Since LARP is an experimental debug-only feature, we should try to warn
about it being in use once per mount, not once per reboot.
Signed-off-by: Darrick J. Wong <djwong@kernel.org>
Reviewed-by: Dave Chinner <dchinner@redhat.com>
Signed-off-by: Dave Chinner <david@fromorbit.com>
Currently, we don't have a consistent story around logging when an
EXPERIMENTAL feature gets turned on at runtime -- online fsck and shrink
log a message once per day across all mounts, and the recently merged
LARP mode only ever does it once per insmod cycle or reboot.
Because EXPERIMENTAL tags are supposed to go away eventually, convert
the existing daily warnings into state flags that travel with the mount,
and warn once per mount. Making this an opstate flag means that we'll
be able to capture the experimental usage in the ftrace output too.
Signed-off-by: Darrick J. Wong <djwong@kernel.org>
Reviewed-by: Dave Chinner <dchinner@redhat.com>
Signed-off-by: Dave Chinner <david@fromorbit.com>
xfs: Large extent counters
The commit xfs: fix inode fork extent count overflow
(3f8a4f1d87) mentions that 10 billion
data fork extents should be possible to create. However the
corresponding on-disk field has a signed 32-bit type. Hence this
patchset extends the per-inode data fork extent counter to 64 bits
(out of which 48 bits are used to store the extent count).
Also, XFS has an attribute fork extent counter which is 16 bits
wide. A workload that,
1. Creates 1 million 255-byte sized xattrs,
2. Deletes 50% of these xattrs in an alternating manner,
3. Tries to insert 400,000 new 255-byte sized xattrs
causes the xattr extent counter to overflow.
Dave tells me that there are instances where a single file has more
than 100 million hardlinks. With parent pointers being stored in
xattrs, we will overflow the signed 16-bits wide attribute extent
counter when large number of hardlinks are created. Hence this
patchset extends the on-disk field to 32-bits.
The following changes are made to accomplish this,
1. A 64-bit inode field is carved out of existing di_pad and
di_flushiter fields to hold the 64-bit data fork extent counter.
2. The existing 32-bit inode data fork extent counter will be used to
hold the attribute fork extent counter.
3. A new incompat superblock flag to prevent older kernels from mounting
the filesystem.
Signed-off-by: Chandan Babu R <chandan.babu@oracle.com>
Signed-off-by: Dave Chinner <david@fromorbit.com>
5.18 w/ std=gnu11 compiled with gcc-5 wants flags stored in unsigned
fields to be unsigned.
Signed-off-by: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Chandan Babu R <chandan.babu@oracle.com>
Signed-off-by: Dave Chinner <david@fromorbit.com>
As mentioned in the previous commit, the kernel misuses sb_frextents in
the incore mount to reflect both incore reservations made by running
transactions as well as the actual count of free rt extents on disk.
This results in the superblock being written to the log with an
underestimate of the number of rt extents that are marked free in the
rtbitmap.
Teaching XFS to recompute frextents after log recovery avoids
operational problems in the current mount, but it doesn't solve the
problem of us writing undercounted frextents which are then recovered by
an older kernel that doesn't have that fix.
Create an incore percpu counter to mirror the ondisk frextents. This
new counter will track transaction reservations and the only time we
will touch the incore super counter (i.e the one that gets logged) is
when those transactions commit updates to the rt bitmap. This is in
contrast to the lazysbcount counters (e.g. fdblocks), where we know that
log recovery will always fix any incorrect counter that we log.
As a bonus, we only take m_sb_lock at transaction commit time.
Signed-off-by: Darrick J. Wong <djwong@kernel.org>
Reviewed-by: Dave Chinner <dchinner@redhat.com>
Signed-off-by: Dave Chinner <david@fromorbit.com>
XFS_SB_FEAT_INCOMPAT_NREXT64 incompat feature bit will be set on filesystems
which support large per-inode extent counters. This commit defines the new
incompat feature bit and the corresponding per-fs feature bit (along with
inline functions to work on it).
