OSDN Git Service

tomoyo/tomoyo-test1.git
3 years agobtrfs: remove fail label in check_compressed_csum
Nikolay Borisov [Thu, 2 Jul 2020 13:46:47 +0000 (16:46 +0300)]
btrfs: remove fail label in check_compressed_csum

Signed-off-by: Nikolay Borisov <nborisov@suse.com>
Reviewed-by: David Sterba <dsterba@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
3 years agobtrfs: raid56: don't opencode swap() in __raid_recover_end_io
Nikolay Borisov [Thu, 2 Jul 2020 13:46:46 +0000 (16:46 +0300)]
btrfs: raid56: don't opencode swap() in __raid_recover_end_io

Reviewed-by: Johannes Thumshirn <johannes.thumshirn@wdc.com>
Signed-off-by: Nikolay Borisov <nborisov@suse.com>
Reviewed-by: David Sterba <dsterba@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
3 years agobtrfs: raid56: use in_range where applicable
Nikolay Borisov [Thu, 2 Jul 2020 13:46:45 +0000 (16:46 +0300)]
btrfs: raid56: use in_range where applicable

While at it use the opportunity to simplify find_logical_bio_stripe by
reducing the scope of 'stripe_start' variable and squash the
sector-to-bytes conversion on one line.

Reviewed-by: Johannes Thumshirn <johannes.thumshirn@wdc.com>
Signed-off-by: Nikolay Borisov <nborisov@suse.com>
Reviewed-by: David Sterba <dsterba@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
3 years agobtrfs: raid56: assign bio in while() when using bio_list_pop
Nikolay Borisov [Thu, 2 Jul 2020 13:46:43 +0000 (16:46 +0300)]
btrfs: raid56: assign bio in while() when using bio_list_pop

Unify the style in the file such that return value of bio_list_pop is
assigned directly in the while loop. This is in line with the rest of
the kernel.

Reviewed-by: Johannes Thumshirn <johannes.thumshirn@wdc.com>
Signed-off-by: Nikolay Borisov <nborisov@suse.com>
Reviewed-by: David Sterba <dsterba@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
3 years agobtrfs: raid56: remove redundant device check in rbio_add_io_page
Nikolay Borisov [Thu, 2 Jul 2020 13:46:42 +0000 (16:46 +0300)]
btrfs: raid56: remove redundant device check in rbio_add_io_page

The merging logic is always executed if the current stripe's device
is not missing. So there's no point in duplicating the check. Simply
remove it, while at it reduce the scope of the 'last_end' variable.
If the current stripe's device is missing we fail the stripe early on.

Reviewed-by: Johannes Thumshirn <johannes.thumshirn@wdc.com>
Signed-off-by: Nikolay Borisov <nborisov@suse.com>
Reviewed-by: David Sterba <dsterba@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
3 years agobtrfs: always initialize btrfs_bio::tgtdev_map/raid_map pointers
Nikolay Borisov [Thu, 2 Jul 2020 13:46:41 +0000 (16:46 +0300)]
btrfs: always initialize btrfs_bio::tgtdev_map/raid_map pointers

Since btrfs_bio always contains the extra space for the tgtdev_map and
raid_maps it's pointless to make the assignment iff specific conditions
are met.

Instead, always assign the pointers to their correct value at allocation
time. To accommodate this change also move code a bit in
__btrfs_map_block so that btrfs_bio::stripes array is always initialized
before the raid_map, subsequently move the call to sort_parity_stripes
in the 'if' building the raid_map, retaining the old behavior.

To better understand the change, there are 2 aspects to this:

1. The original code is harder to grasp because the calculations for
   initializing raid_map/tgtdev ponters are apart from the initial
   allocation of memory. Having them predicated on 2 separate checks
   doesn't help that either... So by moving the initialisation in
   alloc_btrfs_bio puts everything together.

2. tgtdev/raid_maps are now always initialized despite sometimes they
   might be equal i.e __btrfs_map_block_for_discard calls
   alloc_btrfs_bio with tgtdev = 0 but their usage should be predicated
   on external checks i.e. just because those pointers are non-null
   doesn't mean they are valid per-se. And actually while taking another
   look at __btrfs_map_block I saw a discrepancy:

   Original code initialised tgtdev_map if the following check is true:

   if (dev_replace_is_ongoing && dev_replace->tgtdev != NULL)

   However, further down tgtdev_map is only used if the following check
   is true:

if (dev_replace_is_ongoing && dev_replace->tgtdev != NULL && need_full_stripe(op))

  e.g. the additional need_full_stripe(op) predicate is there.

Signed-off-by: Nikolay Borisov <nborisov@suse.com>
Reviewed-by: David Sterba <dsterba@suse.com>
[ copy more details from mail discussion ]
Signed-off-by: David Sterba <dsterba@suse.com>
3 years agobtrfs: sysfs: add bdi link to the fsid directory
Nikolay Borisov [Fri, 3 Jul 2020 08:13:15 +0000 (11:13 +0300)]
btrfs: sysfs: add bdi link to the fsid directory

Since BTRFS uses a private bdi it makes sense to create a link to this
bdi under /sys/fs/btrfs/<UUID>/bdi. This allows size of read ahead to
be controlled. Without this patch it's not possible to uniquely identify
which bdi pertains to which btrfs filesystem in the case of multiple
btrfs filesystems.

It's fine to simply call sysfs_remove_link without checking if the
link indeed has been created. The call path

sysfs_remove_link
 kernfs_remove_by_name
  kernfs_remove_by_name_ns

will simply return -ENOENT in case it doesn't exist.

Signed-off-by: Nikolay Borisov <nborisov@suse.com>
Reviewed-by: David Sterba <dsterba@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
3 years agobtrfs: increment corrupt device counter during compressed read
Nikolay Borisov [Thu, 2 Jul 2020 12:23:34 +0000 (15:23 +0300)]
btrfs: increment corrupt device counter during compressed read

If a compressed read fails due to checksum error only a line is printed
to dmesg, device corrupt counter is not modified.

Reviewed-by: Josef Bacik <josef@toxicpanda.com>
Reviewed-by: Johannes Thumshirn <johannes.thumshirn@wdc.com>
Signed-off-by: Nikolay Borisov <nborisov@suse.com>
Reviewed-by: David Sterba <dsterba@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
3 years agobtrfs: remove needless ASSERT check of orig_bio in end_compressed_bio_read
Nikolay Borisov [Thu, 2 Jul 2020 12:23:33 +0000 (15:23 +0300)]
btrfs: remove needless ASSERT check of orig_bio in end_compressed_bio_read

compressed_bio::orig_bio is always set in btrfs_submit_compressed_read
before any bio submission is performed. Since that function is always
called with a valid bio it renders the ASSERT unnecessary.

Reviewed-by: Josef Bacik <josef@toxicpanda.com>
Reviewed-by: Johannes Thumshirn <johannes.thumshirn@wdc.com>
Signed-off-by: Nikolay Borisov <nborisov@suse.com>
Reviewed-by: David Sterba <dsterba@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
3 years agobtrfs: increment device corruption error in case of checksum error
Nikolay Borisov [Thu, 2 Jul 2020 12:23:32 +0000 (15:23 +0300)]
btrfs: increment device corruption error in case of checksum error

Now that btrfs_io_bio have access to btrfs_device we can safely
increment the device corruption counter on error. There is one notable
exception - repair bios for raid. Since those don't go through the
normal submit_stripe_bio callpath but through raid56_parity_recover thus
repair bios won't have their device set.

Scrub increments the corruption counter for checksum mismatch as well
but does not call this function.

Link: https://lore.kernel.org/linux-btrfs/4857863.FCrPRfMyHP@liv/
Reviewed-by: Josef Bacik <josef@toxicpanda.com>
Signed-off-by: Nikolay Borisov <nborisov@suse.com>
Reviewed-by: David Sterba <dsterba@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
3 years agobtrfs: don't check for btrfs_device::bdev in btrfs_end_bio
Nikolay Borisov [Thu, 2 Jul 2020 12:23:31 +0000 (15:23 +0300)]
btrfs: don't check for btrfs_device::bdev in btrfs_end_bio

btrfs_map_bio ensures that all submitted bios to devices have valid
btrfs_device::bdev so this check can be removed from btrfs_end_bio. This
check was added in june 2012 597a60fadedf ("Btrfs: don't count I/O
statistic read errors for missing devices") but then in October of the
same year another commit de1ee92ac3bc ("Btrfs: recheck bio against
block device when we map the bio") started checking for the presence of
btrfs_device::bdev before actually issuing the bio.

Reviewed-by: Josef Bacik <josef@toxicpanda.com>
Signed-off-by: Nikolay Borisov <nborisov@suse.com>
Reviewed-by: David Sterba <dsterba@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
3 years agobtrfs: record btrfs_device directly in btrfs_io_bio
Nikolay Borisov [Fri, 3 Jul 2020 08:14:27 +0000 (11:14 +0300)]
btrfs: record btrfs_device directly in btrfs_io_bio

Instead of recording stripe_index and using that to access correct
btrfs_device from btrfs_bio::stripes record the btrfs_device in
btrfs_io_bio. This will enable endio handlers to increment device
error counters on checksum errors.

Reviewed-by: Johannes Thumshirn <johannes.thumshirn@wdc.com>
Signed-off-by: Nikolay Borisov <nborisov@suse.com>
Reviewed-by: David Sterba <dsterba@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
3 years agobtrfs: streamline btrfs_get_io_failure_record logic
Nikolay Borisov [Thu, 2 Jul 2020 12:23:29 +0000 (15:23 +0300)]
btrfs: streamline btrfs_get_io_failure_record logic

Make the function directly return a pointer to a failure record and
adjust callers to handle it. Also refactor the logic inside so that
the case which allocates the failure record for the first time is not
handled in an 'if' arm, saving us a level of indentation. Finally make
the function static as it's not used outside of extent_io.c .

