An inode in the AIL can be flush locked and marked stale if
a cluster free transaction occurs at the right time. The
inode item is then marked as flushing, which causes xfsaild
to spin and leaves the filesystem stalled. This is
reproduced by running xfstests 273 in a loop for an
extended period of time.
Check for stale inodes before the flush lock. This marks
the inode as pinned, leads to a log flush and allows the
filesystem to proceed.
Signed-off-by: Brian Foster <bfoster@redhat.com>
Reviewed-by: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Mark Tinguely <tinguely@sgi.com>
Signed-off-by: Ben Myers <bpm@sgi.com>
On filesytems with a block size smaller than PAGE_SIZE we currently have
a problem with unwritten extents. If a we have multi-block page for
which an unwritten extent has been allocated, and only some of the
buffers have been written to, and they are not contiguous, we can expose
stale data from disk in the blocks between the writes after extent
conversion.
Example of a page with unwritten and real data.
buffer content
0 empty b_state = 0
1 DATA b_state = 0x1023 Uptodate,Dirty,Mapped,Unwritten
2 DATA b_state = 0x1023 Uptodate,Dirty,Mapped,Unwritten
3 empty b_state = 0
4 empty b_state = 0
5 DATA b_state = 0x1023 Uptodate,Dirty,Mapped,Unwritten
6 DATA b_state = 0x1023 Uptodate,Dirty,Mapped,Unwritten
7 empty b_state = 0
Buffers 1, 2, 5, and 6 have been written to, leaving 0, 3, 4, and 7
empty. Currently buffers 1, 2, 5, and 6 are added to a single ioend,
and when IO has completed, extent conversion creates a real extent from
block 1 through block 6, leaving 0 and 7 unwritten. However buffers 3
and 4 were not written to disk, so stale data is exposed from those
blocks on a subsequent read.
Fix this by setting iomap_valid = 0 when we find a buffer that is not
Uptodate. This ensures that buffers 5 and 6 are not added to the same
ioend as buffers 1 and 2. Later these blocks will be converted into two
separate real extents, leaving the blocks in between unwritten.
Signed-off-by: Alain Renaud <arenaud@sgi.com>
Reviewed-by: Dave Chinner <dchinner@redhat.com>
Signed-off-by: Ben Myers <bpm@sgi.com>
Btrfs has to make sure we have space to allocate new blocks in order to modify
the inode, so updating time can fail. We've gotten around this by having our
own file_update_time but this is kind of a pain, and Christoph has indicated he
would like to make xfs do something different with atime updates. So introduce
->update_time, where we will deal with i_version an a/m/c time updates and
indicate which changes need to be made. The normal version just does what it
has always done, updates the time and marks the inode dirty, and then
filesystems can choose to do something different.
I've gone through all of the users of file_update_time and made them check for
errors with the exception of the fault code since it's complicated and I wasn't
quite sure what to do there, also Jan is going to be pushing the file time
updates into page_mkwrite for those who have it so that should satisfy btrfs and
make it not a big deal to check the file_update_time() return code in the
generic fault path. Thanks,
Signed-off-by: Josef Bacik <josef@redhat.com>
pass inode + parent's inode or NULL instead of dentry + bool saying
whether we want the parent or not.
NOTE: that needs ceph fix folded in.
Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
-----BEGIN PGP SIGNATURE-----
Version: GnuPG v1.4.11 (GNU/Linux)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=oLXf
-----END PGP SIGNATURE-----
Merge tag 'writeback' of git://git.kernel.org/pub/scm/linux/kernel/git/wfg/linux
Pull writeback tree from Wu Fengguang:
"Mainly from Jan Kara to avoid iput() in the flusher threads."