Reviewed-by: Darrick J. Wong <djwong@kernel.org>
Reviewed-by: Dave Chinner <dchinner@redhat.com>
Signed-off-by: Chandan Babu R <chandan.babu@oracle.com>
xfs_reserve_blocks controls the size of the user-visible free space
reserve pool. Given the difference between the current and requested
pool sizes, it will try to reserve free space from fdblocks. However,
the amount requested from fdblocks is also constrained by the amount of
space that we think xfs_mod_fdblocks will give us. If we forget to
subtract m_allocbt_blks before calling xfs_mod_fdblocks, it will will
return ENOSPC and we'll hang the kernel at mount due to the infinite
loop.
In commit fd43cf600c, we decided that xfs_mod_fdblocks should not hand
out the "free space" used by the free space btrees, because some portion
of the free space btrees hold in reserve space for future btree
expansion. Unfortunately, xfs_reserve_blocks' estimation of the number
of blocks that it could request from xfs_mod_fdblocks was not updated to
include m_allocbt_blks, so if space is extremely low, the caller hangs.
Fix this by creating a function to estimate the number of blocks that
can be reserved from fdblocks, which needs to exclude the set-aside and
m_allocbt_blks.
Found by running xfs/306 (which formats a single-AG 20MB filesystem)
with an fstests configuration that specifies a 1k blocksize and a
specially crafted log size that will consume 7/8 of the space (17920
blocks, specifically) in that AG.
Cc: Brian Foster <bfoster@redhat.com>
Fixes: fd43cf600c ("xfs: set aside allocation btree blocks from block reservation")
Signed-off-by: Darrick J. Wong <djwong@kernel.org>
Reviewed-by: Brian Foster <bfoster@redhat.com>
Reviewed-by: Dave Chinner <dchinner@redhat.com>
Compute the actual maximum AG btree height for deciding if a per-AG
block reservation is critically low. This only affects the sanity check
condition, since we /generally/ will trigger on the 10% threshold. This
is a long-winded way of saying that we're removing one more usage of
XFS_BTREE_MAXLEVELS.
Signed-off-by: Darrick J. Wong <djwong@kernel.org>
Reviewed-by: Dave Chinner <dchinner@redhat.com>
Years ago when XFS was thought to be much more simple, we introduced
m_ag_maxlevels to specify the maximum btree height of per-AG btrees for
a given filesystem mount. Then we observed that inode btrees don't
actually have the same height and split that off; and now we have rmap
and refcount btrees with much different geometries and separate
maxlevels variables.
The 'ag' part of the name doesn't make much sense anymore, so rename
this to m_alloc_maxlevels to reinforce that this is the maximum height
of the *free space* btrees. This sets us up for the next patch, which
will add a variable to track the maximum height of all AG btrees.
(Also take the opportunity to improve adjacent comments and fix minor
style problems.)
Signed-off-by: Darrick J. Wong <djwong@kernel.org>
Reviewed-by: Dave Chinner <dchinner@redhat.com>
Remove the shouty macro and instead use the inline function that
matches other state/feature check wrapper naming. This conversion
was done with sed.
Signed-off-by: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Darrick J. Wong <djwong@kernel.org>
Signed-off-by: Darrick J. Wong <djwong@kernel.org>
The remaining mount flags kept in m_flags are actually runtime state
flags. These change dynamically, so they really should be updated
atomically so we don't potentially lose an update due to racing
modifications.
Convert these remaining flags to be stored in m_opstate and use
atomic bitops to set and clear the flags. This also adds a couple of
simple wrappers for common state checks - read only and shutdown.
Signed-off-by: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Darrick J. Wong <djwong@kernel.org>
Signed-off-by: Darrick J. Wong <djwong@kernel.org>
Replace m_flags feature checks with xfs_has_<feature>() calls and
rework the setup code to set flags in m_features.
