Inside of FB, as well as some user reports, we've had a consistent
problem of occasional ENOSPC transaction aborts. Inside FB we were
seeing ~100-200 ENOSPC aborts per day in the fleet, which is a really
low occurrence rate given the size of our fleet, but it's not nothing.
There are two causes of this particular problem.
First is delayed allocation. The reservation system for delalloc
assumes that contiguous dirty ranges will result in 1 file extent item.
However if there is memory pressure that results in fragmented writeout,
or there is fragmentation in the block groups, this won't necessarily be
true. Consider the case where we do a single 256MiB write to a file and
then close it. We will have 1 reservation for the inode update, the
reservations for the checksum updates, and 1 reservation for the file
extent item. At some point later we decide to write this entire range
out, but we're so fragmented that we break this into 100 different file
extents. Since we've already closed the file and are no longer writing
to it there's nothing to trigger a refill of the delalloc block rsv to
satisfy the 99 new file extent reservations we need. At this point we
exhaust our delalloc reservation, and we begin to steal from the global
reserve. If you have enough of these cases going in parallel you can
easily exhaust the global reserve, get an ENOSPC at
btrfs_alloc_tree_block() time, and then abort the transaction.
The other case is the delayed refs reserve. The delayed refs reserve
updates its size based on outstanding delayed refs and dirty block
groups. However we only refill this block reserve when returning
excess reservations and when we call btrfs_start_transaction(root, X).
We will reserve 2*X credits at transaction start time, and fill in X
into the delayed refs reserve to make sure it stays topped off.
Generally this works well, but clearly has downsides. If we do a
particularly delayed ref heavy operation we may never catch up in our
reservations. Additionally running delayed refs generates more delayed
refs, and at that point we may be committing the transaction and have no
way to trigger a refill of our delayed refs rsv. Then a similar thing
occurs with the delalloc reserve.
Generally speaking we well over-reserve in all of our block rsvs. If we
reserve 1 credit we're usually reserving around 264k of space, but we'll
often not use any of that reservation, or use a few blocks of that
reservation. We can be reasonably sure that as long as you were able to
reserve space up front for your operation you'll be able to find space
on disk for that reservation.
So introduce a new flushing state, BTRFS_RESERVE_FLUSH_EMERGENCY. This
gets used in the case that we've exhausted our reserve and the global
reserve. It simply forces a reservation if we have enough actual space
on disk to make the reservation, which is almost always the case. This
keeps us from hitting ENOSPC aborts in these odd occurrences where we've
not kept up with the delayed work.
Fixing this in a complete way is going to be relatively complicated and
time consuming. This patch is what I discussed with Filipe earlier this
year, and what I put into our kernels inside FB. With this patch we're
down to 1-2 ENOSPC aborts per week, which is a significant reduction.
This is a decent stop gap until we can work out a more wholistic
solution to these two corner cases.
Reviewed-by: Filipe Manana <fdmanana@suse.com>
Signed-off-by: Josef Bacik <josef@toxicpanda.com>
Signed-off-by: David Sterba <dsterba@suse.com>
In order to accommodate NOWAIT IOCB's we need to be able to do NO_FLUSH
data reservations, so plumb this through the delalloc reservation
system.
Reviewed-by: Filipe Manana <fdmanana@suse.com>
Signed-off-by: Josef Bacik <josef@toxicpanda.com>
Signed-off-by: Stefan Roesch <shr@fb.com>
Signed-off-by: David Sterba <dsterba@suse.com>
At space-info.c:__reserve_bytes(), we increment the 'used' variable, but
then we don't use the variable anymore, making the increment pointless.
The increment became useless with commit 2e294c6049 ("btrfs: simplify
the logic in need_preemptive_flushing"), so just remove it.
Reviewed-by: Josef Bacik <josef@toxicpanda.com>
Signed-off-by: Filipe Manana <fdmanana@suse.com>
Reviewed-by: David Sterba <dsterba@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
We have hit some transaction abort due to -ENOSPC internally.
Normally we should always reserve enough space for metadata for every
transaction, thus hitting -ENOSPC should really indicate some cases we
didn't expect.
But unfortunately current error reporting will only give a kernel
warning and stack trace, not really helpful to debug what's causing the
problem.
And mount option debug_enospc can only help when user can reproduce the
problem, but under most cases, such transaction abort by -ENOSPC is
really hard to reproduce.
So this patch will dump all space infos (data, metadata, system) when we
abort the first transaction with -ENOSPC.
This should at least provide some clue to us.
