Expose all metadata structure buffer verifier functions via buf_ops.
These will be used by the online scrub mechanism to look for problems
with buffers that are already sitting around in memory.
Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
Reviewed-by: Dave Chinner <dchinner@redhat.com>
If the xattr leaf block looks corrupt, return -EFSCORRUPTED to userspace
instead of ASSERTing on debug kernels or running off the end of the
buffer on regular kernels.
Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
Reviewed-by: Dave Chinner <dchinner@redhat.com>
Replace the current haphazard dir2 shortform verifier callsites with a
centralized verifier function that can be called either with the default
verifier functions or with a custom set. This helps us strengthen
integrity checking while providing us with flexibility for repair tools.
xfs_repair wants this to be able to supply its own verifier functions
when trying to fix possibly corrupt metadata.
Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
Reviewed-by: Dave Chinner <dchinner@redhat.com>
Change the short form directory structure verifier function to return
the instruction pointer of a failing check or NULL if everything's ok.
Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
Reviewed-by: Dave Chinner <dchinner@redhat.com>
Create a function to check the structure of short form symlink targets.
Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
Reviewed-by: Dave Chinner <dchinner@redhat.com>
Create a function to perform structure verification for short form
extended attributes.
Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
Reviewed-by: Dave Chinner <dchinner@redhat.com>
Consolidate the fork size and format verifiers to xfs_dinode_verify so
that we can reject bad inodes earlier and in a single place.
Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
Reviewed-by: Dave Chinner <dchinner@redhat.com>
Move the v3 inode integrity information (crc, owner, metauuid) before we
look at anything else in the inode so that we don't waste time on a torn
write or a totally garbled block. This makes xfs_dinode_verify more
consistent with the other verifiers.
Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
Reviewed-by: Dave Chinner <dchinner@redhat.com>
Refactor the callers of verifiers to print the instruction address of a
failing check.
Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
Reviewed-by: Dave Chinner <dchinner@redhat.com>
Modify each function that checks the contents of a metadata buffer to
return the instruction address of the failing test so that we can report
more precise failure errors to the log.
Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
Reviewed-by: Dave Chinner <dchinner@redhat.com>
Since all verification errors also mark the buffer as having an error,
we can combine these two calls. Later we'll add a xfs_failaddr_t
parameter to promote the idea of reporting corruption errors and the
address of the failing check to enable better debugging reports.
Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
Reviewed-by: Dave Chinner <dchinner@redhat.com>
Since __xfs_dir3_data_check verifies on-disk metadata, we can't have it
noisily blowing asserts and hanging the system on corrupt data coming in
off the disk. Instead, have it return a boolean like all the other
checker functions, and only have it noisily fail if we fail in debug
mode.
Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
Reviewed-by: Dave Chinner <dchinner@redhat.com>
Now that we have xfs_verify_agbno, use it to verify short form btree
pointers instead of open-coding them.
Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
Reviewed-by: Dave Chinner <dchinner@redhat.com>
Create two helper functions to verify the headers of a long format
btree block. We'll use this later for the realtime rmapbt.
Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
Reviewed-by: Dave Chinner <dchinner@redhat.com>
We already have a function to verify fsb pointers, so get rid of the
last users of the (less robust) macro.
Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
Reviewed-by: Dave Chinner <dchinner@redhat.com>
The create transaction reservation calculation has two different
branches of code depending on whether the filesystem is a v5 format
fs or older. Each branch considers the max reservation between the
allocation case (new chunk allocation + record insert) and the
modify case (chunk exists, record modification) of inode allocation.
The modify case is the same for both superblock versions with the
exception of the finobt. The finobt helper checks the feature bit,
however, and so the modify case already shares the same code.
Now that inode chunk allocation has been refactored into a helper
that checks the superblock version to calculate the appropriate
reservation for the create transaction, the only remaining
difference between the create and icreate branches is the call to
the finobt helper. As noted above, the finobt helper is a no-op when
the feature is not enabled. Therefore, these branches are
effectively duplicate and can be condensed.
