| Age | Commit message (Collapse) | Author |
|---|
|
Add a zoned group with an attribute for the maximum number of open zones.
This allows querying the open zones for data placement tests, or also
for placement aware applications that are in control of the entire
file system.
Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: "Darrick J. Wong" <djwong@kernel.org>
|
|
Extend the error sysfs initialization helper to include the neighbouring
attributes as well.
Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: "Darrick J. Wong" <djwong@kernel.org>
|
|
Add the per-zone life time hint and the used block distribution
for fully written zones, grouping reclaimable zones in fixed-percentage
buckets spanning 0..9%, 10..19% and full zones as 100% used as well as a
few statistics about the zone allocator and open and reclaimable zones
in /proc/*/mountstats.
This gives good insight into data fragmentation and data placement
success rate.
Signed-off-by: Hans Holmberg <hans.holmberg@wdc.com>
Co-developed-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: "Darrick J. Wong" <djwong@kernel.org>
|
|
The show_stats option allows a file system to dump plain text statistic
on a per-mount basis into /proc/*/mountstats. Wire up a no-op version
which will grow useful information for zoned file systems later.
Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: "Darrick J. Wong" <djwong@kernel.org>
|
|
Add a file write life time data placement allocation scheme that aims to
minimize fragmentation and thereby to do two things:
a) separate file data to different zones when possible.
b) colocate file data of similar life times when feasible.
To get best results, average file sizes should align with the zone
capacity that is reported through the XFS_IOC_FSGEOMETRY ioctl.
This improvement in data placement efficiency reduces the number of
blocks requiring relocation by GC, and thus decreases overall write
amplification. The impact on performance varies depending on how full
the file system is.
For RocksDB using leveled compaction, the lifetime hints can improve
throughput for overwrite workloads at 80% file system utilization by
~10%, but for lower file system utilization there won't be as much
benefit in application performance as there is less need for garbage
collection to start with.
Lifetime hints can be disabled using the nolifetime mount option.
Signed-off-by: Hans Holmberg <hans.holmberg@wdc.com>
Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: "Darrick J. Wong" <djwong@kernel.org>
|
|
Allow limiting the number of open zones used below that exported by the
device. This is required to tune the number of write streams when zoned
RT devices are used on conventional devices, and can be useful on zoned
devices that support a very large number of open zones.
Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: "Darrick J. Wong" <djwong@kernel.org>
|
|
Zoned devices can have gaps beyond the usable capacity of a zone and the
end in the LBA/daddr address space. In other words, the hardware
equivalent to the RT groups already takes care of the power of 2
alignment for us. In this case the sparse FSB/RTB address space maps 1:1
to the device address space.
Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: "Darrick J. Wong" <djwong@kernel.org>
|
|
Enable the zoned RT device directory feature. With this feature, RT
groups are written sequentially and always emptied before rewriting
the blocks. This perfectly maps to zoned devices, but can also be
used on conventional block devices.
Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: "Darrick J. Wong" <djwong@kernel.org>
|
|
They'll need a little more work.
Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: "Darrick J. Wong" <djwong@kernel.org>
|
|
While the zoned on-disk format supports reflinks, the GC code currently
always unshares reflinks when moving blocks to new zones, thus making the
feature unusuable. Disable reflinks until the GC code is refcount aware.
Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: "Darrick J. Wong" <djwong@kernel.org>
|
|
File system with internal RT devices are a bit odd in that we need
to report AGs and RGs. To make this happen use separate synthetic
fmr_device values for the different sections instead of the dev_t
mapping used by other XFS configurations.
The data device is reported as file system metadata before the
start of the RGs for the synthetic RT fmr_device.
Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: "Darrick J. Wong" <djwong@kernel.org>
|
|
Space usage is tracked by the rmap, which already is separately
cross-referenced. But on top of that we have the write pointer and can
do a basic sanity check here that the block is not beyond the write
pointer.
Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: "Darrick J. Wong" <djwong@kernel.org>
|
|
Space usage is tracked by the rmap, which already is separately
cross-referenced. But on top of that we have the write pointer and can
do a basic sanity check here that the block is not beyond the write
pointer.
Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: "Darrick J. Wong" <djwong@kernel.org>
|
|
Zoned file systems can have COW forks even without reflinks.
Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: "Darrick J. Wong" <djwong@kernel.org>
|
|
Replace the inner loop growing one RT bitmap block at a time with
one just modifying the superblock counters for growing an entire
zone (aka RTG). The big restriction is just like at mkfs time only
a RT extent size of a single FSB is allowed, and the file system
capacity needs to be aligned to the zone size.
Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: "Darrick J. Wong" <djwong@kernel.org>
|
|
File systems with a zoned RT device have a large number of reserved
blocks that are required for garbage collection, and which can't be
filled with user data. Exclude them from the available blocks reported
through stat(v)fs.
Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: "Darrick J. Wong" <djwong@kernel.org>
|
|
Make xfs_rtextent_free_finish_item call into the zoned allocator to free
blocks on zoned RT devices.
Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: "Darrick J. Wong" <djwong@kernel.org>
|
|
Direct writes to zoned RT devices are extremely simple. After taking the
block reservation before acquiring the iolock, the iomap direct I/O calls
into ->iomap_begin which will return a "fake" iomap for the entire
requested range. The actual block allocation is then done from the
submit_io handler using code shared with the buffered I/O path.
The iomap_dio_ops set the bio_set to the (iomap) ioend one and initialize
the embedded ioend, which allows reusing the existing ioend based buffered
I/O completion path.
Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: "Darrick J. Wong" <djwong@kernel.org>
|
|
Implement buffered writes including page faults and block zeroing for
zoned RT devices. Buffered writes to zoned RT devices are split into
three phases:
1) a reservation for the worst case data block usage is taken before
acquiring the iolock. When there are enough free blocks but not
enough available one, garbage collection is kicked off to free the
space before continuing with the write. If there isn't enough
freeable space, the block reservation is reduced and a short write
will happen as expected by normal Linux write semantics.
2) with the iolock held, the generic iomap buffered write code is
called, which through the iomap_begin operation usually just inserts
delalloc extents for the range in a single iteration. Only for
overwrites of existing data that are not block aligned, or zeroing
operations the existing extent mapping is read to fill out the srcmap
and to figure out if zeroing is required.
3) the ->map_blocks callback to the generic iomap writeback code
calls into the zoned space allocator to actually allocate on-disk
space for the range before kicking of the writeback.
Note that because all writes are out of place, truncate or hole punches
that are not aligned to block size boundaries need to allocate space.
For block zeroing from truncate, ->setattr is called with the iolock
(aka i_rwsem) already held, so a hacky deviation from the above
scheme is needed. In this case the space reservations is called with
the iolock held, but is required not to block and can dip into the
reserved block pool. This can lead to -ENOSPC when truncating a
file, which is unfortunate. But fixing the calling conventions in
the VFS is probably much easier with code requiring it already in
mainline.
Similarly because all writes are out place, the zoned allocator can't
support unwritten extents and thus the FALLOC_FL_ALLOCATE_RANGE range
mode of fallocate. Other fallocate modes that would reserved space
but don't need to to provide proper semantics do work but do not
reserve space.
Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: "Darrick J. Wong" <djwong@kernel.org>
|
|
RT groups on a zoned file system need to be completely empty before their
space can be reused. This means that partially empty groups need to be
emptied entirely to free up space if no entirely free groups are
available.
Add a garbage collection thread that moves all data out of the least used
zone when not enough free zones are available, and which resets all zones
that have been emptied. To find empty zone a simple set of 10 buckets
based on the amount of space used in the zone is used. To empty zones,
the rmap is walked to find the owners and the data is read and then
written to the new place.
To automatically defragment files the rmap records are sorted by inode
and logical offset. This means defragmentation of parallel writes into
a single zone happens automatically when performing garbage collection.
Because holding the iolock over the entire GC cycle would inject very
noticeable latency for other accesses to the inodes, the iolock is not
taken while performing I/O. Instead the I/O completion handler checks
that the mapping hasn't changed over the one recorded at the start of
the GC cycle and doesn't update the mapping if it change.
Co-developed-by: Hans Holmberg <hans.holmberg@wdc.com>
Signed-off-by: Hans Holmberg <hans.holmberg@wdc.com>
Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: "Darrick J. Wong" <djwong@kernel.org>
|
|
For zoned file systems garbage collection (GC) has to take the iolock
and mmaplock after moving data to a new place to synchronize with
readers. This means waiting for garbage collection with the iolock can
deadlock.
To avoid this, the worst case required blocks have to be reserved before
taking the iolock, which is done using a new RTAVAILABLE counter that
tracks blocks that are free to write into and don't require garbage
collection. The new helpers try to take these available blocks, and
if there aren't enough available it wakes and waits for GC. This is
done using a list of on-stack reservations to ensure fairness.
