- Various cleanups
- Livelock fixes for eofblocks scanning
- Improved input verification for on-disk metadata
- Fix races in the copy on write remap mechanism
- Fix buffer io error timeout controls
- Streamlining of directio copy on write
- Asynchronous discard support
- Fix asserts when splitting delalloc reservations
- Don't bloat bmbt when right shifting extents
- Inode alignment fixes for 32k block sizes
-----BEGIN PGP SIGNATURE-----
Version: GnuPG v1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=Ga1F
-----END PGP SIGNATURE-----
Merge tag 'xfs-4.11-merge-7' of git://git.kernel.org/pub/scm/fs/xfs/xfs-linux
Pull xfs updates from Darrick Wong:
"Here are the XFS changes for 4.11. We aren't introducing any major
features in this release cycle except for this being the first merge
window I've managed on my own. :)
Changes since last update:
- Various cleanups
- Livelock fixes for eofblocks scanning
- Improved input verification for on-disk metadata
- Fix races in the copy on write remap mechanism
- Fix buffer io error timeout controls
- Streamlining of directio copy on write
- Asynchronous discard support
- Fix asserts when splitting delalloc reservations
- Don't bloat bmbt when right shifting extents
- Inode alignment fixes for 32k block sizes"
* tag 'xfs-4.11-merge-7' of git://git.kernel.org/pub/scm/fs/xfs/xfs-linux: (39 commits)
xfs: remove XFS_ALLOCTYPE_ANY_AG and XFS_ALLOCTYPE_START_AG
xfs: simplify xfs_rtallocate_extent
xfs: tune down agno asserts in the bmap code
xfs: Use xfs_icluster_size_fsb() to calculate inode chunk alignment
xfs: don't reserve blocks for right shift transactions
xfs: fix len comparison in xfs_extent_busy_trim
xfs: fix uninitialized variable in _reflink_convert_cow
xfs: split indlen reservations fairly when under reserved
xfs: handle indlen shortage on delalloc extent merge
xfs: resurrect debug mode drop buffered writes mechanism
xfs: clear delalloc and cache on buffered write failure
xfs: don't block the log commit handler for discards
xfs: improve busy extent sorting
xfs: improve handling of busy extents in the low-level allocator
xfs: don't fail xfs_extent_busy allocation
xfs: correct null checks and error processing in xfs_initialize_perag
xfs: update ctime and mtime on clone destinatation inodes
xfs: allocate direct I/O COW blocks in iomap_begin
xfs: go straight to real allocations for direct I/O COW writes
xfs: return the converted extent in __xfs_reflink_convert_cow
...
XFS_ALLOCTYPE_ANY_AG was only used for the RT allocator and is unused
now, and XFS_ALLOCTYPE_START_AG has been unused for a while.
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 can deduce the allocation type from the bno argument, and do the
return without prod much simpler internally.
Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Darrick J. Wong <darrick.wong@oracle.com>
[darrick: fix the macro for the non-rt build]
Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
In various places we currently assert that xfs_bmap_btalloc allocates
from the same as the firstblock value passed in, unless it's either
NULLAGNO or the dop_low flag is set. But the reflink code does not
fully follow this convention as it passes in firstblock purely as
a hint for the allocator without actually having previous allocations
in the transaction, and without having a minleft check on the current
AG, leading to the assert firing on a very full and heavily used
file system. As even the reflink code only allocates from equal or
higher AGs for now we can simply the check to always allow for equal
or higher AGs.
Note that we need to eventually split the two meanings of the firstblock
value. At that point we can also allow the reflink code to allocate
from any AG instead of limiting it in any way.
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>
On a ppc64 system, executing generic/256 test with 32k block size gives the following call trace,
XFS: Assertion failed: args->maxlen > 0, file: /root/repos/linux/fs/xfs/libxfs/xfs_alloc.c, line: 2026
kernel BUG at /root/repos/linux/fs/xfs/xfs_message.c:113!
