Whenever we add a ticket to a space_info object we increment the object's
reclaim_size counter witht the ticket's bytes, and we decrement it with
the corresponding amount only when we are able to grant the requested
space to the ticket. When we are not able to grant the space to a ticket,
or when the ticket is removed due to a signal (e.g. an application has
received sigterm from the terminal) we never decrement the counter with
the corresponding bytes from the ticket. This leak can result in the
space reclaim code to later do much more work than necessary. So fix it
by decrementing the counter when those two cases happen as well.
Fixes: db161806dc ("btrfs: account ticket size at add/delete time")
Reviewed-by: Nikolay Borisov <nborisov@suse.com>
Signed-off-by: Filipe Manana <fdmanana@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
Instead of iterating all pending tickets on the normal/priority list to
sum their total size the cost can be amortized across ticket addition/
removal. This turns O(n) + O(m) (where n is the size of the normal list
and m of the priority list) into O(1). This will mostly have effect in
workloads that experience heavy flushing.
Signed-off-by: Nikolay Borisov <nborisov@suse.com>
Reviewed-by: David Sterba <dsterba@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
I noticed while running my snapshot torture test that we were getting a
lot of metadata chunks allocated with very little actually used.
Digging into this we would commit the transaction, still not have enough
space, and then force a chunk allocation.
I noticed that we were barely flushing any delalloc at all, despite the
fact that we had around 13gib of outstanding delalloc reservations. It
turns out this is because of our btrfs_calc_reclaim_metadata_size()
calculation. It _only_ takes into account the outstanding ticket sizes,
which isn't the whole story. In this particular workload we're slowly
filling up the disk, which means our overcommit space will suddenly
become a lot less, and our outstanding reservations will be well more
than what we can handle. However we are only flushing based on our
ticket size, which is much less than we need to actually reclaim.
So fix btrfs_calc_reclaim_metadata_size() to take into account the
overage in the case that we've gotten less available space suddenly.
This makes it so we attempt to reclaim a lot more delalloc space, which
allows us to make our reservations and we no longer are allocating a
bunch of needless metadata chunks.
CC: stable@vger.kernel.org # 4.4+
Signed-off-by: Josef Bacik <josef@toxicpanda.com>
Signed-off-by: David Sterba <dsterba@suse.com>
Add another comment to cover how the space reservation system works
generally. This covers the actual reservation flow, as well as how
flushing is handled.
Signed-off-by: Josef Bacik <josef@toxicpanda.com>
Signed-off-by: David Sterba <dsterba@suse.com>
inc_block_group_ro does a calculation to see if we have enough room left
over if we mark this block group as read only in order to see if it's ok
to mark the block group as read only.
The problem is this calculation _only_ works for data, where our used is
always less than our total. For metadata we will overcommit, so this
will almost always fail for metadata.
Fix this by exporting btrfs_can_overcommit, and then see if we have
enough space to remove the remaining free space in the block group we
are trying to mark read only. If we do then we can mark this block
group as read only.
Reviewed-by: Qu Wenruo <wqu@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 have the space_info, we can just check its flags to see if it's the
system chunk space info.
Reviewed-by: Nikolay Borisov <nborisov@suse.com>
Reviewed-by: Qu Wenruo <wqu@suse.com>
Reviewed-by: Johannes Thumshirn <jthumshirn@suse.de>
Signed-off-by: Josef Bacik <josef@toxicpanda.com>
Reviewed-by: David Sterba <dsterba@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
The type name is misleading, a single entry is named 'cache' while this
normally means a collection of objects. Rename that everywhere. Also the
identifier was quite long, making function prototypes harder to format.
Suggested-by: Nikolay Borisov <nborisov@suse.com>
Reviewed-by: Qu Wenruo <wqu@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
It is not used anymore since commit 957780eb27 ("Btrfs: introduce
ticketed enospc infrastructure"), so just remove it.
