DAX has a page fault serialisation problem with block allocation.
Because it allows concurrent page faults and does not have a page
lock to serialise faults to the same page, it can get two concurrent
faults to the page that race.
When two read faults race, this isn't a huge problem as the data
underlying the page is not changing and so "detect and drop" works
just fine. The issues are to do with write faults.
When two write faults occur, we serialise block allocation in
get_blocks() so only one faul will allocate the extent. It will,
however, be marked as an unwritten extent, and that is where the
problem lies - the DAX fault code cannot differentiate between a
block that was just allocated and a block that was preallocated and
needs zeroing. The result is that both write faults end up zeroing
the block and attempting to convert it back to written.
The problem is that the first fault can zero and convert before the
second fault starts zeroing, resulting in the zeroing for the second
fault overwriting the data that the first fault wrote with zeros.
The second fault then attempts to convert the unwritten extent,
which is then a no-op because it's already written. Data loss occurs
as a result of this race.
Because there is no sane locking construct in the page fault code
that we can use for serialisation across the page faults, we need to
ensure block allocation and zeroing occurs atomically in the
filesystem. This means we can still take concurrent page faults and
the only time they will serialise is in the filesystem
mapping/allocation callback. The page fault code will always see
written, initialised extents, so we will be able to remove the
unwritten extent handling from the DAX code when all filesystems are
converted.
Signed-off-by: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Brian Foster <bfoster@redhat.com>
Signed-off-by: Dave Chinner <david@fromorbit.com>
This patch modifies the stats counting macros and the callers
to those macros to properly increment, decrement, and add-to
the xfs stats counts. The counts for global and per-fs stats
are correctly advanced, and cleared by writing a "1" to the
corresponding clear file.
global counts: /sys/fs/xfs/stats/stats
per-fs counts: /sys/fs/xfs/sda*/stats/stats
global clear: /sys/fs/xfs/stats/stats_clear
per-fs clear: /sys/fs/xfs/sda*/stats/stats_clear
[dchinner: cleaned up macro variables, removed CONFIG_FS_PROC around
stats structures and macros. ]
Signed-off-by: Bill O'Donnell <billodo@redhat.com>
Reviewed-by: Eric Sandeen <sandeen@redhat.com>
Signed-off-by: Dave Chinner <david@fromorbit.com>
The total field from struct xfs_alloc_arg is a bit of an unknown
commodity. It is documented as the total block requirement for the
transaction and is used in this manner from most call sites by virtue of
passing the total block reservation of the transaction associated with
an allocation. Several xfs_bmapi_write() callers pass hardcoded values
of 0 or 1 for the total block requirement, which is a historical oddity
without any clear reasoning.
The xfs_iomap_write_direct() caller, for example, passes 0 for the total
block requirement. This has been determined to cause problems in the
form of ABBA deadlocks of AGF buffers due to incorrect AG selection in
the block allocator. Specifically, the xfs_alloc_space_available()
function incorrectly selects an AG that doesn't actually have sufficient
space for the allocation. This occurs because the args.total field is 0
and thus the remaining free space check on the AG doesn't actually
consider the size of the allocation request. This locks the AGF buffer,
the allocation attempt proceeds and ultimately fails (in
xfs_alloc_fix_minleft()), and xfs_alloc_vexent() moves on to the next
AG. In turn, this can lead to incorrect AG locking order (if the
allocator wraps around, attempting to lock AG 0 after acquiring AG N)
and thus deadlock if racing with another operation. This problem has
been reproduced via generic/299 on smallish (1GB) ramdisk test devices.
To avoid this problem, replace the undocumented hardcoded total
parameters from the iomap and utility callers to pass the block
reservation used for the associated transaction. This is consistent with
other xfs_bmapi_write() callers throughout XFS. The assumption is that
the total field allows the selection of an AG that can handle the entire
operation rather than simply the allocation/range being requested (e.g.,
resulting btree splits, etc.). This addresses the aforementioned
generic/299 hang by ensuring AG selection only occurs when the
allocation can be satisfied by the AG.