* tag 'writeback' of git://git.kernel.org/pub/scm/linux/kernel/git/wfg/linux:
writeback: Avoid iput() from flusher thread
vfs: Rename end_writeback() to clear_inode()
vfs: Move waiting for inode writeback from end_writeback() to evict_inode()
writeback: Refactor writeback_single_inode()
writeback: Remove wb->list_lock from writeback_single_inode()
writeback: Separate inode requeueing after writeback
writeback: Move I_DIRTY_PAGES handling
writeback: Move requeueing when I_SYNC set to writeback_sb_inodes()
writeback: Move clearing of I_SYNC into inode_sync_complete()
writeback: initialize global_dirty_limit
fs: remove 8 bytes of padding from struct writeback_control on 64 bit builds
mm: page-writeback.c: local functions should not be exposed globally
To enable easy tracing of the location of log forces and the
frequency of them via perf, add a pair of trace points to the log
force functions. This will help debug where excessive log forces
are being issued from by simple perf commands like:
# ~/perf/perf top -e xfs:xfs_log_force -G -U
Which gives this sort of output:
Events: 141 xfs:xfs_log_force
- 100.00% [kernel] [k] xfs_log_force
- xfs_log_force
87.04% xfsaild
kthread
kernel_thread_helper
- 12.87% xfs_buf_lock
_xfs_buf_find
xfs_buf_get
xfs_trans_get_buf
xfs_da_do_buf
xfs_da_get_buf
xfs_dir2_data_init
xfs_dir2_leaf_addname
xfs_dir_createname
xfs_create
xfs_vn_mknod
xfs_vn_create
vfs_create
do_last.isra.41
path_openat
do_filp_open
do_sys_open
sys_open
system_call_fastpath
Signed-off-by: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Mark Tinguely <tinguely@sgi.com>
Signed-off-by: Ben Myers <bpm@sig.com>
Note xfs_iget can be called while holding a locked agi buffer. If
it goes into memory reclaim then inode teardown may try to lock the
same buffer. Prevent the deadlock by calling radix_tree_preload
with GFP_NOFS.
Signed-off-by: Peter Watkins <treestem@gmail.com>
Reviewed-by: Dave Chinner <dchinner@redhat.com>
Signed-off-by: Ben Myers <bpm@sgi.com>
xfstest 270 was causing quota reservations way beyond what was sane
(ten to hundreds of TB) for a 4GB filesystem. There's a sign problem
in the error handling path of xfs_bmapi_reserve_delalloc() because
xfs_trans_unreserve_quota_nblks() simple negates the value passed -
which doesn't work for an unsigned variable. This causes
reservations of close to 2^32 block instead of removing a
reservation of a handful of blocks.
Fix the same problem in the other xfs_trans_unreserve_quota_nblks()
callers where unsigned integer variables are used, too.
Signed-off-by: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Eric Sandeen <sandeen@redhat.com>
Signed-off-by: Ben Myers <bpm@sgi.com>
xfs_sync_worker checks the MS_ACTIVE flag in s_flags to avoid doing
work during mount and unmount. This flag can be cleared by unmount
after the xfs_sync_worker checks it but before the work is completed.
The has caused crashes in the completion handler for the dummy
transaction commited by xfs_sync_worker:
PID: 27544 TASK: ffff88013544e040 CPU: 3 COMMAND: "kworker/3:0"
#0 [ffff88016fdff930] machine_kexec at ffffffff810244e9
#1 [ffff88016fdff9a0] crash_kexec at ffffffff8108d053
#2 [ffff88016fdffa70] oops_end at ffffffff813ad1b8
#3 [ffff88016fdffaa0] no_context at ffffffff8102bd48
#4 [ffff88016fdffaf0] __bad_area_nosemaphore at ffffffff8102c04d
#5 [ffff88016fdffb40] bad_area_nosemaphore at ffffffff8102c12e
#6 [ffff88016fdffb50] do_page_fault at ffffffff813afaee
#7 [ffff88016fdffc60] page_fault at ffffffff813ac635
[exception RIP: xlog_get_lowest_lsn+0x30]
RIP: ffffffffa04a9910 RSP: ffff88016fdffd10 RFLAGS: 00010246
RAX: ffffc90014e48000 RBX: ffff88014d879980 RCX: ffff88014d879980
RDX: ffff8802214ee4c0 RSI: 0000000000000000 RDI: 0000000000000000
RBP: ffff88016fdffd10 R8: ffff88014d879a80 R9: 0000000000000000
R10: 0000000000000001 R11: 0000000000000000 R12: ffff8802214ee400
R13: ffff88014d879980 R14: 0000000000000000 R15: ffff88022fd96605
ORIG_RAX: ffffffffffffffff CS: 0010 SS: 0018
#8 [ffff88016fdffd18] xlog_state_do_callback at ffffffffa04aa186 [xfs]
#9 [ffff88016fdffd98] xlog_state_done_syncing at ffffffffa04aa568 [xfs]
Protect xfs_sync_worker by using the s_umount semaphore at the read
level to provide exclusion with unmount while work is progressing.
Reviewed-by: Mark Tinguely <tinguely@sgi.com>
Signed-off-by: Ben Myers <bpm@sgi.com>
This patch adds lseek(2) SEEK_DATA/SEEK_HOLE functionality to xfs.
Signed-off-by: Jie Liu <jeff.liu@oracle.com>
Reviewed-by: Mark Tinguely <tinguely@sgi.com>
Signed-off-by: Ben Myers <bpm@sgi.com>
Commit e459df5, 'xfs: move busy extent handling to it's own file'
moved some code from xfs_alloc.c into xfs_extent_busy.c for
convenience in userspace code merges. One of the functions moved is
xfs_extent_busy_trim (formerly xfs_alloc_busy_trim) which is defined
STATIC. Unfortunately this function is still used in xfs_alloc.c, and
this results in an undefined symbol in xfs.ko.
