Boris Burkov [Thu, 16 Jul 2020 20:29:46 +0000 (13:29 -0700)]
btrfs: fix mount failure caused by race with umount
It is possible to cause a btrfs mount to fail by racing it with a slow
umount. The crux of the sequence is generic_shutdown_super not yet
calling sop->put_super before btrfs_mount_root calls btrfs_open_devices.
If that occurs, btrfs_open_devices will decide the opened counter is
non-zero, increment it, and skip resetting fs_devices->total_rw_bytes to
0. From here, mount will call sget which will result in grab_super
trying to take the super block umount semaphore. That semaphore will be
held by the slow umount, so mount will block. Before up-ing the
semaphore, umount will delete the super block, resulting in mount's sget
reliably allocating a new one, which causes the mount path to dutifully
fill it out, and increment total_rw_bytes a second time, which causes
the mount to fail, as we see double the expected bytes.
To fix this, we clear total_rw_bytes from within btrfs_read_chunk_tree
before the calls to read_one_dev, while holding the sb umount semaphore
and the uuid mutex.
To reproduce, it is sufficient to dirty a decent number of inodes, then
quickly umount and mount.
for i in $(seq 0 500)
do
dd if=/dev/zero of="/mnt/foo/$i" bs=1M count=1
done
umount /mnt/foo&
mount /mnt/foo
Robbie Ko [Mon, 20 Jul 2020 01:42:09 +0000 (09:42 +0800)]
btrfs: fix page leaks after failure to lock page for delalloc
When locking pages for delalloc, we check if it's dirty and mapping still
matches. If it does not match, we need to return -EAGAIN and release all
pages. Only the current page was put though, iterate over all the
remaining pages too.
btrfs: qgroup: fix data leak caused by race between writeback and truncate
[BUG]
When running tests like generic/013 on test device with btrfs quota
enabled, it can normally lead to data leak, detected at unmount time:
BTRFS warning (device dm-3): qgroup 0/5 has unreleased space, type 0 rsv 4096
------------[ cut here ]------------
WARNING: CPU: 11 PID: 16386 at fs/btrfs/disk-io.c:4142 close_ctree+0x1dc/0x323 [btrfs]
RIP: 0010:close_ctree+0x1dc/0x323 [btrfs]
Call Trace:
btrfs_put_super+0x15/0x17 [btrfs]
generic_shutdown_super+0x72/0x110
kill_anon_super+0x18/0x30
btrfs_kill_super+0x17/0x30 [btrfs]
deactivate_locked_super+0x3b/0xa0
deactivate_super+0x40/0x50
cleanup_mnt+0x135/0x190
__cleanup_mnt+0x12/0x20
task_work_run+0x64/0xb0
__prepare_exit_to_usermode+0x1bc/0x1c0
__syscall_return_slowpath+0x47/0x230
do_syscall_64+0x64/0xb0
entry_SYSCALL_64_after_hwframe+0x44/0xa9
---[ end trace caf08beafeca2392 ]---
BTRFS error (device dm-3): qgroup reserved space leaked
[CAUSE]
In the offending case, the offending operations are:
2/6: writev f2X[269 1 0 0 0 0] [1006997,67,288] 0
2/7: truncate f2X[269 1 0 0 48 1026293] 18388 0
The following sequence of events could happen after the writev():
CPU1 (writeback) | CPU2 (truncate)
-----------------------------------------------------------------
btrfs_writepages() |
|- extent_write_cache_pages() |
|- Got page for 1003520 |
| 1003520 is Dirty, no writeback |
| So (!clear_page_dirty_for_io()) |
| gets called for it |
|- Now page 1003520 is Clean. |
| | btrfs_setattr()
| | |- btrfs_setsize()
| | |- truncate_setsize()
| | New i_size is 18388
|- __extent_writepage() |
| |- page_offset() > i_size |
|- btrfs_invalidatepage() |
|- Page is clean, so no qgroup |
callback executed
This means, the qgroup reserved data space is not properly released in
btrfs_invalidatepage() as the page is Clean.
[FIX]
Instead of checking the dirty bit of a page, call
btrfs_qgroup_free_data() unconditionally in btrfs_invalidatepage().
As qgroup rsv are completely bound to the QGROUP_RESERVED bit of
io_tree, not bound to page status, thus we won't cause double freeing
anyway.
Filipe Manana [Mon, 13 Jul 2020 14:11:56 +0000 (15:11 +0100)]
btrfs: fix double free on ulist after backref resolution failure
At btrfs_find_all_roots_safe() we allocate a ulist and set the **roots
argument to point to it. However if later we fail due to an error returned
by find_parent_nodes(), we free that ulist but leave a dangling pointer in
the **roots argument. Upon receiving the error, a caller of this function
can attempt to free the same ulist again, resulting in an invalid memory
access.
One such scenario is during qgroup accounting:
btrfs_qgroup_account_extents()
--> calls btrfs_find_all_roots() passes &new_roots (a stack allocated
pointer) to btrfs_find_all_roots()
--> btrfs_find_all_roots() just calls btrfs_find_all_roots_safe()
passing &new_roots to it
--> allocates ulist and assigns its address to **roots (which
points to new_roots from btrfs_qgroup_account_extents())
--> find_parent_nodes() returns an error, so we free the ulist
and leave **roots pointing to it after returning
--> btrfs_qgroup_account_extents() sees btrfs_find_all_roots() returned
an error and jumps to the label 'cleanup', which just tries to
free again the same ulist
btrfs implements the iter_write op and thus can use the more efficient
iov_iter based splice implementation. For now falling back to the less
efficient default is pretty harmless, but I have a pending series that
removes the default, and thus would cause btrfs to not support splice
at all.
Josef Bacik [Mon, 6 Jul 2020 13:14:12 +0000 (09:14 -0400)]
btrfs: fix double put of block group with nocow
While debugging a patch that I wrote I was hitting use-after-free panics
when accessing block groups on unmount. This turned out to be because
in the nocow case if we bail out of doing the nocow for whatever reason
we need to call btrfs_dec_nocow_writers() if we called the inc. This
puts our block group, but a few error cases does
if (nocow) {
btrfs_dec_nocow_writers();
goto error;
}
unfortunately, error is
error:
if (nocow)
btrfs_dec_nocow_writers();
so we get a double put on our block group. Fix this by dropping the
error cases calling of btrfs_dec_nocow_writers(), as it's handled at the
error label now.
Josef Bacik [Tue, 30 Jun 2020 18:53:02 +0000 (14:53 -0400)]
btrfs: reset tree root pointer after error in init_tree_roots
Eric reported an issue where mounting -o recovery with a fuzzed fs
resulted in a kernel panic. This is because we tried to free the tree
node, except it was an error from the read. Fix this by properly
resetting the tree_root->node == NULL in this case. The panic was the
following
Nik says: this is problematic only if we fail on the last iteration of
the loop as this results in init_tree_roots returning err value with
tree_root->node = -ERR. Subsequently the caller does: fail_tree_roots
which calls free_root_pointers on the bogus value.
Filipe Manana [Sat, 27 Jun 2020 10:40:44 +0000 (11:40 +0100)]
btrfs: fix reclaim_size counter leak after stealing from global reserve
Commit 7f9fe614407692 ("btrfs: improve global reserve stealing logic"),
added in the 5.8 merge window, introduced another leak for the space_info's
reclaim_size counter. This is very often triggered by the test cases
generic/269 and generic/416 from fstests, producing a stack trace like the
following during unmount:
In the past commit d611add48b717a ("btrfs: fix reclaim counter leak of
space_info objects") fixed similar cases. That commit however has a date
more recent (April 7 2020) then the commit mentioned before (March 13
2020), however it was merged in kernel 5.7 while the older commit, which
introduces a new leak, was merged only in the 5.8 merge window. So the
leak sneaked in unnoticed.
Fix this by making steal_from_global_rsv() remove the ticket using the
helper remove_ticket(), which decrements the reclaim_size counter of the
space_info object.
Fixes: 7f9fe614407692 ("btrfs: improve global reserve stealing logic") Signed-off-by: Filipe Manana <[email protected]> Reviewed-by: David Sterba <[email protected]> Signed-off-by: David Sterba <[email protected]>
Boris Burkov [Wed, 17 Jun 2020 18:35:19 +0000 (11:35 -0700)]
btrfs: fix fatal extent_buffer readahead vs releasepage race
Under somewhat convoluted conditions, it is possible to attempt to
release an extent_buffer that is under io, which triggers a BUG_ON in
btrfs_release_extent_buffer_pages.
This relies on a few different factors. First, extent_buffer reads done
as readahead for searching use WAIT_NONE, so they free the local extent
buffer reference while the io is outstanding. However, they should still
be protected by TREE_REF. However, if the system is doing signficant
reclaim, and simultaneously heavily accessing the extent_buffers, it is
possible for releasepage to race with two concurrent readahead attempts
in a way that leaves TREE_REF unset when the readahead extent buffer is
released.
Essentially, if two tasks race to allocate a new extent_buffer, but the
winner who attempts the first io is rebuffed by a page being locked
(likely by the reclaim itself) then the loser will still go ahead with
issuing the readahead. The loser's call to find_extent_buffer must also
race with the reclaim task reading the extent_buffer's refcount as 1 in
a way that allows the reclaim to re-clear the TREE_REF checked by
find_extent_buffer.