Reported-by: Ross Zwisler <ross.zwisler@linux.intel.com>
Signed-off-by: Brian Foster <bfoster@redhat.com>
Reviewed-by: Dave Chinner <dchinner@redhat.com>
Signed-off-by: Dave Chinner <david@fromorbit.com>
The iomap codepath (via get_blocks()) acquires and release the inode
lock in the case of a direct write that requires block allocation. This
is because xfs_iomap_write_direct() allocates a transaction, which means
the ilock must be dropped and reacquired after the transaction is
allocated and reserved.
xfs_iomap_write_direct() invokes xfs_iomap_eof_align_last_fsb() before
the transaction is created and thus before the ilock is reacquired. This
can lead to calls to xfs_iread_extents() and reads of the in-core extent
list without any synchronization (via xfs_bmap_eof() and
xfs_bmap_last_extent()). xfs_iread_extents() assert fails if the ilock
is not held, but this is not currently seen in practice as the current
callers had already invoked xfs_bmapi_read().
What has been seen in practice are reports of crashes down in the
xfs_bmap_eof() codepath on direct writes due to seemingly bogus pointer
references from xfs_iext_get_ext(). While an explicit reproducer is not
currently available to confirm the cause of the problem, crash analysis
and code inspection from David Jeffrey had identified the insufficient
locking.
xfs_iomap_eof_align_last_fsb() is called from other contexts with the
inode lock already held, so we cannot acquire it therein.
__xfs_get_blocks() acquires and drops the ilock with variable flags to
cover the event that the extent list must be read in. The common case is
that __xfs_get_blocks() acquires the shared ilock. To provide locking
around the last extent alignment call without adding more lock cycles to
the dio path, update xfs_iomap_write_direct() to expect the shared ilock
held on entry and do the extent alignment under its protection. Demote
the lock, if necessary, from __xfs_get_blocks() and push the
xfs_qm_dqattach() call outside of the shared lock critical section.
Also, add an assert to document that the extent list is always expected
to be present in this path. Otherwise, we risk a call to
xfs_iread_extents() while under the shared ilock. This is safe as all
current callers have executed an xfs_bmapi_read() call under the current
iolock context.
Reported-by: David Jeffery <djeffery@redhat.com>
Signed-off-by: Brian Foster <bfoster@redhat.com>
Reviewed-by: Dave Chinner <dchinner@redhat.com>
Signed-off-by: Dave Chinner <david@fromorbit.com>
The flags argument to xfs_trans_commit is not useful for most callers, as
a commit of a transaction without a permanent log reservation must pass
0 here, and all callers for a transaction with a permanent log reservation
except for xfs_trans_roll must pass XFS_TRANS_RELEASE_LOG_RES. So remove
the flags argument from the public xfs_trans_commit interfaces, and
introduce low-level __xfs_trans_commit variant just for xfs_trans_roll
that regrants a log reservation instead of releasing it.
Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Dave Chinner <dchinner@redhat.com>
Signed-off-by: Dave Chinner <david@fromorbit.com>
xfs_trans_cancel takes two flags arguments: XFS_TRANS_RELEASE_LOG_RES and
XFS_TRANS_ABORT. Both of them are a direct product of the transaction
state, and can be deducted:
- any dirty transaction needs XFS_TRANS_ABORT to be properly canceled,
and XFS_TRANS_ABORT is a noop for a transaction that is not dirty.
- any transaction with a permanent log reservation needs
XFS_TRANS_RELEASE_LOG_RES to be properly canceled, and passing
XFS_TRANS_RELEASE_LOG_RES for a transaction without a permanent
log reservation is invalid.
So just remove the flags argument and do the right thing.
Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Dave Chinner <dchinner@redhat.com>
Signed-off-by: Dave Chinner <david@fromorbit.com>
Now that the in-core superblock infrastructure has been replaced with
generic per-cpu counters, we don't need it anymore. Nuke it from
orbit so we are sure that it won't haunt us again...
Signed-off-by: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Brian Foster <bfoster@redhat.com>
Signed-off-by: Dave Chinner <david@fromorbit.com>
XFS has hand-rolled per-cpu counters for the superblock since before
there was any generic implementation. The free block counter is
special in that it is used for ENOSPC detection outside transaction
contexts for for delayed allocation. This means that the counter
needs to be accurate at zero. The current per-cpu counter code jumps
through lots of hoops to ensure we never run past zero, but we don't
need to make all those jumps with the generic counter
implementation.
The generic counter implementation allows us to pass a "batch"
threshold at which the addition/subtraction to the counter value
will be folded back into global value under lock. We can use this
feature to reduce the batch size as we approach 0 in a very similar
manner to the existing counters and their rebalance algorithm. If we
use a batch size of 1 as we approach 0, then every addition and
subtraction will be done against the global value and hence allow
accurate detection of zero threshold crossing.
Hence we can replace the handrolled, accurate-at-zero counters with
generic percpu counters.
Note: this removes just enough of the icsb infrastructure to compile
without warnings. The rest will go in subsequent commits.