Signed-off-by: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Darrick J. Wong <djwong@kernel.org>
Signed-off-by: Darrick J. Wong <djwong@kernel.org>
This provides separation of mount time feature flags from runtime
mount flags and mount option state. It also makes the feature
checks use the same interface as the superblock features. i.e. we
don't care if the feature is enabled by superblock flags or mount
options, we just care if it's enabled or not.
Signed-off-by: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Darrick J. Wong <djwong@kernel.org>
Signed-off-by: Darrick J. Wong <djwong@kernel.org>
Convert the xfs_sb_version_hasfoo() to checks against
mp->m_features. Checks of the superblock itself during disk
operations (e.g. in the read/write verifiers and the to/from disk
formatters) are not converted - they operate purely on the
superblock state. Everything else should use the mount features.
Large parts of this conversion were done with sed with commands like
this:
for f in `git grep -l xfs_sb_version_has fs/xfs/*.c`; do
sed -i -e 's/xfs_sb_version_has\(.*\)(&\(.*\)->m_sb)/xfs_has_\1(\2)/' $f
done
With manual cleanups for things like "xfs_has_extflgbit" and other
little inconsistencies in naming.
The result is ia lot less typing to check features and an XFS binary
size reduced by a bit over 3kB:
$ size -t fs/xfs/built-in.a
text data bss dec hex filenam
before 1130866 311352 484 1442702 16038e (TOTALS)
after 1127727 311352 484 1439563 15f74b (TOTALS)
Signed-off-by: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Darrick J. Wong <djwong@kernel.org>
Signed-off-by: Darrick J. Wong <djwong@kernel.org>
Currently on-disk feature checks require decoding the superblock
fileds and so can be non-trivial. We have almost 400 hundred
individual feature checks in the XFS code, so this is a significant
amount of code. To reduce runtime check overhead, pre-process all
the version flags into a features field in the xfs_mount at mount
time so we can convert all the feature checks to a simple flag
check.
There is also a need to convert the dynamic feature flags to update
the m_features field. This is required for attr, attr2 and quota
features. New xfs_mount based wrappers are added for this.
Signed-off-by: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Darrick J. Wong <djwong@kernel.org>
Signed-off-by: Darrick J. Wong <djwong@kernel.org>
Add a tracepoint for fs shutdowns so we can capture that in ftrace
output.
Signed-off-by: Darrick J. Wong <djwong@kernel.org>
Reviewed-by: Christoph Hellwig <hch@lst.de>
We only use the CIL workqueue in the CIL, so it makes no sense to
hang it off the xfs_mount and have to walk multiple pointers back up
to the mount when we have the CIL structures right there.
Signed-off-by: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Darrick J. Wong <djwong@kernel.org>
Signed-off-by: Darrick J. Wong <djwong@kernel.org>
Log incompat feature flags in the superblock exist for one purpose: to
protect the contents of a dirty log from replay on a kernel that isn't
prepared to handle those dirty contents. This means that they can be
cleared if (a) we know the log is clean and (b) we know that there
aren't any other threads in the system that might be setting or relying
upon a log incompat flag.
Therefore, clear the log incompat flags when we've finished recovering
the log, when we're unmounting cleanly, remounting read-only, or
freezing; and provide a function so that subsequent patches can start
using this.
Signed-off-by: Darrick J. Wong <djwong@kernel.org>
Reviewed-by: Allison Henderson <allison.henderson@oracle.com>
Reviewed-by: Chandan Babu R <chandanrlinux@gmail.com>
Now that we defer inode inactivation, we've decoupled the process of
unlinking or closing an inode from the process of inactivating it. In
theory this should lead to better throughput since we now inactivate the
queued inodes in batches instead of one at a time.
Unfortunately, one of the primary risks with this decoupling is the loss
of rate control feedback between the frontend and background threads.