Make xfs_extent_busy_trim not static and add its prototype to
xfs_extent_busy.h.
Signed-off-by: Ben Myers <bpm@sgi.com>
Reviewed-by: Mark Tinguely <tinguely@sgi.com>
Rather than specifying XBF_MAPPED for almost all buffers, introduce
XBF_UNMAPPED for the couple of users that use unmapped buffers.
Signed-off-by: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Mark Tinguely <tinguely@sgi.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Ben Myers <bpm@sgi.com>
When we fail to mount the log in xfs_mountfs(), we tear down all the
infrastructure we have already allocated. However, the process of
mounting the log may have progressed to the point of reading,
caching and modifying buffers in memory. Hence before we can free
all the infrastructure, we have to flush and remove all the buffers
from memory.
Problem first reported by Eric Sandeen, later a different incarnation
was reported by Ben Myers.
Signed-off-by: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Mark Tinguely <tinguely@sgi.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Ben Myers <bpm@sgi.com>
Recent event tracing during a debugging session showed that flags
that define the IO type for a buffer are leaking into the flags on
the buffer incorrectly. Fix the flag exclusion mask in
xfs_buf_alloc() to avoid problems that may be caused by such
leakage.
Signed-off-by: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Mark Tinguely <tinguely@sgi.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Ben Myers <bpm@sgi.com>
With the removal of xfs_rw.h and other changes over time, xfs_bit.h
is being included in many files that don't actually need it. Clean
up the includes as necessary.
Also move the only-used-once xfs_ialloc_find_free() static inline
function out of a header file that is widely included to reduce
the number of needless dependencies on xfs_bit.h.
Signed-off-by: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Mark Tinguely <tinguely@sgi.com>
Signed-off-by: Ben Myers <bpm@sgi.com>
xfs_do_force_shutdown now is the only thing in xfs_rw.c. There is no
need to keep it in it's own file anymore, so move it to xfs_fsops.c
next to xfs_fs_goingdown() and kill xfs_rw.c.
Signed-off-by: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Mark Tinguely <tinguely@sgi.com>
Signed-off-by: Ben Myers <bpm@sgi.com>
The only thing left in xfs_rw.h is a function prototype for an inode
function. Move that to xfs_inode.h, and kill xfs_rw.h.
Also move the function implementing the prototype from xfs_rw.c to
xfs_inode.c so we only have one function left in xfs_rw.c
Signed-off-by: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Mark Tinguely <tinguely@sgi.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Ben Myers <bpm@sgi.com>
This is the only remaining useful function in xfs_rw.h, so move it
to a header file responsible for block mapping functions that the
callers already include. Soon we can get rid of xfs_rw.h.
Signed-off-by: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Mark Tinguely <tinguely@sgi.com>
Signed-off-by: Ben Myers <bpm@sgi.com>
Now that the busy extent tracking has been moved out of the
allocation files, clean up the namespace it uses to
"xfs_extent_busy" rather than a mix of "xfs_busy" and
"xfs_alloc_busy".
Signed-off-by: Dave Chinner<dchinner@redhat.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Mark Tinguely <tinguely@sgi.com>
Signed-off-by: Ben Myers <bpm@sgi.com>
To make it easier to handle userspace code merges, move all the busy
extent handling out of the allocation code and into it's own file.
The userspace code does not need the busy extent code, so this
simplifies the merging of the kernel code into the userspace
xfsprogs library.
Because the busy extent code has been almost completely rewritten
over the past couple of years, also update the copyright on this new
file to include the authors that made all those changes.
Signed-off-by: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Mark Tinguely <tinguely@sgi.com>
Signed-off-by: Ben Myers <bpm@sgi.com>
Untangle the header file includes a bit by moving the definition of
xfs_agino_t to xfs_types.h. This removes the dependency that xfs_ag.h has on
xfs_inum.h, meaning we don't need to include xfs_inum.h everywhere we include
xfs_ag.h.
Signed-off-by: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Mark Tinguely <tinguely@sgi.com>
Signed-off-by: Ben Myers <bpm@sgi.com>
fsstress has a particular effective way of stopping debug XFS
kernels. We keep seeing assert failures due finding delayed
allocation extents where there should be none. This shows up when
extracting extent maps and we are holding all the locks we should be
to prevent races, so this really makes no sense to see these errors.