Signed-off-by: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Brian Foster <bfoster@redhat.com>
Signed-off-by: Dave Chinner <david@fromorbit.com>
The code is already ready for it, and the pnfs layout commit code expects
to be able to pass a larger than 32-bit argument.
Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Dave Chinner <dchinner@redhat.com>
Signed-off-by: Dave Chinner <david@fromorbit.com>
The kernel compile doesn't turn on these checks by default, so it's
only when I do a kernel-user sync that I find that there are lots of
compiler warnings waiting to be fixed. Fix up these set-but-unused
warnings.
Signed-off-by: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Eric Sandeen <sandeen@redhat.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Dave Chinner <david@fromorbit.com>
If extsize is set and new_last_fsb is larger than 32 bits, the
roundup to extsize will overflow the align variable. Instead,
combine alignments by rounding stripe size up to extsize.
Signed-off-by: Peter Watkins <treestem@gmail.com>
Reviewed-by: Nathaniel W. Turner <nate@houseofnate.net>
Reviewed-by: Brian Foster <bfoster@redhat.com>
Signed-off-by: Dave Chinner <david@fromorbit.com>
More on-disk format consolidation.
Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Dave Chinner <dchinner@redhat.com>
Signed-off-by: Dave Chinner <david@fromorbit.com>
More on-disk format consolidation. A few declarations that weren't on-disk
format related move into better suitable spots.
Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Dave Chinner <dchinner@redhat.com>
Signed-off-by: Dave Chinner <david@fromorbit.com>
More consolidatation for the on-disk format defintions. Note that the
XFS_IS_REALTIME_INODE moves to xfs_linux.h instead as it is not related
to the on disk format, but depends on a CONFIG_ option.
Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Dave Chinner <dchinner@redhat.com>
Signed-off-by: Dave Chinner <david@fromorbit.com>
Coverity spotted this.
Granted, we *just* checked xfs_inod_dquot() in the caller (by
calling xfs_quota_need_throttle). However, this is the only place we
don't check the return value but the check is cheap and future-proof
so add it.
Signed-off-by: Eric Sandeen <sandeen@redhat.com>
Reviewed-by: Dave Chinner <dchinner@redhat.com>
Signed-off-by: Dave Chinner <david@fromorbit.com>
From: Brian Foster <bfoster@redhat.com>
Commit 4d559a3b introduced heavy prealloc. squashing to catch the case
of requesting too large a prealloc on smaller filesystems, leading to
repeated flush and retry cycles that occur on ENOSPC. Now that we issue
eofblocks scans on EDQUOT/ENOSPC, squash the prealloc against the
minimum available free space across all applicable quotas as well to
avoid a similar problem of repeated eofblocks scans.
Signed-off-by: Brian Foster <bfoster@redhat.com>
Reviewed-by: Dave Chinner <dchinner@redhat.com>
Signed-off-by: Dave Chinner <david@fromorbit.com>
The allocation stack switch at xfs_bmapi_allocate() has served it's
purpose, but is no longer a sufficient solution to the stack usage
problem we have in the XFS allocation path.
Whilst the kernel stack size is now 16k, that is not a valid reason
for undoing all our "keep stack usage down" modifications. What it
does allow us to do is have the freedom to refine and perfect the
modifications knowing that if we get it wrong it won't blow up in
our faces - we have a safety net now.
This is important because we still have the issue of older kernels
having smaller stacks and that they are still supported and are
demonstrating a wide range of different stack overflows. Red Hat
has several open bugs for allocation based stack overflows from
directory modifications and direct IO block allocation and these
problems still need to be solved. If we can solve them upstream,
then distro's won't need to bake their own unique solutions.
To that end, I've observed that every allocation based stack
overflow report has had a specific characteristic - it has happened
during or directly after a bmap btree block split. That event
requires a new block to be allocated to the tree, and so we
effectively stack one allocation stack on top of another, and that's
when we get into trouble.
A further observation is that bmap btree block splits are much rarer
than writeback allocation - over a range of different workloads I've
observed the ratio of bmap btree inserts to splits ranges from 100:1
(xfstests run) to 10000:1 (local VM image server with sparse files
that range in the hundreds of thousands to millions of extents).
Either way, bmap btree split events are much, much rarer than
allocation events.
Finally, we have to move the kswapd state to the allocation workqueue
work when allocation is done on behalf of kswapd. This is proving to
cause significant perturbation in performance under memory pressure
and appears to be generating allocation deadlock warnings under some
workloads, so avoiding the use of a workqueue for the majority of
kswapd writeback allocation will minimise the impact of such
behaviour.
