openwrt/staging/blogic.git
5 years agobtrfs: Rename find_oldest_super_backup to init_backup_root_slot
Nikolay Borisov [Tue, 15 Oct 2019 15:42:23 +0000 (18:42 +0300)]
btrfs: Rename find_oldest_super_backup to init_backup_root_slot

The old name name was an awful misnomer because it didn't really find
the oldest super backup per-se but rather its slot. For example if we
have:

slot0: gen - 2
slot1: gen - 1
slot2: gen
slot3: empty

init_backup_root_slot will return slot3 and not slot0.

The new name is more appropriate since the function doesn't care whether
there is a valid backup in the returned slot or not.

Signed-off-by: Nikolay Borisov <nborisov@suse.com>
Reviewed-by: David Sterba <dsterba@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
5 years agobtrfs: Remove unused next_root_backup function
Nikolay Borisov [Tue, 15 Oct 2019 15:42:22 +0000 (18:42 +0300)]
btrfs: Remove unused next_root_backup function

This function has been superseded by previous commits and is no longer
used so just remove it.

Signed-off-by: Nikolay Borisov <nborisov@suse.com>
Reviewed-by: David Sterba <dsterba@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
5 years agobtrfs: Don't use objectid_mutex during mount
Nikolay Borisov [Tue, 15 Oct 2019 15:42:21 +0000 (18:42 +0300)]
btrfs: Don't use objectid_mutex during mount

Since the filesystem is not well formed and no trees are loaded it's
pointless holding the objectid_mutex. Just remove its usage.

Signed-off-by: Nikolay Borisov <nborisov@suse.com>
Reviewed-by: David Sterba <dsterba@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
5 years agobtrfs: Factor out tree roots initialization during mount
Nikolay Borisov [Tue, 15 Oct 2019 15:42:20 +0000 (18:42 +0300)]
btrfs: Factor out tree roots initialization during mount

The code responsible for reading and initializing tree roots is
scattered in open_ctree among 2 labels, emulating a loop. This is rather
confusing to reason about. Instead, factor the code to a new function,
init_tree_roots which implements the same logical flow.

There are a couple of notable differences, namely:

* Instead of using next_backup_root it's using the newly introduced
  read_backup_root.

* If read_backup_root returns an error init_tree_roots propagates the
  error and there is no special handling of that case e.g. the code jumps
  straight to 'fail_tree_roots' label. The old code, however, was
  (erroneously) jumping to 'fail_block_groups' label if next_backup_root
  did fail, this was unnecessary since the tree roots init logic doesn't
  modify the state of block groups.

Signed-off-by: Nikolay Borisov <nborisov@suse.com>
Reviewed-by: David Sterba <dsterba@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
5 years agobtrfs: Add read_backup_root
Nikolay Borisov [Tue, 15 Oct 2019 15:42:19 +0000 (18:42 +0300)]
btrfs: Add read_backup_root

This function will replace next_root_backup with a much saner/cleaner
interface.

Signed-off-by: Nikolay Borisov <nborisov@suse.com>
Reviewed-by: David Sterba <dsterba@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
5 years agobtrfs: Remove newest_gen argument from find_oldest_super_backup
Nikolay Borisov [Tue, 15 Oct 2019 15:42:18 +0000 (18:42 +0300)]
btrfs: Remove newest_gen argument from find_oldest_super_backup

It's no longer needed following cleanups around find_newest_backup_root

Signed-off-by: Nikolay Borisov <nborisov@suse.com>
Reviewed-by: David Sterba <dsterba@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
5 years agobtrfs: Cleanup and simplify find_newest_super_backup
Nikolay Borisov [Tue, 15 Oct 2019 15:42:17 +0000 (18:42 +0300)]
btrfs: Cleanup and simplify find_newest_super_backup

Backup roots are always written in a circular manner. By definition we
can only ever have 1 backup root whose generation equals to that of the
superblock. Hence, the 'if' in the for loop will trigger at most once.
This is sufficient to return the newest backup root.

Furthermore the newest_gen parameter is always set to the generation of
the superblock. This value can be obtained from the fs_info.

This patch removes the unnecessary code dealing with the wraparound
case and makes 'newest_gen' a local variable.

Signed-off-by: Nikolay Borisov <nborisov@suse.com>
Reviewed-by: David Sterba <dsterba@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
5 years agoBtrfs: remove unnecessary delalloc mutex for inodes
Filipe Manana [Fri, 25 Oct 2019 09:52:42 +0000 (10:52 +0100)]
Btrfs: remove unnecessary delalloc mutex for inodes

The inode delalloc mutex was added a long time ago by commit f248679e86fea
("Btrfs: add a delalloc mutex to inodes for delalloc reservations"), and
the reason for its introduction is not very clear from the change log. It
claims it solves bogus warnings from lockdep, however it lacks an example
report/warning from lockdep, or any explanation.

Since we have enough concurrentcy protection from the locks of the space
info and block reserve objects, and such lockdep warnings don't seem to
exist anymore (at least on a 5.3 kernel I couldn't get them with fstests,
ltp, fs_mark, etc), remove it, simplifying things a bit and decreasing
the size of the btrfs_inode structure. With some quick fio tests doing
direct IO and mmap writes I couldn't observe any significant performance
increase either (direct IO writes that don't increase the file's size
don't hold the inode's lock for their entire duration and mmap writes
don't hold the inode's lock at all), which are the only type of writes
that could see any performance gain due to less serialization.

Review feedback from Josef:

The problem was taking the i_mutex in mmap, which is how I was
protecting delalloc reservations originally.  The delalloc mutex didn't
come with all of the other dependencies.  That's what the lockdep
messages were about, removing the lock isn't going to make them appear
again.

We _had_ to lock around this because we used to do tricks to keep from
over-reserving, and if we didn't serialize delalloc reservations we'd
end up with ugly accounting problems when we tried to clean things up.

However with my recentish changes this isn't the case anymore.  Every
operation is responsible for reserving its space, and then adding it to
the inode.  Then cleaning up is straightforward and can't be mucked up
by other users.  So we no longer need the delalloc mutex to safe us from
ourselves.

Reviewed-by: Josef Bacik <josef@toxicpanda.com>
Signed-off-by: Filipe Manana <fdmanana@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
5 years agoBtrfs: remove wait queue from space_info structure
Filipe Manana [Fri, 25 Oct 2019 09:53:12 +0000 (10:53 +0100)]
Btrfs: remove wait queue from space_info structure

It is not used anymore since commit 957780eb2788d8 ("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>
5 years agobtrfs: remove cached space_info in btrfs_statfs()
Johannes Thumshirn [Thu, 24 Oct 2019 15:44:54 +0000 (17:44 +0200)]
btrfs: remove cached space_info in btrfs_statfs()

In btrfs_statfs() we cache fs_info::space_info in a local variable only
to use it once in a list_for_each_rcu() statement.

Not only is the local variable unnecessary it even makes the code harder
to follow as it's not clear which list it is iterating.

Signed-off-by: Johannes Thumshirn <jthumshirn@suse.de>
Reviewed-by: David Sterba <dsterba@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
5 years agobtrfs: add dedicated members for start and length of a block group
David Sterba [Wed, 23 Oct 2019 16:48:22 +0000 (18:48 +0200)]
btrfs: add dedicated members for start and length of a block group

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>
5 years agobtrfs: rename extent buffer block group item accessors
David Sterba [Wed, 23 Oct 2019 16:48:20 +0000 (18:48 +0200)]
btrfs: rename extent buffer block group item accessors

Accessors defined by BTRFS_SETGET_FUNCS take a raw extent buffer and
manipulate the items there, there's no special prefix required. The
block group accessors had _disk_ because previously the names were
occupied by the on-stack accessors. As this has been addressed in the
previous patch, we can now unify the naming.

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>
5 years agobtrfs: rename block_group_item on-stack accessors to follow naming
David Sterba [Wed, 23 Oct 2019 16:48:18 +0000 (18:48 +0200)]
btrfs: rename block_group_item on-stack accessors to follow naming

All accessors defined by BTRFS_SETGET_STACK_FUNCS contain _stack_ in the
name, the block group ones were not following that scheme, so let's
switch them.

Signed-off-by: David Sterba <dsterba@suse.com>
5 years agobtrfs: remove embedded block_group_cache::item
David Sterba [Wed, 23 Oct 2019 16:48:15 +0000 (18:48 +0200)]
btrfs: remove embedded block_group_cache::item

The members ::used and ::flags are now in the block group cache
structure, the last one is chunk_objectid, but that's set to a fixed
value and otherwise unused. The item is constructed from a local
variable before write, so we can remove the embedded one from block
group.

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>
5 years agobtrfs: move block_group_item::flags to block group
David Sterba [Wed, 23 Oct 2019 16:48:13 +0000 (18:48 +0200)]
btrfs: move block_group_item::flags to block group

The flags are read from the item that's embedded to block group struct,
but the item will be removed. Use the ::flags after read and before
write.

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>
5 years agobtrfs: move block_group_item::used to block group
David Sterba [Wed, 23 Oct 2019 16:48:11 +0000 (18:48 +0200)]
btrfs: move block_group_item::used to block group

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>
5 years agobtrfs: Remove btrfs_bio::flags member
Qu Wenruo [Thu, 24 Oct 2019 01:38:29 +0000 (09:38 +0800)]
btrfs: Remove btrfs_bio::flags member

The last user of btrfs_bio::flags was removed in commit 326e1dbb5736
("block: remove management of bi_remaining when restoring original
bi_end_io"), remove it.