Oops: Exception in kernel mode, sig: 5 [#1]
SMP NR_CPUS=2048
DEBUG_PAGEALLOC
NUMA
pSeries
Modules linked in:
CPU: 2 PID: 19361 Comm: mkdir Not tainted 4.10.0-rc5 #58
task: c000000102606d80 task.stack: c0000001026b8000
NIP: c0000000004ef798 LR: c0000000004ef798 CTR: c00000000082b290
REGS: c0000001026bb090 TRAP: 0700 Not tainted (4.10.0-rc5)
MSR: 8000000000029032 <SF,EE,ME,IR,DR,RI>
CR: 28004428 XER: 00000000
CFAR: c0000000004ef180 SOFTE: 1
GPR00: c0000000004ef798 c0000001026bb310 c000000001157300 ffffffffffffffea
GPR04: 000000000000000a c0000001026bb130 0000000000000000 ffffffffffffffc0
GPR08: 00000000000000d1 0000000000000021 00000000ffffffd1 c000000000dd4990
GPR12: 0000000022004444 c00000000fe00800 0000000020000000 0000000000000000
GPR16: 0000000000000000 0000000043a606fc 0000000043a76c08 0000000043a1b3d0
GPR20: 000001002a35cd60 c0000001026bbb80 0000000000000000 0000000000000001
GPR24: 0000000000000240 0000000000000004 c00000062dc55000 0000000000000000
GPR28: 0000000000000004 c00000062ecd9200 0000000000000000 c0000001026bb6c0
NIP [c0000000004ef798] .assfail+0x28/0x30
LR [c0000000004ef798] .assfail+0x28/0x30
Call Trace:
[c0000001026bb310] [c0000000004ef798] .assfail+0x28/0x30 (unreliable)
[c0000001026bb380] [c000000000455d74] .xfs_alloc_space_available+0x194/0x1b0
[c0000001026bb410] [c00000000045b914] .xfs_alloc_fix_freelist+0x144/0x480
[c0000001026bb580] [c00000000045c368] .xfs_alloc_vextent+0x698/0xa90
[c0000001026bb650] [c0000000004a6200] .xfs_ialloc_ag_alloc+0x170/0x820
[c0000001026bb7c0] [c0000000004a9098] .xfs_dialloc+0x158/0x320
[c0000001026bb8a0] [c0000000004e628c] .xfs_ialloc+0x7c/0x610
[c0000001026bb990] [c0000000004e8138] .xfs_dir_ialloc+0xa8/0x2f0
[c0000001026bbaa0] [c0000000004e8814] .xfs_create+0x494/0x790
[c0000001026bbbf0] [c0000000004e5ebc] .xfs_generic_create+0x2bc/0x410
[c0000001026bbce0] [c0000000002b4a34] .vfs_mkdir+0x154/0x230
[c0000001026bbd70] [c0000000002bc444] .SyS_mkdirat+0x94/0x120
[c0000001026bbe30] [c00000000000b760] system_call+0x38/0xfc
Instruction dump:
4e800020 60000000 7c0802a6 7c862378 3c82ffca 7ca72b78 38841c18 7c651b78
38600000 f8010010 f821ff91 4bfff94d <0fe00000> 60000000 7c0802a6 7c892378
When block size is larger than inode cluster size, the call to
XFS_B_TO_FSBT(mp, mp->m_inode_cluster_size) returns 0. Also, mkfs.xfs
would have set xfs_sb->sb_inoalignmt to 0. This causes
xfs_ialloc_cluster_alignment() to return 0. Due to this
args.minalignslop (in xfs_ialloc_ag_alloc()) gets the unsigned
equivalent of -1 assigned to it. This later causes alloc_len in
xfs_alloc_space_available() to have a value of 0. In such a scenario
when args.total is also 0, the assert statement "ASSERT(args->maxlen >
0);" fails.
This commit fixes the bug by replacing the call to XFS_B_TO_FSBT() in
xfs_ialloc_cluster_alignment() with a call to xfs_icluster_size_fsb().
Suggested-by: Darrick J. Wong <darrick.wong@oracle.com>
Signed-off-by: Chandan Rajendra <chandan@linux.vnet.ibm.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>
The block reservation for the transaction allocated in
xfs_shift_file_space() is an artifact of the original collapse range
support. It exists to handle the case where a collapse range occurs,
the initial extent is left shifted into a location that forms a
contiguous boundary with the previous extent and thus the extents
are merged. This code was subsequently refactored and reused for
insert range (right shift) support.
