner <david@fromorbit.com>
7 years agoxfs: remote attribute blocks aren't really userdata
Dave Chinner [Sun, 25 Sep 2016 22:21:28 +0000 (08:21 +1000)]
xfs: remote attribute blocks aren't really userdata

When adding a new remote attribute, we write the attribute to the
new extent before the allocation transaction is committed. This
means we cannot reuse busy extents as that violates crash
consistency semantics. Hence we currently treat remote attribute
extent allocation like userdata because it has the same overwrite
ordering constraints as userdata.

Unfortunately, this also allows the allocator to incorrectly apply
extent size hints to the remote attribute extent allocation. This
results in interesting failures, such as transaction block
reservation overruns and in-memory inode attribute fork corruption.

To fix this, we need to separate the busy extent reuse configuration
from the userdata configuration. This changes the definition of
XFS_BMAPI_METADATA slightly - it now means that allocation is
metadata and reuse of busy extents is acceptible due to the metadata
ordering semantics of the journal. If this flag is not set, it
means the allocation is that has unordered data writeback, and hence
busy extent reuse is not allowed. It no longer implies the
allocation is for user data, just that the data write will not be
strictly ordered. This matches the semantics for both user data
and remote attribute block allocation.

As such, This patch changes the "userdata" field to a "datatype"
field, and adds a "no busy reuse" flag to the field.
When we detect an unordered data extent allocation, we immediately set
the no reuse flag. We then set the "user data" flags based on the
inode fork we are allocating the extent to. Hence we only set
userdata flags on data fork allocations now and consider attribute
fork remote extents to be an unordered metadata extent.

The result is that remote attribute extents now have the expected
allocation semantics, and the data fork allocation behaviour is
completely unchanged.

It should be noted that there may be other ways to fix this (e.g.
use ordered metadata buffers for the remote attribute extent data
write) but they are more invasive and difficult to validate both
from a design and implementation POV. Hence this patch takes the
simple, obvious route to fixing the problem...

Reported-and-tested-by: Ross Zwisler <ross.zwisler@linux.intel.com>
Signed-off-by: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Dave Chinner <david@fromorbit.com>
7 years agoext2: use iomap to implement DAX
Christoph Hellwig [Mon, 19 Sep 2016 01:30:29 +0000 (11:30 +1000)]
ext2: use iomap to implement DAX

Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Ross Zwisler <ross.zwisler@linux.intel.com>
Signed-off-by: Dave Chinner <david@fromorbit.com>
7 years agoext2: stop passing buffer_head to ext2_get_blocks
Christoph Hellwig [Mon, 19 Sep 2016 01:28:39 +0000 (11:28 +1000)]
ext2: stop passing buffer_head to ext2_get_blocks

Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Ross Zwisler <ross.zwisler@linux.intel.com>
Signed-off-by: Dave Chinner <david@fromorbit.com>
7 years agoxfs: use iomap to implement DAX
Christoph Hellwig [Mon, 19 Sep 2016 01:28:38 +0000 (11:28 +1000)]
xfs: use iomap to implement DAX

Another users of buffer_heads bytes the dust.

Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Ross Zwisler <ross.zwisler@linux.intel.com>
Signed-off-by: Dave Chinner <david@fromorbit.com>
7 years agoxfs: refactor xfs_setfilesize
Christoph Hellwig [Mon, 19 Sep 2016 01:26:41 +0000 (11:26 +1000)]
xfs: refactor xfs_setfilesize

Rename the current function to __xfs_setfilesize and add a non-static
wrapper that also takes care of creating the transaction.  This new
helper will be used by the new iomap-based DAX path.

Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Dave Chinner <dchinner@redhat.com>
Signed-off-by: Dave Chinner <david@fromorbit.com>
7 years agoxfs: take the ilock shared if possible in xfs_file_iomap_begin
Christoph Hellwig [Mon, 19 Sep 2016 01:26:39 +0000 (11:26 +1000)]
xfs: take the ilock shared if possible in xfs_file_iomap_begin

We always just read the extent first, and will later lock exlusively
after first dropping the lock in case we actually allocate blocks.

Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Dave Chinner <dchinner@redhat.com>
Signed-off-by: Dave Chinner <david@fromorbit.com>
7 years agoxfs: fix locking for DAX writes
Christoph Hellwig [Mon, 19 Sep 2016 01:24:50 +0000 (11:24 +1000)]
xfs: fix locking for DAX writes

So far DAX writes inherited the locking from direct I/O writes, but
the direct I/O model of using shared locks for writes is actually
wrong for DAX.  For direct I/O we're out of any standards and don't
have to provide the Posix required exclusion between writers, but
for DAX which gets transparently enable on applications without any
knowledge of it we can't simply drop the requirement.  Even worse
this only happens for aligned writes and thus doesn't show up for
many typical use cases.

Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Dave Chinner <dchinner@redhat.com>
Signed-off-by: Dave Chinner <david@fromorbit.com>
7 years agodax: provide an iomap based fault handler
Christoph Hellwig [Mon, 19 Sep 2016 01:24:50 +0000 (11:24 +1000)]
dax: provide an iomap based fault handler

Very similar to the existing dax_fault function, but instead of using
the get_block callback we rely on the iomap_ops vector from iomap.c.
That also avoids having to do two calls into the file system for write
faults.

Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Ross Zwisler <ross.zwisler@linux.intel.com>
Signed-off-by: Dave Chinner <david@fromorbit.com>
7 years agodax: provide an iomap based dax read/write path
Christoph Hellwig [Mon, 19 Sep 2016 01:24:49 +0000 (11:24 +1000)]
dax: provide an iomap based dax read/write path

This is a much simpler implementation of the DAX read/write path
that makes use of the iomap infrastructure.  It does not try to
mirror the direct I/O calling conventions and thus doesn't have to
deal with i_dio_count or the end_io handler, but instead leaves
locking and filesystem-specific I/O completion to the caller.

Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Ross Zwisler <ross.zwisler@linux.intel.com>
Signed-off-by: Dave Chinner <david@fromorbit.com>
7 years agodax: don't pass buffer_head to copy_user_dax
Christoph Hellwig [Mon, 19 Sep 2016 01:24:49 +0000 (11:24 +1000)]
dax: don't pass buffer_head to copy_user_dax

This way we can use this helper for the iomap based DAX implementation
as well.

Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Ross Zwisler <ross.zwisler@linux.intel.com>
Signed-off-by: Dave Chinner <david@fromorbit.com>
7 years agodax: don't pass buffer_head to dax_insert_mapping
Christoph Hellwig [Mon, 19 Sep 2016 01:24:49 +0000 (11:24 +1000)]
dax: don't pass buffer_head to dax_insert_mapping

This way we can use this helper for the iomap based DAX implementation
as well.

Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Ross Zwisler <ross.zwisler@linux.intel.com>
Signed-off-by: Dave Chinner <david@fromorbit.com>
7 years agoiomap: expose iomap_apply outside iomap.c
Christoph Hellwig [Mon, 19 Sep 2016 01:24:49 +0000 (11:24 +1000)]
iomap: expose iomap_apply outside iomap.c

This allows the DAX code to use it.

Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Ross Zwisler <ross.zwisler@linux.intel.com>
Signed-off-by: Dave Chinner <david@fromorbit.com>
7 years agoiomap: add IOMAP_F_NEW flag
Christoph Hellwig [Mon, 19 Sep 2016 01:24:37 +0000 (11:24 +1000)]
iomap: add IOMAP_F_NEW flag

Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Ross Zwisler <ross.zwisler@linux.intel.com>
Signed-off-by: Dave Chinner <david@fromorbit.com>
7 years agoxfs: rewrite and optimize the delalloc write path
Christoph Hellwig [Mon, 19 Sep 2016 01:10:21 +0000 (11:10 +1000)]
xfs: rewrite and optimize the delalloc write path

Currently xfs_iomap_write_delay does up to lookups in the inode
extent tree, which is rather costly especially with the new iomap
based write path and small write sizes.

