btrfs: simple quotas
Get series with b4
b4 am [email protected]Versions newest first · linked by author and title
- v4 posting 2/26 patches 11 repliesIn review
- v3 posting 3/36 patches 3 repliesSuperseded
- v2 posting 3/36 patches 0 repliesSuperseded
Earlier versions · 8
| Version | Posted | Patches | Replies | Status |
|---|---|---|---|---|
| v3 posting 2/3 | 8 patches | 2 replies | Superseded | |
| v6 Latest | 18 patches | 0 replies | New | |
| v5 | 18 patches | 34 replies | Superseded | |
| v4 posting 1/2 | 18 patches | 0 replies | Superseded | |
| v3 posting 1/3 | 20 patches | 1 reply | Superseded | |
| v2 posting 2/3 | 8 patches | 0 replies | Superseded | |
| v2 posting 1/3 | 20 patches | 0 replies | Superseded | |
| v1 | 5 patches | 5 replies | Superseded |
Changes in v6 from the cover letter · 21 lines
* cleanup stray newlines
* new named mode helpers for "not disabled" and "full accounting"
* lots of boolean logic order cleanup
* fixes to reduce struct memory waste from padding
* improve name of new status flag
* don't assign strings in sysfs; just call sysfs_emit
* don't handle disabled mode in sysfs
* s/simple_quota_delta/squota_delta/g
* use Qu's awesome new ll iterator for recording squota delta
* made comments sentences
* added incompat bit assertions where we see the new OWNER_REF
* document OWNER_REF inline ref key, renumber it to 188
* auto inherit uses struct_size instead of unchecked kzalloc
* auto inherit uses more named variables for aliased pointer readability
* check auto inherit return value
* free auto inherit struct if it was allocated by us
* moved enable gen in fs_info with other qgroup fields
* renamed to qgroup_enable_gen to match other qgroup fields
* cleaned up simple variable while reading status item
* added helper for enable_gen read to harden the version change v5:
* fix btrfs/187 failure in squota mode: relocation+dedupe led to drop refs with the wrong owning root coming first, followed by drop refs with the real owning root. The "bad" ones are never last, so fix it by letting the "good" ones set it on the head ref.
Patches in v6 No applies recorded yet
| # | Subject and files | Size | Replies | Applied |
|---|---|---|---|---|
| 1/18 | btrfs: introduce quota modefs/ fs/ | +13 | 0 replies | |
| 2/18 | btrfs: add new quota mode for simple quotasfs/ fs/ fs/ +6 more | +114 | 0 replies | |
| 3/18 | btrfs: expose quota mode via sysfsfs/ | +28 | 0 replies | |
| 4/18 | btrfs: add simple_quota incompat feature to sysfsfs/ | +2 | 0 replies | |
| 5/18 | btrfs: flush reservations during quota disablefs/ | +44 | 0 replies | |
| 6/18 | btrfs: create qgroup earlier in snapshot creationfs/ fs/ | +9 | 0 replies | |
| 7/18 | btrfs: function for recording simple quota deltasfs/ fs/ | +60 | 0 replies | |
| 8/18 | btrfs: rename tree_ref and data_ref owning_rootfs/ fs/ fs/ +1 more | +18 | 0 replies | |
| 9/18 | btrfs: track owning root in btrfs_reffs/ fs/ fs/ +3 more | +34 | 0 replies | |
| 10/18 | btrfs: track original extent owner in head_reffs/ fs/ | +23 | 0 replies | |
| 11/18 | btrfs: new inline ref storing owning subvol of data extentsfs/ fs/ fs/ +4 more | +83 | 0 replies | |
| 12/18 | btrfs: inline owner ref lookup helperfs/ fs/ | +53 | 0 replies | |
| 13/18 | btrfs: record simple quota deltasfs/ fs/ fs/ | +82 | 0 replies | |
| 14/18 | btrfs: simple quota auto hierarchy for nested subvolsfs/ fs/ fs/ +1 more | +66 | 0 replies | |
| 15/18 | btrfs: check generation when recording simple quota deltafs/ fs/ fs/ +3 more | +40 | 0 replies | |
| 16/18 | btrfs: track metadata relocation cow with simple quotafs/ fs/ fs/ +2 more | +23 | 0 replies | |
| 17/18 | btrfs: track data relocation with simple quotafs/ fs/ fs/ | +57 | 0 replies | |
| 18/18 | btrfs: only set QUOTA_ENABLED when done reading qgroupsfs/ fs/ | +7 | 0 replies |
Reviews trailers per patch
5 of 18 patches have a Reviewed-by or Acked-by.