If an insert range occurs under low free space conditions, the
extent at the starting offset is split before the first shift
transaction is allocated. If the block reservation fails, this
leaves separate, but contiguous extents around in the inode. While
not a fatal problem, this is unexpected and will flag a warning on
subsequent insert range operations on the inode. This problem has
been reproduce intermittently by generic/270 running against a
ramdisk device.
Since right shift does not create new extent boundaries in the
inode, a block reservation for extent merge is unnecessary. Update
xfs_shift_file_space() to conditionally reserve fs blocks for left
shift transactions only. This avoids the warning reproduced by
generic/270.
Reported-by: Ross Zwisler <ross.zwisler@linux.intel.com>
Signed-off-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>
The length is now passed by reference, so the assertion has to be updated
to match the other changes, as pointed out by this W=1 warning:
fs/xfs/xfs_extent_busy.c: In function 'xfs_extent_busy_trim':
fs/xfs/xfs_extent_busy.c:356:13: error: ordered comparison of pointer with integer zero [-Werror=extra]
Fixes: ebf5587261 ("xfs: improve handling of busy extents in the low-level allocator")
Signed-off-by: Arnd Bergmann <arnd@arndb.de>
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>
Fix an uninitialize variable.
Reported-by: Dan Carpenter <dan.carpenter@oracle.com>
Reviewed-by: Brian Foster <bfoster@redhat.com>
Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
Certain workoads that punch holes into speculative preallocation can
cause delalloc indirect reservation splits when the delalloc extent is
split in two. If further splits occur, an already short-handed extent
can be split into two in a manner that leaves zero indirect blocks for
one of the two new extents. This occurs because the shortage is large
enough that the xfs_bmap_split_indlen() algorithm completely drains the
requested indlen of one of the extents before it honors the existing
reservation.
This ultimately results in a warning from xfs_bmap_del_extent(). This
has been observed during file copies of large, sparse files using 'cp
--sparse=always.'
To avoid this problem, update xfs_bmap_split_indlen() to explicitly
apply the reservation shortage fairly between both extents. This smooths
out the overall indlen shortage and defers the situation where we end up
with a delalloc extent with zero indlen reservation to extreme
circumstances.
Reported-by: Patrick Dung <mpatdung@gmail.com>
Signed-off-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>
When a delalloc extent is created, it can be merged with pre-existing,
contiguous, delalloc extents. When this occurs,
xfs_bmap_add_extent_hole_delay() merges the extents along with the
associated indirect block reservations. The expectation here is that the
combined worst case indlen reservation is always less than or equal to
the indlen reservation for the individual extents.
This is not always the case, however, as existing extents can less than
the expected indlen reservation if the extent was previously split due
to a hole punch. If a new extent merges with such an extent, the total
indlen requirement may be larger than the sum of the indlen reservations
held by both extents.
xfs_bmap_add_extent_hole_delay() assumes that the worst case indlen
reservation is always available and assigns it to the merged extent
without consideration for the indlen held by the pre-existing extent. As
a result, the subsequent xfs_mod_fdblocks() call can attempt an
unintentional allocation rather than a free (indicated by an ASSERT()
failure). Further, if the allocation happens to fail in this context,
the failure goes unhandled and creates a filesystem wide block
accounting inconsistency.
Fix xfs_bmap_add_extent_hole_delay() to function as designed. Cap the
indlen reservation assigned to the merged extent to the sum of the
indlen reservations held by each of the individual extents.
Signed-off-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>
A debug mode write failure mechanism was introduced to XFS in commit
801cc4e17a ("xfs: debug mode forced buffered write failure") to
facilitate targeted testing of delalloc indirect reservation management
from userspace. This code was subsequently rendered ineffective by the
move to iomap based buffered writes in commit 68a9f5e700 ("xfs:
implement iomap based buffered write path"). This likely went unnoticed
because the associated userspace code had not made it into xfstests.