The example of a dump would look like this:
BTRFS: Transaction aborted (error -28)
WARNING: CPU: 8 PID: 3366 at fs/btrfs/transaction.c:2137 btrfs_commit_transaction+0xf81/0xfb0 [btrfs]
<call trace skipped>
---[ end trace 0000000000000000 ]---
BTRFS info (device dm-1: state A): dumping space info:
BTRFS info (device dm-1: state A): space_info DATA has 6791168 free, is not full
BTRFS info (device dm-1: state A): space_info total=8388608, used=1597440, pinned=0, reserved=0, may_use=0, readonly=0 zone_unusable=0
BTRFS info (device dm-1: state A): space_info METADATA has 257114112 free, is not full
BTRFS info (device dm-1: state A): space_info total=268435456, used=131072, pinned=180224, reserved=65536, may_use=10878976, readonly=65536 zone_unusable=0
BTRFS info (device dm-1: state A): space_info SYSTEM has 8372224 free, is not full
BTRFS info (device dm-1: state A): space_info total=8388608, used=16384, pinned=0, reserved=0, may_use=0, readonly=0 zone_unusable=0
BTRFS info (device dm-1: state A): global_block_rsv: size 3670016 reserved 3670016
BTRFS info (device dm-1: state A): trans_block_rsv: size 0 reserved 0
BTRFS info (device dm-1: state A): chunk_block_rsv: size 0 reserved 0
BTRFS info (device dm-1: state A): delayed_block_rsv: size 4063232 reserved 4063232
BTRFS info (device dm-1: state A): delayed_refs_rsv: size 3145728 reserved 3145728
BTRFS: error (device dm-1: state A) in btrfs_commit_transaction:2137: errno=-28 No space left
BTRFS info (device dm-1: state EA): forced readonly
Reviewed-by: Johannes Thumshirn <johannes.thumshirn@wdc.com>
Signed-off-by: Qu Wenruo <wqu@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
For btrfs_space_info, its flags has only 4 possible values:
- BTRFS_BLOCK_GROUP_SYSTEM
- BTRFS_BLOCK_GROUP_METADATA | BTRFS_BLOCK_GROUP_DATA
- BTRFS_BLOCK_GROUP_METADATA
- BTRFS_BLOCK_GROUP_DATA
Make the output more human readable, now it looks like:
BTRFS info (device dm-1: state A): space_info METADATA has 251494400 free, is not full
Reviewed-by: Johannes Thumshirn <johannes.thumshirn@wdc.com>
Signed-off-by: Qu Wenruo <wqu@suse.com>
Reviewed-by: David Sterba <dsterba@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
We use a bit field in the btrfs_block_group for different flags, however
this is awkward because we have to hold the block_group->lock for any
modification of any of these fields, and makes the code clunky for a few
of these flags. Convert these to a properly flags setup so we can
utilize the bit helpers.
Reviewed-by: Johannes Thumshirn <johannes.thumshirn@wdc.com>
Signed-off-by: Josef Bacik <josef@toxicpanda.com>
Reviewed-by: David Sterba <dsterba@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
We previously had the pattern of
btrfs_update_space_info(all, the, bg, fields, &space_info);
link_block_group(bg);
bg->space_info = space_info;
Now that we're passing the bg into btrfs_add_bg_to_space_info we can do
the linking in that function, transforming this to simply
btrfs_add_bg_to_space_info(fs_info, bg);
and put the link_block_group() and bg->space_info assignment directly in
btrfs_add_bg_to_space_info.
Signed-off-by: Josef Bacik <josef@toxicpanda.com>
Reviewed-by: David Sterba <dsterba@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
This function has grown a bunch of new arguments, and it just boils down
to passing in all the block group fields as arguments. Simplify this by
passing in the block group itself and updating the space_info fields
based on the block group fields directly.
Signed-off-by: Josef Bacik <josef@toxicpanda.com>
Reviewed-by: David Sterba <dsterba@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
[BEHAVIOR CHANGE]
Since commit f6fca3917b ("btrfs: store chunk size in space-info
struct"), btrfs no longer can create larger data chunks than 1G:
mkfs.btrfs -f -m raid1 -d raid0 $dev1 $dev2 $dev3 $dev4
mount $dev1 $mnt
btrfs balance start --full $mnt
btrfs balance start --full $mnt
umount $mnt
btrfs ins dump-tree -t chunk $dev1 | grep "DATA|RAID0" -C 2
Before that offending commit, what we got is a 4G data chunk:
item 6 key (FIRST_CHUNK_TREE CHUNK_ITEM 9492758528) itemoff 15491 itemsize 176
length 4294967296 owner 2 stripe_len 65536 type DATA|RAID0
io_align 65536 io_width 65536 sector_size 4096
num_stripes 4 sub_stripes 1
Now what we got is only 1G data chunk:
item 6 key (FIRST_CHUNK_TREE CHUNK_ITEM 6271533056) itemoff 15491 itemsize 176
length 1073741824 owner 2 stripe_len 65536 type DATA|RAID0
io_align 65536 io_width 65536 sector_size 4096
num_stripes 4 sub_stripes 1
This will increase the number of data chunks by the number of devices,
not only increase system chunk usage, but also greatly increase mount
time.
