linux/fs/btrfs/delayed-ref.h

459 lines
13 KiB

Btrfs: do extent allocation and reference count updates in the background The extent allocation tree maintains a reference count and full back reference information for every extent allocated in the filesystem. For subvolume and snapshot trees, every time a block goes through COW, the new copy of the block adds a reference on every block it points to. If a btree node points to 150 leaves, then the COW code needs to go and add backrefs on 150 different extents, which might be spread all over the extent allocation tree. These updates currently happen during btrfs_cow_block, and most COWs happen during btrfs_search_slot. btrfs_search_slot has locks held on both the parent and the node we are COWing, and so we really want to avoid IO during the COW if we can. This commit adds an rbtree of pending reference count updates and extent allocations. The tree is ordered by byte number of the extent and byte number of the parent for the back reference. The tree allows us to: 1) Modify back references in something close to disk order, reducing seeks 2) Significantly reduce the number of modifications made as block pointers are balanced around 3) Do all of the extent insertion and back reference modifications outside of the performance critical btrfs_search_slot code. #3 has the added benefit of greatly reducing the btrfs stack footprint. The extent allocation tree modifications are done without the deep (and somewhat recursive) call chains used in the past. These delayed back reference updates must be done before the transaction commits, and so the rbtree is tied to the transaction. Throttling is implemented to help keep the queue of backrefs at a reasonable size. Since there was a similar mechanism in place for the extent tree extents, that is removed and replaced by the delayed reference tree. Yan Zheng <yan.zheng@oracle.com> helped review and fixup this code. Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-03-13 10:10:06 -04:00
Btrfs: do extent allocation and reference count updates in the background The extent allocation tree maintains a reference count and full back reference information for every extent allocated in the filesystem. For subvolume and snapshot trees, every time a block goes through COW, the new copy of the block adds a reference on every block it points to. If a btree node points to 150 leaves, then the COW code needs to go and add backrefs on 150 different extents, which might be spread all over the extent allocation tree. These updates currently happen during btrfs_cow_block, and most COWs happen during btrfs_search_slot. btrfs_search_slot has locks held on both the parent and the node we are COWing, and so we really want to avoid IO during the COW if we can. This commit adds an rbtree of pending reference count updates and extent allocations. The tree is ordered by byte number of the extent and byte number of the parent for the back reference. The tree allows us to: 1) Modify back references in something close to disk order, reducing seeks 2) Significantly reduce the number of modifications made as block pointers are balanced around 3) Do all of the extent insertion and back reference modifications outside of the performance critical btrfs_search_slot code. #3 has the added benefit of greatly reducing the btrfs stack footprint. The extent allocation tree modifications are done without the deep (and somewhat recursive) call chains used in the past. These delayed back reference updates must be done before the transaction commits, and so the rbtree is tied to the transaction. Throttling is implemented to help keep the queue of backrefs at a reasonable size. Since there was a similar mechanism in place for the extent tree extents, that is removed and replaced by the delayed reference tree. Yan Zheng <yan.zheng@oracle.com> helped review and fixup this code. Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-03-13 10:10:06 -04:00
Btrfs: do extent allocation and reference count updates in the background The extent allocation tree maintains a reference count and full back reference information for every extent allocated in the filesystem. For subvolume and snapshot trees, every time a block goes through COW, the new copy of the block adds a reference on every block it points to. If a btree node points to 150 leaves, then the COW code needs to go and add backrefs on 150 different extents, which might be spread all over the extent allocation tree. These updates currently happen during btrfs_cow_block, and most COWs happen during btrfs_search_slot. btrfs_search_slot has locks held on both the parent and the node we are COWing, and so we really want to avoid IO during the COW if we can. This commit adds an rbtree of pending reference count updates and extent allocations. The tree is ordered by byte number of the extent and byte number of the parent for the back reference. The tree allows us to: 1) Modify back references in something close to disk order, reducing seeks 2) Significantly reduce the number of modifications made as block pointers are balanced around 3) Do all of the extent insertion and back reference modifications outside of the performance critical btrfs_search_slot code. #3 has the added benefit of greatly reducing the btrfs stack footprint. The extent allocation tree modifications are done without the deep (and somewhat recursive) call chains used in the past. These delayed back reference updates must be done before the transaction commits, and so the rbtree is tied to the transaction. Throttling is implemented to help keep the queue of backrefs at a reasonable size. Since there was a similar mechanism in place for the extent tree extents, that is removed and replaced by the delayed reference tree. Yan Zheng <yan.zheng@oracle.com> helped review and fixup this code. Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-03-13 10:10:06 -04:00
Btrfs: do extent allocation and reference count updates in the background The extent allocation tree maintains a reference count and full back reference information for every extent allocated in the filesystem. For subvolume and snapshot trees, every time a block goes through COW, the new copy of the block adds a reference on every block it points to. If a btree node points to 150 leaves, then the COW code needs to go and add backrefs on 150 different extents, which might be spread all over the extent allocation tree. These updates currently happen during btrfs_cow_block, and most COWs happen during btrfs_search_slot. btrfs_search_slot has locks held on both the parent and the node we are COWing, and so we really want to avoid IO during the COW if we can. This commit adds an rbtree of pending reference count updates and extent allocations. The tree is ordered by byte number of the extent and byte number of the parent for the back reference. The tree allows us to: 1) Modify back references in something close to disk order, reducing seeks 2) Significantly reduce the number of modifications made as block pointers are balanced around 3) Do all of the extent insertion and back reference modifications outside of the performance critical btrfs_search_slot code. #3 has the added benefit of greatly reducing the btrfs stack footprint. The extent allocation tree modifications are done without the deep (and somewhat recursive) call chains used in the past. These delayed back reference updates must be done before the transaction commits, and so the rbtree is tied to the transaction. Throttling is implemented to help keep the queue of backrefs at a reasonable size. Since there was a similar mechanism in place for the extent tree extents, that is removed and replaced by the delayed reference tree. Yan Zheng <yan.zheng@oracle.com> helped review and fixup this code. Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-03-13 10:10:06 -04:00
btrfs: improve delayed refs iterations This issue was found when I tried to delete a heavily reflinked file, when deleting such files, other transaction operation will not have a chance to make progress, for example, start_transaction() will blocked in wait_current_trans(root) for long time, sometimes it even triggers soft lockups, and the time taken to delete such heavily reflinked file is also very large, often hundreds of seconds. Using perf top, it reports that: PerfTop: 7416 irqs/sec kernel:99.8% exact: 0.0% [4000Hz cpu-clock], (all, 4 CPUs) --------------------------------------------------------------------------------------- 84.37% [btrfs] [k] __btrfs_run_delayed_refs.constprop.80 11.02% [kernel] [k] delay_tsc 0.79% [kernel] [k] _raw_spin_unlock_irq 0.78% [kernel] [k] _raw_spin_unlock_irqrestore 0.45% [kernel] [k] do_raw_spin_lock 0.18% [kernel] [k] __slab_alloc It seems __btrfs_run_delayed_refs() took most cpu time, after some debug work, I found it's select_delayed_ref() causing this issue, for a delayed head, in our case, it'll be full of BTRFS_DROP_DELAYED_REF nodes, but select_delayed_ref() will firstly try to iterate node list to find BTRFS_ADD_DELAYED_REF nodes, obviously it's a disaster in this case, and waste much time. To fix this issue, we introduce a new ref_add_list in struct btrfs_delayed_ref_head, then in select_delayed_ref(), if this list is not empty, we can directly use nodes in this list. With this patch, it just took about 10~15 seconds to delte the same file. Now using perf top, it reports that: PerfTop: 2734 irqs/sec kernel:99.5% exact: 0.0% [4000Hz cpu-clock], (all, 4 CPUs) ---------------------------------------------------------------------------------------- 20.74% [kernel] [k] _raw_spin_unlock_irqrestore 16.33% [kernel] [k] __slab_alloc 5.41% [kernel] [k] lock_acquired 4.42% [kernel] [k] lock_acquire 4.05% [kernel] [k] lock_release 3.37% [kernel] [k] _raw_spin_unlock_irq For normal files, this patch also gives help, at least we do not need to iterate whole list to found BTRFS_ADD_DELAYED_REF nodes. Signed-off-by: Wang Xiaoguang <wangxg.fnst@cn.fujitsu.com> Reviewed-by: Liu Bo <bo.li.liu@oracle.com> Tested-by: Holger Hoffstätte <holger@applied-asynchrony.com> Signed-off-by: David Sterba <dsterba@suse.com>