After checking that fsstress does not use mmap, it occurred to me
that fsstress uses something that no sane application uses - the
XFS_IOC_ALLOCSP ioctl interfaces for preallocation. These interfaces
do allocation of blocks beyond EOF without using preallocation, and
then call setattr to extend and zero the allocated blocks.
THe problem here is this is a buffered write, and hence the
allocation is a delayed allocation. Unlike the buffered IO path, the
allocation and zeroing are not serialised using the IOLOCK. Hence
the ALLOCSP operation can race with operations holding the iolock to
prevent buffered IO operations from occurring.
Signed-off-by: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Mark Tinguely <tinguely@sgi.com>
Signed-off-by: Ben Myers <bpm@sgi.com>
Just about all callers of xfs_buf_read() and xfs_buf_get() use XBF_DONTBLOCK.
This is used to make memory allocation use GFP_NOFS rather than GFP_KERNEL to
avoid recursion through memory reclaim back into the filesystem.
All the blocking get calls in growfs occur inside a transaction, even though
they are no part of the transaction, so all allocation will be GFP_NOFS due to
the task flag PF_TRANS being set. The blocking read calls occur during log
recovery, so they will probably be unaffected by converting to GFP_NOFS
allocations.
Hence make XBF_DONTBLOCK behaviour always occur for buffers and kill the flag.
Signed-off-by: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Mark Tinguely <tinguely@sgi.com>
Signed-off-by: Ben Myers <bpm@sgi.com>
xfs_read_buf() is effectively the same as xfs_trans_read_buf() when called
outside a transaction context. The error handling is slightly different in that
xfs_read_buf stales the errored buffer it gets back, but there is probably good
reason for xfs_trans_read_buf() for doing this.
Hence update xfs_trans_read_buf() to the same error handling as xfs_read_buf(),
and convert all the callers of xfs_read_buf() to use the former function. We can
then remove xfs_read_buf().
Signed-off-by: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Mark Tinguely <tinguely@sgi.com>
Signed-off-by: Ben Myers <bpm@sgi.com>
Buffers are always returned locked from the lookup routines. Hence
we don't need to tell the lookup routines to return locked buffers,
on to try and lock them. Remove XBF_LOCK from all the callers and
from internal buffer cache usage.
Signed-off-by: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Mark Tinguely <tinguely@sgi.com>
Signed-off-by: Ben Myers <bpm@sgi.com>
xfs_buf_btoc and friends are simple macros that do basic block
to page index conversion and vice versa. These aren't widely used,
and we use open coded masking and shifting everywhere else. Hence
remove the macros and open code the work they do.
Also, use of PAGE_CACHE_{SIZE|SHIFT|MASK} for these macros is now
incorrect - we are using pages directly and not the page cache, so
use PAGE_{SIZE|MASK|SHIFT} instead.
Signed-off-by: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Mark Tinguely <tinguely@sgi.com>
Signed-off-by: Ben Myers <bpm@sgi.com>
Now that we pass block counts everywhere, and index buffers by block
number and length in units of blocks, convert the desired IO size
into block counts rather than bytes. Convert the code to use block
counts, and those that need byte counts get converted at the time of
use.
Rename the b_desired_count variable to something closer to it's
purpose - b_io_length - as it is only used to specify the length of
an IO for a subset of the buffer. The only time this is used is for
log IO - both writing iclogs and during log recovery. In all other
cases, the b_io_length matches b_length, and hence a lot of code
confuses the two. e.g. the buf item code uses the io count
exclusively when it should be using the buffer length. Fix these
apprpriately as they are found.
Also, remove the XFS_BUF_{SET_}COUNT() macros that are just wrappers
around the desired IO length. They only serve to make the code
shouty loud, don't actually add any real value, and are often used
incorrectly.
Signed-off-by: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Mark Tinguely <tinguely@sgi.com>
Signed-off-by: Ben Myers <bpm@sgi.com>
Now that we pass block counts everywhere, and index buffers by block
number, track the length of the buffer in units of blocks rather
than bytes. Convert the code to use block counts, and those that
need byte counts get converted at the time of use.
Also, remove the XFS_BUF_{SET_}SIZE() macros that are just wrappers
around the buffer length. They only serve to make the code shouty
loud and don't actually add any real value.
Signed-off-by: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Mark Tinguely <tinguely@sgi.com>
Signed-off-by: Ben Myers <bpm@sgi.com>
Seeing as we pass block numbers around everywhere in the buffer
cache now, it makes no sense to index everything by byte offset.
Replace all the byte offset indexing with block number based
indexing, and replace all uses of the byte offset with direct
conversion from the block index.
