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>
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>
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>
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>
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>
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>
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>
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>
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>
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>
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>
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>
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>
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>
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>
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>
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>
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>
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>
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>
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>
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>
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>
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>
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>
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>
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>
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>
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>
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>
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>
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>
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>
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>
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>
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>
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>
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>
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>
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>
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>
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>
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>
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>
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>
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>