2016-10-26 18:07:33 +08:00
btrfs: improve delayed refs iterations This issue was found when I tried to delete a heavily reflinked file, when deleting such files, other transaction operation will not have a chance to make progress, for example, start_transaction() will blocked in wait_current_trans(root) for long time, sometimes it even triggers soft lockups, and the time taken to delete such heavily reflinked file is also very large, often hundreds of seconds. Using perf top, it reports that: PerfTop: 7416 irqs/sec kernel:99.8% exact: 0.0% [4000Hz cpu-clock], (all, 4 CPUs) --------------------------------------------------------------------------------------- 84.37% [btrfs] [k] __btrfs_run_delayed_refs.constprop.80 11.02% [kernel] [k] delay_tsc 0.79% [kernel] [k] _raw_spin_unlock_irq 0.78% [kernel] [k] _raw_spin_unlock_irqrestore 0.45% [kernel] [k] do_raw_spin_lock 0.18% [kernel] [k] __slab_alloc It seems __btrfs_run_delayed_refs() took most cpu time, after some debug work, I found it's select_delayed_ref() causing this issue, for a delayed head, in our case, it'll be full of BTRFS_DROP_DELAYED_REF nodes, but select_delayed_ref() will firstly try to iterate node list to find BTRFS_ADD_DELAYED_REF nodes, obviously it's a disaster in this case, and waste much time. To fix this issue, we introduce a new ref_add_list in struct btrfs_delayed_ref_head, then in select_delayed_ref(), if this list is not empty, we can directly use nodes in this list. With this patch, it just took about 10~15 seconds to delte the same file. Now using perf top, it reports that: PerfTop: 2734 irqs/sec kernel:99.5% exact: 0.0% [4000Hz cpu-clock], (all, 4 CPUs) ---------------------------------------------------------------------------------------- 20.74% [kernel] [k] _raw_spin_unlock_irqrestore 16.33% [kernel] [k] __slab_alloc 5.41% [kernel] [k] lock_acquired 4.42% [kernel] [k] lock_acquire 4.05% [kernel] [k] lock_release 3.37% [kernel] [k] _raw_spin_unlock_irq For normal files, this patch also gives help, at least we do not need to iterate whole list to found BTRFS_ADD_DELAYED_REF nodes. Signed-off-by: Wang Xiaoguang <wangxg.fnst@cn.fujitsu.com> Reviewed-by: Liu Bo <bo.li.liu@oracle.com> Tested-by: Holger Hoffstätte <holger@applied-asynchrony.com> Signed-off-by: David Sterba <dsterba@suse.com>
2016-10-26 18:07:33 +08:00
Btrfs: do extent allocation and reference count updates in the background The extent allocation tree maintains a reference count and full back reference information for every extent allocated in the filesystem. For subvolume and snapshot trees, every time a block goes through COW, the new copy of the block adds a reference on every block it points to. If a btree node points to 150 leaves, then the COW code needs to go and add backrefs on 150 different extents, which might be spread all over the extent allocation tree. These updates currently happen during btrfs_cow_block, and most COWs happen during btrfs_search_slot. btrfs_search_slot has locks held on both the parent and the node we are COWing, and so we really want to avoid IO during the COW if we can. This commit adds an rbtree of pending reference count updates and extent allocations. The tree is ordered by byte number of the extent and byte number of the parent for the back reference. The tree allows us to: 1) Modify back references in something close to disk order, reducing seeks 2) Significantly reduce the number of modifications made as block pointers are balanced around 3) Do all of the extent insertion and back reference modifications outside of the performance critical btrfs_search_slot code. #3 has the added benefit of greatly reducing the btrfs stack footprint. The extent allocation tree modifications are done without the deep (and somewhat recursive) call chains used in the past. These delayed back reference updates must be done before the transaction commits, and so the rbtree is tied to the transaction. Throttling is implemented to help keep the queue of backrefs at a reasonable size. Since there was a similar mechanism in place for the extent tree extents, that is removed and replaced by the delayed reference tree. Yan Zheng <yan.zheng@oracle.com> helped review and fixup this code. Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-03-13 10:10:06 -04:00
Btrfs: do extent allocation and reference count updates in the background The extent allocation tree maintains a reference count and full back reference information for every extent allocated in the filesystem. For subvolume and snapshot trees, every time a block goes through COW, the new copy of the block adds a reference on every block it points to. If a btree node points to 150 leaves, then the COW code needs to go and add backrefs on 150 different extents, which might be spread all over the extent allocation tree. These updates currently happen during btrfs_cow_block, and most COWs happen during btrfs_search_slot. btrfs_search_slot has locks held on both the parent and the node we are COWing, and so we really want to avoid IO during the COW if we can. This commit adds an rbtree of pending reference count updates and extent allocations. The tree is ordered by byte number of the extent and byte number of the parent for the back reference. The tree allows us to: 1) Modify back references in something close to disk order, reducing seeks 2) Significantly reduce the number of modifications made as block pointers are balanced around 3) Do all of the extent insertion and back reference modifications outside of the performance critical btrfs_search_slot code. #3 has the added benefit of greatly reducing the btrfs stack footprint. The extent allocation tree modifications are done without the deep (and somewhat recursive) call chains used in the past. These delayed back reference updates must be done before the transaction commits, and so the rbtree is tied to the transaction. Throttling is implemented to help keep the queue of backrefs at a reasonable size. Since there was a similar mechanism in place for the extent tree extents, that is removed and replaced by the delayed reference tree. Yan Zheng <yan.zheng@oracle.com> helped review and fixup this code. Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-03-13 10:10:06 -04:00
Btrfs: do extent allocation and reference count updates in the background The extent allocation tree maintains a reference count and full back reference information for every extent allocated in the filesystem. For subvolume and snapshot trees, every time a block goes through COW, the new copy of the block adds a reference on every block it points to. If a btree node points to 150 leaves, then the COW code needs to go and add backrefs on 150 different extents, which might be spread all over the extent allocation tree. These updates currently happen during btrfs_cow_block, and most COWs happen during btrfs_search_slot. btrfs_search_slot has locks held on both the parent and the node we are COWing, and so we really want to avoid IO during the COW if we can. This commit adds an rbtree of pending reference count updates and extent allocations. The tree is ordered by byte number of the extent and byte number of the parent for the back reference. The tree allows us to: 1) Modify back references in something close to disk order, reducing seeks 2) Significantly reduce the number of modifications made as block pointers are balanced around 3) Do all of the extent insertion and back reference modifications outside of the performance critical btrfs_search_slot code. #3 has the added benefit of greatly reducing the btrfs stack footprint. The extent allocation tree modifications are done without the deep (and somewhat recursive) call chains used in the past. These delayed back reference updates must be done before the transaction commits, and so the rbtree is tied to the transaction. Throttling is implemented to help keep the queue of backrefs at a reasonable size. Since there was a similar mechanism in place for the extent tree extents, that is removed and replaced by the delayed reference tree. Yan Zheng <yan.zheng@oracle.com> helped review and fixup this code. Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-03-13 10:10:06 -04:00