The following represents an example execution demonstrating the race:
CPU0 CPU1 CPU2
reada_for_search reada_for_search
readahead_tree_block readahead_tree_block
find_create_tree_block find_create_tree_block
alloc_extent_buffer alloc_extent_buffer
find_extent_buffer // not found
allocates eb
lock pages
associate pages to eb
insert eb into radix tree
set TREE_REF, refs == 2
unlock pages
read_extent_buffer_pages // WAIT_NONE
not uptodate (brand new eb)
lock_page
if !trylock_page
goto unlock_exit // not an error
free_extent_buffer
release_extent_buffer
atomic_dec_and_test refs to 1
find_extent_buffer // found
try_release_extent_buffer
take refs_lock
reads refs == 1; no io
atomic_inc_not_zero refs to 2
mark_buffer_accessed
check_buffer_tree_ref
// not STALE, won't take refs_lock
refs == 2; TREE_REF set // no action
read_extent_buffer_pages // WAIT_NONE
clear TREE_REF
release_extent_buffer
atomic_dec_and_test refs to 1
unlock_page
still not uptodate (CPU1 read failed on trylock_page)
locks pages
set io_pages > 0
submit io
return
free_extent_buffer
release_extent_buffer
dec refs to 0
delete from radix tree
btrfs_release_extent_buffer_pages
BUG_ON(io_pages > 0)!!!
We observe this at a very low rate in production and were also able to
reproduce it in a test environment by introducing some spurious delays
and by introducing probabilistic trylock_page failures.
To fix it, we apply check_tree_ref at a point where it could not
possibly be unset by a competing task: after io_pages has been
incremented. All the codepaths that clear TREE_REF check for io, so they
would not be able to clear it after this point until the io is done.
Waiman Long [Tue, 16 Jun 2020 15:31:59 +0000 (11:31 -0400)]
btrfs: use kfree() in btrfs_ioctl_get_subvol_info()
In btrfs_ioctl_get_subvol_info(), there is a classic case where kzalloc()
was incorrectly paired with kzfree(). According to David Sterba, there
isn't any sensitive information in the subvol_info that needs to be
cleared before freeing. So kzfree() isn't really needed, use kfree()
instead.
Filipe Manana [Mon, 15 Jun 2020 17:49:39 +0000 (18:49 +0100)]
btrfs: fix RWF_NOWAIT writes blocking on extent locks and waiting for IO
A RWF_NOWAIT write is not supposed to wait on filesystem locks that can be
held for a long time or for ongoing IO to complete.
However when calling check_can_nocow(), if the inode has prealloc extents
or has the NOCOW flag set, we can block on extent (file range) locks
through the call to btrfs_lock_and_flush_ordered_range(). Such lock can
take a significant amount of time to be available. For example, a fiemap
task may be running, and iterating through the entire file range checking
all extents and doing backref walking to determine if they are shared,
or a readpage operation may be in progress.
Also at btrfs_lock_and_flush_ordered_range(), called by check_can_nocow(),
after locking the file range we wait for any existing ordered extent that
is in progress to complete. Another operation that can take a significant
amount of time and defeat the purpose of RWF_NOWAIT.
So fix this by trying to lock the file range and if it's currently locked
return -EAGAIN to user space. If we are able to lock the file range without
waiting and there is an ordered extent in the range, return -EAGAIN as
well, instead of waiting for it to complete. Finally, don't bother trying
to lock the snapshot lock of the root when attempting a RWF_NOWAIT write,
as that is only important for buffered writes.
Fixes: edf064e7c6fec3 ("btrfs: nowait aio support") Signed-off-by: Filipe Manana <[email protected]> Signed-off-by: David Sterba <[email protected]>
Filipe Manana [Mon, 15 Jun 2020 17:49:13 +0000 (18:49 +0100)]
btrfs: fix RWF_NOWAIT write not failling when we need to cow
If we attempt to do a RWF_NOWAIT write against a file range for which we
can only do NOCOW for a part of it, due to the existence of holes or
shared extents for example, we proceed with the write as if it were
possible to NOCOW the whole range.
Example:
$ mkfs.btrfs -f /dev/sdb
$ mount /dev/sdb /mnt
$ touch /mnt/sdj/bar
$ chattr +C /mnt/sdj/bar
$ xfs_io -d -c "pwrite -S 0xab -b 256K 0 256K" /mnt/bar
wrote 262144/262144 bytes at offset 0
256 KiB, 1 ops; 0.0003 sec (694.444 MiB/sec and 2777.7778 ops/sec)
This last write should fail with -EAGAIN since the file range from 64K to
128K is a hole. On xfs it fails, as expected, but on ext4 it currently
succeeds because apparently it is expensive to check if there are extents
allocated for the whole range, but I'll check with the ext4 people.
Fix the issue by checking if check_can_nocow() returns a number of
NOCOW'able bytes smaller then the requested number of bytes, and if it
does return -EAGAIN.
Filipe Manana [Mon, 15 Jun 2020 17:48:58 +0000 (18:48 +0100)]
btrfs: fix failure of RWF_NOWAIT write into prealloc extent beyond eof
If we attempt to write to prealloc extent located after eof using a
RWF_NOWAIT write, we always fail with -EAGAIN.
We do actually check if we have an allocated extent for the write at
the start of btrfs_file_write_iter() through a call to check_can_nocow(),
but later when we go into the actual direct IO write path we simply
return -EAGAIN if the write starts at or beyond EOF.
Trivial to reproduce:
$ mkfs.btrfs -f /dev/sdb
$ mount /dev/sdb /mnt
$ touch /mnt/foo
$ chattr +C /mnt/foo
$ xfs_io -d -c "pwrite -S 0xab 0 64K" /mnt/foo
wrote 65536/65536 bytes at offset 0
64 KiB, 16 ops; 0.0004 sec (135.575 MiB/sec and 34707.1584 ops/sec)
Filipe Manana [Mon, 15 Jun 2020 09:38:44 +0000 (10:38 +0100)]
btrfs: check if a log root exists before locking the log_mutex on unlink
This brings back an optimization that commit e678934cbe5f02 ("btrfs:
Remove unnecessary check from join_running_log_trans") removed, but in
a different form. So it's almost equivalent to a revert.
That commit removed an optimization where we avoid locking a root's
log_mutex when there is no log tree created in the current transaction.
The affected code path is triggered through unlink operations.
That commit was based on the assumption that the optimization was not
necessary because we used to have the following checks when the patch
was authored:
int btrfs_del_dir_entries_in_log(...)
{
(...)
if (dir->logged_trans < trans->transid)
return 0;
ret = join_running_log_trans(root);
(...)
}
int btrfs_del_inode_ref_in_log(...)
{
(...)
if (inode->logged_trans < trans->transid)
return 0;
ret = join_running_log_trans(root);
(...)
}
However before that patch was merged, another patch was merged first which
replaced those checks because they were buggy.
That other patch corresponds to commit 803f0f64d17769 ("Btrfs: fix fsync
not persisting dentry deletions due to inode evictions"). The assumption
that if the logged_trans field of an inode had a smaller value then the
current transaction's generation (transid) meant that the inode was not
logged in the current transaction was only correct if the inode was not
evicted and reloaded in the current transaction. So the corresponding bug
fix changed those checks and replaced them with the following helper
function:
So if we have a subvolume without a log tree in the current transaction
(because we had no fsyncs), every time we unlink an inode we can end up
trying to lock the log_mutex of the root through join_running_log_trans()
twice, once for the inode being unlinked (by btrfs_del_inode_ref_in_log())
and once for the parent directory (with btrfs_del_dir_entries_in_log()).
This means if we have several unlink operations happening in parallel for
inodes in the same subvolume, and the those inodes and/or their parent
inode were changed in the current transaction, we end up having a lot of
contention on the log_mutex.
The test robots from intel reported a -30.7% performance regression for
a REAIM test after commit e678934cbe5f02 ("btrfs: Remove unnecessary check
from join_running_log_trans").
So just bring back the optimization to join_running_log_trans() where we
check first if a log root exists before trying to lock the log_mutex. This
is done by checking for a bit that is set on the root when a log tree is
created and removed when a log tree is freed (at transaction commit time).
Commit e678934cbe5f02 ("btrfs: Remove unnecessary check from
join_running_log_trans") was merged in the 5.4 merge window while commit 803f0f64d17769 ("Btrfs: fix fsync not persisting dentry deletions due to
inode evictions") was merged in the 5.3 merge window. But the first
commit was actually authored before the second commit (May 23 2019 vs
June 19 2019).
Filipe Manana [Mon, 8 Jun 2020 12:33:05 +0000 (13:33 +0100)]
btrfs: fix bytes_may_use underflow when running balance and scrub in parallel
When balance and scrub are running in parallel it is possible to end up
with an underflow of the bytes_may_use counter of the data space_info
object, which triggers a warning like the following:
When relocating a data block group, for each extent allocated in the
block group we preallocate another extent with the same size for the
data relocation inode (we do it at prealloc_file_extent_cluster()).
We reserve space by calling btrfs_check_data_free_space(), which ends
up incrementing the data space_info's bytes_may_use counter, and
then call btrfs_prealloc_file_range() to allocate the extent, which
always decrements the bytes_may_use counter by the same amount.