In other words, a rm -rf /* thread can run the system out of memory if
it can queue inodes for inactivation and jump to a new CPU faster than
the background threads can actually clear the deferred work. The
workers can get scheduled off the CPU if they have to do IO, etc.
To solve this problem, we configure a shrinker so that it will activate
the /second/ time the shrinkers are called. The custom shrinker will
queue all percpu deferred inactivation workers immediately and set a
flag to force frontend callers who are releasing a vfs inode to wait for
the inactivation workers.
On my test VM with 560M of RAM and a 2TB filesystem, this seems to solve
most of the OOMing problem when deleting 10 million inodes.
Signed-off-by: Darrick J. Wong <djwong@kernel.org>
Reviewed-by: Dave Chinner <dchinner@redhat.com>
Now that we have the infrastructure to switch background workers on and
off at will, fix the block gc worker code so that we don't actually run
the worker when the filesystem is frozen, same as we do for deferred
inactivation.
Signed-off-by: Darrick J. Wong <djwong@kernel.org>
Reviewed-by: Dave Chinner <dchinner@redhat.com>
Now that we have made the inactivation of unlinked inodes a background
task to increase the throughput of file deletions, we need to be a
little more careful about how long of a delay we can tolerate.
Similar to the patch doing this for free space on the data device, if
the file being inactivated is a realtime file and the realtime volume is
running low on free extents, we want to run the worker ASAP so that the
realtime allocator can make better decisions.
Signed-off-by: Darrick J. Wong <djwong@kernel.org>
Reviewed-by: Dave Chinner <dchinner@redhat.com>
Now that we have made the inactivation of unlinked inodes a background
task to increase the throughput of file deletions, we need to be a
little more careful about how long of a delay we can tolerate.
On a mostly empty filesystem, the risk of the allocator making poor
decisions due to fragmentation of the free space on account a lengthy
delay in background updates is minimal because there's plenty of space.
However, if free space is tight, we want to deallocate unlinked inodes
as quickly as possible to avoid fallocate ENOSPC and to give the
allocator the best shot at optimal allocations for new writes.
Therefore, queue the percpu worker immediately if the filesystem is more
than 95% full. This follows the same principle that XFS becomes less
aggressive about speculative allocations and lazy cleanup (and more
precise about accounting) when nearing full.
Signed-off-by: Darrick J. Wong <djwong@kernel.org>
Reviewed-by: Dave Chinner <dchinner@redhat.com>
Move inode inactivation to background work contexts so that it no
longer runs in the context that releases the final reference to an
inode. This will allow process work that ends up blocking on
inactivation to continue doing work while the filesytem processes
the inactivation in the background.
A typical demonstration of this is unlinking an inode with lots of
extents. The extents are removed during inactivation, so this blocks
the process that unlinked the inode from the directory structure. By
moving the inactivation to the background process, the userspace
applicaiton can keep working (e.g. unlinking the next inode in the
directory) while the inactivation work on the previous inode is
done by a different CPU.
The implementation of the queue is relatively simple. We use a
per-cpu lockless linked list (llist) to queue inodes for
inactivation without requiring serialisation mechanisms, and a work
item to allow the queue to be processed by a CPU bound worker
thread. We also keep a count of the queue depth so that we can
trigger work after a number of deferred inactivations have been
queued.
The use of a bound workqueue with a single work depth allows the
workqueue to run one work item per CPU. We queue the work item on
the CPU we are currently running on, and so this essentially gives
us affine per-cpu worker threads for the per-cpu queues. THis
maintains the effective CPU affinity that occurs within XFS at the
AG level due to all objects in a directory being local to an AG.
Hence inactivation work tends to run on the same CPU that last
accessed all the objects that inactivation accesses and this
maintains hot CPU caches for unlink workloads.
A depth of 32 inodes was chosen to match the number of inodes in an
inode cluster buffer. This hopefully allows sequential
allocation/unlink behaviours to defering inactivation of all the
inodes in a single cluster buffer at a time, further helping
maintain hot CPU and buffer cache accesses while running
inactivations.
