xfs: timestamp updates cause excessive fdatasync log traffic
Sage Weil reported that a ceph test workload was writing to the
log on every fdatasync during an overwrite workload. Event tracing
showed that the only metadata modification being made was the
timestamp updates during the write(2) syscall, but fdatasync(2)
is supposed to ignore them. The key observation was that the
transactions in the log all looked like this:
INODE: #regs: 4 ino: 0x8b flags: 0x45 dsize: 32
And contained a flags field of 0x45 or 0x85, and had data and
attribute forks following the inode core. This means that the
timestamp updates were triggering dirty relogging of previously
logged parts of the inode that hadn't yet been flushed back to
disk.
There are two parts to this problem. The first is that XFS relogs
dirty regions in subsequent transactions, so it carries around the
fields that have been dirtied since the last time the inode was
written back to disk, not since the last time the inode was forced
into the log.
The second part is that on v5 filesystems, the inode change count
update during inode dirtying also sets the XFS_ILOG_CORE flag, so
on v5 filesystems this makes a timestamp update dirty the entire
inode.
As a result when fdatasync is run, it looks at the dirty fields in
the inode, and sees more than just the timestamp flag, even though
the only metadata change since the last fdatasync was just the
timestamps. Hence we force the log on every subsequent fdatasync
even though it is not needed.
To fix this, add a new field to the inode log item that tracks
changes since the last time fsync/fdatasync forced the log to flush
the changes to the journal. This flag is updated when we dirty the
inode, but we do it before updating the change count so it does not
carry the "core dirty" flag from timestamp updates. The fields are
zeroed when the inode is marked clean (due to writeback/freeing) or
when an fsync/datasync forces the log. Hence if we only dirty the
timestamps on the inode between fsync/fdatasync calls, the fdatasync
will not trigger another log force.
Over 100 runs of the test program:
Ext4 baseline:
runtime: 1.63s +/- 0.24s
avg lat: 1.59ms +/- 0.24ms
iops: ~2000
XFS, vanilla kernel:
runtime: 2.45s +/- 0.18s
avg lat: 2.39ms +/- 0.18ms
log forces: ~400/s
iops: ~1000
XFS, patched kernel:
runtime: 1.49s +/- 0.26s
avg lat: 1.46ms +/- 0.25ms
log forces: ~30/s
iops: ~1500
Reported-by: Sage Weil <sage@redhat.com>
Signed-off-by: Dave Chinner <dchinner@redhat.com>
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>
The typedef for timespecs and nanotime() are completely unnecessary,
and delay() can be moved to fs/xfs/linux.h, which means this file
can go away.
Signed-off-by: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Dave Chinner <david@fromorbit.com>
Michael L Semon reported that generic/069 runtime increased on v5
superblocks by 100% compared to v4 superblocks. his perf-based
analysis pointed directly at the timestamp updates being done by the
write path in this workload. The append writers are doing 4-byte
writes, so there are lots of timestamp updates occurring.
The thing is, they aren't being triggered by timestamp changes -
they are being triggered by the inode change counter needing to be
updated. That is, every write(2) system call needs to bump the inode
version count, and it does that through the timestamp update
mechanism. Hence for v5 filesystems, test generic/069 is running 3
orders of magnitude more timestmap update transactions on v5
filesystems due to the fact it does a huge number of *4 byte*
write(2) calls.
This isn't a real world scenario we really need to address - anyone
doing such sequential IO should be using fwrite(3), not write(2).
i.e. fwrite(3) buffers the writes in userspace to minimise the
number of write(2) syscalls, and the problem goes away.
However, there is a small change we can make to improve the
situation - removing the expensive lock operation on the change
counter update. All inode version counter changes in XFS occur
under the ip->i_ilock during a transaction, and therefore we
don't actually need the spin lock that provides exclusive access to
it through inc_inode_iversion().
Hence avoid the lock and just open code the increment ourselves when
logging the inode.
Reported-by: Michael L. Semon <mlsemon35@gmail.com>
Signed-off-by: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Ben Myers <bpm@sgi.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>
For CRC enabled filesystems, add support for the monotonic inode
version change counter that is needed by protocols like NFSv4 for
determining if the inode has changed in any way at all between two
unrelated operations on the inode.
This bumps the change count the first time an inode is dirtied in a
transaction. Since all modifications to the inode are logged, this
will catch all changes that are made to the inode, including
timestamp updates that occur during data writes.
Signed-off-by: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Mark Tinguely <tinguely@sgi.com>
Reviewed-by: Chandra Seetharaman <sekharan@us.ibm.com>
Signed-off-by: Ben Myers <bpm@sgi.com>
Remove the XFS_TRANS_DEBUG routines. They are no longer appropriate
and have not been used in years
Signed-off-by: Mark Tinguely <tinguely@sgi.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Dave Chinner <dchinner@redhat.com>
Signed-off-by: Ben Myers <bpm@sgi.com>
With the removal of xfs_rw.h and other changes over time, xfs_bit.h
is being included in many files that don't actually need it. Clean
up the includes as necessary.
