Skip to contentSkip to conversation

Revert "btrfs: balance dirty metadata pages in btrfs_finish_ordered_io"

PATCH New David Sterba 1 participant 1 message Jan 9, 2019 Btrfs file system

lore
Share & tools
lore

David Sterba

author maintainer
#0 lore raw

PATCH 1/1 Revert "btrfs: balance dirty metadata pages in btrfs_finish_ordered_io"

This reverts commit e73e81b6d0114d4a303205a952ab2e87c44bd279.

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.

---
 fs/btrfs/inode.c | 3 ---
 1 file changed, 3 deletions(-)
fs/btrfs/inode.c +0 -3
index 39c9585fe2a9..78205fb56a88 100644
@@ -3138,9 +3138,6 @@ 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;
 }
 
Signature · 1 line
2.20.1