Since Linux 2.6.36 the writeback code has introduces various measures for
live lock prevention during sync(). Unfortunately some of these are
actively harmful for the XFS model, where the inode gets marked dirty for
metadata from the data I/O handler.
The older_than_this checks that are now more strictly enforced since
writeback: avoid livelocking WB_SYNC_ALL writeback
by only calling into __writeback_inodes_sb and thus only sampling the
current cut off time once. But on a slow enough devices the previous
asynchronous sync pass might not have fully completed yet, and thus XFS
might mark metadata dirty only after that sampling of the cut off time for
the blocking pass already happened. I have not myself reproduced this
myself on a real system, but by introducing artificial delay into the
XFS I/O completion workqueues it can be reproduced easily.
Fix this by iterating over all XFS inodes in ->sync_fs and log all that
are dirty. This might log inode that only got redirtied after the
previous pass, but given how cheap delayed logging of inodes is it
isn't a major concern for performance.
Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Dave Chinner <dchinner@redhat.com>
Tested-by: Mark Tinguely <tinguely@sgi.com>
Reviewed-by: Mark Tinguely <tinguely@sgi.com>
Signed-off-by: Ben Myers <bpm@sgi.com>
If the writeback code writes back an inode because it has expired we currently
use the non-blockin ->write_inode path. This means any inode that is pinned
is skipped. With delayed logging and a workload that has very little log
traffic otherwise it is very likely that an inode that gets constantly
written to is always pinned, and thus we keep refusing to write it. The VM
writeback code at that point redirties it and doesn't try to write it again
for another 30 seconds. This means under certain scenarious time based
metadata writeback never happens.
Fix this by calling into xfs_log_inode for kupdate in addition to data
integrity syncs, and thus transfer the inode to the log ASAP.
Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Dave Chinner <dchinner@redhat.com>
Tested-by: Mark Tinguely <tinguely@sgi.com>
Reviewed-by: Mark Tinguely <tinguely@sgi.com>
Signed-off-by: Ben Myers <bpm@sgi.com>
Apply the scheme used in log_regrant_write_log_space to wake up any other
threads waiting for log space before the newly added one to
log_regrant_write_log_space as well, and factor the code into readable
helpers. For each of the queues we have add two helpers:
- one to try to wake up all waiting threads. This helper will also be
usable by xfs_log_move_tail once we remove the current opportunistic
wakeups in it.
- one to sleep on t_wait until enough log space is available, loosely
modelled after Linux waitqueues.
And use them to reimplement the guts of log_regrant_write_log_space and
log_regrant_write_log_space. These two function now use one and the same
algorithm for waiting on log space instead of subtly different ones before,
with an option to completely unify them in the near future.
Also move the filesystem shutdown handling to the common caller given
that we had to touch it anyway.
Based on hard debugging and an earlier patch from
Chandra Seetharaman <sekharan@us.ibm.com>.
Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Chandra Seetharaman <sekharan@us.ibm.com>
Tested-by: Chandra Seetharaman <sekharan@us.ibm.com>
Signed-off-by: Ben Myers <bpm@sgi.com>
The i_ino field in the VFS inode is of type unsigned long and thus can't
hold the full 64-bit inode number on 32-bit kernels. We have the full
inode number in the XFS inode, so use that one for nfs exports. Note
that I've also switched the 32-bit file handles types to it, just to make
the code more consistent and copy & paste errors less likely to happen.
Reported-by: Guoquan Yang <ygq51@hotmail.com>
Reported-by: Hank Peng <pengxihan@gmail.com>
Signed-off-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Ben Myers <bpm@sgi.com>
When testing the new xfstests --large-fs option that does very large
file preallocations, this assert was tripped deep in
xfs_alloc_vextent():
XFS: Assertion failed: args->minlen <= args->maxlen, file: fs/xfs/xfs_alloc.c, line: 2239
The allocation was trying to allocate a zero length extent because
the lower 32 bits of the allocation length was zero. The remaining
length of the allocation to be done was an exact multiple of 2^32 -
the first case I saw was at 496TB remaining to be allocated.