A hard per-cpu queue throttle of 256 inode has been set to avoid
runaway queuing when inodes that take a long to time inactivate are
being processed. For example, when unlinking inodes with large
numbers of extents that can take a lot of processing to free.
Signed-off-by: Dave Chinner <dchinner@redhat.com>
[djwong: tweak comments and tracepoints, convert opflags to state bits]
Reviewed-by: Darrick J. Wong <djwong@kernel.org>
Signed-off-by: Darrick J. Wong <djwong@kernel.org>
The inode inactivation and CIL tracking percpu structures are
per-xfs_mount structures. That means when we get a CPU dead
notification, we need to then iterate all the per-cpu structure
instances to process them. Rather than keeping linked lists of
per-cpu structures in each subsystem, add a list of all xfs_mounts
that the generic xfs_cpu_dead() function will iterate and call into
each subsystem appropriately.
This allows us to handle both per-mount and global XFS percpu state
from xfs_cpu_dead(), and avoids the need to link subsystem
structures that can be easily found from the xfs_mount into their
own global lists.
Signed-off-by: Dave Chinner <dchinner@redhat.com>
[djwong: expand some comments about mount list setup ordering rules]
Reviewed-by: Darrick J. Wong <djwong@kernel.org>
Signed-off-by: Darrick J. Wong <djwong@kernel.org>
Move the xfs_perag infrastructure to the libxfs files that contain
all the per AG infrastructure. This helps set up for passing perags
around all the code instead of bare agnos with minimal extra
includes for existing files.
Signed-off-by: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Brian Foster <bfoster@redhat.com>
Reviewed-by: Darrick J. Wong <djwong@kernel.org>
The perag structures really need to be defined with the rest of the
AG support infrastructure. The struct xfs_perag and init/teardown
has been placed in xfs_mount.[ch] because there are differences in
the structure between kernel and userspace. Mainly that userspace
doesn't have a lot of the internal stuff that the kernel has for
caches and discard and other such structures.
However, it makes more sense to move this to libxfs than to keep
this separation because we are now moving to use struct perags
everywhere in the code instead of passing raw agnumber_t values
about. Hence we shoudl really move the support infrastructure to
libxfs/xfs_ag.[ch].
To do this without breaking userspace, first we need to rearrange
the structures and code so that all the kernel specific code is
located together. This makes it simple for userspace to ifdef out
the all the parts it does not need, minimising the code differences
between kernel and userspace. The next commit will do the move...
Signed-off-by: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Brian Foster <bfoster@redhat.com>
Reviewed-by: Darrick J. Wong <djwong@kernel.org>
Introduce an in-core counter to track the sum of all allocbt blocks
used by the filesystem. This value is currently tracked per-ag via
the ->agf_btreeblks field in the AGF, which also happens to include
rmapbt blocks. A global, in-core count of allocbt blocks is required
to identify the subset of global ->m_fdblocks that consists of
unavailable blocks currently used for allocation btrees. To support
this calculation at block reservation time, construct a similar
global counter for allocbt blocks, populate it on first read of each
AGF and update it as allocbt blocks are used and released.
Signed-off-by: Brian Foster <bfoster@redhat.com>
Reviewed-by: Chandan Babu R <chandanrlinux@gmail.com>
Reviewed-by: Allison Henderson <allison.henderson@oracle.com>
Reviewed-by: Darrick J. Wong <djwong@kernel.org>
Signed-off-by: Darrick J. Wong <djwong@kernel.org>
Since we're about to start using the blockgc workqueue to dispose of
inactivated inodes, strip the "block" prefix from the name; now it's
merely the general garbage collection (gc) workqueue.
Signed-off-by: Darrick J. Wong <djwong@kernel.org>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Split the block preallocation garbage collection work into per-AG work
items so that we can take advantage of parallelization.