The expectation is that writeback of the data relocation inode always
follows a NOCOW path, by writing into the preallocated extents. However,
when starting writeback we might end up falling back into the COW path,
because the block group that contains the preallocated extent was turned
into RO mode by a scrub running in parallel. The COW path then calls the
extent allocator which ends up calling btrfs_add_reserved_bytes(), and
this function decrements the bytes_may_use counter of the data space_info
object by an amount corresponding to the size of the allocated extent,
despite we haven't previously incremented it. When the counter currently
has a value smaller then the allocated extent we reset the counter to 0
and emit a warning, otherwise we just decrement it and slowly mess up
with this counter which is crucial for space reservation, the end result
can be granting reserved space to tasks when there isn't really enough
free space, and having the tasks fail later in critical places where
error handling consists of a transaction abort or hitting a BUG_ON().
Fix this by making sure that if we fallback to the COW path for a data
relocation inode, we increment the bytes_may_use counter of the data
space_info object. The COW path will then decrement it at
btrfs_add_reserved_bytes() on success or through its error handling part
by a call to extent_clear_unlock_delalloc() (which ends up calling
btrfs_clear_delalloc_extent() that does the decrement operation) in case
of an error.
Test case btrfs/061 from fstests could sporadically trigger this.
Filipe Manana [Mon, 8 Jun 2020 12:32:55 +0000 (13:32 +0100)]
btrfs: fix data block group relocation failure due to concurrent scrub
When running relocation of a data block group while scrub is running in
parallel, it is possible that the relocation will fail and abort the
current transaction with an -EINVAL error:
When relocating a data block group, for each allocated extent of the block
group, we preallocate another extent (at prealloc_file_extent_cluster()),
associated with the data relocation inode, and then dirty all its pages.
These preallocated extents have, and must have, the same size that extents
from the data block group being relocated have.
Later before we start the relocation stage that updates pointers (bytenr
field of file extent items) to point to the the new extents, we trigger
writeback for the data relocation inode. The expectation is that writeback
will write the pages to the previously preallocated extents, that it
follows the NOCOW path. That is generally the case, however, if a scrub
is running it may have turned the block group that contains those extents
into RO mode, in which case writeback falls back to the COW path.
However in the COW path instead of allocating exactly one extent with the
expected size, the allocator may end up allocating several smaller extents
due to free space fragmentation - because we tell it at cow_file_range()
that the minimum allocation size can match the filesystem's sector size.
This later breaks the relocation's expectation that an extent associated
to a file extent item in the data relocation inode has the same size as
the respective extent pointed by a file extent item in another tree - in
this case the extent to which the relocation inode poins to is smaller,
causing relocation.c:get_new_location() to return -EINVAL.
For example, if we are relocating a data block group X that has a logical
address of X and the block group has an extent allocated at the logical
address X + 128KiB with a size of 64KiB:
1) At prealloc_file_extent_cluster() we allocate an extent for the data
relocation inode with a size of 64KiB and associate it to the file
offset 128KiB (X + 128KiB - X) of the data relocation inode. This
preallocated extent was allocated at block group Z;
2) A scrub running in parallel turns block group Z into RO mode and
starts scrubing its extents;
3) Relocation triggers writeback for the data relocation inode;
4) When running delalloc (btrfs_run_delalloc_range()), we try first the
NOCOW path because the data relocation inode has BTRFS_INODE_PREALLOC
set in its flags. However, because block group Z is in RO mode, the
NOCOW path (run_delalloc_nocow()) falls back into the COW path, by
calling cow_file_range();
5) At cow_file_range(), in the first iteration of the while loop we call
btrfs_reserve_extent() to allocate a 64KiB extent and pass it a minimum
allocation size of 4KiB (fs_info->sectorsize). Due to free space
fragmentation, btrfs_reserve_extent() ends up allocating two extents
of 32KiB each, each one on a different iteration of that while loop;
6) Writeback of the data relocation inode completes;
7) Relocation proceeds and ends up at relocation.c:replace_file_extents(),
with a leaf which has a file extent item that points to the data extent
from block group X, that has a logical address (bytenr) of X + 128KiB
and a size of 64KiB. Then it calls get_new_location(), which does a
lookup in the data relocation tree for a file extent item starting at
offset 128KiB (X + 128KiB - X) and belonging to the data relocation
inode. It finds a corresponding file extent item, however that item
points to an extent that has a size of 32KiB, which doesn't match the
expected size of 64KiB, resuling in -EINVAL being returned from this
function and propagated up to __btrfs_cow_block(), which aborts the
current transaction.
To fix this make sure that at cow_file_range() when we call the allocator
we pass it a minimum allocation size corresponding the desired extent size
if the inode belongs to the data relocation tree, otherwise pass it the
filesystem's sector size as the minimum allocation size.
Filipe Manana [Mon, 1 Jun 2020 18:12:19 +0000 (19:12 +0100)]
btrfs: fix race between block group removal and block group creation
There is a race between block group removal and block group creation
when the removal is completed by a task running fitrim or scrub. When
this happens we end up failing the block group creation with an error
-EEXIST since we attempt to insert a duplicate block group item key
in the extent tree. That results in a transaction abort.
The race happens like this:
1) Task A is doing a fitrim, and at btrfs_trim_block_group() it freezes
block group X with btrfs_freeze_block_group() (until very recently
that was named btrfs_get_block_group_trimming());
2) Task B starts removing block group X, either because it's now unused
or due to relocation for example. So at btrfs_remove_block_group(),
while holding the chunk mutex and the block group's lock, it sets
the 'removed' flag of the block group and it sets the local variable
'remove_em' to false, because the block group is currently frozen
(its 'frozen' counter is > 0, until very recently this counter was
named 'trimming');
3) Task B unlocks the block group and the chunk mutex;
4) Task A is done trimming the block group and unfreezes the block group
by calling btrfs_unfreeze_block_group() (until very recently this was
named btrfs_put_block_group_trimming()). In this function we lock the
block group and set the local variable 'cleanup' to true because we
were able to decrement the block group's 'frozen' counter down to 0 and
the flag 'removed' is set in the block group.
Since 'cleanup' is set to true, it locks the chunk mutex and removes
the extent mapping representing the block group from the mapping tree;
5) Task C allocates a new block group Y and it picks up the logical address
that block group X had as the logical address for Y, because X was the
block group with the highest logical address and now the second block
group with the highest logical address, the last in the fs mapping tree,
ends at an offset corresponding to block group X's logical address (this
logical address selection is done at volumes.c:find_next_chunk()).
At this point the new block group Y does not have yet its item added
to the extent tree (nor the corresponding device extent items and
chunk item in the device and chunk trees). The new group Y is added to
the list of pending block groups in the transaction handle;
6) Before task B proceeds to removing the block group item for block
group X from the extent tree, which has a key matching:
task C while ending its transaction handle calls
btrfs_create_pending_block_groups(), which finds block group Y and
tries to insert the block group item for Y into the exten tree, which
fails with -EEXIST since logical offset is the same that X had and
task B hasn't yet deleted the key from the extent tree.
This failure results in a transaction abort, producing a stack like
the following:
Fix this simply by making btrfs_remove_block_group() remove the block
group's item from the extent tree before it flags the block group as
removed. Also make the free space deletion from the free space tree
before flagging the block group as removed, to avoid a similar race
with adding and removing free space entries for the free space tree.
Fixes: 04216820fe83d5 ("Btrfs: fix race between fs trimming and block group remove/allocation") CC: [email protected] # 4.4+ Signed-off-by: Filipe Manana <[email protected]> Signed-off-by: David Sterba <[email protected]>
Filipe Manana [Mon, 1 Jun 2020 18:12:06 +0000 (19:12 +0100)]
btrfs: fix a block group ref counter leak after failure to remove block group
When removing a block group, if we fail to delete the block group's item
from the extent tree, we jump to the 'out' label and end up decrementing
the block group's reference count once only (by 1), resulting in a counter
leak because the block group at that point was already removed from the
block group cache rbtree - so we have to decrement the reference count
twice, once for the rbtree and once for our lookup at the start of the
function.
There is a second bug where if removing the free space tree entries (the
call to remove_block_group_free_space()) fails we end up jumping to the
'out_put_group' label but end up decrementing the reference count only
once, when we should have done it twice, since we have already removed
the block group from the block group cache rbtree. This happens because
the reference count decrement for the rbtree reference happens after
attempting to remove the free space tree entries, which is far away from
the place where we remove the block group from the rbtree.
To make things less error prone, decrement the reference count for the
rbtree immediately after removing the block group from it. This also
eleminates the need for two different exit labels on error, renaming
'out_put_label' to just 'out' and removing the old 'out'.
This patch reverts the main part of switching direct io implementation
to iomap infrastructure. There's a problem in invalidate page that
couldn't be solved as regression in this development cycle.
The problem occurs when buffered and direct io are mixed, and the ranges
overlap. Although this is not recommended, filesystems implement
measures or fallbacks to make it somehow work. In this case, fallback to
buffered IO would be an option for btrfs (this already happens when
direct io is done on compressed data), but the change would be needed in
the iomap code, bringing new semantics to other filesystems.
Another problem arises when again the buffered and direct ios are mixed,
invalidation fails, then -EIO is set on the mapping and fsync will fail,
though there's no real error.
There have been discussions how to fix that, but revert seems to be the
least intrusive option.
Filipe Manana [Wed, 27 May 2020 10:16:19 +0000 (11:16 +0100)]
btrfs: fix space_info bytes_may_use underflow during space cache writeout
We always preallocate a data extent for writing a free space cache, which
causes writeback to always try the nocow path first, since the free space
inode has the prealloc bit set in its flags.
However if the block group that contains the data extent for the space
cache has been turned to RO mode due to a running scrub or balance for
example, we have to fallback to the cow path. In that case once a new data
extent is allocated we end up calling btrfs_add_reserved_bytes(), which
decrements the counter named bytes_may_use from the data space_info object
with the expection that this counter was previously incremented with the
same amount (the size of the data extent).