Without a proper reason, we should not change the max chunk size.
[CAUSE]
Previously, we set max data chunk size to 10G, while max data stripe
length to 1G.
Commit f6fca3917b ("btrfs: store chunk size in space-info struct")
completely ignored the 10G limit, but use 1G max stripe limit instead,
causing above shrink in max data chunk size.
[FIX]
Fix the max data chunk size to 10G, and in decide_stripe_size_regular()
we limit stripe_size to 1G manually.
This should only affect data chunks, as for metadata chunks we always
set the max stripe size the same as max chunk size (256M or 1G
depending on fs size).
Now the same script result the same old result:
item 6 key (FIRST_CHUNK_TREE CHUNK_ITEM 9492758528) itemoff 15491 itemsize 176
length 4294967296 owner 2 stripe_len 65536 type DATA|RAID0
io_align 65536 io_width 65536 sector_size 4096
num_stripes 4 sub_stripes 1
Reported-by: Wang Yugui <wangyugui@e16-tech.com>
Fixes: f6fca3917b ("btrfs: store chunk size in space-info struct")
Signed-off-by: Qu Wenruo <wqu@suse.com>
Reviewed-by: David Sterba <dsterba@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
For metadata space on zoned filesystem, reaching ALLOC_CHUNK{,_FORCE}
means we don't have enough space left in the active_total_bytes. Before
allocating a new chunk, we can try to activate an existing block group
in this case.
Also, allocating a chunk is not enough to grant a ticket for metadata
space on zoned filesystem we need to activate the block group to
increase the active_total_bytes.
btrfs_zoned_activate_one_bg() implements the activation feature. It will
activate a block group by (maybe) finishing a block group. It will give up
activating a block group if it cannot finish any block group.
CC: stable@vger.kernel.org # 5.16+
Fixes: afba2bc036 ("btrfs: zoned: implement active zone tracking")
Signed-off-by: Naohiro Aota <naohiro.aota@wdc.com>
Signed-off-by: David Sterba <dsterba@suse.com>
The metadata overcommit makes the space reservation flexible but it is also
harmful to active zone tracking. Since we cannot finish a block group from
the metadata allocation context, we might not activate a new block group
and might not be able to actually write out the overcommit reservations.
So, disable metadata overcommit for zoned filesystems. We will ensure
the reservations are under active_total_bytes in the following patches.
CC: stable@vger.kernel.org # 5.16+
Fixes: afba2bc036 ("btrfs: zoned: implement active zone tracking")
Signed-off-by: Naohiro Aota <naohiro.aota@wdc.com>
Reviewed-by: Johannes Thumshirn <johannes.thumshirn@wdc.com>
Signed-off-by: David Sterba <dsterba@suse.com>
The active_total_bytes, like the total_bytes, accounts for the total bytes
of active block groups in the space_info.
With an introduction of active_total_bytes, we can check if the reserved
bytes can be written to the block groups without activating a new block
group. The check is necessary for metadata allocation on zoned
filesystem. We cannot finish a block group, which may require waiting
for the current transaction, from the metadata allocation context.
Instead, we need to ensure the ongoing allocation (reserved bytes) fits
in active block groups.
Signed-off-by: Naohiro Aota <naohiro.aota@wdc.com>
Signed-off-by: David Sterba <dsterba@suse.com>
The chunk size is stored in the btrfs_space_info structure. It is
initialized at the start and is then used.
A new API is added to update the current chunk size. This API is used
to be able to expose the chunk_size as a sysfs setting.
Signed-off-by: Stefan Roesch <shr@fb.com>
Reviewed-by: David Sterba <dsterba@suse.com>
[ rename and merge helpers, switch atomic type to u64, style fixes ]
Signed-off-by: David Sterba <dsterba@suse.com>
For non-zoned file systems it's useful to have the auto reclaim feature,
however there are different use cases for non-zoned, for example we may
not want to reclaim metadata chunks ever, only data chunks. Move this
sysfs flag to per-space_info. This won't affect current users because
this tunable only ever did anything for zoned, and that is currently
hidden behind BTRFS_CONFIG_DEBUG.
Tested-by: Pankaj Raghav <p.raghav@samsung.com>
Signed-off-by: Josef Bacik <josef@toxicpanda.com>
[ jth restore global bg_reclaim_threshold ]
Signed-off-by: Johannes Thumshirn <johannes.thumshirn@wdc.com>
Reviewed-by: David Sterba <dsterba@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
Explicit type casts are not necessary when it's void* to another pointer
type.
Signed-off-by: Yu Zhe <yuzhe@nfschina.com>
Reviewed-by: David Sterba <dsterba@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
In a previous patch ("btrfs: extend locking to all space_info members
accesses") the locking for the space_info members was extended in
btrfs_preempt_reclaim_metadata_space because not all the member
accesses that needed locks were actually locked (bytes_pinned et al).