This turns out to be an overflow when converting the allocation
length (a 64 bit quantity) into the extent length to allocate (a 32
bit quantity), and it requires the length to be allocated an exact
multiple of 2^32 blocks to trip the assert.
Fix it by limiting the extent lenth to allocate to MAXEXTLEN.
Signed-off-by: Dave Chinner <dchinner@redhat.com>
Signed-off-by: Ben Myers <bpm@sgi.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
With Dmitry fsstress updates I've seen very reproducible crashes in
xfs_attr_shortform_remove because xfs_attr_shortform_bytesfit claims that
the attributes would not fit inline into the inode after removing an
attribute. It turns out that we were operating on an inode with lots
of delalloc extents, and thus an if_bytes values for the data fork that
is larger than biggest possible on-disk storage for it which utterly
confuses the code near the end of xfs_attr_shortform_bytesfit.
Fix this by always allowing the current attribute fork, like we already
do for the attr1 format, given that delalloc conversion will take care
for moving either the data or attribute area out of line if it doesn't
fit at that point - or making the point moot by merging extents at this
point.
Also document the function better, and clean up some loose bits.
Reviewed-by: Dave Chinner <dchinner@redhat.com>
Signed-off-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Ben Myers <bpm@sgi.com>
If we are doing synchronous inode reclaim we block the VM from making
progress in memory reclaim. So if we encouter a flush locked inode
promote it in the delwri list and wake up xfsbufd to write it out now.
Without this we can get hangs of up to 30 seconds during workloads hitting
synchronous inode reclaim.
The scheme is copied from what we do for dquot reclaims.
Reported-by: Simon Kirby <sim@hostway.ca>
Signed-off-by: Christoph Hellwig <hch@lst.de>
Tested-by: Simon Kirby <sim@hostway.ca>
Signed-off-by: Ben Myers <bpm@sgi.com>
This prevents in-memory corruption and possible panics if the on-disk
ACL is badly corrupted.
Signed-off-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Ben Myers <bpm@sgi.com>
The doalloc arg in xfs_qm_dqattach_one() is a flag that indicates
whether a new area to handle quota information will be allocated
if needed. Originally, it was passed to xfs_qm_dqget(), but has
been removed by the following commit (probably by mistake):
commit 8e9b6e7fa4
Author: Christoph Hellwig <hch@lst.de>
Date: Sun Feb 8 21:51:42 2009 +0100
xfs: remove the unused XFS_QMOPT_DQLOCK flag
As the result, xfs_qm_dqget() called from xfs_qm_dqattach_one()
never allocates the new area even if it is needed.
This patch gives the doalloc arg to xfs_qm_dqget() in
xfs_qm_dqattach_one() to fix this problem.
Signed-off-by: Mitsuo Hayasaka <mitsuo.hayasaka.hu@hitachi.com>
Cc: Alex Elder <aelder@sgi.com>
Cc: Christoph Hellwig <hch@infradead.org>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Ben Myers <bpm@sgi.com>
Ensure ioend->io_error gets propagated back to e.g. AIO completions.
Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Alex Elder <aelder@sgi.com>
The log item ops aren't nessecarily the biggest exploit vector, but marking
them const is easy enough. Also remove the unused xfs_item_ops_t typedef
while we're at it.
Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Alex Elder <aelder@sgi.com>
Fixes a possible memory corruption when the link is larger than
MAXPATHLEN and XFS_DEBUG is not enabled. This also remove the
S_ISLNK assert, since the inode mode is checked previously in
xfs_readlink_by_handle() and via VFS.
Updated to address concerns raised by Ben Hutchings about the loose
attention paid to 32- vs 64-bit values, and the lack of handling a
potentially negative pathlen value:
- Changed type of "pathlen" to be xfs_fsize_t, to match that of
ip->i_d.di_size
- Added checking for a negative pathlen to the too-long pathlen
test, and generalized the message that gets reported in that case
to reflect the change
As a result, if a negative pathlen were encountered, this function
would return EFSCORRUPTED (and would fail an assertion for a debug
build)--just as would a too-long pathlen.