(Tagged for stable as the structure is heavily used and space savings
are desirable.)

CC: stable@vger.kernel.org # 4.4+
Signed-off-by: Qu Wenruo <wqu@suse.com>
Reviewed-by: David Sterba <dsterba@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
5 years agobtrfs: add blake2b to checksumming algorithms
David Sterba [Mon, 7 Oct 2019 09:11:02 +0000 (11:11 +0200)]
btrfs: add blake2b to checksumming algorithms

Add blake2b (with 256 bit digest) to the list of possible checksumming
algorithms used by BTRFS.

Signed-off-by: David Sterba <dsterba@suse.com>
5 years agobtrfs: add member for a specific checksum driver
David Sterba [Tue, 8 Oct 2019 16:41:33 +0000 (18:41 +0200)]
btrfs: add member for a specific checksum driver

Currently all the checksum algorithms generate a fixed size digest size
and we use it.  The on-disk format can hold up to BTRFS_CSUM_SIZE bytes
and BLAKE2b produces digest of 512 bits by default. We can't do that and
will use the blake2b-256, this needs to be passed to the crypto API.

Separate that from the base algorithm name and add a member to request
specific driver, in this case with the digest size.

The only place that uses the driver name is the crypto API setup.

Signed-off-by: David Sterba <dsterba@suse.com>
5 years agobtrfs: sysfs: show used checksum driver per filesystem
Johannes Thumshirn [Mon, 7 Oct 2019 09:11:04 +0000 (11:11 +0200)]
btrfs: sysfs: show used checksum driver per filesystem

Show the used driver for the checksum algorithm for the filesystem in
sysfs file /sys/fs/btrfs/UUID/features/checksum, eg.

  crc32c (crc32c-generic)

Signed-off-by: Johannes Thumshirn <jthumshirn@suse.de>
Reviewed-by: David Sterba <dsterba@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
5 years agobtrfs: sysfs: export supported checksums
David Sterba [Mon, 7 Oct 2019 09:11:03 +0000 (11:11 +0200)]
btrfs: sysfs: export supported checksums

Export supported checksum algorithms via sysfs in the list of static
features:

  /sys/fs/btrfs/features/supported_checksums

Space spearated list of checksum algorithm names.

Co-developed-by: Johannes Thumshirn <jthumshirn@suse.de>
Signed-off-by: David Sterba <dsterba@suse.com>
5 years agobtrfs: add sha256 to checksumming algorithm
Johannes Thumshirn [Mon, 7 Oct 2019 09:11:02 +0000 (11:11 +0200)]
btrfs: add sha256 to checksumming algorithm

Add sha256 to the list of possible checksumming algorithms used by BTRFS.

Signed-off-by: Johannes Thumshirn <jthumshirn@suse.de>
Reviewed-by: David Sterba <dsterba@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
5 years agobtrfs: add xxhash64 to checksumming algorithms
Johannes Thumshirn [Mon, 7 Oct 2019 09:11:01 +0000 (11:11 +0200)]
btrfs: add xxhash64 to checksumming algorithms

Add xxhash64 to the list of possible checksumming algorithms used by
BTRFS.

Signed-off-by: Johannes Thumshirn <jthumshirn@suse.de>
Reviewed-by: David Sterba <dsterba@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
5 years agobtrfs: get bdev from latest_dev for dio bh_result
David Sterba [Fri, 30 Aug 2019 13:42:07 +0000 (15:42 +0200)]
btrfs: get bdev from latest_dev for dio bh_result

To remove use of extent_map::bdev we need to find a replacement, and the
latest_bdev is the only one we can use here, because inode::i_bdev and
superblock::s_bdev are NULL.

The DIO code uses bdev in two places:

* to read blocksize to perform alignment checks in
  do_blockdev_direct_IO, but we do them in btrfs code before any call to
  DIO

* in the following call chain:

  do_direct_IO
    get_more_blocks
     sdio->get_block() <-- this is btrfs_get_blocks_direct

  subsequently the map_bh->b_dev member is used in clean_bdev_aliases
  and dio_new_bio to set the bio's bdev to that of the buffer_head.
  However, because we have provided a submit function dio_bio_submit
  calls our submission function and ignores the bdev.

So it's safe to pass any valid bdev that's used within the filesystem.

Reviewed-by: Nikolay Borisov <nborisov@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
5 years agobtrfs: assert extent_map bdevs and lookup_map and split
David Sterba [Fri, 10 May 2019 15:48:30 +0000 (17:48 +0200)]
btrfs: assert extent_map bdevs and lookup_map and split

This is a preparatory patch for removing extent_map::bdev. There's some
history behind the code so this is only precaution to catch if things
break before the actual removal happens.

Logically, comparing a raw low-level block device (bdev) does not make
sense for extent maps (high-level objects). This had no effect in
practice but was quite confusing in the code.  The lookup_map is set iff
EXTENT_FLAG_FS_MAPPING is set.

The two pointers were stored in the same bytes and used potentially in
two meanings. Now they're split, so the asserts are in place to check
that the condition will not change.

The lookup map pointer misused bdev, this has been changed in commit
95617d69326c ("btrfs: cleanup, stop casting for extent_map->lookup
everywhere") to the explicit type. But the semantics hasn't changed and
bdev was not actually used to decide if maps are mergeable.

Signed-off-by: David Sterba <dsterba@suse.com>
5 years agobtrfs: remove pointless indentation in btrfs_read_sys_array()
Johannes Thumshirn [Fri, 18 Oct 2019 09:58:23 +0000 (11:58 +0200)]
btrfs: remove pointless indentation in btrfs_read_sys_array()

Instead of checking if we've read a BTRFS_CHUNK_ITEM_KEY from disk and
then process it we could just bail out early if the read disk key wasn't
a BTRFS_CHUNK_ITEM_KEY.

This removes a level of indentation and makes the code nicer to read.

Reviewed-by: Nikolay Borisov <nborisov@suse.com>
Signed-off-by: Johannes Thumshirn <jthumshirn@suse.de>
Reviewed-by: David Sterba <dsterba@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
5 years agobtrfs: reduce indentation in btrfs_may_alloc_data_chunk
Johannes Thumshirn [Fri, 18 Oct 2019 09:58:22 +0000 (11:58 +0200)]
btrfs: reduce indentation in btrfs_may_alloc_data_chunk

In btrfs_may_alloc_data_chunk() we're checking if the chunk type is of
type BTRFS_BLOCK_GROUP_DATA and if it is we process it.

Instead of checking if the chunk type is a BTRFS_BLOCK_GROUP_DATA chunk
we can negate the check and bail out early if it isn't.

This makes the code a bit more readable.

Reviewed-by: Nikolay Borisov <nborisov@suse.com>
Signed-off-by: Johannes Thumshirn <jthumshirn@suse.de>
Reviewed-by: David Sterba <dsterba@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
5 years agobtrfs: remove pointless local variable in lock_stripe_add()
Johannes Thumshirn [Fri, 18 Oct 2019 09:58:21 +0000 (11:58 +0200)]
btrfs: remove pointless local variable in lock_stripe_add()

In lock_stripe_add() we're caching the bucket for the stripe hash table
just for a single call to dereference the stripe hash.

If we just directly call rbio_bucket() we can safe the pointless local
variable.

Also move the dereferencing of the stripe hash outside of the variable
declaration block to not break over the 80 characters limit.

Reviewed-by: Nikolay Borisov <nborisov@suse.com>
Signed-off-by: Johannes Thumshirn <jthumshirn@suse.de>
Reviewed-by: David Sterba <dsterba@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
5 years agobtrfs: raid56: reduce indentation in lock_stripe_add
Johannes Thumshirn [Fri, 18 Oct 2019 09:58:20 +0000 (11:58 +0200)]
btrfs: raid56: reduce indentation in lock_stripe_add

In lock_stripe_add() we're traversing the stripe hash list and check if
the current list element's raid_map equals is equal to the raid bio's
raid_map. If both are equal we continue processing.

If we'd check for inequality instead of equality we can reduce one level
of indentation.

Reviewed-by: Nikolay Borisov <nborisov@suse.com>
Signed-off-by: Johannes Thumshirn <jthumshirn@suse.de>
Reviewed-by: David Sterba <dsterba@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
5 years agobtrfs: tracepoints: constify all pointers
David Sterba [Thu, 17 Oct 2019 11:28:57 +0000 (13:28 +0200)]
btrfs: tracepoints: constify all pointers

We don't modify the data passed to tracepoints, some of the declarations
are already const, add it to the rest.

Signed-off-by: David Sterba <dsterba@suse.com>
5 years agobtrfs: tracepoints: drop typecasts from printk
David Sterba [Thu, 17 Oct 2019 11:28:55 +0000 (13:28 +0200)]
btrfs: tracepoints: drop typecasts from printk

Remove typecasts from trace printk, adjust types and move typecast to
the assignment if necessary. When assigning, the types are more obvious
compared to matching the variables to the format strings.

Signed-off-by: David Sterba <dsterba@suse.com>
5 years agobtrfs: Return offset from find_desired_extent
Nikolay Borisov [Fri, 27 Sep 2019 10:23:18 +0000 (13:23 +0300)]
btrfs: Return offset from find_desired_extent

Instead of using an input pointer parameter as the return value and have
an int as the return type of find_desired_extent, rework the function to
directly return the found offset. Doing that the 'ret' variable in
btrfs_llseek_file can be removed. Additional (subjective) benefit is
that btrfs' llseek function now resemebles those of the other major
filesystems.