Co-developed-by: Hans Holmberg <hans.holmberg@wdc.com>
Signed-off-by: Hans Holmberg <hans.holmberg@wdc.com>
Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: "Darrick J. Wong" <djwong@kernel.org>
|
|
For zoned RT devices space is always allocated at the write pointer, that
is right after the last written block and only recorded on I/O completion.
Because the actual allocation algorithm is very simple and just involves
picking a good zone - preferably the one used for the last write to the
inode. As the number of zones that can written at the same time is
usually limited by the hardware, selecting a zone is done as late as
possible from the iomap dio and buffered writeback bio submissions
helpers just before submitting the bio.
Given that the writers already took a reservation before acquiring the
iolock, space will always be readily available if an open zone slot is
available. A new structure is used to track these open zones, and
pointed to by the xfs_rtgroup. Because zoned file systems don't have
a rsum cache the space for that pointer can be reused.
Allocations are only recorded at I/O completion time. The scheme used
for that is very similar to the reflink COW end I/O path.
Co-developed-by: Hans Holmberg <hans.holmberg@wdc.com>
Signed-off-by: Hans Holmberg <hans.holmberg@wdc.com>
Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: "Darrick J. Wong" <djwong@kernel.org>
|
|
Add support to validate and parse reported hardware zone state.
Co-developed-by: Hans Holmberg <hans.holmberg@wdc.com>
Signed-off-by: Hans Holmberg <hans.holmberg@wdc.com>
Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: "Darrick J. Wong" <djwong@kernel.org>
|
|
The zoned allocator never performs speculative preallocations, so don't
bother queueing up zoned inodes here.
Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: "Darrick J. Wong" <djwong@kernel.org>
|
|
Zoned file systems require out of place writes and thus can't support
post-EOF speculative preallocations. Avoid the pointless ilock critical
section to find out that none can be freed.
Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: "Darrick J. Wong" <djwong@kernel.org>
|
|
The zoned allocator unconditionally issues zone resets or discards after
emptying an entire zone, so supporting FITRIM for a zoned RT device is
not useful.
Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: "Darrick J. Wong" <djwong@kernel.org>
|
|
Zoned file systems not only don't use the global frextents counter, but
for them the in-memory percpu counter also includes reservations taken
before even allocating delalloc extent records, so it will never match
the per-zone used information. Disable all updates and verification of
the sb counter for zoned file systems as it isn't useful for them.
Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: "Darrick J. Wong" <djwong@kernel.org>
|
|
Export the zoned geometry information so that userspace can query it.
Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: "Darrick J. Wong" <djwong@kernel.org>
|
|
Allow creating an RT subvolume on the same device as the main data
device. This is mostly used for SMR HDDs where the conventional zones
are used for the data device and the sequential write required zones
for the zoned RT section.
Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: "Darrick J. Wong" <djwong@kernel.org>
|
|
Zone file systems reuse the basic RT group enabled XFS file system
structure to support a mode where each RT group is always written from
start to end and then reset for reuse (after moving out any remaining
data). There are few minor but important changes, which are indicated
by a new incompat flag:
1) there are no bitmap and summary inodes, thus the
/rtgroups/{rgno}.{bitmap,summary} metadir files do not exist and the
sb_rbmblocks superblock field must be cleared to zero.
2) there is a new superblock field that specifies the start of an
internal RT section. This allows supporting SMR HDDs that have random
writable space at the beginning which is used for the XFS data device
(which really is the metadata device for this configuration), directly
followed by a RT device on the same block device. While something
similar could be achieved using dm-linear just having a single device
directly consumed by XFS makes handling the file systems a lot easier.
3) Another superblock field that tracks the amount of reserved space (or
overprovisioning) that is never used for user capacity, but allows GC
to run more smoothly.
4) an overlay of the cowextsize field for the rtrmap inode so that we
can persistently track the total amount of rtblocks currently used in
a RT group. There is no data structure other than the rmap that
tracks used space in an RT group, and this counter is used to decide
when a RT group has been entirely emptied, and to select one that
is relatively empty if garbage collection needs to be performed.
While this counter could be tracked entirely in memory and rebuilt
from the rmap at mount time, that would lead to very long mount times
with the large number of RT groups implied by the number of hardware
zones especially on SMR hard drives with 256MB zone sizes.
Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: "Darrick J. Wong" <djwong@kernel.org>
|
|
Add a helper to find the last offset mapped in the rtrmap. This will be
used by the zoned code to find out where to start writing again on
conventional devices without hardware zone support.
Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: "Darrick J. Wong" <djwong@kernel.org>
|
|
The zone allocator wants to be able to remove a delalloc mapping in the
COW fork while keeping the block reservation. To support that pass the
flags argument down to xfs_bmap_del_extent_delay and support the
XFS_BMAPI_REMAP flag to keep the reservation.
Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: "Darrick J. Wong" <djwong@kernel.org>
|
|
For always COW inodes we also must check the alignment of each individual
iovec segment, as they could end up with different I/Os due to the way
bio_iov_iter_get_pages works, and we'd then overwrite an already written
block. The existing always_cow sysctl based code doesn't catch this
because nothing enforces that blocks aren't rewritten, but for zoned XFS
on sequential write required zones this is a hard error.
Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: "Darrick J. Wong" <djwong@kernel.org>
|
|
xfs_reflink_trim_around_shared tries to find shared blocks in the
refcount btree. Always_cow inodes don't have that tree, so don't
bother.
For the existing always_cow code this is a minor optimization. For
the upcoming zoned code that can do COW without the rtreflink code it
avoids triggering a NULL pointer dereference.
Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: "Darrick J. Wong" <djwong@kernel.org>
|
|
Delalloc reservations are not supported in userspace, and thus it doesn't
make sense to share this helper with xfsprogs.c. Move it to xfs_iomap.c
toward the two callers.
Note that there rest of the delalloc handling should probably eventually
also move out of xfs_bmap.c, but that will require a bit more surgery.
Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: "Darrick J. Wong" <djwong@kernel.org>
|
|
Shortcut dereferencing the xg_block_count field in the generic group
structure.
Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: "Darrick J. Wong" <djwong@kernel.org>
|
|
Add a helper to check that the last block of a RT device is readable
to share the code between mount and growfs. This also adds the mount
time overflow check to growfs and improves the error messages.
Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: "Darrick J. Wong" <djwong@kernel.org>
|
|
There is no point in reserving more space than actually available
on the data device for the worst case scenario that is unlikely to
happen. Reserve at most 1/4th of the data device blocks, which is
still a heuristic.
Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: "Darrick J. Wong" <djwong@kernel.org>
|
|
Currently each metabtree inode has it's own space reservation to ensure
it can be expanded to the maximum size, mirroring what is done for the
AG-based btrees. But unlike the AG-based btrees the metabtree inodes
aren't restricted to allocate from a single AG but can use free space
form the entire file system. And unlike AG-based btrees where the
required reservation shrinks with the available free space due to this,
the metabtree reservations for the rtrmap and rtfreflink trees are not
bound in any way by the data device free space as they track RT extent
allocations. This is not very efficient as it requires a large number
of blocks to be set aside that can't be used at all by other btrees.
Switch to a model that uses a global pool instead in preparation for
reducing the amount of reserved space, which now also removes the
overloading of the i_nblocks field for metabtree inodes, which would
create problems if metabtree inodes ever had a big enough xattr fork
to require xattr blocks outside the inode.
Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: "Darrick J. Wong" <djwong@kernel.org>
|
|
All callers of xrep_reap_metadir_fsblocks need to fix up the metabtree
reservation, otherwise they'd leave the reservations in an incoherent
state. Move the call to xrep_reset_metafile_resv into
xrep_reap_metadir_fsblocks so it always is taken care of, and remove
now superfluous helper functions in the callers.
Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: "Darrick J. Wong" <djwong@kernel.org>
|
|
Add two tracepoints when the freecounter dips into the reserved pool
and when it is entirely out of space.
Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: "Darrick J. Wong" <djwong@kernel.org>
|
|
The zoned space allocator will need reserved RT extents for garbage
collection and zeroing of partial blocks. Move the resblks related
fields into the freecounter array so that they can be used for all
counters.
Co-developed-by: Hans Holmberg <hans.holmberg@wdc.com>
Signed-off-by: Hans Holmberg <hans.holmberg@wdc.com>
Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: "Darrick J. Wong" <djwong@kernel.org>
|
|
xfs_{add,dec}_freecounter already handles the block and RT extent
percpu counters, but it currently hardcodes the passed in counter.