Hence it makes sense to move the stack switch to xfs_btree_split()
and only do it for bmap btree splits. Stack switches during
allocation will be much rarer, so there won't be significant
performacne overhead caused by switching stacks. The worse case
stack from all allocation paths will be split, not just writeback.
And the majority of memory allocations will be done in the correct
context (e.g. kswapd) without causing additional latency, and so we
simplify the memory reclaim interactions between processes,
workqueues and kswapd.
The worst stack I've been able to generate with this patch in place
is 5600 bytes deep. It's very revealing because we exit XFS at:
37) 1768 64 kmem_cache_alloc+0x13b/0x170
about 1800 bytes of stack consumed, and the remaining 3800 bytes
(and 36 functions) is memory reclaim, swap and the IO stack. And
this occurs in the inode allocation from an open(O_CREAT) syscall,
not writeback.
The amount of stack being used is much less than I've previously be
able to generate - fs_mark testing has been able to generate stack
usage of around 7k without too much trouble; with this patch it's
only just getting to 5.5k. This is primarily because the metadata
allocation paths (e.g. directory blocks) are no longer causing
double splits on the same stack, and hence now stack tracing is
showing swapping being the worst stack consumer rather than XFS.
Performance of fs_mark inode create workloads is unchanged.
Performance of fs_mark async fsync workloads is consistently good
with context switches reduced by around 150,000/s (30%).
Performance of dbench, streaming IO and postmark is unchanged.
Allocation deadlock warnings have not been seen on the workloads
that generated them since adding this patch.
Signed-off-by: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Brian Foster <bfoster@redhat.com>
Signed-off-by: Dave Chinner <david@fromorbit.com>
Convert all the errors the core XFs code to negative error signs
like the rest of the kernel and remove all the sign conversion we
do in the interface layers.
Errors for conversion (and comparison) found via searches like:
$ git grep " E" fs/xfs
$ git grep "return E" fs/xfs
$ git grep " E[A-Z].*;$" fs/xfs
Negation points found via searches like:
$ git grep "= -[a-z,A-Z]" fs/xfs
$ git grep "return -[a-z,A-D,F-Z]" fs/xfs
$ git grep " -[a-z].*;" fs/xfs
[ with some bits I missed from Brian Foster ]
Signed-off-by: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Brian Foster <bfoster@redhat.com>
Signed-off-by: Dave Chinner <david@fromorbit.com>
XFS_ERROR was designed long ago to trap return values, but it's not
runtime configurable, it's not consistently used, and we can do
similar error trapping with ftrace scripts and triggers from
userspace.
Just nuke XFS_ERROR and associated bits.
Signed-off-by: Eric Sandeen <sandeen@redhat.com>
Reviewed-by: Dave Chinner <dchinner@redhat.com>
Signed-off-by: Dave Chinner <david@fromorbit.com>
remove unused transaction pointer from various
callchains leading to xfs_bmap_last_offset().
Signed-off-by: Eric Sandeen <sandeen@redhat.com>
Reviewed-by: Dave Chinner <dchinner@redhat.com>
Signed-off-by: Dave Chinner <david@fromorbit.com>
To allow aio writes beyond i_size we need to create unwritten extents for
newly allocated blocks, similar to how we already do inside i_size.
Instead of adding another special case we now use unwritten extents
unconditionally. This also marks the end of directly allocation data
extents in all of XFS - we now always use either delalloc or unwritten
extents.
Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Dave Chinner <dchinner@redhat.com>
Signed-off-by: Dave Chinner <david@fromorbit.com>
Currently the xfs_inode.h header has a dependency on the definition
of the BMAP btree records as the inode fork includes an array of
xfs_bmbt_rec_host_t objects in it's definition.
Move all the btree format definitions from xfs_btree.h,
xfs_bmap_btree.h, xfs_alloc_btree.h and xfs_ialloc_btree.h to
xfs_format.h to continue the process of centralising the on-disk
format definitions. With this done, the xfs inode definitions are no
longer dependent on btree header files.
The enables a massive culling of unnecessary includes, with close to
200 #include directives removed from the XFS kernel code base.
Signed-off-by: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Ben Myers <bpm@sgi.com>
Signed-off-by: Ben Myers <bpm@sgi.com>
xfs_trans.h has a dependency on xfs_log.h for a couple of
structures. Most code that does transactions doesn't need to know
anything about the log, but this dependency means that they have to
include xfs_log.h. Decouple the xfs_trans.h and xfs_log.h header
files and clean up the includes to be in dependency order.