Reviewed-by: Johannes Thumshirn <jthumshirn@suse.de>
Signed-off-by: Nikolay Borisov <nborisov@suse.com>
Reviewed-by: David Sterba <dsterba@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
5 years agobtrfs: Simplify btrfs_file_llseek
Nikolay Borisov [Fri, 27 Sep 2019 10:23:17 +0000 (13:23 +0300)]
btrfs: Simplify btrfs_file_llseek

Handle SEEK_END/SEEK_CUR in a single 'default' case by directly
returning from generic_file_llseek. This makes the 'out' label
redundant.  Finally return directly the vale from vfs_setpos. No
semantic changes.

Reviewed-by: Johannes Thumshirn <jthumshirn@suse.de>
Signed-off-by: Nikolay Borisov <nborisov@suse.com>
Reviewed-by: David Sterba <dsterba@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
5 years agobtrfs: Speed up btrfs_file_llseek
Nikolay Borisov [Fri, 27 Sep 2019 10:23:16 +0000 (13:23 +0300)]
btrfs: Speed up btrfs_file_llseek

Modifying the file position is done on a per-file basis. This renders
holding the inode lock for writing useless and makes the performance of
concurrent llseek's abysmal.

Fix this by holding the inode for read. This provides protection against
concurrent truncates and find_desired_extent already includes proper
extent locking for the range which ensures proper locking against
concurrent writes. SEEK_CUR and SEEK_END can be done lockessly.

The former is synchronized by file::f_lock spinlock. SEEK_END is not
synchronized but atomic, but that's OK since there is not guarantee that
SEEK_END will always be at the end of the file in the face of tail
modifications.

This change brings ~82% performance improvement when doing a lot of
parallel fseeks. The workload essentially does:

    for (d=0; d<num_seek_read; d++)
      {
/* offset %= 16777216; */
fseek (f, 256 * d % 16777216, SEEK_SET);
fread (buffer, 64, 1, f);
      }

Without patch:

num workprocesses = 16
num fseek/fread = 8000000
step = 256
fork 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15

real 0m41.412s
user 0m28.777s
sys 2m16.510s

With patch:

num workprocesses = 16
num fseek/fread = 8000000
step = 256
fork 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15

real 0m11.479s
user 0m27.629s
sys 0m21.040s

Signed-off-by: Nikolay Borisov <nborisov@suse.com>
Reviewed-by: David Sterba <dsterba@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
5 years agobtrfs: compression: remove ops pointer from workspace_manager
David Sterba [Fri, 4 Oct 2019 01:09:55 +0000 (03:09 +0200)]
btrfs: compression: remove ops pointer from workspace_manager

We can infer the ops from the type that is now passed to all functions
that would need it, this makes workspace_manager::ops redundant and can
be removed.

Reviewed-by: Johannes Thumshirn <jthumshirn@suse.de>
Reviewed-by: Nikolay Borisov <nborisov@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
5 years agobtrfs: compression: inline free_workspace
David Sterba [Fri, 4 Oct 2019 00:57:22 +0000 (02:57 +0200)]
btrfs: compression: inline free_workspace

Replace indirect calls to free_workspace by switch and calls to the
specific callbacks. This is mainly to get rid of the indirection due to
spectre vulnerability mitigations.

Reviewed-by: Johannes Thumshirn <jthumshirn@suse.de>
Reviewed-by: Nikolay Borisov <nborisov@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
5 years agobtrfs: compression: pass type to btrfs_put_workspace
David Sterba [Fri, 4 Oct 2019 00:50:28 +0000 (02:50 +0200)]
btrfs: compression: pass type to btrfs_put_workspace

We can infer the workspace_manager from type and the type will be used
in the following patch to call a common helper for free_workspace.

Reviewed-by: Johannes Thumshirn <jthumshirn@suse.de>
Reviewed-by: Nikolay Borisov <nborisov@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
5 years agobtrfs: compression: inline alloc_workspace
David Sterba [Fri, 4 Oct 2019 00:47:39 +0000 (02:47 +0200)]
btrfs: compression: inline alloc_workspace

Replace indirect calls to alloc_workspace by switch and calls to the
specific callbacks. This is mainly to get rid of the indirection due to
spectre vulnerability mitigations.

Reviewed-by: Johannes Thumshirn <jthumshirn@suse.de>
Reviewed-by: Nikolay Borisov <nborisov@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
5 years agobtrfs: compression: pass type to btrfs_get_workspace
David Sterba [Fri, 4 Oct 2019 00:50:28 +0000 (02:50 +0200)]
btrfs: compression: pass type to btrfs_get_workspace

We can infer the workspace_manager from type and the type will be used
in the following patch to call a common helper for alloc_workspace.

Reviewed-by: Johannes Thumshirn <jthumshirn@suse.de>
Reviewed-by: Nikolay Borisov <nborisov@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
5 years agobtrfs: compression: inline put_workspace
David Sterba [Fri, 4 Oct 2019 00:42:03 +0000 (02:42 +0200)]
btrfs: compression: inline put_workspace

Similar to get_workspace, majority of the callbacks is trivial, we don't
gain anything by the indirection, so replace them by a switch function.
Trivial callback implementations use the helper.

Reviewed-by: Johannes Thumshirn <jthumshirn@suse.de>
Reviewed-by: Nikolay Borisov <nborisov@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
5 years agobtrfs: compression: inline get_workspace
David Sterba [Fri, 4 Oct 2019 00:36:16 +0000 (02:36 +0200)]
btrfs: compression: inline get_workspace

Majority of the callbacks is trivial, we don't gain anything by the
indirection, so replace them by a switch function.

ZLIB needs to adjust level in the callback and ZSTD workspace management
is complex, the rest is call to the helper.

Reviewed-by: Johannes Thumshirn <jthumshirn@suse.de>
Reviewed-by: Nikolay Borisov <nborisov@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
5 years agobtrfs: compression: export alloc/free/get/put callbacks of all algos
David Sterba [Fri, 4 Oct 2019 00:21:48 +0000 (02:21 +0200)]
btrfs: compression: export alloc/free/get/put callbacks of all algos

The indirect calls will be replaced by a switch in compression.c.
(Switch is faster than indirect calls with when Spectre mitigations are
enabled).

Reviewed-by: Johannes Thumshirn <jthumshirn@suse.de>
Reviewed-by: Nikolay Borisov <nborisov@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
5 years agobtrfs: compression: inline cleanup_workspace_manager
David Sterba [Tue, 1 Oct 2019 23:08:03 +0000 (01:08 +0200)]
btrfs: compression: inline cleanup_workspace_manager

Replace loop calling to all algos with a list of direct calls to the
cleanup manager callback. When that becomes trivial it is replaced by
direct call to the helper.

Reviewed-by: Johannes Thumshirn <jthumshirn@suse.de>
Reviewed-by: Nikolay Borisov <nborisov@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
5 years agobtrfs: compression: let workspace manager cleanup take only the type
David Sterba [Thu, 3 Oct 2019 23:40:58 +0000 (01:40 +0200)]
btrfs: compression: let workspace manager cleanup take only the type

With the access to the workspace structures, we can look it up together
with the compression ops inside the workspace manager cleanup helper.

Reviewed-by: Johannes Thumshirn <jthumshirn@suse.de>
Reviewed-by: Nikolay Borisov <nborisov@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
5 years agobtrfs: compression: inline init_workspace_manager
David Sterba [Tue, 1 Oct 2019 23:08:03 +0000 (01:08 +0200)]
btrfs: compression: inline init_workspace_manager

Replace loop calling to all algos with a list of direct calls to the
init manager callback. When that becomes trivial it is replaced by
direct call to the helper.

Reviewed-by: Johannes Thumshirn <jthumshirn@suse.de>
Reviewed-by: Nikolay Borisov <nborisov@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
5 years agobtrfs: compression: let workspace manager init take only the type
David Sterba [Thu, 3 Oct 2019 23:40:58 +0000 (01:40 +0200)]
btrfs: compression: let workspace manager init take only the type

With the access to the workspace structures, we can look it up together
with the compression ops inside the workspace manager init helper.

Reviewed-by: Johannes Thumshirn <jthumshirn@suse.de>
Reviewed-by: Nikolay Borisov <nborisov@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
5 years agobtrfs: compression: attach workspace manager to the ops
David Sterba [Tue, 1 Oct 2019 22:53:31 +0000 (00:53 +0200)]
btrfs: compression: attach workspace manager to the ops

There's a lot of indirection when the generic code calls into
algo-specific callbacks to reach the private workspace manager structure
and back to the generic code.

To simplify that, export the workspace manager for heuristic, LZO and
ZLIB, while ZSTD is going to use it's own manager.

Reviewed-by: Johannes Thumshirn <jthumshirn@suse.de>
Reviewed-by: Nikolay Borisov <nborisov@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
5 years agobtrfs: switch compression callbacks to direct calls
David Sterba [Tue, 1 Oct 2019 22:06:15 +0000 (00:06 +0200)]
btrfs: switch compression callbacks to direct calls

The indirect calls bring some overhead due to spectre vulnerability
mitigations. The number of cases is small and below the threshold
(10-20) where indirect call would be better.

