[BUG]
Btrfs/139 will fail with a high probability if the testing machine (VM)
has only 2G RAM.
Resulting the final write success while it should fail due to EDQUOT,
and the fs will have quota exceeding the limit by 16K.
The simplified reproducer will be: (needs a 2G ram VM)
$ mkfs.btrfs -f $dev
$ mount $dev $mnt
$ btrfs subv create $mnt/subv
$ btrfs quota enable $mnt
$ btrfs quota rescan -w $mnt
$ btrfs qgroup limit -e 1G $mnt/subv
$ for i in $(seq -w 1 8); do
xfs_io -f -c "pwrite 0 128M" $mnt/subv/file_$i > /dev/null
echo "file $i written" > /dev/kmsg
done
$ sync
$ btrfs qgroup show -pcre --raw $mnt
The last pwrite will not trigger EDQUOT and final 'qgroup show' will
show something like:
qgroupid rfer excl max_rfer max_excl parent child
-------- ---- ---- -------- -------- ------ -----
0/5 16384 16384 none none --- ---
0/256 1073758208 1073758208 none 1073741824 --- ---
And 1073758208 is larger than
> 1073741824.
[CAUSE]
It's a bug in btrfs qgroup data reserved space management.
For quota limit, we must ensure that:
reserved (data + metadata) + rfer/excl <= limit
Since rfer/excl is only updated at transaction commmit time, reserved
space needs to be taken special care.
One important part of reserved space is data, and for a new data extent
written to disk, we still need to take the reserved space until
rfer/excl numbers get updated.
Originally when an ordered extent finishes, we migrate the reserved
qgroup data space from extent_io tree to delayed ref head of the data
extent, expecting delayed ref will only be cleaned up at commit
transaction time.
However for small RAM machine, due to memory pressure dirty pages can be
flushed back to disk without committing a transaction.
The related events will be something like:
file 1 written
btrfs_finish_ordered_io: ino=258 ordered offset=0 len=54947840
btrfs_finish_ordered_io: ino=258 ordered offset=54947840 len=5636096
btrfs_finish_ordered_io: ino=258 ordered offset=61153280 len=57344
btrfs_finish_ordered_io: ino=258 ordered offset=61210624 len=8192
btrfs_finish_ordered_io: ino=258 ordered offset=60583936 len=569344
cleanup_ref_head: num_bytes=54947840
cleanup_ref_head: num_bytes=5636096
cleanup_ref_head: num_bytes=569344
cleanup_ref_head: num_bytes=57344
cleanup_ref_head: num_bytes=8192
^^^^^^^^^^^^^^^^ This will free qgroup data reserved space
file 2 written
...
file 8 written
cleanup_ref_head: num_bytes=8192
...
btrfs_commit_transaction <<< the only transaction committed during
the test
When file 2 is written, we have already freed 128M reserved qgroup data
space for ino 258. Thus later write won't trigger EDQUOT.
This allows us to write more data beyond qgroup limit.
In my 2G ram VM, it could reach about 1.2G before hitting EDQUOT.
[FIX]
By moving reserved qgroup data space from btrfs_delayed_ref_head to
btrfs_qgroup_extent_record, we can ensure that reserved qgroup data
space won't be freed half way before commit transaction, thus fix the
problem.
Fixes: f64d5ca86821 ("btrfs: delayed_ref: Add new function to record reserved space into delayed ref")
Signed-off-by: Qu Wenruo <wqu@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
Since it's replaced by new delayed subtree swap code, remove the original code. The cleanup is small since most of its core function is still used by delayed subtree swap trace. Signed-off-by: Qu Wenruo <wqu@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
Before this patch, qgroup code traces the whole subtree of subvolume and
reloc trees unconditionally.
This makes qgroup numbers consistent, but it could cause tons of
unnecessary extent tracing, which causes a lot of overhead.
However for subtree swap of balance, just swap both subtrees because
they contain the same contents and tree structure, so qgroup numbers
won't change.
It's the race window between subtree swap and transaction commit could
cause qgroup number change.