However when we started writeout of the space cache at cache_save_setup(),
we incremented the value of the bytes_may_use counter through a call to
btrfs_check_data_free_space() and then decremented it through a call to
btrfs_prealloc_file_range_trans() immediately after. So when starting the
writeback if we fallback to cow mode we have to increment the counter
bytes_may_use of the data space_info again to compensate for the extent
allocation done by the cow path.
When this issue happens we are incorrectly decrementing the bytes_may_use
counter and when its current value is smaller then the amount we try to
subtract we end up with the following warning:
So fix this by incrementing the bytes_may_use counter of the data
space_info when we fallback to the cow path. If the cow path is successful
the counter is decremented after extent allocation (by
btrfs_add_reserved_bytes()), if it fails it ends up being decremented as
well when clearing the delalloc range (extent_clear_unlock_delalloc()).
This could be triggered sporadically by the test case btrfs/061 from
fstests.
Fixes: 82d5902d9c681b ("Btrfs: Support reading/writing on disk free ino cache") CC: [email protected] # 4.4+ Signed-off-by: Filipe Manana <[email protected]> Signed-off-by: David Sterba <[email protected]>
Filipe Manana [Wed, 27 May 2020 10:16:07 +0000 (11:16 +0100)]
btrfs: fix space_info bytes_may_use underflow after nocow buffered write
When doing a buffered write we always try to reserve data space for it,
even when the file has the NOCOW bit set or the write falls into a file
range covered by a prealloc extent. This is done both because it is
expensive to check if we can do a nocow write (checking if an extent is
shared through reflinks or if there's a hole in the range for example),
and because when writeback starts we might actually need to fallback to
COW mode (for example the block group containing the target extents was
turned into RO mode due to a scrub or balance).
When we are unable to reserve data space we check if we can do a nocow
write, and if we can, we proceed with dirtying the pages and setting up
the range for delalloc. In this case the bytes_may_use counter of the
data space_info object is not incremented, unlike in the case where we
are able to reserve data space (done through btrfs_check_data_free_space()
which calls btrfs_alloc_data_chunk_ondemand()).
Later when running delalloc we attempt to start writeback in nocow mode
but we might revert back to cow mode, for example because in the meanwhile
a block group was turned into RO mode by a scrub or relocation. The cow
path after successfully allocating an extent ends up calling
btrfs_add_reserved_bytes(), which expects the bytes_may_use counter of
the data space_info object to have been incremented before - but we did
not do it when the buffered write started, since there was not enough
available data space. So btrfs_add_reserved_bytes() ends up decrementing
the bytes_may_use counter anyway, and when the counter's current value
is smaller then the size of the allocated extent we get a stack trace
like the following:
So to fix this, when falling back into cow mode check if space was not
reserved, by testing for the bit EXTENT_NORESERVE in the respective file
range, and if not, increment the bytes_may_use counter for the data
space_info object. Also clear the EXTENT_NORESERVE bit from the range, so
that if the cow path fails it decrements the bytes_may_use counter when
clearing the delalloc range (through the btrfs_clear_delalloc_extent()
callback).
Fixes: 7ee9e4405f264e ("Btrfs: check if we can nocow if we don't have data space") CC: [email protected] # 4.4+ Signed-off-by: Filipe Manana <[email protected]> Signed-off-by: David Sterba <[email protected]>
Filipe Manana [Wed, 27 May 2020 10:15:53 +0000 (11:15 +0100)]
btrfs: fix wrong file range cleanup after an error filling dealloc range
If an error happens while running dellaloc in COW mode for a range, we can
end up calling extent_clear_unlock_delalloc() for a range that goes beyond
our range's end offset by 1 byte, which affects 1 extra page. This results
in clearing bits and doing page operations (such as a page unlock) outside
our target range.
Fix that by calling extent_clear_unlock_delalloc() with an inclusive end
offset, instead of an exclusive end offset, at cow_file_range().
Fixes: a315e68f6e8b30 ("Btrfs: fix invalid attempt to free reserved space on failure to cow range") CC: [email protected] # 4.14+ Signed-off-by: Filipe Manana <[email protected]> Signed-off-by: David Sterba <[email protected]>
Nikolay Borisov [Wed, 27 May 2020 10:10:59 +0000 (13:10 +0300)]
btrfs: remove redundant local variable in read_block_for_search
The local 'b' variable is only used to directly read values from passed
extent buffer. So eliminate it and directly use the input parameter.
Furthermore this shrinks the size of the following functions:
./scripts/bloat-o-meter ctree.orig fs/btrfs/ctree.o
add/remove: 0/0 grow/shrink: 0/2 up/down: 0/-73 (-73)
Function old new delta
read_block_for_search.isra 876 871 -5
push_node_left 1112 1044 -68
Total: Before=50348, After=50275, chg -0.14%
Nikolay Borisov [Wed, 27 May 2020 10:10:53 +0000 (13:10 +0300)]
btrfs: open code key_search
This function wraps the optimisation implemented by d7396f07358a
("Btrfs: optimize key searches in btrfs_search_slot") however this
optimisation is really used in only one place - btrfs_search_slot.
Just open code the optimisation and also add a comment explaining how it
works since it's not clear just by looking at the code - the key point
here is it depends on an internal invariant that BTRFS' btree provides,
namely intermediate pointers always contain the key at slot0 at the
child node. So in the case of exact match we can safely assume that the
given key will always be in slot 0 on lower levels.
Furthermore this results in a reduction of btrfs_search_slot's size:
./scripts/bloat-o-meter ctree.orig fs/btrfs/ctree.o
add/remove: 0/0 grow/shrink: 0/1 up/down: 0/-75 (-75)
Function old new delta
btrfs_search_slot 2783 2708 -75
Total: Before=50423, After=50348, chg -0.15%
btrfs: split btrfs_direct_IO to read and write part
The read and write versions don't have anything in common except for the
call to iomap_dio_rw. So split this function, and merge each half into
its only caller.
Switch from __blockdev_direct_IO() to iomap_dio_rw().
Rename btrfs_get_blocks_direct() to btrfs_dio_iomap_begin() and use it
as iomap_begin() for iomap direct I/O functions. This function
allocates and locks all the blocks required for the I/O.
btrfs_submit_direct() is used as the submit_io() hook for direct I/O
ops.
Since we need direct I/O reads to go through iomap_dio_rw(), we change
file_operations.read_iter() to a btrfs_file_read_iter() which calls
btrfs_direct_IO() for direct reads and falls back to
generic_file_buffered_read() for incomplete reads and buffered reads.
We don't need address_space.direct_IO() anymore so set it to noop.
Similarly, we don't need flags used in __blockdev_direct_IO(). iomap is
capable of direct I/O reads from a hole, so we don't need to return
-ENOENT.
BTRFS direct I/O is now done under i_rwsem, shared in case of reads and
exclusive in case of writes. This guards against simultaneous truncates.
Use iomap->iomap_end() to check for failed or incomplete direct I/O:
- for writes, call __endio_write_update_ordered()
- for reads, unlock extents
btrfs_dio_data is now hooked in iomap->private and not
current->journal_info. It carries the reservation variable and the
amount of data submitted, so we can calculate the amount of data to call
__endio_write_update_ordered in case of an error.
This patch removes last use of struct buffer_head from btrfs.
Filesystems such as btrfs can perform direct I/O without holding the
inode->i_rwsem in some of the cases like writing within i_size. So,
remove the check for lockdep_assert_held() in iomap_dio_rw().
iomap: add a filesystem hook for direct I/O bio submission
This helps filesystems to perform tasks on the bio while submitting for
I/O. This could be post-write operations such as data CRC or data
replication for fs-handled RAID.
Filipe Manana [Mon, 18 May 2020 16:34:23 +0000 (17:34 +0100)]
btrfs: turn space cache writeout failure messages into debug messages
Since commit 1afb648e945428 ("btrfs: use standard debug config option to
enable free-space-cache debug prints"), we started to log error messages
that were never logged before since there was no DEBUG macro defined
anywhere. This started to make test case btrfs/187 to fail very often,
as it greps for any btrfs error messages in dmesg/syslog and fails if
any is found:
(...)
btrfs/186 1s ... 2s
btrfs/187 - output mismatch (see .../results//btrfs/187.out.bad)
\--- tests/btrfs/187.out 2019-05-17 12:48:32.537340749 +0100
\+++ /home/fdmanana/git/hub/xfstests/results//btrfs/187.out.bad ...
\@@ -1,3 +1,8 @@
QA output created by 187
Create a readonly snapshot of 'SCRATCH_MNT' in 'SCRATCH_MNT/snap1'
Create a readonly snapshot of 'SCRATCH_MNT' in 'SCRATCH_MNT/snap2'
+[268364.139958] BTRFS error (device sdc): failed to write free space cache for block group 30408704
+[268380.156503] BTRFS error (device sdc): failed to write free space cache for block group 30408704
+[268380.161703] BTRFS error (device sdc): failed to write free space cache for block group 30408704
+[268380.253180] BTRFS error (device sdc): failed to write free space cache for block group 30408704
...
(Run 'diff -u /home/fdmanana/git/hub/xfstests/tests/btrfs/187.out ...
btrfs/188 4s ... 2s
(...)