Reviewed-by: Johannes Thumshirn <jthumshirn@suse.de>
Reviewed-by: Nikolay Borisov <nborisov@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
5 years agobtrfs: export compression and decompression callbacks
David Sterba [Tue, 1 Oct 2019 20:38:34 +0000 (22:38 +0200)]
btrfs: export compression and decompression callbacks

Export compress_pages, decompress_bio and decompress callbacks for all
compression algos. The indirect calls will be replaced by a switch.

Reviewed-by: Johannes Thumshirn <jthumshirn@suse.de>
Reviewed-by: Nikolay Borisov <nborisov@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
5 years agobtrfs: use btrfs_block_group_cache_done in update_block_group
Josef Bacik [Tue, 24 Sep 2019 20:50:44 +0000 (16:50 -0400)]
btrfs: use btrfs_block_group_cache_done in update_block_group

When free'ing extents in a block group we check to see if the block
group is not cached, and then cache it if we need to.  However we'll
just carry on as long as we're loading the cache.  This is problematic
because we are dirtying the block group here.  If we are fast enough we
could do a transaction commit and clear the free space cache while we're
still loading the space cache in another thread.  This truncates the
free space inode, which will keep it from loading the space cache.

Fix this by using the btrfs_block_group_cache_done helper so that we try
to load the space cache unconditionally here, which will result in the
caller waiting for the fast caching to complete and keep us from
truncating the free space inode.

CC: stable@vger.kernel.org # 4.4+
Signed-off-by: Josef Bacik <josef@toxicpanda.com>
Reviewed-by: Nikolay Borisov <nborisov@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
5 years agobtrfs: check page->mapping when loading free space cache
Josef Bacik [Tue, 24 Sep 2019 20:50:43 +0000 (16:50 -0400)]
btrfs: check page->mapping when loading free space cache

While testing 5.2 we ran into the following panic

[52238.017028] BUG: kernel NULL pointer dereference, address: 0000000000000001
[52238.105608] RIP: 0010:drop_buffers+0x3d/0x150
[52238.304051] Call Trace:
[52238.308958]  try_to_free_buffers+0x15b/0x1b0
[52238.317503]  shrink_page_list+0x1164/0x1780
[52238.325877]  shrink_inactive_list+0x18f/0x3b0
[52238.334596]  shrink_node_memcg+0x23e/0x7d0
[52238.342790]  ? do_shrink_slab+0x4f/0x290
[52238.350648]  shrink_node+0xce/0x4a0
[52238.357628]  balance_pgdat+0x2c7/0x510
[52238.365135]  kswapd+0x216/0x3e0
[52238.371425]  ? wait_woken+0x80/0x80
[52238.378412]  ? balance_pgdat+0x510/0x510
[52238.386265]  kthread+0x111/0x130
[52238.392727]  ? kthread_create_on_node+0x60/0x60
[52238.401782]  ret_from_fork+0x1f/0x30

The page we were trying to drop had a page->private, but had no
page->mapping and so called drop_buffers, assuming that we had a
buffer_head on the page, and then panic'ed trying to deref 1, which is
our page->private for data pages.

This is happening because we're truncating the free space cache while
we're trying to load the free space cache.  This isn't supposed to
happen, and I'll fix that in a followup patch.  However we still
shouldn't allow those sort of mistakes to result in messing with pages
that do not belong to us.  So add the page->mapping check to verify that
we still own this page after dropping and re-acquiring the page lock.

This page being unlocked as:
btrfs_readpage
  extent_read_full_page
    __extent_read_full_page
      __do_readpage
        if (!nr)
   unlock_page  <-- nr can be 0 only if submit_extent_page
    returns an error

CC: stable@vger.kernel.org # 4.4+
Reviewed-by: Filipe Manana <fdmanana@suse.com>
Reviewed-by: Nikolay Borisov <nborisov@suse.com>
Signed-off-by: Josef Bacik <josef@toxicpanda.com>
[ add callchain ]
Signed-off-by: David Sterba <dsterba@suse.com>
5 years agoBtrfs: fix metadata space leak on fixup worker failure to set range as delalloc
Filipe Manana [Wed, 9 Oct 2019 16:43:59 +0000 (17:43 +0100)]
Btrfs: fix metadata space leak on fixup worker failure to set range as delalloc

In the fixup worker, if we fail to mark the range as delalloc in the io
tree, we must release the previously reserved metadata, as well as update
the outstanding extents counter for the inode, otherwise we leak metadata
space.

In pratice we can't return an error from btrfs_set_extent_delalloc(),
which is just a wrapper around __set_extent_bit(), as for most errors
__set_extent_bit() does a BUG_ON() (or panics which hits a BUG_ON() as
well) and returning an -EEXIST error doesn't happen in this case since
the exclusive bits parameter always has a value of 0 through this code
path. Nevertheless, just fix the error handling in the fixup worker,
in case one day __set_extent_bit() can return an error to this code
path.

Fixes: f3038ee3a3f101 ("btrfs: Handle btrfs_set_extent_delalloc failure in fixup worker")
CC: stable@vger.kernel.org # 4.19+
Reviewed-by: Nikolay Borisov <nborisov@suse.com>
Signed-off-by: Filipe Manana <fdmanana@suse.com>
Reviewed-by: David Sterba <dsterba@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
5 years agoBtrfs: fix negative subv_writers counter and data space leak after buffered write
Filipe Manana [Fri, 11 Oct 2019 15:41:20 +0000 (16:41 +0100)]
Btrfs: fix negative subv_writers counter and data space leak after buffered write

When doing a buffered write it's possible to leave the subv_writers
counter of the root, used for synchronization between buffered nocow
writers and snapshotting. This happens in an exceptional case like the
following:

1) We fail to allocate data space for the write, since there's not
   enough available data space nor enough unallocated space for allocating
   a new data block group;

2) Because of that failure, we try to go to NOCOW mode, which succeeds
   and therefore we set the local variable 'only_release_metadata' to true
   and set the root's sub_writers counter to 1 through the call to
   btrfs_start_write_no_snapshotting() made by check_can_nocow();

3) The call to btrfs_copy_from_user() returns zero, which is very unlikely
   to happen but not impossible;

4) No pages are copied because btrfs_copy_from_user() returned zero;

5) We call btrfs_end_write_no_snapshotting() which decrements the root's
   subv_writers counter to 0;

6) We don't set 'only_release_metadata' back to 'false' because we do
   it only if 'copied', the value returned by btrfs_copy_from_user(), is
   greater than zero;

7) On the next iteration of the while loop, which processes the same
   page range, we are now able to allocate data space for the write (we
   got enough data space released in the meanwhile);

8) After this if we fail at btrfs_delalloc_reserve_metadata(), because
   now there isn't enough free metadata space, or in some other place
   further below (prepare_pages(), lock_and_cleanup_extent_if_need(),
   btrfs_dirty_pages()), we break out of the while loop with
   'only_release_metadata' having a value of 'true';

9) Because 'only_release_metadata' is 'true' we end up decrementing the
   root's subv_writers counter to -1 (through a call to
   btrfs_end_write_no_snapshotting()), and we also end up not releasing the
   data space previously reserved through btrfs_check_data_free_space().
   As a consequence the mechanism for synchronizing NOCOW buffered writes
   with snapshotting gets broken.

Fix this by always setting 'only_release_metadata' to false at the start
of each iteration.

Fixes: 8257b2dc3c1a ("Btrfs: introduce btrfs_{start, end}_nocow_write() for each subvolume")
Fixes: 7ee9e4405f26 ("Btrfs: check if we can nocow if we don't have data space")
CC: stable@vger.kernel.org # 4.4+
Reviewed-by: Josef Bacik <josef@toxicpanda.com>
Signed-off-by: Filipe Manana <fdmanana@suse.com>
Reviewed-by: David Sterba <dsterba@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
5 years agobtrfs: ioctl: Try to use btrfs_fs_info instead of *file
Marcos Paulo de Souza [Fri, 11 Oct 2019 00:23:11 +0000 (21:23 -0300)]
btrfs: ioctl: Try to use btrfs_fs_info instead of *file

Some functions are doing some unnecessary indirection to reach the
btrfs_fs_info struct. Change these functions to receive a btrfs_fs_info
struct instead of a *file.

Reviewed-by: Anand Jain <anand.jain@oracle.com>
Reviewed-by: Nikolay Borisov <nborisov@suse.com>
Signed-off-by: Marcos Paulo de Souza <mpdesouza@suse.com>
Reviewed-by: David Sterba <dsterba@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
5 years agobtrfs: use bool argument in free_root_pointers()
Anand Jain [Thu, 10 Oct 2019 02:39:25 +0000 (10:39 +0800)]
btrfs: use bool argument in free_root_pointers()

We don't need int argument bool shall do in free_root_pointers().  And
rename the argument as it confused two people.

Reviewed-by: Qu Wenruo <wqu@suse.com>
Signed-off-by: Anand Jain <anand.jain@oracle.com>
Reviewed-by: David Sterba <dsterba@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
5 years agobtrfs: use better definition of number of compression type
Chengguang Xu [Thu, 10 Oct 2019 07:59:57 +0000 (15:59 +0800)]
btrfs: use better definition of number of compression type

The compression type upper limit constant is the same as the last value
and this is confusing.  In order to keep coding style consistent, use
BTRFS_NR_COMPRESS_TYPES as the total number that follows the idom of
'NR' being one more than the last value.

Signed-off-by: Chengguang Xu <cgxu519@mykernel.net>
Reviewed-by: David Sterba <dsterba@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
5 years agobtrfs: use enum for extent type defines
Chengguang Xu [Thu, 10 Oct 2019 07:59:58 +0000 (15:59 +0800)]
btrfs: use enum for extent type defines

Use enum to replace macro definitions of extent types.