In doing this, remove the direct include of xfs_trans_reserve.h from
xfs_trans.h so that we remove the dependency between xfs_trans.h and
xfs_mount.h. Hence the xfs_trans.h include can be moved to the
indicate the actual dependencies other header files have on it.
Note that these are kernel only header files, so this does not
translate to any userspace changes at all.
Signed-off-by: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Ben Myers <bpm@sgi.com>
Signed-off-by: Ben Myers <bpm@sgi.com>
All of the buffer operations structures are needed to be exported
for xfs_db, so move them all to a common location rather than
spreading them all over the place. They are verifying the on-disk
format, so while xfs_format.h might be a good place, it is not part
of the on disk format.
Hence we need to create a new header file that we centralise these
related definitions. Start by moving the bffer operations
structures, and then also move all the other definitions that have
crept into xfs_log_format.h and xfs_format.h as there was no other
shared header file to put them in.
Signed-off-by: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Ben Myers <bpm@sgi.com>
__xfs_printk adds its own "\n". Having it in the original string
leads to unintentional blank lines from these messages.
Most format strings have no newline, but a few do, leading to
i.e.:
[ 7347.119911] XFS (sdb2): Access to block zero in inode 132 start_block: 0 start_off: 0 blkcnt: 0 extent-state: 0 lastx: 1a05
[ 7347.119911]
[ 7347.119919] XFS (sdb2): Access to block zero in inode 132 start_block: 0 start_off: 0 blkcnt: 0 extent-state: 0 lastx: 1a05
[ 7347.119919]
Fix them all.
Signed-off-by: Eric Sandeen <sandeen@redhat.com>
Reviewed-by: Mark Tinguely <tinguely@sgi.com>
Signed-off-by: Ben Myers <bpm@sgi.com>
Get rid of function variable count from xfs_iomap_write_allocate() as
it is unused.
Additionally, checkpatch warn me of the following for this change:
WARNING: extern prototypes should be avoided in .h files
+extern int xfs_iomap_write_allocate(struct xfs_inode *, xfs_off_t,
So this patch also remove all extern function prototypes at xfs_iomap.h
to suppress it to make this code style in consistent manner in this file.
Signed-off-by: Jie Liu <jeff.liu@oracle.com>
Reviewed-by: Dave Chinner <dchinner@redhat.com>
Signed-off-by: Ben Myers <bpm@sgi.com>
With the new xfs_trans_res structure has been introduced, the log
reservation size, log count as well as log flags are pre-initialized
at mount time. So it's time to refine xfs_trans_reserve() interface
to be more neat.
Also, introduce a new helper M_RES() to return a pointer to the
mp->m_resv structure to simplify the input.
Signed-off-by: Jie Liu <jeff.liu@oracle.com>
Signed-off-by: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Mark Tinguely <tinguely@sgi.com>
Signed-off-by: Ben Myers <bpm@sgi.com>
There are a few small helper functions in xfs_util, all related to
xfs_inode modifications. Move them all to xfs_inode.c so all
xfs_inode operations are consiolidated in the one place.
Signed-off-by: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Mark Tinguely <tinguely@sgi.com>
Signed-off-by: Ben Myers <bpm@sgi.com>
There is a bunch of code in xfs_bmap.c that is kernel specific and
not shared with userspace. To minimise the difference between the
kernel and userspace code, shift this unshared code to
xfs_bmap_util.c, and the declarations to xfs_bmap_util.h.
The biggest issue here is xfs_bmap_finish() - userspace has it's own
definition of this function, and so we need to move it out of
xfs_bmap.[ch]. This means several other files need to include
xfs_bmap_util.h as well.
It also introduces and interesting dance for the stack switching
code in xfs_bmapi_allocate(). The stack switching/workqueue code is
actually moved to xfs_bmap_util.c, so that userspace can simply use
a #define in a header file to connect the dots without needing to
know about the stack switch code at all.
Signed-off-by: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Mark Tinguely <tinguely@sgi.com>
Signed-off-by: Ben Myers <bpm@sgi.com>
The on disk format definitions of the on-disk dquot, log formats and
quota off log formats are all intertwined with other definitions for
quotas. Separate them out into their own header file so they can
easily be shared with userspace.
