On Wed, Feb 24, 2021 at 01:29:29PM -0800, Darrick J. Wong wrote:

On Tue, Feb 23, 2021 at 03:46:34PM +1100, Dave Chinner wrote:

quoted

From: Dave Chinner <redacted>

When we modify btrees repeatedly, we regularly increase the size of
the logged region by a single chunk at a time (per transaction
commit). This results in the CIL formatting code having to
reallocate the log vector buffer every time the buffer dirty region
grows. Hence over a typical 4kB btree buffer, we might grow the log
vector 4096/128 = 32x over a short period where we repeatedly add
or remove records to/from the buffer over a series of running
transaction. This means we are doing 32 memory allocations and frees
over this time during a performance critical path in the journal.

The amount of space tracked in the CIL for the object is calculated
during the ->iop_format() call for the buffer log item, but the
buffer memory allocated for it is calculated by the ->iop_size()
call. The size callout determines the size of the buffer, the format
call determines the space used in the buffer.

Hence we can oversize the buffer space required in the size
calculation without impacting the amount of space used and accounted
to the CIL for the changes being logged. This allows us to reduce
the number of allocations by rounding up the buffer size to allow
for future growth. This can safe a substantial amount of CPU time in
this path:

-   46.52%     2.02%  [kernel]                  [k] xfs_log_commit_cil
   - 44.49% xfs_log_commit_cil
      - 30.78% _raw_spin_lock
         - 30.75% do_raw_spin_lock
              30.27% __pv_queued_spin_lock_slowpath

(oh, ouch!)
....
      - 1.05% kmem_alloc_large
         - 1.02% kmem_alloc
              0.94% __kmalloc

This overhead here us what this patch is aimed at. After:

      - 0.76% kmem_alloc_large
         - 0.75% kmem_alloc
              0.70% __kmalloc

Signed-off-by: Dave Chinner <redacted>
Reviewed-by: Chandan Babu R <redacted>

Any particular reason for 512?  It looks like you simply picked an
arbitrary power of 2, but was there a particular target in mind? i.e.
we never need to realloc for the usual 4k filesystem?

It is based on the bitmap chunk size being 128 bytes and that random
directory entry updates almost never require more than 3-4 128 byte
regions to be logged in the directory block.

The other observation is for per-ag btrees. When we are inserting
into a new btree block, we'll pack it from the front. Hence the
first few records land in the first 128 bytes so we log only 128
bytes, the next 8-16 records land in the second region so now we log
256 bytes. And so on.  If we are doing random updates, it will only
allocate every 4 random 128 byte regions that are dirtied instead of
every single one.

Any larger than this and I noticed an increase in memory footprint
in my scalability workloads. Any less than this and I didn't really
see any significant benefit to CPU usage.

Cheers,

Dave.
-- 
Dave Chinner
david@fromorbit.com