Signed-off-by: Chengguang Xu <cgxu519@mykernel.net>
Reviewed-by: David Sterba <dsterba@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
5 years agobtrfs: props: remove unnecessary hash_init()
Chengguang Xu [Thu, 10 Oct 2019 07:59:56 +0000 (15:59 +0800)]
btrfs: props: remove unnecessary hash_init()

DEFINE_HASHTABLE itself has already included initialization code,
we don't have to call hash_init() again, so remove it.

Signed-off-by: Chengguang Xu <cgxu519@mykernel.net>
Reviewed-by: David Sterba <dsterba@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
5 years agobtrfs: Rename btrfs_join_transaction_nolock
Nikolay Borisov [Tue, 8 Oct 2019 17:43:06 +0000 (20:43 +0300)]
btrfs: Rename btrfs_join_transaction_nolock

This function is used only during the final phase of freespace cache
writeout. This is necessary since using the plain btrfs_join_transaction
api is deadlock prone. The deadlock looks like:

T1:
btrfs_commit_transaction
  commit_cowonly_roots
    btrfs_write_dirty_block_groups
      btrfs_wait_cache_io
        __btrfs_wait_cache_io
       btrfs_wait_ordered_range <-- Triggers ordered IO for freespace
                                    inode and blocks transaction commit
    until freespace cache writeout

T2: <-- after T1 has triggered the writeout
finish_ordered_fn
  btrfs_finish_ordered_io
    btrfs_join_transaction <--- this would block waiting for current
                                transaction to commit, but since trans
commit is waiting for this writeout to
finish

The special purpose functions prevents it by simply skipping the "wait
for writeout" since it's guaranteed the transaction won't proceed until
we are done.

Reviewed-by: Qu Wenruo <wqu@suse.com>
Signed-off-by: Nikolay Borisov <nborisov@suse.com>
Reviewed-by: David Sterba <dsterba@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
5 years agobtrfs: User assert to document transaction requirement
Nikolay Borisov [Tue, 8 Oct 2019 13:26:16 +0000 (16:26 +0300)]
btrfs: User assert to document transaction requirement

Using an ASSERT in btrfs_pin_extent allows to more stringently observe
whether the function is called under a transaction or not.

Reviewed-by: Johannes Thumshirn <jthumshirn@suse.de>
Signed-off-by: Nikolay Borisov <nborisov@suse.com>
Reviewed-by: David Sterba <dsterba@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
5 years agobtrfs: opencode extent_buffer_get
David Sterba [Tue, 8 Oct 2019 11:28:47 +0000 (13:28 +0200)]
btrfs: opencode extent_buffer_get

The helper is trivial and we can understand what the atomic_inc on
something named refs does.

Reviewed-by: Johannes Thumshirn <jthumshirn@suse.de>
Reviewed-by: Anand Jain <anand.jain@oracle.com>
Signed-off-by: David Sterba <dsterba@suse.com>
5 years agobtrfs: Avoid getting stuck during cyclic writebacks
Tejun Heo [Thu, 3 Oct 2019 14:27:13 +0000 (07:27 -0700)]
btrfs: Avoid getting stuck during cyclic writebacks

During a cyclic writeback, extent_write_cache_pages() uses done_index
to update the writeback_index after the current run is over.  However,
instead of current index + 1, it gets to to the current index itself.

Unfortunately, this, combined with returning on EOF instead of looping
back, can lead to the following pathlogical behavior.

1. There is a single file which has accumulated enough dirty pages to
   trigger balance_dirty_pages() and the writer appending to the file
   with a series of short writes.

2. balance_dirty_pages kicks in, wakes up background writeback and sleeps.

3. Writeback kicks in and the cursor is on the last page of the dirty
   file.  Writeback is started or skipped if already in progress.  As
   it's EOF, extent_write_cache_pages() returns and the cursor is set
   to done_index which is pointing to the last page.

4. Writeback is done.  Nothing happens till balance_dirty_pages
   finishes, at which point we go back to #1.

This can almost completely stall out writing back of the file and keep
the system over dirty threshold for a long time which can mess up the
whole system.  We encountered this issue in production with a package
handling application which can reliably reproduce the issue when
running under tight memory limits.

Reading the comment in the error handling section, this seems to be to
avoid accidentally skipping a page in case the write attempt on the
page doesn't succeed.  However, this concern seems bogus.

On each page, the code either:

* Skips and moves onto the next page.

* Fails issue and sets done_index to index + 1.

* Successfully issues and continue to the next page if budget allows
  and not EOF.

IOW, as long as it's not EOF and there's budget, the code never
retries writing back the same page.  Only when a page happens to be
the last page of a particular run, we end up retrying the page, which
can't possibly guarantee anything data integrity related.  Besides,
cyclic writes are only used for non-syncing writebacks meaning that
there's no data integrity implication to begin with.

Fix it by always setting done_index past the current page being
processed.

Note that this problem exists in other writepages too.

CC: stable@vger.kernel.org # 4.19+
Signed-off-by: Tejun Heo <tj@kernel.org>
Reviewed-by: David Sterba <dsterba@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
5 years agobtrfs: block-group: Rework documentation of check_system_chunk function
Marcos Paulo de Souza [Tue, 8 Oct 2019 00:50:38 +0000 (21:50 -0300)]
btrfs: block-group: Rework documentation of check_system_chunk function

Commit 4617ea3a52cf (" Btrfs: fix necessary chunk tree space calculation
when allocating a chunk") removed the is_allocation argument from
check_system_chunk, since the formula for reserving the necessary space
for allocation or removing a chunk would be the same.

So, rework the comment by removing the mention of is_allocation
argument.

Signed-off-by: Marcos Paulo de Souza <marcos.souza.org@gmail.com>
Reviewed-by: Nikolay Borisov <nborisov@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
5 years agobtrfs: Enhance error output for write time tree checker
Qu Wenruo [Fri, 4 Oct 2019 09:31:33 +0000 (17:31 +0800)]
btrfs: Enhance error output for write time tree checker

Unlike read time tree checker errors, write time error can't be
inspected by "btrfs inspect dump-tree", so we need extra information to
determine what's going wrong.

The patch will add the following output for write time tree checker
error:

- The content of the offending tree block
  To help determining if it's a false alert.

- Kernel WARN_ON() for debug build
  This is helpful for us to detect unexpected write time tree checker
  error, especially fstests could catch the dmesg.
  Since the WARN_ON() is only triggered for write time tree checker,
  test cases utilizing dm-error won't trigger this WARN_ON(), thus no
  extra noise.

Signed-off-by: Qu Wenruo <wqu@suse.com>
Reviewed-by: David Sterba <dsterba@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
5 years agobtrfs: tree-checker: Refactor prev_key check for ino into a function
Qu Wenruo [Fri, 4 Oct 2019 09:31:32 +0000 (17:31 +0800)]
btrfs: tree-checker: Refactor prev_key check for ino into a function

Refactor the check for prev_key->objectid of the following key types
into one function, check_prev_ino():

- EXTENT_DATA
- INODE_REF
- DIR_INDEX
- DIR_ITEM
- XATTR_ITEM

Also add the check of prev_key for INODE_REF.

Signed-off-by: Qu Wenruo <wqu@suse.com>
Reviewed-by: David Sterba <dsterba@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
5 years agoBtrfs: extent_write_locked_range() should attach inode->i_wb
Chris Mason [Wed, 10 Jul 2019 19:28:18 +0000 (12:28 -0700)]
Btrfs: extent_write_locked_range() should attach inode->i_wb

extent_write_locked_range() is used when we're falling back to buffered
IO from inside of compression.  It allocates its own wbc and should
associate it with the inode's i_wb to make sure the IO goes down from
the correct cgroup.

Reviewed-by: Josef Bacik <josef@toxicpanda.com>
Signed-off-by: Chris Mason <clm@fb.com>
Signed-off-by: David Sterba <dsterba@suse.com>
5 years agoBtrfs: use REQ_CGROUP_PUNT for worker thread submitted bios
Chris Mason [Wed, 10 Jul 2019 19:28:17 +0000 (12:28 -0700)]
Btrfs: use REQ_CGROUP_PUNT for worker thread submitted bios

Async CRCs and compression submit IO through helper threads, which means
they have IO priority inversions when cgroup IO controllers are in use.

This flags all of the writes submitted by btrfs helper threads as
REQ_CGROUP_PUNT.  submit_bio() will punt these to dedicated per-blkcg
work items to avoid the priority inversion.

For the compression code, we take a reference on the wbc's blkg css and
pass it down to the async workers.

For the async CRCs, the bio already has the correct css, we just need to
tell the block layer to use REQ_CGROUP_PUNT.

Reviewed-by: Josef Bacik <josef@toxicpanda.com>
Signed-off-by: Chris Mason <clm@fb.com>
Modified-and-reviewed-by: Tejun Heo <tj@kernel.org>
Signed-off-by: David Sterba <dsterba@suse.com>
5 years agoBtrfs: only associate the locked page with one async_chunk struct
Chris Mason [Wed, 10 Jul 2019 19:28:16 +0000 (12:28 -0700)]
Btrfs: only associate the locked page with one async_chunk struct

The btrfs writepages function collects a large range of pages flagged
for delayed allocation, and then sends them down through the COW code
for processing.  When compression is on, we allocate one async_chunk
structure for every 512K, and then run those pages through the
compression code for IO submission.

writepages starts all of this off with a single page, locked by the
original call to extent_write_cache_pages(), and it's important to keep
track of this page because it has already been through
clear_page_dirty_for_io().

