btrfs: speedup mount time with force readahead chunk tree
PATCH 1/1 btrfs: speedup mount time with force readahead chunk tree
From: Robbie Ko <[email protected]>
When mounting, we always need to read the whole chunk tree, when there are too many chunk items, most of the time is spent on btrfs_read_chunk_tree, because we only read one leaf at a time.
We fix this by adding a new readahead mode READA_FORWARD_FORCE, which reads all the leaves after the key in the node when reading a level 1 node.
--- fs/btrfs/ctree.c | 7 +++++-- fs/btrfs/ctree.h | 2 +- fs/btrfs/volumes.c | 1 + 3 files changed, 7 insertions(+), 3 deletions(-)
fs/btrfs/ctree.c +5 -2
@@ -2194,7 +2194,7 @@ static void reada_for_search(struct btrfs_fs_info *fs_info,if (nr ==0 )break ; nr--; - }else if (path->reada ==READA_FORWARD ) { + }else if (path->reada ==READA_FORWARD || path->reada ==READA_FORWARD_FORCE ) { nr++;if (nr >= nritems)break ;
@@ -2205,12 +2205,15 @@ static void reada_for_search(struct btrfs_fs_info *fs_info,break ; } search =btrfs_node_blockptr (node, nr); -if ((search <= target && target - search <=65536 ) || +if ((path->reada ==READA_FORWARD_FORCE ) || + (search <= target && target - search <=65536 ) || (search > target && search - target <=65536 )) {readahead_tree_block (fs_info, search); nread += blocksize; } nscan++; +if (path->reada ==READA_FORWARD_FORCE ) +continue ;if ((nread >65536 || nscan >32 ))break ; }
fs/btrfs/ctree.h +1 -1
@@ -353,7 +353,7 @@ struct btrfs_node {* The slots array records the index of the item or block pointer * used while walking the tree. */ -enum {READA_NONE ,READA_BACK ,READA_FORWARD }; +enum {READA_NONE ,READA_BACK ,READA_FORWARD ,READA_FORWARD_FORCE };struct btrfs_path {struct extent_buffer *nodes[BTRFS_MAX_LEVEL ];int slots[BTRFS_MAX_LEVEL ];
fs/btrfs/volumes.c +1 -0
@@ -7043,6 +7043,7 @@ int btrfs_read_chunk_tree(struct btrfs_fs_info *fs_info) path =btrfs_alloc_path ();if (!path)return -ENOMEM ; + path->reada =READA_FORWARD_FORCE ;/* * uuid_mutex is needed only if we are mounting a sprout FS
Signature · 1 line
2.17.1
Does anyone have any suggestions?
robbieko 於 2020/7/1 下午5:29 寫道:
Quoted 73 lines
From: Robbie Ko <[email protected]>
When mounting, we always need to read the whole chunk tree, when there are too many chunk items, most of the time is spent on
btrfs_read_chunk_tree, because we only read one leaf at a time.We fix this by adding a new readahead mode
READA_FORWARD_FORCE, which reads all the leaves after the key in the node when reading a level 1 node.Signed-off-by Robbie Ko <[email protected]>--- fs/btrfs/ctree.c | 7 +++++-- fs/btrfs/ctree.h | 2 +- fs/btrfs/volumes.c | 1 + 3 files changed, 7 insertions(+), 3 deletions(-)fs/btrfs/ctree.c +5
-2@@ -2194,7 +2194,7 @@ static void reada_for_search(struct btrfs_fs_info *fs_info,if (nr ==0 )break ; nr--; - }else if (path->reada ==READA_FORWARD ) { + }else if (path->reada ==READA_FORWARD || path->reada ==READA_FORWARD_FORCE ) { nr++;if (nr >= nritems)break ;@@ -2205,12 +2205,15 @@ static void reada_for_search(struct btrfs_fs_info *fs_info,break ; } search =btrfs_node_blockptr (node, nr); -if ((search <= target && target - search <=65536 ) || +if ((path->reada ==READA_FORWARD_FORCE ) || + (search <= target && target - search <=65536 ) || (search > target && search - target <=65536 )) {readahead_tree_block (fs_info, search); nread += blocksize; } nscan++; +if (path->reada ==READA_FORWARD_FORCE ) +continue ;if ((nread >65536 || nscan >32 ))break ; }fs/btrfs/ctree.h +1
-1@@ -353,7 +353,7 @@ struct btrfs_node {* The slots array records the index of the item or block pointer * used while walking the tree. */ -enum {READA_NONE ,READA_BACK ,READA_FORWARD }; +enum {READA_NONE ,READA_BACK ,READA_FORWARD ,READA_FORWARD_FORCE };struct btrfs_path {struct extent_buffer *nodes[BTRFS_MAX_LEVEL ];int slots[BTRFS_MAX_LEVEL ];fs/btrfs/volumes.c +1
-0@@ -7043,6 +7043,7 @@ int btrfs_read_chunk_tree(struct btrfs_fs_info *fs_info) path =btrfs_alloc_path ();if (!path)return -ENOMEM ; + path->reada =READA_FORWARD_FORCE ;/* * uuid_mutex is needed only if we are mounting a sprout FS
Quoted 1 line · Robbie Ko
Does anyone have any suggestions?