But it turns out that the low-level xfs_bmap_search_extents gives us
all the information we need in the regular delalloc buffered write
path:

 - it will return us an extent covering the block we are looking up
   if it exists.  In that case we can simply return that extent to
   the caller and are done
 - it will tell us if we are beyoned the last current allocated
   block with an eof return parameter.  In that case we can create a
   delalloc reservation and use the also returned information about
   the last extent in the file as the hint to size our delalloc
   reservation.
 - it can tell us that we are writing into a hole, but that there is
   an extent beyoned this hole.  In this case we can create a
   delalloc reservation that covers the requested size (possible
   capped to the next existing allocation).

All that can be done in one single routine instead of bouncing up
and down a few layers.  This reduced the CPU overhead of the block
mapping routines and also simplified the code a lot.

Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Dave Chinner <dchinner@redhat.com>
Signed-off-by: Dave Chinner <david@fromorbit.com>
7 years agoxfs: make xfs_inode_set_eofblocks_tag cheaper for the common case
Christoph Hellwig [Mon, 19 Sep 2016 01:09:48 +0000 (11:09 +1000)]
xfs: make xfs_inode_set_eofblocks_tag cheaper for the common case

For long growing file writes we will usually already have the
eofblocks tag set when adding more speculative preallocations.  Add
a flag in the inode to allow us to skip the the fairly expensive
AG-wide spinlocks and multiple radix tree operations in that case.

Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Dave Chinner <dchinner@redhat.com>
Signed-off-by: Dave Chinner <david@fromorbit.com>
7 years agoxfs: factor our a helper to calculate the EOF alignment
Christoph Hellwig [Mon, 19 Sep 2016 01:09:28 +0000 (11:09 +1000)]
xfs: factor our a helper to calculate the EOF alignment

And drop the pointless mp argument to xfs_iomap_eof_align_last_fsb,
while we're at it.

Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Dave Chinner <dchinner@redhat.com>
Signed-off-by: Dave Chinner <david@fromorbit.com>
7 years agoxfs: move xfs_bmbt_to_iomap up
Christoph Hellwig [Mon, 19 Sep 2016 01:09:12 +0000 (11:09 +1000)]
xfs: move xfs_bmbt_to_iomap up

We'll need it earlier in the file soon, so the unchanged function to
the top of xfs_iomap.c

Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Dave Chinner <dchinner@redhat.com>
Signed-off-by: Dave Chinner <david@fromorbit.com>
7 years agoxfs: set up per-AG free space reservations
Darrick J. Wong [Mon, 19 Sep 2016 00:30:52 +0000 (10:30 +1000)]
xfs: set up per-AG free space reservations

One unfortunate quirk of the reference count and reverse mapping
btrees -- they can expand in size when blocks are written to *other*
allocation groups if, say, one large extent becomes a lot of tiny
extents.  Since we don't want to start throwing errors in the middle
of CoWing, we need to reserve some blocks to handle future expansion.
The transaction block reservation counters aren't sufficient here
because we have to have a reserve of blocks in every AG, not just
somewhere in the filesystem.

Therefore, create two per-AG block reservation pools.  One feeds the
AGFL so that rmapbt expansion always succeeds, and the other feeds all
other metadata so that refcountbt expansion never fails.

Use the count of how many reserved blocks we need to have on hand to
create a virtual reservation in the AG.  Through selective clamping of
the maximum length of allocation requests and of the length of the
longest free extent, we can make it look like there's less free space
in the AG unless the reservation owner is asking for blocks.

In other words, play some accounting tricks in-core to make sure that
we always have blocks available.  On the plus side, there's nothing to
clean up if we crash, which is contrast to the strategy that the rough
draft used (actually removing extents from the freespace btrees).

Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
Reviewed-by: Dave Chinner <dchinner@redhat.com>
Signed-off-by: Dave Chinner <david@fromorbit.com>
7 years agoxfs: defer should allow ->finish_item to request a new transaction
Darrick J. Wong [Mon, 19 Sep 2016 00:26:25 +0000 (10:26 +1000)]
xfs: defer should allow ->finish_item to request a new transaction

When xfs_defer_finish calls ->finish_item, it's possible that
(refcount) won't be able to finish all the work in a single
transaction.  When this happens, the ->finish_item handler should
shorten the log done item's list count, update the work item to
reflect where work should continue, and return -EAGAIN so that
defer_finish knows to retain the pending item on the pending list,
roll the transaction, and restart processing where we left off.

Plumb in the code and document how this mechanism is supposed to work.

Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
7 years agoxfs: count the blocks in a btree
Darrick J. Wong [Mon, 19 Sep 2016 00:25:20 +0000 (10:25 +1000)]
xfs: count the blocks in a btree

Provide a helper method to count the number of blocks in a short form
btree.  The refcount and rmap btrees need to know the number of blocks
already in use to set up their per-AG block reservations during mount.

Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Dave Chinner <david@fromorbit.com>
7 years agoxfs: create a standard btree size calculator code
Darrick J. Wong [Mon, 19 Sep 2016 00:25:03 +0000 (10:25 +1000)]
xfs: create a standard btree size calculator code

Create a helper to generate AG btree height calculator functions.
This will be used (much) later when we get to the refcount btree.

Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Dave Chinner <david@fromorbit.com>
7 years agoxfs: remove xfs_btree_bigkey
Darrick J. Wong [Mon, 19 Sep 2016 00:24:36 +0000 (10:24 +1000)]
xfs: remove xfs_btree_bigkey

Remove the xfs_btree_bigkey mess and simply make xfs_btree_key big enough
to hold both keys in-core.

Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Dave Chinner <david@fromorbit.com>
7 years agoxfs: convert RUI log formats to use variable length arrays
Darrick J. Wong [Mon, 19 Sep 2016 00:24:27 +0000 (10:24 +1000)]
xfs: convert RUI log formats to use variable length arrays

Use variable length array declarations for RUI log items,
and replace the open coded sizeof formulae with a single function.

Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
Reviewed-by: Dave Chinner <dchinner@redhat.com>
Signed-off-by: Dave Chinner <david@fromorbit.com>
7 years agoiomap: add a flag to report shared extents
Darrick J. Wong [Mon, 19 Sep 2016 00:13:02 +0000 (10:13 +1000)]
iomap: add a flag to report shared extents

Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Dave Chinner <david@fromorbit.com>
7 years agofs: add iomap_file_dirty
Christoph Hellwig [Mon, 19 Sep 2016 00:12:45 +0000 (10:12 +1000)]
fs: add iomap_file_dirty

Originally-From: Christoph Hellwig <hch@lst.de>

This function uses the iomap infrastructure to re-write all pages
in a given range.  This is useful for doing a copy-up of COW ranges,
and might be useful for scrubbing in the future.

Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Dave Chinner <dchinner@redhat.com>
Signed-off-by: Dave Chinner <david@fromorbit.com>
7 years agoxfs: Document error handlers behavior
Carlos Maiolino [Sun, 18 Sep 2016 23:38:25 +0000 (09:38 +1000)]
xfs: Document error handlers behavior

Document the implementation of error handlers into sysfs.

[dchinner: Added lots more detail.]

Signed-off-by: Carlos Maiolino <cmaiolino@redhat.com>
Signed-off-by: Dave Chinner <dchinner@redhat.com>
7 years agoxfs: normalize "infinite" retries in error configs
Eric Sandeen [Tue, 13 Sep 2016 21:51:30 +0000 (07:51 +1000)]
xfs: normalize "infinite" retries in error configs

As it stands today, the "fail immediately" vs. "retry forever"
values for max_retries and retry_timeout_seconds in the xfs metadata
error configurations are not consistent.

A retry_timeout_seconds of 0 means "retry forever," but a
max_retries of 0 means "fail immediately."

retry_timeout_seconds < 0 is disallowed, while max_retries == -1
means "retry forever."

Make this consistent across the error configs, such that a value of
0 means "fail immediately" (i.e. wait 0 seconds, or retry 0 times),
and a value of -1 always means "retry forever."

This makes retry_timeout a signed long to accommodate the -1, even
though it stores jiffies.  Given our limit of a 1 day maximum
timeout, this should be sufficient even at much higher HZ values
than we have available today.

Signed-off-by: Eric Sandeen <sandeen@redhat.com>
Reviewed-by: Dave Chinner <dchinner@redhat.com>
Signed-off-by: Dave Chinner <david@fromorbit.com>
7 years agoxfs: fix signed integer overflow
Xie XiuQi [Tue, 13 Sep 2016 21:41:16 +0000 (07:41 +1000)]
xfs: fix signed integer overflow

Use 1U for unsigned int to avoid a overflow warning from UBSAN.