The btrfs async_chunk struct has a pointer to the locked_page, and when
we're redirtying the page because compression had to fallback to
uncompressed IO, we use page->index to decide if a given async_chunk
struct really owns that page.

But, this is racey.  If a given delalloc range is broken up into two
async_chunks (chunkA and chunkB), we can end up with something like
this:

 compress_file_range(chunkA)
 submit_compress_extents(chunkA)
 submit compressed bios(chunkA)
 put_page(locked_page)

 compress_file_range(chunkB)
 ...

Or:

 async_cow_submit
  submit_compressed_extents <--- falls back to buffered writeout
   cow_file_range
    extent_clear_unlock_delalloc
     __process_pages_contig
       put_page(locked_pages)

    async_cow_submit

The end result is that chunkA is completed and cleaned up before chunkB
even starts processing.  This means we can free locked_page() and reuse
it elsewhere.  If we get really lucky, it'll have the same page->index
in its new home as it did before.

While we're processing chunkB, we might decide we need to fall back to
uncompressed IO, and so compress_file_range() will call
__set_page_dirty_nobufers() on chunkB->locked_page.

Without cgroups in use, this creates as a phantom dirty page, which
isn't great but isn't the end of the world. What can happen, it can go
through the fixup worker and the whole COW machinery again:

in submit_compressed_extents():
  while (async extents) {
  ...
    cow_file_range
    if (!page_started ...)
      extent_write_locked_range
    else if (...)
      unlock_page
    continue;

This hasn't been observed in practice but is still possible.

With cgroups in use, we might crash in the accounting code because
page->mapping->i_wb isn't set.

  BUG: unable to handle kernel NULL pointer dereference at 00000000000000d0
  IP: percpu_counter_add_batch+0x11/0x70
  PGD 66534e067 P4D 66534e067 PUD 66534f067 PMD 0
  Oops: 0000 [#1] SMP DEBUG_PAGEALLOC
  CPU: 16 PID: 2172 Comm: rm Not tainted
  RIP: 0010:percpu_counter_add_batch+0x11/0x70
  RSP: 0018:ffffc9000a97bbe0 EFLAGS: 00010286
  RAX: 0000000000000005 RBX: 0000000000000090 RCX: 0000000000026115
  RDX: 0000000000000030 RSI: ffffffffffffffff RDI: 0000000000000090
  RBP: 0000000000000000 R08: fffffffffffffff5 R09: 0000000000000000
  R10: 00000000000260c0 R11: ffff881037fc26c0 R12: ffffffffffffffff
  R13: ffff880fe4111548 R14: ffffc9000a97bc90 R15: 0000000000000001
  FS:  00007f5503ced480(0000) GS:ffff880ff7200000(0000) knlGS:0000000000000000
  CS:  0010 DS: 0000 ES: 0000 CR0: 0000000080050033
  CR2: 00000000000000d0 CR3: 00000001e0459005 CR4: 0000000000360ee0
  DR0: 0000000000000000 DR1: 0000000000000000 DR2: 0000000000000000
  DR3: 0000000000000000 DR6: 00000000fffe0ff0 DR7: 0000000000000400
  Call Trace:
   account_page_cleaned+0x15b/0x1f0
   __cancel_dirty_page+0x146/0x200
   truncate_cleanup_page+0x92/0xb0
   truncate_inode_pages_range+0x202/0x7d0
   btrfs_evict_inode+0x92/0x5a0
   evict+0xc1/0x190
   do_unlinkat+0x176/0x280
   do_syscall_64+0x63/0x1a0
   entry_SYSCALL_64_after_hwframe+0x42/0xb7

The fix here is to make asyc_chunk->locked_page NULL everywhere but the
one async_chunk struct that's allowed to do things to the locked page.

Link: https://lore.kernel.org/linux-btrfs/c2419d01-5c84-3fb4-189e-4db519d08796@suse.com/
Fixes: 771ed689d2cd ("Btrfs: Optimize compressed writeback and reads")
Reviewed-by: Josef Bacik <josef@toxicpanda.com>
Signed-off-by: Chris Mason <clm@fb.com>
[ update changelog from mail thread discussion ]
Signed-off-by: David Sterba <dsterba@suse.com>
5 years agoBtrfs: delete the entire async bio submission framework
Chris Mason [Wed, 10 Jul 2019 19:28:15 +0000 (12:28 -0700)]
Btrfs: delete the entire async bio submission framework

Now that we're not using btrfs_schedule_bio() anymore, delete all the
code that supported it.

Reviewed-by: Josef Bacik <josef@toxicpanda.com>
Reviewed-by: Nikolay Borisov <nborisov@suse.com>
Signed-off-by: Chris Mason <clm@fb.com>
Reviewed-by: David Sterba <dsterba@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
5 years agoBtrfs: stop using btrfs_schedule_bio()
Chris Mason [Wed, 10 Jul 2019 19:28:14 +0000 (12:28 -0700)]
Btrfs: stop using btrfs_schedule_bio()

btrfs_schedule_bio() hands IO off to a helper thread to do the actual
submit_bio() call.  This has been used to make sure async crc and
compression helpers don't get stuck on IO submission.  To maintain good
performance, over time the IO submission threads duplicated some IO
scheduler characteristics such as high and low priority IOs and they
also made some ugly assumptions about request allocation batch sizes.

All of this cost at least one extra context switch during IO submission,
and doesn't fit well with the modern blkmq IO stack.  So, this commit stops
using btrfs_schedule_bio().  We may need to adjust the number of async
helper threads for crcs and compression, but long term it's a better
path.

Reviewed-by: Josef Bacik <josef@toxicpanda.com>
Reviewed-by: Nikolay Borisov <nborisov@suse.com>
Signed-off-by: Chris Mason <clm@fb.com>
Reviewed-by: David Sterba <dsterba@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
5 years agobtrfs: add __pure attribute to functions
David Sterba [Tue, 1 Oct 2019 17:57:39 +0000 (19:57 +0200)]
btrfs: add __pure attribute to functions

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>
5 years agobtrfs: add const function attribute
David Sterba [Tue, 1 Oct 2019 17:57:37 +0000 (19:57 +0200)]
btrfs: add const function attribute

For some reason the attribute is called __attribute_const__ and not
__const, marks functions that have no observable effects on program
state, IOW not reading pointers, just the arguments and calculating a
value. Allows the compiler to do some optimizations, based on
-Wsuggest-attribute=const . The effects are rather small, though, about
60 bytes decrese of btrfs.ko.

Reviewed-by: Nikolay Borisov <nborisov@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
5 years agobtrfs: add __cold attribute to more functions
David Sterba [Tue, 1 Oct 2019 17:57:35 +0000 (19:57 +0200)]
btrfs: add __cold attribute to more functions

The attribute can mark functions supposed to be called rarely if at all
and the text can be moved to sections far from the other code. The
attribute has been added to several functions already, this patch is
based on hints given by gcc -Wsuggest-attribute=cold.

The net effect of this patch is decrease of btrfs.ko by 1000-1300,
depending on the config options.

Reviewed-by: Nikolay Borisov <nborisov@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
5 years agobtrfs: drop unused parameter is_new from btrfs_iget
David Sterba [Thu, 3 Oct 2019 17:09:35 +0000 (19:09 +0200)]
btrfs: drop unused parameter is_new from btrfs_iget

The parameter is now always set to NULL and could be dropped. The last
user was get_default_root but that got reworked in 05dbe6837b60 ("Btrfs:
unify subvol= and subvolid= mounting") and the parameter became unused.

Reviewed-by: Anand Jain <anand.jain@oracle.com>
Signed-off-by: David Sterba <dsterba@suse.com>
5 years agobtrfs: use refcount_inc_not_zero in kill_all_nodes
Josef Bacik [Thu, 26 Sep 2019 12:29:32 +0000 (08:29 -0400)]
btrfs: use refcount_inc_not_zero in kill_all_nodes

We hit the following warning while running down a different problem

[ 6197.175850] ------------[ cut here ]------------
[ 6197.185082] refcount_t: underflow; use-after-free.
[ 6197.194704] WARNING: CPU: 47 PID: 966 at lib/refcount.c:190 refcount_sub_and_test_checked+0x53/0x60
[ 6197.521792] Call Trace:
[ 6197.526687]  __btrfs_release_delayed_node+0x76/0x1c0
[ 6197.536615]  btrfs_kill_all_delayed_nodes+0xec/0x130
[ 6197.546532]  ? __btrfs_btree_balance_dirty+0x60/0x60
[ 6197.556482]  btrfs_clean_one_deleted_snapshot+0x71/0xd0
[ 6197.566910]  cleaner_kthread+0xfa/0x120
[ 6197.574573]  kthread+0x111/0x130
[ 6197.581022]  ? kthread_create_on_node+0x60/0x60
[ 6197.590086]  ret_from_fork+0x1f/0x30
[ 6197.597228] ---[ end trace 424bb7ae00509f56 ]---

This is because the free side drops the ref without the lock, and then
takes the lock if our refcount is 0.  So you can have nodes on the tree
that have a refcount of 0.  Fix this by zero'ing out that element in our
temporary array so we don't try to kill it again.