Signed-off-by: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Mark Tinguely <tinguely@sgi.com>
Signed-off-by: Ben Myers <bpm@sgi.com>
The xfs_buf_get/read API is not consistent in the units it uses, and
does not use appropriate or consistent units/types for the
variables.
Convert the API to use disk addresses and block counts for all
buffer get and read calls. Use consistent naming for all the
functions and their declarations, and convert the internal functions
to use disk addresses and block counts to avoid need to convert them
from one type to another and back again.
Fix all the callers to use disk addresses and block counts. In many
cases, this removes an additional conversion from the function call
as the callers already have a block count.
Signed-off-by: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Mark Tinguely <tinguely@sgi.com>
Signed-off-by: Ben Myers <bpm@sgi.com>
To replace the alloc/memset pair.
Signed-off-by: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Mark Tinguely <tinguely@sgi.com>
Signed-off-by: Ben Myers <bpm@sgi.com>
Because we no longer use the page cache for buffering, there is no
direct block number to page offset relationship anymore.
xfs_buf_get_pages is still setting up b_offset as if there was some
relationship, and that is leading to incorrectly setting up
*uncached* buffers that don't overwrite b_offset once they've had
pages allocated.
For cached buffers, the first block of the buffer is always at offset
zero into the allocated memory. This is true for sub-page sized
buffers, as well as for multiple-page buffers.
For uncached buffers, b_offset is only non-zero when we are
associating specific memory to the buffers, and that is set
correctly by the code setting up the buffer.
Hence remove the setting of b_offset in xfs_buf_get_pages, because
it is now always the wrong thing to do.
Signed-off-by: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Mark Tinguely <tinguely@sgi.com>
Signed-off-by: Ben Myers <bpm@sgi.com>
If we call xfs_buf_iowait() on a buffer that failed dispatch due to
an IO error, it will wait forever for an Io that does not exist.
This is hndled in xfs_buf_read, but there is other code that calls
xfs_buf_iowait directly that doesn't.
Rather than make the call sites have to handle checking for dispatch
errors and then checking for completion errors, make
xfs_buf_iowait() check for dispatch errors on the buffer before
waiting. This means we handle both dispatch and completion errors
with one set of error handling at the caller sites.
Signed-off-by: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Mark Tinguely <tinguely@sgi.com>
Signed-off-by: Ben Myers <bpm@sgi.com>
When memory allocation fails to add the page array or tht epages to
a buffer during xfs_buf_get(), the buffer is left in the cache in a
partially initialised state. There is enough state left for the next
lookup on that buffer to find the buffer, and for the buffer to then
be used without finishing the initialisation. As a result, when an
attempt to do IO on the buffer occurs, it fails with EIO because
there are no pages attached to the buffer.
We cannot remove the buffer from the cache immediately and free it,
because there may already be a racing lookup that is blocked on the
buffer lock. Hence the moment we unlock the buffer to then free it,
the other user is woken and we have a use-after-free situation.
To avoid this race condition altogether, allocate the pages for the
buffer before we insert it into the cache. This then means that we
don't have an allocation failure case to deal after the buffer is
already present in the cache, and hence avoid the problem
altogether. In most cases we won't have racing inserts for the same
buffer, and so won't increase the memory pressure allocation before
insertion may entail.
Signed-off-by: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Mark Tinguely <tinguely@sgi.com>
Signed-off-by: Ben Myers <bpm@sgi.com>
xfstest 229 exposes a problem with buffered IO, delayed allocation
and extent size hints. That is when we do delayed allocation during
buffered IO, we reserve space for the extent size hint alignment and
allocate the physical space to align the extent, but we do not zero
the regions of the extent that aren't written by the write(2)
syscall. The result is that we expose stale data in unwritten
regions of the extent size hints.
There are two ways to fix this. The first is to detect that we are
doing unaligned writes, check if there is already a mapping or data
over the extent size hint range, and if not zero the page cache
first before then doing the real write. This can be very expensive
for large extent size hints, especially if the subsequent writes
fill then entire extent size before the data is written to disk.
The second, and simpler way, is simply to turn off delayed
allocation when the extent size hint is set and use preallocation
instead. This results in unwritten extents being laid down on disk
and so only the written portions will be converted. This matches the
behaviour for direct IO, and will also work for the real time
device. The disadvantage of this approach is that for small extent
size hints we can get file fragmentation, but in general extent size
hints are fairly large (e.g. stripe width sized) so this isn't a big
deal.
Implement the second approach as it is simple and effective.