The space cache write failures happen due to ENOSPC when attempting to
update the free space cache items in the root tree. This happens because
when starting or joining a transaction we don't know how many block
groups we will end up changing (due to extent allocation or release) and
therefore never reserve space for updating free space cache items.
More often than not, the free space cache writeout succeeds since the
metadata space info is not yet full nor very close to being full, but
when it is, the space cache writeout fails with ENOSPC.
Occasional failures to write space caches are not considered critical
since they can be rebuilt when mounting the filesystem or the next
attempt to write a free space cache in the next transaction commit might
succeed, so we used to hide those error messages with a preprocessor
check for the existence of the DEBUG macro that was never enabled
anywhere.
A few other generic test cases also trigger the error messages due to
ENOSPC failure when writing free space caches as well, however they don't
fail since they don't grep dmesg/syslog for any btrfs specific error
messages.
So change the messages from 'error' level to 'debug' level, as it doesn't
make much sense to have error messages triggered only if the debug macro
is enabled plus, more importantly, the error is not serious nor highly
unexpected.
Filipe Manana [Mon, 18 May 2020 16:34:11 +0000 (17:34 +0100)]
btrfs: include error on messages about failure to write space/inode caches
Currently the error messages logged when we fail to write a free space
cache or an inode cache are not very useful as they don't mention what
was the error. So include the error number in the messages.
Filipe Manana [Mon, 18 May 2020 11:15:09 +0000 (12:15 +0100)]
btrfs: do not ignore error from btrfs_next_leaf() when inserting checksums
We are currently treating any non-zero return value from btrfs_next_leaf()
the same way, by going to the code that inserts a new checksum item in the
tree. However if btrfs_next_leaf() returns an error (a value < 0), we
should just stop and return the error, and not behave as if nothing has
happened, since in that case we do not have a way to know if there is a
next leaf or we are currently at the last leaf already.
So fix that by returning the error from btrfs_next_leaf().
Filipe Manana [Mon, 18 May 2020 11:15:00 +0000 (12:15 +0100)]
btrfs: make checksum item extension more efficient
When we want to add checksums into the checksums tree, or a log tree, we
try whenever possible to extend existing checksum items, as this helps
reduce amount of metadata space used, since adding a new item uses extra
metadata space for a btrfs_item structure (25 bytes).
However we have two inefficiencies in the current approach:
1) After finding a checksum item that covers a range with an end offset
that matches the start offset of the checksum range we want to insert,
we release the search path populated by btrfs_lookup_csum() and then
do another COW search on tree with the goal of getting additional
space for at least one checksum. Doing this path release and then
searching again is a waste of time because very often the leaf already
has enough free space for at least one more checksum;
2) After the COW search that guarantees we get free space in the leaf for
at least one more checksum, we end up not doing the extension of the
previous checksum item, and fallback to insertion of a new checksum
item, if the leaf doesn't have an amount of free space larger then the
space required for 2 checksums plus one btrfs_item structure - this is
pointless for two reasons:
a) We want to extend an existing item, so we don't need to account for
a btrfs_item structure (25 bytes);
b) We made the COW search with an insertion size for 1 single checksum,
so if the leaf ends up with a free space amount smaller then 2
checksums plus the size of a btrfs_item structure, we give up on the
extension of the existing item and jump to the 'insert' label, where
we end up releasing the path and then doing yet another search to
insert a new checksum item for a single checksum.
Fix these inefficiencies by doing the following:
- For case 1), before releasing the path just check if the leaf already
has enough space for at least 1 more checksum, and if it does, jump
directly to the item extension code, with releasing our current path,
which was already COWed by btrfs_lookup_csum();
- For case 2), fix the logic so that for item extension we require only
that the leaf has enough free space for 1 checksum, and not a minimum
of 2 checksums plus space for a btrfs_item structure.
Filipe Manana [Mon, 18 May 2020 11:14:50 +0000 (12:14 +0100)]
btrfs: fix corrupt log due to concurrent fsync of inodes with shared extents
When we have extents shared amongst different inodes in the same subvolume,
if we fsync them in parallel we can end up with checksum items in the log
tree that represent ranges which overlap.
For example, consider we have inodes A and B, both sharing an extent that
covers the logical range from X to X + 64KiB:
1) Task A starts an fsync on inode A;
2) Task B starts an fsync on inode B;
3) Task A calls btrfs_csum_file_blocks(), and the first search in the
log tree, through btrfs_lookup_csum(), returns -EFBIG because it
finds an existing checksum item that covers the range from X - 64KiB
to X;
4) Task A checks that the checksum item has not reached the maximum
possible size (MAX_CSUM_ITEMS) and then releases the search path
before it does another path search for insertion (through a direct
call to btrfs_search_slot());
5) As soon as task A releases the path and before it does the search
for insertion, task B calls btrfs_csum_file_blocks() and gets -EFBIG
too, because there is an existing checksum item that has an end
offset that matches the start offset (X) of the checksum range we want
to log;
6) Task B releases the path;
7) Task A does the path search for insertion (through btrfs_search_slot())
and then verifies that the checksum item that ends at offset X still
exists and extends its size to insert the checksums for the range from
X to X + 64KiB;
8) Task A releases the path and returns from btrfs_csum_file_blocks(),
having inserted the checksums into an existing checksum item that got
its size extended. At this point we have one checksum item in the log
tree that covers the logical range from X - 64KiB to X + 64KiB;
9) Task B now does a search for insertion using btrfs_search_slot() too,
but it finds that the previous checksum item no longer ends at the
offset X, it now ends at an of offset X + 64KiB, so it leaves that item
untouched.
Then it releases the path and calls btrfs_insert_empty_item()
that inserts a checksum item with a key offset corresponding to X and
a size for inserting a single checksum (4 bytes in case of crc32c).
Subsequent iterations end up extending this new checksum item so that
it contains the checksums for the range from X to X + 64KiB.
So after task B returns from btrfs_csum_file_blocks() we end up with
two checksum items in the log tree that have overlapping ranges, one
for the range from X - 64KiB to X + 64KiB, and another for the range
from X to X + 64KiB.
Having checksum items that represent ranges which overlap, regardless of
being in the log tree or in the chekcsums tree, can lead to problems where
checksums for a file range end up not being found. This type of problem
has happened a few times in the past and the following commits fixed them
and explain in detail why having checksum items with overlapping ranges is
problematic:
27b9a8122ff71a "Btrfs: fix csum tree corruption, duplicate and outdated checksums" b84b8390d6009c "Btrfs: fix file read corruption after extent cloning and fsync" 40e046acbd2f36 "Btrfs: fix missing data checksums after replaying a log tree"
Since this specific instance of the problem can only happen when logging
inodes, because it is the only case where concurrent attempts to insert
checksums for the same range can happen, fix the issue by using an extent
io tree as a range lock to serialize checksum insertion during inode
logging.
This issue could often be reproduced by the test case generic/457 from
fstests. When it happens it produces the following trace:
David Sterba [Fri, 15 May 2020 17:35:59 +0000 (19:35 +0200)]
btrfs: simplify iget helpers
The inode lookup starting at btrfs_iget takes the full location key,
while only the objectid is used to match the inode, because the lookup
happens inside the given root thus the inode number is unique.
The entire location key is properly set up in btrfs_init_locked_inode.
Simplify the helpers and pass only inode number, renaming it to 'ino'
instead of 'objectid'. This allows to remove temporary variables key,
saving some stack space.
David Sterba [Fri, 15 May 2020 17:35:55 +0000 (19:35 +0200)]
btrfs: simplify root lookup by id
The main function to lookup a root by its id btrfs_get_fs_root takes the
whole key, while only using the objectid. The value of offset is preset
to (u64)-1 but not actually used until btrfs_find_root that does the
actual search.
Switch btrfs_get_fs_root to use only objectid and remove all local
variables that existed just for the lookup. The actual key for search is
set up in btrfs_get_fs_root, reusing another key variable.
Qu Wenruo [Wed, 20 May 2020 06:58:51 +0000 (14:58 +0800)]
btrfs: reloc: clear DEAD_RELOC_TREE bit for orphan roots to prevent runaway balance
[BUG]
There are several reported runaway balance, that balance is flooding the
log with "found X extents" where the X never changes.
[CAUSE]
Commit d2311e698578 ("btrfs: relocation: Delay reloc tree deletion after
merge_reloc_roots") introduced BTRFS_ROOT_DEAD_RELOC_TREE bit to
indicate that one subvolume has finished its tree blocks swap with its
reloc tree.
However if balance is canceled or hits ENOSPC halfway, we didn't clear
the BTRFS_ROOT_DEAD_RELOC_TREE bit, leaving that bit hanging forever
until unmount.
Any subvolume root with that bit, would cause backref cache to skip this
tree block, as it has finished its tree block swap. This would cause
all tree blocks of that root be ignored by balance, leading to runaway
balance.
[FIX]
Fix the problem by also clearing the BTRFS_ROOT_DEAD_RELOC_TREE bit for
the original subvolume of orphan reloc root.
Add an umount check for the stale bit still set.
Fixes: d2311e698578 ("btrfs: relocation: Delay reloc tree deletion after merge_reloc_roots") Signed-off-by: Qu Wenruo <[email protected]> Signed-off-by: David Sterba <[email protected]>
Robbie Ko [Thu, 14 May 2020 09:19:18 +0000 (17:19 +0800)]
btrfs: reduce lock contention when creating snapshot
When creating a snapshot, ordered extents need to be flushed and this
can take a long time.
In create_snapshot there are two locks held when this happens:
1. Destination directory inode lock
2. Global subvolume semaphore
This will unnecessarily block other operations like subvolume destroy,
create, or setflag until the snapshot is created.
