Commit 131e404a authored by Filipe David Borba Manana's avatar Filipe David Borba Manana Committed by Chris Mason
Browse files

Btrfs: fix very slow inode eviction and fs unmount 



The inode eviction can be very slow, because during eviction we
tell the VFS to truncate all of the inode's pages. This results
in calls to btrfs_invalidatepage() which in turn does calls to
lock_extent_bits() and clear_extent_bit(). These calls result in
too many merges and splits of extent_state structures, which
consume a lot of time and cpu when the inode has many pages. In
some scenarios I have experienced umount times higher than 15
minutes, even when there's no pending IO (after a btrfs fs sync).

A quick way to reproduce this issue:

$ mkfs.btrfs -f /dev/sdb3
$ mount /dev/sdb3 /mnt/btrfs
$ cd /mnt/btrfs
$ sysbench --test=fileio --file-num=128 --file-total-size=16G \
    --file-test-mode=seqwr --num-threads=128 \
    --file-block-size=16384 --max-time=60 --max-requests=0 run
$ time btrfs fi sync .
FSSync '.'

real	0m25.457s
user	0m0.000s
sys	0m0.092s
$ cd ..
$ time umount /mnt/btrfs

real	1m38.234s
user	0m0.000s
sys	1m25.760s

The same test on ext4 runs much faster:

$ mkfs.ext4 /dev/sdb3
$ mount /dev/sdb3 /mnt/ext4
$ cd /mnt/ext4
$ sysbench --test=fileio --file-num=128 --file-total-size=16G \
    --file-test-mode=seqwr --num-threads=128 \
    --file-block-size=16384 --max-time=60 --max-requests=0 run
$ sync
$ cd ..
$ time umount /mnt/ext4

real	0m3.626s
user	0m0.004s
sys	0m3.012s

After this patch, the unmount (inode evictions) is much faster:

$ mkfs.btrfs -f /dev/sdb3
$ mount /dev/sdb3 /mnt/btrfs
$ cd /mnt/btrfs
$ sysbench --test=fileio --file-num=128 --file-total-size=16G \
    --file-test-mode=seqwr --num-threads=128 \
    --file-block-size=16384 --max-time=60 --max-requests=0 run
$ time btrfs fi sync .
FSSync '.'

real	0m26.774s
user	0m0.000s
sys	0m0.084s
$ cd ..
$ time umount /mnt/btrfs

real	0m1.811s
user	0m0.000s
sys	0m1.564s

Signed-off-by: default avatarFilipe David Borba Manana <fdmanana@gmail.com>
Signed-off-by: default avatarJosef Bacik <jbacik@fb.com>
Signed-off-by: default avatarChris Mason <clm@fb.com>
parent 0647bf56

fs/btrfs/inode.c

+84 −14
Original line number Diff line number Diff line
@@ -4488,6 +4488,62 @@ static int btrfs_setattr(struct dentry *dentry, struct iattr *attr) 
	return err;

}



/*

 * While truncating the inode pages during eviction, we get the VFS calling

 * btrfs_invalidatepage() against each page of the inode. This is slow because

 * the calls to btrfs_invalidatepage() result in a huge amount of calls to

 * lock_extent_bits() and clear_extent_bit(), which keep merging and splitting

 * extent_state structures over and over, wasting lots of time.

 *

 * Therefore if the inode is being evicted, let btrfs_invalidatepage() skip all

 * those expensive operations on a per page basis and do only the ordered io

 * finishing, while we release here the extent_map and extent_state structures,

 * without the excessive merging and splitting.

 */

static void evict_inode_truncate_pages(struct inode *inode)

{

	struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;

	struct extent_map_tree *map_tree = &BTRFS_I(inode)->extent_tree;

	struct rb_node *node;



	ASSERT(inode->i_state & I_FREEING);

	truncate_inode_pages(&inode->i_data, 0);



	write_lock(&map_tree->lock);

	while (!RB_EMPTY_ROOT(&map_tree->map)) {

		struct extent_map *em;



		node = rb_first(&map_tree->map);

		em = rb_entry(node, struct extent_map, rb_node);

		remove_extent_mapping(map_tree, em);

		free_extent_map(em);