Signed-off-by: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Mark Tinguely <tinguely@sgi.com>
Signed-off-by: Ben Myers <bpm@sgi.com>
Speculative delayed allocation beyond EOF near the maximum supported
file offset can result in creating delalloc extents beyond
mp->m_maxioffset (8EB). These can never be trimmed during
xfs_free_eof_blocks() because they are beyond mp->m_maxioffset, and
that results in assert failures in xfs_fs_destroy_inode() due to
delalloc blocks still being present. xfstests 071 exposes this
problem.
Limit speculative delalloc to mp->m_maxioffset to avoid this
problem.
Signed-off-by: Dave Chinner <dchinner@redhat.com>
Signed-off-by: Ben Myers <bpm@sgi.com>
When we are doing speculative delayed allocation beyond EOF,
conversion of the region allocated beyond EOF is dependent on the
largest free space extent available. If the largest free extent is
smaller than the delalloc range, then after allocation we leave
a delalloc extent that starts beyond EOF. This extent cannot *ever*
be converted by flushing data, and so will remain there until either
the EOF moves into the extent or it is truncated away.
Hence if xfs_getbmap() runs on such an inode and is asked to return
extents beyond EOF, it will assert fail on this extent even though
there is nothing xfs_getbmap() can do to convert it to a real
extent. Hence we should simply report these delalloc extents rather
than assert that there should be none.
Signed-off-by: Dave Chinner <dchinner@redhat.com>
Signed-off-by: Ben Myers <bpm@sgi.com>
Often mounting small filesystem with small logs will emit a warning
such as:
XFS (vdb): Invalid block length (0x2000) for buffer
during log recovery. This causes tests to randomly fail because this
output causes the clean filesystem checks on test completion to
think the filesystem is inconsistent.
The cause of the error is simply that log recovery is asking for a
buffer size that is larger than the log when zeroing the tail. This
is because the buffer size is rounded up, and if the right head and
tail conditions exist then the buffer size can be larger than the log.
Limit the variable size xlog_get_bp() callers to requesting buffers
smaller than the log.
Signed-off-by: Dave Chinner <dchinner@redhat.com>
Signed-off-by: Ben Myers <bpm@sgi.com>
When a partial write inside EOF fails, it can leave delayed
allocation blocks lying around because they don't get punched back
out. This leads to assert failures like:
XFS: Assertion failed: XFS_FORCED_SHUTDOWN(ip->i_mount) || ip->i_delayed_blks == 0, file: fs/xfs/xfs_super.c, line: 847
when evicting inodes from the cache. This can be trivially triggered
by xfstests 083, which takes between 5 and 15 executions on a 512
byte block size filesystem to trip over this. Debugging shows a
failed write due to ENOSPC calling xfs_vm_write_failed such as:
[ 5012.329024] ino 0xa0026: vwf to 0x17000, sze 0x1c85ae
and no action is taken on it. This leaves behind a delayed
allocation extent that has no page covering it and no data in it:
[ 5015.867162] ino 0xa0026: blks: 0x83 delay blocks 0x1, size 0x2538c0
[ 5015.868293] ext 0: off 0x4a, fsb 0x50306, len 0x1
[ 5015.869095] ext 1: off 0x4b, fsb 0x7899, len 0x6b
[ 5015.869900] ext 2: off 0xb6, fsb 0xffffffffe0008, len 0x1
^^^^^^^^^^^^^^^
[ 5015.871027] ext 3: off 0x36e, fsb 0x7a27, len 0xd
[ 5015.872206] ext 4: off 0x4cf, fsb 0x7a1d, len 0xa
So the delayed allocation extent is one block long at offset
0x16c00. Tracing shows that a bigger write:
xfs_file_buffered_write: size 0x1c85ae offset 0x959d count 0x1ca3f ioflags
allocates the block, and then fails with ENOSPC trying to allocate
the last block on the page, leading to a failed write with stale
delalloc blocks on it.
Because we've had an ENOSPC when trying to allocate 0x16e00, it
means that we are never goinge to call ->write_end on the page and
so the allocated new buffer will not get marked dirty or have the
buffer_new state cleared. In other works, what the above write is
supposed to end up with is this mapping for the page:
+------+------+------+------+------+------+------+------+
UMA UMA UMA UMA UMA UMA UND FAIL
where: U = uptodate
M = mapped
N = new
A = allocated
D = delalloc
FAIL = block we ENOSPC'd on.
and the key point being the buffer_new() state for the newly
allocated delayed allocation block. Except it doesn't - we're not
marking buffers new correctly.
That buffer_new() problem goes back to the xfs_iomap removal days,
where xfs_iomap() used to return a "new" status for any map with
newly allocated blocks, so that __xfs_get_blocks() could call
set_buffer_new() on it. We still have the "new" variable and the
check for it in the set_buffer_new() logic - except we never set it
now!