It was then suggested to also add a call to lockdep_assert_held to
need_preemptive_reclaim. This function also works with space_info
members. As of now, it has only two call sites which both hold the lock.
Suggested-by: Johannes Thumshirn <johannes.thumshirn@wdc.com>
Reviewed-by: Johannes Thumshirn <johannes.thumshirn@wdc.com>
Signed-off-by: Niels Dossche <dossche.niels@gmail.com>
Reviewed-by: David Sterba <dsterba@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
bytes_pinned is always accessed under space_info->lock, except in
btrfs_preempt_reclaim_metadata_space, however the other members are
accessed under that lock. The reserved member of the rsv's are also
partially accessed under a lock and partially not. Move all these
accesses into the same lock to ensure consistency.
This could potentially race and lead to a flush instead of a commit but
it's not a big problem as it's only for preemptive flush.
CC: stable@vger.kernel.org # 5.15+
Reviewed-by: Johannes Thumshirn <johannes.thumshirn@wdc.com>
Reviewed-by: Josef Bacik <josef@toxicpanda.com>
Signed-off-by: Niels Dossche <niels.dossche@ugent.be>
Signed-off-by: Niels Dossche <dossche.niels@gmail.com>
Reviewed-by: David Sterba <dsterba@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
The warnings were found by running scripts/kernel-doc, which is
caused by using 'make W=1'.
fs/btrfs/extent_io.c:3210: warning: Function parameter or member
'bio_ctrl' not described in 'btrfs_bio_add_page'
fs/btrfs/extent_io.c:3210: warning: Excess function parameter 'bio'
description in 'btrfs_bio_add_page'
fs/btrfs/extent_io.c:3210: warning: Excess function parameter
'prev_bio_flags' description in 'btrfs_bio_add_page'
fs/btrfs/space-info.c:1602: warning: Excess function parameter 'root'
description in 'btrfs_reserve_metadata_bytes'
fs/btrfs/space-info.c:1602: warning: Function parameter or member
'fs_info' not described in 'btrfs_reserve_metadata_bytes'
Note: this is fixing only the warnings regarding parameter list, the
first line is not strictly conforming to the kdoc format as the btrfs
codebase does not stick to that and keeps the first line more free form
(because it's only for internal use).
Reported-by: Abaci Robot <abaci@linux.alibaba.com>
Signed-off-by: Yang Li <yang.lee@linux.alibaba.com>
Reviewed-by: David Sterba <dsterba@suse.com>
[ add note ]
Signed-off-by: David Sterba <dsterba@suse.com>
We only need the root to start a transaction, and since it's a global
root we can pick anything, change to the tree_root as we'll have a lot
of extent roots in the future.
Signed-off-by: Josef Bacik <josef@toxicpanda.com>
Reviewed-by: David Sterba <dsterba@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
We used to need the root for btrfs_reserve_metadata_bytes to check the
orphan cleanup state, but we no longer need that, we simply need the
fs_info. Change btrfs_reserve_metadata_bytes() to use the fs_info, and
change both btrfs_block_rsv_refill() and btrfs_block_rsv_add() to do the
same as they simply call btrfs_reserve_metadata_bytes() and then
manipulate the block_rsv that is being used.
Reviewed-by: Nikolay Borisov <nborisov@suse.com>
Signed-off-by: Josef Bacik <josef@toxicpanda.com>
Reviewed-by: David Sterba <dsterba@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
This is very old code before we were stealing from the global reserve
during evict. We have proper ways to steal from the global reserve
while we're evicting, so rip out this code as it's no longer necessary.
Reviewed-by: Nikolay Borisov <nborisov@suse.com>
Signed-off-by: Josef Bacik <josef@toxicpanda.com>
Signed-off-by: David Sterba <dsterba@suse.com>
I forgot to convert this over when I introduced the global reserve
stealing code to the space flushing code. Evict was simply trying to
make its reservation and then if it failed it would steal from the
global rsv, which is racey because it's outside of the normal ticketing
code.
Fix this by setting ticket->steal if we are BTRFS_RESERVE_FLUSH_EVICT,
and then make the priority flushing path do the steal for us.
Reviewed-by: Nikolay Borisov <nborisov@suse.com>
Signed-off-by: Josef Bacik <josef@toxicpanda.com>
Signed-off-by: David Sterba <dsterba@suse.com>
We're going to use this helper in the priority flushing loop, move this
check into the helper to simplify the logic.
Reviewed-by: Nikolay Borisov <nborisov@suse.com>
Signed-off-by: Josef Bacik <josef@toxicpanda.com>
Reviewed-by: David Sterba <dsterba@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
Since we're dropping locks before we enter the priority flushing loops
we could have had our ticket granted before we got the space_info->lock.
So add this check to avoid doing some extra flushing in the priority
flushing cases.