Signed-off-by: Filipe Manana <fdmanana@suse.com>
Reviewed-by: David Sterba <dsterba@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
The on-disk format of block group item makes use of the key that stores
the offset and length. This is further used in the code, although this
makes thing harder to understand. The key is also packed so the
offset/length is not properly aligned as u64.
Add start (key.objectid) and length (key.offset) members to block group
and remove the embedded key. When the item is searched or written, a
local variable for key is used.
Reviewed-by: Johannes Thumshirn <jthumshirn@suse.de>
Reviewed-by: Nikolay Borisov <nborisov@suse.com>
Reviewed-by: Qu Wenruo <wqu@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
For unknown reasons, the member 'used' in the block group struct is
stored in the b-tree item and accessed everywhere using the special
accessor helper. Let's unify it and make it a regular member and only
update the item before writing it to the tree.
The item is still being used for flags and chunk_objectid, there's some
duplication until the item is removed in following patches.
Reviewed-by: Johannes Thumshirn <jthumshirn@suse.de>
Reviewed-by: Nikolay Borisov <nborisov@suse.com>
Reviewed-by: Qu Wenruo <wqu@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
The attribute is more relaxed than const and the functions could
dereference pointers, as long as the observable state is not changed. We
do have such functions, based on -Wsuggest-attribute=pure .
The visible effects of this patch are negligible, there are differences
in the assembly but hard to summarize.
Reviewed-by: Nikolay Borisov <nborisov@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
When a task that is allocating metadata needs to wait for the async
reclaim job to process its ticket and gets a signal (because it was killed
for example) before doing the wait, the task ends up erroring out but
with space reserved for its ticket, which never gets released, resulting
in a metadata space leak (more specifically a leak in the bytes_may_use
counter of the metadata space_info object).
Here's the sequence of steps leading to the space leak:
1) A task tries to create a file for example, so it ends up trying to
start a transaction at btrfs_create();
2) The filesystem is currently in a state where there is not enough
metadata free space to satisfy the transaction's needs. So at
space-info.c:__reserve_metadata_bytes() we create a ticket and
add it to the list of tickets of the space info object. Also,
because the metadata async reclaim job is not running, we queue
a job ro run metadata reclaim;
3) In the meanwhile the task receives a signal (like SIGTERM from
a kill command for example);
4) After queing the async reclaim job, at __reserve_metadata_bytes(),
we unlock the metadata space info and call handle_reserve_ticket();
5) That last function calls wait_reserve_ticket(), which acquires the
lock from the metadata space info. Then in the first iteration of
its while loop, it calls prepare_to_wait_event(), which returns
-ERESTARTSYS because the task has a pending signal. As a result,
we set the error field of the ticket to -EINTR and exit the while
loop without deleting the ticket from the list of tickets (in the
space info object). After exiting the loop we unlock the space info;
6) The async reclaim job is able to release enough metadata, acquires
the metadata space info's lock and then reserves space for the ticket,
since the ticket is still in the list of (non-priority) tickets. The
space reservation happens at btrfs_try_granting_tickets(), called from
maybe_fail_all_tickets(). This increments the bytes_may_use counter
from the metadata space info object, sets the ticket's bytes field to
zero (meaning success, that space was reserved) and removes it from
the list of tickets;
7) wait_reserve_ticket() returns, with the error field of the ticket
set to -EINTR. Then handle_reserve_ticket() just propagates that error
to the caller. Because an error was returned, the caller does not
release the reserved space, since the expectation is that any error
means no space was reserved.
Fix this by removing the ticket from the list, while holding the space
info lock, at wait_reserve_ticket() when prepare_to_wait_event() returns
an error.
Also add some comments and an assertion to guarantee we never end up with
a ticket that has an error set and a bytes counter field set to zero, to
more easily detect regressions in the future.
This issue could be triggered sporadically by some test cases from fstests
such as generic/269 for example, which tries to fill a filesystem and then
kills fsstress processes running in the background.