Reviewed-by: Josef Bacik <josef@toxicpanda.com>
Signed-off-by: Nikolay Borisov <nborisov@suse.com>
Reviewed-by: David Sterba <dsterba@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
3 years agobtrfs: make get_state_failrec return failrec directly
Nikolay Borisov [Thu, 2 Jul 2020 12:23:28 +0000 (15:23 +0300)]
btrfs: make get_state_failrec return failrec directly

Only failure that get_state_failrec can get is if there is no failure
for the given address. There is no reason why the function should return
a status code and use a separate parameter for returning the actual
failure rec (if one is found). Simplify it by making the return type
a pointer and return ERR_PTR value in case of errors.

Signed-off-by: Nikolay Borisov <nborisov@suse.com>
Reviewed-by: David Sterba <dsterba@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
3 years agobtrfs: remove deprecated mount option subvolrootid
David Sterba [Wed, 1 Jul 2020 15:08:43 +0000 (17:08 +0200)]
btrfs: remove deprecated mount option subvolrootid

The option subvolrootid used to be a workaround for mounting subvolumes
and ineffective since 5e2a4b25da23 ("btrfs: deprecate subvolrootid mount
option"). We have subvol= that works and we don't need to keep the
cruft, let's remove it.

Signed-off-by: David Sterba <dsterba@suse.com>
3 years agobtrfs: remove deprecated mount option alloc_start
David Sterba [Wed, 1 Jul 2020 15:02:34 +0000 (17:02 +0200)]
btrfs: remove deprecated mount option alloc_start

The mount option alloc_start has no effect since 0d0c71b31720 ("btrfs:
obsolete and remove mount option alloc_start") which has details why
it's been deprecated. We can remove it.

Signed-off-by: David Sterba <dsterba@suse.com>
3 years agobtrfs: remove no longer needed use of log_writers for the log root tree
Filipe Manana [Thu, 2 Jul 2020 11:32:40 +0000 (12:32 +0100)]
btrfs: remove no longer needed use of log_writers for the log root tree

When syncing the log, we used to update the log root tree without holding
neither the log_mutex of the subvolume root nor the log_mutex of log root
tree.

We used to have two critical sections delimited by the log_mutex of the
log root tree, so in the first one we incremented the log_writers of the
log root tree and on the second one we decremented it and waited for the
log_writers counter to go down to zero. This was because the update of
the log root tree happened between the two critical sections.

The use of two critical sections allowed a little bit more of parallelism
and required the use of the log_writers counter, necessary to make sure
we didn't miss any log root tree update when we have multiple tasks trying
to sync the log in parallel.

However after commit 06989c799f0481 ("Btrfs: fix race updating log root
item during fsync") the log root tree update was moved into a critical
section delimited by the subvolume's log_mutex. Later another commit
moved the log tree update from that critical section into the second
critical section delimited by the log_mutex of the log root tree. Both
commits addressed different bugs.

The end result is that the first critical section delimited by the
log_mutex of the log root tree became pointless, since there's nothing
done between it and the second critical section, we just have an unlock
of the log_mutex followed by a lock operation. This means we can merge
both critical sections, as the first one does almost nothing now, and we
can stop using the log_writers counter of the log root tree, which was
incremented in the first critical section and decremented in the second
criticial section, used to make sure no one in the second critical section
started writeback of the log root tree before some other task updated it.

So just remove the mutex_unlock() followed by mutex_lock() of the log root
tree, as well as the use of the log_writers counter for the log root tree.

This patch is part of a series that has the following patches:

1/4 btrfs: only commit the delayed inode when doing a full fsync
2/4 btrfs: only commit delayed items at fsync if we are logging a directory
3/4 btrfs: stop incremening log_batch for the log root tree when syncing log
4/4 btrfs: remove no longer needed use of log_writers for the log root tree

After the entire patchset applied I saw about 12% decrease on max latency
reported by dbench. The test was done on a qemu vm, with 8 cores, 16Gb of
ram, using kvm and using a raw NVMe device directly (no intermediary fs on
the host). The test was invoked like the following:

  mkfs.btrfs -f /dev/sdk
  mount -o ssd -o nospace_cache /dev/sdk /mnt/sdk
  dbench -D /mnt/sdk -t 300 8
  umount /mnt/dsk

CC: stable@vger.kernel.org # 5.4+
Reviewed-by: Josef Bacik <josef@toxicpanda.com>
Signed-off-by: Filipe Manana <fdmanana@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
3 years agobtrfs: stop incremening log_batch for the log root tree when syncing log
Filipe Manana [Thu, 2 Jul 2020 11:32:31 +0000 (12:32 +0100)]
btrfs: stop incremening log_batch for the log root tree when syncing log

We are incrementing the log_batch atomic counter of the root log tree but
we never use that counter, it's used only for the log trees of subvolume
roots. We started doing it when we moved the log_batch and log_write
counters from the global, per fs, btrfs_fs_info structure, into the
btrfs_root structure in commit 7237f1833601dc ("Btrfs: fix tree logs
parallel sync").

So just stop doing it for the log root tree and add a comment over the
field declaration so inform it's used only for log trees of subvolume
roots.

This patch is part of a series that has the following patches:

1/4 btrfs: only commit the delayed inode when doing a full fsync
2/4 btrfs: only commit delayed items at fsync if we are logging a directory
3/4 btrfs: stop incremening log_batch for the log root tree when syncing log
4/4 btrfs: remove no longer needed use of log_writers for the log root tree

After the entire patchset applied I saw about 12% decrease on max latency
reported by dbench. The test was done on a qemu vm, with 8 cores, 16Gb of
ram, using kvm and using a raw NVMe device directly (no intermediary fs on
the host). The test was invoked like the following:

  mkfs.btrfs -f /dev/sdk
  mount -o ssd -o nospace_cache /dev/sdk /mnt/sdk
  dbench -D /mnt/sdk -t 300 8
  umount /mnt/dsk

CC: stable@vger.kernel.org # 5.4+
Reviewed-by: Josef Bacik <josef@toxicpanda.com>
Signed-off-by: Filipe Manana <fdmanana@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
3 years agobtrfs: only commit delayed items at fsync if we are logging a directory
Filipe Manana [Thu, 2 Jul 2020 11:32:20 +0000 (12:32 +0100)]
btrfs: only commit delayed items at fsync if we are logging a directory

When logging an inode we are committing its delayed items if either the
inode is a directory or if it is a new inode, created in the current
transaction.

We need to do it for directories, since new directory indexes are stored
as delayed items of the inode and when logging a directory we need to be
able to access all indexes from the fs/subvolume tree in order to figure
out which index ranges need to be logged.

However for new inodes that are not directories, we do not need to do it
because the only type of delayed item they can have is the inode item, and
we are guaranteed to always log an up to date version of the inode item:

*) for a full fsync we do it by committing the delayed inode and then
   copying the item from the fs/subvolume tree with
   copy_inode_items_to_log();

*) for a fast fsync we always log the inode item based on the contents of
   the in-memory struct btrfs_inode. We guarantee this is always done since
   commit e4545de5b035c7 ("Btrfs: fix fsync data loss after append write").

So stop running delayed items for a new inodes that are not directories,
since that forces committing the delayed inode into the fs/subvolume tree,
wasting time and adding contention to the tree when a full fsync is not
required. We will only do it in case a fast fsync is needed.

This patch is part of a series that has the following patches:

1/4 btrfs: only commit the delayed inode when doing a full fsync
2/4 btrfs: only commit delayed items at fsync if we are logging a directory
3/4 btrfs: stop incremening log_batch for the log root tree when syncing log
4/4 btrfs: remove no longer needed use of log_writers for the log root tree

After the entire patchset applied I saw about 12% decrease on max latency
reported by dbench. The test was done on a qemu vm, with 8 cores, 16Gb of
ram, using kvm and using a raw NVMe device directly (no intermediary fs on
the host). The test was invoked like the following:

  mkfs.btrfs -f /dev/sdk
  mount -o ssd -o nospace_cache /dev/sdk /mnt/sdk
  dbench -D /mnt/sdk -t 300 8
  umount /mnt/dsk

CC: stable@vger.kernel.org # 5.4+
Reviewed-by: Josef Bacik <josef@toxicpanda.com>
Signed-off-by: Filipe Manana <fdmanana@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
3 years agobtrfs: only commit the delayed inode when doing a full fsync
Filipe Manana [Thu, 2 Jul 2020 11:31:59 +0000 (12:31 +0100)]
btrfs: only commit the delayed inode when doing a full fsync

Commit 2c2c452b0cafdc ("Btrfs: fix fsync when extend references are added
to an inode") forced a commit of the delayed inode when logging an inode
in order to ensure we would end up logging the inode item during a full
fsync. By committing the delayed inode, we updated the inode item in the
fs/subvolume tree and then later when copying items from leafs modified in
the current transaction into the log tree (with copy_inode_items_to_log())
we ended up copying the inode item from the fs/subvolume tree into the log
tree. Logging an up to date version of the inode item is required to make
sure at log replay time we get the link count fixup triggered among other
things (replay xattr deletes, etc). The test case generic/040 from fstests
exercises the bug which that commit fixed.

However for a fast fsync we don't need to commit the delayed inode because
we always log an up to date version of the inode item based on the struct
btrfs_inode we have in-memory. We started doing this for fast fsyncs since
commit e4545de5b035c7 ("Btrfs: fix fsync data loss after append write").

So just stop committing the delayed inode if we are doing a fast fsync,
we are only wasting time and adding contention on fs/subvolume tree.