We can fix that by moving the flush outside the locked section as this
does not depend on the aforementioned locks. The code factors out the
snapshot related work from create_snapshot to btrfs_mksnapshot.
Qu Wenruo [Fri, 15 May 2020 06:01:42 +0000 (14:01 +0800)]
btrfs: don't set SHAREABLE flag for data reloc tree
SHAREABLE flag is set for subvolumes because users can create snapshot
for subvolumes, thus sharing tree blocks of them.
But data reloc tree is not exposed to user space, as it's only an
internal tree for data relocation, thus it doesn't need the full path
replacement handling at all.
This patch will make data reloc tree a non-shareable tree, and add
btrfs_fs_info::data_reloc_root for data reloc tree, so relocation code
can grab it from fs_info directly.
This would slightly improve tree relocation, as now data reloc tree
can go through regular COW routine to get relocated, without bothering
the complex tree reloc tree routine.
Qu Wenruo [Fri, 15 May 2020 06:01:41 +0000 (14:01 +0800)]
btrfs: inode: cleanup the log-tree exceptions in btrfs_truncate_inode_items()
There are a lot of root owner checks in btrfs_truncate_inode_items()
like:
if (test_bit(BTRFS_ROOT_SHAREABLE, &root->state) ||
root == fs_info->tree_root)
But considering that, only these trees can have INODE_ITEMs:
- tree root (for v1 space cache)
- subvolume trees
- tree reloc trees
- data reloc tree
- log trees
And since subvolume/tree reloc/data reloc trees all have SHAREABLE bit,
and we're checking tree root manually, so above check is just excluding
log trees.
This patch will replace two of such checks to a simpler one:
if (root->root_key.objectid != BTRFS_TREE_LOG_OBJECTID)
This would merge btrfs_drop_extent_cache() and lock_extent_bits() call
into the same if branch.
Qu Wenruo [Fri, 15 May 2020 06:01:40 +0000 (14:01 +0800)]
btrfs: rename BTRFS_ROOT_REF_COWS to BTRFS_ROOT_SHAREABLE
The name BTRFS_ROOT_REF_COWS is not very clear about the meaning.
In fact, that bit can only be set to those trees:
- Subvolume roots
- Data reloc root
- Reloc roots for above roots
All other trees won't get this bit set. So just by the result, it is
obvious that, roots with this bit set can have tree blocks shared with
other trees. Either shared by snapshots, or by reloc roots (an special
snapshot created by relocation).
This patch will rename BTRFS_ROOT_REF_COWS to BTRFS_ROOT_SHAREABLE to
make it easier to understand, and update all comment mentioning
"reference counted" to follow the rename.
David Sterba [Thu, 30 Apr 2020 15:57:55 +0000 (17:57 +0200)]
btrfs: optimize split page write in btrfs_set_token_##bits
The fallback path calls helper write_extent_buffer to do write of the
data spanning two extent buffer pages. As the size is known, we can do
the write directly in two steps. This removes one function call and
compiler can optimize memcpy as the sizes are known at compile time. The
cached token address is set to the second page.
David Sterba [Thu, 30 Apr 2020 15:57:55 +0000 (17:57 +0200)]
btrfs: optimize split page write in btrfs_set_##bits
The helper write_extent_buffer is called to do write of the data
spanning two extent buffer pages. As the size is known, we can do the
write directly in two steps. This removes one function call and
compiler can optimize memcpy as the sizes are known at compile time.
David Sterba [Thu, 30 Apr 2020 15:57:55 +0000 (17:57 +0200)]
btrfs: optimize split page read in btrfs_get_token_##bits
The fallback path calls helper read_extent_buffer to do read of the data
spanning two extent buffer pages. As the size is known, we can do the
read directly in two steps. This removes one function call and compiler
can optimize memcpy as the sizes are known at compile time. The cached
token address is set to the second page.
David Sterba [Thu, 30 Apr 2020 15:57:55 +0000 (17:57 +0200)]
btrfs: optimize split page read in btrfs_get_##bits
The helper read_extent_buffer is called to do read of the data spanning
two extent buffer pages. As the size is known, we can do the read
directly in two steps. This removes one function call and compiler can
optimize memcpy as the sizes are known at compile time.
David Sterba [Wed, 29 Apr 2020 21:41:20 +0000 (23:41 +0200)]
btrfs: drop unnecessary offset_in_page in extent buffer helpers
Helpers that iterate over extent buffer pages set up several variables,
one of them is finding out offset of the extent buffer start within a
page. Right now we have extent buffers aligned to page sizes so this is
effectively storing zero. This makes the code harder the follow and can
be simplified.
David Sterba [Wed, 29 Apr 2020 01:04:10 +0000 (03:04 +0200)]
btrfs: constify extent_buffer in the API functions
There are many helpers around extent buffers, found in extent_io.h and
ctree.h. Most of them can be converted to take constified eb as there
are no changes to the extent buffer structure itself but rather the
pages.
David Sterba [Wed, 29 Apr 2020 21:36:03 +0000 (23:36 +0200)]
btrfs: remove unused map_private_extent_buffer
All uses of map_private_extent_buffer have been replaced by more
effective way. The set/get helpers have their own bounds checker.
The function name was confusing since the non-private helper was removed
in a65917156e34 ("Btrfs: stop using highmem for extent_buffers") many
years ago.
David Sterba [Wed, 29 Apr 2020 21:23:37 +0000 (23:23 +0200)]
btrfs: speed up and simplify generic_bin_search
The bin search jumps over the extent buffer item keys, comparing
directly the bytes if the key is in one page, or storing it in a
temporary buffer in case it spans two pages.
The mapping start and length are obtained from map_private_extent_buffer,
which is heavy weight compared to what we need. We know the key size and
can find out the eb page in a simple way. For keys spanning two pages
the fallback read_extent_buffer is used.
The temporary variables are reduced and moved to the scope of use.
David Sterba [Wed, 29 Apr 2020 16:23:37 +0000 (18:23 +0200)]
btrfs: speed up btrfs_set_token_##bits helpers
The set/get token helpers either use the cached address in the token or
unconditionally call map_private_extent_buffer to get the address of
page containing the requested offset plus the mapping start and length.
Depending on the return value, the fast path uses unaligned put to write
data within a page, or fall back to write_extent_buffer that can handle
writes spanning more pages.
This is all wasteful. We know the number of bytes to write, 1/2/4/8 and
can find out the page. Then simply check if it's contained or the
fallback is needed. The token address is updated to the page, or the on
the next index, expecting that the next write will use that.
This saves one function call to map_private_extent_buffer and several
unnecessary temporary variables.
David Sterba [Wed, 29 Apr 2020 16:07:04 +0000 (18:07 +0200)]
btrfs: speed up btrfs_set_##bits helpers
The helpers unconditionally call map_private_extent_buffer to get the
address of page containing the requested offset plus the mapping start
and length. Depending on the return value, the fast path uses unaligned
put to write data within a page, or fall back to write_extent_buffer
that can handle writes spanning more pages.
This is all wasteful. We know the number of bytes to write, 1/2/4/8 and
can find out the page. Then simply check if it's contained or the
fallback is needed.
This saves one function call to map_private_extent_buffer and several
unnecessary temporary variables.
David Sterba [Wed, 29 Apr 2020 15:45:33 +0000 (17:45 +0200)]
btrfs: speed up btrfs_get_token_##bits helpers
The set/get token helpers either use the cached address in the token or
unconditionally call map_private_extent_buffer to get the address of
page containing the requested offset plus the mapping start and length.
Depending on the return value, the fast path uses unaligned read to get
data within a page, or fall back to read_extent_buffer that can handle
reads spanning more pages.
This is all wasteful. We know the number of bytes to read, 1/2/4/8 and
can find out the page. Then simply check if it's contained or the
fallback is needed. The token address is updated to the page, or the on
the next index, expecting that the next read will use that.
This saves one function call to map_private_extent_buffer and several
unnecessary temporary variables.
David Sterba [Wed, 29 Apr 2020 14:04:44 +0000 (16:04 +0200)]
btrfs: speed up btrfs_get_##bits helpers
The helpers unconditionally call map_private_extent_buffer to get the
address of page containing the requested offset plus the mapping start
and length. Depending on the return value, the fast path uses unaligned
read to get data within a page, or fall back to read_extent_buffer that
can handle reads spanning more pages.
This is all wasteful. We know the number of bytes to read, 1/2/4/8 and
can find out the page. Then simply check if it's contained or the
fallback is needed.
This saves one function call to map_private_extent_buffer and several
unnecessary temporary variables.
David Sterba [Thu, 30 Apr 2020 21:38:11 +0000 (23:38 +0200)]
btrfs: add separate bounds checker for set/get helpers
The bounds checking is now done in map_private_extent_buffer but that
will be removed in following patches and some sanity checks should still
be done.
There are two separate checks to see the kind of out of bounds access:
partial (start offset is in the buffer) or complete (both start and end
are out).
David Sterba [Wed, 29 Apr 2020 17:29:04 +0000 (19:29 +0200)]
btrfs: preset set/get token with first page and drop condition
All the set/get helpers first check if the token contains a cached
address. After first use the address is always valid, but the extra
check is done for each call.
The token initialization can optimistically set it to the first extent
buffer page, that we know always exists. Then the condition in all
btrfs_token_*/btrfs_set_token_* can be simplified by removing the
address check from the condition, but for development the assertion
still makes sure it's valid.