Signed-off-by: Alex Elder <aelder@sgi.com>
Signed-off-by: Carlos Maiolino <cmaiolino@redhat.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Replace remaining direct i_nlink updates with a new set_nlink()
updater function.
Signed-off-by: Miklos Szeredi <mszeredi@suse.cz>
Tested-by: Toshiyuki Okajima <toshi.okajima@jp.fujitsu.com>
Signed-off-by: Christoph Hellwig <hch@lst.de>
Direct reclaim should never writeback pages. For now, handle the
situation and warn about it. Ultimately, this will be a BUG_ON.
Signed-off-by: Mel Gorman <mgorman@suse.de>
Cc: Dave Chinner <david@fromorbit.com>
Cc: Christoph Hellwig <hch@infradead.org>
Cc: Johannes Weiner <jweiner@redhat.com>
Cc: Wu Fengguang <fengguang.wu@intel.com>
Cc: Jan Kara <jack@suse.cz>
Cc: Minchan Kim <minchan.kim@gmail.com>
Cc: Rik van Riel <riel@redhat.com>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Alex Elder <aelder@sgi.com>
Cc: Theodore Ts'o <tytso@mit.edu>
Cc: Chris Mason <chris.mason@oracle.com>
Cc: Dave Hansen <dave@linux.vnet.ibm.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
I intended to do this as part of fixing part of the conflict with
the merge with Linus' tree, but evidently it didn't get included in
the commit.
Signed-off-by: Alex Elder <aelder@sgi.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
When we call xfs_flush_buftarg (generally from sync or umount) it already
is too late to flush the data workqueues, as I/O completion is signalled
for them and we are thus already done with the data we would flush here.
There are places where flushing them might be useful, but the current
sync interface doesn't give us that opportunity.
Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Dave Chinner <dchinner@redhat.com>
Signed-off-by: Alex Elder <aelder@sgi.com>
The calling convention that returns a pointer to a static buffer is
fairly nasty, so just opencode it in the only caller that is left.
Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Dave Chinner <dchinner@redhat.com>
Signed-off-by: Alex Elder <aelder@sgi.com>
Use xfs_ioerror_alert instead of opencoding a very similar error
message.
Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Dave Chinner <dchinner@redhat.com>
Signed-off-by: Alex Elder <aelder@sgi.com>
Instead of passing the block number and mount structure explicitly
get them off the bp and fix make the argument order more natural.
Also move it to xfs_buf.c and stop printing the device name given
that we already get the fs name as part of xfs_alert, and we know
what device is operates on because of the caller that gets printed,
finally rename it to xfs_buf_ioerror_alert and pass __func__ as
argument where it makes sense.
Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Dave Chinner <dchinner@redhat.com>
Signed-off-by: Alex Elder <aelder@sgi.com>
Change _xfs_buf_initialize to allocate the buffer directly and rename it to
xfs_buf_alloc now that is the only buffer allocation routine. Also remove
the xfs_buf_deallocate wrapper around the kmem_zone_free calls for buffers.
Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Dave Chinner <dchinner@redhat.com>
Signed-off-by: Alex Elder <aelder@sgi.com>
For each call to xfs_buf_stale we call xfs_buf_delwri_dequeue either
directly before or after it, or are guaranteed by the surrounding
conditionals that we are never called on delwri buffers. Simply
this situation by moving the call to xfs_buf_delwri_dequeue into
xfs_buf_stale.
Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Dave Chinner <dchinner@redhat.com>
Signed-off-by: Alex Elder <aelder@sgi.com>
The code is unused and under a config option that doesn't exist, remove it.
Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Dave Chinner <dchinner@redhat.com>
Signed-off-by: Alex Elder <aelder@sgi.com>
The code to flush buffers in the umount code is a bit iffy: we first
flush all delwri buffers out, but then might be able to queue up a
new one when logging the sb counts. On a normal shutdown that one
would get flushed out when doing the synchronous superblock write in
xfs_unmountfs_writesb, but we skip that one if the filesystem has
been shut down.