Add a freecounter abstraction that uses an enum to designate the counter
and add wrappers that hide the actual percpu_counters. This will allow
expanding the reserved block handling to the RT extent counter in the
next step, and also prepares for adding yet another such counter that
can share the code. Both these additions will be needed for the zoned
allocator.
Also switch the flooring of the frextents counter to 0 in statfs for the
rthinherit case to a manual min_t call to match the handling of the
fdblocks counter for normal file systems.
Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: "Darrick J. Wong" <djwong@kernel.org>
|
|
Let the successful allocation be the main path through the function
with exception handling in branches to make the code easier to
follow.
Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: "Darrick J. Wong" <djwong@kernel.org>
|
|
git://git.kernel.org/pub/scm/linux/kernel/git/vfs/vfs into xfs-6.15-merge
|
|
Some filesystems, such as NFS, cifs, ceph, and fuse, do not have
complete control of sequencing on the actual filesystem (e.g. on a
different server) and may find that the inode created for a mkdir
request already exists in the icache and dcache by the time the mkdir
request returns. For example, if the filesystem is mounted twice the
directory could be visible on the other mount before it is on the
original mount, and a pair of name_to_handle_at(), open_by_handle_at()
calls could instantiate the directory inode with an IS_ROOT() dentry
before the first mkdir returns.
This means that the dentry passed to ->mkdir() may not be the one that
is associated with the inode after the ->mkdir() completes. Some
callers need to interact with the inode after the ->mkdir completes and
they currently need to perform a lookup in the (rare) case that the
dentry is no longer hashed.
This lookup-after-mkdir requires that the directory remains locked to
avoid races. Planned future patches to lock the dentry rather than the
directory will mean that this lookup cannot be performed atomically with
the mkdir.
To remove this barrier, this patch changes ->mkdir to return the
resulting dentry if it is different from the one passed in.
Possible returns are:
NULL - the directory was created and no other dentry was used
ERR_PTR() - an error occurred
non-NULL - this other dentry was spliced in
This patch only changes file-systems to return "ERR_PTR(err)" instead of
"err" or equivalent transformations. Subsequent patches will make
further changes to some file-systems to return a correct dentry.
Not all filesystems reliably result in a positive hashed dentry:
- NFS, cifs, hostfs will sometimes need to perform a lookup of
the name to get inode information. Races could result in this
returning something different. Note that this lookup is
non-atomic which is what we are trying to avoid. Placing the
lookup in filesystem code means it only happens when the filesystem
has no other option.
- kernfs and tracefs leave the dentry negative and the ->revalidate
operation ensures that lookup will be called to correctly populate
the dentry. This could be fixed but I don't think it is important
to any of the users of vfs_mkdir() which look at the dentry.
The recommendation to use
d_drop();d_splice_alias()
is ugly but fits with current practice. A planned future patch will
change this.
Reviewed-by: Jeff Layton <jlayton@kernel.org>
Reviewed-by: Jan Kara <jack@suse.cz>
Signed-off-by: NeilBrown <neilb@suse.de>
Link: https://lore.kernel.org/r/20250227013949.536172-2-neilb@suse.de
Signed-off-by: Christian Brauner <brauner@kernel.org>
|
|
Read side was already fully supported, and with the write side
appropriately punted to the worker queue, all that's needed now is
setting FOP_DONTCACHE in the file_operations structure to enable full
support for read and write uncached IO.
This provides similar benefits to using RWF_DONTCACHE with reads. Testing
buffered writes on 32 files:
writing bs 65536, uncached 0
1s: 196035MB/sec
2s: 132308MB/sec
3s: 132438MB/sec
4s: 116528MB/sec
5s: 103898MB/sec
6s: 108893MB/sec
7s: 99678MB/sec
8s: 106545MB/sec
9s: 106826MB/sec
10s: 101544MB/sec
11s: 111044MB/sec
12s: 124257MB/sec
13s: 116031MB/sec
14s: 114540MB/sec
15s: 115011MB/sec
16s: 115260MB/sec
17s: 116068MB/sec
18s: 116096MB/sec
where it's quite obvious where the page cache filled, and performance
dropped from to about half of where it started, settling in at around
115GB/sec. Meanwhile, 32 kswapds were running full steam trying to
reclaim pages.