I believe David's suggestion on using regular readahead is already good enough for chunk tree.
Especially since chunk tree is not really the main cause for slow mount.
Thanks,
Qu
Quoted 78 lines
robbieko 於 2020/7/1 下午5:29 寫道:
Quoted 76 lines
From: Robbie Ko <[email protected]>
When mounting, we always need to read the whole chunk tree, when there are too many chunk items, most of the time is spent on
btrfs_read_chunk_tree, because we only read one leaf at a time.We fix this by adding a new readahead mode
READA_FORWARD_FORCE, which reads all the leaves after the key in the node when reading a level 1 node.Signed-off-by Robbie Ko <[email protected]>---
fs/btrfs/ctree.c | 7 +++++--
fs/btrfs/ctree.h | 2 +-
fs/btrfs/volumes.c| 1 +3 files changed, 7 insertions(+), 3 deletions(-)fs/btrfs/ctree.c +1
-1@@ -2194,7 +2194,7 @@ static void reada_for_search(struct btrfs_fs_info *fs_info, if (nr == 0) break; nr--; - } else if (path->reada == READA_FORWARD) { + } else if (path->reada == READA_FORWARD || path->reada ==READA_FORWARD_FORCE ) { nr++;if (nr >= nritems)
break;fs/btrfs/ctree.c +4
-1@@ -2205,12 +2205,15 @@ static void reada_for_search(structbtrfs_fs_info *fs_info, break; } search = btrfs_node_blockptr(node, nr); - if ((search <= target && target - search <= 65536) || + if ((path->reada == READA_FORWARD_FORCE) || + (search <= target && target - search <= 65536) || (search > target && search - target <= 65536)) { readahead_tree_block(fs_info, search); nread += blocksize; } nscan++; + if (path->reada == READA_FORWARD_FORCE) + continue; if ((nread > 65536 || nscan > 32)) break;}
fs/btrfs/ctree.h +1
-1@@ -353,7 +353,7 @@ struct btrfs_node {* The slots array records the index of the item or block pointer * used while walking the tree. */ -enum {READA_NONE ,READA_BACK ,READA_FORWARD }; +enum {READA_NONE ,READA_BACK ,READA_FORWARD ,READA_FORWARD_FORCE }; struct btrfs_path { struct extent_buffer *nodes[BTRFS_MAX_LEVEL]; int slots[BTRFS_MAX_LEVEL];fs/btrfs/volumes.c +1
-0@@ -7043,6 +7043,7 @@ int btrfs_read_chunk_tree(struct btrfs_fs_info *fs_info) path = btrfs_alloc_path(); if (!path) return -ENOMEM; + path->reada = READA_FORWARD_FORCE; /** uuid_mutex is needed only if we are mounting a sprout FS
I've known btrfs_read_block_groups for a long time,
we can use BG_TREE freature to speed up btrfs_read_block_groups.
https://lwn.net/Articles/801990/
But reading the chunk tree also takes some time,
we can speed up the chunk tree by using the readahead mechanism.
Why we not just use regular forward readahead?
- Because the regular forward readahead ,
reads only the logical address adjacent to the 64k, but the logical address of the next leaf may not be in 64k.
