Time to try again to get this code merged.
This series aims to improve the scalability of XFS transaction
commits on large CPU count machines. My 32p machine hits contention
limits in xlog_cil_commit() at about 700,000 transaction commits a
section. It hits this at 16 thread workloads, and 32 thread
workloads go no faster and just burn CPU on the CIL spinlocks.
This patchset gets rid of spinlocks and global serialisation points
in the xlog_cil_commit() path. It does this by moving to a
combination of per-cpu counters, unordered per-cpu lists and
post-ordered per-cpu lists.
This results in transaction commit rates exceeding 1.6 million
commits/s under unlink certain workloads, and while the log lock
contention is largely gone there is still significant lock
contention at the VFS at 600,000 transactions/s:
19.39% [kernel] [k] __pv_queued_spin_lock_slowpath
6.40% [kernel] [k] do_raw_spin_lock
4.07% [kernel] [k] __raw_callee_save___pv_queued_spin_unlock
3.08% [kernel] [k] memcpy_erms
1.93% [kernel] [k] xfs_buf_find
1.69% [kernel] [k] xlog_cil_commit
1.50% [kernel] [k] syscall_exit_to_user_mode
1.18% [kernel] [k] memset_erms
- 64.23% 0.22% [kernel] [k] path_openat
- 64.01% path_openat
- 48.69% xfs_vn_create
- 48.60% xfs_generic_create
- 40.96% xfs_create
- 20.39% xfs_dir_ialloc
- 7.05% xfs_setup_inode
>>>>> - 6.87% inode_sb_list_add
- 6.54% _raw_spin_lock
- 6.53% do_raw_spin_lock
6.08% __pv_queued_spin_lock_slowpath
.....
- 11.27% xfs_trans_commit
- 11.23% __xfs_trans_commit
- 10.85% xlog_cil_commit
2.47% memcpy_erms
- 1.77% xfs_buf_item_committing
- 1.70% xfs_buf_item_release
- 0.79% xfs_buf_unlock
0.68% up
0.61% xfs_buf_rele
0.80% xfs_buf_item_format
0.73% xfs_inode_item_format
0.68% xfs_buf_item_size
- 0.55% kmem_alloc_large
- 0.55% kmem_alloc
0.52% __kmalloc
.....
- 7.08% d_instantiate
- 6.66% security_d_instantiate
>>>>>> - 6.63% selinux_d_instantiate
- 6.48% inode_doinit_with_dentry
- 6.11% _raw_spin_lock
- 6.09% do_raw_spin_lock
5.60% __pv_queued_spin_lock_slowpath
....
- 1.77% terminate_walk
>>>>>> - 1.69% dput
- 1.55% _raw_spin_lock
- do_raw_spin_lock
1.19% __pv_queued_spin_lock_slowpath
But when we extend out to 1.5M commits/s we see that the contention
starts to shift to the atomics in the lockless log reservation path:
14.81% [kernel] [k] __pv_queued_spin_lock_slowpath
7.88% [kernel] [k] xlog_grant_add_space
7.18% [kernel] [k] xfs_log_ticket_ungrant
4.82% [kernel] [k] do_raw_spin_lock
3.58% [kernel] [k] xlog_space_left
3.51% [kernel] [k] xlog_cil_commit
There's still substantial spin lock contention occurring at the VFS,
too, but it's indicating that multiple atomic variable updates per
transaction reservation/commit pair is starting to reach scalability
limits here.
This is largely a re-implementation of a past RFC patchsets. While
that were good enough proof of concept to perf test, they did not
preserve transaction order correctly and failed shutdown tests all
the time. The changes to the CIL accounting and behaviour, combined
with the structural changes to xlog_write() in prior patchsets make
the per-cpu restructuring possible and sane.
Instead of trying to account for continuation log opheaders on a
"growth" basis, we pre-calculate how many iclogs we'll need to write
out a maximally sized CIL checkpoint and just reserve that space one
per commit until the CIL has a full reservation. If we ever run a
commit when we are already at the hard limit (because
post-throttling) we simply take an extra reservation from each
commit that is run when over the limit. Hence we don't need to do
space usage math in the fast path and so never need to sum the
per-cpu counters in this path.
Similarly, per-cpu lists have the problem of ordering - we can't
remove an item from a per-cpu list if we want to move it forward in
the CIL. We solve this problem by using an atomic counter to give
every commit a sequence number that is copied into the log items in
that transaction. Hence relogging items just overwrites the sequence
number in the log item, and does not move it in the per-cpu lists.
Once we reaggregate the per-cpu lists back into a single list in the
CIL push work, we can run it through list-sort() and reorder it back
into a globally ordered list. This costs a bit of CPU time, but now
that the CIL can run multiple works and pipelines properly, this is
not a limiting factor for performance. It does increase fsync
latency when the CIL is full, but workloads issuing large numbers of
fsync()s or sync transactions end up with very small CILs and so the
latency impact or sorting is not measurable for such workloads.
git://git.kernel.org/pub/scm/linux/kernel/git/dgc/linux-xfs.git xfs-cil-scale-3
Version 6:
- split out from aggregated patchset
- rebase on linux-xfs/for-next + dgc/xlog-write-rework
Version 5:
- https://lore.kernel.org/linux-xfs/20210603052240.171998-1-david@xxxxxxxxxxxxx/