When this issue happens, we get a warning in syslog/dmesg when unmounting
the filesystem, like the following:
------------[ cut here ]------------
WARNING: CPU: 0 PID: 13240 at fs/btrfs/block-group.c:3186 btrfs_free_block_groups+0x314/0x470 [btrfs]
(...)
CPU: 0 PID: 13240 Comm: umount Tainted: G W L 5.3.0-rc8-btrfs-next-48+ #1
Hardware name: QEMU Standard PC (i440FX + PIIX, 1996), BIOS rel-1.12.0-0-ga698c8995f-prebuilt.qemu.org 04/01/2014
RIP: 0010:btrfs_free_block_groups+0x314/0x470 [btrfs]
(...)
RSP: 0018:ffff9910c14cfdb8 EFLAGS: 00010286
RAX: 0000000000000024 RBX: ffff89cd8a4d55f0 RCX: 0000000000000000
RDX: 0000000000000000 RSI: ffff89cdf6a178a8 RDI: ffff89cdf6a178a8
RBP: ffff9910c14cfde8 R08: 0000000000000000 R09: 0000000000000001
R10: ffff89cd4d618040 R11: 0000000000000000 R12: ffff89cd8a4d5508
R13: ffff89cde7c4a600 R14: dead000000000122 R15: dead000000000100
FS: 00007f42754432c0(0000) GS:ffff89cdf6a00000(0000) knlGS:0000000000000000
CS: 0010 DS: 0000 ES: 0000 CR0: 0000000080050033
CR2: 00007fd25a47f730 CR3: 000000021f8d6006 CR4: 00000000003606f0
DR0: 0000000000000000 DR1: 0000000000000000 DR2: 0000000000000000
DR3: 0000000000000000 DR6: 00000000fffe0ff0 DR7: 0000000000000400
Call Trace:
close_ctree+0x1ad/0x390 [btrfs]
generic_shutdown_super+0x6c/0x110
kill_anon_super+0xe/0x30
btrfs_kill_super+0x12/0xa0 [btrfs]
deactivate_locked_super+0x3a/0x70
cleanup_mnt+0xb4/0x160
task_work_run+0x7e/0xc0
exit_to_usermode_loop+0xfa/0x100
do_syscall_64+0x1cb/0x220
entry_SYSCALL_64_after_hwframe+0x49/0xbe
RIP: 0033:0x7f4274d2cb37
(...)
RSP: 002b:00007ffcff701d38 EFLAGS: 00000246 ORIG_RAX: 00000000000000a6
RAX: 0000000000000000 RBX: 0000557ebde2f060 RCX: 00007f4274d2cb37
RDX: 0000000000000001 RSI: 0000000000000000 RDI: 0000557ebde2f240
RBP: 0000557ebde2f240 R08: 0000557ebde2f270 R09: 0000000000000015
R10: 00000000000006b4 R11: 0000000000000246 R12: 00007f427522ee64
R13: 0000000000000000 R14: 0000000000000000 R15: 00007ffcff701fc0
irq event stamp: 0
hardirqs last enabled at (0): [<0000000000000000>] 0x0
hardirqs last disabled at (0): [<ffffffffb12b561e>] copy_process+0x75e/0x1fd0
softirqs last enabled at (0): [<ffffffffb12b561e>] copy_process+0x75e/0x1fd0
softirqs last disabled at (0): [<0000000000000000>] 0x0
---[ end trace bcf4b235461b26f6 ]---
BTRFS info (device sdb): space_info 4 has 19116032 free, is full
BTRFS info (device sdb): space_info total=33554432, used=14176256, pinned=0, reserved=0, may_use=196608, readonly=65536
BTRFS info (device sdb): global_block_rsv: size 0 reserved 0
BTRFS info (device sdb): trans_block_rsv: size 0 reserved 0
BTRFS info (device sdb): chunk_block_rsv: size 0 reserved 0
BTRFS info (device sdb): delayed_block_rsv: size 0 reserved 0
BTRFS info (device sdb): delayed_refs_rsv: size 0 reserved 0
Fixes: 374bf9c5cd ("btrfs: unify error handling for ticket flushing")
Reviewed-by: Josef Bacik <josef@toxicpanda.com>
Signed-off-by: Filipe Manana <fdmanana@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
When debugging weird enospc problems it's handy to be able to dump the
space info when we wake up all tickets, and see what the ticket values
are. This helped me figure out cases where we were enospc'ing when we
shouldn't have been.
