btrfs: reduce lock contention with not always use keep locks when insert extent backref
PATCH 1/1 btrfs: reduce lock contention with not always use keep locks when insert extent backref
From: Robbie Ko <[email protected]>
When inserting extent backref, in order to check whether refs other than inline refs are used, we always use keep locks for tree search, which will increase the lock contention of extent-tree.
We do not need the parent node every time to determine whether normal refs are used.
It is only needed when the extent-item is the last item in leaf.
Therefore, we change to first use keep_locks=0 for search. If the extent-item happens to be the last item in leaf, we then change to keep_locks=1 for the second search to slow down lock contention.
--- fs/btrfs/extent-tree.c | 27 ++++++++++++++++++++++++--- 1 file changed, 24 insertions(+), 3 deletions(-)
fs/btrfs/extent-tree.c +24 -3
@@ -786,6 +786,8 @@ int lookup_inline_extent_backref(struct btrfs_trans_handle *trans,int ret;bool skinny_metadata =btrfs_fs_incompat (fs_info,SKINNY_METADATA );int needed; +bool keep_locks = false; +struct btrfs_key tmp_key; key.objectid = bytenr; key.type =BTRFS_EXTENT_ITEM_KEY ;
@@ -795,7 +797,6 @@ int lookup_inline_extent_backref(struct btrfs_trans_handle *trans,if (insert) { extra_size =btrfs_extent_inline_ref_size (want); path->search_for_extension =1 ; - path->keep_locks =1 ; }else extra_size =-1 ;
@@ -946,6 +947,24 @@ int lookup_inline_extent_backref(struct btrfs_trans_handle *trans, ret = -EAGAIN ; goto out; } + +if (path->slots[0 ] +1 <btrfs_header_nritems (path->nodes[0 ])) { +btrfs_item_key_to_cpu (path->nodes[0 ], &tmp_key, path->slots[0 ] +1 ); +if (tmp_key.objectid == bytenr && + tmp_key.type <BTRFS_BLOCK_GROUP_ITEM_KEY ) { + ret = -EAGAIN ; + goto out; + } + goto enoent; + } + +if (!keep_locks) { +btrfs_release_path (path); + path->keep_locks =1 ; + keep_locks = true; + goto again; + } +/* * To add new inline back ref, we have to make sure * there is no corresponding back ref item.
@@ -959,13 +978,15 @@ int lookup_inline_extent_backref(struct btrfs_trans_handle *trans, goto out; } } +enoent: *ref_ret = (struct btrfs_extent_inline_ref *)ptr; out: -if (insert) { +if (keep_locks) { path->keep_locks =0 ; - path->search_for_extension =0 ;btrfs_unlock_up_safe (path,1 ); } +if (insert) + path->search_for_extension =0 ;return ret; }
Signature · 1 line
2.17.1
Quoted 7 lines · robbieko
From: Robbie Ko <[email protected]>
When inserting extent backref, in order to check whether refs other than inline refs are used, we always use keep locks for tree search, which will increase the lock contention of extent-tree.
The line length limit is 75, you're using about 50 characters per line. You can make the lines use more characters which makes things a bit easier to read.
On the other hand the subject is kind of too long and phrased a bit oddly:
btrfs: reduce lock contention with not always use keep locks when insert extent backref
I would suggest something like:
btrfs: reduce extent tree lock contention when searching for inline backref
Quoted 4 lines
We do not need the parent node every time to determine whether normal refs are used.
It is only needed when the extent-item is the last item in leaf.
in leaf -> in a leaf
Quoted 3 lines
Therefore, we change to first use keep_locks=0 for search. If the extent-item happens to be the last item in leaf,
in leaf -> in the leaf
Quoted 2 lines
we then change to keep_locks=1 for the second search to slow down lock contention.
slow down -> reduce
Quoted 11 lines
Signed-off-by Robbie Ko <[email protected]>--- fs/btrfs/extent-tree.c | 27 ++++++++++++++++++++++++--- 1 file changed, 24 insertions(+), 3 deletions(-)fs/btrfs/extent-tree.c +0
-0
fs/btrfs/extent-tree.c int lookup_inline_extent_backref(struct btrfs_trans_handle *trans, · line 789
Quoted 4 lines
fs/btrfs/extent-tree.c +1
-0@@ -786,6 +786,8 @@ int lookup_inline_extent_backref(struct btrfs_trans_handle *trans,int ret;bool skinny_metadata =btrfs_fs_incompat (fs_info,SKINNY_METADATA );int needed; +bool keep_locks = false;
There's no need to use this variable, we can use path->keep_locks directly.
fs/btrfs/extent-tree.c int lookup_inline_extent_backref(struct btrfs_trans_handle *trans, · line 790
Quoted 1 line
fs/btrfs/extent-tree.c +1
-0+struct btrfs_key tmp_key;
Please move the declaration of tmp_key to the block where it's used, see below.
Quoted 12 lines
fs/btrfs/extent-tree.c +0
-1key.objectid = bytenr; key.type =BTRFS_EXTENT_ITEM_KEY ;@@ -795,7 +797,6 @@ int lookup_inline_extent_backref(struct btrfs_trans_handle *trans,if (insert) { extra_size =btrfs_extent_inline_ref_size (want); path->search_for_extension =1 ; - path->keep_locks =1 ; }else extra_size =-1 ;
fs/btrfs/extent-tree.c int lookup_inline_extent_backref(struct btrfs_trans_handle *trans, · line 951
Quoted 5 lines
fs/btrfs/extent-tree.c +2
-0@@ -946,6 +947,24 @@ int lookup_inline_extent_backref(struct btrfs_trans_handle *trans, ret = -EAGAIN ; goto out; } + +if (path->slots[0 ] +1 <btrfs_header_nritems (path->nodes[0 ])) {
Here place the declaration of tmp_key, since it's not used outside this block.
