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>