Reviewed-by: Nikolay Borisov <nborisov@suse.com>
Signed-off-by: Josef Bacik <josef@toxicpanda.com>
Signed-off-by: David Sterba <dsterba@suse.com>
We ran into a problem in production where a box with plenty of space was
getting wedged doing ENOSPC flushing. These boxes only had 20% of the
disk allocated, but their metadata space + global reserve was right at
the size of their metadata chunk.
In this case can_overcommit should be allowing allocations without
problem, but there's logic in can_overcommit that doesn't allow us to
overcommit if there's not enough real space to satisfy the global
reserve.
This is for historical reasons. Before there were only certain places
we could allocate chunks. We could go to commit the transaction and not
have enough space for our pending delayed refs and such and be unable to
allocate a new chunk. This would result in a abort because of ENOSPC.
This code was added to solve this problem.
However since then we've gained the ability to always be able to
allocate a chunk. So we can easily overcommit in these cases without
risking a transaction abort because of ENOSPC.
Also prior to now the global reserve really would be used because that's
the space we relied on for delayed refs. With delayed refs being
tracked separately we no longer have to worry about running out of
delayed refs space while committing. We are much less likely to
exhaust our global reserve space during transaction commit.
Fix the can_overcommit code to simply see if our current usage + what we
want is less than our current free space plus whatever slack space we
have in the disk is. This solves the problem we were seeing in
production and keeps us from flushing as aggressively as we approach our
actual metadata size usage.
Reviewed-by: Nikolay Borisov <nborisov@suse.com>
Signed-off-by: Josef Bacik <josef@toxicpanda.com>
Signed-off-by: David Sterba <dsterba@suse.com>
Now that we do not do partial filling of tickets simply remove
orig_bytes, it is no longer needed.
Reviewed-by: Nikolay Borisov <nborisov@suse.com>
Signed-off-by: Josef Bacik <josef@toxicpanda.com>
Signed-off-by: David Sterba <dsterba@suse.com>
Now that we aren't partially filling tickets we may have some slack
space left in the space_info. We need to account for this in
may_commit_transaction, otherwise we may choose to not commit the
transaction despite it actually having enough space to satisfy our
ticket.
Calculate the free space we have in the space_info, if any, and subtract
this from the ticket we have and use that amount to determine if we will
need to commit to reclaim enough space.
Reviewed-by: Nikolay Borisov <nborisov@suse.com>
Signed-off-by: Josef Bacik <josef@toxicpanda.com>
Signed-off-by: David Sterba <dsterba@suse.com>
Now that we no longer partially fill tickets we need to rework
wake_all_tickets to call btrfs_try_to_wakeup_tickets() in order to see
if any subsequent tickets are able to be satisfied. If our tickets_id
changes we know something happened and we can keep flushing.
Also if we find a ticket that is smaller than the first ticket in our
queue then we want to retry the flushing loop again in case
may_commit_transaction() decides we could satisfy the ticket by
committing the transaction.
Rename this to maybe_fail_all_tickets() while we're at it, to better
reflect what the function is actually doing.
Signed-off-by: Josef Bacik <josef@toxicpanda.com>
Signed-off-by: David Sterba <dsterba@suse.com>
Now that btrfs_space_info_add_old_bytes simply checks if we can make the
reservation and updates bytes_may_use, there's no reason to have both
helpers in place.