I have a test environment as follows:
200TB btrfs volume: used 192TB
Data, single: total=192.00TiB, used=192.00TiB System, DUP: total=40.00MiB, used=19.91MiB Metadata, DUP: total=63.00GiB, used=46.46GiB GlobalReserve, single: total=2.00GiB, used=0.00B
chunk tree level : 2
chunk tree tree:
nodes: 4
leaves: 1270
total: 1274
chunk tree size: 19.9 MB
SYSTEM chunks count : 2 (8MB, 32MB)
btrfs_read_chunk_tree spends the following time :
before: 1.89s
patch: 0.27s
Speed increase of about 85%.
Between the chunk tree leaves, there will be a different SYSTEM chunk,
when the The more frequent the chunk allocate/remove, the larger the gap between the leaves.
e.g. chunk tree node :
key (FIRST_CHUNK_TREE CHUNK_ITEM 85020014280704) block 79866020003840 (4874635010) gen 12963 key (FIRST_CHUNK_TREE CHUNK_ITEM 85185370521600) block 28999680 (1770) gen 12964 key (FIRST_CHUNK_TREE CHUNK_ITEM 85351800504320) block 79866020347904 (4874635031) gen 12964 key (FIRST_CHUNK_TREE CHUNK_ITEM 85518230487040) block 79866020102144 (4874635016) gen 12964 key (FIRST_CHUNK_TREE CHUNK_ITEM 85684660469760) block 79866020118528 (4874635017) gen 12964 key (FIRST_CHUNK_TREE CHUNK_ITEM 85851090452480) block 79866020134912 (4874635018) gen 12964 key (FIRST_CHUNK_TREE CHUNK_ITEM 86017520435200) block 29261824 (1786) gen 12964 key (FIRST_CHUNK_TREE CHUNK_ITEM 86183950417920) block 79866020397056 (4874635034) gen 12965 key (FIRST_CHUNK_TREE CHUNK_ITEM 86350380400640) block 79866020151296 (4874635019) gen 12965 key (FIRST_CHUNK_TREE CHUNK_ITEM 86516810383360) block 79866020167680 (4874635020) gen 12965 key (FIRST_CHUNK_TREE CHUNK_ITEM 86683240366080) block 79866020184064 (4874635021) gen 12965 key (FIRST_CHUNK_TREE CHUNK_ITEM 86849670348800) block 79866020200448 (4874635022) gen 12965 key (FIRST_CHUNK_TREE CHUNK_ITEM 87016100331520) block 29065216 (1774) gen 12966 key (FIRST_CHUNK_TREE CHUNK_ITEM 87182530314240) block 79866020315136 (4874635029) gen 12966 key (FIRST_CHUNK_TREE CHUNK_ITEM 87348960296960) block 79866020331520 (4874635030) gen 12966 key (FIRST_CHUNK_TREE CHUNK_ITEM 87515390279680) block 79866020413440 (4874635035) gen 12966 key (FIRST_CHUNK_TREE CHUNK_ITEM 87681820262400) block 79866020429824 (4874635036) gen 12966 key (FIRST_CHUNK_TREE CHUNK_ITEM 87848250245120) block 29294592 (1788) gen 12968 key (FIRST_CHUNK_TREE CHUNK_ITEM 88014680227840) block 79866020544512 (4874635043) gen 12968
With 1PB of btrfs volume, we will have more SYSTEM CHUNK,
and each chunk tree leaf will be more fragmented,
and the time difference will be larger.
Qu Wenruo 於 2020/7/6 下午2:16 寫道:
Quoted 88 lines · Robbie Ko
Quoted 1 line
Does anyone have any suggestions?
I believe David's suggestion on using regular readahead is already good enough for chunk tree.
Especially since chunk tree is not really the main cause for slow mount.