Hence that newly allocated delalloc block doesn't have the new flag
set on it, so when the write fails we cannot tell which blocks we
are supposed to punch out. WHy do we need the buffer_new flag? Well,
that's because we can have this case:
+------+------+------+------+------+------+------+------+
UMD UMD UMD UMD UMD UMD UND FAIL
where all the UMD buffers contain valid data from a previously
successful write() system call. We only want to punch the UND buffer
because that's the only one that we added in this write and it was
only this write that failed.
That implies that even the old buffer_new() logic was wrong -
because it would result in all those UMD buffers on the page having
set_buffer_new() called on them even though they aren't new. Hence
we shoul donly be calling set_buffer_new() for delalloc buffers that
were allocated (i.e. were a hole before xfs_iomap_write_delay() was
called).
So, fix this set_buffer_new logic according to how we need it to
work for handling failed writes correctly. Also, restore the new
buffer logic handling for blocks allocated via
xfs_iomap_write_direct(), because it should still set the buffer_new
flag appropriately for newly allocated blocks, too.
SO, now we have the buffer_new() being set appropriately in
__xfs_get_blocks(), we can detect the exact delalloc ranges that
we allocated in a failed write, and hence can now do a walk of the
buffers on a page to find them.
Except, it's not that easy. When block_write_begin() fails, it
unlocks and releases the page that we just had an error on, so we
can't use that page to handle errors anymore. We have to get access
to the page while it is still locked to walk the buffers. Hence we
have to open code block_write_begin() in xfs_vm_write_begin() to be
able to insert xfs_vm_write_failed() is the right place.
With that, we can pass the page and write range to
xfs_vm_write_failed() and walk the buffers on the page, looking for
delalloc buffers that are either new or beyond EOF and punch them
out. Handling buffers beyond EOF ensures we still handle the
existing case that xfs_vm_write_failed() handles.
Of special note is the truncate_pagecache() handling - that only
should be done for pages outside EOF - pages within EOF can still
contain valid, dirty data so we must not punch them out of the
cache.
That just leaves the xfs_vm_write_end() failure handling.
The only failure case here is that we didn't copy the entire range,
and generic_write_end() handles that by zeroing the region of the
page that wasn't copied, we don't have to punch out blocks within
the file because they are guaranteed to contain zeros. Hence we only
have to handle the existing "beyond EOF" case and don't need access
to the buffers on the page. Hence it remains largely unchanged.
Note that xfs_getbmap() can still trip over delalloc blocks beyond
EOF that are left there by speculative delayed allocation. Hence
this bug fix does not solve all known issues with bmap vs delalloc,
but it does fix all the the known accidental occurances of the
problem.
Signed-off-by: Dave Chinner <david@fromorbit.com>
Signed-off-by: Ben Myers <bpm@sgi.com>
xfs_is_delayed_page() checks to see if a page has buffers matching
the given IO type passed in. It does so by walking the buffer heads
on the page and checking if the state flags match the IO type.
However, the "acceptable" variable that is calculated is overwritten
every time a new buffer is checked. Hence if the first buffer on the
page is of the right type, this state is lost if the second buffer
is not of the correct type. This means that xfs_aops_discard_page()
may not discard delalloc regions when it is supposed to, and
xfs_convert_page() may not cluster IO as efficiently as possible.
This problem only occurs on filesystems with a block size smaller
than page size.
Also, rename xfs_is_delayed_page() to xfs_check_page_type() to
better describe what it is doing - it is not delalloc specific
anymore.
The problem was first noticed by Peter Watkins.
Signed-off-by: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Mark Tinguely <tinguely@sgi.com>
Signed-off-by: Ben Myers <bpm@sgi.com>
Doing background CIL flushes adds significant latency to whatever
async transaction that triggers it. To avoid blocking async
transactions on things like waiting for log buffer IO to complete,
move the CIL push off into a workqueue. By moving the push work
into a workqueue, we remove all the latency that the commit adds
from the foreground transaction commit path. This also means that
single threaded workloads won't do the CIL push procssing, leaving
them more CPU to do more async transactions.
To do this, we need to keep track of the sequence number we have
pushed work for. This avoids having many transaction commits
attempting to schedule work for the same sequence, and ensures that
we only ever have one push (background or forced) in progress at a
time. It also means that we don't need to take the CIL lock in write
mode to check for potential background push races, which reduces
lock contention.
To avoid potential issues with "smart" IO schedulers, don't use the
workqueue for log force triggered flushes. Instead, do them directly
so that the log IO is done directly by the process issuing the log
force and so doesn't get stuck on IO elevator queue idling
incorrectly delaying the log IO from the workqueue.