Signed-off-by: Darrick J. Wong <djwong@kernel.org>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Remove the separate cowblocks work items and knob so that we can control
and run everything from a single blockgc work queue. Note that the
speculative_prealloc_lifetime sysfs knob retains its historical name
even though the functions move to prefix xfs_blockgc_*.
Signed-off-by: Darrick J. Wong <djwong@kernel.org>
Reviewed-by: Christoph Hellwig <hch@lst.de>
xfs_log_sbcount() calls xfs_sync_sb() to sync superblock counters to
disk when lazy superblock accounting is enabled. This occurs on
unmount, freeze, and read-only (re)mount and ensures the final
values are calculated and persisted to disk before each form of
quiesce completes.
Now that log covering occurs in all of these contexts and uses the
same xfs_sync_sb() mechanism to update log state, there is no need
to log the superblock separately for any reason. Update the log
quiesce path to sync the superblock at least once for any mount
where lazy superblock accounting is enabled. If the log is already
covered, it will remain in the covered state. Otherwise, the next
sync as part of the normal covering sequence will carry the
associated superblock update with it. Remove xfs_log_sbcount() now
that it is no longer needed.
Signed-off-by: Brian Foster <bfoster@redhat.com>
Reviewed-by: Darrick J. Wong <darrick.wong@oracle.com>
Reviewed-by: Darrick J. Wong <djwong@kernel.org>
Signed-off-by: Darrick J. Wong <djwong@kernel.org>
Reviewed-by: Allison Henderson <allison.henderson@oracle.com>
Merge xfs_getsb into its only caller, and clean that one up a little bit
as well.
Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Darrick J. Wong <darrick.wong@oracle.com>
Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
Inode reclaim will still throttle direct reclaim on the per-ag
reclaim locks. This is no longer necessary as reclaim can run
non-blocking now. Hence we can remove these locks so that we don't
arbitrarily block reclaimers just because there are more direct
reclaimers than there are AGs.
This can result in multiple reclaimers working on the same range of
an AG, but this doesn't cause any apparent issues. Optimising the
spread of concurrent reclaimers for best efficiency can be done in a
future patchset.
Signed-off-by: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Darrick J. Wong <darrick.wong@oracle.com>
Reviewed-by: Brian Foster <bfoster@redhat.com>
Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
It's a global atomic counter, and we are hitting it at a rate of
half a million transactions a second, so it's bouncing the counter
cacheline all over the place on large machines. We don't actually
need it anymore - it used to be required because the VFS freeze code
could not track/prevent filesystem transactions that were running,
but that problem no longer exists.
Hence to remove the counter, we simply have to ensure that nothing
calls xfs_sync_sb() while we are trying to quiesce the filesytem.
That only happens if the log worker is still running when we call
xfs_quiesce_attr(). The log worker is cancelled at the end of
xfs_quiesce_attr() by calling xfs_log_quiesce(), so just call it
early here and then we can remove the counter altogether.
Concurrent create, 50 million inodes, identical 16p/16GB virtual
machines on different physical hosts. Machine A has twice the CPU
cores per socket of machine B:
unpatched patched
machine A: 3m16s 2m00s
machine B: 4m04s 4m05s
Create rates:
unpatched patched
machine A: 282k+/-31k 468k+/-21k
machine B: 231k+/-8k 233k+/-11k
Concurrent rm of same 50 million inodes:
unpatched patched
machine A: 6m42s 2m33s
machine B: 4m47s 4m47s
The transaction rate on the fast machine went from just under
300k/sec to 700k/sec, which indicates just how much of a bottleneck
this atomic counter was.