Btrfs: Mixed back reference (FORWARD ROLLING FORMAT CHANGE) This commit introduces a new kind of back reference for btrfs metadata. Once a filesystem has been mounted with this commit, IT WILL NO LONGER BE MOUNTABLE BY OLDER KERNELS. When a tree block in subvolume tree is cow'd, the reference counts of all extents it points to are increased by one. At transaction commit time, the old root of the subvolume is recorded in a "dead root" data structure, and the btree it points to is later walked, dropping reference counts and freeing any blocks where the reference count goes to 0. The increments done during cow and decrements done after commit cancel out, and the walk is a very expensive way to go about freeing the blocks that are no longer referenced by the new btree root. This commit reduces the transaction overhead by avoiding the need for dead root records. When a non-shared tree block is cow'd, we free the old block at once, and the new block inherits old block's references. When a tree block with reference count > 1 is cow'd, we increase the reference counts of all extents the new block points to by one, and decrease the old block's reference count by one. This dead tree avoidance code removes the need to modify the reference counts of lower level extents when a non-shared tree block is cow'd. But we still need to update back ref for all pointers in the block. This is because the location of the block is recorded in the back ref item. We can solve this by introducing a new type of back ref. The new back ref provides information about pointer's key, level and in which tree the pointer lives. This information allow us to find the pointer by searching the tree. The shortcoming of the new back ref is that it only works for pointers in tree blocks referenced by their owner trees. This is mostly a problem for snapshots, where resolving one of these fuzzy back references would be O(number_of_snapshots) and quite slow. The solution used here is to use the fuzzy back references in the common case where a given tree block is only referenced by one root, and use the full back references when multiple roots have a reference on a given block. This commit adds per subvolume red-black tree to keep trace of cached inodes. The red-black tree helps the balancing code to find cached inodes whose inode numbers within a given range. This commit improves the balancing code by introducing several data structures to keep the state of balancing. The most important one is the back ref cache. It caches how the upper level tree blocks are referenced. This greatly reduce the overhead of checking back ref. The improved balancing code scales significantly better with a large number of snapshots. This is a very large commit and was written in a number of pieces. But, they depend heavily on the disk format change and were squashed together to make sure git bisect didn't end up in a bad state wrt space balancing or the format change. Signed-off-by: Yan Zheng <zheng.yan@oracle.com> Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-06-10 10:45:14 -04:00
Btrfs: do extent allocation and reference count updates in the background The extent allocation tree maintains a reference count and full back reference information for every extent allocated in the filesystem. For subvolume and snapshot trees, every time a block goes through COW, the new copy of the block adds a reference on every block it points to. If a btree node points to 150 leaves, then the COW code needs to go and add backrefs on 150 different extents, which might be spread all over the extent allocation tree. These updates currently happen during btrfs_cow_block, and most COWs happen during btrfs_search_slot. btrfs_search_slot has locks held on both the parent and the node we are COWing, and so we really want to avoid IO during the COW if we can. This commit adds an rbtree of pending reference count updates and extent allocations. The tree is ordered by byte number of the extent and byte number of the parent for the back reference. The tree allows us to: 1) Modify back references in something close to disk order, reducing seeks 2) Significantly reduce the number of modifications made as block pointers are balanced around 3) Do all of the extent insertion and back reference modifications outside of the performance critical btrfs_search_slot code. #3 has the added benefit of greatly reducing the btrfs stack footprint. The extent allocation tree modifications are done without the deep (and somewhat recursive) call chains used in the past. These delayed back reference updates must be done before the transaction commits, and so the rbtree is tied to the transaction. Throttling is implemented to help keep the queue of backrefs at a reasonable size. Since there was a similar mechanism in place for the extent tree extents, that is removed and replaced by the delayed reference tree. Yan Zheng <yan.zheng@oracle.com> helped review and fixup this code. Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-03-13 10:10:06 -04:00
Btrfs: Mixed back reference (FORWARD ROLLING FORMAT CHANGE) This commit introduces a new kind of back reference for btrfs metadata. Once a filesystem has been mounted with this commit, IT WILL NO LONGER BE MOUNTABLE BY OLDER KERNELS. When a tree block in subvolume tree is cow'd, the reference counts of all extents it points to are increased by one. At transaction commit time, the old root of the subvolume is recorded in a "dead root" data structure, and the btree it points to is later walked, dropping reference counts and freeing any blocks where the reference count goes to 0. The increments done during cow and decrements done after commit cancel out, and the walk is a very expensive way to go about freeing the blocks that are no longer referenced by the new btree root. This commit reduces the transaction overhead by avoiding the need for dead root records. When a non-shared tree block is cow'd, we free the old block at once, and the new block inherits old block's references. When a tree block with reference count > 1 is cow'd, we increase the reference counts of all extents the new block points to by one, and decrease the old block's reference count by one. This dead tree avoidance code removes the need to modify the reference counts of lower level extents when a non-shared tree block is cow'd. But we still need to update back ref for all pointers in the block. This is because the location of the block is recorded in the back ref item. We can solve this by introducing a new type of back ref. The new back ref provides information about pointer's key, level and in which tree the pointer lives. This information allow us to find the pointer by searching the tree. The shortcoming of the new back ref is that it only works for pointers in tree blocks referenced by their owner trees. This is mostly a problem for snapshots, where resolving one of these fuzzy back references would be O(number_of_snapshots) and quite slow. The solution used here is to use the fuzzy back references in the common case where a given tree block is only referenced by one root, and use the full back references when multiple roots have a reference on a given block. This commit adds per subvolume red-black tree to keep trace of cached inodes. The red-black tree helps the balancing code to find cached inodes whose inode numbers within a given range. This commit improves the balancing code by introducing several data structures to keep the state of balancing. The most important one is the back ref cache. It caches how the upper level tree blocks are referenced. This greatly reduce the overhead of checking back ref. The improved balancing code scales significantly better with a large number of snapshots. This is a very large commit and was written in a number of pieces. But, they depend heavily on the disk format change and were squashed together to make sure git bisect didn't end up in a bad state wrt space balancing or the format change. Signed-off-by: Yan Zheng <zheng.yan@oracle.com> Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-06-10 10:45:14 -04:00
Btrfs: Mixed back reference (FORWARD ROLLING FORMAT CHANGE) This commit introduces a new kind of back reference for btrfs metadata. Once a filesystem has been mounted with this commit, IT WILL NO LONGER BE MOUNTABLE BY OLDER KERNELS. When a tree block in subvolume tree is cow'd, the reference counts of all extents it points to are increased by one. At transaction commit time, the old root of the subvolume is recorded in a "dead root" data structure, and the btree it points to is later walked, dropping reference counts and freeing any blocks where the reference count goes to 0. The increments done during cow and decrements done after commit cancel out, and the walk is a very expensive way to go about freeing the blocks that are no longer referenced by the new btree root. This commit reduces the transaction overhead by avoiding the need for dead root records. When a non-shared tree block is cow'd, we free the old block at once, and the new block inherits old block's references. When a tree block with reference count > 1 is cow'd, we increase the reference counts of all extents the new block points to by one, and decrease the old block's reference count by one. This dead tree avoidance code removes the need to modify the reference counts of lower level extents when a non-shared tree block is cow'd. But we still need to update back ref for all pointers in the block. This is because the location of the block is recorded in the back ref item. We can solve this by introducing a new type of back ref. The new back ref provides information about pointer's key, level and in which tree the pointer lives. This information allow us to find the pointer by searching the tree. The shortcoming of the new back ref is that it only works for pointers in tree blocks referenced by their owner trees. This is mostly a problem for snapshots, where resolving one of these fuzzy back references would be O(number_of_snapshots) and quite slow. The solution used here is to use the fuzzy back references in the common case where a given tree block is only referenced by one root, and use the full back references when multiple roots have a reference on a given block. This commit adds per subvolume red-black tree to keep trace of cached inodes. The red-black tree helps the balancing code to find cached inodes whose inode numbers within a given range. This commit improves the balancing code by introducing several data structures to keep the state of balancing. The most important one is the back ref cache. It caches how the upper level tree blocks are referenced. This greatly reduce the overhead of checking back ref. The improved balancing code scales significantly better with a large number of snapshots. This is a very large commit and was written in a number of pieces. But, they depend heavily on the disk format change and were squashed together to make sure git bisect didn't end up in a bad state wrt space balancing or the format change. Signed-off-by: Yan Zheng <zheng.yan@oracle.com> Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-06-10 10:45:14 -04:00