The case in priority_reclaim_metadata_space is an optimization. Think
we came in to reserve, we didn't have the space, we added our ticket to
the list. But at the same time somebody was waiting on the space_info
lock to add space and do btrfs_try_granting_ticket(), so we drop the
lock, get satisfied, come in to do our loop, and we have been
satisfied.
This is the priority reclaim path, so to_reclaim could be !0 still
because we may have only satisfied the priority tickets and still left
non priority tickets on the list. We would then have to_reclaim but
->bytes == 0.
Reviewed-by: Nikolay Borisov <nborisov@suse.com>
Signed-off-by: Josef Bacik <josef@toxicpanda.com>
[ add note about the optimization ]
Signed-off-by: David Sterba <dsterba@suse.com>
Currently the error case for the priority tickets is handled where we
deal with all of the tickets, priority and non-priority. This is OK in
general, but it makes for some awkward locking. We take and drop the
space_info->lock back to back because of these different types of
tickets.
Rework the code to handle priority ticket failures in their respective
helpers. This allows us to be less wonky with our space_info->lock
usage, and means that the main handler simply has to check
ticket->error, as the ticket is guaranteed to be off any list and
completely handled by the time it exits one of the handlers.
Reviewed-by: Nikolay Borisov <nborisov@suse.com>
Signed-off-by: Josef Bacik <josef@toxicpanda.com>
Reviewed-by: David Sterba <dsterba@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
Error injection stressing uncovered a busy loop in our data reclaim
loop. There are two cases here, one where we loop creating block groups
until space_info->full is set, or in the main loop we will skip erroring
out any tickets if space_info->full == 0. Unfortunately if we aborted
the transaction then we will never allocate chunks or reclaim any space
and thus never get ->full, and you'll see stack traces like this:
watchdog: BUG: soft lockup - CPU#0 stuck for 26s! [kworker/u4:4:139]
CPU: 0 PID: 139 Comm: kworker/u4:4 Tainted: G W 5.13.0-rc1+ #328
Hardware name: QEMU Standard PC (Q35 + ICH9, 2009), BIOS 1.13.0-2.fc32 04/01/2014
Workqueue: events_unbound btrfs_async_reclaim_data_space
RIP: 0010:btrfs_join_transaction+0x12/0x20
RSP: 0018:ffffb2b780b77de0 EFLAGS: 00000246
RAX: ffffb2b781863d58 RBX: 0000000000000000 RCX: 0000000000000000
RDX: 0000000000000801 RSI: ffff987952b57400 RDI: ffff987940aa3000
RBP: ffff987954d55000 R08: 0000000000000001 R09: ffff98795539e8f0
R10: 000000000000000f R11: 000000000000000f R12: ffffffffffffffff
R13: ffff987952b574c8 R14: ffff987952b57400 R15: 0000000000000008
FS: 0000000000000000(0000) GS:ffff9879bbc00000(0000) knlGS:0000000000000000
CS: 0010 DS: 0000 ES: 0000 CR0: 0000000080050033
CR2: 00007f0703da4000 CR3: 0000000113398004 CR4: 0000000000370ef0
Call Trace:
flush_space+0x4a8/0x660
btrfs_async_reclaim_data_space+0x55/0x130
process_one_work+0x1e9/0x380
worker_thread+0x53/0x3e0
? process_one_work+0x380/0x380
kthread+0x118/0x140
? __kthread_bind_mask+0x60/0x60
ret_from_fork+0x1f/0x30
Fix this by checking to see if we have a btrfs fs error in either of the
reclaim loops, and if so fail the tickets and bail. In addition to
this, fix maybe_fail_all_tickets() to not try to grant tickets if we've
aborted, simply fail everything.
Reviewed-by: Nikolay Borisov <nborisov@suse.com>
Reviewed-by: Filipe Manana <fdmanana@suse.com>
Signed-off-by: Josef Bacik <josef@toxicpanda.com>
Signed-off-by: David Sterba <dsterba@suse.com>
It's not uncommon where __btrfs_dump_space_info() gets called
under over-commit situations.
In that case free space would underflow as total allocated space is not
enough to handle all the over-committed space.
Such underflow values can sometimes cause confusion for users enabled
enospc_debug mount option, and takes some seconds for developers to
convert the underflow value to signed result.
Just output the free space as s64 to avoid such problem.
Reported-by: Eli V <eliventer@gmail.com>
Link: https://lore.kernel.org/linux-btrfs/CAJtFHUSy4zgyhf-4d9T+KdJp9w=UgzC2A0V=VtmaeEpcGgm1-Q@mail.gmail.com/
CC: stable@vger.kernel.org # 5.4+
Reviewed-by: Anand Jain <anand.jain@oracle.com>
Signed-off-by: Qu Wenruo <wqu@suse.com>
Reviewed-by: David Sterba <dsterba@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
A common characteristic of the bug report where preemptive flushing was
going full tilt was the fact that the vast majority of the free metadata
space was used up by the global reserve. The hard 90% threshold would
cover the majority of these cases, but to be even smarter we should take
into account how much of the outstanding reservations are covered by the
global block reserve. If the global block reserve accounts for the vast
majority of outstanding reservations, skip preemptive flushing, as it
will likely just cause churn and pain.