| Patch | Josef Bacik |
|---|---|
| 1/18 | R-b |
| 2/18 | |
| 3/18 | |
| 4/18 | |
| 5/18 | R-b |
| 6/18 | |
| 7/18 | |
| 8/18 | R-b |
| 9/18 | |
| 10/18 | |
| 11/18 | |
| 12/18 | R-b |
| 13/18 | |
| 14/18 | |
| 15/18 | |
| 16/18 | R-b |
| 17/18 | |
| 18/18 |
Trailers given in replies are filled; those carried in the patches from earlier versions are outlined with dashes.
Cover letter [PATCH v6 00/18] btrfs: simple quotas
btrfs quota groups (qgroups) are a compelling feature of btrfs that allow flexible control for limiting subvolume data and metadata usage. However, due to btrfs's high level decision to tradeoff snapshot performance against ref-counting performance, qgroups suffer from non-trivial performance issues that make them unattractive in certain workloads. Particularly, frequent backref walking during writes and during commits can make operations increasingly expensive as the number of snapshots scales up. For that reason, we have never been able to commit to using qgroups in production at Meta, despite significant interest from people running container workloads, where we would benefit from protecting the rest of the host from a buggy application in a container running away with disk usage. This patch series introduces a simplified version of qgroups called
simple quotas (squotas) which never computes global reference counts for extents, and thus has similar performance characteristics to normal, quotas disabled, btrfs. The "trick" is that in simple quotas mode, we account all extents permanently to the subvolume in which they were originally created. That allows us to make all accounting 1:1 with extent item lifetime, removing the need to walk backrefs. However, this sacrifices the ability to compute shared vs. exclusive usage. It also results in counter-intuitive, though still predictable and simple accounting in the cases where an original extent is removed while a shared copy still exists. Qgroups is able to detect that case and count the remaining copy as an exclusive owner, while squotas is not. As a result, squotas works best when the original extent is immutable and outlives any clones.
==Format Change==
In order to track the original creating subvolume of a data extent in the face of reflinks, it is necessary to add additional accounting to the extent item. To save space, this is done with a new inline ref item. However, the downside of this approach is that it makes enabling squota an incompat change, denoted by the new incompat bit SIMPLE_QUOTA. When this bit is set and quotas are enabled, new extent items get the extra accounting, and freed extent items check for the accounting to find their creating subvolume. In addition, 1:1 with this incompat bit, the quota status item now tracks a "quota enablement generation" needed for properly handling deleting extents with predate enablement.
==API==
Squotas reuses the api of qgroups. The only difference is that when you enable quotas via btrfs quota enable, you pass the --simple flag. Squotas will always report exclusive == shared for each qgroup. Squotas deal with extent_item/metadata_item sizes and thus do not do anything special with compression. Squotas also introduce auto inheritance for nested subvols. The API is documented more fully in the documentation patches in btrfs-progs.
==Testing methodology==
Using updated btrfs-progs and fstests (relevant matching patch sets to be sent ASAP)
btrfs-progs: https://github.com/boryas/btrfs-progs/tree/squota-progs
fstests: https://github.com/boryas/fstests/tree/squota-test
I ran '-g auto' on fstests on the following configurations:
1a) baseline kernel/progs/fstests.