Factor out the ticket wakeup logic into it's own helper, make
btrfs_space_info_add_old_bytes() update bytes_may_use and then call the
wakeup helper, and replace all calls to btrfs_space_info_add_new_bytes()
with the wakeup helper.
Reviewed-by: Nikolay Borisov <nborisov@suse.com>
Signed-off-by: Josef Bacik <josef@toxicpanda.com>
Signed-off-by: David Sterba <dsterba@suse.com>
btrfs_space_info_add_old_bytes is used when adding the extra space from
an existing reservation back into the space_info to be used by any
waiting tickets. In order to keep us from overcommitting we check to
make sure that we can still use this space for our reserve ticket, and
if we cannot we'll simply subtract it from space_info->bytes_may_use.
However this is problematic, because it assumes that only changes to
bytes_may_use would affect our ability to make reservations. Any
changes to bytes_reserved would be missed. If we were unable to make a
reservation prior because of reserved space, but that reserved space was
free'd due to unlink or truncate and we were allowed to immediately
reclaim that metadata space we would still ENOSPC.
Consider the example where we create a file with a bunch of extents,
using up 2MiB of actual space for the new tree blocks. Then we try to
make a reservation of 2MiB but we do not have enough space to make this
reservation. The iput() occurs in another thread and we remove this
space, and since we did not write the blocks we simply do
space_info->bytes_reserved -= 2MiB. We would never see this because we
do not check our space info used, we just try to re-use the freed
reservations.
To fix this problem, and to greatly simplify the wakeup code, do away
with this partial refilling nonsense. Use
btrfs_space_info_add_old_bytes to subtract the reservation from
space_info->bytes_may_use, and then check the ticket against the total
used of the space_info the same way we do with the initial reservation
attempt.
This keeps the reservation logic consistent and solves the problem of
early ENOSPC in the case that we free up space in places other than
bytes_may_use and bytes_pinned. Thanks,
Signed-off-by: Josef Bacik <josef@toxicpanda.com>
Signed-off-by: David Sterba <dsterba@suse.com>
We duplicate this tracepoint everywhere we call these helpers, so update
the helper to have the tracepoint as well.
Signed-off-by: Josef Bacik <josef@toxicpanda.com>
Reviewed-by: David Sterba <dsterba@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
If we already have tickets on the list we don't want to steal their
reservations. This is a preparation patch for upcoming changes,
technically this shouldn't happen today because of the way we add bytes
to tickets before adding them to the space_info in most cases.
This does not change the FIFO nature of reserve tickets, it simply
allows us to enforce it in a different way. Previously it was enforced
because any new space would be added to the first ticket on the list,
which would result in new reservations getting a reserve ticket. This
replaces that mechanism by simply checking to see if we have outstanding
reserve tickets and skipping straight to adding a ticket for our
reservation.
Reviewed-by: Nikolay Borisov <nborisov@suse.com>
Signed-off-by: Josef Bacik <josef@toxicpanda.com>
Signed-off-by: David Sterba <dsterba@suse.com>
btrfs_calc_trunc_metadata_size differs from trans_metadata_size in that
it doesn't take into account any splitting at the levels, because
truncate will never split nodes. However truncate _and_ changing will
never split nodes, so rename btrfs_calc_trunc_metadata_size to
btrfs_calc_metadata_size. Also btrfs_calc_trans_metadata_size is purely
for inserting items, so rename this to btrfs_calc_insert_metadata_size.
Making these clearer will help when I start using them differently in
upcoming patches.
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 have this weird space flushing loop inside inode.c for evict where
we'll do the normal LIMIT flush, and then commit the transaction and
hope we get our space. This is super janky, and in fact there's really
nothing stopping us from using FLUSH_ALL except that we run delayed
iputs, which means we could deadlock. So introduce a new flush state
for eviction that does the normal priority flushing with all of the
states that are safe for eviction.
The nice side-effect of this is that we'll try harder for evictions.