Bugzilla: https://bugzilla.kernel.org/show_bug.cgi?id=212185
Signed-off-by: Josef Bacik <josef@toxicpanda.com>
Signed-off-by: David Sterba <dsterba@suse.com>
The preemptive flushing code was added in order to avoid needing to
synchronously wait for ENOSPC flushing to recover space. Once we're
almost full however we can essentially flush constantly. We were using
98% as a threshold to determine if we were simply full, however in
practice this is a really high bar to hit. For example reports of
systems running into this problem had around 94% usage and thus
continued to flush. Fix this by lowering the threshold to 90%, which is
a more sane value, especially for smaller file systems.
Bugzilla: https://bugzilla.kernel.org/show_bug.cgi?id=212185
CC: stable@vger.kernel.org # 5.12+
Fixes: 576fa34830 ("btrfs: improve preemptive background space flushing")
Signed-off-by: Josef Bacik <josef@toxicpanda.com>
Signed-off-by: David Sterba <dsterba@suse.com>
I've been debugging an early ENOSPC problem in production and finally
root caused it to this problem. When we switched to the per-inode in
38d715f494 ("btrfs: use btrfs_start_delalloc_roots in
shrink_delalloc") I pulled out the async extent handling, because we
were doing the correct thing by calling filemap_flush() if we had async
extents set. This would properly wait on any async extents by locking
the page in the second flush, thus making sure our ordered extents were
properly set up.
However when I switched us back to page based flushing, I used
sync_inode(), which allows us to pass in our own wbc. The problem here
is that sync_inode() is smarter than the filemap_* helpers, it tries to
avoid calling writepages at all. This means that our second call could
skip calling do_writepages altogether, and thus not wait on the pagelock
for the async helpers. This means we could come back before any ordered
extents were created and then simply continue on in our flushing
mechanisms and ENOSPC out when we have plenty of space to use.
Fix this by putting back the async pages logic in shrink_delalloc. This
allows us to bulk write out everything that we need to, and then we can
wait in one place for the async helpers to catch up, and then wait on
any ordered extents that are created.
Fixes: e076ab2a2c ("btrfs: shrink delalloc pages instead of full inodes")
CC: stable@vger.kernel.org # 5.10+
Reviewed-by: Nikolay Borisov <nborisov@suse.com>
Signed-off-by: Josef Bacik <josef@toxicpanda.com>
Signed-off-by: David Sterba <dsterba@suse.com>
We have been hitting some early ENOSPC issues in production with more
recent kernels, and I tracked it down to us simply not flushing delalloc
as aggressively as we should be. With tracing I was seeing us failing
all tickets with all of the block rsvs at or around 0, with very little
pinned space, but still around 120MiB of outstanding bytes_may_used.
Upon further investigation I saw that we were flushing around 14 pages
per shrink call for delalloc, despite having around 2GiB of delalloc
outstanding.
Consider the example of a 8 way machine, all CPUs trying to create a
file in parallel, which at the time of this commit requires 5 items to
do. Assuming a 16k leaf size, we have 10MiB of total metadata reclaim
size waiting on reservations. Now assume we have 128MiB of delalloc
outstanding. With our current math we would set items to 20, and then
set to_reclaim to 20 * 256k, or 5MiB.
Assuming that we went through this loop all 3 times, for both
FLUSH_DELALLOC and FLUSH_DELALLOC_WAIT, and then did the full loop
twice, we'd only flush 60MiB of the 128MiB delalloc space. This could
leave a fair bit of delalloc reservations still hanging around by the
time we go to ENOSPC out all the remaining tickets.
Fix this two ways. First, change the calculations to be a fraction of
the total delalloc bytes on the system. Prior to this change we were
calculating based on dirty inodes so our math made more sense, now it's
just completely unrelated to what we're actually doing.
Second add a FLUSH_DELALLOC_FULL state, that we hold off until we've
gone through the flush states at least once. This will empty the system
of all delalloc so we're sure to be truly out of space when we start
failing tickets.
I'm tagging stable 5.10 and forward, because this is where we started
using the page stuff heavily again. This affects earlier kernel
versions as well, but would be a pain to backport to them as the
flushing mechanisms aren't the same.
CC: stable@vger.kernel.org # 5.10+
Reviewed-by: Nikolay Borisov <nborisov@suse.com>
Signed-off-by: Josef Bacik <josef@toxicpanda.com>
Signed-off-by: David Sterba <dsterba@suse.com>
When debugging early enospc problems it was useful to have a tracepoint
where we failed all tickets so I could check the state of the enospc
counters at failure time to validate my fixes. This adds the tracpoint
so you can easily get that information.