Signed-off-by: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Mark Tinguely <tinguely@sgi.com>
Signed-off-by: Ben Myers <bpm@sgi.com>
xfs_trans_ail_delete_bulk() can be called from different contexts so
if the item is not in the AIL we need different shutdown for each
context. Pass in the shutdown method needed so the correct action
can be taken.
Signed-off-by: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Mark Tinguely <tinguely@sgi.com>
Signed-off-by: Ben Myers <bpm@sgi.com>
Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Mark Tinguely <tinguely@sgi.com>
Signed-off-by: Ben Myers <bpm@sgi.com>
Queue delwri buffers on a local on-stack list instead of a per-buftarg one,
and write back the buffers per-process instead of by waking up xfsbufd.
This is now easily doable given that we have very few places left that write
delwri buffers:
- log recovery:
Only done at mount time, and already forcing out the buffers
synchronously using xfs_flush_buftarg
- quotacheck:
Same story.
- dquot reclaim:
Writes out dirty dquots on the LRU under memory pressure. We might
want to look into doing more of this via xfsaild, but it's already
more optimal than the synchronous inode reclaim that writes each
buffer synchronously.
- xfsaild:
This is the main beneficiary of the change. By keeping a local list
of buffers to write we reduce latency of writing out buffers, and
more importably we can remove all the delwri list promotions which
were hitting the buffer cache hard under sustained metadata loads.
The implementation is very straight forward - xfs_buf_delwri_queue now gets
a new list_head pointer that it adds the delwri buffers to, and all callers
need to eventually submit the list using xfs_buf_delwi_submit or
xfs_buf_delwi_submit_nowait. Buffers that already are on a delwri list are
skipped in xfs_buf_delwri_queue, assuming they already are on another delwri
list. The biggest change to pass down the buffer list was done to the AIL
pushing. Now that we operate on buffers the trylock, push and pushbuf log
item methods are merged into a single push routine, which tries to lock the
item, and if possible add the buffer that needs writeback to the buffer list.
This leads to much simpler code than the previous split but requires the
individual IOP_PUSH instances to unlock and reacquire the AIL around calls
to blocking routines.
Given that xfsailds now also handle writing out buffers, the conditions for
log forcing and the sleep times needed some small changes. The most
important one is that we consider an AIL busy as long we still have buffers
to push, and the other one is that we do increment the pushed LSN for
buffers that are under flushing at this moment, but still count them towards
the stuck items for restart purposes. Without this we could hammer on stuck
items without ever forcing the log and not make progress under heavy random
delete workloads on fast flash storage devices.
[ Dave Chinner:
- rebase on previous patches.
- improved comments for XBF_DELWRI_Q handling
- fix XBF_ASYNC handling in queue submission (test 106 failure)
- rename delwri submit function buffer list parameters for clarity
- xfs_efd_item_push() should return XFS_ITEM_PINNED ]
Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Mark Tinguely <tinguely@sgi.com>
Signed-off-by: Ben Myers <bpm@sgi.com>
Instead of adding buffers to the delwri list as soon as they are logged,
even if they can't be written until commited because they are pinned
defer adding them to the delwri list until xfsaild pushes them. This
makes the code more similar to other log items and prepares for writing
buffers directly from xfsaild.
The complication here is that we need to fail buffers that were added
but not logged yet in xfs_buf_item_unpin, borrowing code from
xfs_bioerror.
Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Mark Tinguely <tinguely@sgi.com>
Signed-off-by: Ben Myers <bpm@sgi.com>
Instead of writing the buffer directly from inside xfs_qm_dqflush return it
to the caller and let the caller decide what to do with the buffer. Also
remove the pincount check in xfs_qm_dqflush that all non-blocking callers
already implement and the now unused flags parameter and the XFS_DQ_IS_DIRTY
check that all callers already perform.
[ Dave Chinner: fixed build error cause by missing '{'. ]
Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Mark Tinguely <tinguely@sgi.com>
Signed-off-by: Ben Myers <bpm@sgi.com>
Instead of writing the buffer directly from inside xfs_iflush return it to
the caller and let the caller decide what to do with the buffer. Also
remove the pincount check in xfs_iflush that all non-blocking callers already
implement and the now unused flags parameter.
Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Mark Tinguely <tinguely@sgi.com>
Signed-off-by: Ben Myers <bpm@sgi.com>
We already flush dirty inodes throug the AIL regularly, there is no reason
to have second thread compete with it and disturb the I/O pattern. We still
do write inodes when doing a synchronous reclaim from the shrinker or during
unmount for now.
Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Mark Tinguely <tinguely@sgi.com>
Signed-off-by: Ben Myers <bpm@sgi.com>