Remove the xfs_calc_create_*() branch of functions and update the
various callers to use the xfs_calc_icreate_*() variant. The latter
creates the same reservation size for v4 create transactions as the
removed branch. As such, this patch does not result in transaction
reservation changes.
Signed-off-by: Brian Foster <bfoster@redhat.com>
Reviewed-by: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Darrick J. Wong <darrick.wong@oracle.com>
Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
The reservation for the various forms of inode allocation is
scattered across several different functions. This includes two
variants of chunk allocation (v5 icreate transactions vs. older
create transactions) and the inode free transaction.
To clean up some of this code and clarify the purpose of specific
allocfree reservations, continue the pattern of defining helper
functions for smaller operational units of broader transactions.
Refactor the reservation into an inode chunk alloc/free helper that
considers the various conditions based on filesystem format.
An inode chunk free involves an extent free and buffer
invalidations. The latter requires reservation for log headers only.
An inode chunk allocation modifies the free space btrees and logs
the chunk on v4 supers. v5 supers initialize the inode chunk using
ordered buffers and so do not log the chunk.
As a side effect of this refactoring, add one more allocfree res to
the ifree transaction. Technically this does not serve a specific
purpose because inode chunks are freed via deferred operations and
thus occur after a transaction roll. tr_ifree has a bit of a history
of tx overruns caused by too many agfl fixups during sustained file
deletion workloads, so add this extra reservation as a form of
padding nonetheless.
Signed-off-by: Brian Foster <bfoster@redhat.com>
Reviewed-by: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Darrick J. Wong <darrick.wong@oracle.com>
Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
Analysis of recent reports of log reservation overruns and code
inspection has uncovered that the reservations associated with inode
operations may not cover the worst case scenarios. In particular,
many cases only include one allocfree res. for a particular
operation even though said operations may also entail AGFL fixups
and inode btree block allocations in addition to the actual inode
chunk allocation. This can easily turn into two or three block
allocations (or frees) per operation.
In theory, the only way to define the worst case reservation is to
include an allocfree res for each individual allocation in a
transaction. Since that is impractical (we can perform multiple agfl
fixups per tx and not every allocation results in a full tree
operation), we need to find a reasonable compromise that addresses
the deficiency in practice without blowing out the size of the
transactions.
Since the inode btrees are not filled by the AGFL, record insertion
and removal can directly result in block allocations and frees
depending on the shape of the tree. These allocations and frees
occur in the same transaction context as the inobt update itself,
but are separate from the allocation/free that might be required for
an inode chunk. Therefore, it makes sense to assume that an [f]inobt
insert/remove can directly result in one or more block allocations
on behalf of the tree.
Refactor the inode transaction reservations to include one allocfree
res. per inode btree modification to cover allocations required by
the tree itself. This separates the reservation required to allocate
the inode chunk from the reservation required for inobt record
insertion/removal. Apply the same logic to the finobt. This results
in killing off the finobt modify condition because we no longer
assume that the broader transaction reservation will cover finobt
block allocations and finobt shape changes can occur in either of
the inobt allocation or modify situations.
Suggested-by: Dave Chinner <david@fromorbit.com>
Signed-off-by: Brian Foster <bfoster@redhat.com>
Reviewed-by: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Darrick J. Wong <darrick.wong@oracle.com>
Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
The truncate transaction does not ever modify the inode btree, but
includes an associated log reservation. Update
xfs_calc_itruncate_reservation() to remove the reservation
associated with inobt updates.
Signed-off-by: Brian Foster <bfoster@redhat.com>
Reviewed-by: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Darrick J. Wong <darrick.wong@oracle.com>
Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
The current AGI unlinked list addition and removal reservations do
not reflect the worst case log usage. An unlinked list removal can
log up to two on-disk inode clusters but only includes reservation
for one. An unlinked list addition logs the on-disk cluster but
includes reservation for an in-core inode.
Update the AGI unlinked list reservation helpers to calculate the
correct worst case reservation for the associated operations.