Resurrect this mechanism to facilitate effective indlen reservation
testing from xfstests. The move to iomap based buffered writes relocated
the hook this mechanism needs to return write failure from XFS to
generic code. The failure trigger must remain in XFS. Given that
limitation, convert this from a write failure mechanism to one that
simply drops writes without returning failure to userspace. Rename all
"fail_writes" references to "drop_writes" to illustrate the point. This
is more hacky than preferred, but still triggers the XFS error handling
behavior required to drive the indlen tests. This is only available in
DEBUG mode and for testing purposes only.
Signed-off-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>
The buffered write failure handling code in
xfs_file_iomap_end_delalloc() has a couple minor problems. First, if
written == 0, start_fsb is not rounded down and it fails to kill off a
delalloc block if the start offset is block unaligned. This results in a
lingering delalloc block and broken delalloc block accounting detected
at unmount time. Fix this by rounding down start_fsb in the unlikely
event that written == 0.
Second, it is possible for a failed overwrite of a delalloc extent to
leave dirty pagecache around over a hole in the file. This is because is
possible to hit ->iomap_end() on write failure before the iomap code has
attempted to allocate pagecache, and thus has no need to clean it up. If
the targeted delalloc extent was successfully written by a previous
write, however, then it does still have dirty pages when ->iomap_end()
punches out the underlying blocks. This ultimately results in writeback
over a hole. To fix this problem, unconditionally punch out the
pagecache from XFS before the associated delalloc range.
Signed-off-by: Brian Foster <bfoster@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>
Instead we submit the discard requests and use another workqueue to
release the extents from the extent busy list.
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>
Sort busy extents by the full block number instead of just the AGNO so
that we can issue consecutive discard requests that the block layer could
merge (although we'll need additional block layer fixes for fast devices).
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>
Currently we force the log and simply try again if we hit a busy extent,
but especially with online discard enabled it might take a while after
the log force for the busy extents to disappear, and we might have
already completed our second pass.
So instead we add a new waitqueue and a generation counter to the pag
structure so that we can do wakeups once we've removed busy extents,
and we replace the single retry with an unconditional one - after
all we hold the AGF buffer lock, so no other allocations or frees
can be racing with us in this AG.
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 don't just need the structure to track busy extents which can be
avoided with a synchronous transaction, but also to keep track of
pending discard.
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>
If pag cannot be allocated, the current error exit path will trip
a null pointer deference error when calling xfs_buf_hash_destroy
with a null pag. Fix this by adding a new error exit labels and
jumping to those accordingly, avoiding the hash destroy and
unnecessary kmem_free on pag.
Up to three things need to be properly unwound:
1) pag memory allocation
2) xfs_buf_hash_init
3) radix_tree_insert
For any given iteration through the loop, any of the above which
succeed must be unwound for /this/ pag, and then all prior
initialized pags must be unwound.
Addresses-Coverity-Id: 1397628 ("Dereference after null check")
Reported-by: Colin Ian King <colin.king@canonical.com>
Signed-off-by: Bill O'Donnell <billodo@redhat.com>
Reviewed-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>
We're changing both metadata and data, so we need to update the
timestamps for clone operations. Dedupe on the other hand does
not change file data, and only changes invisible metadata so the
timestamps should not be updated.
This follows existing btrfs behavior.
Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Darrick J. Wong <darrick.wong@oracle.com>
[darrick: remove redundant is_dedupe test]
Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
Instead of preallocating all the required COW blocks in the high-level
write code do it inside the iomap code, like we do for all other I/O.
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>
When we allocate COW fork blocks for direct I/O writes we currently first
create a delayed allocation, and then convert it to a real allocation
once we've got the delayed one.
As there is no good reason for that this patch instead makes use call
xfs_bmapi_write from the COW allocation path. The only interesting bits
are a few tweaks the low-level allocator to allow for this, most notably
the need to remove the call to xfs_bmap_extsize_align for the cowextsize
in xfs_bmap_btalloc - for the existing convert case it's a no-op, but
for the direct allocation case it would blow up our block reservation
way beyond what we reserved for the transaction.
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'll need it for the direct I/O code. Also rename the function to
xfs_reflink_convert_cow_extent to describe it a bit better.