Reviewed-by: Nikolay Borisov <nborisov@suse.com>
Signed-off-by: Josef Bacik <josef@toxicpanda.com>
Reviewed-by: David Sterba <dsterba@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
We used this in may_commit_transaction() in order to determine if we
needed to commit the transaction. However we no longer have that logic
and thus have no use of this counter anymore, so delete it.
Reviewed-by: Nikolay Borisov <nborisov@suse.com>
Signed-off-by: Josef Bacik <josef@toxicpanda.com>
Reviewed-by: David Sterba <dsterba@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
This was a trick implemented to handle the case where we had a giant
reservation in front of a bunch of little reservations in the ticket
queue. If the giant reservation was too large for the transaction
commit to make a difference we'd ENOSPC everybody out instead of
committing the transaction. This logic was put in to force us to go
back and re-try the transaction commit logic to see if we could make
progress.
Instead now we know we've committed the transaction, so any space that
would have been recovered is now available, and would be caught by the
btrfs_try_granting_tickets() in this loop, so we no longer need this
code and can simply delete it.
Reviewed-by: Nikolay Borisov <nborisov@suse.com>
Signed-off-by: Josef Bacik <josef@toxicpanda.com>
Signed-off-by: David Sterba <dsterba@suse.com>
Since we unconditionally commit the transaction now we no longer need to
run the delayed refs to make sure our total_bytes_pinned value is
uptodate, we can simply commit the transaction. Remove this stage from
the data flushing list.
Reviewed-by: Nikolay Borisov <nborisov@suse.com>
Signed-off-by: Josef Bacik <josef@toxicpanda.com>
Signed-off-by: David Sterba <dsterba@suse.com>
may_commit_transaction was introduced before the ticketing
infrastructure existed. There was a problem where we'd legitimately be
out of space, but every reservation would trigger a transaction commit
and then fail. Thus if you had 1000 things trying to make a
reservation, they'd all do the flushing loop and thus commit the
transaction 1000 times before they'd get their ENOSPC.
This helper was introduced to short circuit this, if there wasn't space
that could be reclaimed by committing the transaction then simply ENOSPC
out. This made true ENOSPC tests much faster as we didn't waste a bunch
of time.
However many of our bugs over the years have been from cases where we
didn't account for some space that would be reclaimed by committing a
transaction. The delayed refs rsv space, delayed rsv, many pinned bytes
miscalculations, etc. And in the meantime the original problem has been
solved with ticketing. We no longer will commit the transaction 1000
times. Instead we'll get 1000 waiters, we will go through the flushing
mechanisms, and if there's no progress after 2 loops we ENOSPC everybody
out. The ticketing infrastructure gives us a deterministic way to see
if we're making progress or not, thus we avoid a lot of extra work.
So simplify this step by simply unconditionally committing the
transaction. This removes what is arguably our most common source of
early ENOSPC bugs and will allow us to drastically simplify many of the
things we track because we simply won't need them with this stuff gone.
Reviewed-by: Nikolay Borisov <nborisov@suse.com>
Signed-off-by: Josef Bacik <josef@toxicpanda.com>
Signed-off-by: David Sterba <dsterba@suse.com>
If we decide to flush delalloc from the preemptive flusher, we really do
not want to wait on ordered extents, as it gains us nothing. However
there was logic to go ahead and wait on ordered extents if there was
more ordered bytes than delalloc bytes. We do not want this behavior,
so pass through whether this flushing is for preemption, and do not wait
for ordered extents if that's the case. Also break out of the shrink
loop after the first flushing, as we just want to one shot shrink
delalloc.
Signed-off-by: Josef Bacik <josef@toxicpanda.com>
Signed-off-by: David Sterba <dsterba@suse.com>
While testing heavy delalloc workloads I noticed that sometimes we'd
just stop preemptively flushing when we had loads of delalloc available
to flush. This is because we skip preemptive flushing if delalloc <=
ordered. However if we start with say 4gib of delalloc, and we flush
2gib of that, we'll stop flushing there, when we still have 2gib of
delalloc to flush.
Instead adjust the ordered bytes down by half, this way if 2/3 of our
outstanding delalloc reservations are tied up by ordered extents we
don't bother preemptive flushing, as we're getting close to the state
where we need to wait on ordered extents.
Signed-off-by: Josef Bacik <josef@toxicpanda.com>
Signed-off-by: David Sterba <dsterba@suse.com>
When deciding if we should preemptively flush space, we will add in the
amount of space used by all block rsvs. However this also includes the
global block rsv, which isn't flushable so shouldn't be accounted for in
this calculation. If we decide to use ->bytes_may_use in our used
calculation we need to subtract the global rsv size from this amount so
it most closely matches the flushable space.
