Revert "btrfs: balance dirty metadata pages in btrfs_finish_ordered_io"

commit 77b7aad upstream.

This reverts commit e73e81b.

This patch causes a few problems:

- adds latency to btrfs_finish_ordered_io
- as btrfs_finish_ordered_io is used for free space cache, generating
  more work from btrfs_btree_balance_dirty_nodelay could end up in the
  same workque, effectively deadlocking

12260 kworker/u96:16+btrfs-freespace-write D
[<0>] balance_dirty_pages+0x6e6/0x7ad
[<0>] balance_dirty_pages_ratelimited+0x6bb/0xa90
[<0>] btrfs_finish_ordered_io+0x3da/0x770
[<0>] normal_work_helper+0x1c5/0x5a0
[<0>] process_one_work+0x1ee/0x5a0
[<0>] worker_thread+0x46/0x3d0
[<0>] kthread+0xf5/0x130
[<0>] ret_from_fork+0x24/0x30
[<0>] 0xffffffffffffffff

Transaction commit will wait on the freespace cache:

838 btrfs-transacti D
[<0>] btrfs_start_ordered_extent+0x154/0x1e0
[<0>] btrfs_wait_ordered_range+0xbd/0x110
[<0>] __btrfs_wait_cache_io+0x49/0x1a0
[<0>] btrfs_write_dirty_block_groups+0x10b/0x3b0
[<0>] commit_cowonly_roots+0x215/0x2b0
[<0>] btrfs_commit_transaction+0x37e/0x910
[<0>] transaction_kthread+0x14d/0x180
[<0>] kthread+0xf5/0x130
[<0>] ret_from_fork+0x24/0x30
[<0>] 0xffffffffffffffff

And then writepages ends up waiting on transaction commit:

9520 kworker/u96:13+flush-btrfs-1 D
[<0>] wait_current_trans+0xac/0xe0
[<0>] start_transaction+0x21b/0x4b0
[<0>] cow_file_range_inline+0x10b/0x6b0
[<0>] cow_file_range.isra.69+0x329/0x4a0
[<0>] run_delalloc_range+0x105/0x3c0
[<0>] writepage_delalloc+0x119/0x180
[<0>] __extent_writepage+0x10c/0x390
[<0>] extent_write_cache_pages+0x26f/0x3d0
[<0>] extent_writepages+0x4f/0x80
[<0>] do_writepages+0x17/0x60
[<0>] __writeback_single_inode+0x59/0x690
[<0>] writeback_sb_inodes+0x291/0x4e0
[<0>] __writeback_inodes_wb+0x87/0xb0
[<0>] wb_writeback+0x3bb/0x500
[<0>] wb_workfn+0x40d/0x610
[<0>] process_one_work+0x1ee/0x5a0
[<0>] worker_thread+0x1e0/0x3d0
[<0>] kthread+0xf5/0x130
[<0>] ret_from_fork+0x24/0x30
[<0>] 0xffffffffffffffff

Eventually, we have every process in the system waiting on
balance_dirty_pages(), and nobody is able to make progress on page
writeback.

The original patch tried to fix an OOM condition, that happened on 4.4 but no
success reproducing that on later kernels (4.19 and 4.20). This is more likely
a problem in OOM itself.

Link: https://lore.kernel.org/linux-btrfs/[email protected]/
Reported-by: Chris Mason <[email protected]>
CC: [email protected] # 4.18+
CC: ethanlien <[email protected]>
Signed-off-by: David Sterba <[email protected]>
Signed-off-by: Greg Kroah-Hartman <[email protected]>

Author: kdave

fs/btrfs/inode.c

@@ -3147,9 +3147,6 @@ static int btrfs_finish_ordered_io(struct btrfs_ordered_extent *ordered_extent)
 	/* once for the tree */
 	btrfs_put_ordered_extent(ordered_extent);
 
-	/* Try to release some metadata so we don't get an OOM but don't wait */
-	btrfs_btree_balance_dirty_nodelay(fs_info);
-
 	return ret;
 }