Signed-off-by: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Brian Foster <bfoster@redhat.com>
Reviewed-by: Mark Tinguely <tinguely@sgi.com>
Signed-off-by: Ben Myers <bpm@sgi.com>
Dedicated small file workloads have been seeing significant free
space fragmentation causing premature inode allocation failure
when large inode sizes are in use. A particular test case showed
that a workload that runs to a real ENOSPC on 256 byte inodes would
fail inode allocation with ENOSPC about about 80% full with 512 byte
inodes, and at about 50% full with 1024 byte inodes.
The same workload, when run with -o allocsize=4096 on 1024 byte
inodes would run to being 100% full before giving ENOSPC. That is,
no freespace fragmentation at all.
The issue was caused by the specific IO pattern the application had
- the framework it was using did not support direct IO, and so it
was emulating it by using fadvise(DONT_NEED). The result was that
the data was getting written back before the speculative prealloc
had been trimmed from memory by the close(), and so small single
block files were being allocated with 2 blocks, and then having one
truncated away. The result was lots of small 4k free space extents,
and hence each new 8k allocation would take another 8k from
contiguous free space and turn it into 4k of allocated space and 4k
of free space.
Hence inode allocation, which requires contiguous, aligned
allocation of 16k (256 byte inodes), 32k (512 byte inodes) or 64k
(1024 byte inodes) can fail to find sufficiently large freespace and
hence fail while there is still lots of free space available.
There's a simple fix for this, and one that has precendence in the
allocator code already - don't do speculative allocation unless the
size of the file is larger than a certain size. In this case, that
size is the minimum default preallocation size:
mp->m_writeio_blocks. And to keep with the concept of being nice to
people when the files are still relatively small, cap the prealloc
to mp->m_writeio_blocks until the file goes over a stripe unit is
size, at which point we'll fall back to the current behaviour based
on the last extent size.
This will effectively turn off speculative prealloc for very small
files, keep preallocation low for small files, and behave as it
currently does for any file larger than a stripe unit. This
completely avoids the freespace fragmentation problem this
particular IO pattern was causing.
Signed-off-by: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Brian Foster <bfoster@redhat.com>
Reviewed-by: Mark Tinguely <tinguely@sgi.com>
Signed-off-by: Ben Myers <bpm@sgi.com>
Add a tracepoint to provide some feedback on preallocation size
calculation.
Signed-off-by: Brian Foster <bfoster@redhat.com>
Reviewed-by: Mark Tinguely <tinguely@sgi.com>
Signed-off-by: Ben Myers <bpm@sgi.com>
Introduce the need_throttle() and calc_throttle() functions to
independently check whether throttling is required for a particular
dquot and if so, calculate the associated throttling metrics based
on the state of the quota. We use the same general algorithm to
calculate the throttle shift as for global free space with the
exception of using three stages rather than five.
Update xfs_iomap_prealloc_size() to use the smallest available
prealloc size based on each of the constraints and apply the
maximum shift to obtain the throttled preallocation size.
Signed-off-by: Brian Foster <bfoster@redhat.com>
Reviewed-by: Mark Tinguely <tinguely@sgi.com>
Signed-off-by: Ben Myers <bpm@sgi.com>
The round down occurs towards the beginning of the function. Push
it down after throttling has occurred. This is to support adding
further transformations to 'alloc_blocks' that might not preserve
power-of-two alignment (and thus could lead to rounding down
multiple times).
Signed-off-by: Brian Foster <bfoster@redhat.com>
Reviewed-by: Ben Myers <bpm@sgi.com>
Reviewed-by: Mark Tinguely <tinguely@sgi.com>
Signed-off-by: Ben Myers <bpm@sgi.com>
The majority of xfs_iomap_prealloc_size() executes within the
check for lack of default I/O size. Reverse the logic to remove the
extra indentation.
Signed-off-by: Brian Foster <bfoster@redhat.com>
Reviewed-by: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Ben Myers <bpm@sgi.com>
Reviewed-by: Mark Tinguely <tinguely@sgi.com>
Signed-off-by: Ben Myers <bpm@sgi.com>
Fix the return type of xfs_iomap_eof_prealloc_initial_size() to
xfs_fsblock_t to reflect the fact that the return value may be an
unsigned 64 bits if XFS_BIG_BLKNOS is defined.
Signed-off-by: Mark Tinguely <tinguely@sgi.com>
Reviewed-by: Dave Chinner <dchinner@redhat.com>
Signed-off-by: Ben Myers <bpm@sgi.com>
The updated speculative preallocation algorithm for handling sparse
files can becomes less effective in situations with a high number of
concurrent, sequential writers. The number of writers and amount of
available RAM affect the writeback bandwidth slicing algorithm,
which in turn affects the block allocation pattern of XFS. For
example, running 32 sequential writers on a system with 32GB RAM,
preallocs become fixed at a value of around 128MB (instead of
steadily increasing to the 8GB maximum as sequential writes
proceed).