This patch is part of a series that has the following patches:

1/4 btrfs: only commit the delayed inode when doing a full fsync
2/4 btrfs: only commit delayed items at fsync if we are logging a directory
3/4 btrfs: stop incremening log_batch for the log root tree when syncing log
4/4 btrfs: remove no longer needed use of log_writers for the log root tree

After the entire patchset applied I saw about 12% decrease on max latency
reported by dbench. The test was done on a qemu vm, with 8 cores, 16Gb of
ram, using kvm and using a raw NVMe device directly (no intermediary fs on
the host). The test was invoked like the following:

  mkfs.btrfs -f /dev/sdk
  mount -o ssd -o nospace_cache /dev/sdk /mnt/sdk
  dbench -D /mnt/sdk -t 300 8
  umount /mnt/dsk

CC: stable@vger.kernel.org # 5.4+
Reviewed-by: Josef Bacik <josef@toxicpanda.com>
Signed-off-by: Filipe Manana <fdmanana@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
3 years agobtrfs: preallocate anon block device at first phase of snapshot creation
Qu Wenruo [Tue, 16 Jun 2020 02:17:36 +0000 (10:17 +0800)]
btrfs: preallocate anon block device at first phase of snapshot creation

[BUG]
When the anonymous block device pool is exhausted, subvolume/snapshot
creation fails with EMFILE (Too many files open). This has been reported
by a user. The allocation happens in the second phase during transaction
commit where it's only way out is to abort the transaction

  BTRFS: Transaction aborted (error -24)
  WARNING: CPU: 17 PID: 17041 at fs/btrfs/transaction.c:1576 create_pending_snapshot+0xbc4/0xd10 [btrfs]
  RIP: 0010:create_pending_snapshot+0xbc4/0xd10 [btrfs]
  Call Trace:
   create_pending_snapshots+0x82/0xa0 [btrfs]
   btrfs_commit_transaction+0x275/0x8c0 [btrfs]
   btrfs_mksubvol+0x4b9/0x500 [btrfs]
   btrfs_ioctl_snap_create_transid+0x174/0x180 [btrfs]
   btrfs_ioctl_snap_create_v2+0x11c/0x180 [btrfs]
   btrfs_ioctl+0x11a4/0x2da0 [btrfs]
   do_vfs_ioctl+0xa9/0x640
   ksys_ioctl+0x67/0x90
   __x64_sys_ioctl+0x1a/0x20
   do_syscall_64+0x5a/0x110
   entry_SYSCALL_64_after_hwframe+0x44/0xa9
  ---[ end trace 33f2f83f3d5250e9 ]---
  BTRFS: error (device sda1) in create_pending_snapshot:1576: errno=-24 unknown
  BTRFS info (device sda1): forced readonly
  BTRFS warning (device sda1): Skipping commit of aborted transaction.
  BTRFS: error (device sda1) in cleanup_transaction:1831: errno=-24 unknown

[CAUSE]
When the global anonymous block device pool is exhausted, the following
call chain will fail, and lead to transaction abort:

 btrfs_ioctl_snap_create_v2()
 |- btrfs_ioctl_snap_create_transid()
    |- btrfs_mksubvol()
       |- btrfs_commit_transaction()
          |- create_pending_snapshot()
             |- btrfs_get_fs_root()
                |- btrfs_init_fs_root()
                   |- get_anon_bdev()

[FIX]
Although we can't enlarge the anonymous block device pool, at least we
can preallocate anon_dev for subvolume/snapshot in the first phase,
outside of transaction context and exactly at the moment the user calls
the creation ioctl.

Reported-by: Greed Rong <greedrong@gmail.com>
Link: https://lore.kernel.org/linux-btrfs/CA+UqX+NTrZ6boGnWHhSeZmEY5J76CTqmYjO2S+=tHJX7nb9DPw@mail.gmail.com/
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>
3 years agobtrfs: free anon block device right after subvolume deletion
Qu Wenruo [Tue, 16 Jun 2020 02:17:37 +0000 (10:17 +0800)]
btrfs: free anon block device right after subvolume deletion

[BUG]
When a lot of subvolumes are created, there is a user report about
transaction aborted caused by slow anonymous block device reclaim:

  BTRFS: Transaction aborted (error -24)
  WARNING: CPU: 17 PID: 17041 at fs/btrfs/transaction.c:1576 create_pending_snapshot+0xbc4/0xd10 [btrfs]
  RIP: 0010:create_pending_snapshot+0xbc4/0xd10 [btrfs]
  Call Trace:
   create_pending_snapshots+0x82/0xa0 [btrfs]
   btrfs_commit_transaction+0x275/0x8c0 [btrfs]
   btrfs_mksubvol+0x4b9/0x500 [btrfs]
   btrfs_ioctl_snap_create_transid+0x174/0x180 [btrfs]
   btrfs_ioctl_snap_create_v2+0x11c/0x180 [btrfs]
   btrfs_ioctl+0x11a4/0x2da0 [btrfs]
   do_vfs_ioctl+0xa9/0x640
   ksys_ioctl+0x67/0x90
   __x64_sys_ioctl+0x1a/0x20
   do_syscall_64+0x5a/0x110
   entry_SYSCALL_64_after_hwframe+0x44/0xa9
  ---[ end trace 33f2f83f3d5250e9 ]---
  BTRFS: error (device sda1) in create_pending_snapshot:1576: errno=-24 unknown
  BTRFS info (device sda1): forced readonly
  BTRFS warning (device sda1): Skipping commit of aborted transaction.
  BTRFS: error (device sda1) in cleanup_transaction:1831: errno=-24 unknown

[CAUSE]
The anonymous device pool is shared and its size is 1M. It's possible to
hit that limit if the subvolume deletion is not fast enough and the
subvolumes to be cleaned keep the ids allocated.

[WORKAROUND]
We can't avoid the anon device pool exhaustion but we can shorten the
time the id is attached to the subvolume root once the subvolume becomes
invisible to the user.

Reported-by: Greed Rong <greedrong@gmail.com>
Link: https://lore.kernel.org/linux-btrfs/CA+UqX+NTrZ6boGnWHhSeZmEY5J76CTqmYjO2S+=tHJX7nb9DPw@mail.gmail.com/
CC: stable@vger.kernel.org # 4.4+
Reviewed-by: Josef Bacik <josef@toxicpanda.com>
Signed-off-by: Qu Wenruo <wqu@suse.com>
Reviewed-by: David Sterba <dsterba@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
3 years agobtrfs: don't allocate anonymous block device for user invisible roots
Qu Wenruo [Tue, 16 Jun 2020 02:17:34 +0000 (10:17 +0800)]
btrfs: don't allocate anonymous block device for user invisible roots

[BUG]
When a lot of subvolumes are created, there is a user report about
transaction aborted:

  BTRFS: Transaction aborted (error -24)
  WARNING: CPU: 17 PID: 17041 at fs/btrfs/transaction.c:1576 create_pending_snapshot+0xbc4/0xd10 [btrfs]
  RIP: 0010:create_pending_snapshot+0xbc4/0xd10 [btrfs]
  Call Trace:
   create_pending_snapshots+0x82/0xa0 [btrfs]
   btrfs_commit_transaction+0x275/0x8c0 [btrfs]
   btrfs_mksubvol+0x4b9/0x500 [btrfs]
   btrfs_ioctl_snap_create_transid+0x174/0x180 [btrfs]
   btrfs_ioctl_snap_create_v2+0x11c/0x180 [btrfs]
   btrfs_ioctl+0x11a4/0x2da0 [btrfs]
   do_vfs_ioctl+0xa9/0x640
   ksys_ioctl+0x67/0x90
   __x64_sys_ioctl+0x1a/0x20
   do_syscall_64+0x5a/0x110
   entry_SYSCALL_64_after_hwframe+0x44/0xa9
  ---[ end trace 33f2f83f3d5250e9 ]---
  BTRFS: error (device sda1) in create_pending_snapshot:1576: errno=-24 unknown
  BTRFS info (device sda1): forced readonly
  BTRFS warning (device sda1): Skipping commit of aborted transaction.
  BTRFS: error (device sda1) in cleanup_transaction:1831: errno=-24 unknown

[CAUSE]
The error is EMFILE (Too many files open) and comes from the anonymous
block device allocation. The ids are in a shared pool of size 1<<20.

The ids are assigned to live subvolumes, ie. the root structure exists
in memory (eg. after creation or after the root appears in some path).
The pool could be exhausted if the numbers are not reclaimed fast
enough, after subvolume deletion or if other system component uses the
anon block devices.

[WORKAROUND]
Since it's not possible to completely solve the problem, we can only
minimize the time the id is allocated to a subvolume root.

Firstly, we can reduce the use of anon_dev by trees that are not
subvolume roots, like data reloc tree.

This patch will do extra check on root objectid, to skip roots that
don't need anon_dev.  Currently it's only data reloc tree and orphan
roots.

Reported-by: Greed Rong <greedrong@gmail.com>
Link: https://lore.kernel.org/linux-btrfs/CA+UqX+NTrZ6boGnWHhSeZmEY5J76CTqmYjO2S+=tHJX7nb9DPw@mail.gmail.com/
CC: stable@vger.kernel.org # 4.4+
Reviewed-by: Josef Bacik <josef@toxicpanda.com>
Signed-off-by: Qu Wenruo <wqu@suse.com>
Reviewed-by: David Sterba <dsterba@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
3 years agobtrfs: qgroup: export qgroups in sysfs
Qu Wenruo [Sun, 28 Jun 2020 05:07:15 +0000 (13:07 +0800)]
btrfs: qgroup: export qgroups in sysfs

This patch will add the following sysfs interface:

  /sys/fs/btrfs/<UUID>/qgroups/<qgroup_id>/referenced
  /sys/fs/btrfs/<UUID>/qgroups/<qgroup_id>/exclusive
  /sys/fs/btrfs/<UUID>/qgroups/<qgroup_id>/max_referenced
  /sys/fs/btrfs/<UUID>/qgroups/<qgroup_id>/max_exclusive
  /sys/fs/btrfs/<UUID>/qgroups/<qgroup_id>/limit_flags

Which is also available in output of "btrfs qgroup show".

  /sys/fs/btrfs/<UUID>/qgroups/<qgroup_id>/rsv_data
  /sys/fs/btrfs/<UUID>/qgroups/<qgroup_id>/rsv_meta_pertrans
  /sys/fs/btrfs/<UUID>/qgroups/<qgroup_id>/rsv_meta_prealloc

The last 3 rsv related members are not visible to users, but can be very
useful to debug qgroup limit related bugs.

Also, to avoid '/' used in <qgroup_id>, the separator between qgroup
level and qgroup id is changed to '_'.

The interface is not hidden behind 'debug' as we want this interface to
be included into production build and to provide another way to read the
qgroup information besides the ioctls.