Btrfs: do extent allocation and reference count updates in the background The extent allocation tree maintains a reference count and full back reference information for every extent allocated in the filesystem. For subvolume and snapshot trees, every time a block goes through COW, the new copy of the block adds a reference on every block it points to. If a btree node points to 150 leaves, then the COW code needs to go and add backrefs on 150 different extents, which might be spread all over the extent allocation tree. These updates currently happen during btrfs_cow_block, and most COWs happen during btrfs_search_slot. btrfs_search_slot has locks held on both the parent and the node we are COWing, and so we really want to avoid IO during the COW if we can. This commit adds an rbtree of pending reference count updates and extent allocations. The tree is ordered by byte number of the extent and byte number of the parent for the back reference. The tree allows us to: 1) Modify back references in something close to disk order, reducing seeks 2) Significantly reduce the number of modifications made as block pointers are balanced around 3) Do all of the extent insertion and back reference modifications outside of the performance critical btrfs_search_slot code. #3 has the added benefit of greatly reducing the btrfs stack footprint. The extent allocation tree modifications are done without the deep (and somewhat recursive) call chains used in the past. These delayed back reference updates must be done before the transaction commits, and so the rbtree is tied to the transaction. Throttling is implemented to help keep the queue of backrefs at a reasonable size. Since there was a similar mechanism in place for the extent tree extents, that is removed and replaced by the delayed reference tree. Yan Zheng <yan.zheng@oracle.com> helped review and fixup this code. Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-03-13 10:10:06 -04:00
Btrfs: do extent allocation and reference count updates in the background The extent allocation tree maintains a reference count and full back reference information for every extent allocated in the filesystem. For subvolume and snapshot trees, every time a block goes through COW, the new copy of the block adds a reference on every block it points to. If a btree node points to 150 leaves, then the COW code needs to go and add backrefs on 150 different extents, which might be spread all over the extent allocation tree. These updates currently happen during btrfs_cow_block, and most COWs happen during btrfs_search_slot. btrfs_search_slot has locks held on both the parent and the node we are COWing, and so we really want to avoid IO during the COW if we can. This commit adds an rbtree of pending reference count updates and extent allocations. The tree is ordered by byte number of the extent and byte number of the parent for the back reference. The tree allows us to: 1) Modify back references in something close to disk order, reducing seeks 2) Significantly reduce the number of modifications made as block pointers are balanced around 3) Do all of the extent insertion and back reference modifications outside of the performance critical btrfs_search_slot code. #3 has the added benefit of greatly reducing the btrfs stack footprint. The extent allocation tree modifications are done without the deep (and somewhat recursive) call chains used in the past. These delayed back reference updates must be done before the transaction commits, and so the rbtree is tied to the transaction. Throttling is implemented to help keep the queue of backrefs at a reasonable size. Since there was a similar mechanism in place for the extent tree extents, that is removed and replaced by the delayed reference tree. Yan Zheng <yan.zheng@oracle.com> helped review and fixup this code. Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-03-13 10:10:06 -04:00
btrfs: reorder some members of struct btrfs_delayed_ref_head Currently struct delayed_ref_head has its 'bytenr' and 'href_node' members in different cache lines (even on a release, non-debug, kernel). This is not optimal because when iterating the red black tree of delayed ref heads for inserting a new delayed ref head (htree_insert()) we have to pull in 2 cache lines of delayed ref heads we find in a patch, one for the tree node (struct rb_node) and another one for the 'bytenr' field. The same applies when searching for an existing delayed ref head (find_ref_head()). On a release (non-debug) kernel, the structure also has two 4 bytes holes, which makes it 8 bytes longer than necessary. Its current layout is the following: struct btrfs_delayed_ref_head { u64 bytenr; /* 0 8 */ u64 num_bytes; /* 8 8 */ refcount_t refs; /* 16 4 */ /* XXX 4 bytes hole, try to pack */ struct mutex mutex; /* 24 32 */ spinlock_t lock; /* 56 4 */ /* XXX 4 bytes hole, try to pack */ /* --- cacheline 1 boundary (64 bytes) --- */ struct rb_root_cached ref_tree; /* 64 16 */ struct list_head ref_add_list; /* 80 16 */ struct rb_node href_node __attribute__((__aligned__(8))); /* 96 24 */ struct btrfs_delayed_extent_op * extent_op; /* 120 8 */ /* --- cacheline 2 boundary (128 bytes) --- */ int total_ref_mod; /* 128 4 */ int ref_mod; /* 132 4 */ unsigned int must_insert_reserved:1; /* 136: 0 4 */ unsigned int is_data:1; /* 136: 1 4 */ unsigned int is_system:1; /* 136: 2 4 */ unsigned int processing:1; /* 136: 3 4 */ /* size: 144, cachelines: 3, members: 15 */ /* sum members: 128, holes: 2, sum holes: 8 */ /* sum bitfield members: 4 bits (0 bytes) */ /* padding: 4 */ /* bit_padding: 28 bits */ /* forced alignments: 1 */ /* last cacheline: 16 bytes */ } __attribute__((__aligned__(8))); This change reorders the 'href_node' and 'refs' members so that we have the 'href_node' in the same cache line as the 'bytenr' field, while also eliminating the two holes and reducing the structure size from 144 bytes down to 136 bytes, so we can now have 30 ref heads per 4K page (on x86_64) instead of 28. The new structure layout after this change is now: struct btrfs_delayed_ref_head { u64 bytenr; /* 0 8 */ u64 num_bytes; /* 8 8 */ struct rb_node href_node __attribute__((__aligned__(8))); /* 16 24 */ struct mutex mutex; /* 40 32 */ /* --- cacheline 1 boundary (64 bytes) was 8 bytes ago --- */ refcount_t refs; /* 72 4 */ spinlock_t lock; /* 76 4 */ struct rb_root_cached ref_tree; /* 80 16 */ struct list_head ref_add_list; /* 96 16 */ struct btrfs_delayed_extent_op * extent_op; /* 112 8 */ int total_ref_mod; /* 120 4 */ int ref_mod; /* 124 4 */ /* --- cacheline 2 boundary (128 bytes) --- */ unsigned int must_insert_reserved:1; /* 128: 0 4 */ unsigned int is_data:1; /* 128: 1 4 */ unsigned int is_system:1; /* 128: 2 4 */ unsigned int processing:1; /* 128: 3 4 */ /* size: 136, cachelines: 3, members: 15 */ /* padding: 4 */ /* bit_padding: 28 bits */ /* forced alignments: 1 */ /* last cacheline: 8 bytes */ } __attribute__((__aligned__(8))); Running the following fs_mark test shows some significant improvement. $ cat test.sh #!/bin/bash # 15G null block device DEV=/dev/nullb0 MNT=/mnt/nullb0 FILES=100000 THREADS=$(nproc --all) FILE_SIZE=0 echo "performance" | \ tee /sys/devices/system/cpu/cpu*/cpufreq/scaling_governor mkfs.btrfs -f $DEV mount -o ssd $DEV $MNT OPTS="-S 0 -L 5 -n $FILES -s $FILE_SIZE -t $THREADS -k" for ((i = 1; i <= $THREADS; i++)); do OPTS="$OPTS -d $MNT/d$i" done fs_mark $OPTS umount $MNT Before this change: FSUse% Count Size Files/sec App Overhead 10 1200000 0 112631.3 11928055 16 2400000 0 189943.8 12140777 23 3600000 0 150719.2 13178480 50 4800000 0 99137.3 12504293 53 6000000 0 111733.9 12670836 Total files/sec: 664165.5 After this change: FSUse% Count Size Files/sec App Overhead 10 1200000 0 148589.5 11565889 16 2400000 0 227743.8 11561596 23 3600000 0 191590.5 12550755 30 4800000 0 179812.3 12629610 53 6000000 0 92471.4 12352383 Total files/sec: 840207.5 Measuring the execution times of htree_insert(), in nanoseconds, during those fs_mark runs: Before this change: Range: 0.000 - 940647.000; Mean: 619.733; Median: 548.000; Stddev: 1834.231 Percentiles: 90th: 980.000; 95th: 1208.000; 99th: 2090.000 0.000 - 6.384: 257 | 6.384 - 26.259: 977 | 26.259 - 99.635: 4963 | 99.635 - 370.526: 136800 ############# 370.526 - 1370.603: 566110 ##################################################### 1370.603 - 5062.704: 24945 ## 5062.704 - 18693.248: 944 | 18693.248 - 69014.670: 211 | 69014.670 - 254791.959: 30 | 254791.959 - 940647.000: 4 | After this change: Range: 0.000 - 299200.000; Mean: 587.754; Median: 542.000; Stddev: 1030.422 Percentiles: 90th: 918.000; 95th: 1113.000; 99th: 1987.000 0.000 - 5.585: 163 | 5.585 - 20.678: 452 | 20.678 - 70.369: 1806 | 70.369 - 233.965: 26268 #### 233.965 - 772.564: 333519 ##################################################### 772.564 - 2545.771: 91820 ############### 2545.771 - 8383.615: 2238 | 8383.615 - 27603.280: 170 | 27603.280 - 90879.297: 68 | 90879.297 - 299200.000: 12 | Mean, percentiles, maximum times are all better, as well as a lower standard deviation. Signed-off-by: Filipe Manana <fdmanana@suse.com> Reviewed-by: David Sterba <dsterba@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2023-05-29 16:16:57 +01:00