Update the speculative prealloc heuristic to base the size of the
next prealloc on double the size of the preceding extent. This
preserves the original aggressive speculative preallocation
behavior and continues to accomodate sparse files at a slight cost
of increasing the size of preallocated data regions following holes
of sparse files.
Signed-off-by: Brian Foster <bfoster@redhat.com>
Reviewed-by: Dave Chinner <dchinner@redhat.com>
Signed-off-by: Ben Myers <bpm@sgi.com>
If freesp == 0, we could end up in an infinite loop while squashing
the preallocation. Break the loop when we've killed the prealloc
entirely.
Signed-off-by: Brian Foster <bfoster@redhat.com>
Reviewed-by: Dave Chinner <dchinner@redhat.com>
Signed-off-by: Ben Myers <bpm@sgi.com>
Speculative preallocation based on the current file size works well
for contiguous files, but is sub-optimal for sparse files where the
EOF preallocation can fill holes and result in large amounts of
zeros being written when it is not necessary.
The algorithm is modified to prevent EOF speculative preallocation
from triggering larger allocations on IO patterns of
truncate--to-zero-seek-write-seek-write-.... which results in
non-sparse files for large files. This, unfortunately, is the way cp
now behaves when copying sparse files and so needs to be fixed.
What this code does is that it looks at the existing extent adjacent
to the current EOF and if it determines that it is a hole we disable
speculative preallocation altogether. To avoid the next write from
doing a large prealloc, it takes the size of subsequent
preallocations from the current size of the existing EOF extent.
IOWs, if you leave a hole in the file, it resets preallocation
behaviour to the same as if it was a zero size file.
Example new behaviour:
$ xfs_io -f -c "pwrite 0 31m" \
-c "pwrite 33m 1m" \
-c "pwrite 128m 1m" \
-c "fiemap -v" /mnt/scratch/blah
wrote 32505856/32505856 bytes at offset 0
31 MiB, 7936 ops; 0.0000 sec (1.608 GiB/sec and 421432.7439 ops/sec)
wrote 1048576/1048576 bytes at offset 34603008
1 MiB, 256 ops; 0.0000 sec (1.462 GiB/sec and 383233.5329 ops/sec)
wrote 1048576/1048576 bytes at offset 134217728
1 MiB, 256 ops; 0.0000 sec (1.719 GiB/sec and 450704.2254 ops/sec)
/mnt/scratch/blah:
EXT: FILE-OFFSET BLOCK-RANGE TOTAL FLAGS
0: [0..65535]: 96..65631 65536 0x0
1: [65536..67583]: hole 2048
2: [67584..69631]: 67680..69727 2048 0x0
3: [69632..262143]: hole 192512
4: [262144..264191]: 262240..264287 2048 0x1
Signed-off-by: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Mark Tinguely <tinguely@sgi.com>
Reviewed-by: Brian Foster <bfoster@redhat.com>
Signed-off-by: Ben Myers <bpm@sgi.com>
There is a window on small filesytsems where specualtive
preallocation can be larger than that ENOSPC throttling thresholds,
resulting in specualtive preallocation trying to reserve more space
than there is space available. This causes immediate ENOSPC to be
triggered, prealloc to be turned off and flushing to occur. One the
next write (i.e. next 4k page), we do exactly the same thing, and so
effective drive into synchronous 4k writes by triggering ENOSPC
flushing on every page while in the window between the prealloc size
and the ENOSPC prealloc throttle threshold.
Fix this by checking to see if the prealloc size would consume all
free space, and throttle it appropriately to avoid premature
ENOSPC...
Signed-off-by: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Brian Foster <bfoster@redhat.com>
Signed-off-by: Ben Myers <bpm@sgi.com>
Add the XFS_ICI_EOFBLOCKS_TAG inode tag to identify inodes with
speculatively preallocated blocks beyond EOF. An inode is tagged
when speculative preallocation occurs and untagged either via
truncate down or when post-EOF blocks are freed via release or
reclaim.
The tag management is intentionally not aggressive to prefer
simplicity over the complexity of handling all the corner cases
under which post-EOF blocks could be freed (i.e., forward
truncation, fallocate, write error conditions, etc.). This means
that a tagged inode may or may not have post-EOF blocks after a
period of time. The tag is eventually cleared when the inode is
released or reclaimed.