Signed-off-by: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Darrick J. Wong <darrick.wong@oracle.com>
Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
Seeing massive cpu usage from xfs_agino_range() on one machine;
instruction level profiles look similar to another machine running
the same workload, only one machine is consuming 10x as much CPU as
the other and going much slower. The only real difference between
the two machines is core count per socket. Both are running
identical 16p/16GB virtual machine configurations
Machine A:
25.83% [k] xfs_agino_range
12.68% [k] __xfs_dir3_data_check
6.95% [k] xfs_verify_ino
6.78% [k] xfs_dir2_data_entry_tag_p
3.56% [k] xfs_buf_find
2.31% [k] xfs_verify_dir_ino
2.02% [k] xfs_dabuf_map.constprop.0
1.65% [k] xfs_ag_block_count
And takes around 13 minutes to remove 50 million inodes.
Machine B:
13.90% [k] __pv_queued_spin_lock_slowpath
3.76% [k] do_raw_spin_lock
2.83% [k] xfs_dir3_leaf_check_int
2.75% [k] xfs_agino_range
2.51% [k] __raw_callee_save___pv_queued_spin_unlock
2.18% [k] __xfs_dir3_data_check
2.02% [k] xfs_log_commit_cil
And takes around 5m30s to remove 50 million inodes.
Suspect is cacheline contention on m_sectbb_log which is used in one
of the macros in xfs_agino_range. This is a read-only variable but
shares a cacheline with m_active_trans which is a global atomic that
gets bounced all around the machine.
The workload is trying to run hundreds of thousands of transactions
per second and hence cacheline contention will be occurring on this
atomic counter. Hence xfs_agino_range() is likely just be an
innocent bystander as the cache coherency protocol fights over the
cacheline between CPU cores and sockets.
On machine A, this rearrangement of the struct xfs_mount
results in the profile changing to:
9.77% [kernel] [k] xfs_agino_range
6.27% [kernel] [k] __xfs_dir3_data_check
5.31% [kernel] [k] __pv_queued_spin_lock_slowpath
4.54% [kernel] [k] xfs_buf_find
3.79% [kernel] [k] do_raw_spin_lock
3.39% [kernel] [k] xfs_verify_ino
2.73% [kernel] [k] __raw_callee_save___pv_queued_spin_unlock
Vastly less CPU usage in xfs_agino_range(), but still 3x the amount
of machine B and still runs substantially slower than it should.
Current rm -rf of 50 million files:
vanilla patched
machine A 13m20s 6m42s
machine B 5m30s 5m02s
It's an improvement, hence indicating that separation and further
optimisation of read-only global filesystem data is worthwhile, but
it clearly isn't the underlying issue causing this specific
performance degradation.
Signed-off-by: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Darrick J. Wong <darrick.wong@oracle.com>
Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
Shaokun Zhang reported that XFS was using substantial CPU time in
percpu_count_sum() when running a single threaded benchmark on
a high CPU count (128p) machine from xfs_mod_ifree(). The issue
is that the filesystem is empty when the benchmark runs, so inode
allocation is running with a very low inode free count.
With the percpu counter batching, this means comparisons when the
counter is less that 128 * 256 = 32768 use the slow path of adding
up all the counters across the CPUs, and this is expensive on high
CPU count machines.
The summing in xfs_mod_ifree() is only used to fire an assert if an
underrun occurs. The error is ignored by the higher level code.
Hence this is really just debug code and we don't need to run it
on production kernels, nor do we need such debug checks to return
error values just to trigger an assert.
Finally, xfs_mod_icount/xfs_mod_ifree are only called from
xfs_trans_unreserve_and_mod_sb(), so get rid of them and just
directly call the percpu_counter_add/percpu_counter_compare
functions. The compare functions are now run only on debug builds as
they are internal to ASSERT() checks and so only compiled in when
ASSERTs are active (CONFIG_XFS_DEBUG=y or CONFIG_XFS_WARN=y).
Reported-by: Shaokun Zhang <zhangshaokun@hisilicon.com>
Signed-off-by: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Darrick J. Wong <darrick.wong@oracle.com>
Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
Both types control shutdown messaging and neither is used in the
current codebase.