btrfs: improve delayed refs iterations This issue was found when I tried to delete a heavily reflinked file, when deleting such files, other transaction operation will not have a chance to make progress, for example, start_transaction() will blocked in wait_current_trans(root) for long time, sometimes it even triggers soft lockups, and the time taken to delete such heavily reflinked file is also very large, often hundreds of seconds. Using perf top, it reports that: PerfTop: 7416 irqs/sec kernel:99.8% exact: 0.0% [4000Hz cpu-clock], (all, 4 CPUs) --------------------------------------------------------------------------------------- 84.37% [btrfs] [k] __btrfs_run_delayed_refs.constprop.80 11.02% [kernel] [k] delay_tsc 0.79% [kernel] [k] _raw_spin_unlock_irq 0.78% [kernel] [k] _raw_spin_unlock_irqrestore 0.45% [kernel] [k] do_raw_spin_lock 0.18% [kernel] [k] __slab_alloc It seems __btrfs_run_delayed_refs() took most cpu time, after some debug work, I found it's select_delayed_ref() causing this issue, for a delayed head, in our case, it'll be full of BTRFS_DROP_DELAYED_REF nodes, but select_delayed_ref() will firstly try to iterate node list to find BTRFS_ADD_DELAYED_REF nodes, obviously it's a disaster in this case, and waste much time. To fix this issue, we introduce a new ref_add_list in struct btrfs_delayed_ref_head, then in select_delayed_ref(), if this list is not empty, we can directly use nodes in this list. With this patch, it just took about 10~15 seconds to delte the same file. Now using perf top, it reports that: PerfTop: 2734 irqs/sec kernel:99.5% exact: 0.0% [4000Hz cpu-clock], (all, 4 CPUs) ---------------------------------------------------------------------------------------- 20.74% [kernel] [k] _raw_spin_unlock_irqrestore 16.33% [kernel] [k] __slab_alloc 5.41% [kernel] [k] lock_acquired 4.42% [kernel] [k] lock_acquire 4.05% [kernel] [k] lock_release 3.37% [kernel] [k] _raw_spin_unlock_irq For normal files, this patch also gives help, at least we do not need to iterate whole list to found BTRFS_ADD_DELAYED_REF nodes. Signed-off-by: Wang Xiaoguang <wangxg.fnst@cn.fujitsu.com> Reviewed-by: Liu Bo <bo.li.liu@oracle.com> Tested-by: Holger Hoffstätte <holger@applied-asynchrony.com> Signed-off-by: David Sterba <dsterba@suse.com>
2016-10-26 18:07:33 +08:00
Btrfs: Mixed back reference (FORWARD ROLLING FORMAT CHANGE) This commit introduces a new kind of back reference for btrfs metadata. Once a filesystem has been mounted with this commit, IT WILL NO LONGER BE MOUNTABLE BY OLDER KERNELS. When a tree block in subvolume tree is cow'd, the reference counts of all extents it points to are increased by one. At transaction commit time, the old root of the subvolume is recorded in a "dead root" data structure, and the btree it points to is later walked, dropping reference counts and freeing any blocks where the reference count goes to 0. The increments done during cow and decrements done after commit cancel out, and the walk is a very expensive way to go about freeing the blocks that are no longer referenced by the new btree root. This commit reduces the transaction overhead by avoiding the need for dead root records. When a non-shared tree block is cow'd, we free the old block at once, and the new block inherits old block's references. When a tree block with reference count > 1 is cow'd, we increase the reference counts of all extents the new block points to by one, and decrease the old block's reference count by one. This dead tree avoidance code removes the need to modify the reference counts of lower level extents when a non-shared tree block is cow'd. But we still need to update back ref for all pointers in the block. This is because the location of the block is recorded in the back ref item. We can solve this by introducing a new type of back ref. The new back ref provides information about pointer's key, level and in which tree the pointer lives. This information allow us to find the pointer by searching the tree. The shortcoming of the new back ref is that it only works for pointers in tree blocks referenced by their owner trees. This is mostly a problem for snapshots, where resolving one of these fuzzy back references would be O(number_of_snapshots) and quite slow. The solution used here is to use the fuzzy back references in the common case where a given tree block is only referenced by one root, and use the full back references when multiple roots have a reference on a given block. This commit adds per subvolume red-black tree to keep trace of cached inodes. The red-black tree helps the balancing code to find cached inodes whose inode numbers within a given range. This commit improves the balancing code by introducing several data structures to keep the state of balancing. The most important one is the back ref cache. It caches how the upper level tree blocks are referenced. This greatly reduce the overhead of checking back ref. The improved balancing code scales significantly better with a large number of snapshots. This is a very large commit and was written in a number of pieces. But, they depend heavily on the disk format change and were squashed together to make sure git bisect didn't end up in a bad state wrt space balancing or the format change. Signed-off-by: Yan Zheng <zheng.yan@oracle.com> Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-06-10 10:45:14 -04:00
Btrfs: do extent allocation and reference count updates in the background The extent allocation tree maintains a reference count and full back reference information for every extent allocated in the filesystem. For subvolume and snapshot trees, every time a block goes through COW, the new copy of the block adds a reference on every block it points to. If a btree node points to 150 leaves, then the COW code needs to go and add backrefs on 150 different extents, which might be spread all over the extent allocation tree. These updates currently happen during btrfs_cow_block, and most COWs happen during btrfs_search_slot. btrfs_search_slot has locks held on both the parent and the node we are COWing, and so we really want to avoid IO during the COW if we can. This commit adds an rbtree of pending reference count updates and extent allocations. The tree is ordered by byte number of the extent and byte number of the parent for the back reference. The tree allows us to: 1) Modify back references in something close to disk order, reducing seeks 2) Significantly reduce the number of modifications made as block pointers are balanced around 3) Do all of the extent insertion and back reference modifications outside of the performance critical btrfs_search_slot code. #3 has the added benefit of greatly reducing the btrfs stack footprint. The extent allocation tree modifications are done without the deep (and somewhat recursive) call chains used in the past. These delayed back reference updates must be done before the transaction commits, and so the rbtree is tied to the transaction. Throttling is implemented to help keep the queue of backrefs at a reasonable size. Since there was a similar mechanism in place for the extent tree extents, that is removed and replaced by the delayed reference tree. Yan Zheng <yan.zheng@oracle.com> helped review and fixup this code. Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-03-13 10:10:06 -04:00
btrfs: track delayed ref heads in an xarray Currently we use a red black tree (rb-tree) to track the delayed ref heads (in struct btrfs_delayed_ref_root::href_root). This however is not very efficient when the number of delayed ref heads is large (and it's very common to be at least in the order of thousands) since rb-trees are binary trees. For example for 10K delayed ref heads, the tree has a depth of 13. Besides that, inserting into the tree requires navigating through it and pulling useless cache lines in the process since the red black tree nodes are embedded within the delayed ref head structure - on the other hand, by being embedded, it requires no extra memory allocations. We can improve this by using an xarray instead which has a much higher branching factor than a red black tree (binary balanced tree) and is more cache friendly and behaves like a resizable array, with a much better search and insertion complexity than a red black tree. This only has one small disadvantage which is that insertion will sometimes require allocating memory for the xarray - which may fail (not that often since it uses a kmem_cache) - but on the other hand we can reduce the delayed ref head structure size by 24 bytes (from 152 down to 128 bytes) after removing the embedded red black tree node, meaning than we can now fit 32 delayed ref heads per 4K page instead of 26, and that gain compensates for the occasional memory allocations needed for the xarray nodes. We also end up using only 2 cache lines instead of 3 per delayed ref head. Running the following fs_mark test showed some improvements: $ cat test.sh #!/bin/bash DEV=/dev/nullb0 MNT=/mnt/nullb0 MOUNT_OPTIONS="-o ssd" FILES=100000 THREADS=$(nproc --all) echo "performance" | \ tee /sys/devices/system/cpu/cpu*/cpufreq/scaling_governor mkfs.btrfs -f $DEV mount $MOUNT_OPTIONS $DEV $MNT OPTS="-S 0 -L 5 -n $FILES -s 0 -t $THREADS -k" for ((i = 1; i <= $THREADS; i++)); do OPTS="$OPTS -d $MNT/d$i" done fs_mark $OPTS umount $MNT Before this patch: FSUse% Count Size Files/sec App Overhead 10 1200000 0 171845.7 12253839 16 2400000 0 230898.7 12308254 23 3600000 0 212292.9 12467768 30 4800000 0 195737.8 12627554 46 6000000 0 171055.2 12783329 After this patch: FSUse% Count Size Files/sec App Overhead 10 1200000 0 173835.0 12246131 16 2400000 0 233537.8 12271746 23 3600000 0 220398.7 12307737 30 4800000 0 204483.6 12392318 40 6000000 0 182923.3 12771843 Reviewed-by: Boris Burkov <boris@bur.io> Reviewed-by: Qu Wenruo <wqu@suse.com> Signed-off-by: Filipe Manana <fdmanana@suse.com> Reviewed-by: David Sterba <dsterba@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2024-10-23 16:27:14 +01:00