This patch will delay the qgroup subtree scan until COW happens for the
subtree root.
So if there is no other operations for the fs, balance won't cause extra
qgroup overhead. (best case scenario)
Depending on the workload, most of the subtree scan can still be
avoided.
Only for worst case scenario, it will fall back to old subtree swap
overhead. (scan all swapped subtrees)
[[Benchmark]]
Hardware:
VM 4G vRAM, 8 vCPUs,
disk is using 'unsafe' cache mode,
backing device is SAMSUNG 850 evo SSD.
Host has 16G ram.
Mkfs parameter:
--nodesize 4K (To bump up tree size)
Initial subvolume contents:
4G data copied from /usr and /lib.
(With enough regular small files)
Snapshots:
16 snapshots of the original subvolume.
each snapshot has 3 random files modified.
balance parameter:
-m
So the content should be pretty similar to a real world root fs layout.
And after file system population, there is no other activity, so it
should be the best case scenario.
| v4.20-rc1 | w/ patchset | diff
-----------------------------------------------------------------------
relocated extents | 22615 | 22457 | -0.1%
qgroup dirty extents | 163457 | 121606 | -25.6%
time (sys) | 22.884s | 18.842s | -17.6%
time (real) | 27.724s | 22.884s | -17.5%
Signed-off-by: Qu Wenruo <wqu@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
To allow delayed subtree swap rescan, btrfs needs to record per-root
information about which tree blocks get swapped. This patch introduces
the required infrastructure.
The designed workflow will be:
1) Record the subtree root block that gets swapped.
During subtree swap:
O = Old tree blocks
N = New tree blocks
reloc tree subvolume tree X
Root Root
/ \ / \
NA OB OA OB
/ | | \ / | | \
NC ND OE OF OC OD OE OF
In this case, NA and OA are going to be swapped, record (NA, OA) into
subvolume tree X.
2) After subtree swap.
reloc tree subvolume tree X
Root Root
/ \ / \
OA OB NA OB
/ | | \ / | | \
OC OD OE OF NC ND OE OF
3a) COW happens for OB
If we are going to COW tree block OB, we check OB's bytenr against
tree X's swapped_blocks structure.
If it doesn't fit any, nothing will happen.
3b) COW happens for NA
Check NA's bytenr against tree X's swapped_blocks, and get a hit.
Then we do subtree scan on both subtrees OA and NA.
Resulting 6 tree blocks to be scanned (OA, OC, OD, NA, NC, ND).
Then no matter what we do to subvolume tree X, qgroup numbers will
still be correct.
Then NA's record gets removed from X's swapped_blocks.
4) Transaction commit
Any record in X's swapped_blocks gets removed, since there is no
modification to swapped subtrees, no need to trigger heavy qgroup
subtree rescan for them.
This will introduce 128 bytes overhead for each btrfs_root even qgroup
is not enabled. This is to reduce memory allocations and potential
failures.
Signed-off-by: Qu Wenruo <wqu@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
The typos accumulate over time so once in a while time they get fixed in a large patch. Signed-off-by: Andrea Gelmini <andrea.gelmini@gelma.net> Reviewed-by: David Sterba <dsterba@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
The first auto-assigned value to enum is 0, we can use that and not initialize all members where the auto-increment does the same. This is used for values that are not part of on-disk format. Reviewed-by: Omar Sandoval <osandov@fb.com> Reviewed-by: Qu Wenruo <wqu@suse.com> Reviewed-by: Johannes Thumshirn <jthumshirn@suse.de> Signed-off-by: David Sterba <dsterba@suse.com>
Some qgroup trace events like btrfs_qgroup_release_data() and btrfs_qgroup_free_delayed_ref() can still be triggered even if qgroup is not enabled. This is caused by the lack of qgroup status check before calling some qgroup functions. Thankfully the functions can handle quota disabled case well and just do nothing for qgroup disabled case. This patch will do earlier check before triggering related trace events. And for enabled <-> disabled race case: 1) For enabled->disabled case Disable will wipe out all qgroups data including reservation and excl/rfer. Even if we leak some reservation or numbers, it will still be cleared, so nothing will go wrong. 2) For disabled -> enabled case Current btrfs_qgroup_release_data() will use extent_io tree to ensure we won't underflow reservation. And for delayed_ref we use head->qgroup_reserved to record the reserved space, so in that case head->qgroup_reserved should be 0 and we won't underflow. CC: stable@vger.kernel.org # 4.14+ Reported-by: Chris Murphy <lists@colorremedies.com> Link: https://lore.kernel.org/linux-btrfs/CAJCQCtQau7DtuUUeycCkZ36qjbKuxNzsgqJ7+sJ6W0dK_NLE3w@mail.gmail.com/ Signed-off-by: Qu Wenruo <wqu@suse.com> Reviewed-by: David Sterba <dsterba@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
For qgroup_trace_extent_swap(), if we find one leaf that needs to be
traced, we will also iterate all file extents and trace them.