1b) squota kernel baseline progs/fstests.
2a) baseline kernel/progs/fstests. fstests configured to mkfs with quota 2b) squota kernel/progs/fstests. fstests configured to mkfs with squota
I compared 1a against 1b and 2a against 2b and detected no regressions. 2a/2b both exhibit regressions against 1a/1b which are largely issues with quota reservations in various complicated cases. I intend to run those down in the future, but they are not simple quota specific, as they are already broken with plain qgroups.
==Performance Testing==
I measured the performance of the change using fsperf. I ran with 3 configurations using the squota kernel:
- plain mkfs
- qgroup mkfs
- squota mkfs
And added a new performance test which creates 1000 files in a subvol, creates 100 snapshots of that subvol, then unshares extents in files in the snapshots. I measured write performance with fio and btrfs commit critical section performance side effects with bpftrace on 'wait_current_trans'.
The results for the test which measures unshare perf (unshare.py) with qgroup and squota compared to the baseline:
group test results
unshare results
metric baseline current stdev diff
========================================================================================avg_commit_ms 162.13 285.75 3.14 76.24%bg_count 16 16 0 0.00%
commits 378.20 379 1.92 0.21% elapsed 201.40 270.40 1.34 34.26%
end_state_mount_ns 26036211.60 26004593.60 2281065.40 -0.12%end_state_umount_ns 2.45e+09 2.55e+09 20740154.41 3.93%max_commit_ms 425.80 594 53.34 39.50%sys_cpu 0.10 0.06 0.06 -42.15%wait_current_trans_calls 2945.60 3405.20 47.08 15.60%wait_current_trans_ns_max 1.56e+08 3.43e+08 32659393.25 120.07%wait_current_trans_ns_mean 1974875.35 28588482.55 1557588.84 1347.61%
wait_current_trans_ns_min 232 232 25.88 0.00% wait_current_trans_ns_p50 718 740 22.80 3.06% wait_current_trans_ns_p95 7711770.20 2.21e+08 17241032.09 2761.19% wait_current_trans_ns_p99 67744932.29 2.68e+08 41275815.87 295.16% write_bw_bytes 653008.80 486344.40 4209.91 -25.52% write_clat_ns_mean 6251404.78 8406837.89 39779.15 34.48% write_clat_ns_p50 1656422.40 1643315.20 27415.68 -0.79% write_clat_ns_p99 1.90e+08 3.20e+08 2097152 68.62% write_io_kbytes 128000 128000 0 0.00% write_iops 159.43 118.74 1.03 -25.52% write_lat_ns_max 7.06e+08 9.80e+08 47324816.61 38.88% write_lat_ns_mean 6251503.06 8406936.06 39780.83 34.48% write_lat_ns_min 3354 4648 616.06 38.58%
squota test results
unshare results
metric baseline current stdev diff
========================================================================================avg_commit_ms 162.13 164.16 3.14 1.25%bg_count 16 0 0 -100.00%
commits 378.20 380.80 1.92 0.69% elapsed 201.40 208.20 1.34 3.38% end_state_mount_ns 26036211.60 25840729.60 2281065.40 -0.75% end_state_umount_ns 2.45e+09 3.01e+09 20740154.41 22.80% max_commit_ms 425.80 415.80 53.34 -2.35%
sys_cpu 0.10 0.08 0.06 -23.36%wait_current_trans_calls 2945.60 2981.60 47.08 1.22%wait_current_trans_ns_max 1.56e+08 1.12e+08 32659393.25 -27.86%wait_current_trans_ns_mean 1974875.35 1064734.76 1557588.84 -46.09%
wait_current_trans_ns_min 232 238 25.88 2.59% wait_current_trans_ns_p50 718 746 22.80 3.90% wait_current_trans_ns_p95 7711770.20 1567.60 17241032.09 -99.98% wait_current_trans_ns_p99 67744932.29 49880514.27 41275815.87 -26.37% write_bw_bytes 653008.80 631256 4209.91 -3.33% write_clat_ns_mean 6251404.78 6476816.06 39779.15 3.61% write_clat_ns_p50 1656422.40 1581056 27415.68 -4.55% write_clat_ns_p99 1.90e+08 1.94e+08 2097152 2.21% write_io_kbytes 128000 128000 0 0.00% write_iops 159.43 154.12 1.03 -3.33%