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>
Factor a helper to calculate the extent-size aligned block out of the
iomap code, so that it can be reused by the upcoming reflink dio code.
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 currently fall back from direct to buffered writes if we detect a
remaining shared extent in the iomap_begin callback. But by the time
iomap_begin is called for the potentially unaligned end block we might
have already written most of the data to disk, which we'd now write
again using buffered I/O. To avoid this reject all writes to reflinked
files before starting I/O so that we are guaranteed to only write the
data once.
The alternative would be to unshare the unaligned start and/or end block
before doing the I/O. I think that's doable, and will actually be
required to support reflinks on DAX file system. But it will take a
little more time and I'd rather get rid of the double write ASAP.
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>
After successful IO or permanent error, b_first_retry_time also
needs to be cleared, else the invalid first retry time will be
used by the next retry check.
Signed-off-by: Hou Tao <houtao1@huawei.com>
Reviewed-by: Darrick J. Wong <darrick.wong@oracle.com>
Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
Christoph Hellwig pointed out that there's a potentially nasty race when
performing simultaneous nearby directio cow writes:
"Thread 1 writes a range from B to c
" B --------- C
p
"a little later thread 2 writes from A to B
" A --------- B
p
[editor's note: the 'p' denote cowextsize boundaries, which I added to
make this more clear]
"but the code preallocates beyond B into the range where thread
"1 has just written, but ->end_io hasn't been called yet.
"But once ->end_io is called thread 2 has already allocated
"up to the extent size hint into the write range of thread 1,
"so the end_io handler will splice the unintialized blocks from
"that preallocation back into the file right after B."
We can avoid this race by ensuring that thread 1 cannot accidentally
remap the blocks that thread 2 allocated (as part of speculative
preallocation) as part of t2's write preparation in t1's end_io handler.
The way we make this happen is by taking advantage of the unwritten
extent flag as an intermediate step.
Recall that when we begin the process of writing data to shared blocks,
we create a delayed allocation extent in the CoW fork:
D: --RRRRRRSSSRRRRRRRR---
C: ------DDDDDDD---------
When a thread prepares to CoW some dirty data out to disk, it will now
convert the delalloc reservation into an /unwritten/ allocated extent in
the cow fork. The da conversion code tries to opportunistically
allocate as much of a (speculatively prealloc'd) extent as possible, so
we may end up allocating a larger extent than we're actually writing
out:
D: --RRRRRRSSSRRRRRRRR---
U: ------UUUUUUU---------
Next, we convert only the part of the extent that we're actively
planning to write to normal (i.e. not unwritten) status:
D: --RRRRRRSSSRRRRRRRR---
U: ------UURRUUU---------
If the write succeeds, the end_cow function will now scan the relevant
range of the CoW fork for real extents and remap only the real extents
into the data fork:
D: --RRRRRRRRSRRRRRRRR---
U: ------UU--UUU---------
This ensures that we never obliterate valid data fork extents with
unwritten blocks from the CoW fork.
Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
In the data fork, we only allow extents to perform the following state
transitions:
delay -> real <-> unwritten
There's no way to move directly from a delalloc reservation to an
/unwritten/ allocated extent. However, for the CoW fork we want to be
able to do the following to each extent:
delalloc -> unwritten -> written -> remapped to data fork
This will help us to avoid a race in the speculative CoW preallocation
code between a first thread that is allocating a CoW extent and a second
thread that is remapping part of a file after a write. In order to do
this, however, we need two things: first, we have to be able to
transition from da to unwritten, and second the function that converts
between real and unwritten has to be made aware of the cow fork. Do
both of those things.
Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Perform basic sanity checking of the directory free block header
fields so that we avoid hanging the system on invalid data.
(Granted that just means that now we shutdown on directory write,
but that seems better than hanging...)
Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
We can't handle a bmbt that's taller than BTREE_MAXLEVELS, and there's
no such thing as a zero-level bmbt (for that we have extents format),
so if we see this, send back an error code.
Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Don't let anybody load an obviously bad btree pointer. Since the values
come from disk, we must return an error, not just ASSERT.
Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
Reviewed-by: Eric Sandeen <sandeen@redhat.com>
When we open a directory, we try to readahead block 0 of the directory
on the assumption that we're going to need it soon. If the bmbt is
corrupt, the directory will never be usable and the readahead fails
immediately, so we might as well prevent the directory from being opened
at all. This prevents a subsequent read or modify operation from
hitting it and taking the fs offline.
NOTE: We're only checking for early failures in the block mapping, not
the readahead directory block itself.
Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
Reviewed-by: Eric Sandeen <sandeen@redhat.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
We use di_format and if_flags to decide whether we're grabbing the ilock
in btree mode (btree extents not loaded) or shared mode (anything else),
but the state of those fields can be changed by other threads that are
also trying to load the btree extents -- IFEXTENTS gets set before the
_bmap_read_extents call and cleared if it fails.
We don't actually need to have IFEXTENTS set until after the bmbt
records are successfully loaded and validated, which will fix the race
between multiple threads trying to read the same directory. The next
patch strengthens directory bmbt validation by refusing to open the
directory if reading the bmbt to start directory readahead fails.
Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
blk_get_backing_dev_info() is now a simple dereference. Remove that
function and simplify some code around that.
Signed-off-by: Jan Kara <jack@suse.cz>
Signed-off-by: Jens Axboe <axboe@fb.com>
The "full" argument was used only by the fiemap formatter,
which is now gone with the iomap updates.
Remove the unused arg.
Signed-off-by: Eric Sandeen <sandeen@redhat.com>
Reviewed-by: Alex Elder <elder@linaro.org>
Reviewed-by: Darrick J. Wong <darrick.wong@oracle.com>
Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
It's possible for post-eof blocks to end up being used for direct I/O
writes. dio write performs an upfront unwritten extent allocation, sends
the dio and then updates the inode size (if necessary) on write
completion. If a file release occurs while a file extending dio write is
in flight, it is possible to mistake the post-eof blocks for speculative
preallocation and incorrectly truncate them from the inode. This means
that the resulting dio write completion can discover a hole and allocate
new blocks rather than perform unwritten extent conversion.
This requires a strange mix of I/O and is thus not likely to reproduce
in real world workloads. It is intermittently reproduced by generic/299.
The error manifests as an assert failure due to transaction overrun
because the aforementioned write completion transaction has only
reserved enough blocks for btree operations:
XFS: Assertion failed: tp->t_blk_res_used <= tp->t_blk_res, \
file: fs/xfs//xfs_trans.c, line: 309
The root cause is that xfs_free_eofblocks() uses i_size to truncate
post-eof blocks from the inode, but async, file extending direct writes
do not update i_size until write completion, long after inode locks are
dropped. Therefore, xfs_free_eofblocks() effectively truncates the inode
to the incorrect size.
Update xfs_free_eofblocks() to serialize against dio similar to how
extending writes are serialized against i_size updates before post-eof
block zeroing. Specifically, wait on dio while under the iolock. This
ensures that dio write completions have updated i_size before post-eof
blocks are processed.
Signed-off-by: Brian Foster <bfoster@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>
The xfs_eofblocks.eof_scan_owner field is an internal field to
facilitate invoking eofb scans from the kernel while under the iolock.
This is necessary because the eofb scan acquires the iolock of each
inode. Synchronous scans are invoked on certain buffered write failures
while under iolock. In such cases, the scan owner indicates that the
context for the scan already owns the particular iolock and prevents a
double lock deadlock.
eofblocks scans while under iolock are still livelock prone in the event
of multiple parallel scans, however. If multiple buffered writes to
different inodes fail and invoke eofblocks scans at the same time, each
scan avoids a deadlock with its own inode by virtue of the
eof_scan_owner field, but will never be able to acquire the iolock of
the inode from the parallel scan. Because the low free space scans are
invoked with SYNC_WAIT, the scan will not return until it has processed
every tagged inode and thus both scans will spin indefinitely on the
iolock being held across the opposite scan. This problem can be
reproduced reliably by generic/224 on systems with higher cpu counts
(x16).