Signed-off-by: Qu Wenruo <wqu@suse.com>
Reviewed-by: David Sterba <dsterba@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
3 years agobtrfs: use __u16 for the return value of btrfs_qgroup_level()
Qu Wenruo [Sun, 28 Jun 2020 05:07:14 +0000 (13:07 +0800)]
btrfs: use __u16 for the return value of btrfs_qgroup_level()

The qgroup level is limited to u16, so no need to use u64 for it.

Signed-off-by: Qu Wenruo <wqu@suse.com>
Reviewed-by: David Sterba <dsterba@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
3 years agobtrfs: make btrfs_qgroup_check_reserved_leak take btrfs_inode
Nikolay Borisov [Wed, 3 Jun 2020 05:55:46 +0000 (08:55 +0300)]
btrfs: make btrfs_qgroup_check_reserved_leak take btrfs_inode

vfs_inode is used only for the inode number everything else requires
btrfs_inode.

Signed-off-by: Nikolay Borisov <nborisov@suse.com>
Reviewed-by: David Sterba <dsterba@suse.com>
[ use btrfs_ino ]
Signed-off-by: David Sterba <dsterba@suse.com>
3 years agobtrfs: make btrfs_set_inode_last_trans take btrfs_inode
Nikolay Borisov [Fri, 5 Jun 2020 07:41:13 +0000 (10:41 +0300)]
btrfs: make btrfs_set_inode_last_trans take btrfs_inode

Instead of making multiple calls to BTRFS_I simply take btrfs_inode as
an input paramter.

Signed-off-by: Nikolay Borisov <nborisov@suse.com>
Reviewed-by: David Sterba <dsterba@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
3 years agobtrfs: make prealloc_file_extent_cluster take btrfs_inode
Nikolay Borisov [Wed, 3 Jun 2020 05:55:44 +0000 (08:55 +0300)]
btrfs: make prealloc_file_extent_cluster take btrfs_inode

The vfs inode is only used for a pair of inode_lock/unlock calls all
other uses call for btrfs_inode.

Signed-off-by: Nikolay Borisov <nborisov@suse.com>
Reviewed-by: David Sterba <dsterba@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
3 years agobtrfs: remove BTRFS_I calls in btrfs_writepage_fixup_worker
Nikolay Borisov [Fri, 5 Jun 2020 07:51:51 +0000 (10:51 +0300)]
btrfs: remove BTRFS_I calls in btrfs_writepage_fixup_worker

All of its children functions use btrfs_inode.

Signed-off-by: Nikolay Borisov <nborisov@suse.com>
Reviewed-by: David Sterba <dsterba@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
3 years agobtrfs: make btrfs_delalloc_reserve_space take btrfs_inode
Nikolay Borisov [Wed, 3 Jun 2020 05:55:42 +0000 (08:55 +0300)]
btrfs: make btrfs_delalloc_reserve_space take btrfs_inode

All of its children take btrfs_inode so bubble up this requirement to
btrfs_delalloc_reserve_space's interface and stop calling BTRFS_I
internally.

Signed-off-by: Nikolay Borisov <nborisov@suse.com>
Reviewed-by: David Sterba <dsterba@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
3 years agobtrfs: make btrfs_check_data_free_space take btrfs_inode
Nikolay Borisov [Wed, 3 Jun 2020 05:55:41 +0000 (08:55 +0300)]
btrfs: make btrfs_check_data_free_space take btrfs_inode

Instead of calling BTRFS_I on the passed vfs_inode take btrfs_inode
directly.

Signed-off-by: Nikolay Borisov <nborisov@suse.com>
Reviewed-by: David Sterba <dsterba@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
3 years agobtrfs: make btrfs_delalloc_release_space take btrfs_inode
Nikolay Borisov [Wed, 3 Jun 2020 05:55:40 +0000 (08:55 +0300)]
btrfs: make btrfs_delalloc_release_space take btrfs_inode

It needs btrfs_inode so take it as a parameter directly.

Signed-off-by: Nikolay Borisov <nborisov@suse.com>
Reviewed-by: David Sterba <dsterba@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
3 years agobtrfs: make btrfs_free_reserved_data_space take btrfs_inode
Nikolay Borisov [Wed, 3 Jun 2020 05:55:39 +0000 (08:55 +0300)]
btrfs: make btrfs_free_reserved_data_space take btrfs_inode

It only uses btrfs_inode internally so take it as a parameter.

Signed-off-by: Nikolay Borisov <nborisov@suse.com>
Reviewed-by: David Sterba <dsterba@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
3 years agobtrfs: make btrfs_free_reserved_data_space_noquota take btrfs_fs_info
Nikolay Borisov [Wed, 3 Jun 2020 05:55:38 +0000 (08:55 +0300)]
btrfs: make btrfs_free_reserved_data_space_noquota take btrfs_fs_info

No point in taking an inode only to get btrfs_fs_info from it, instead
take btrfs_fs_info directly.

Signed-off-by: Nikolay Borisov <nborisov@suse.com>
Reviewed-by: David Sterba <dsterba@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
3 years agobtrfs: make btrfs_qgroup_reserve_data take btrfs_inode
Nikolay Borisov [Wed, 3 Jun 2020 05:55:37 +0000 (08:55 +0300)]
btrfs: make btrfs_qgroup_reserve_data take btrfs_inode

There's only a single use of vfs_inode in a tracepoint so let's take
btrfs_inode directly.

Signed-off-by: Nikolay Borisov <nborisov@suse.com>
Reviewed-by: David Sterba <dsterba@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
3 years agobtrfs: make btrfs_dirty_pages take btrfs_inode
Nikolay Borisov [Wed, 3 Jun 2020 05:55:36 +0000 (08:55 +0300)]
btrfs: make btrfs_dirty_pages take btrfs_inode

There is a single use of the generic vfs_inode so let's take btrfs_inode
as a parameter and remove couple of redundant BTRFS_I() calls.

Signed-off-by: Nikolay Borisov <nborisov@suse.com>
Reviewed-by: David Sterba <dsterba@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
3 years agobtrfs: make btrfs_set_extent_delalloc take btrfs_inode
Nikolay Borisov [Wed, 3 Jun 2020 05:55:35 +0000 (08:55 +0300)]
btrfs: make btrfs_set_extent_delalloc take btrfs_inode

Preparation to make btrfs_dirty_pages take btrfs_inode as parameter.

Signed-off-by: Nikolay Borisov <nborisov@suse.com>
Reviewed-by: David Sterba <dsterba@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
3 years agobtrfs: make writepage_delalloc take btrfs_inode
Nikolay Borisov [Fri, 5 Jun 2020 07:42:10 +0000 (10:42 +0300)]
btrfs: make writepage_delalloc take btrfs_inode

Only find_lock_delalloc_range uses vfs_inode so let's take the
btrfs_inode as a parameter.

Signed-off-by: Nikolay Borisov <nborisov@suse.com>
Reviewed-by: David Sterba <dsterba@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
3 years agobtrfs: make __extent_writepage_io take btrfs_inode
Nikolay Borisov [Wed, 3 Jun 2020 05:55:33 +0000 (08:55 +0300)]
btrfs: make __extent_writepage_io take btrfs_inode

It has only a single use for a generic vfs inode vs 3 for btrfs_inode.

Signed-off-by: Nikolay Borisov <nborisov@suse.com>
Reviewed-by: David Sterba <dsterba@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
3 years agobtrfs: make btrfs_new_extent_direct take btrfs_inode
Nikolay Borisov [Wed, 3 Jun 2020 05:55:32 +0000 (08:55 +0300)]
btrfs: make btrfs_new_extent_direct take btrfs_inode

This function really needs a btrfs_inode and not a generic vfs one. Take
it as a parameter and get rid of superfluous BTRFS_I() calls.

Signed-off-by: Nikolay Borisov <nborisov@suse.com>
Reviewed-by: David Sterba <dsterba@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
3 years agobtrfs: make btrfs_create_dio_extent take btrfs_inode
Nikolay Borisov [Wed, 3 Jun 2020 05:55:31 +0000 (08:55 +0300)]
btrfs: make btrfs_create_dio_extent take btrfs_inode

Take btrfs_inode directly and stop using superfulous BTRFS_I calls.

Signed-off-by: Nikolay Borisov <nborisov@suse.com>
Reviewed-by: David Sterba <dsterba@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
3 years agobtrfs: make btrfs_add_ordered_extent_dio take btrfs_inode
Nikolay Borisov [Wed, 3 Jun 2020 05:55:30 +0000 (08:55 +0300)]
btrfs: make btrfs_add_ordered_extent_dio take btrfs_inode

Simply forwards its argument so let's get rid of one extra BTRFS_I by
taking btrfs_inode directly.

Signed-off-by: Nikolay Borisov <nborisov@suse.com>
Reviewed-by: David Sterba <dsterba@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
3 years agobtrfs: make btrfs_run_delalloc_range take btrfs_inode
Nikolay Borisov [Wed, 3 Jun 2020 05:55:29 +0000 (08:55 +0300)]
btrfs: make btrfs_run_delalloc_range take btrfs_inode

All children now take btrfs_inode so convert it to taking it as a
parameter as well.

Signed-off-by: Nikolay Borisov <nborisov@suse.com>
Reviewed-by: David Sterba <dsterba@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
3 years agobtrfs: make need_force_cow take btrfs_inode
Nikolay Borisov [Wed, 3 Jun 2020 05:55:28 +0000 (08:55 +0300)]
btrfs: make need_force_cow take btrfs_inode

Gets rid of superfulous BTRFS_I() calls and prepare for converting
btrfs_run_delalloc_range to using btrfs_inode.

Signed-off-by: Nikolay Borisov <nborisov@suse.com>
Reviewed-by: David Sterba <dsterba@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
3 years agobtrfs: make inode_need_compress take btrfs_inode
Nikolay Borisov [Wed, 3 Jun 2020 05:55:27 +0000 (08:55 +0300)]
btrfs: make inode_need_compress take btrfs_inode

Simply gets rid of superfluous BTRFS_I() calls.