Btrfs: do extent allocation and reference count updates in the background The extent allocation tree maintains a reference count and full back reference information for every extent allocated in the filesystem. For subvolume and snapshot trees, every time a block goes through COW, the new copy of the block adds a reference on every block it points to. If a btree node points to 150 leaves, then the COW code needs to go and add backrefs on 150 different extents, which might be spread all over the extent allocation tree. These updates currently happen during btrfs_cow_block, and most COWs happen during btrfs_search_slot. btrfs_search_slot has locks held on both the parent and the node we are COWing, and so we really want to avoid IO during the COW if we can. This commit adds an rbtree of pending reference count updates and extent allocations. The tree is ordered by byte number of the extent and byte number of the parent for the back reference. The tree allows us to: 1) Modify back references in something close to disk order, reducing seeks 2) Significantly reduce the number of modifications made as block pointers are balanced around 3) Do all of the extent insertion and back reference modifications outside of the performance critical btrfs_search_slot code. #3 has the added benefit of greatly reducing the btrfs stack footprint. The extent allocation tree modifications are done without the deep (and somewhat recursive) call chains used in the past. These delayed back reference updates must be done before the transaction commits, and so the rbtree is tied to the transaction. Throttling is implemented to help keep the queue of backrefs at a reasonable size. Since there was a similar mechanism in place for the extent tree extents, that is removed and replaced by the delayed reference tree. Yan Zheng <yan.zheng@oracle.com> helped review and fixup this code. Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-03-13 10:10:06 -04:00
btrfs: only let one thread pre-flush delayed refs in commit I've been running a stress test that runs 20 workers in their own subvolume, which are running an fsstress instance with 4 threads per worker, which is 80 total fsstress threads. In addition to this I'm running balance in the background as well as creating and deleting snapshots. This test takes around 12 hours to run normally, going slower and slower as the test goes on. The reason for this is because fsstress is running fsync sometimes, and because we're messing with block groups we often fall through to btrfs_commit_transaction, so will often have 20-30 threads all calling btrfs_commit_transaction at the same time. These all get stuck contending on the extent tree while they try to run delayed refs during the initial part of the commit. This is suboptimal, really because the extent tree is a single point of failure we only want one thread acting on that tree at once to reduce lock contention. Fix this by making the flushing mechanism a bit operation, to make it easy to use test_and_set_bit() in order to make sure only one task does this initial flush. Once we're into the transaction commit we only have one thread doing delayed ref running, it's just this initial pre-flush that is problematic. With this patch my stress test takes around 90 minutes to run, instead of 12 hours. The memory barrier is not necessary for the flushing bit as it's ordered, unlike plain int. The transaction state accessed in btrfs_should_end_transaction could be affected by that too as it's not always used under transaction lock. Upon Nikolay's analysis in [1] it's not necessary: In should_end_transaction it's read without holding any locks. (U) It's modified in btrfs_cleanup_transaction without holding the fs_info->trans_lock (U), but the STATE_ERROR flag is going to be set. set in cleanup_transaction under fs_info->trans_lock (L) set in btrfs_commit_trans to COMMIT_START under fs_info->trans_lock.(L) set in btrfs_commit_trans to COMMIT_DOING under fs_info->trans_lock.(L) set in btrfs_commit_trans to COMMIT_UNBLOCK under fs_info->trans_lock.(L) set in btrfs_commit_trans to COMMIT_COMPLETED without locks but at this point the transaction is finished and fs_info->running_trans is NULL (U but irrelevant). So by the looks of it we can have a concurrent READ race with a WRITE, due to reads not taking a lock. In this case what we want to ensure is we either see new or old state. I consulted with Will Deacon and he said that in such a case we'd want to annotate the accesses to ->state with (READ|WRITE)_ONCE so as to avoid a theoretical tear, in this case I don't think this could happen but I imagine at some point KCSAN would flag such an access as racy (which it is). [1] https://lore.kernel.org/linux-btrfs/e1fd5cc1-0f28-f670-69f4-e9958b4964e6@suse.com Reviewed-by: Nikolay Borisov <nborisov@suse.com> Signed-off-by: Josef Bacik <josef@toxicpanda.com> [ add comments regarding memory barrier ] Signed-off-by: David Sterba <dsterba@suse.com>
2020-12-18 14:24:20 -05:00
Btrfs: do extent allocation and reference count updates in the background The extent allocation tree maintains a reference count and full back reference information for every extent allocated in the filesystem. For subvolume and snapshot trees, every time a block goes through COW, the new copy of the block adds a reference on every block it points to. If a btree node points to 150 leaves, then the COW code needs to go and add backrefs on 150 different extents, which might be spread all over the extent allocation tree. These updates currently happen during btrfs_cow_block, and most COWs happen during btrfs_search_slot. btrfs_search_slot has locks held on both the parent and the node we are COWing, and so we really want to avoid IO during the COW if we can. This commit adds an rbtree of pending reference count updates and extent allocations. The tree is ordered by byte number of the extent and byte number of the parent for the back reference. The tree allows us to: 1) Modify back references in something close to disk order, reducing seeks 2) Significantly reduce the number of modifications made as block pointers are balanced around 3) Do all of the extent insertion and back reference modifications outside of the performance critical btrfs_search_slot code. #3 has the added benefit of greatly reducing the btrfs stack footprint. The extent allocation tree modifications are done without the deep (and somewhat recursive) call chains used in the past. These delayed back reference updates must be done before the transaction commits, and so the rbtree is tied to the transaction. Throttling is implemented to help keep the queue of backrefs at a reasonable size. Since there was a similar mechanism in place for the extent tree extents, that is removed and replaced by the delayed reference tree. Yan Zheng <yan.zheng@oracle.com> helped review and fixup this code. Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-03-13 10:10:06 -04:00
btrfs: track delayed ref heads in an xarray Currently we use a red black tree (rb-tree) to track the delayed ref heads (in struct btrfs_delayed_ref_root::href_root). This however is not very efficient when the number of delayed ref heads is large (and it's very common to be at least in the order of thousands) since rb-trees are binary trees. For example for 10K delayed ref heads, the tree has a depth of 13. Besides that, inserting into the tree requires navigating through it and pulling useless cache lines in the process since the red black tree nodes are embedded within the delayed ref head structure - on the other hand, by being embedded, it requires no extra memory allocations. We can improve this by using an xarray instead which has a much higher branching factor than a red black tree (binary balanced tree) and is more cache friendly and behaves like a resizable array, with a much better search and insertion complexity than a red black tree. This only has one small disadvantage which is that insertion will sometimes require allocating memory for the xarray - which may fail (not that often since it uses a kmem_cache) - but on the other hand we can reduce the delayed ref head structure size by 24 bytes (from 152 down to 128 bytes) after removing the embedded red black tree node, meaning than we can now fit 32 delayed ref heads per 4K page instead of 26, and that gain compensates for the occasional memory allocations needed for the xarray nodes. We also end up using only 2 cache lines instead of 3 per delayed ref head. Running the following fs_mark test showed some improvements: $ cat test.sh #!/bin/bash DEV=/dev/nullb0 MNT=/mnt/nullb0 MOUNT_OPTIONS="-o ssd" FILES=100000 THREADS=$(nproc --all) echo "performance" | \ tee /sys/devices/system/cpu/cpu*/cpufreq/scaling_governor mkfs.btrfs -f $DEV mount $MOUNT_OPTIONS $DEV $MNT OPTS="-S 0 -L 5 -n $FILES -s 0 -t $THREADS -k" for ((i = 1; i <= $THREADS; i++)); do OPTS="$OPTS -d $MNT/d$i" done fs_mark $OPTS umount $MNT Before this patch: FSUse% Count Size Files/sec App Overhead 10 1200000 0 171845.7 12253839 16 2400000 0 230898.7 12308254 23 3600000 0 212292.9 12467768 30 4800000 0 195737.8 12627554 46 6000000 0 171055.2 12783329 After this patch: FSUse% Count Size Files/sec App Overhead 10 1200000 0 173835.0 12246131 16 2400000 0 233537.8 12271746 23 3600000 0 220398.7 12307737 30 4800000 0 204483.6 12392318 40 6000000 0 182923.3 12771843 Reviewed-by: Boris Burkov <boris@bur.io> Reviewed-by: Qu Wenruo <wqu@suse.com> Signed-off-by: Filipe Manana <fdmanana@suse.com> Reviewed-by: David Sterba <dsterba@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2024-10-23 16:27:14 +01:00