Thanks,
QuQuoted 77 lines
robbieko 於 2020/7/1 下午5:29 寫道:
Quoted 76 lines
From: Robbie Ko <[email protected]>
When mounting, we always need to read the whole chunk tree, when there are too many chunk items, most of the time is spent on
btrfs_read_chunk_tree, because we only read one leaf at a time.We fix this by adding a new readahead mode
READA_FORWARD_FORCE, which reads all the leaves after the key in the node when reading a level 1 node.Signed-off-by Robbie Ko <[email protected]>--- fs/btrfs/ctree.c | 7 +++++-- fs/btrfs/ctree.h | 2 +- fs/btrfs/volumes.c | 1 + 3 files changed, 7 insertions(+), 3 deletions(-)fs/btrfs/ctree.c +1
-1@@ -2194,7 +2194,7 @@ static void reada_for_search(struct btrfs_fs_info *fs_info, if (nr == 0) break; nr--; - } else if (path->reada == READA_FORWARD) { + } else if (path->reada == READA_FORWARD || path->reada ==READA_FORWARD_FORCE ) { nr++;if (nr >= nritems)
break;fs/btrfs/ctree.c +4
-1@@ -2205,12 +2205,15 @@ static void reada_for_search(structbtrfs_fs_info *fs_info, break; } search = btrfs_node_blockptr(node, nr); - if ((search <= target && target - search <= 65536) || + if ((path->reada == READA_FORWARD_FORCE) || + (search <= target && target - search <= 65536) || (search > target && search - target <= 65536)) { readahead_tree_block(fs_info, search); nread += blocksize; } nscan++; + if (path->reada == READA_FORWARD_FORCE) + continue; if ((nread > 65536 || nscan > 32)) break;}
fs/btrfs/ctree.h +1
-1@@ -353,7 +353,7 @@ struct btrfs_node {* The slots array records the index of the item or block pointer * used while walking the tree. */ -enum {READA_NONE ,READA_BACK ,READA_FORWARD }; +enum {READA_NONE ,READA_BACK ,READA_FORWARD ,READA_FORWARD_FORCE }; struct btrfs_path { struct extent_buffer *nodes[BTRFS_MAX_LEVEL]; int slots[BTRFS_MAX_LEVEL];fs/btrfs/volumes.c +1
-0@@ -7043,6 +7043,7 @@ int btrfs_read_chunk_tree(struct btrfs_fs_info *fs_info) path = btrfs_alloc_path(); if (!path) return -ENOMEM; + path->reada = READA_FORWARD_FORCE; /** uuid_mutex is needed only if we are mounting a sprout FS
Quoted 8 lines · Robbie Ko
I've known
btrfs_read_block_groupsfor a long time,we can use
BG_TREEfreature to speed upbtrfs_read_block_groups.
Quoted 8 lines
But reading the chunk tree also takes some time,
we can speed up the chunk tree by using the readahead mechanism.
Why we not just use regular forward readahead?
- Because the regular forward readahead ,
reads only the logical address adjacent to the 64k, but the logical address of the next leaf may not be in 64k.
Great, please add these into the changelog for the need of READA_FORWARD_FORCE.
But on the other hand, it looks like we could change READA_FORWARD without introducing the _FORCE version,
as _FORCE variant is what we really want.
That logical bytenr limit looks not reasonable to me, which makes that READA_FORWARD near useless to me.
Although in that direction, we also need to verify all existing READA_FORWARD are really correctly.
Personally I tend to prefer change existing READA_FORWARD and review all existing callers.
Quoted 18 lines
I have a test environment as follows:
200TB btrfs volume: used 192TB
Data, single: total=192.00TiB, used=192.00TiB System, DUP: total=40.00MiB, used=19.91MiB Metadata, DUP: total=63.00GiB, used=46.46GiB GlobalReserve, single: total=2.00GiB, used=0.00B
chunk tree level : 2
chunk tree tree:
nodes: 4
leaves: 1270
total: 1274
chunk tree size: 19.9 MB
SYSTEM chunks count : 2 (8MB, 32MB)
Quoted 5 lines
btrfs_read_chunk_treespends the following time :before: 1.89s
patch: 0.27s
That's great, although 2s is already fast enough to me :)
Thanks,
Qu
Quoted 145 lines
Speed increase of about 85%.