write_lat_ns_max 7.06e+08 7.65e+08 47324816.61 8.38%write_lat_ns_mean 6251503.06 6476912.76 39780.83 3.61%write_lat_ns_min 3354 4062 616.06 21.11%
And the same, but only showing results where the deviation was outside of a 95% confidence interval for the mean (default significance highlighting in fsperf):
qgroup test results
unshare results
metric baseline current stdev diff
========================================================================================
avg_commit_ms 162.13 285.75 3.14 76.24% elapsed 201.40 270.40 1.34 34.26% end_state_umount_ns 2.45e+09 2.55e+09 20740154.41 3.93% max_commit_ms 425.80 594 53.34 39.50% wait_current_trans_calls 2945.60 3405.20 47.08 15.60% wait_current_trans_ns_max 1.56e+08 3.43e+08 32659393.25 120.07% wait_current_trans_ns_mean 1974875.35 28588482.55 1557588.84 1347.61% wait_current_trans_ns_p95 7711770.20 2.21e+08 17241032.09 2761.19%
wait_current_trans_ns_p99 67744932.29 2.68e+08 41275815.87 295.16%write_bw_bytes 653008.80 486344.40 4209.91 -25.52%write_clat_ns_mean 6251404.78 8406837.89 39779.15 34.48%write_clat_ns_p99 1.90e+08 3.20e+08 2097152 68.62%write_iops 159.43 118.74 1.03 -25.52%
write_lat_ns_max 7.06e+08 9.80e+08 47324816.61 38.88% write_lat_ns_mean 6251503.06 8406936.06 39780.83 34.48% write_lat_ns_min 3354 4648 616.06 38.58%
squota test results
unshare results
metric baseline current stdev diff
========================================================================================
elapsed 201.40 208.20 1.34 3.38%end_state_umount_ns 2.45e+09 3.01e+09 20740154.41 22.80%write_bw_bytes 653008.80 631256 4209.91 -3.33%
write_clat_ns_mean 6251404.78 6476816.06 39779.15 3.61% write_clat_ns_p50 1656422.40 1581056 27415.68 -4.55%
write_clat_ns_p99 1.90e+08 1.94e+08 2097152 2.21%write_iops 159.43 154.12 1.03 -3.33%write_lat_ns_mean 6251503.06 6476912.76 39780.83 3.61%
Particularly noteworthy are the massive regressions to wait_current_trans in qgroup mode as well as the solid regressions to bandwidth, iops and write latency. The regressions/improvements in squotas are modest in comparison in line with the expectation. I am still investigating the squota umount regression, particularly whether it is in the umount's final commit and represents a real performance problem with squotas.
---
Changelog:
v6:
* cleanup stray newlines
* new named mode helpers for "not disabled" and "full accounting"
* lots of boolean logic order cleanup
* fixes to reduce struct memory waste from padding
* improve name of new status flag
* don't assign strings in sysfs; just call sysfs_emit
* don't handle disabled mode in sysfs
* s/simple_quota_delta/squota_delta/g
* use Qu's awesome new ll iterator for recording squota delta
* made comments sentences
* added incompat bit assertions where we see the new OWNER_REF
* document OWNER_REF inline ref key, renumber it to 188
* auto inherit uses struct_size instead of unchecked kzalloc
* auto inherit uses more named variables for aliased pointer readability
* check auto inherit return value
* free auto inherit struct if it was allocated by us
* moved enable gen in fs_info with other qgroup fields
* renamed to qgroup_enable_gen to match other qgroup fields
* cleaned up simple variable while reading status item
* added helper for enable_gen read to harden the version change v5:
* fix btrfs/187 failure in squota mode: relocation+dedupe led to drop refs with the wrong owning root coming first, followed by drop refs with the real owning root. The "bad" ones are never last, so fix it by letting the "good" ones set it on the head ref.