CC: stable@vger.kernel.org # 4.14+
Reviewed-by: Nikolay Borisov <nborisov@suse.com>
Signed-off-by: Josef Bacik <josef@toxicpanda.com>
Reviewed-by: David Sterba <dsterba@suse.com>
[ add comment ]
Signed-off-by: David Sterba <dsterba@suse.com>
5 years agobtrfs: print process name and pid that calls device scanning
Anand Jain [Wed, 2 Oct 2019 10:30:48 +0000 (18:30 +0800)]
btrfs: print process name and pid that calls device scanning

Its very helpful if we had logged the device scanner process name to
debug the race condition between the systemd-udevd scan and the user
initiated device forget command.

This patch adds process name and pid to the scan message.

Signed-off-by: Anand Jain <anand.jain@oracle.com>
Reviewed-by: David Sterba <dsterba@suse.com>
[ add pid to the message ]
Signed-off-by: David Sterba <dsterba@suse.com>
5 years agobtrfs: Open-code name_in_log_ref in replay_one_name
Nikolay Borisov [Fri, 30 Aug 2019 14:44:49 +0000 (17:44 +0300)]
btrfs: Open-code name_in_log_ref in replay_one_name

That function adds unnecessary indirection between backref_in_log and
the caller. Furthermore it also "downgrades" backref_in_log's return
value to a boolean, when in fact it could very well be an error.

Rectify the situation by simply opencoding name_in_log_ref in
replay_one_name and properly handling possible return codes from
backref_in_log.

Signed-off-by: Nikolay Borisov <nborisov@suse.com>
Reviewed-by: David Sterba <dsterba@suse.com>
[ update comment ]
Signed-off-by: David Sterba <dsterba@suse.com>
5 years agobtrfs: Properly handle backref_in_log retval
Nikolay Borisov [Wed, 25 Sep 2019 11:03:03 +0000 (14:03 +0300)]
btrfs: Properly handle backref_in_log retval

This function can return a negative error value if btrfs_search_slot
errors for whatever reason or if btrfs_alloc_path runs out of memory.
This is currently problemattic because backref_in_log is treated by its
callers as if it returns boolean.

Fix this by adding proper error handling in callers. That also enables
the function to return the direct error code from btrfs_search_slot.

Signed-off-by: Nikolay Borisov <nborisov@suse.com>
Reviewed-by: David Sterba <dsterba@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
5 years agobtrfs: Don't opencode btrfs_find_name_in_backref in backref_in_log
Nikolay Borisov [Fri, 30 Aug 2019 14:44:47 +0000 (17:44 +0300)]
btrfs: Don't opencode btrfs_find_name_in_backref in backref_in_log

Direct replacement, though note that the inside of the loop in
btrfs_find_name_in_backref is organized in a slightly different way but
is equvalent.

Signed-off-by: Nikolay Borisov <nborisov@suse.com>
Reviewed-by: David Sterba <dsterba@suse.com>
[ add changelog ]
Signed-off-by: David Sterba <dsterba@suse.com>
5 years agobtrfs: transaction: Cleanup unused TRANS_STATE_BLOCKED
Qu Wenruo [Thu, 22 Aug 2019 07:25:00 +0000 (15:25 +0800)]
btrfs: transaction: Cleanup unused TRANS_STATE_BLOCKED

The state was introduced in commit 4a9d8bdee368 ("Btrfs: make the state
of the transaction more readable"), then in commit 302167c50b32
("btrfs: don't end the transaction for delayed refs in throttle") the
state is completely removed.

So we can just clean up the state since it's only compared but never
set.

Signed-off-by: Qu Wenruo <wqu@suse.com>
Reviewed-by: David Sterba <dsterba@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
5 years agobtrfs: transaction: describe transaction states and transitions
Qu Wenruo [Thu, 22 Aug 2019 07:24:59 +0000 (15:24 +0800)]
btrfs: transaction: describe transaction states and transitions

Add an overview of the basic btrfs transaction transitions, including
the following states:

- No transaction states
- Transaction N [[TRANS_STATE_RUNNING]]
- Transaction N [[TRANS_STATE_COMMIT_START]]
- Transaction N [[TRANS_STATE_COMMIT_DOING]]
- Transaction N [[TRANS_STATE_UNBLOCKED]]
- Transaction N [[TRANS_STATE_COMPLETED]]

For each state, the comment will include:

- Basic explaination about current state
- How to go next stage
- What will happen if we call various start_transaction() functions
- Relationship to transaction N+1

This doesn't provide tech details, but serves as a cheat sheet for
reader to get into the code a little easier.

Signed-off-by: Qu Wenruo <wqu@suse.com>
Reviewed-by: David Sterba <dsterba@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
5 years agobtrfs: use has_single_bit_set for clarity
David Sterba [Tue, 1 Oct 2019 17:44:42 +0000 (19:44 +0200)]
btrfs: use has_single_bit_set for clarity

Replace is_power_of_2 with the helper that is self-documenting and
remove the open coded call in alloc_profile_is_valid.

Signed-off-by: David Sterba <dsterba@suse.com>
5 years agobtrfs: add 64bit safe helper for power of two checks
David Sterba [Tue, 1 Oct 2019 17:40:15 +0000 (19:40 +0200)]
btrfs: add 64bit safe helper for power of two checks

As is_power_of_two takes unsigned long, it's not safe on 32bit
architectures, but we could pass any u64 value in seveal places. Add a
separate helper and also an alias that better expresses the purpose for
which the helper is used.

Signed-off-by: David Sterba <dsterba@suse.com>
5 years agobtrfs: balance: use term redundancy instead of integrity in message
Anand Jain [Wed, 25 Sep 2019 06:29:28 +0000 (14:29 +0800)]
btrfs: balance: use term redundancy instead of integrity in message

When balance reduces the number of copies of metadata, it reduces the
redundancy, use the term redundancy instead of integrity.

Signed-off-by: Anand Jain <anand.jain@oracle.com>
Reviewed-by: David Sterba <dsterba@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
5 years agobtrfs: move btrfs_unlock_up_safe to other locking functions
David Sterba [Tue, 24 Sep 2019 17:17:17 +0000 (19:17 +0200)]
btrfs: move btrfs_unlock_up_safe to other locking functions

The function belongs to the family of locking functions, so move it
there. The 'noinline' keyword is dropped as it's now an exported
function that does not need it.

Reviewed-by: Josef Bacik <josef@toxicpanda.com>
Signed-off-by: David Sterba <dsterba@suse.com>
5 years agobtrfs: move btrfs_set_path_blocking to other locking functions
David Sterba [Tue, 24 Sep 2019 17:17:17 +0000 (19:17 +0200)]
btrfs: move btrfs_set_path_blocking to other locking functions

The function belongs to the family of locking functions, so move it
there. The 'noinline' keyword is dropped as it's now an exported
function that does not need it.

Reviewed-by: Josef Bacik <josef@toxicpanda.com>
Signed-off-by: David Sterba <dsterba@suse.com>
5 years agobtrfs: make btrfs_assert_tree_locked static inline
David Sterba [Tue, 24 Sep 2019 16:44:24 +0000 (18:44 +0200)]
btrfs: make btrfs_assert_tree_locked static inline

The function btrfs_assert_tree_locked is used outside of the locking
code so it is exported, however we can make it static inine as it's
fairly trivial.

This is the only locking assertion used in release builds, inlining
improves the text size by 174 bytes and reduces stack consumption in the
callers.

Reviewed-by: Josef Bacik <josef@toxicpanda.com>
Signed-off-by: David Sterba <dsterba@suse.com>
5 years agobtrfs: make locking assertion helpers static inline
David Sterba [Tue, 24 Sep 2019 16:29:10 +0000 (18:29 +0200)]
btrfs: make locking assertion helpers static inline

I've noticed that none of the btrfs_assert_*lock* debugging helpers is
inlined, despite they're short and mostly a value update. Making them
inline shaves 67 from the text size, reduces stack consumption and
perhaps also slightly improves the performance due to avoiding
unnecessary calls.

Reviewed-by: Josef Bacik <josef@toxicpanda.com>
Signed-off-by: David Sterba <dsterba@suse.com>
5 years agobtrfs: get rid of pointless wtag variable in async-thread.c
Omar Sandoval [Mon, 16 Sep 2019 18:30:58 +0000 (11:30 -0700)]
btrfs: get rid of pointless wtag variable in async-thread.c

Commit ac0c7cf8be00 ("btrfs: fix crash when tracepoint arguments are
freed by wq callbacks") added a void pointer, wtag, which is passed into
trace_btrfs_all_work_done() instead of the freed work item. This is
silly for a few reasons:

1. The freed work item still has the same address.
2. work is still in scope after it's freed, so assigning wtag doesn't
   stop anyone from using it.
3. The tracepoint has always taken a void * argument, so assigning wtag
   doesn't actually make things any more type-safe. (Note that the
   original bug in commit bc074524e123 ("btrfs: prefix fsid to all trace
   events") was that the void * was implicitly casted when it was passed
   to btrfs_work_owner() in the trace point itself).

Instead, let's add some clearer warnings as comments.

Reviewed-by: Nikolay Borisov <nborisov@suse.com>
Reviewed-by: Filipe Manana <fdmanana@suse.com>
Signed-off-by: Omar Sandoval <osandov@fb.com>
Reviewed-by: David Sterba <dsterba@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
5 years agobtrfs: get rid of unique workqueue helper functions
Omar Sandoval [Mon, 16 Sep 2019 18:30:57 +0000 (11:30 -0700)]
btrfs: get rid of unique workqueue helper functions

Commit 9e0af2376434 ("Btrfs: fix task hang under heavy compressed
write") worked around the issue that a recycled work item could get a
false dependency on the original work item due to how the workqueue code
guarantees non-reentrancy. It did so by giving different work functions
to different types of work.