Between the chunk tree leaves, there will be a different SYSTEM chunk,
when the The more frequent the chunk allocate/remove, the larger the gap between the leaves.
e.g. chunk tree node :
key (FIRST_CHUNK_TREECHUNK_ITEM85020014280704) block 79866020003840 (4874635010) gen 12963
key (FIRST_CHUNK_TREECHUNK_ITEM85185370521600) block 28999680 (1770) gen 12964
key (FIRST_CHUNK_TREECHUNK_ITEM85351800504320) block 79866020347904 (4874635031) gen 12964
key (FIRST_CHUNK_TREECHUNK_ITEM85518230487040) block 79866020102144 (4874635016) gen 12964
key (FIRST_CHUNK_TREECHUNK_ITEM85684660469760) block 79866020118528 (4874635017) gen 12964
key (FIRST_CHUNK_TREECHUNK_ITEM85851090452480) block 79866020134912 (4874635018) gen 12964
key (FIRST_CHUNK_TREECHUNK_ITEM86017520435200) block 29261824 (1786) gen 12964
key (FIRST_CHUNK_TREECHUNK_ITEM86183950417920) block 79866020397056 (4874635034) gen 12965
key (FIRST_CHUNK_TREECHUNK_ITEM86350380400640) block 79866020151296 (4874635019) gen 12965
key (FIRST_CHUNK_TREECHUNK_ITEM86516810383360) block 79866020167680 (4874635020) gen 12965
key (FIRST_CHUNK_TREECHUNK_ITEM86683240366080) block 79866020184064 (4874635021) gen 12965
key (FIRST_CHUNK_TREECHUNK_ITEM86849670348800) block 79866020200448 (4874635022) gen 12965
key (FIRST_CHUNK_TREECHUNK_ITEM87016100331520) block 29065216 (1774) gen 12966
key (FIRST_CHUNK_TREECHUNK_ITEM87182530314240) block 79866020315136 (4874635029) gen 12966
key (FIRST_CHUNK_TREECHUNK_ITEM87348960296960) block 79866020331520 (4874635030) gen 12966
key (FIRST_CHUNK_TREECHUNK_ITEM87515390279680) block 79866020413440 (4874635035) gen 12966
key (FIRST_CHUNK_TREECHUNK_ITEM87681820262400) block 79866020429824 (4874635036) gen 12966
key (FIRST_CHUNK_TREECHUNK_ITEM87848250245120) block 29294592 (1788) gen 12968
key (FIRST_CHUNK_TREECHUNK_ITEM88014680227840) block 79866020544512 (4874635043) gen 12968With 1PB of btrfs volume, we will have more SYSTEM CHUNK,
and each chunk tree leaf will be more fragmented,
and the time difference will be larger.
Qu Wenruo 於 2020/7/6 下午2:16 寫道:
Quoted 88 lines · Robbie Ko
Quoted 1 line
Does anyone have any suggestions?
I believe David's suggestion on using regular readahead is already good enough for chunk tree.
Especially since chunk tree is not really the main cause for slow mount.
Thanks,
QuQuoted 77 lines
robbieko 於 2020/7/1 下午5:29 寫道:
Quoted 76 lines
From: Robbie Ko <[email protected]>
When mounting, we always need to read the whole chunk tree, when there are too many chunk items, most of the time is spent on
btrfs_read_chunk_tree, because we only read one leaf at a time.We fix this by adding a new readahead mode
READA_FORWARD_FORCE, which reads all the leaves after the key in the node when reading a level 1 node.Signed-off-by Robbie Ko <[email protected]>---
fs/btrfs/ctree.c | 7 +++++--
fs/btrfs/ctree.h | 2 +-
fs/btrfs/volumes.c| 1 +3 files changed, 7 insertions(+), 3 deletions(-)fs/btrfs/ctree.c +1
-1@@ -2194,7 +2194,7 @@ static void reada_for_search(struct btrfs_fs_info *fs_info, if (nr == 0) break; nr--; - } else if (path->reada == READA_FORWARD) { + } else if (path->reada == READA_FORWARD || path->reada ==READA_FORWARD_FORCE ) { nr++;if (nr >= nritems)
break;fs/btrfs/ctree.c +4
-1@@ -2205,12 +2205,15 @@ static void reada_for_search(structbtrfs_fs_info *fs_info, break; } search = btrfs_node_blockptr(node, nr); - if ((search <= target && target - search <= 65536) || + if ((path->reada == READA_FORWARD_FORCE) || + (search <= target && target - search <= 65536) || (search > target && search - target <= 65536)) { readahead_tree_block(fs_info, search); nread += blocksize; } nscan++; + if (path->reada == READA_FORWARD_FORCE) + continue; if ((nread > 65536 || nscan > 32)) break;}
fs/btrfs/ctree.h +1