Signed-off-by: Brian Foster <bfoster@redhat.com>
Reviewed-by: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Darrick J. Wong <darrick.wong@oracle.com>
Reviewed-by: Allison Collins <allison.henderson@oracle.com>
Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
As agreed upon[1]. We make the dax mount option a tri-state. '-o dax'
continues to operate the same. We add 'always', 'never', and 'inode'
(default).
[1] https://lore.kernel.org/lkml/20200405061945.GA94792@iweiny-DESK2.sc.intel.com/
Signed-off-by: Ira Weiny <ira.weiny@intel.com>
Reviewed-by: Darrick J. Wong <darrick.wong@oracle.com>
Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
In prep for the new tri-state mount option which then introduces
XFS_MOUNT_DAX_NEVER.
Reviewed-by: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Darrick J. Wong <darrick.wong@oracle.com>
Signed-off-by: Ira Weiny <ira.weiny@intel.com>
Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
Move the inode dirty data flushing to a workqueue so that multiple
threads can take advantage of a single thread's flushing work. The
ratelimiting technique used in bdd4ee4 was not successful, because
threads that skipped the inode flush scan due to ratelimiting would
ENOSPC early, which caused occasional (but noticeable) changes in
behavior and sporadic fstest regressions.
Therefore, make all the writer threads wait on a single inode flush,
which eliminates both the stampeding hordes of flushers and the small
window in which a write could fail with ENOSPC because it lost the
ratelimit race after even another thread freed space.
Fixes: c6425702f2 ("xfs: ratelimit inode flush on buffered write ENOSPC")
Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
Reviewed-by: Brian Foster <bfoster@redhat.com>
A customer reported rcu stalls and softlockup warnings on a computer
with many CPU cores and many many more IO threads trying to write to a
filesystem that is totally out of space. Subsequent analysis pointed to
the many many IO threads calling xfs_flush_inodes -> sync_inodes_sb,
which causes a lot of wb_writeback_work to be queued. The writeback
worker spends so much time trying to wake the many many threads waiting
for writeback completion that it trips the softlockup detector, and (in
this case) the system automatically reboots.
In addition, they complain that the lengthy xfs_flush_inodes scan traps
all of those threads in uninterruptible sleep, which hampers their
ability to kill the program or do anything else to escape the situation.
If there's thousands of threads trying to write to files on a full
filesystem, each of those threads will start separate copies of the
inode flush scan. This is kind of pointless since we only need one
scan, so rate limit the inode flush.
Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
Reviewed-by: Dave Chinner <dchinner@redhat.com>
Remove some unused typedef'd simple types, and some unused
structure members.
Signed-off-by: Eric Sandeen <sandeen@redhat.com>
Reviewed-by: Darrick J. Wong <darrick.wong@oracle.com>
Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
Instead of causing a relatively expensive indirect call for each
hashing and comparism of a file name in a directory just use an
inline function and a simple branch on the ASCII CI bit.
Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Darrick J. Wong <darrick.wong@oracle.com>
[darrick: fix unused variable warning]
Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Darrick J. Wong <darrick.wong@oracle.com>
Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Darrick J. Wong <darrick.wong@oracle.com>
Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
Move the node header size field to struct xfs_da_geometry, and remove
the now unused non-directory dir ops infrastructure.
Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Darrick J. Wong <darrick.wong@oracle.com>
Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
Eliminate struct xfs_mount field m_fsname by using the super block s_id
field directly.
Signed-off-by: Ian Kent <raven@themaw.net>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Darrick J. Wong <darrick.wong@oracle.com>
Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
The struct xfs_mount field m_fsname_len is not used anywhere, remove it.
Signed-off-by: Ian Kent <raven@themaw.net>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Darrick J. Wong <darrick.wong@oracle.com>
Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
Replace XFS_MOUNT_COMPAT_IOSIZE with an inverted XFS_MOUNT_LARGEIO flag
that makes the usage more clear.
Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Eric Sandeen <sandeen@redhat.com>
Reviewed-by: Darrick J. Wong <darrick.wong@oracle.com>
Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