Signed-off-by: Brian Foster <bfoster@redhat.com>
Reviewed-by: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Darrick J. Wong <darrick.wong@oracle.com>
Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
The tr_ifree transaction handles inode unlinks and inode chunk
frees. The current transaction calculation does not accurately
reflect worst case changes to the inode btree, however. The inobt
portion of the current transaction reservation only covers
modification of a single inobt buffer (for the particular inode
record). This is a historical artifact from the days before XFS
supported full inode chunk removal.
When support for inode chunk removal was added in commit
254f6311ed1b ("Implement deletion of inode clusters in XFS."), the
additional log reservation required for chunk removal was not added
correctly. The new reservation only considered the header overhead
of associated buffers rather than the full contents of the btrees
and AGF and AGFL buffers affected by the transaction. The
reservation for the free space btrees was subsequently fixed up in
commit 5fe6abb82f76 ("Add space for inode and allocation btrees to
ITRUNCATE log reservation"), but the res. for full inobt joins has
never been added.
Further review of the ifree reservation uncovered a couple more
problems:
- The undocumented +2 blocks are intended for the AGF and AGFL, but
are also not sized correctly and should be logged as full sectors
(not FSBs).
- The additional single block header is undocumented and serves no
apparent purpose.
Update xfs_calc_ifree_reservation() to include a full inobt join in
the reservation calculation. Refactor the undocumented blocks
appropriately and fix up the comments to reflect the current
calculation.
Signed-off-by: Brian Foster <bfoster@redhat.com>
Reviewed-by: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Darrick J. Wong <darrick.wong@oracle.com>
Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
For rmap removal, refactor the rmap owner checks into a separate
function, then skip the checks if we are performing an unknown-owner
removal.
Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Calling xfs_rmap_free with an unknown owner is supposed to remove any
rmaps covering that range regardless of owner. This is used by the EFI
recovery code to say "we're freeing this, it mustn't be owned by
anything anymore", but for whatever reason xfs_free_ag_extent filters
them out.
Therefore, remove the filter and make xfs_rmap_unmap actually treat it
as a wildcard owner -- free anything that's already there, and if
there's no owner at all then that's fine too.
There are two existing callers of bmap_add_free that take care the rmap
deferred ops themselves and use OWN_UNKNOWN to skip the EFI-based rmap
cleanup; convert these to use OWN_NULL (via helpers), and now we really
require that an RUI (if any) gets added to the defer ops before any EFI.
Lastly, now that xfs_free_extent filters out OWN_NULL rmap free requests,
growfs will have to consult directly with the rmap to ensure that there
aren't any rmaps in the grown region.
Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Under the deferred rmap operation scheme, there's a certain order in
which the rmap deferred ops have to be queued to maintain integrity
during log replay. For alloc/map operations that order is cui -> rui;
for free/unmap operations that order is cui -> rui -> efi. However, the
initial refcount code got the ordering wrong in the free side of things
because it queued refcount free op and an EFI and the refcount free op
queued a rmap free op, resulting in the order cui -> efi -> rui.
If we fail before the efd finishes, the efi recovery will try to do a
wildcard rmap removal and the subsequent rui will fail to find the rmap
and blow up. This didn't ever happen due to other screws up in handling
unknown owner rmap removals, but those other screw ups broke recovery in
other ways, so fix the ordering to follow the intended rules.
Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Move the tracepoint in xfs_iext_insert to after the point where we've
inserted the extent because otherwise we report stale extent data in
the ftrace output.
Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
In e1a4e37cc7 ("xfs: try to avoid blowing out the transaction
reservation when bunmaping a shared extent"), we try to constrain the
amount of real extents we unmap from the data fork in a given call so
that we don't blow out transaction reservations.
However, not all bunmapi operations require a transaction -- if we're
only removing a delalloc extent, no transaction is needed, so we have to
code against that.
Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
The new attribute leaf buffer is not held locked across the transaction
roll between the shortform->leaf modification and the addition of the
new entry. As a result, the attribute buffer modification being made is
not atomic from an operational perspective. Hence the AIL push can grab
it in the transient state of "just created" after the initial
transaction is rolled, because the buffer has been released. This leads
to xfs_attr3_leaf_verify() asserting that hdr.count is zero, treating
this as in-memory corruption, and shutting down the filesystem.