David Sterba [Wed, 29 Apr 2020 20:56:01 +0000 (22:56 +0200)]
btrfs: don't use set/get token in leaf_space_used
The token is supposed to cache the last page used by the set/get
helpers. In leaf_space_used the first and last items are accessed, it's
not likely they'd be on the same page so there's some overhead caused
updating the token address but not using it.
David Sterba [Wed, 29 Apr 2020 13:29:53 +0000 (15:29 +0200)]
btrfs: don't use set/get token for single assignment in overwrite_item
The set/get token is supposed to cache the last page that was accessed
so it speeds up subsequential access to the eb. It does not make sense
to use that for just one change, which is the case of inode size in
overwrite_item.
David Sterba [Wed, 29 Apr 2020 00:15:56 +0000 (02:15 +0200)]
btrfs: drop eb parameter from set/get token helpers
Now that all set/get helpers use the eb from the token, we don't need to
pass it to many btrfs_token_*/btrfs_set_token_* helpers, saving some
stack space.
David Sterba [Wed, 29 Apr 2020 00:13:57 +0000 (02:13 +0200)]
btrfs: use the token::eb for all set/get helpers
The token stores a copy of the extent buffer pointer but does not make
any use of it besides sanity checks. We can use it and drop the eb
parameter from several functions, this patch only switches the use
inside the set/get helpers.
Qu Wenruo [Mon, 4 May 2020 23:58:22 +0000 (07:58 +0800)]
btrfs: block-group: refactor how we insert a block group item
Currently the block group item insert is pretty straight forward, fill
the block group item structure and insert it into extent tree.
However the incoming skinny block group feature is going to change this,
so this patch will refactor insertion into a new function,
insert_block_group_item(), to make the incoming feature easier to add.
Qu Wenruo [Mon, 4 May 2020 23:58:21 +0000 (07:58 +0800)]
btrfs: block-group: refactor how we delete one block group item
When deleting a block group item, it's pretty straight forward, just
delete the item pointed by the key. However it will not be that
straight-forward for incoming skinny block group item.
So refactor the block group item deletion into a new function,
remove_block_group_item(), also to make the already lengthy
btrfs_remove_block_group() a little shorter.
Qu Wenruo [Mon, 4 May 2020 23:58:20 +0000 (07:58 +0800)]
btrfs: block-group: refactor how we read one block group item
Structure btrfs_block_group has the following members which are
currently read from on-disk block group item and key:
- length - from item key
- used
- flags - from block group item
However for incoming skinny block group tree, we are going to read those
members from different sources.
This patch will refactor such read by:
- Don't initialize btrfs_block_group::length at allocation
Caller should initialize them manually.
Also to avoid possible (well, only two callers) missing
initialization, add extra ASSERT() in btrfs_add_block_group_cache().
- Refactor length/used/flags initialization into one function
The new function, fill_one_block_group() will handle the
initialization of such members.
- Use btrfs_block_group::length to replace key::offset
Since skinny block group item would have a different meaning for its
key offset.
Whenever a chown is executed, all capabilities of the file being touched
are lost. When doing incremental send with a file with capabilities,
there is a situation where the capability can be lost on the receiving
side. The sequence of actions bellow shows the problem:
At this point, only a chown was emitted by "btrfs send" since only the
group was changed. This makes the cap_sys_nice capability to be dropped
from fs2/snap_incremental/foo.bar
To fix that, only emit capabilities after chown is emitted. The current
code first checks for xattrs that are new/changed, emits them, and later
emit the chown. Now, __process_new_xattr skips capabilities, letting
only finish_inode_if_needed to emit them, if they exist, for the inode
being processed.
This behavior was being worked around in "btrfs receive" side by caching
the capability and only applying it after chown. Now, xattrs are only
emmited _after_ chown, making that workaround not needed anymore.
Filipe Manana [Fri, 8 May 2020 10:02:07 +0000 (11:02 +0100)]
btrfs: scrub, only lookup for csums if we are dealing with a data extent
When scrubbing a stripe, whenever we find an extent we lookup for its
checksums in the checksum tree. However we do it even for metadata extents
which don't have checksum items stored in the checksum tree, that is
only for data extents.
So make the lookup for checksums only if we are processing with a data
extent.
Filipe Manana [Fri, 8 May 2020 10:01:59 +0000 (11:01 +0100)]
btrfs: move the block group freeze/unfreeze helpers into block-group.c
The helpers btrfs_freeze_block_group() and btrfs_unfreeze_block_group()
used to be named btrfs_get_block_group_trimming() and
btrfs_put_block_group_trimming() respectively.
At the time they were added to free-space-cache.c, by commit e33e17ee1098
("btrfs: add missing discards when unpinning extents with -o discard")
because all the trimming related functions were in free-space-cache.c.
Now that the helpers were renamed and are used in scrub context as well,
move them to block-group.c, a much more logical location for them.
Filipe Manana [Fri, 8 May 2020 10:01:47 +0000 (11:01 +0100)]
btrfs: rename member 'trimming' of block group to a more generic name
Back in 2014, commit 04216820fe83d5 ("Btrfs: fix race between fs trimming
and block group remove/allocation"), I added the 'trimming' member to the
block group structure. Its purpose was to prevent races between trimming
and block group deletion/allocation by pinning the block group in a way
that prevents its logical address and device extents from being reused
while trimming is in progress for a block group, so that if another task
deletes the block group and then another task allocates a new block group
that gets the same logical address and device extents while the trimming
task is still in progress.
After the previous fix for scrub (patch "btrfs: fix a race between scrub
and block group removal/allocation"), scrub now also has the same needs that
trimming has, so the member name 'trimming' no longer makes sense.
Since there is already a 'pinned' member in the block group that refers
to space reservations (pinned bytes), rename the member to 'frozen',
add a comment on top of it to describe its general purpose and rename
the helpers to increment and decrement the counter as well, to match
the new member name.
The next patch in the series will move the helpers into a more suitable
file (from free-space-cache.c to block-group.c).
Filipe Manana [Fri, 8 May 2020 10:01:10 +0000 (11:01 +0100)]
btrfs: fix a race between scrub and block group removal/allocation
When scrub is verifying the extents of a block group for a device, it is
possible that the corresponding block group gets removed and its logical
address and device extents get used for a new block group allocation.
When this happens scrub incorrectly reports that errors were detected
and, if the the new block group has a different profile then the old one,
deleted block group, we can crash due to a null pointer dereference.
Possibly other unexpected and weird consequences can happen as well.
Consider the following sequence of actions that leads to the null pointer
dereference crash when scrub is running in parallel with balance:
1) Balance sets block group X to read-only mode and starts relocating it.
Block group X is a metadata block group, has a raid1 profile (two
device extents, each one in a different device) and a logical address
of 19424870400;
2) Scrub is running and finds device extent E, which belongs to block
group X. It enters scrub_stripe() to find all extents allocated to
block group X, the search is done using the extent tree;
3) Balance finishes relocating block group X and removes block group X;
4) Balance starts relocating another block group and when trying to
commit the current transaction as part of the preparation step
(prepare_to_relocate()), it blocks because scrub is running;
5) The scrub task finds the metadata extent at the logical address 19425001472 and marks the pages of the extent to be read by a bio
(struct scrub_bio). The extent item's flags, which have the bit
BTRFS_EXTENT_FLAG_TREE_BLOCK set, are added to each page (struct
scrub_page). It is these flags in the scrub pages that tells the
bio's end io function (scrub_bio_end_io_worker) which type of extent
it is dealing with. At this point we end up with 4 pages in a bio
which is ready for submission (the metadata extent has a size of
16Kb, so that gives 4 pages on x86);
6) At the next iteration of scrub_stripe(), scrub checks that there is a
pause request from the relocation task trying to commit a transaction,
therefore it submits the pending bio and pauses, waiting for the
transaction commit to complete before resuming;
7) The relocation task commits the transaction. The device extent E, that
was used by our block group X, is now available for allocation, since
the commit root for the device tree was swapped by the transaction
commit;
8) Another task doing a direct IO write allocates a new data block group Y
which ends using device extent E. This new block group Y also ends up
getting the same logical address that block group X had: 19424870400.
This happens because block group X was the block group with the highest
logical address and, when allocating Y, find_next_chunk() returns the
end offset of the current last block group to be used as the logical
address for the new block group, which is
So our new block group Y has the same logical address and device extent
that block group X had. However Y is a data block group, while X was
a metadata one, and Y has a raid0 profile, while X had a raid1 profile;
9) After allocating block group Y, the direct IO submits a bio to write
to device extent E;
10) The read bio submitted by scrub reads the 4 pages (16Kb) from device
extent E, which now correspond to the data written by the task that
did a direct IO write. Then at the end io function associated with
the bio, scrub_bio_end_io_worker(), we call scrub_block_complete()
which calls scrub_checksum(). This later function checks the flags
of the first page, and sees that the bit BTRFS_EXTENT_FLAG_TREE_BLOCK
is set in the flags, so it assumes it has a metadata extent and
then calls scrub_checksum_tree_block(). That functions returns an
error, since interpreting data as a metadata extent causes the
checksum verification to fail.
So this makes scrub_checksum() call scrub_handle_errored_block(),
which determines 'failed_mirror_index' to be 1, since the device
extent E was allocated as the second mirror of block group X.
It allocates BTRFS_MAX_MIRRORS scrub_block structures as an array at
'sblocks_for_recheck', and all the memory is initialized to zeroes by
kcalloc().