Signed-off-by: Brian Foster <bfoster@redhat.com>
Reviewed-by: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Mark Tinguely <tinguely@sgi.com>
Signed-off-by: Ben Myers <bpm@sgi.com>
Certain allocation paths through xfs_bmapi_write() are in situations
where we have limited stack available. These are almost always in
the buffered IO writeback path when convertion delayed allocation
extents to real extents.
The current stack switch occurs for userdata allocations, which
means we also do stack switches for preallocation, direct IO and
unwritten extent conversion, even those these call chains have never
been implicated in a stack overrun.
Hence, let's target just the single stack overun offended for stack
switches. To do that, introduce a XFS_BMAPI_STACK_SWITCH flag that
the caller can pass xfs_bmapi_write() to indicate it should switch
stacks if it needs to do allocation.
Signed-off-by: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Mark Tinguely <tinguely@sgi.com>
Signed-off-by: Ben Myers <bpm@sgi.com>
We don't do any data writeback from XFS any more - the VFS is
completely responsible for that, including for freeze. We can
replace the remaining caller with a VFS level function that
achieves the same thing, but without conflicting with current
writeback work.
This means we can remove the flush_work and xfs_flush_inodes() - the
VFS functionality completely replaces the internal flush queue for
doing this writeback work in a separate context to avoid stack
overruns.
This does have one complication - it cannot be called with page
locks held. Hence move the flushing of delalloc space when ENOSPC
occurs back up into xfs_file_aio_buffered_write when we don't hold
any locks that will stall writeback.
Unfortunately, writeback_inodes_sb_if_idle() is not sufficient to
trigger delalloc conversion fast enough to prevent spurious ENOSPC
whent here are hundreds of writers, thousands of small files and GBs
of free RAM. Hence we need to use sync_sb_inodes() to block callers
while we wait for writeback like the previous xfs_flush_inodes
implementation did.
That means we have to hold the s_umount lock here, but because this
call can nest inside i_mutex (the parent directory in the create
case, held by the VFS), we have to use down_read_trylock() to avoid
potential deadlocks. In practice, this trylock will succeed on
almost every attempt as unmount/remount type operations are
exceedingly rare.
Note: we always need to pass a count of zero to
generic_file_buffered_write() as the previously written byte count.
We only do this by accident before this patch by the virtue of ret
always being zero when there are no errors. Make this explicit
rather than needing to specifically zero ret in the ENOSPC retry
case.
Signed-off-by: Dave Chinner <dchinner@redhat.com>
Tested-by: Brian Foster <bfoster@redhat.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Ben Myers <bpm@sgi.com>
Pull second vfs pile from Al Viro:
"The stuff in there: fsfreeze deadlock fixes by Jan (essentially, the
deadlock reproduced by xfstests 068), symlink and hardlink restriction
patches, plus assorted cleanups and fixes.
Note that another fsfreeze deadlock (emergency thaw one) is *not*
dealt with - the series by Fernando conflicts a lot with Jan's, breaks
userland ABI (FIFREEZE semantics gets changed) and trades the deadlock
for massive vfsmount leak; this is going to be handled next cycle.
There probably will be another pull request, but that stuff won't be
in it."
Fix up trivial conflicts due to unrelated changes next to each other in
drivers/{staging/gdm72xx/usb_boot.c, usb/gadget/storage_common.c}
* 'for-linus' of git://git.kernel.org/pub/scm/linux/kernel/git/viro/vfs: (54 commits)
delousing target_core_file a bit
Documentation: Correct s_umount state for freeze_fs/unfreeze_fs
fs: Remove old freezing mechanism
ext2: Implement freezing
btrfs: Convert to new freezing mechanism
nilfs2: Convert to new freezing mechanism
ntfs: Convert to new freezing mechanism
fuse: Convert to new freezing mechanism
gfs2: Convert to new freezing mechanism
ocfs2: Convert to new freezing mechanism
xfs: Convert to new freezing code
ext4: Convert to new freezing mechanism
fs: Protect write paths by sb_start_write - sb_end_write
fs: Skip atime update on frozen filesystem
fs: Add freezing handling to mnt_want_write() / mnt_drop_write()
fs: Improve filesystem freezing handling
switch the protection of percpu_counter list to spinlock
nfsd: Push mnt_want_write() outside of i_mutex
btrfs: Push mnt_want_write() outside of i_mutex
fat: Push mnt_want_write() outside of i_mutex
...
Dave Chinner