Signed-off-by: Nikolay Borisov <nborisov@suse.com>
Reviewed-by: David Sterba <dsterba@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
3 years agobtrfs: make inode_can_compress take btrfs_inode
Nikolay Borisov [Wed, 3 Jun 2020 05:55:26 +0000 (08:55 +0300)]
btrfs: make inode_can_compress take btrfs_inode

Gets rid of superfluous BTRFS_I() calls.

Signed-off-by: Nikolay Borisov <nborisov@suse.com>
Reviewed-by: David Sterba <dsterba@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
3 years agobtrfs: make btrfs_cleanup_ordered_extents take btrfs_inode
Nikolay Borisov [Wed, 3 Jun 2020 05:55:25 +0000 (08:55 +0300)]
btrfs: make btrfs_cleanup_ordered_extents take btrfs_inode

Preparation to converting btrfs_run_delalloc_range to using btrfs_inode
without BTRFS_I() calls.

Signed-off-by: Nikolay Borisov <nborisov@suse.com>
Reviewed-by: David Sterba <dsterba@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
3 years agobtrfs: make __endio_write_update_ordered take btrfs_inode
Nikolay Borisov [Wed, 3 Jun 2020 05:55:24 +0000 (08:55 +0300)]
btrfs: make __endio_write_update_ordered take btrfs_inode

It really wants btrfs_inode.

Signed-off-by: Nikolay Borisov <nborisov@suse.com>
Reviewed-by: David Sterba <dsterba@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
3 years agobtrfs: make btrfs_dec_test_first_ordered_pending take btrfs_inode
Nikolay Borisov [Wed, 3 Jun 2020 05:55:23 +0000 (08:55 +0300)]
btrfs: make btrfs_dec_test_first_ordered_pending take btrfs_inode

It doesn't really need vfs_inode but btrfs_inode.

Signed-off-by: Nikolay Borisov <nborisov@suse.com>
Reviewed-by: David Sterba <dsterba@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
3 years agobtrfs: make cow_file_range_async take btrfs_inode
Nikolay Borisov [Wed, 3 Jun 2020 05:55:22 +0000 (08:55 +0300)]
btrfs: make cow_file_range_async take btrfs_inode

It only uses vfs inode for assigning it to the async_chunk function.

Signed-off-by: Nikolay Borisov <nborisov@suse.com>
Reviewed-by: David Sterba <dsterba@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
3 years agobtrfs: make run_delalloc_nocow take btrfs_inode
Nikolay Borisov [Wed, 3 Jun 2020 05:55:21 +0000 (08:55 +0300)]
btrfs: make run_delalloc_nocow take btrfs_inode

It only really uses btrfs_inode.

Signed-off-by: Nikolay Borisov <nborisov@suse.com>
Reviewed-by: David Sterba <dsterba@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
3 years agobtrfs: make fallback_to_cow take btrfs_inode
Nikolay Borisov [Wed, 3 Jun 2020 05:55:20 +0000 (08:55 +0300)]
btrfs: make fallback_to_cow take btrfs_inode

It really wants btrfs_inode and is prepration to converting
run_delalloc_nocow to taking btrfs_inode.

Signed-off-by: Nikolay Borisov <nborisov@suse.com>
Reviewed-by: David Sterba <dsterba@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
3 years agobtrfs: make insert_reserved_file_extent take btrfs_inode
Nikolay Borisov [Wed, 3 Jun 2020 05:55:19 +0000 (08:55 +0300)]
btrfs: make insert_reserved_file_extent take btrfs_inode

Signed-off-by: Nikolay Borisov <nborisov@suse.com>
Reviewed-by: David Sterba <dsterba@suse.com>c
Signed-off-by: David Sterba <dsterba@suse.com>
3 years agobtrfs: make btrfs_qgroup_release_data take btrfs_inode
Nikolay Borisov [Wed, 3 Jun 2020 05:55:18 +0000 (08:55 +0300)]
btrfs: make btrfs_qgroup_release_data take btrfs_inode

It just forwards its argument to __btrfs_qgroup_release_data.

Signed-off-by: Nikolay Borisov <nborisov@suse.com>
Reviewed-by: David Sterba <dsterba@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
3 years agobtrfs: make submit_compressed_extents take btrfs_inode
Nikolay Borisov [Wed, 3 Jun 2020 05:55:17 +0000 (08:55 +0300)]
btrfs: make submit_compressed_extents take btrfs_inode

All but 3 uses require vfs_inode so convert the logic to have
btrfs_inode be the main inode struct.

Signed-off-by: Nikolay Borisov <nborisov@suse.com>
Reviewed-by: David Sterba <dsterba@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
3 years agobtrfs: make btrfs_submit_compressed_write take btrfs_inode
Nikolay Borisov [Wed, 3 Jun 2020 05:55:16 +0000 (08:55 +0300)]
btrfs: make btrfs_submit_compressed_write take btrfs_inode

Majority of its uses are for btrfs_inode so take it as an argument
directly.

Signed-off-by: Nikolay Borisov <nborisov@suse.com>
Reviewed-by: David Sterba <dsterba@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
3 years agobtrfs: make btrfs_add_ordered_extent_compress take btrfs_inode
Nikolay Borisov [Wed, 3 Jun 2020 05:55:15 +0000 (08:55 +0300)]
btrfs: make btrfs_add_ordered_extent_compress take btrfs_inode

It simpy forwards its inode argument to __btrfs_add_ordered_extent which
already takes btrfs_inode.

Signed-off-by: Nikolay Borisov <nborisov@suse.com>
Reviewed-by: David Sterba <dsterba@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
3 years agobtrfs: make cow_file_range take btrfs_inode
Nikolay Borisov [Wed, 3 Jun 2020 05:55:14 +0000 (08:55 +0300)]
btrfs: make cow_file_range take btrfs_inode

All its children functions take btrfs_inode so convert it to taking
btrfs_inode.

Signed-off-by: Nikolay Borisov <nborisov@suse.com>
Reviewed-by: David Sterba <dsterba@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
3 years agobtrfs: make btrfs_add_ordered_extent take btrfs_inode
Nikolay Borisov [Wed, 3 Jun 2020 05:55:13 +0000 (08:55 +0300)]
btrfs: make btrfs_add_ordered_extent take btrfs_inode

Preparation to converting its callers to taking btrfs_inode.

Signed-off-by: Nikolay Borisov <nborisov@suse.com>
Reviewed-by: David Sterba <dsterba@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
3 years agobtrfs: make cow_file_range_inline take btrfs_inode
Nikolay Borisov [Wed, 3 Jun 2020 05:55:12 +0000 (08:55 +0300)]
btrfs: make cow_file_range_inline take btrfs_inode

It has only 2 uses for the vfs_inode - insert_inline_extent and
i_size_read.  On the flipside it will allow converting its callers to
btrfs_inode, so convert it to taking btrfs_inode.

Signed-off-by: Nikolay Borisov <nborisov@suse.com>
Reviewed-by: David Sterba <dsterba@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
3 years agobtrfs: make btrfs_qgroup_free_data take btrfs_inode
Nikolay Borisov [Wed, 3 Jun 2020 05:55:11 +0000 (08:55 +0300)]
btrfs: make btrfs_qgroup_free_data take btrfs_inode

It passes btrfs_inode to its callee so change the interface.

Signed-off-by: Nikolay Borisov <nborisov@suse.com>
Reviewed-by: David Sterba <dsterba@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
3 years agobtrfs: make __btrfs_qgroup_release_data take btrfs_inode
Nikolay Borisov [Wed, 3 Jun 2020 05:55:10 +0000 (08:55 +0300)]
btrfs: make __btrfs_qgroup_release_data take btrfs_inode

It uses vfs_inode only for a tracepoint so convert its interface to take
btrfs_inode directly.

Signed-off-by: Nikolay Borisov <nborisov@suse.com>
Reviewed-by: David Sterba <dsterba@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
3 years agobtrfs: make qgroup_free_reserved_data take btrfs_inode
Nikolay Borisov [Wed, 3 Jun 2020 05:55:09 +0000 (08:55 +0300)]
btrfs: make qgroup_free_reserved_data take btrfs_inode

It only uses btrfs_inode so can just as easily take it as an argument.

Signed-off-by: Nikolay Borisov <nborisov@suse.com>
Reviewed-by: David Sterba <dsterba@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
3 years agobtrfs: tracepoints: convert flush states to using EM macros
Nikolay Borisov [Fri, 19 Jun 2020 12:24:51 +0000 (15:24 +0300)]
btrfs: tracepoints: convert flush states to using EM macros

Only 6 out of all flush states were being printed correctly since
only they were exported via the TRACE_DEFINE_ENUM macro. This patch
converts all flush states to use the newly introduced EM macro so that
they can all be printed correctly.

Signed-off-by: Nikolay Borisov <nborisov@suse.com>
Reviewed-by: David Sterba <dsterba@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
3 years agobtrfs: tracepoints: switch extent_io_tree_owner to using EM macro
Nikolay Borisov [Fri, 19 Jun 2020 12:24:50 +0000 (15:24 +0300)]
btrfs: tracepoints: switch extent_io_tree_owner to using EM macro

This fixes correct pint out of the extent io tree owner in
btrfs_set_extent_bit/btrfs_clear_extent_bit/btrfs_convert_extent_bit
tracepoints.

Signed-off-by: Nikolay Borisov <nborisov@suse.com>
Reviewed-by: David Sterba <dsterba@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
3 years agobtrfs: tracepoints: fix qgroup reservation type printing
Nikolay Borisov [Fri, 19 Jun 2020 12:24:49 +0000 (15:24 +0300)]
btrfs: tracepoints: fix qgroup reservation type printing

Since qgroup's reservation types are define in a macro they must be
exported to user space in order for user space tools to convert raw
binary data to symbolic names. Currently trace-cmd report produces
the following output:

kworker/u8:2-459   [003]  1208.543587: qgroup_update_reserve:
2b742cae-e0e5-4def-9ef7-28a9b34a951e: qgid=5 type=0x2 cur_reserved=54870016 diff=-32768

With this fix the output is:

kworker/u8:2-459   [003]  1208.543587: qgroup_update_reserve:
2b742cae-e0e5-4def-9ef7-28a9b34a951e: qgid=5 type=BTRFS_QGROUP_RSV_META_PREALLOC cur_reserved=54870016 diff=-32768

Signed-off-by: Nikolay Borisov <nborisov@suse.com>
Reviewed-by: David Sterba <dsterba@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
3 years agobtrfs: tracepoints: move FLUSH_ACTIONS define
Nikolay Borisov [Fri, 19 Jun 2020 12:24:48 +0000 (15:24 +0300)]
btrfs: tracepoints: move FLUSH_ACTIONS define

Since all enums used in btrfs' tracepoints are going to be redefined
to allow proper parsing of their values by userspace tools let's
rearrange when they are defined. This will allow to use only a single
set of #define EM/#undef EM sequence. No functional changes.