-1@@ -353,7 +353,7 @@ struct btrfs_node {* The slots array records the index of the item or block pointer * used while walking the tree. */ -enum {READA_NONE ,READA_BACK ,READA_FORWARD }; +enum {READA_NONE ,READA_BACK ,READA_FORWARD ,READA_FORWARD_FORCE }; struct btrfs_path { struct extent_buffer *nodes[BTRFS_MAX_LEVEL]; int slots[BTRFS_MAX_LEVEL];fs/btrfs/volumes.c +1
-0@@ -7043,6 +7043,7 @@ int btrfs_read_chunk_tree(struct btrfs_fs_info *fs_info) path = btrfs_alloc_path(); if (!path) return -ENOMEM; + path->reada = READA_FORWARD_FORCE; /** uuid_mutex is needed only if we are mounting a sprout FS
Quoted 15 lines · Qu Wenruo
Quoted 12 lines · Robbie Ko
I've known
btrfs_read_block_groupsfor a long time,we can use
BG_TREEfreature to speed upbtrfs_read_block_groups.https://lwn.net/Articles/801990/
But reading the chunk tree also takes some time,
we can speed up the chunk tree by using the readahead mechanism.
Quoted 7 lines
Quoted 4 lines
Why we not just use regular forward readahead?
- Because the regular forward readahead ,
reads only the logical address adjacent to the 64k, but the logical address of the next leaf may not be in 64k.Great, please add these into the changelog for the need of
READA_FORWARD_FORCE.
<nod> Performance patches should come with data backing their performance improvements claims.
Quoted 6 lines
But on the other hand, it looks like we could change
READA_FORWARDwithout introducing the_FORCEversion,
as_FORCEvariant is what we really want.
Quoted 8 lines
That logical bytenr limit looks not reasonable to me, which makes that
READA_FORWARDnear useless to me.Although in that direction, we also need to verify all existing
READA_FORWARDare really correctly.Personally I tend to prefer change existing
READA_FORWARDand review all existing callers.
I agree, we should ideally revise the existing READA_FORWARD and give it better semantics if the current ones are needlessly rigid. Only if this improves way too cumbersome i.e breaking assumption which are made should we introduce yet another mode. Because this other mode is really just special casing and we should avoid special cases as much as possible.
<snip>
Nikolay Borisov 於 2020/7/6 下午4:37 寫道:
Quoted 15 lines · Qu Wenruo
Quoted 13 lines · Robbie Ko
Quoted 11 lines
I've known
btrfs_read_block_groupsfor a long time,we can use
BG_TREEfreature to speed upbtrfs_read_block_groups.https://lwn.net/Articles/801990/
But reading the chunk tree also takes some time,
we can speed up the chunk tree by using the readahead mechanism.
Quoted 8 lines
Quoted 6 lines
Quoted 4 lines
Why we not just use regular forward readahead?
- Because the regular forward readahead ,
reads only the logical address adjacent to the 64k, but the logical address of the next leaf may not be in 64k.Great, please add these into the changelog for the need of
READA_FORWARD_FORCE.<nod> Performance patches should come with data backing their performance improvements claims.
Ok, I will add these into the changelog.
Quoted 12 lines
Quoted 12 lines
But on the other hand, it looks like we could change
READA_FORWARDwithout introducing the_FORCEversion,
as_FORCEvariant is what we really want.That logical bytenr limit looks not reasonable to me, which makes that
READA_FORWARDnear useless to me.Although in that direction, we also need to verify all existing
READA_FORWARDare really correctly.
Quoted 7 lines
Quoted 2 lines
Personally I tend to prefer change existing
READA_FORWARDand review all existing callers.I agree, we should ideally revise the existing
READA_FORWARDand give it better semantics if the current ones are needlessly rigid. Only if this improves way too cumbersome i.e breaking assumption which are made should we introduce yet another mode. Because this other mode is really just special casing and we should avoid special cases as much as possible.
Ok, I'll modify the existing READA_FORWARD mechanism to make it better
and send a V2 patch.
Quoted 3 lines
<snip>