To avoid this problem, simplify the semantics of eofblocks scans to
never invoke a scan while under iolock. This means that the buffered
write context must drop the iolock before the scan. It must reacquire
the lock before the write retry and also repeat the initial write
checks, as the original state might no longer be valid once the iolock
was dropped.
Signed-off-by: Brian Foster <bfoster@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>
xfs_free_eofblocks() requires the IOLOCK_EXCL lock, but is called from
different contexts where the lock may or may not be held. The
need_iolock parameter exists for this reason, to indicate whether
xfs_free_eofblocks() must acquire the iolock itself before it can
proceed.
This is ugly and confusing. Simplify the semantics of
xfs_free_eofblocks() to require the caller to acquire the iolock
appropriately and kill the need_iolock parameter. While here, the mp
param can be removed as well as the xfs_mount is accessible from the
xfs_inode structure. This patch does not change behavior.
Signed-off-by: Brian Foster <bfoster@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>
After scratching my head looking for "xfs_busy_extent" I realized
it's not used; it's xfs_extent_busy, and the declaration for the
other name is bogus. Remove that and a few others as well.
(struct xfs_log_callback is used, but the 2nd declaration is
unnecessary).
Signed-off-by: Eric Sandeen <sandeen@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>
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>
Now that xfs_btree_init_block_int is able to determine crc
status from the passed-in mp, we can determine the proper
magic as well if we are given a btree number, rather than
an explicit magic value.
Change xfs_btree_init_block[_int] callers to pass in the
btree number, and let xfs_btree_init_block_int use the
xfs_magics array via the xfs_btree_magic macro to determine
which magic value is needed. This makes all of the
if (crc) / else stanzas identical, and the if/else can be
removed, leading to a single, common init_block call.
Signed-off-by: Eric Sandeen <sandeen@redhat.com>
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>
Right now the xfs_btree_magic() define takes only a cursor;
change this to take crc and btnum args to make it more generically
useful, and move to a function.
This will allow xfs_btree_init_block_int callers which don't
have a cursor to make use of the xfs_magics array, which will
happen in the next patch.
Signed-off-by: Eric Sandeen <sandeen@redhat.com>
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>
xfs_btree_init_block_int() can determine whether crcs are
in effect without the passed-in XFS_BTREE_CRC_BLOCKS flag;
the mp argument allows us to determine this from the
superblock. Remove the flag from callers, and use
xfs_sb_version_hascrc(&mp->m_sb) internally instead.
This removes one difference between the if & else cases
in the callers.
Signed-off-by: Eric Sandeen <sandeen@redhat.com>
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>
Quotacheck runs at mount time in situations where quota accounting must
be recalculated. In doing so, it uses bulkstat to visit every inode in
the filesystem. Historically, every inode processed during quotacheck
was released and immediately tagged for reclaim because quotacheck runs
before the superblock is marked active by the VFS. In other words,
the final iput() lead to an immediate ->destroy_inode() call, which
allowed the XFS background reclaim worker to start reclaiming inodes.
Commit 17c12bcd3 ("xfs: when replaying bmap operations, don't let
unlinked inodes get reaped") marks the XFS superblock active sooner as
part of the mount process to support caching inodes processed during log
recovery. This occurs before quotacheck and thus means all inodes
processed by quotacheck are inserted to the LRU on release. The
s_umount lock is held until the mount has completed and thus prevents
the shrinkers from operating on the sb. This means that quotacheck can
excessively populate the inode LRU and lead to OOM conditions on systems
without sufficient RAM.
Update the quotacheck bulkstat handler to set XFS_IGET_DONTCACHE on
inodes processed by quotacheck. This causes ->drop_inode() to return 1
and in turn causes iput_final() to evict the inode. This preserves the
original quotacheck behavior and prevents it from overloading the LRU
and running out of memory.
CC: stable@vger.kernel.org # v4.9
Reported-by: Martin Svec <martin.svec@zoner.cz>
Signed-off-by: Brian Foster <bfoster@redhat.com>
Reviewed-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>
In a bmapx call, bmv_count is the total size of the array, including the
zeroth element that userspace uses to supply the search key. The output
array starts at offset 1 so that we can set up the user for the next
invocation. Since we now can split an extent into multiple bmap records
due to shared/unshared status, we have to be careful that we don't
overflow the output array.
