[PATCH v2 0/4] btrfs: fix error handling of cow_file_range(unlock = 0)

[Naohiro Aota <naohiro.aota@wdc.com>]

This series is a revisit of patch "btrfs: ensure pages are unlocked on
cow_file_range() failure" [1].

[1] https://lore.kernel.org/linux-btrfs/20211213034338.949507-1-naohiro.aota@wdc.com/

Changes
- v2:
  - Rephrase comment in patch 1 (Filipe)
  - Add stable target to patch 3
    - I choose 5.18+ as the target as they are after refactoring and we can
      apply the series cleanly.
  - Rephrase subject of patch 4 (Filipe)

We have a hang report like below which is caused by a locked page.

[  726.328648] INFO: task rocksdb:high0:11085 blocked for more than 241 seconds.
[  726.329839]       Not tainted 5.16.0-rc1+ #1
[  726.330484] "echo 0 > /proc/sys/kernel/hung_task_timeout_secs" disables this message.
[  726.331603] task:rocksdb:high0   state:D stack:    0 pid:11085 ppid: 11082 flags:0x00000000
[  726.331608] Call Trace:
[  726.331611]  <TASK>
[  726.331614]  __schedule+0x2e5/0x9d0
[  726.331622]  schedule+0x58/0xd0
[  726.331626]  io_schedule+0x3f/0x70
[  726.331629]  __folio_lock+0x125/0x200
[  726.331634]  ? find_get_entries+0x1bc/0x240
[  726.331638]  ? filemap_invalidate_unlock_two+0x40/0x40
[  726.331642]  truncate_inode_pages_range+0x5b2/0x770
[  726.331649]  truncate_inode_pages_final+0x44/0x50
[  726.331653]  btrfs_evict_inode+0x67/0x480
[  726.331658]  evict+0xd0/0x180
[  726.331661]  iput+0x13f/0x200
[  726.331664]  do_unlinkat+0x1c0/0x2b0
[  726.331668]  __x64_sys_unlink+0x23/0x30
[  726.331670]  do_syscall_64+0x3b/0xc0
[  726.331674]  entry_SYSCALL_64_after_hwframe+0x44/0xae

When cow_file_range(unlock = 0), we never unlock a page after an ordered
extent is allocated for the page. When the allocation loop fails after some
ordered extents are allocated, we have three things to do: (1) unlock the
pages in the successfully allocated ordered extents, (2) clean-up the
allocated ordered extents, and (3) propagate the error to the userland.

However, current code fails to do (1) in the all cow_file_range(unlock = 0)
case and cause the above hang. Also, it fails to do (2) and (3) in
submit_uncompressed_range() case.

This series addresses these three issues on error handling of
cow_file_range(unlock = 0) case.

To test the series, I applied the following two patches to stress
submit_uncompressed_range() and to inject an error.

The following first diff forces the compress_type to be
BTRFS_COMPRESS_NONE, so that all "compress" writes go through
submit_uncompressed_range path.

    diff --git a/fs/btrfs/inode.c b/fs/btrfs/inode.c
    index 7a54f964ff37..eb12c47d02b8 100644
    --- a/fs/btrfs/inode.c
    +++ b/fs/btrfs/inode.c
    @@ -706,6 +706,7 @@ static noinline int compress_file_range(struct async_chunk *async_chunk)
     			compress_type = BTRFS_I(inode)->defrag_compress;
     		else if (BTRFS_I(inode)->prop_compress)
     			compress_type = BTRFS_I(inode)->prop_compress;
    +		compress_type = BTRFS_COMPRESS_NONE;
     
     		/*
     		 * we need to call clear_page_dirty_for_io on each

The following second diff limits the allocation size at most 256KB to run
the loop many times. Also, it fails the allocation at the specific offset
(fail_offset).

    diff --git a/fs/btrfs/inode.c b/fs/btrfs/inode.c
    index eb12c47d02b8..1247690e7021 100644
    --- a/fs/btrfs/inode.c
    +++ b/fs/btrfs/inode.c
    @@ -1155,6 +1155,8 @@ static noinline int cow_file_range(struct btrfs_inode *inode,
     	bool extent_reserved = false;
     	int ret = 0;
     
    +	u64 fail_offset = SZ_1M + (u64)SZ_256K * 0;
    +
     	if (btrfs_is_free_space_inode(inode)) {
     		ret = -EINVAL;
     		goto out_unlock;
    @@ -1239,9 +1241,13 @@ static noinline int cow_file_range(struct btrfs_inode *inode,
     
     	while (num_bytes > 0) {
     		cur_alloc_size = num_bytes;
    -		ret = btrfs_reserve_extent(root, cur_alloc_size, cur_alloc_size,
    -					   min_alloc_size, 0, alloc_hint,
    -					   &ins, 1, 1);
    +		cur_alloc_size = min_t(u64, SZ_256K, num_bytes);
    +		if (start != fail_offset)
    +			ret = btrfs_reserve_extent(root, cur_alloc_size, cur_alloc_size,
    +						   min_alloc_size, 0, alloc_hint,
    +						   &ins, 1, 1);
    +		else
    +			ret = -ENOSPC;
     		if (ret < 0)
     			goto out_unlock;
     		cur_alloc_size = ins.offset;

I ran the following script with these patches + the series applied changing
the fail_offset from "1MB + 256KB * 0" to "1MB + 256KB * 15" step by 256KB,
and confirmed the above three error handlings are properly done.

    run() {
    	local mkfs_opts=$1
    	local mount_opts=$2
    
    	for x in $(seq 100); do
    		echo $x
    		mkfs.btrfs -f -d single -m single ${mkfs_opts} /dev/nullb0
    		mount ${mount_opts} /dev/nullb0 /mnt/test
    		dd if=/dev/zero of=/mnt/test/file bs=1M count=4 seek=1 oflag=sync 2>&1 | tee /tmp/err
    		# check error propagation
    		grep -q 'No space left' /tmp/err || exit 1
    		sync
    		umount /mnt/test
    		dmesg | grep -q WARN && exit 1
    	done
    }
    
    run ""         ""
    run ""         "-o compress-force"
    run "-O zoned" ""
    run "-O zoned" "-o compress-force"

Patch 1 addresses the (1) by unlocking the pages in the error case. Also,
it adds a figure to clarify what to do for each range in the error case.

Patches 2 and 3 do (2) and (3) by fixing the error case of
submit_uncompressed_range().

Patch 4 is a refactoring patch to replace unnecessary "goto out"s with
direct return.

Naohiro Aota (4):
  btrfs: ensure pages are unlocked on cow_file_range() failure
  btrfs: extend btrfs_cleanup_ordered_extens for NULL locked_page
  btrfs: fix error handling of fallbacked uncompress write
  btrfs: replace unnecessary goto with direct return at cow_file_range()

 fs/btrfs/inode.c | 132 ++++++++++++++++++++++++++++++++++++-----------
 1 file changed, 103 insertions(+), 29 deletions(-)

-- 
2.35.1