Signed-off-by: Nikolay Borisov <nborisov@suse.com>
Reviewed-by: David Sterba <dsterba@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
3 years agobtrfs: tracepoints: fix extent type symbolic name print
Nikolay Borisov [Fri, 19 Jun 2020 12:24:47 +0000 (15:24 +0300)]
btrfs: tracepoints: fix extent type symbolic name print

extent's type is an enum and this requires that the enum values be
exported to user space so that user space tools can correctly map raw
binary data to the symbolic name. Currently tracepoints using
btrfs__file_extent_item_regular or btrfs__file_extent_item_inline result
in the following output:

fio-443   [002]   586.609450: btrfs_get_extent_show_fi_regular: f0c3bf8e-0174-4bcc-92aa-6c2d62430420:i
root=5(FS_TREE) inode=258 size=2136457216 disk_isize=0
file extent range=[2126946304 2136457216] (num_bytes=9510912
ram_bytes=9510912 disk_bytenr=0 disk_num_bytes=0 extent_offset=0
type=0x1 compression=0

E.g type is 0x1 . With this patch applie the output is:

<ommitted for brevity>  disk_bytenr=141348864 disk_num_bytes=4096 extent_offset=0 type=REG compression=0

Signed-off-by: Nikolay Borisov <nborisov@suse.com>
Reviewed-by: David Sterba <dsterba@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
3 years agobtrfs: tracepoints: fix btrfs_trigger_flush symbolic string for flags
Nikolay Borisov [Fri, 19 Jun 2020 12:24:46 +0000 (15:24 +0300)]
btrfs: tracepoints: fix btrfs_trigger_flush symbolic string for flags

When tracepoints use __print_symbolic to print textual representation of
a value that comes from an ENUM each enum value needs to be exported
to user space so that user space tools can convert the binary value
data to the trings as user space does not know what those enums are
about.

Doing a trace-cmd record && trace-cmd report currently results in:

kworker/u8:1-61    [000]    66.299527:
btrfs_flush_space:    5302ee13-c65e-45bb-98ef-8fe3835bd943:
state=3(0x3) flags=4(METADATA) num_bytes=2621440 ret=0

I.e state is not translated to its symbolic counterpart. With this patch
applied the output is:

fio-370   [002]    56.762402: btrfs_trigger_flush:  d04cd7ac-38e2-452f-a7f5-8157529fd5f0:
preempt: flush=3(BTRFS_RESERVE_FLUSH_ALL) flags=4(METADATA) bytes=655360

See also 190f0b76ca49 ("mm: tracing: Export enums in tracepoints to user
space").

Signed-off-by: Nikolay Borisov <nborisov@suse.com>
Reviewed-by: David Sterba <dsterba@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
3 years agobtrfs: allow use of global block reserve for balance item deletion
David Sterba [Thu, 25 Jun 2020 10:35:28 +0000 (12:35 +0200)]
btrfs: allow use of global block reserve for balance item deletion

On a filesystem with exhausted metadata, but still enough to start
balance, it's possible to hit this error:

[324402.053842] BTRFS info (device loop0): 1 enospc errors during balance
[324402.060769] BTRFS info (device loop0): balance: ended with status: -28
[324402.172295] BTRFS: error (device loop0) in reset_balance_state:3321: errno=-28 No space left

It fails inside reset_balance_state and turns the filesystem to
read-only, which is unnecessary and should be fixed too, but the problem
is caused by lack for space when the balance item is deleted. This is a
one-time operation and from the same rank as unlink that is allowed to
use the global block reserve. So do the same for the balance item.

Status of the filesystem (100GiB) just after the balance fails:

$ btrfs fi df mnt
Data, single: total=80.01GiB, used=38.58GiB
System, single: total=4.00MiB, used=16.00KiB
Metadata, single: total=19.99GiB, used=19.48GiB
GlobalReserve, single: total=512.00MiB, used=50.11MiB

CC: stable@vger.kernel.org # 4.4+
Reviewed-by: Johannes Thumshirn <johannes.thumshirn@wdc.com>
Signed-off-by: David Sterba <dsterba@suse.com>
3 years agobtrfs: refactor btrfs_check_can_nocow() into two variants
Qu Wenruo [Tue, 23 Jun 2020 23:23:52 +0000 (07:23 +0800)]
btrfs: refactor btrfs_check_can_nocow() into two variants

The function btrfs_check_can_nocow() now has two completely different
call patterns.

For nowait variant, callers don't need to do any cleanup.  While for
wait variant, callers need to release the lock if they can do nocow
write.

This is somehow confusing, and is already a problem for the exported
btrfs_check_can_nocow().

So this patch will separate the different patterns into different
functions.
For nowait variant, the function will be called check_nocow_nolock().
For wait variant, the function pair will be btrfs_check_nocow_lock()
btrfs_check_nocow_unlock().

Reviewed-by: Anand Jain <anand.jain@oracle.com>
Signed-off-by: Qu Wenruo <wqu@suse.com>
Reviewed-by: David Sterba <dsterba@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
3 years agobtrfs: add comments for btrfs_check_can_nocow() and can_nocow_extent()
Qu Wenruo [Tue, 23 Jun 2020 23:23:51 +0000 (07:23 +0800)]
btrfs: add comments for btrfs_check_can_nocow() and can_nocow_extent()

These two functions have extra conditions that their callers need to
meet, and some not-that-common parameters used for return value.

So adding some comments may save reviewers some time.

Signed-off-by: Qu Wenruo <wqu@suse.com>
Reviewed-by: David Sterba <dsterba@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
3 years agobtrfs: allow btrfs_truncate_block() to fallback to nocow for data space reservation
Qu Wenruo [Tue, 23 Jun 2020 23:23:50 +0000 (07:23 +0800)]
btrfs: allow btrfs_truncate_block() to fallback to nocow for data space reservation

[BUG]
When the data space is exhausted, even if the inode has NOCOW attribute,
we will still refuse to truncate unaligned range due to ENOSPC.

The following script can reproduce it pretty easily:
  #!/bin/bash

  dev=/dev/test/test
  mnt=/mnt/btrfs

  umount $dev &> /dev/null
  umount $mnt &> /dev/null

  mkfs.btrfs -f $dev -b 1G
  mount -o nospace_cache $dev $mnt
  touch $mnt/foobar
  chattr +C $mnt/foobar

  xfs_io -f -c "pwrite -b 4k 0 4k" $mnt/foobar > /dev/null
  xfs_io -f -c "pwrite -b 4k 0 1G" $mnt/padding &> /dev/null
  sync

  xfs_io -c "fpunch 0 2k" $mnt/foobar
  umount $mnt

Currently this will fail at the fpunch part.

[CAUSE]
Because btrfs_truncate_block() always reserves space without checking
the NOCOW attribute.

Since the writeback path follows NOCOW bit, we only need to bother the
space reservation code in btrfs_truncate_block().

[FIX]
Make btrfs_truncate_block() follow btrfs_buffered_write() to try to
reserve data space first, and fall back to NOCOW check only when we
don't have enough space.

Such always-try-reserve is an optimization introduced in
btrfs_buffered_write(), to avoid expensive btrfs_check_can_nocow() call.

This patch will export check_can_nocow() as btrfs_check_can_nocow(), and
use it in btrfs_truncate_block() to fix the problem.

Reported-by: Martin Doucha <martin.doucha@suse.com>
Reviewed-by: Filipe Manana <fdmanana@suse.com>
Reviewed-by: Anand Jain <anand.jain@oracle.com>
Signed-off-by: Qu Wenruo <wqu@suse.com>
Reviewed-by: David Sterba <dsterba@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
3 years agobtrfs: start deprecation of mount option inode_cache
David Sterba [Thu, 18 Jun 2020 12:54:56 +0000 (14:54 +0200)]
btrfs: start deprecation of mount option inode_cache

Estimated time of removal of the functionality is 5.11, the option will
be still parsed but will have no effect.

Reasons for deprecation and removal:

- very poor naming choice of the mount option, it's supposed to cache
  and reuse the inode _numbers_, but it sounds a some generic cache for
  inodes

- the only known usecase where this option would make sense is on a
  32bit architecture where inode numbers in one subvolume would be
  exhausted due to 32bit inode::i_ino

- the cache is stored on disk, consumes space, needs to be loaded and
  written back

- new inode number allocation is slower due to lookups into the cache
  (compared to a simple increment which is the default)

- uses the free-space-cache code that is going to be deprecated as well
  in the future

Known problems:

- since 2011, returning EEXIST when there's not enough space in a page
  to store all checksums, see commit 4b9465cb9e38 ("Btrfs: add mount -o
  inode_cache")

Remaining issues:

- if the option was enabled, new inodes created, the option disabled
  again, the cache is still stored on the devices and there's currently
  no way to remove it

Signed-off-by: David Sterba <dsterba@suse.com>
3 years agobtrfs: remove unused btrfs_root::defrag_trans_start
David Sterba [Tue, 23 Jun 2020 19:23:54 +0000 (21:23 +0200)]
btrfs: remove unused btrfs_root::defrag_trans_start

Last touched in 2013 by commit de78b51a2852 ("btrfs: remove cache only
arguments from defrag path") that was the only code that used the value.
Now it's only set but never used for anything, so we can remove it.