Also move the only-used-once xfs_ialloc_find_free() static inline
function out of a header file that is widely included to reduce
the number of needless dependencies on xfs_bit.h.
Signed-off-by: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Mark Tinguely <tinguely@sgi.com>
Signed-off-by: Ben Myers <bpm@sgi.com>
Untangle the header file includes a bit by moving the definition of
xfs_agino_t to xfs_types.h. This removes the dependency that xfs_ag.h has on
xfs_inum.h, meaning we don't need to include xfs_inum.h everywhere we include
xfs_ag.h.
Signed-off-by: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Mark Tinguely <tinguely@sgi.com>
Signed-off-by: Ben Myers <bpm@sgi.com>
Add a new ili_fields member to the inode log item to isolate the in-memory
flags from the ones that actually go to the log. This will allow tracking
timestamp-only updates for fdatasync and O_DSYNC in the next patch and
prepares for divorcing the on-disk log format from the in-memory log item
a little further down the road.
Reviewed-by: Dave Chinner <dchinner@redhat.com>
Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Mark Tinguely <tinguely@sgi.com>
Signed-off-by: Ben Myers <bpm@sgi.com>
Timestamps on regular files are the last metadata that XFS does not update
transactionally. Now that we use the delaylog mode exclusively and made
the log scode scale extremly well there is no need to bypass that code for
timestamp updates. Logging all updates allows to drop a lot of code, and
will allow for further performance improvements later on.
Note that this patch drops optimized handling of fdatasync - it will be
added back in a separate commit.
Reviewed-by: Dave Chinner <dchinner@redhat.com>
Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Mark Tinguely <tinguely@sgi.com>
Signed-off-by: Ben Myers <bpm@sgi.com>
There is no reason to keep a reference to the inode even if we unlock
it during transaction commit because we never drop a reference between
the ijoin and commit. Also use this fact to merge xfs_trans_ijoin_ref
back into xfs_trans_ijoin - the third argument decides if an unlock
is needed now.
I'm actually starting to wonder if allowing inodes to be unlocked
at transaction commit really is worth the effort. The only real
benefit is that they can be unlocked earlier when commiting a
synchronous transactions, but that could be solved by doing the
log force manually after the unlock, too.
Signed-off-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Alex Elder <aelder@sgi.com>
Remove the transaction pointer in the inode. It's only used to avoid
passing down an argument in the bmap code, and for a few asserts in
the transaction code right now.
Also use the local variable ip in a few more places in xfs_inode_item_unlock,
so that it isn't only used for debug builds after the above change.
Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Alex Elder <aelder@sgi.com>
Reviewed-by: Dave Chinner <dchinner@redhat.com>
Currently we return iodes from xfs_ialloc with just a single reference held.
But we need two references, as one is dropped during transaction commit and
the second needs to be transfered to the VFS. Change xfs_ialloc to use
xfs_iget plus xfs_trans_ijoin_ref to grab two references to the inode,
and remove the now superflous IHOLD calls from all callers. This also
greatly simplifies the error handling in xfs_create and also allow to remove
xfs_trans_iget as no other callers are left.
Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Dave Chinner <dchinner@redhat.com>
Signed-off-by: Alex Elder <aelder@sgi.com>
Under heavy multi-way parallel create workloads, the VFS struggles
to write back all the inodes that have been changed in age order.
The bdi flusher thread becomes CPU bound, spending 85% of it's time
in the VFS code, mostly traversing the superblock dirty inode list
to separate dirty inodes old enough to flush.
We already keep an index of all metadata changes in age order - in
the AIL - and continued log pressure will do age ordered writeback
without any extra overhead at all. If there is no pressure on the
log, the xfssyncd will periodically write back metadata in ascending
disk address offset order so will be very efficient.
Hence we can stop marking VFS inodes dirty during transaction commit
or when changing timestamps during transactions. This will keep the
inodes in the superblock dirty list to those containing data or
unlogged metadata changes.
However, the timstamp changes are slightly more complex than this -
there are a couple of places that do unlogged updates of the
timestamps, and the VFS need to be informed of these. Hence add a
new function xfs_trans_ichgtime() for transactional changes,
and leave xfs_ichgtime() for the non-transactional changes.
Signed-off-by: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Alex Elder <aelder@sgi.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Currently we need to either call IHOLD or xfs_trans_ihold on an inode when
joining it to a transaction via xfs_trans_ijoin.