However, the fixes in the previous few patches are more complete, as
they prevent a work item from being recycled at all (except for a tiny
window that the kernel workqueue code handles for us). This obsoletes
the previous fix, so we don't need the unique helpers for correctness.
The only other reason to keep them would be so they show up in stack
traces, but they always seem to be optimized to a tail call, so they
don't show up anyways. So, let's just get rid of the extra indirection.

While we're here, rename normal_work_helper() to the more informative
btrfs_work_helper().

Reviewed-by: Nikolay Borisov <nborisov@suse.com>
Reviewed-by: Filipe Manana <fdmanana@suse.com>
Signed-off-by: Omar Sandoval <osandov@fb.com>
Reviewed-by: David Sterba <dsterba@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
5 years agobtrfs: don't prematurely free work in scrub_missing_raid56_worker()
Omar Sandoval [Mon, 16 Sep 2019 18:30:56 +0000 (11:30 -0700)]
btrfs: don't prematurely free work in scrub_missing_raid56_worker()

Currently, scrub_missing_raid56_worker() puts and potentially frees
sblock (which embeds the work item) and then submits a bio through
scrub_wr_submit(). This is another potential instance of the bug in
"btrfs: don't prematurely free work in run_ordered_work()". Fix it by
dropping the reference after we submit the bio.

Reviewed-by: Johannes Thumshirn <jthumshirn@suse.de>
Signed-off-by: Omar Sandoval <osandov@fb.com>
Reviewed-by: David Sterba <dsterba@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
5 years agobtrfs: don't prematurely free work in reada_start_machine_worker()
Omar Sandoval [Mon, 16 Sep 2019 18:30:55 +0000 (11:30 -0700)]
btrfs: don't prematurely free work in reada_start_machine_worker()

Currently, reada_start_machine_worker() frees the reada_machine_work and
then calls __reada_start_machine() to do readahead. This is another
potential instance of the bug in "btrfs: don't prematurely free work in
run_ordered_work()".

There _might_ already be a deadlock here: reada_start_machine_worker()
can depend on itself through stacked filesystems (__read_start_machine()
-> reada_start_machine_dev() -> reada_tree_block_flagged() ->
read_extent_buffer_pages() -> submit_one_bio() ->
btree_submit_bio_hook() -> btrfs_map_bio() -> submit_stripe_bio() ->
submit_bio() onto a loop device can trigger readahead on the lower
filesystem).

Either way, let's fix it by freeing the work at the end.

Reviewed-by: Johannes Thumshirn <jthumshirn@suse.de>
Signed-off-by: Omar Sandoval <osandov@fb.com>
Reviewed-by: David Sterba <dsterba@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
5 years agobtrfs: don't prematurely free work in end_workqueue_fn()
Omar Sandoval [Mon, 16 Sep 2019 18:30:54 +0000 (11:30 -0700)]
btrfs: don't prematurely free work in end_workqueue_fn()

Currently, end_workqueue_fn() frees the end_io_wq entry (which embeds
the work item) and then calls bio_endio(). This is another potential
instance of the bug in "btrfs: don't prematurely free work in
run_ordered_work()".

In particular, the endio call may depend on other work items. For
example, btrfs_end_dio_bio() can call btrfs_subio_endio_read() ->
__btrfs_correct_data_nocsum() -> dio_read_error() ->
submit_dio_repair_bio(), which submits a bio that is also completed
through a end_workqueue_fn() work item. However,
__btrfs_correct_data_nocsum() waits for the newly submitted bio to
complete, thus it depends on another work item.

This example currently usually works because we use different workqueue
helper functions for BTRFS_WQ_ENDIO_DATA and BTRFS_WQ_ENDIO_DIO_REPAIR.
However, it may deadlock with stacked filesystems and is fragile
overall. The proper fix is to free the work item at the very end of the
work function, so let's do that.

Reviewed-by: Johannes Thumshirn <jthumshirn@suse.de>
Signed-off-by: Omar Sandoval <osandov@fb.com>
Reviewed-by: David Sterba <dsterba@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
5 years agobtrfs: don't prematurely free work in run_ordered_work()
Omar Sandoval [Mon, 16 Sep 2019 18:30:53 +0000 (11:30 -0700)]
btrfs: don't prematurely free work in run_ordered_work()

We hit the following very strange deadlock on a system with Btrfs on a
loop device backed by another Btrfs filesystem:

1. The top (loop device) filesystem queues an async_cow work item from
   cow_file_range_async(). We'll call this work X.
2. Worker thread A starts work X (normal_work_helper()).
3. Worker thread A executes the ordered work for the top filesystem
   (run_ordered_work()).
4. Worker thread A finishes the ordered work for work X and frees X
   (work->ordered_free()).
5. Worker thread A executes another ordered work and gets blocked on I/O
   to the bottom filesystem (still in run_ordered_work()).
6. Meanwhile, the bottom filesystem allocates and queues an async_cow
   work item which happens to be the recently-freed X.
7. The workqueue code sees that X is already being executed by worker
   thread A, so it schedules X to be executed _after_ worker thread A
   finishes (see the find_worker_executing_work() call in
   process_one_work()).

Now, the top filesystem is waiting for I/O on the bottom filesystem, but
the bottom filesystem is waiting for the top filesystem to finish, so we
deadlock.

This happens because we are breaking the workqueue assumption that a
work item cannot be recycled while it still depends on other work. Fix
it by waiting to free the work item until we are done with all of the
related ordered work.

P.S.:

One might ask why the workqueue code doesn't try to detect a recycled
work item. It actually does try by checking whether the work item has
the same work function (find_worker_executing_work()), but in our case
the function is the same. This is the only key that the workqueue code
has available to compare, short of adding an additional, layer-violating
"custom key". Considering that we're the only ones that have ever hit
this, we should just play by the rules.

Unfortunately, we haven't been able to create a minimal reproducer other
than our full container setup using a compress-force=zstd filesystem on
top of another compress-force=zstd filesystem.

Suggested-by: Tejun Heo <tj@kernel.org>
Reviewed-by: Johannes Thumshirn <jthumshirn@suse.de>
Signed-off-by: Omar Sandoval <osandov@fb.com>
Reviewed-by: David Sterba <dsterba@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
5 years agobtrfs: get rid of unnecessary memset() of work item
Omar Sandoval [Mon, 16 Sep 2019 18:30:52 +0000 (11:30 -0700)]
btrfs: get rid of unnecessary memset() of work item

Commit fc97fab0ea59 ("btrfs: Replace fs_info->qgroup_rescan_worker
workqueue with btrfs_workqueue.") converted qgroup_rescan_work to be
initialized with btrfs_init_work(), but it left behind an unnecessary
memset(). Get rid of the memset().

Reviewed-by: Johannes Thumshirn <jthumshirn@suse.de>
Signed-off-by: Omar Sandoval <osandov@fb.com>
Reviewed-by: David Sterba <dsterba@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
5 years agobtrfs: move the failrec tree stuff into extent-io-tree.h
Josef Bacik [Mon, 23 Sep 2019 14:05:21 +0000 (10:05 -0400)]
btrfs: move the failrec tree stuff into extent-io-tree.h

This needs to be cleaned up in the future, but for now it belongs to the
extent-io-tree stuff since it uses the internal tree search code.
Needed to export get_state_failrec and set_state_failrec as well since
we're not going to move the actual IO part of the failrec stuff out at
this point.

Signed-off-by: Josef Bacik <josef@toxicpanda.com>
Reviewed-by: David Sterba <dsterba@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
5 years agobtrfs: export find_delalloc_range
Josef Bacik [Mon, 23 Sep 2019 14:05:20 +0000 (10:05 -0400)]
btrfs: export find_delalloc_range

This utilizes internal stuff to the extent_io_tree, so we need to export
it before we move it.

Signed-off-by: Josef Bacik <josef@toxicpanda.com>
Reviewed-by: David Sterba <dsterba@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
5 years agobtrfs: move extent_io_tree defs to their own header
Josef Bacik [Mon, 23 Sep 2019 14:05:19 +0000 (10:05 -0400)]
btrfs: move extent_io_tree defs to their own header

extent_io.c/h are huge, encompassing a bunch of different things.  The
extent_io_tree code can live on its own, so separate this out.

Signed-off-by: Josef Bacik <josef@toxicpanda.com>
Reviewed-by: David Sterba <dsterba@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
5 years agobtrfs: separate out the extent io init function
Josef Bacik [Mon, 23 Sep 2019 14:05:18 +0000 (10:05 -0400)]
btrfs: separate out the extent io init function

We are moving extent_io_tree into it's on file, so separate out the
extent_state init stuff from extent_io_tree_init().

Signed-off-by: Josef Bacik <josef@toxicpanda.com>
Reviewed-by: David Sterba <dsterba@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
5 years agobtrfs: separate out the extent leak code
Josef Bacik [Mon, 23 Sep 2019 14:05:17 +0000 (10:05 -0400)]
btrfs: separate out the extent leak code

We check both extent buffer and extent state leaks in the same function,
separate these two functions out so we can move them around.

Signed-off-by: Josef Bacik <josef@toxicpanda.com>
Reviewed-by: David Sterba <dsterba@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>