Reviewed-by: Nikolay Borisov <nborisov@suse.com>
Reviewed-by: Anand Jain <anand.jain@oracle.com>
Signed-off-by: David Sterba <dsterba@suse.com>
3 years agobtrfs: don't use UAPI types for fiemap callback
David Sterba [Tue, 23 Jun 2020 18:56:12 +0000 (20:56 +0200)]
btrfs: don't use UAPI types for fiemap callback

The fiemap callback is not part of UAPI interface and the prototypes
don't have the __u64 types either.

Reviewed-by: Nikolay Borisov <nborisov@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
3 years agobtrfs: tests: remove if duplicate in __check_free_space_extents()
Denis Efremov [Mon, 22 Jun 2020 20:18:41 +0000 (23:18 +0300)]
btrfs: tests: remove if duplicate in __check_free_space_extents()

num_extents is already checked in the next if condition and can
be safely removed.

Signed-off-by: Denis Efremov <efremov@linux.com>
Reviewed-by: David Sterba <dsterba@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
3 years agobtrfs: use free_root_extent_buffer to free root
Johannes Thumshirn [Tue, 23 Jun 2020 08:40:07 +0000 (17:40 +0900)]
btrfs: use free_root_extent_buffer to free root

In btrfs_put_root() we're freeing a btrfs_root's 'node' and 'commit_root'
extent buffers manually via kfree(), while we're using
free_root_extent_buffers() in the free_root_pointers() function above.

free_root_extent_buffers() also NULLs the pointers after freeing, which
mitigates potential double frees.

Reviewed-by: Anand Jain <anand.jain@oracle.com>
Signed-off-by: Johannes Thumshirn <johannes.thumshirn@wdc.com>
Reviewed-by: David Sterba <dsterba@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
3 years agobtrfs: use for loop in prealloc_file_extent_cluster
Nikolay Borisov [Wed, 17 Jun 2020 09:10:44 +0000 (12:10 +0300)]
btrfs: use for loop in prealloc_file_extent_cluster

This function iterates all extents in the extent cluster, make this
intention obvious by using a for loop. No functional chanes.

Reviewed-by: Johannes Thumshirn <johannes.thumshirn@wdc.com>
Signed-off-by: Nikolay Borisov <nborisov@suse.com>
Reviewed-by: David Sterba <dsterba@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
3 years agobtrfs: perform data management operations outside of inode lock
Nikolay Borisov [Wed, 17 Jun 2020 09:10:43 +0000 (12:10 +0300)]
btrfs: perform data management operations outside of inode lock

btrfs_alloc_data_chunk_ondemand and btrfs_free_reserved_data_space_noquota
don't really use the guts of the inodes being passed to them. This
implies it's not required to call them under extent lock. Move code
around in prealloc_file_extent_cluster to do the heavy, data alloc/free
operations outside of the lock. This also makes the 'out' label
unnecessary, so remove it.

Reviewed-by: Johannes Thumshirn <johannes.thumshirn@wdc.com>
Signed-off-by: Nikolay Borisov <nborisov@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
3 years agobtrfs: remove hole check in prealloc_file_extent_cluster
Nikolay Borisov [Wed, 17 Jun 2020 09:10:42 +0000 (12:10 +0300)]
btrfs: remove hole check in prealloc_file_extent_cluster

Extents in the extent cluster are guaranteed to be contiguous as such
the hole check inside the loop can never trigger. In fact this check was
never functional since it was added in 18513091af94 ("btrfs: update
btrfs_space_info's bytes_may_use timely") which came after the commit
introducing clustered/contiguous extents 0257bb82d21b ("Btrfs: relocate
file extents in clusters").

Let's just remove it as it adds noise to the source.

Reviewed-by: Johannes Thumshirn <johannes.thumshirn@wdc.com>
Signed-off-by: Nikolay Borisov <nborisov@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
3 years agobtrfs: make __btrfs_drop_extents take btrfs_inode
Nikolay Borisov [Wed, 3 Jun 2020 05:55:08 +0000 (08:55 +0300)]
btrfs: make __btrfs_drop_extents take btrfs_inode

It has only 4 uses of a vfs_inode for inode_sub_bytes but unifies the
interface with the non  __ prefixed version. Will also makes converting
its callers to btrfs_inode easier.

Signed-off-by: Nikolay Borisov <nborisov@suse.com>
Reviewed-by: David Sterba <dsterba@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
3 years agobtrfs: make btrfs_csum_one_bio takae btrfs_inode
Nikolay Borisov [Wed, 3 Jun 2020 05:55:07 +0000 (08:55 +0300)]
btrfs: make btrfs_csum_one_bio takae btrfs_inode

Will enable converting btrfs_submit_compressed_write to btrfs_inode more
easily.

Signed-off-by: Nikolay Borisov <nborisov@suse.com>
Reviewed-by: David Sterba <dsterba@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
3 years agobtrfs: make extent_clear_unlock_delalloc take btrfs_inode
Nikolay Borisov [Wed, 3 Jun 2020 05:55:06 +0000 (08:55 +0300)]
btrfs: make extent_clear_unlock_delalloc take btrfs_inode

It has one VFS and 1 btrfs inode usages but converting it to btrfs_inode
interface will allow seamless conversion of its callers.

Signed-off-by: Nikolay Borisov <nborisov@suse.com>
Reviewed-by: David Sterba <dsterba@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
3 years agobtrfs: make create_io_em take btrfs_inode
Nikolay Borisov [Wed, 3 Jun 2020 05:55:05 +0000 (08:55 +0300)]
btrfs: make create_io_em take btrfs_inode

It really wants a btrfs_inode and will allow submit_compressed_extents
to be completely converted to btrfs_inode in follow up patches.

Signed-off-by: Nikolay Borisov <nborisov@suse.com>
Reviewed-by: David Sterba <dsterba@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
3 years agobtrfs: make btrfs_reloc_clone_csums take btrfs_inode
Nikolay Borisov [Wed, 3 Jun 2020 05:55:04 +0000 (08:55 +0300)]
btrfs: make btrfs_reloc_clone_csums take btrfs_inode

It really wants btrfs_inode and not a vfs inode.

Signed-off-by: Nikolay Borisov <nborisov@suse.com>
Reviewed-by: David Sterba <dsterba@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
3 years agobtrfs: make btrfs_lookup_ordered_extent take btrfs_inode
Nikolay Borisov [Wed, 3 Jun 2020 05:55:03 +0000 (08:55 +0300)]
btrfs: make btrfs_lookup_ordered_extent take btrfs_inode

It doesn't use the generic vfs inode for anything use btrfs_inode
directly.

Signed-off-by: Nikolay Borisov <nborisov@suse.com>
Reviewed-by: David Sterba <dsterba@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
3 years agobtrfs: make get_extent_allocation_hint take btrfs_inode
Nikolay Borisov [Wed, 3 Jun 2020 05:55:02 +0000 (08:55 +0300)]
btrfs: make get_extent_allocation_hint take btrfs_inode

It doesn't use the vfs inode for anything, can just as easily take
btrfs_inode.  Follow up patches will convert callers as well.

Signed-off-by: Nikolay Borisov <nborisov@suse.com>
Reviewed-by: David Sterba <dsterba@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
3 years agobtrfs: make __btrfs_add_ordered_extent take struct btrfs_inode
Nikolay Borisov [Wed, 3 Jun 2020 05:55:01 +0000 (08:55 +0300)]
btrfs: make __btrfs_add_ordered_extent take struct btrfs_inode

This is internal btrfs function what really needs the vfs_inode only for
igrab and a tracepoint.

Signed-off-by: Nikolay Borisov <nborisov@suse.com>
Reviewed-by: David Sterba <dsterba@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
3 years agobtrfs: remove no longer used trans_list member of struct btrfs_ordered_extent
Filipe Manana [Mon, 15 Jun 2020 09:36:58 +0000 (10:36 +0100)]
btrfs: remove no longer used trans_list member of struct btrfs_ordered_extent

The 'trans_list' member of an ordered extent was used to keep track of the
ordered extents for which a transaction commit had to wait. These were
ordered extents that were started and logged by an fsync. However we don't
do that anymore and before we stopped doing it we changed the approach to
wait for the ordered extents in commit 161c3549b45aee ("Btrfs: change how
we wait for pending ordered extents"), which stopped using that list and
therefore the 'trans_list' member is not used anymore since that commit.
So just remove it since it's doing nothing and making each ordered extent
structure waste memory (2 pointers).

Reviewed-by: Johannes Thumshirn <johannes.thumshirn@wdc.com>
Signed-off-by: Filipe Manana <fdmanana@suse.com>
Reviewed-by: David Sterba <dsterba@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
3 years agobtrfs: remove no longer used log_list member of struct btrfs_ordered_extent
Filipe Manana [Mon, 15 Jun 2020 09:36:48 +0000 (10:36 +0100)]
btrfs: remove no longer used log_list member of struct btrfs_ordered_extent

The 'log_list' member of an ordered extent was used keep track of which
ordered extents we needed to wait after logging metadata, but is not used
anymore since commit 5636cf7d6dc86f ("btrfs: remove the logged extents
infrastructure"), as we now always wait on ordered extent completion
before logging metadata. So just remove it since it's doing nothing and
making each ordered extent structure waste more memory (2 pointers).

Reviewed-by: Johannes Thumshirn <johannes.thumshirn@wdc.com>
Signed-off-by: Filipe Manana <fdmanana@suse.com>
Reviewed-by: David Sterba <dsterba@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
3 years agobtrfs: add little-endian optimized key helpers
David Sterba [Mon, 8 Jun 2020 14:06:07 +0000 (16:06 +0200)]
btrfs: add little-endian optimized key helpers

The CPU and on-disk keys are mapped to two different structures because
of the endianness. There's an intermediate buffer used to do the
conversion, but this is not necessary when CPU and on-disk endianness
match.

Add optimized versions of helpers that take disk_key and use the buffer
directly for CPU keys or drop the intermediate buffer and conversion.