[   31.910858] UBSAN: Undefined behaviour in fs/xfs/xfs_buf_item.c:889:25
[   31.911252] signed integer overflow:
[   31.911478] -2147483648 - 1 cannot be represented in type 'int'
[   31.911846] CPU: 1 PID: 1011 Comm: tuned Tainted: G    B          ---- -------   3.10.0-327.28.3.el7.x86_64 #1
[   31.911857] Hardware name: VMware, Inc. VMware Virtual Platform/440BX Desktop Reference Platform, BIOS 6.00 01/07/2011
[   31.911866]  1ffff1004069cd3b 0000000076bec3fd ffff8802034e69a0 ffffffff81ee3140
[   31.911883]  ffff8802034e69b8 ffffffff81ee31fd ffffffffa0ad79e0 ffff8802034e6b20
[   31.911898]  ffffffff81ee46e2 0000002d515470c0 0000000000000001 0000000041b58ab3
[   31.911913] Call Trace:
[   31.911932]  [<ffffffff81ee3140>] dump_stack+0x1e/0x20
[   31.911947]  [<ffffffff81ee31fd>] ubsan_epilogue+0x12/0x55
[   31.911964]  [<ffffffff81ee46e2>] handle_overflow+0x1ba/0x215
[   31.912083]  [<ffffffff81ee4798>] __ubsan_handle_sub_overflow+0x2a/0x31
[   31.912204]  [<ffffffffa08676fb>] xfs_buf_item_log+0x34b/0x3f0 [xfs]
[   31.912314]  [<ffffffffa0880490>] xfs_trans_log_buf+0x120/0x260 [xfs]
[   31.912402]  [<ffffffffa079a890>] xfs_btree_log_recs+0x80/0xc0 [xfs]
[   31.912490]  [<ffffffffa07a29f8>] xfs_btree_delrec+0x11a8/0x2d50 [xfs]
[   31.913589]  [<ffffffffa07a86f9>] xfs_btree_delete+0xc9/0x260 [xfs]
[   31.913762]  [<ffffffffa075b5cf>] xfs_free_ag_extent+0x63f/0xe20 [xfs]
[   31.914339]  [<ffffffffa075ec0f>] xfs_free_extent+0x2af/0x3e0 [xfs]
[   31.914641]  [<ffffffffa0801b2b>] xfs_bmap_finish+0x32b/0x4b0 [xfs]
[   31.914841]  [<ffffffffa083c2e7>] xfs_itruncate_extents+0x3b7/0x740 [xfs]
[   31.915216]  [<ffffffffa08342fa>] xfs_setattr_size+0x60a/0x860 [xfs]
[   31.915471]  [<ffffffffa08345ea>] xfs_vn_setattr+0x9a/0xe0 [xfs]
[   31.915590]  [<ffffffff8149ad38>] notify_change+0x5c8/0x8a0
[   31.915607]  [<ffffffff81450f22>] do_truncate+0x122/0x1d0
[   31.915640]  [<ffffffff8147beee>] do_last+0x15de/0x2c80
[   31.915707]  [<ffffffff8147d777>] path_openat+0x1e7/0xcc0
[   31.915802]  [<ffffffff81480824>] do_filp_open+0xa4/0x160
[   31.915848]  [<ffffffff81453127>] do_sys_open+0x1b7/0x3f0
[   31.915879]  [<ffffffff81453392>] SyS_open+0x32/0x40
[   31.915897]  [<ffffffff81f08989>] system_call_fastpath+0x16/0x1b