Fix this by moving the delwri list flushing until just before unmounting
the log, and while we're at it also remove the superflous delwri list
and buffer lru flusing for the rt and log device that can never have
cached or delwri buffers.
Signed-off-by: Christoph Hellwig <hch@lst.de>
Reported-by: Amit Sahrawat <amit.sahrawat83@gmail.com>
Tested-by: Amit Sahrawat <amit.sahrawat83@gmail.com>
Signed-off-by: Alex Elder <aelder@sgi.com>
Directories are only updated transactionally, which means fsync only
needs to flush the log the inode is currently dirty, but not bother
with checking for dirty data, non-transactional updates, and most
importanly doesn't have to flush disk caches except as part of a
transaction commit.
While the first two optimizations can't easily be measured, the
latter actually makes a difference when doing lots of fsync that do
not actually have to commit the inode, e.g. because an earlier fsync
already pushed the log far enough.
The new xfs_dir_fsync is identical to xfs_nfs_commit_metadata except
for the prototype, but I'm not sure creating a common helper for the
two is worth it given how simple the functions are.
Signed-off-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Alex Elder <aelder@sgi.com>
The AIL push code will issue a log force on ever single push loop
that it exits and has encountered pinned items. It doesn't rescan
these pinned items until it revisits the AIL from the start. Hence
we only need to force the log once per walk from the start of the
AIL to the target LSN.
This results in numbers like this:
xs_push_ail_flush..... 1456
xs_log_force......... 1485
For an 8-way 50M inode create workload - almost all the log forces
are coming from the AIL pushing code.
Reduce the number of log forces by only forcing the log if the
previous walk found pinned buffers. This reduces the numbers to:
xs_push_ail_flush..... 665
xs_log_force......... 682
For the same test.
Signed-off-by: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Alex Elder <aelder@sgi.com>
Stats show that for an 8-way unlink @ ~80,000 unlinks/s we are doing
~1 million cache hit lookups to ~3000 buffer creates. That's almost
3 orders of magnitude more cahce hits than misses, so optimising for
cache hits is quite important. In the cache hit case, we do not need
to allocate a new buffer in case of a cache miss, so we are
effectively hitting the allocator for no good reason for vast the
majority of calls to _xfs_buf_find. 8-way create workloads are
showing similar cache hit/miss ratios.
The result is profiles that look like this:
samples pcnt function DSO
_______ _____ _______________________________ _________________
1036.00 10.0% _xfs_buf_find [kernel.kallsyms]
582.00 5.6% kmem_cache_alloc [kernel.kallsyms]
519.00 5.0% __memcpy [kernel.kallsyms]
468.00 4.5% __ticket_spin_lock [kernel.kallsyms]
388.00 3.7% kmem_cache_free [kernel.kallsyms]
331.00 3.2% xfs_log_commit_cil [kernel.kallsyms]
Further, there is a fair bit of work involved in initialising a new
buffer once a cache miss has occurred and we currently do that under
the rbtree spinlock. That increases spinlock hold time on what are
heavily used trees.
To fix this, remove the initialisation of the buffer from
_xfs_buf_find() and only allocate the new buffer once we've had a
cache miss. Initialise the buffer immediately after allocating it in
xfs_buf_get, too, so that is it ready for insert if we get another
cache miss after allocation. This minimises lock hold time and
avoids unnecessary allocator churn. The resulting profiles look
like:
samples pcnt function DSO
_______ _____ ___________________________ _________________
8111.00 9.1% _xfs_buf_find [kernel.kallsyms]
4380.00 4.9% __memcpy [kernel.kallsyms]
4341.00 4.8% __ticket_spin_lock [kernel.kallsyms]
3401.00 3.8% kmem_cache_alloc [kernel.kallsyms]
2856.00 3.2% xfs_log_commit_cil [kernel.kallsyms]
2625.00 2.9% __kmalloc [kernel.kallsyms]
2380.00 2.7% kfree [kernel.kallsyms]
2016.00 2.3% kmem_cache_free [kernel.kallsyms]
Showing a significant reduction in time spent doing allocation and
freeing from slabs (kmem_cache_alloc and kmem_cache_free).