Running the same test with uncached buffered writes:
writing bs 65536, uncached 1
1s: 198974MB/sec
2s: 189618MB/sec
3s: 193601MB/sec
4s: 188582MB/sec
5s: 193487MB/sec
6s: 188341MB/sec
7s: 194325MB/sec
8s: 188114MB/sec
9s: 192740MB/sec
10s: 189206MB/sec
11s: 193442MB/sec
12s: 189659MB/sec
13s: 191732MB/sec
14s: 190701MB/sec
15s: 191789MB/sec
16s: 191259MB/sec
17s: 190613MB/sec
18s: 191951MB/sec
and the behavior is fully predictable, performing the same throughout
even after the page cache would otherwise have fully filled with dirty
data. It's also about 65% faster, and using half the CPU of the system
compared to the normal buffered write.
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Link: https://lore.kernel.org/r/20250204184047.356762-3-axboe@kernel.dk
Reviewed-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: "Darrick J. Wong" <djwong@kernel.org>
Signed-off-by: Christian Brauner <brauner@kernel.org>
|
|
Read side was already fully supported, and with the write side
appropriately punted to the worker queue, all that's needed now is
setting FOP_DONTCACHE in the file_operations structure to enable full
support for read and write uncached IO.
This provides similar benefits to using RWF_DONTCACHE with reads. Testing
buffered writes on 32 files:
writing bs 65536, uncached 0
1s: 196035MB/sec
2s: 132308MB/sec
3s: 132438MB/sec
4s: 116528MB/sec
5s: 103898MB/sec
6s: 108893MB/sec
7s: 99678MB/sec
8s: 106545MB/sec
9s: 106826MB/sec
10s: 101544MB/sec
11s: 111044MB/sec
12s: 124257MB/sec
13s: 116031MB/sec
14s: 114540MB/sec
15s: 115011MB/sec
16s: 115260MB/sec
17s: 116068MB/sec
18s: 116096MB/sec
where it's quite obvious where the page cache filled, and performance
dropped from to about half of where it started, settling in at around
115GB/sec. Meanwhile, 32 kswapds were running full steam trying to
reclaim pages.
Running the same test with uncached buffered writes:
writing bs 65536, uncached 1
1s: 198974MB/sec
2s: 189618MB/sec
3s: 193601MB/sec
4s: 188582MB/sec
5s: 193487MB/sec
6s: 188341MB/sec
7s: 194325MB/sec
8s: 188114MB/sec
9s: 192740MB/sec
10s: 189206MB/sec
11s: 193442MB/sec
12s: 189659MB/sec
13s: 191732MB/sec
14s: 190701MB/sec
15s: 191789MB/sec
16s: 191259MB/sec
17s: 190613MB/sec
18s: 191951MB/sec
and the behavior is fully predictable, performing the same throughout
even after the page cache would otherwise have fully filled with dirty
data. It's also about 65% faster, and using half the CPU of the system
compared to the normal buffered write.
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Link: https://lore.kernel.org/r/20250204184047.356762-3-axboe@kernel.dk
Reviewed-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: "Darrick J. Wong" <djwong@kernel.org>
Signed-off-by: Christian Brauner <brauner@kernel.org>
|
|
xfs_buf_stale already set b_lru_ref to 0, and thus prevents the buffer
from moving to the LRU. Remove the duplicate check.
Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Darrick J. Wong <djwong@kernel.org>
Signed-off-by: Carlos Maiolino <cem@kernel.org>
|
|
The buffer cache keeps a bt_io_count per-CPU counter to track all
in-flight I/O, which is used to ensure no I/O is in flight when
unmounting the file system.
For most I/O we already keep track of inflight I/O at higher levels:
- for synchronous I/O (xfs_buf_read/xfs_bwrite/xfs_buf_delwri_submit),
the caller has a reference and waits for I/O completions using
xfs_buf_iowait
- for xfs_buf_delwri_submit_nowait the only caller (AIL writeback)
tracks the log items that the buffer attached to
This only leaves only xfs_buf_readahead_map as a submitter of
asynchronous I/O that is not tracked by anything else. Replace the
bt_io_count per-cpu counter with a more specific bt_readahead_count
counter only tracking readahead I/O. This allows to simply increment
it when submitting readahead I/O and decrementing it when it completed,
and thus simplify xfs_buf_rele and remove the needed for the
XBF_NO_IOACCT flags and the XFS_BSTATE_IN_FLIGHT buffer state.
Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Darrick J. Wong <djwong@kernel.org>
Signed-off-by: Carlos Maiolino <cem@kernel.org>
|