[  240.086809] UBSAN: Undefined behaviour in fs/xfs/xfs_buf_item.c:866:34
[  240.086820] signed integer overflow:
[  240.086830] -2147483648 - 1 cannot be represented in type 'int'
[  240.086846] CPU: 1 PID: 12969 Comm: rm Tainted: G    B          ---- -------   3.10.0-327.28.3.el7.x86_64 #1
[  240.086857] Hardware name: VMware, Inc. VMware Virtual Platform/440BX Desktop Reference Platform, BIOS 6.00 01/07/2011
[  240.086868]  1ffff10040491def 00000000e2ea59c1 ffff88020248ef40 ffffffff81ee3140
[  240.086885]  ffff88020248ef58 ffffffff81ee31fd ffffffffa0ad79e0 ffff88020248f0c0
[  240.086901]  ffffffff81ee46e2 0000002d02488000 0000000000000001 0000000041b58ab3
[  240.086915] Call Trace:
[  240.086938]  [<ffffffff81ee3140>] dump_stack+0x1e/0x20
[  240.086953]  [<ffffffff81ee31fd>] ubsan_epilogue+0x12/0x55
[  240.086971]  [<ffffffff81ee46e2>] handle_overflow+0x1ba/0x215
...

Signed-off-by: Xie XiuQi <xiexiuqi@huawei.com>
Reviewed-by: Dave Chinner <dchinner@redhat.com>
Signed-off-by: Dave Chinner <david@fromorbit.com>
7 years agoMake __xfs_xattr_put_listen preperly report errors.
Artem Savkov [Tue, 13 Sep 2016 21:40:35 +0000 (07:40 +1000)]
Make __xfs_xattr_put_listen preperly report errors.

Commit 2a6fba6 "xfs: only return -errno or success from attr ->put_listent"
changes the returnvalue of __xfs_xattr_put_listen to 0 in case when there is
insufficient space in the buffer assuming that setting context->count to -1
would be enough, but all of the ->put_listent callers only check seen_enough.
This results in a failed assertion:
XFS: Assertion failed: context->count >= 0, file: fs/xfs/xfs_xattr.c, line: 175
in insufficient buffer size case.

This is only reproducible with at least 2 xattrs and only when the buffer
gets depleted before the last one.

Furthermore if buffersize is such that it is enough to hold the last xattr's
name, but not enough to hold the sum of preceeding xattr names listxattr won't
fail with ERANGE, but will suceed returning last xattr's name without the
first character. The first character end's up overwriting data stored at
(context->alist - 1).

Signed-off-by: Artem Savkov <asavkov@redhat.com>
Reviewed-by: Dave Chinner <dchinner@redhat.com>
Signed-off-by: Dave Chinner <david@fromorbit.com>
7 years agoxfs: change mailing list address
Dave Chinner [Tue, 13 Sep 2016 21:40:21 +0000 (07:40 +1000)]
xfs: change mailing list address

oss.sgi.com is going away, move contact details over to vger.

cc: <stable@vger.kernel.org>
Signed-off-by: Dave Chinner <david@fromorbit.com>
7 years agoxfs: undo block reservation correctly in xfs_trans_reserve()
Eryu Guan [Tue, 13 Sep 2016 21:39:07 +0000 (07:39 +1000)]
xfs: undo block reservation correctly in xfs_trans_reserve()

"blocks" should be added back to fdblocks at undo time, not taken
away, i.e. the minus sign should not be used.

This is a regression introduced by commit 0d485ada404b ("xfs: use
generic percpu counters for free block counter"). And it's found by
code inspection, I didn't it in real world, so there's no
reproducer.

Signed-off-by: Eryu Guan <eguan@redhat.com>
Reviewed-by: Dave Chinner <dchinner@redhat.com>
Signed-off-by: Dave Chinner <david@fromorbit.com>
7 years agoxfs: track log done items directly in the deferred pending work item
Darrick J. Wong [Tue, 30 Aug 2016 03:51:39 +0000 (13:51 +1000)]
xfs: track log done items directly in the deferred pending work item

Christoph reports slab corruption when a deferred refcount update
aborts during _defer_finish().  The cause of this was broken log item
state tracking in xfs_defer_pending -- upon an abort,
_defer_trans_abort() will call abort_intent on all intent items,
including the ones that have already had a done item attached.

This is incorrect because each intent item has 2 refcount: the first
is released when the intent item is committed to the log; and the
second is released when the _done_ item is committed to the log, or
by the intent creator if there is no done item.  In other words, once
we log the done item, responsibility for releasing the intent item's
second refcount is transferred to the done item and /must not/ be
performed by anything else.

The dfp_committed flag should have been tracking whether or not we had
a done item so that _defer_trans_abort could decide if it needs to
abort the intent item, but due to a thinko this was not the case.  Rip
it out and track the done item directly so that we do the right thing
w.r.t. intent item freeing.

Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
Reported-by: Christoph Hellwig <hch@infradead.org>
Reviewed-by: Dave Chinner <dchinner@redhat.com>
Signed-off-by: Dave Chinner <david@fromorbit.com>
7 years agoiomap: don't set FIEMAP_EXTENT_MERGED for extent based filesystems
Christoph Hellwig [Mon, 29 Aug 2016 01:33:58 +0000 (11:33 +1000)]
iomap: don't set FIEMAP_EXTENT_MERGED for extent based filesystems

Filesystems like XFS that use extents should not set the
FIEMAP_EXTENT_MERGED flag in the fiemap extent structures.  To allow
for both behaviors for the upcoming gfs2 usage split the iomap
type field into type and flags, and only set FIEMAP_EXTENT_MERGED if
the IOMAP_F_MERGED flag is set.  The flags field will also come in
handy for future features such as shared extents on reflink-enabled
file systems.

Reported-by: Andreas Gruenbacher <agruenba@redhat.com>
Signed-off-by: Christoph Hellwig <hch@lst.de>
Acked-by: Darrick J. Wong <darrick.wong@oracle.com>
Signed-off-by: Dave Chinner <david@fromorbit.com>
7 years agoxfs: prevent dropping ioend completions during buftarg wait
Brian Foster [Fri, 26 Aug 2016 06:01:59 +0000 (16:01 +1000)]
xfs: prevent dropping ioend completions during buftarg wait

xfs_wait_buftarg() waits for all pending I/O, drains the ioend
completion workqueue and walks the LRU until all buffers in the cache
have been released. This is traditionally an unmount operation` but the
mechanism is also reused during filesystem freeze.

xfs_wait_buftarg() invokes drain_workqueue() as part of the quiesce,
which is intended more for a shutdown sequence in that it indicates to
the queue that new operations are not expected once the drain has begun.
New work jobs after this point result in a WARN_ON_ONCE() and are
otherwise dropped.

With filesystem freeze, however, read operations are allowed and can
proceed during or after the workqueue drain. If such a read occurs
during the drain sequence, the workqueue infrastructure complains about
the queued ioend completion work item and drops it on the floor. As a
result, the buffer remains on the LRU and the freeze never completes.

Despite the fact that the overall buffer cache cleanup is not necessary
during freeze, fix up this operation such that it is safe to invoke
during non-unmount quiesce operations. Replace the drain_workqueue()
call with flush_workqueue(), which runs a similar serialization on
pending workqueue jobs without causing new jobs to be dropped. This is
safe for unmount as unmount independently locks out new operations by
the time xfs_wait_buftarg() is invoked.

cc: <stable@vger.kernel.org>
Signed-off-by: Brian Foster <bfoster@redhat.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Dave Chinner <david@fromorbit.com>
7 years agoxfs: fix superblock inprogress check
Dave Chinner [Fri, 26 Aug 2016 06:01:30 +0000 (16:01 +1000)]
xfs: fix superblock inprogress check

From inspection, the superblock sb_inprogress check is done in the
verifier and triggered only for the primary superblock via a
"bp->b_bn == XFS_SB_DADDR" check.

Unfortunately, the primary superblock is an uncached buffer, and
hence it is configured by xfs_buf_read_uncached() with:

bp->b_bn = XFS_BUF_DADDR_NULL;  /* always null for uncached buffers */

And so this check never triggers. Fix it.

cc: <stable@vger.kernel.org>
Signed-off-by: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Brian Foster <bfoster@redhat.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Dave Chinner <david@fromorbit.com>
7 years agoxfs: simple btree query range should look right if LE lookup fails
Darrick J. Wong [Fri, 26 Aug 2016 06:00:10 +0000 (16:00 +1000)]
xfs: simple btree query range should look right if LE lookup fails

If the initial LOOKUP_LE in the simple query range fails to find
anything, we should attempt to increment the btree cursor to see
if there actually /are/ records for what we're trying to find.
Without this patch, a bnobt range query of (0, $agsize) returns
no results because the leftmost record never has a startblock
of zero.

Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Dave Chinner <david@fromorbit.com>