Reviewed-by: Johannes Thumshirn <johannes.thumshirn@wdc.com>
Signed-off-by: Josef Bacik <josef@toxicpanda.com>
Signed-off-by: David Sterba <dsterba@suse.com>
We calculate the amount of "free" space available for normal
reservations by taking the total space and subtracting out the hard used
space, which is readonly, used, and reserved space.
However we weren't taking into account the global block rsv, which is
essentially hard used space. Handle this by subtracting it from the
available free space, so that our threshold more closely mirrors
reality.
We need to do the check because it's possible that the global_rsv_size +
used is > total_bytes, sometimes the global reserve can end up being
calculated as larger than the available size (think small filesystems
where we only have the original 8MiB chunk of metadata). It doesn't
usually happen, but that can get us into trouble so this is safer.
Signed-off-by: Josef Bacik <josef@toxicpanda.com>
Signed-off-by: David Sterba <dsterba@suse.com>
Global rsv can't be used for normal allocations, and for very full file
systems we can decide to try and async flush constantly even though
there's really not a lot of space to reclaim. Deal with this by
including the global block rsv size in the "total used" calculation.
Reviewed-by: Johannes Thumshirn <johannes.thumshirn@wdc.com>
Signed-off-by: Josef Bacik <josef@toxicpanda.com>
Signed-off-by: David Sterba <dsterba@suse.com>
We were clamping the threshold for preemptive reclaim any time we added
a ticket to wait on, which if we have a lot of threads means we'd
essentially max out the clamp the first time we start to flush.
Instead of doing this, simply do it every time we have to start
flushing, this will make us ramp up gradually instead of going to max
clamping as soon as we start needing to do flushing.
Signed-off-by: Josef Bacik <josef@toxicpanda.com>
Signed-off-by: David Sterba <dsterba@suse.com>
need_preemptive_reclaim() does some calculations, which aren't heavy,
but if we're already running preemptive reclaim there's no reason to do
them at all, so re-order the checks so that we don't do the calculation
if we're already doing reclaim.
Reviewed-by: Johannes Thumshirn <johannes.thumshirn@wdc.com>
Signed-off-by: Josef Bacik <josef@toxicpanda.com>
Reviewed-by: David Sterba <dsterba@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
In a zoned filesystem a once written then freed region is not usable
until the underlying zone has been reset. So we need to distinguish such
unusable space from usable free space.
Therefore we need to introduce the "zone_unusable" field to the block
group structure, and "bytes_zone_unusable" to the space_info structure
to track the unusable space.
Pinned bytes are always reclaimed to the unusable space. But, when an
allocated region is returned before using e.g., the block group becomes
read-only between allocation time and reservation time, we can safely
return the region to the block group. For the situation, this commit
introduces "btrfs_add_free_space_unused". This behaves the same as
btrfs_add_free_space() on regular filesystem. On zoned filesystems, it
rewinds the allocation offset.
Because the read-only bytes tracks free but unusable bytes when the block
group is read-only, we need to migrate the zone_unusable bytes to
read-only bytes when a block group is marked read-only.
Reviewed-by: Josef Bacik <josef@toxicpanda.com>
Signed-off-by: Naohiro Aota <naohiro.aota@wdc.com>
Signed-off-by: David Sterba <dsterba@suse.com>
Often when I'm debugging ENOSPC related issues I have to resort to
printing the entire ENOSPC state with trace_printk() in different spots.
This gets pretty annoying, so add a trace state that does this for us.
Then add a trace point at the end of preemptive flushing so you can see
the state of the space_info when we decide to exit preemptive flushing.
This helped me figure out we weren't kicking in the preemptive flushing
soon enough.
Reviewed-by: Nikolay Borisov <nborisov@suse.com>
Signed-off-by: Josef Bacik <josef@toxicpanda.com>
Reviewed-by: David Sterba <dsterba@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
Since we have normal ticketed flushing and preemptive flushing, adjust
the tracepoint so that we know the source of the flushing action to make
it easier to debug problems.
Reviewed-by: Nikolay Borisov <nborisov@suse.com>
Signed-off-by: Josef Bacik <josef@toxicpanda.com>
Reviewed-by: David Sterba <dsterba@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
Starting preemptive flushing at 50% of available free space is a good
start, but some workloads are particularly abusive and can quickly
overwhelm the preemptive flushing code and drive us into using tickets.
Handle this by clamping down on our threshold for starting and
continuing to run preemptive flushing. This is particularly important
for our overcommit case, as we can really drive the file system into
overages and then it's more difficult to pull it back as we start to
actually fill up the file system.
The clamping is essentially 2^CLAMP, but we start at 1 so whatever we
calculate for overcommit is the baseline.
Reviewed-by: Nikolay Borisov <nborisov@suse.com>
Signed-off-by: Josef Bacik <josef@toxicpanda.com>
Signed-off-by: David Sterba <dsterba@suse.com>