btrfs: track delayed ref heads in an xarray Currently we use a red black tree (rb-tree) to track the delayed ref heads (in struct btrfs_delayed_ref_root::href_root). This however is not very efficient when the number of delayed ref heads is large (and it's very common to be at least in the order of thousands) since rb-trees are binary trees. For example for 10K delayed ref heads, the tree has a depth of 13. Besides that, inserting into the tree requires navigating through it and pulling useless cache lines in the process since the red black tree nodes are embedded within the delayed ref head structure - on the other hand, by being embedded, it requires no extra memory allocations. We can improve this by using an xarray instead which has a much higher branching factor than a red black tree (binary balanced tree) and is more cache friendly and behaves like a resizable array, with a much better search and insertion complexity than a red black tree. This only has one small disadvantage which is that insertion will sometimes require allocating memory for the xarray - which may fail (not that often since it uses a kmem_cache) - but on the other hand we can reduce the delayed ref head structure size by 24 bytes (from 152 down to 128 bytes) after removing the embedded red black tree node, meaning than we can now fit 32 delayed ref heads per 4K page instead of 26, and that gain compensates for the occasional memory allocations needed for the xarray nodes. We also end up using only 2 cache lines instead of 3 per delayed ref head. Running the following fs_mark test showed some improvements: $ cat test.sh #!/bin/bash DEV=/dev/nullb0 MNT=/mnt/nullb0 MOUNT_OPTIONS="-o ssd" FILES=100000 THREADS=$(nproc --all) echo "performance" | \ tee /sys/devices/system/cpu/cpu*/cpufreq/scaling_governor mkfs.btrfs -f $DEV mount $MOUNT_OPTIONS $DEV $MNT OPTS="-S 0 -L 5 -n $FILES -s 0 -t $THREADS -k" for ((i = 1; i <= $THREADS; i++)); do OPTS="$OPTS -d $MNT/d$i" done fs_mark $OPTS umount $MNT Before this patch: FSUse% Count Size Files/sec App Overhead 10 1200000 0 171845.7 12253839 16 2400000 0 230898.7 12308254 23 3600000 0 212292.9 12467768 30 4800000 0 195737.8 12627554 46 6000000 0 171055.2 12783329 After this patch: FSUse% Count Size Files/sec App Overhead 10 1200000 0 173835.0 12246131 16 2400000 0 233537.8 12271746 23 3600000 0 220398.7 12307737 30 4800000 0 204483.6 12392318 40 6000000 0 182923.3 12771843 Reviewed-by: Boris Burkov <boris@bur.io> Reviewed-by: Qu Wenruo <wqu@suse.com> Signed-off-by: Filipe Manana <fdmanana@suse.com> Reviewed-by: David Sterba <dsterba@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2024-10-23 16:27:14 +01:00
btrfs: track delayed ref heads in an xarray Currently we use a red black tree (rb-tree) to track the delayed ref heads (in struct btrfs_delayed_ref_root::href_root). This however is not very efficient when the number of delayed ref heads is large (and it's very common to be at least in the order of thousands) since rb-trees are binary trees. For example for 10K delayed ref heads, the tree has a depth of 13. Besides that, inserting into the tree requires navigating through it and pulling useless cache lines in the process since the red black tree nodes are embedded within the delayed ref head structure - on the other hand, by being embedded, it requires no extra memory allocations. We can improve this by using an xarray instead which has a much higher branching factor than a red black tree (binary balanced tree) and is more cache friendly and behaves like a resizable array, with a much better search and insertion complexity than a red black tree. This only has one small disadvantage which is that insertion will sometimes require allocating memory for the xarray - which may fail (not that often since it uses a kmem_cache) - but on the other hand we can reduce the delayed ref head structure size by 24 bytes (from 152 down to 128 bytes) after removing the embedded red black tree node, meaning than we can now fit 32 delayed ref heads per 4K page instead of 26, and that gain compensates for the occasional memory allocations needed for the xarray nodes. We also end up using only 2 cache lines instead of 3 per delayed ref head. Running the following fs_mark test showed some improvements: $ cat test.sh #!/bin/bash DEV=/dev/nullb0 MNT=/mnt/nullb0 MOUNT_OPTIONS="-o ssd" FILES=100000 THREADS=$(nproc --all) echo "performance" | \ tee /sys/devices/system/cpu/cpu*/cpufreq/scaling_governor mkfs.btrfs -f $DEV mount $MOUNT_OPTIONS $DEV $MNT OPTS="-S 0 -L 5 -n $FILES -s 0 -t $THREADS -k" for ((i = 1; i <= $THREADS; i++)); do OPTS="$OPTS -d $MNT/d$i" done fs_mark $OPTS umount $MNT Before this patch: FSUse% Count Size Files/sec App Overhead 10 1200000 0 171845.7 12253839 16 2400000 0 230898.7 12308254 23 3600000 0 212292.9 12467768 30 4800000 0 195737.8 12627554 46 6000000 0 171055.2 12783329 After this patch: FSUse% Count Size Files/sec App Overhead 10 1200000 0 173835.0 12246131 16 2400000 0 233537.8 12271746 23 3600000 0 220398.7 12307737 30 4800000 0 204483.6 12392318 40 6000000 0 182923.3 12771843 Reviewed-by: Boris Burkov <boris@bur.io> Reviewed-by: Qu Wenruo <wqu@suse.com> Signed-off-by: Filipe Manana <fdmanana@suse.com> Reviewed-by: David Sterba <dsterba@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2024-10-23 16:27:14 +01:00
Btrfs: do extent allocation and reference count updates in the background The extent allocation tree maintains a reference count and full back reference information for every extent allocated in the filesystem. For subvolume and snapshot trees, every time a block goes through COW, the new copy of the block adds a reference on every block it points to. If a btree node points to 150 leaves, then the COW code needs to go and add backrefs on 150 different extents, which might be spread all over the extent allocation tree. These updates currently happen during btrfs_cow_block, and most COWs happen during btrfs_search_slot. btrfs_search_slot has locks held on both the parent and the node we are COWing, and so we really want to avoid IO during the COW if we can. This commit adds an rbtree of pending reference count updates and extent allocations. The tree is ordered by byte number of the extent and byte number of the parent for the back reference. The tree allows us to: 1) Modify back references in something close to disk order, reducing seeks 2) Significantly reduce the number of modifications made as block pointers are balanced around 3) Do all of the extent insertion and back reference modifications outside of the performance critical btrfs_search_slot code. #3 has the added benefit of greatly reducing the btrfs stack footprint. The extent allocation tree modifications are done without the deep (and somewhat recursive) call chains used in the past. These delayed back reference updates must be done before the transaction commits, and so the rbtree is tied to the transaction. Throttling is implemented to help keep the queue of backrefs at a reasonable size. Since there was a similar mechanism in place for the extent tree extents, that is removed and replaced by the delayed reference tree. Yan Zheng <yan.zheng@oracle.com> helped review and fixup this code. Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-03-13 10:10:06 -04:00