In the original patch f86f403794 ("xfs: teach get_bmapx about shared
extents and the CoW fork") I used cur_ext (the output index) to check
for overflows, albeit with an off-by-one error. Since nexleft no longer
describes the number of unfilled slots in the output, we can rip all
that out and use cur_ext for the overflow check directly.
Failure to do this causes heap corruption in bmapx callers such as
xfs_io and xfs_scrub. xfs/328 can reproduce this problem.
Reviewed-by: Eric Sandeen <sandeen@redhat.com>
Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
If we try to allocate memory pages to back an xfs_buf that we're trying
to read, it's possible that we'll be so short on memory that the page
allocation fails. For a blocking read we'll just wait, but for
readahead we simply dump all the pages we've collected so far.
Unfortunately, after dumping the pages we neglect to clear the
_XBF_PAGES state, which means that the subsequent call to xfs_buf_free
thinks that b_pages still points to pages we own. It then double-frees
the b_pages pages.
This results in screaming about negative page refcounts from the memory
manager, which xfs oughtn't be triggering. To reproduce this case,
mount a filesystem where the size of the inodes far outweighs the
availalble memory (a ~500M inode filesystem on a VM with 300MB memory
did the trick here) and run bulkstat in parallel with other memory
eating processes to put a huge load on the system. The "check summary"
phase of xfs_scrub also works for this purpose.
Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
Reviewed-by: Eric Sandeen <sandeen@redhat.com>
With COW files they are the hotpath, just like for files with the
extent size hint attribute. We really shouldn't micro-manage anything
but failure cases with unlikely.
Additionally Arnd Bergmann recently reported that one of these two
unlikely annotations causes link failures together with an upcoming
kernel instrumentation patch, so let's get rid of it ASAP.
Signed-off-by: Christoph Hellwig <hch@lst.de>
Reported-by: Arnd Bergmann <arnd@arndb.de>
Reviewed-by: Darrick J. Wong <darrick.wong@oracle.com>
Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
xfs_attr_[get|remove]() have unlocked attribute fork checks to optimize
away a lock cycle in cases where the fork does not exist or is otherwise
empty. This check is not safe, however, because an attribute fork short
form to extent format conversion includes a transient state that causes
the xfs_inode_hasattr() check to fail. Specifically,
xfs_attr_shortform_to_leaf() creates an empty extent format attribute
fork and then adds the existing shortform attributes to it.
This means that lookup of an existing xattr can spuriously return
-ENOATTR when racing against a setxattr that causes the associated
format conversion. This was originally reproduced by an untar on a
particularly configured glusterfs volume, but can also be reproduced on
demand with properly crafted xattr requests.
The format conversion occurs under the exclusive ilock. xfs_attr_get()
and xfs_attr_remove() already have the proper locking and checks further
down in the functions to handle this situation correctly. Drop the
unlocked checks to avoid the spurious failure and rely on the existing
logic.
Signed-off-by: Brian Foster <bfoster@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>
Currently we try to rely on the global reserved block pool for block
allocations for the free inode btree, but I have customer reports
(fairly complex workload, need to find an easier reproducer) where that
is not enough as the AG where we free an inode that requires a new
finobt block is entirely full. This causes us to cancel a dirty
transaction and thus a file system shutdown.
I think the right way to guard against this is to treat the finot the same
way as the refcount btree and have a per-AG reservations for the possible
worst case size of it, and the patch below implements that.
Note that this could increase mount times with large finobt trees. In
an ideal world we would have added a field for the number of finobt
fields to the AGI, similar to what we did for the refcount blocks.
We should do add it next time we rev the AGI or AGF format by adding
new fields.
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>
Try to reserve the blocks first and only then update the fields in
or hanging off the mount structure. This way we can call __xfs_ag_resv_init
again after a previous failure.
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>
sb_dirblklog is added to sb_blocklog to compute the directory block size
in bytes. Therefore, we must compare the sum of both those values
against XFS_MAX_BLOCKSIZE_LOG, not just dirblklog.
Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
Reviewed-by: Eric Sandeen <sandeen@redhat.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