After that it calls scrub_setup_recheck_block(), which is responsible
for filling each of those structures. However, when that function
calls btrfs_map_sblock() against the logical address of the metadata
extent, 19425001472, it gets a struct btrfs_bio ('bbio') that matches
the current block group Y. However block group Y has a raid0 profile
and not a raid1 profile like X had, so the following call returns 1:
scrub_nr_raid_mirrors(bbio)
And as a result scrub_setup_recheck_block() only initializes the
first (index 0) scrub_block structure in 'sblocks_for_recheck'.
Then scrub_recheck_block() is called by scrub_handle_errored_block()
with the second (index 1) scrub_block structure as the argument,
because 'failed_mirror_index' was previously set to 1.
This scrub_block was not initialized by scrub_setup_recheck_block(),
so it has zero pages, its 'page_count' member is 0 and its 'pagev'
page array has all members pointing to NULL.
Finally when scrub_recheck_block() calls scrub_recheck_block_checksum()
we have a NULL pointer dereference when accessing the flags of the first
page, as pavev[0] is NULL:
This issue has been around for a long time, possibly since scrub exists.
The last time I ran into it was over 2 years ago. After recently fixing
fstests to pass the "--full-balance" command line option to btrfs-progs
when doing balance, several tests started to more heavily exercise balance
with fsstress, scrub and other operations in parallel, and therefore
started to hit this issue again (with btrfs/061 for example).
Fix this by having scrub increment the 'trimming' counter of the block
group, which pins the block group in such a way that it guarantees neither
its logical address nor device extents can be reused by future block group
allocations until we decrement the 'trimming' counter. Also make sure that
on each iteration of scrub_stripe() we stop scrubbing the block group if
it was removed already.
A later patch in the series will rename the block group's 'trimming'
counter and its helpers to a more generic name, since now it is not used
exclusively for pinning while trimming anymore.
David Sterba [Mon, 11 May 2020 12:49:10 +0000 (14:49 +0200)]
btrfs: remove more obsolete v0 extent ref declarations
The extent references v0 have been superseded long time go, there are
some unused declarations of access helpers. We can safely remove them
now. The struct btrfs_extent_ref_v0 is not used anywhere, but struct
btrfs_extent_item_v0 is still part of a backward compatibility check in
relocation.c and thus not removed.
[CAUSE]
The problem is in btrfs_qgroup_inherit(), we don't have good enough
check to determine if the new relation would break the existing
accounting.
Unlike btrfs_add_qgroup_relation(), which has proper check to determine
if we can do quick update without a rescan, in btrfs_qgroup_inherit() we
can even assign a snapshot to multiple qgroups.
[FIX]
Fix it by manually marking qgroup inconsistent for snapshot inheritance.
For subvolume creation, since all its extents are exclusively owned, we
don't need to rescan.
In theory, we should call relation check like quick_update_accounting()
when doing qgroup inheritance and inform user about qgroup accounting
inconsistency.
But we don't have good mechanism to relay that back to the user in the
snapshot creation context, thus we can only silently mark the qgroup
inconsistent.
Anyway, user shouldn't use qgroup inheritance during snapshot creation,
and should add qgroup relationship after snapshot creation by 'btrfs
qgroup assign', which has a much better UI to inform user about qgroup
inconsistent and kick in rescan automatically.
Robbie Ko [Thu, 7 May 2020 02:54:40 +0000 (10:54 +0800)]
btrfs: speedup dead root detection during orphan cleanup
When mounting, we handle deleted subvolume and orphan items. First,
find add orphan roots, then add them to fs_root radix tree. Second, in
tree-root, process each orphan item, skip if it is dead root.
The original algorithm is based on the list of dead_roots, one by one to
visit and check whether the objectid is consistent, the time complexity
is O (n ^ 2). When processing 50000 deleted subvols, it takes about
120s.
Because btrfs_find_orphan_roots has already ran before us, and added
deleted subvol to fs_roots radix tree.
The fs root will only be removed from the fs_roots radix tree after the
cleaner process is started, and the cleaner will only start execution
after the mount is complete.
btrfs_orphan_cleanup can be called during the whole filesystem mount
lifetime, but only "tree root" will be used in this section of code, and
only mount time will be brought into tree root.
So we can quickly check whether the orphan item is dead root through the
fs_roots radix tree.
David Sterba [Tue, 28 Apr 2020 15:10:29 +0000 (17:10 +0200)]
btrfs: add more codes to decoder table
I've grepped logs for 'errno=.*unknown' and found -95, -117 and -122,
now added to the table. The wording is adjusted so it makes sense in
context of filesystem.
Anand Jain [Mon, 4 May 2020 18:58:26 +0000 (02:58 +0800)]
btrfs: free alien device after device add
When an old device has new fsid through 'btrfs device add -f <dev>' our
fs_devices list has an alien device in one of the fs_devices lists.
By having an alien device in fs_devices, we have two issues so far
1. missing device does not not show as missing in the userland
2. degraded mount will fail
Both issues are caused by the fact that there's an alien device in the
fs_devices list. (Alien means that it does not belong to the filesystem,
identified by fsid, or does not contain btrfs filesystem at all, eg. due
to overwrite).
A device can be scanned/added through the control device ioctls
SCAN_DEV, DEVICES_READY or by ADD_DEV.
And device coming through the control device is checked against the all
other devices in the lists, but this was not the case for ADD_DEV.
This patch fixes both issues above by removing the alien device.
Anand Jain [Mon, 4 May 2020 18:58:25 +0000 (02:58 +0800)]
btrfs: include non-missing as a qualifier for the latest_bdev
btrfs_free_extra_devids() updates fs_devices::latest_bdev to point to
the bdev with greatest device::generation number. For a typical-missing
device the generation number is zero so fs_devices::latest_bdev will
never point to it.
But if the missing device is due to alienation [1], then
device::generation is not zero and if it is greater or equal to the rest
of device generations in the list, then fs_devices::latest_bdev ends up
pointing to the missing device and reports the error like [2].
[1] We maintain devices of a fsid (as in fs_device::fsid) in the
fs_devices::devices list, a device is considered as an alien device
if its fsid does not match with the fs_device::fsid
Filipe Manana [Tue, 21 Apr 2020 10:25:31 +0000 (11:25 +0100)]
btrfs: remove useless check for copy_items() return value
At btrfs_log_prealloc_extents() we are checking if copy_items() returns a
value greater than 0. That used to happen in the past to signal the caller
that the path given to it was released and reused for other searches, but
as of commit 0e56315ca147b3 ("Btrfs: fix missing hole after hole punching
and fsync when using NO_HOLES"), the copy_items() function does not have
that behaviour anymore and always returns 0 or a negative value. So just
remove that check at btrfs_log_prealloc_extents(), which the previously
mentioned commit forgot to remove.
Currently, direct I/O has its own versions of bio_readpage_error() and
btrfs_check_repairable() (dio_read_error() and
btrfs_check_dio_repairable(), respectively). The main difference is that
the direct I/O version doesn't do read validation. The rework of direct
I/O repair makes it possible to do validation, so we can get rid of
btrfs_check_dio_repairable() and combine bio_readpage_error() and
dio_read_error() into a new helper, btrfs_submit_read_repair().
This was originally added in commit 8b110e393c5a ("Btrfs: implement
repair function when direct read fails") to avoid a deadlock. In that
commit, the direct I/O read endio executes on the endio_workers
workqueue, submits a repair bio, and waits for it to complete. The
repair bio endio must execute on a different workqueue, otherwise it
could block on the endio_workers workqueue becoming available, which
won't happen because the original endio is blocked on the repair bio.
As of the previous commit, the original endio doesn't wait for the
repair bio, so this separate workqueue is unnecessary.
Direct I/O read repair was originally implemented in commit 8b110e393c5a
("Btrfs: implement repair function when direct read fails"). This
implementation is unnecessarily complicated. There is major code
duplication between __btrfs_subio_endio_read() (checks checksums and
handles I/O errors for files with checksums),
__btrfs_correct_data_nocsum() (handles I/O errors for files without
checksums), btrfs_retry_endio() (checks checksums and handles I/O errors
for retries of files with checksums), and btrfs_retry_endio_nocsum()
(handles I/O errors for retries of files without checksum). If it sounds
like these should be one function, that's because they should.
Additionally, these functions are very hard to follow due to their
excessive use of goto.
This commit replaces the original implementation. After the previous
commit getting rid of orig_bio, we can reuse the same endio callback for
repair I/O and the original I/O, we just need to track the file offset
and original iterator in the repair bio. We can also unify the handling
of files with and without checksums and simplify the control flow. We
also no longer have to wait for each repair I/O to complete one by one.
In the worst case, there are _4_ layers of bios in the Btrfs direct I/O
path:
1. The bio created by the generic direct I/O code (dio_bio).
2. A clone of dio_bio we create in btrfs_submit_direct() to represent
the entire direct I/O range (orig_bio).
3. A partial clone of orig_bio limited to the size of a RAID stripe that
we create in btrfs_submit_direct_hook().
4. Clones of each of those split bios for each RAID stripe that we
create in btrfs_map_bio().
As of the previous commit, the second layer (orig_bio) is no longer
needed for anything: we can split dio_bio instead, and complete dio_bio
directly when all of the cloned bios complete. This lets us clean up a
bunch of cruft, including dip->subio_endio and dip->errors (we can use
dio_bio->bi_status instead). It also enables the next big cleanup of
direct I/O read repair.
projects / linux.git / log