v4:
* drop unrelated patches folded into misc-next
* fix crash where check_committed_extent was reading the inline ref type on an extent item with no inline extents. (btrfs/192 *without* squotas enabled)
v3:
* u64 -> __le64 in new owner_ref_item (as caught by kernel test bot) v2:
* fix dumb formatting errors, unexpected/unrelated edits
* use command instead of status in ioctl
* fix the illegal GFP_KERNEL in delta fn (punted on pulling allocations out from the spin lock and using GFP_ATOMIC like other qgroups use cases for now. Plan to fix that in either v3 or a follow up series, as there are other places this is an issue for qgroups/squotas)
* improve boolean logic in head_ref init
* use list_count helper function instead of rolling my own bad one
* fixed the adjacent extents reloc cluster bug Josef noticed
* fixed a qgroups bug I introduced: it needs to be able to account extents while shutting down to not hit a warning in commit_transaction
* added a qgroup_status flag for simple quotas to not rely on the incompat bit directly. This allows disabling simple quotas and enabling qgroups.
Boris Burkov (18):
btrfs: introduce quota mode
btrfs: add new quota mode for simple quotas
btrfs: expose quota mode via sysfs
btrfs: add simple_quota incompat feature to sysfs
btrfs: flush reservations during quota disable
btrfs: create qgroup earlier in snapshot creation
btrfs: function for recording simple quota deltas
btrfs: rename tree_ref and data_ref owning_root
btrfs: track owning root in btrfs_ref
btrfs: track original extent owner in head_ref
btrfs: new inline ref storing owning subvol of data extents
btrfs: inline owner ref lookup helper
btrfs: record simple quota deltas
btrfs: simple quota auto hierarchy for nested subvols
btrfs: check generation when recording simple quota delta
btrfs: track metadata relocation cow with simple quota
btrfs: track data relocation with simple quota
btrfs: only set QUOTA_ENABLED when done reading qgroups
fs/btrfs/accessors.h | 6 + fs/btrfs/backref.c | 3 + fs/btrfs/ctree.c | 22 ++- fs/btrfs/ctree.h | 1 + fs/btrfs/delayed-ref.c | 37 ++-- fs/btrfs/delayed-ref.h | 32 +++- fs/btrfs/disk-io.c | 5 +- fs/btrfs/extent-tree.c | 240 ++++++++++++++++++++++---- fs/btrfs/extent-tree.h | 4 + fs/btrfs/file.c | 10 +- fs/btrfs/fs.h | 4 +- fs/btrfs/inode-item.c | 2 +- fs/btrfs/ioctl.c | 7 +- fs/btrfs/print-tree.c | 12 ++ fs/btrfs/qgroup.c | 297 +++++++++++++++++++++++++++----- fs/btrfs/qgroup.h | 46 ++++- fs/btrfs/ref-verify.c | 7 +- fs/btrfs/relocation.c | 66 ++++++- fs/btrfs/root-tree.c | 2 +- fs/btrfs/sysfs.c | 30 ++++ fs/btrfs/transaction.c | 18 +- fs/btrfs/tree-checker.c | 3 + fs/btrfs/tree-log.c | 3 +- include/uapi/linux/btrfs.h | 2 + include/uapi/linux/btrfs_tree.h | 31 +++- 25 files changed, 745 insertions(+), 145 deletions(-)
Signature · 1 line
2.41.0