Previously if (for example generic/269) you had a bunch of other
operations happening on the fs you could race with those reservations
when committing the transaction, and eventually miss getting a
reservation for the evict. With this code we'll have our ticket in
place through the transaction commit, so any pinned bytes will go to our
pending evictions first.
Signed-off-by: Josef Bacik <josef@toxicpanda.com>
Signed-off-by: David Sterba <dsterba@suse.com>
With the eviction flushing stuff we'll want to allow for different
states, but still work basically the same way that
priority_reclaim_metadata_space works currently. Refactor this to take
the flushing states and size as an argument so we can use the same logic
for limit flushing and eviction flushing.
Signed-off-by: Josef Bacik <josef@toxicpanda.com>
Reviewed-by: David Sterba <dsterba@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
We're going to make this logic a little more complicated for evict, so
factor the ticket flushing/waiting code out of __reserve_metadata_bytes.
This has no functional change.
Signed-off-by: Josef Bacik <josef@toxicpanda.com>
Reviewed-by: David Sterba <dsterba@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
Currently we handle the cleanup of errored out tickets in both the
priority flush path and the normal flushing path. This is the same code
in both places, so just refactor so we don't duplicate the cleanup work.
Signed-off-by: Josef Bacik <josef@toxicpanda.com>
Reviewed-by: David Sterba <dsterba@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
Delayed iputs could very well free up enough space without needing to
commit the transaction, so make this step it's own step. This will
allow us to skip the step for evictions in a later patch.
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 prep work for moving all of the block group cache code into its
own file.
Signed-off-by: Josef Bacik <josef@toxicpanda.com>
Reviewed-by: David Sterba <dsterba@suse.com>
[ minor comment updates ]
Signed-off-by: David Sterba <dsterba@suse.com>
We don't need an if-else-if chain where we can use a simple OR since
both conditions are performing the same action. The short-circuit for OR
will ensure that if the first condition is true, can_overcommit() is not
called.
Reviewed-by: Nikolay Borisov <nborisov@suse.com>
Signed-off-by: Goldwyn Rodrigues <rgoldwyn@suse.com>
Reviewed-by: David Sterba <dsterba@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
Now that we've moved all of the users to space-info.c, unexport it and
name it back to can_overcommit.
Reviewed-by: Nikolay Borisov <nborisov@suse.com>
Signed-off-by: Josef Bacik <josef@toxicpanda.com>
Signed-off-by: David Sterba <dsterba@suse.com>
This moves all of the metadata reservation code into space-info.c.
Reviewed-by: Nikolay Borisov <nborisov@suse.com>
Signed-off-by: Josef Bacik <josef@toxicpanda.com>
Signed-off-by: David Sterba <dsterba@suse.com>
We'll need this exported so we can use it in all the various was we need
to use it. This is prep work to move reserve_metadata_bytes.
Reviewed-by: Nikolay Borisov <nborisov@suse.com>
Signed-off-by: Josef Bacik <josef@toxicpanda.com>
Signed-off-by: David Sterba <dsterba@suse.com>
Now that we've moved all the pre-requisite stuff, move these two
functions.
Reviewed-by: Nikolay Borisov <nborisov@suse.com>
Signed-off-by: Josef Bacik <josef@toxicpanda.com>
Signed-off-by: David Sterba <dsterba@suse.com>
This is the first piece of moving the space reservation code to
space-info.c
Reviewed-by: Nikolay Borisov <nborisov@suse.com>
Signed-off-by: Josef Bacik <josef@toxicpanda.com>
Signed-off-by: David Sterba <dsterba@suse.com>
These are the basic init and lookup functions and some helper functions,
fairly straightforward before the bad stuff starts.
Reviewed-by: Nikolay Borisov <nborisov@suse.com>
Signed-off-by: Josef Bacik <josef@toxicpanda.com>
Signed-off-by: David Sterba <dsterba@suse.com>