Darrick ported the original patch to 4.15 and reworked it use the
xfs_defer_bjoin helper and hold/join the buffer correctly across the
second transaction roll.
Signed-off-by: Alex Lyakas <alex@zadarastorage.com>
Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
In certain cases, defer_ops callers will lock a buffer and want to hold
the lock across transaction rolls. Similar to ijoined inodes, we want
to dirty & join the buffer with each transaction roll in defer_finish so
that afterwards the caller still owns the buffer lock and we haven't
inadvertently pinned the log.
Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Use _GOTO instead of _RETURN so we can free the allocated
cursor on error.
Fixes: bf80628 ("xfs: remove xfs_bmse_shift_one")
Fixes-coverity-id: 1423813, 1423676
Signed-off-by: Eric Sandeen <sandeen@redhat.com>
Reviewed-by: Darrick J. Wong <darrick.wong@oracle.com>
Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
And move them to xfs_linux.h so that xfsprogs can stub them out more
easily.
Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Darrick J. Wong <darrick.wong@oracle.com>
Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
found the issue by kmemleak.
unreferenced object 0xffff8800674611c0 (size 16):
xfs_iext_insert+0x82a/0xa90 [xfs]
xfs_bmap_add_extent_hole_delay+0x1e5/0x5b0 [xfs]
xfs_bmapi_reserve_delalloc+0x483/0x530 [xfs]
xfs_file_iomap_begin+0xac8/0xd40 [xfs]
iomap_apply+0xb8/0x1b0
iomap_file_buffered_write+0xac/0xe0
xfs_file_buffered_aio_write+0x198/0x420 [xfs]
xfs_file_write_iter+0x23f/0x2a0 [xfs]
__vfs_write+0x23e/0x340
vfs_write+0xe9/0x240
SyS_write+0xa1/0x120
do_syscall_64+0xda/0x260
Signed-off-by: Shu Wang <shuwang@redhat.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Darrick J. Wong <darrick.wong@oracle.com>
Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
Be consistent about using uint32_t/uint8_t instead of u32/u8. This is
more so that we don't have to maintain /those/ types in xfsprogs.
Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
Reviewed-by: Eric Sandeen <sandeen@redhat.com>
- Refactor the incore extent map manipulations to use a cursor instead of
directly modifying extent data.
- Refactor the incore extent map cursor to use an in-memory btree instead
of a single high-order allocation. This eliminates a major source of
complaints about insufficient memory when opening a heavily fragmented
file into a system whose memory is also heavily fragmented.
- Fix a longstanding bug where deleting a file with a complex extended
attribute btree incorrectly handled memory pointers, which could lead
to memory corruption.
- Improve metadata validation to eliminate crashing problems found while
fuzzing xfs.
- Move the error injection tag definitions into libxfs to be shared with
userspace components.
- Fix some log recovery bugs where we'd underflow log block position
vector and incorrectly fail log recovery.
- Drain the buffer lru after log recovery to force recovered buffers back
through the verifiers after mount. On a v4 filesystem the log never
attaches verifiers during log replay (v5 does), so we could end up with
buffers marked verified but without having ever been verified.
- Fix various other bugs.
- Introduce the first part of a new online fsck tool. The new fsck tool
will be able to iterate every piece of metadata in the filesystem to
look for obvious errors and corruptions. In the next release cycle
the checking will be extended to cross-reference with the other fs
metadata, so this feature should only be used by the developers in the
mean time.
-----BEGIN PGP SIGNATURE-----
Version: GnuPG v1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=vQrJ
-----END PGP SIGNATURE-----
Merge tag 'xfs-4.15-merge-1' of git://git.kernel.org/pub/scm/fs/xfs/xfs-linux
Pull xfs updates from Darrick Wong:
"xfs: great scads of new stuff for 4.15.