Signed-off-by: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Alex Elder <aelder@sgi.com>
There is no reason to keep a reference to the inode even if we unlock
it during transaction commit because we never drop a reference between
the ijoin and commit. Also use this fact to merge xfs_trans_ijoin_ref
back into xfs_trans_ijoin - the third argument decides if an unlock
is needed now.
I'm actually starting to wonder if allowing inodes to be unlocked
at transaction commit really is worth the effort. The only real
benefit is that they can be unlocked earlier when commiting a
synchronous transactions, but that could be solved by doing the
log force manually after the unlock, too.
Signed-off-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Alex Elder <aelder@sgi.com>
Let the transaction commit unlock the inode before it potentially causes
a synchronous log force.
Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Dave Chinner <dchinner@redhat.com>
Signed-off-by: Alex Elder <aelder@sgi.com>
Only read the LSN we need to push to with the ilock held, and then release
it before we do the log force to improve concurrency.
Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Dave Chinner <dchinner@redhat.com>
Signed-off-by: Alex Elder <aelder@sgi.com>
Only read the LSN we need to push to with the ilock held, and then release
it before we do the log force to improve concurrency.
This also removes the only direct caller of _xfs_trans_commit, thus
allowing it to be merged into the plain xfs_trans_commit again.
Signed-off-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Alex Elder <aelder@sgi.com>
XFS_TRANS_SWAPEXT is a transaction type, not a flag for xfs_trans_commit, so
don't pass it in xfs_swap_extents.
Signed-off-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Alex Elder <aelder@sgi.com>
In xfs_ioc_trim it is possible that computing the last allocation group
to discard might overflow for big start & len values, because the result
might be bigger then xfs_agnumber_t which is 32 bit long. Fix this by not
allowing the start and end block of the range to be beyond the end of the
file system.
Note that if the start is beyond the end of the file system we have to
return -EINVAL, but in the "end" case we have to truncate it to the fs
size.
Also introduce "end" variable, rather than using start+len which which
might be more confusing to get right as this bug shows.
Signed-off-by: Lukas Czerner <lczerner@redhat.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Alex Elder <aelder@sgi.com>
Convert all function prototypes to the short form used elsewhere,
and remove duplicates of comments already placed at the function
body.
Signed-off-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Alex Elder <aelder@sgi.com>
Fix a case in xfs_bmap_add_extent_unwritten_real where we aren't
passing the returned error on.
Signed-off-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Alex Elder <aelder@sgi.com>
All the parameters passed to xfs_bmap_add_extent_hole_real() are in
the xfs_bmalloca structure now. Just pass the bmalloca parameter to
the function instead of 8 separate parameters.
Signed-off-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Alex Elder <aelder@sgi.com>
All the parameters passed to xfs_bmap_add_extent_delay_real() are in
the xfs_bmalloca structure now. Just pass the bmalloca parameter to
the function instead of 8 separate parameters.
Signed-off-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Alex Elder <aelder@sgi.com>
Rather than passing the firstblock and freelist structure around,
embed it into the bmalloca structure and remove it from the function
parameters.
This also enables the minleft parameter to be set only once in
xfs_bmapi_write(), and the freelist cursor directly queried in
xfs_bmapi_allocate to clear it when the lowspace algorithm is
activated.
Signed-off-by: Dave Chinner <dchinner@redhat.com>
Signed-off-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Alex Elder <aelder@sgi.com>
Rather that putting extent records on the stack and then pointing to
them in the bmalloca structure which is in the same stack frame, put
the extent records directly in the bmalloca structure. This reduces
the number of args that need to be passed around.
Signed-off-by: Dave Chinner <dchinner@redhat.com>
Signed-off-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Alex Elder <aelder@sgi.com>
All the variables xfs_bmap_isaeof() is passed are contained within
the xfs_bmalloca structure. Pass that instead.
Signed-off-by: Dave Chinner <dchinner@redhat.com>
Signed-off-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Alex Elder <aelder@sgi.com>
Christoph Hellwig