This saves a lot of stack space accross many functions and removes about
6K of generated binary code:

   text    data     bss     dec     hex filename
1090439   17468   14912 1122819  112203 pre/btrfs.ko
1084613   17456   14912 1116981  110b35 post/btrfs.ko

Delta: -5826

Reviewed-by: Johannes Thumshirn <johannes.thumshirn@wdc.com>
Reviewed-by: Nikolay Borisov <nborisov@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
3 years agobtrfs: qgroup: catch reserved space leaks at unmount time
Qu Wenruo [Wed, 10 Jun 2020 01:04:44 +0000 (09:04 +0800)]
btrfs: qgroup: catch reserved space leaks at unmount time

Before this patch, qgroup completely relies on per-inode extent io tree
to detect reserved data space leak.

However previous bug has already shown how release page before
btrfs_finish_ordered_io() could lead to leak, and since it's
QGROUP_RESERVED bit cleared without triggering qgroup rsv, it can't be
detected by per-inode extent io tree.

So this patch adds another (and hopefully the final) safety net to catch
qgroup data reserved space leak.  At least the new safety net catches
all the leaks during development, so it should be pretty useful in the
real world.

Reviewed-by: Josef Bacik <josef@toxicpanda.com>
Signed-off-by: Qu Wenruo <wqu@suse.com>
Reviewed-by: David Sterba <dsterba@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
3 years agobtrfs: change timing for qgroup reserved space for ordered extents to fix reserved...
Qu Wenruo [Wed, 10 Jun 2020 01:04:43 +0000 (09:04 +0800)]
btrfs: change timing for qgroup reserved space for ordered extents to fix reserved space leak

[BUG]
The following simple workload from fsstress can lead to qgroup reserved
data space leak:
  0/0: creat f0 x:0 0 0
  0/0: creat add id=0,parent=-1
  0/1: write f0[259 1 0 0 0 0] [600030,27288] 0
  0/4: dwrite - xfsctl(XFS_IOC_DIOINFO) f0[259 1 0 0 64 627318] return 25, fallback to stat()
  0/4: dwrite f0[259 1 0 0 64 627318] [610304,106496] 0

This would cause btrfs qgroup to leak 20480 bytes for data reserved
space.  If btrfs qgroup limit is enabled, such leak can lead to
unexpected early EDQUOT and unusable space.

[CAUSE]
When doing direct IO, kernel will try to writeback existing buffered
page cache, then invalidate them:
  generic_file_direct_write()
  |- filemap_write_and_wait_range();
  |- invalidate_inode_pages2_range();

However for btrfs, the bi_end_io hook doesn't finish all its heavy work
right after bio ends.  In fact, it delays its work further:

  submit_extent_page(end_io_func=end_bio_extent_writepage);
  end_bio_extent_writepage()
  |- btrfs_writepage_endio_finish_ordered()
     |- btrfs_init_work(finish_ordered_fn);

  <<< Work queue execution >>>
  finish_ordered_fn()
  |- btrfs_finish_ordered_io();
     |- Clear qgroup bits

This means, when filemap_write_and_wait_range() returns,
btrfs_finish_ordered_io() is not guaranteed to be executed, thus the
qgroup bits for related range are not cleared.

Now into how the leak happens, this will only focus on the overlapping
part of buffered and direct IO part.

1. After buffered write
   The inode had the following range with QGROUP_RESERVED bit:
    596 616K
|///////////////|
   Qgroup reserved data space: 20K

2. Writeback part for range [596K, 616K)
   Write back finished, but btrfs_finish_ordered_io() not get called
   yet.
   So we still have:
    596K 616K
|///////////////|
   Qgroup reserved data space: 20K

3. Pages for range [596K, 616K) get released
   This will clear all qgroup bits, but don't update the reserved data
   space.
   So we have:
    596K 616K
| |
   Qgroup reserved data space: 20K
   That number doesn't match the qgroup bit range anymore.

4. Dio prepare space for range [596K, 700K)
   Qgroup reserved data space for that range, we got:
    596K 616K 700K
|///////////////|///////////////////////|
   Qgroup reserved data space: 20K + 104K = 124K

5. btrfs_finish_ordered_range() gets executed for range [596K, 616K)
   Qgroup free reserved space for that range, we got:
    596K 616K 700K
| |///////////////////////|
   We need to free that range of reserved space.
   Qgroup reserved data space: 124K - 20K = 104K

6. btrfs_finish_ordered_range() gets executed for range [596K, 700K)
   However qgroup bit for range [596K, 616K) is already cleared in
   previous step, so we only free 84K for qgroup reserved space.
    596K 616K 700K
| | |
   We need to free that range of reserved space.
   Qgroup reserved data space: 104K - 84K = 20K

   Now there is no way to release that 20K unless disabling qgroup or
   unmounting the fs.

[FIX]
This patch will change the timing of btrfs_qgroup_release/free_data()
call.  Here it uses buffered COW write as an example.

The new timing | The old timing
----------------------------------------+---------------------------------------
 btrfs_buffered_write() | btrfs_buffered_write()
 |- btrfs_qgroup_reserve_data()  | |- btrfs_qgroup_reserve_data()
|
 btrfs_run_delalloc_range() | btrfs_run_delalloc_range()
 |- btrfs_add_ordered_extent()   |
    |- btrfs_qgroup_release_data() |
       The reserved is passed into |
       btrfs_ordered_extent structure |
|
 btrfs_finish_ordered_io() | btrfs_finish_ordered_io()
 |- The reserved space is passed to  | |- btrfs_qgroup_release_data()
    btrfs_qgroup_record |    The resereved space is passed
|    to btrfs_qgroup_recrod
|
 btrfs_qgroup_account_extents() | btrfs_qgroup_account_extents()
 |- btrfs_qgroup_free_refroot() | |- btrfs_qgroup_free_refroot()

The point of such change is to ensure, when ordered extents are
submitted, the qgroup reserved space is already released, to keep the
timing aligned with file_write_and_wait_range().

So that qgroup data reserved space is all bound to btrfs_ordered_extent
and solve the timing mismatch.

Fixes: f695fdcef83a ("btrfs: qgroup: Introduce functions to release/free qgroup reserve data space")
Suggested-by: Josef Bacik <josef@toxicpanda.com>
Reviewed-by: Josef Bacik <josef@toxicpanda.com>
Signed-off-by: Qu Wenruo <wqu@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
3 years agobtrfs: file: reserve qgroup space after the hole punch range is locked
Qu Wenruo [Wed, 10 Jun 2020 01:04:42 +0000 (09:04 +0800)]
btrfs: file: reserve qgroup space after the hole punch range is locked

The incoming qgroup reserved space timing will move the data reservation
to ordered extent completely.

However in btrfs_punch_hole_lock_range() will call
btrfs_invalidate_page(), which will clear QGROUP_RESERVED bit for the
range.

In current stage it's OK, but if we're making ordered extents handle the
reserved space, then btrfs_punch_hole_lock_range() can clear the
QGROUP_RESERVED bit before we submit ordered extent, leading to qgroup
reserved space leakage.

So here change the timing to make reserve data space after
btrfs_punch_hole_lock_range().
The new timing is fine for either current code or the new code.

Reviewed-by: Josef Bacik <josef@toxicpanda.com>
Signed-off-by: Qu Wenruo <wqu@suse.com>
Reviewed-by: David Sterba <dsterba@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
3 years agobtrfs: inode: move qgroup reserved space release to the callers of insert_reserved_fi...
Qu Wenruo [Wed, 10 Jun 2020 01:04:41 +0000 (09:04 +0800)]
btrfs: inode: move qgroup reserved space release to the callers of insert_reserved_file_extent()

This is to prepare for the incoming timing change of qgroup reserved
data space and ordered extent.

Reviewed-by: Josef Bacik <josef@toxicpanda.com>
Signed-off-by: Qu Wenruo <wqu@suse.com>
Reviewed-by: David Sterba <dsterba@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
3 years agobtrfs: inode: refactor the parameters of insert_reserved_file_extent()
Qu Wenruo [Wed, 10 Jun 2020 01:04:40 +0000 (09:04 +0800)]
btrfs: inode: refactor the parameters of insert_reserved_file_extent()

Function insert_reserved_file_extent() takes a long list of parameters,
which are all for btrfs_file_extent_item, even including two reserved
members, encryption and other_encoding.

This makes the parameter list unnecessary long for a function which only
gets called twice.

This patch will refactor the parameter list, by using
btrfs_file_extent_item as parameter directly to hugely reduce the number
of parameters.

Also, since there are only two callers, one in btrfs_finish_ordered_io()
which inserts file extent for ordered extent, and one
__btrfs_prealloc_file_range().

These two call sites have completely different context, where ordered
extent can be compressed, but will always be regular extent, while the
preallocated one is never going to be compressed and always has PREALLOC
type.

So use two small wrapper for these two different call sites to improve
readability.

Reviewed-by: Josef Bacik <josef@toxicpanda.com>
Signed-off-by: Qu Wenruo <wqu@suse.com>
Reviewed-by: David Sterba <dsterba@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
3 years agobtrfs: scrub: clean up temporary page variables in scrub_checksum_tree_block
David Sterba [Fri, 29 May 2020 14:20:35 +0000 (16:20 +0200)]
btrfs: scrub: clean up temporary page variables in scrub_checksum_tree_block

Add proper variable for the scrub page and use it instead of repeatedly
dereferencing the other structures.

Signed-off-by: David Sterba <dsterba@suse.com>
3 years agobtrfs: scrub: simplify tree block checksum calculation
David Sterba [Fri, 29 May 2020 13:54:41 +0000 (15:54 +0200)]
btrfs: scrub: simplify tree block checksum calculation

Use a simpler iteration over tree block pages, same what csum_tree_block
does: first page always exists, loop over the rest.

Signed-off-by: David Sterba <dsterba@suse.com>
3 years agobtrfs: scrub: clean up temporary page variables in scrub_checksum_data
David Sterba [Fri, 29 May 2020 14:20:35 +0000 (16:20 +0200)]
btrfs: scrub: clean up temporary page variables in scrub_checksum_data

Add proper variable for the scrub page and use it instead of repeatedly
dereferencing the other structures.

Signed-off-by: David Sterba <dsterba@suse.com>