	}

	write_unlock(&map_tree->lock);



	spin_lock(&io_tree->lock);

	while (!RB_EMPTY_ROOT(&io_tree->state)) {

		struct extent_state *state;

		struct extent_state *cached_state = NULL;



		node = rb_first(&io_tree->state);

		state = rb_entry(node, struct extent_state, rb_node);

		atomic_inc(&state->refs);

		spin_unlock(&io_tree->lock);



		lock_extent_bits(io_tree, state->start, state->end,

				 0, &cached_state);

		clear_extent_bit(io_tree, state->start, state->end,

				 EXTENT_LOCKED | EXTENT_DIRTY |

				 EXTENT_DELALLOC | EXTENT_DO_ACCOUNTING |

				 EXTENT_DEFRAG, 1, 1,

				 &cached_state, GFP_NOFS);

		free_extent_state(state);



		spin_lock(&io_tree->lock);

	}

	spin_unlock(&io_tree->lock);

}



void btrfs_evict_inode(struct inode *inode)

{

	struct btrfs_trans_handle *trans;
@@ -4498,7 +4554,8 @@ void btrfs_evict_inode(struct inode *inode) 


	trace_btrfs_inode_evict(inode);



	truncate_inode_pages(&inode->i_data, 0);

	evict_inode_truncate_pages(inode);



	if (inode->i_nlink &&

	    ((btrfs_root_refs(&root->root_item) != 0 &&

	      root->root_key.objectid != BTRFS_ROOT_TREE_OBJECTID) ||
@@ -7379,6 +7436,7 @@ static void btrfs_invalidatepage(struct page *page, unsigned int offset, 
	struct extent_state *cached_state = NULL;

	u64 page_start = page_offset(page);

	u64 page_end = page_start + PAGE_CACHE_SIZE - 1;

	int inode_evicting = inode->i_state & I_FREEING;



	/*

	 * we have the page locked, so new writeback can't start,
@@ -7394,17 +7452,21 @@ static void btrfs_invalidatepage(struct page *page, unsigned int offset, 
		btrfs_releasepage(page, GFP_NOFS);

		return;

	}



	if (!inode_evicting)

		lock_extent_bits(tree, page_start, page_end, 0, &cached_state);

	ordered = btrfs_lookup_ordered_extent(inode, page_offset(page));

	ordered = btrfs_lookup_ordered_extent(inode, page_start);

	if (ordered) {

		/*

		 * IO on this page will never be started, so we need

		 * to account for any ordered extents now

		 */

		if (!inode_evicting)

			clear_extent_bit(tree, page_start, page_end,

					 EXTENT_DIRTY | EXTENT_DELALLOC |

					 EXTENT_LOCKED | EXTENT_DO_ACCOUNTING |

				 EXTENT_DEFRAG, 1, 0, &cached_state, GFP_NOFS);

					 EXTENT_DEFRAG, 1, 0, &cached_state,

					 GFP_NOFS);

		/*

		 * whoever cleared the private bit is responsible

		 * for the finish_ordered_io
@@ -7428,14 +7490,22 @@ static void btrfs_invalidatepage(struct page *page, unsigned int offset, 
				btrfs_finish_ordered_io(ordered);

		}

		btrfs_put_ordered_extent(ordered);

		if (!inode_evicting) {

			cached_state = NULL;

		lock_extent_bits(tree, page_start, page_end, 0, &cached_state);

			lock_extent_bits(tree, page_start, page_end, 0,

					 &cached_state);

		}

	}



	if (!inode_evicting) {

		clear_extent_bit(tree, page_start, page_end,

		 EXTENT_LOCKED | EXTENT_DIRTY | EXTENT_DELALLOC |

		 EXTENT_DO_ACCOUNTING | EXTENT_DEFRAG, 1, 1,

				 EXTENT_LOCKED | EXTENT_DIRTY |

				 EXTENT_DELALLOC | EXTENT_DO_ACCOUNTING |

				 EXTENT_DEFRAG, 1, 1,

				 &cached_state, GFP_NOFS);



		__btrfs_releasepage(page, GFP_NOFS);

	}



	ClearPageChecked(page);

	if (PagePrivate(page)) {

CRA Git | Maintained and supported by SUSTech CRA and CCSE