This merge cycle, we're making some substantive changes to XFS. The
in-core extent mappings have been refactored to use proper iterators
and a btree to handle heavily fragmented files without needing
high-order memory allocations; some important log recovery bug fixes;
and the first part of the online fsck functionality.
(The online fsck feature is disabled by default and more pieces of it
will be coming in future release cycles.)
This giant pile of patches has been run through a full xfstests run
over the weekend and through a quick xfstests run against this
morning's master, with no major failures reported.
New in this version:
- Refactor the incore extent map manipulations to use a cursor
instead of directly modifying extent data.
- Refactor the incore extent map cursor to use an in-memory btree
instead of a single high-order allocation. This eliminates a major
source of complaints about insufficient memory when opening a
heavily fragmented file into a system whose memory is also heavily
fragmented.
- Fix a longstanding bug where deleting a file with a complex
extended attribute btree incorrectly handled memory pointers, which
could lead to memory corruption.
- Improve metadata validation to eliminate crashing problems found
while fuzzing xfs.
- Move the error injection tag definitions into libxfs to be shared
with userspace components.
- Fix some log recovery bugs where we'd underflow log block position
vector and incorrectly fail log recovery.
- Drain the buffer lru after log recovery to force recovered buffers
back through the verifiers after mount. On a v4 filesystem the log
never attaches verifiers during log replay (v5 does), so we could
end up with buffers marked verified but without having ever been
verified.
- Fix various other bugs.
- Introduce the first part of a new online fsck tool. The new fsck
tool will be able to iterate every piece of metadata in the
filesystem to look for obvious errors and corruptions. In the next
release cycle the checking will be extended to cross-reference with
the other fs metadata, so this feature should only be used by the
developers in the mean time"
* tag 'xfs-4.15-merge-1' of git://git.kernel.org/pub/scm/fs/xfs/xfs-linux: (131 commits)
xfs: on failed mount, force-reclaim inodes after unmounting quota controls
xfs: check the uniqueness of the AGFL entries
xfs: remove u_int* type usage
xfs: handle zero entries case in xfs_iext_rebalance_leaf
xfs: add comments documenting the rebalance algorithm
xfs: trivial indentation fixup for xfs_iext_remove_node
xfs: remove a superflous assignment in xfs_iext_remove_node
xfs: add some comments to xfs_iext_insert/xfs_iext_insert_node
xfs: fix number of records handling in xfs_iext_split_leaf
fs/xfs: Remove NULL check before kmem_cache_destroy
xfs: only check da node header padding on v5 filesystems
xfs: fix btree scrub deref check
xfs: fix uninitialized return values in scrub code
xfs: pass inode number to xfs_scrub_ino_set_{preen,warning}
xfs: refactor the directory data block bestfree checks
xfs: mark xlog_verify_dest_ptr STATIC
xfs: mark xlog_recover_check_summary STATIC
xfs: mark xfs_btree_check_lblock and xfs_btree_check_ptr static
xfs: remove unreachable error injection code in xfs_qm_dqget
xfs: remove unused debug counts for xfs_lock_inodes
...
Use the uint* types instead of the u_int* types. This will (hopefully)
pair with an xfsprogs cleanup.
Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
Reviewed-by: Dave Chinner <dchinner@redhat.com>
And also rename fill to nr_entries to match the rest of the code.
Reported-by: Brian Foster <bfoster@redhat.com>
Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Brian Foster <bfoster@redhat.com>
Reviewed-by: Darrick J. Wong <darrick.wong@oracle.com>
Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
Reported-by: Brian Foster <bfoster@redhat.com>
Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Brian Foster <bfoster@redhat.com>
Reviewed-by: Darrick J. Wong <darrick.wong@oracle.com>
Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Brian Foster <bfoster@redhat.com>
Reviewed-by: Darrick J. Wong <darrick.wong@oracle.com>
Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
Reported-by: Brian Foster <bfoster@redhat.com>
Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Brian Foster <bfoster@redhat.com>
Reviewed-by: Darrick J. Wong <darrick.wong@oracle.com>
Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
Reported-by: Brian Foster <bfoster@redhat.com>
Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Brian Foster <bfoster@redhat.com>
Reviewed-by: Darrick J. Wong <darrick.wong@oracle.com>
Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
Fix to check the correct value, and remove a duplicate handling of the
uneven record number split algorith,
Reported-by: Brian Foster <bfoster@redhat.com>
Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Brian Foster <bfoster@redhat.com>
Reviewed-by: Darrick J. Wong <darrick.wong@oracle.com>
Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Darrick J. Wong <darrick.wong@oracle.com>
Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
Neither defines an on-disk format, so move them out of xfs_format.h.
Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Darrick J. Wong <darrick.wong@oracle.com>
Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
This removed an unaligned load per extent, as well as the manual poking
into the on-disk extent format.
Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Darrick J. Wong <darrick.wong@oracle.com>
Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
We only have two places that remove 2 extents at the same time, so unroll
the loop there.
Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Darrick J. Wong <darrick.wong@oracle.com>
Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
We only have two places that insert 2 extents at the same time, so unroll
the loop there.
Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Darrick J. Wong <darrick.wong@oracle.com>
Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
Replace the current linear list and the indirection array for the in-core
extent list with a b+tree to avoid the need for larger memory allocations
for the indirection array when lots of extents are present. The current
extent list implementations leads to heavy pressure on the memory
allocator when modifying files with a high extent count, and can lead
to high latencies because of that.
The replacement is a b+tree with a few quirks. The leaf nodes directly
store the extent record in two u64 values. The encoding is a little bit
different from the existing in-core extent records so that the start
offset and length which are required for lookups can be retreived with
simple mask operations. The inner nodes store a 64-bit key containing
the start offset in the first half of the node, and the pointers to the
next lower level in the second half. In either case we walk the node
from the beginninig to the end and do a linear search, as that is more
efficient for the low number of cache lines touched during a search
(2 for the inner nodes, 4 for the leaf nodes) than a binary search.
We store termination markers (zero length for the leaf nodes, an
otherwise impossible high bit for the inner nodes) to terminate the key
list / records instead of storing a count to use the available cache
lines as efficiently as possible.
One quirk of the algorithm is that while we normally split a node half and
half like usual btree implementations we just spill over entries added at
the very end of the list to a new node on its own. This means we get a
100% fill grade for the common cases of bulk insertion when reading an
inode into memory, and when only sequentially appending to a file. The
downside is a slightly higher chance of splits on the first random
insertions.
Both insert and removal manually recurse into the lower levels, but
the bulk deletion of the whole tree is still implemented as a recursive
function call, although one limited by the overall depth and with very
little stack usage in every iteration.
For the first few extents we dynamically grow the list from a single
extent to the next powers of two until we have a first full leaf block
and that building the actual tree.
The code started out based on the generic lib/btree.c code from Joern
Engel based on earlier work from Peter Zijlstra, but has since been
rewritten beyond recognition.
Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Darrick J. Wong <darrick.wong@oracle.com>
Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
To make life a little simpler make xfs_bmbt_set_all unaligned access
aware so that we can use it directly on the destination buffer.
Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Brian Foster <bfoster@redhat.com>
Reviewed-by: Darrick J. Wong <darrick.wong@oracle.com>
Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
Supporting a small bit of data inside the inode fork blows up the fork size
a lot, removing the 32 bytes of inline data halves the effective size of
the inode fork (and it still has a lot of unused padding left), and the
performance of a single kmalloc doesn't show up compared to the size to read
an inode or create one.
It also simplifies the fork management code a lot.
Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Darrick J. Wong <darrick.wong@oracle.com>
Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
Instead of looking up extents to convert and calling xfs_bmapi_write on
each of them just let xfs_bmapi_write handle the full range. To make
this robust add a new XFS_BMAPI_CONVERT_ONLY that only converts ranges
and never allocates blocks.
[darrick: shorten the stringified CONVERT_ONLY trace flag]
Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Darrick J. Wong <darrick.wong@oracle.com>
Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
Darrick J. Wong