On 3/17/20 4:10 AM, Qu Wenruo wrote:
This patchset is based on an OLD misc-next branch, please inform me
before trying to merge, so I can rebase it to latest misc-next.
(There will be tons of conflicts)
The branch can be fetched from github for review/testing.
https://github.com/adam900710/linux/tree/backref_cache_all
The patchset survives all the existing qgroup tests.
=== BACKGROUND ===
One of the biggest problem for qgroup is its performance impact.
Although we have improved it in since v5.0 kernel, there is still
something slowing down qgroup, the backref walk.
Before this patchset, we use btrfs_find_all_roots() to iterate all roots
referring to one extent.
That function is doing a pretty good job, but it doesn't has any cache,
which means even we're looking up the same extent, we still need to do
the full backref walk.
On the other hand, relocation is doing its own backref cache, and
provides a much faster backref walk.
So the patchset is mostly trying to make qgroup backref walk (at least
commit root backref walk) to use the same mechanism provided by
relocation.
=== BENCHMARK ===
For the performance improvement, the last patch has a benchmark.
The following content is completely copied from that patch:
------
Here is a small script to test it:
mkfs.btrfs -f $dev
mount $dev -o space_cache=v2 $mnt
btrfs subvolume create $mnt/src
for ((i = 0; i < 64; i++)); do
for (( j = 0; j < 16; j++)); do
xfs_io -f -c "pwrite 0 2k" $mnt/src/file_inline_$(($i * 16 + $j)) > /dev/null
done
xfs_io -f -c "pwrite 0 1M" $mnt/src/file_reg_$i > /dev/null
sync
btrfs subvol snapshot $mnt/src $mnt/snapshot_$i
done
sync
btrfs quota enable $mnt
btrfs quota rescan -w $mnt
Here is the benchmark for above small tests.
The performance material is the total execution time of get_old_roots()
for patched kernel (*), and find_all_roots() for original kernel.
*: With CONFIG_BTRFS_FS_CHECK_INTEGRITY disabled, as get_old_roots()
will call find_all_roots() to verify the result if that config is
enabled.
| Number of calls | Total exec time |
------------------------------------------------------
find_all_roots()| 732 | 529991034ns
get_old_roots() | 732 | 127998312ns
------------------------------------------------------
diff | 0.00 % | -75.8 %
------
Patch 01~30 are mostly refactors and code movement, which exposes no
behavior change.
Patch 31~32 are small behavior change only for qgroup backref cache.
Patch 33~39 are the implementation of qgroup backref cache.
=== REASON FOR RFC ===
The naming is currently my biggest concern.
Since the code movement involves exporting quite a lot of functions, in
theory they should have btrfs_ prefix.
(For all newly exported functions in backref.h)
But some functions like alloc_backref_node(), adding "btrfs_" prefix
doesn't make it more clear, but just making it unnecessary long.
My current plan is to rename them using "btrfs_brc_" prefix (BackRef
Cache), and remove the "cache" in the original name.
E.g:
alloc_backref_node => btrfs_brc_alloc_node()
backref_cache_release => btrfs_brc_release()
link_backref_edge => btrfs_brc_link_edge()
But the abbr "brc" is pretty confusing and makes no sense by itself, so
I'm not sure what's the best practice here.
btrfs_backref_node_alloc()
btrfs_backref_node_free()
btrfs_backref_edge_link()
etc, rename the structures to btrfs_backref_* and then make the function names
follow suit. Thanks,
Josef