This is OK if we're relocating data block groups, but if we're
relocating metadata block groups, balance code itself has ensured that
both subtree of file tree and reloc tree contain the same contents.
That's to say, if we're relocating metadata block groups, all file
extents in reloc and file tree should match, thus no need to trace them.
This should reduce the total number of dirty extents processed in metadata
block group balance.
[[Benchmark]] (with all previous enhancement)
Hardware:
VM 4G vRAM, 8 vCPUs,
disk is using 'unsafe' cache mode,
backing device is SAMSUNG 850 evo SSD.
Host has 16G ram.
Mkfs parameter:
--nodesize 4K (To bump up tree size)
Initial subvolume contents:
4G data copied from /usr and /lib.
(With enough regular small files)
Snapshots:
16 snapshots of the original subvolume.
each snapshot has 3 random files modified.
balance parameter:
-m
So the content should be pretty similar to a real world root fs layout.
| v4.19-rc1 | w/ patchset | diff (*)
---------------------------------------------------------------
relocated extents | 22929 | 22851 | -0.3%
qgroup dirty extents | 227757 | 140886 | -38.1%
time (sys) | 65.253s | 37.464s | -42.6%
time (real) | 74.032s | 44.722s | -39.6%
Signed-off-by: Qu Wenruo <wqu@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
Before this patch, with quota enabled during balance, we need to mark
the whole subtree dirty for quota.
E.g.
OO = Old tree blocks (from file tree)
NN = New tree blocks (from reloc tree)
File tree (src) Reloc tree (dst)
OO (a) NN (a)
/ \ / \
(b) OO OO (c) (b) NN NN (c)
/ \ / \ / \ / \
OO OO OO OO (d) OO OO OO NN (d)
For old balance + quota case, quota will mark the whole src and dst tree
dirty, including all the 3 old tree blocks in reloc tree.
It's doable for small file tree or new tree blocks are all located at
lower level.
But for large file tree or new tree blocks are all located at higher
level, this will lead to mark the whole tree dirty, and be unbelievably
slow.
This patch will change how we handle such balance with quota enabled
case.
Now we will search from (b) and (c) for any new tree blocks whose
generation is equal to @last_snapshot, and only mark them dirty.
In above case, we only need to trace tree blocks NN(b), NN(c) and NN(d).
(NN(a) will be traced when COW happens for nodeptr modification). And
also for tree blocks OO(b), OO(c), OO(d). (OO(a) will be traced when COW
happens for nodeptr modification.)
For above case, we could skip 3 tree blocks, but for larger tree, we can
skip tons of unmodified tree blocks, and hugely speed up balance.
This patch will introduce a new function,
btrfs_qgroup_trace_subtree_swap(), which will do the following main
work:
1) Read out real root eb
And setup basic dst_path for later calls
2) Call qgroup_trace_new_subtree_blocks()
To trace all new tree blocks in reloc tree and their counter
parts in the file tree.
Signed-off-by: Qu Wenruo <wqu@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
It can be fetched from the transaction handle. Signed-off-by: Lu Fengqi <lufq.fnst@cn.fujitsu.com> Signed-off-by: David Sterba <dsterba@suse.com>