Everything else looks good, thanks.
Quoted 41 lines
fs/btrfs/extent-tree.c +20
-3+btrfs_item_key_to_cpu (path->nodes[0 ], &tmp_key, path->slots[0 ] +1 ); +if (tmp_key.objectid == bytenr && + tmp_key.type <BTRFS_BLOCK_GROUP_ITEM_KEY ) { + ret = -EAGAIN ; + goto out; + } + goto enoent; + } + +if (!keep_locks) { +btrfs_release_path (path); + path->keep_locks =1 ; + keep_locks = true; + goto again; + } +/* * To add new inline back ref, we have to make sure * there is no corresponding back ref item. @@ -959,13 +978,15 @@ int lookup_inline_extent_backref(struct btrfs_trans_handle *trans, goto out; } } +enoent: *ref_ret = (struct btrfs_extent_inline_ref *)ptr; out: -if (insert) { +if (keep_locks) { path->keep_locks =0 ; - path->search_for_extension =0 ;btrfs_unlock_up_safe (path,1 ); } +if (insert) + path->search_for_extension =0 ;return ret; } --2.17.1
Filipe Manana 於 2024/10/23 下午6:34 寫道:
Quoted 11 lines · robbieko
Quoted 6 lines
From: Robbie Ko <[email protected]>
When inserting extent backref, in order to check whether refs other than inline refs are used, we always use keep locks for tree search, which will increase the lock contention of extent-tree.
The line length limit is 75, you're using about 50 characters per line. You can make the lines use more characters which makes things a bit easier to read.
Quoted 8 lines
On the other hand the subject is kind of too long and phrased a bit oddly:
btrfs: reduce lock contention with not always use keep locks when insert extent backref
I would suggest something like:
btrfs: reduce extent tree lock contention when searching for inline backref
Ok, I will modify and send a new patch again.
Thanks.
Quoted 93 lines
Quoted 3 lines
We do not need the parent node every time to determine whether normal refs are used.
It is only needed when the extent-item is the last item in leaf.in leaf -> in a leaf
Quoted 2 lines
Therefore, we change to first use keep_locks=0 for search. If the extent-item happens to be the last item in leaf,
in leaf -> in the leaf
Quoted 2 lines
we then change to keep_locks=1 for the second search to slow down lock contention.
slow down -> reduce
Quoted 10 lines
Signed-off-by Robbie Ko <[email protected]>--- fs/btrfs/extent-tree.c | 27 ++++++++++++++++++++++++--- 1 file changed, 24 insertions(+), 3 deletions(-)fs/btrfs/extent-tree.c +0
-0fs/btrfs/extent-tree.c int lookup_inline_extent_backref(struct btrfs_trans_handle *trans, · line 789
Quoted 4 lines
fs/btrfs/extent-tree.c +1
-0@@ -786,6 +786,8 @@ int lookup_inline_extent_backref(struct btrfs_trans_handle *trans,int ret;bool skinny_metadata =btrfs_fs_incompat (fs_info,SKINNY_METADATA );int needed; +bool keep_locks = false;There's no need to use this variable, we can use
path->keep_locksdirectly.fs/btrfs/extent-tree.c int lookup_inline_extent_backref(struct btrfs_trans_handle *trans, · line 790
Quoted 1 line
fs/btrfs/extent-tree.c +1
-0+struct btrfs_key tmp_key;Please move the declaration of
tmp_keyto the block where it's used, see below.Quoted 11 lines
fs/btrfs/extent-tree.c +0
-1key.objectid = bytenr; key.type =BTRFS_EXTENT_ITEM_KEY ;@@ -795,7 +797,6 @@ int lookup_inline_extent_backref(struct btrfs_trans_handle *trans,if (insert) { extra_size =btrfs_extent_inline_ref_size (want); path->search_for_extension =1 ; - path->keep_locks =1 ; }else extra_size =-1 ;fs/btrfs/extent-tree.c int lookup_inline_extent_backref(struct btrfs_trans_handle *trans, · line 951
Quoted 5 lines
fs/btrfs/extent-tree.c +2
-0@@ -946,6 +947,24 @@ int lookup_inline_extent_backref(struct btrfs_trans_handle *trans, ret = -EAGAIN ; goto out; } + +if (path->slots[0 ] +1 <btrfs_header_nritems (path->nodes[0 ])) {Here place the declaration of
tmp_key, since it's not used outside this block.Everything else looks good, thanks.
Quoted 41 lines
fs/btrfs/extent-tree.c +20
-3+btrfs_item_key_to_cpu (path->nodes[0 ], &tmp_key, path->slots[0 ] +1 ); +if (tmp_key.objectid == bytenr && + tmp_key.type <BTRFS_BLOCK_GROUP_ITEM_KEY ) { + ret = -EAGAIN ; + goto out; + } + goto enoent; + } + +if (!keep_locks) { +btrfs_release_path (path); + path->keep_locks =1 ; + keep_locks = true; + goto again; + } +/* * To add new inline back ref, we have to make sure * there is no corresponding back ref item. @@ -959,13 +978,15 @@ int lookup_inline_extent_backref(struct btrfs_trans_handle *trans, goto out; } } +enoent: *ref_ret = (struct btrfs_extent_inline_ref *)ptr; out: -if (insert) { +if (keep_locks) { path->keep_locks =0 ; - path->search_for_extension =0 ;btrfs_unlock_up_safe (path,1 ); } +if (insert) + path->search_for_extension =0 ;return ret; } --2.17.1