This patches instead makes xfs_trans_ijoin usable on it's own by doing
an implicity xfs_trans_ihold, which also allows us to drop the third
argument. For the case where we want to hold a reference on the inode
a xfs_trans_ijoin_ref wrapper is added which does the IHOLD and marks
the inode for needing an xfs_iput. In addition to the cleaner interface
to the caller this also simplifies the implementation.
Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Dave Chinner <dchinner@redhat.com>
Currently we track log item descriptor belonging to a transaction using a
complex opencoded chunk allocator. This code has been there since day one
and seems to work around the lack of an efficient slab allocator.
This patch replaces it with dynamically allocated log item descriptors
from a dedicated slab pool, linked to the transaction by a linked list.
This allows to greatly simplify the log item descriptor tracking to the
point where it's just a couple hundred lines in xfs_trans.c instead of
a separate file. The external API has also been simplified while we're
at it - the xfs_trans_add_item and xfs_trans_del_item functions to add/
delete items from a transaction have been simplified to the bare minium,
and the xfs_trans_find_item function is replaced with a direct dereference
of the li_desc field. All debug code walking the list of log items in
a transaction is down to a simple list_for_each_entry.
Note that we could easily use a singly linked list here instead of the
double linked list from list.h as the fastpath only does deletion from
sequential traversal. But given that we don't have one available as
a library function yet I use the list.h functions for simplicity.
Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Dave Chinner <dchinner@redhat.com>
Dmapi support was never merged upstream, but we still have a lot of hooks
bloating XFS for it, all over the fast pathes of the filesystem.
This patch drops over 700 lines of dmapi overhead. If we'll ever get HSM
support in mainline at least the namespace events can be done much saner
in the VFS instead of the individual filesystem, so it's not like this
is much help for future work.
Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Dave Chinner <dchinner@redhat.com>
The block number comes from bulkstat based inode lookups to shortcut
the mapping calculations. We ar enot able to trust anything from
bulkstat, so drop the block number as well so that the correct
lookups and mappings are always done.
Signed-off-by: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
xfs_trans_iget is a wrapper for xfs_iget that adds the inode to the
transaction after it is read. Except when the inode already is in the
inode cache, in which case it returns the existing locked inode with
increment lock recursion counts.
Now, no one in the tree every decrements these lock recursion counts,
so any user of this gets a potential double unlock when both the original
owner of the inode and the xfs_trans_iget caller unlock it. When looking
back in a git bisect in the historic XFS tree there was only one place
that decremented these counts, xfs_trans_iput. Introduced in commit
ca25df7a840f426eb566d52667b6950b92bb84b5 by Adam Sweeney in 1993,
and removed in commit 19f899a3ab155ff6a49c0c79b06f2f61059afaf3 by
Steve Lord in 2003. And as long as it didn't slip through git bisects
cracks never actually used in that time frame.
A quick audit of the callers of xfs_trans_iget shows that no caller
really relies on this behaviour fortunately - xfs_ialloc allows this
inode from disk so it must not be there before, and all the RT allocator
routines only every add each RT bitmap inode once.
In addition to removing lots of code and reducing the size of the inode
item this patch also avoids the double inode cache lookup in each
create/mkdir/mknod transaction.
Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Alex Elder <aelder@sgi.com>
Signed-off-by: Felix Blyakher <felixb@sgi.com>
Use xfs_trans_ijoin in xfs_trans_iget in case we need to join an inode into
a transaction instead of opencoding it. Based on a discussion with and an
incomplete patch from Niv Sardi.
Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Dave Chinner <david@fromorbit.com>
Signed-off-by: Niv Sardi <xaiki@sgi.com>
kmem_free() function takes (ptr, size) arguments but doesn't actually use
second one.
This patch removes size argument from all callsites.
SGI-PV: 981498
SGI-Modid: xfs-linux-melb:xfs-kern:31050a
Signed-off-by: Denys Vlasenko <vda.linux@googlemail.com>
Signed-off-by: David Chinner <dgc@sgi.com>
Signed-off-by: Lachlan McIlroy <lachlan@sgi.com>
The writer field is not needed for non_DEBU builds so remove it. While
we're at i also clean up the interface for is locked asserts to go through
and xfs_iget.c helper with an interface like the xfs_ilock routines to
isolated the XFS codebase from mrlock internals. That way we can kill
mrlock_t entirely once rw_semaphores grow an islocked facility. Also
remove unused flags to the ilock family of functions.
SGI-PV: 976035
SGI-Modid: xfs-linux-melb:xfs-kern:30902a
Signed-off-by: Christoph Hellwig <hch@infradead.org>
Signed-off-by: Lachlan McIlroy <lachlan@sgi.com>
Initial git repository build. I'm not bothering with the full history,
even though we have it. We can create a separate "historical" git
archive of that later if we want to, and in the meantime it's about
3.2GB when imported into git - space that would just make the early
git days unnecessarily complicated, when we don't have a lot of good
infrastructure for it.
Let it rip!