btrfs: only let one thread pre-flush delayed refs in commit I've been running a stress test that runs 20 workers in their own subvolume, which are running an fsstress instance with 4 threads per worker, which is 80 total fsstress threads. In addition to this I'm running balance in the background as well as creating and deleting snapshots. This test takes around 12 hours to run normally, going slower and slower as the test goes on. The reason for this is because fsstress is running fsync sometimes, and because we're messing with block groups we often fall through to btrfs_commit_transaction, so will often have 20-30 threads all calling btrfs_commit_transaction at the same time. These all get stuck contending on the extent tree while they try to run delayed refs during the initial part of the commit. This is suboptimal, really because the extent tree is a single point of failure we only want one thread acting on that tree at once to reduce lock contention. Fix this by making the flushing mechanism a bit operation, to make it easy to use test_and_set_bit() in order to make sure only one task does this initial flush. Once we're into the transaction commit we only have one thread doing delayed ref running, it's just this initial pre-flush that is problematic. With this patch my stress test takes around 90 minutes to run, instead of 12 hours. The memory barrier is not necessary for the flushing bit as it's ordered, unlike plain int. The transaction state accessed in btrfs_should_end_transaction could be affected by that too as it's not always used under transaction lock. Upon Nikolay's analysis in [1] it's not necessary: In should_end_transaction it's read without holding any locks. (U) It's modified in btrfs_cleanup_transaction without holding the fs_info->trans_lock (U), but the STATE_ERROR flag is going to be set. set in cleanup_transaction under fs_info->trans_lock (L) set in btrfs_commit_trans to COMMIT_START under fs_info->trans_lock.(L) set in btrfs_commit_trans to COMMIT_DOING under fs_info->trans_lock.(L) set in btrfs_commit_trans to COMMIT_UNBLOCK under fs_info->trans_lock.(L) set in btrfs_commit_trans to COMMIT_COMPLETED without locks but at this point the transaction is finished and fs_info->running_trans is NULL (U but irrelevant). So by the looks of it we can have a concurrent READ race with a WRITE, due to reads not taking a lock. In this case what we want to ensure is we either see new or old state. I consulted with Will Deacon and he said that in such a case we'd want to annotate the accesses to ->state with (READ|WRITE)_ONCE so as to avoid a theoretical tear, in this case I don't think this could happen but I imagine at some point KCSAN would flag such an access as racy (which it is). [1] https://lore.kernel.org/linux-btrfs/e1fd5cc1-0f28-f670-69f4-e9958b4964e6@suse.com Reviewed-by: Nikolay Borisov <nborisov@suse.com> Signed-off-by: Josef Bacik <josef@toxicpanda.com> [ add comments regarding memory barrier ] Signed-off-by: David Sterba <dsterba@suse.com>
2020-12-18 14:24:20 -05:00
Btrfs: do extent allocation and reference count updates in the background The extent allocation tree maintains a reference count and full back reference information for every extent allocated in the filesystem. For subvolume and snapshot trees, every time a block goes through COW, the new copy of the block adds a reference on every block it points to. If a btree node points to 150 leaves, then the COW code needs to go and add backrefs on 150 different extents, which might be spread all over the extent allocation tree. These updates currently happen during btrfs_cow_block, and most COWs happen during btrfs_search_slot. btrfs_search_slot has locks held on both the parent and the node we are COWing, and so we really want to avoid IO during the COW if we can. This commit adds an rbtree of pending reference count updates and extent allocations. The tree is ordered by byte number of the extent and byte number of the parent for the back reference. The tree allows us to: 1) Modify back references in something close to disk order, reducing seeks 2) Significantly reduce the number of modifications made as block pointers are balanced around 3) Do all of the extent insertion and back reference modifications outside of the performance critical btrfs_search_slot code. #3 has the added benefit of greatly reducing the btrfs stack footprint. The extent allocation tree modifications are done without the deep (and somewhat recursive) call chains used in the past. These delayed back reference updates must be done before the transaction commits, and so the rbtree is tied to the transaction. Throttling is implemented to help keep the queue of backrefs at a reasonable size. Since there was a similar mechanism in place for the extent tree extents, that is removed and replaced by the delayed reference tree. Yan Zheng <yan.zheng@oracle.com> helped review and fixup this code. Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-03-13 10:10:06 -04:00
btrfs: delayed-ref: Introduce better documented delayed ref structures Current delayed ref interface has several problems: - Longer and longer parameter lists bytenr num_bytes parent ---------- so far so good ref_root owner offset ---------- I don't feel good now - Different interpretation of the same parameter Above @owner for data ref is inode number (u64), while for tree ref, it's level (int). They are even in different size range. For level we only need 0 ~ 8, while for ino it's BTRFS_FIRST_FREE_OBJECTID ~ BTRFS_LAST_FREE_OBJECTID. And @offset doesn't even make sense for tree ref. Such parameter reuse may look clever as an hidden union, but it destroys code readability. To solve both problems, we introduce a new structure, btrfs_ref to solve them: - Structure instead of long parameter list This makes later expansion easier, and is better documented. - Use btrfs_ref::type to distinguish data and tree ref - Use proper union to store data/tree ref specific structures. - Use separate functions to fill data/tree ref data, with a common generic function to fill common bytenr/num_bytes members. All parameters will find its place in btrfs_ref, and an extra member, @real_root, inspired by ref-verify code, is newly introduced for later qgroup code, to record which tree is triggered by this extent modification. This patch doesn't touch any code, but provides the basis for further refactoring. Signed-off-by: Qu Wenruo <wqu@suse.com> Reviewed-by: David Sterba <dsterba@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2019-04-04 14:45:29 +08:00
btrfs: delayed-ref: Introduce better documented delayed ref structures Current delayed ref interface has several problems: - Longer and longer parameter lists bytenr num_bytes parent ---------- so far so good ref_root owner offset ---------- I don't feel good now - Different interpretation of the same parameter Above @owner for data ref is inode number (u64), while for tree ref, it's level (int). They are even in different size range. For level we only need 0 ~ 8, while for ino it's BTRFS_FIRST_FREE_OBJECTID ~ BTRFS_LAST_FREE_OBJECTID. And @offset doesn't even make sense for tree ref. Such parameter reuse may look clever as an hidden union, but it destroys code readability. To solve both problems, we introduce a new structure, btrfs_ref to solve them: - Structure instead of long parameter list This makes later expansion easier, and is better documented. - Use btrfs_ref::type to distinguish data and tree ref - Use proper union to store data/tree ref specific structures. - Use separate functions to fill data/tree ref data, with a common generic function to fill common bytenr/num_bytes members. All parameters will find its place in btrfs_ref, and an extra member, @real_root, inspired by ref-verify code, is newly introduced for later qgroup code, to record which tree is triggered by this extent modification. This patch doesn't touch any code, but provides the basis for further refactoring. Signed-off-by: Qu Wenruo <wqu@suse.com> Reviewed-by: David Sterba <dsterba@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2019-04-04 14:45:29 +08:00
btrfs: delayed-ref: Introduce better documented delayed ref structures Current delayed ref interface has several problems: - Longer and longer parameter lists bytenr num_bytes parent ---------- so far so good ref_root owner offset ---------- I don't feel good now - Different interpretation of the same parameter Above @owner for data ref is inode number (u64), while for tree ref, it's level (int). They are even in different size range. For level we only need 0 ~ 8, while for ino it's BTRFS_FIRST_FREE_OBJECTID ~ BTRFS_LAST_FREE_OBJECTID. And @offset doesn't even make sense for tree ref. Such parameter reuse may look clever as an hidden union, but it destroys code readability. To solve both problems, we introduce a new structure, btrfs_ref to solve them: - Structure instead of long parameter list This makes later expansion easier, and is better documented. - Use btrfs_ref::type to distinguish data and tree ref - Use proper union to store data/tree ref specific structures. - Use separate functions to fill data/tree ref data, with a common generic function to fill common bytenr/num_bytes members. All parameters will find its place in btrfs_ref, and an extra member, @real_root, inspired by ref-verify code, is newly introduced for later qgroup code, to record which tree is triggered by this extent modification. This patch doesn't touch any code, but provides the basis for further refactoring. Signed-off-by: Qu Wenruo <wqu@suse.com> Reviewed-by: David Sterba <dsterba@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2019-04-04 14:45:29 +08:00
btrfs: delayed-ref: Introduce better documented delayed ref structures Current delayed ref interface has several problems: - Longer and longer parameter lists bytenr num_bytes parent ---------- so far so good ref_root owner offset ---------- I don't feel good now - Different interpretation of the same parameter Above @owner for data ref is inode number (u64), while for tree ref, it's level (int). They are even in different size range. For level we only need 0 ~ 8, while for ino it's BTRFS_FIRST_FREE_OBJECTID ~ BTRFS_LAST_FREE_OBJECTID. And @offset doesn't even make sense for tree ref. Such parameter reuse may look clever as an hidden union, but it destroys code readability. To solve both problems, we introduce a new structure, btrfs_ref to solve them: - Structure instead of long parameter list This makes later expansion easier, and is better documented. - Use btrfs_ref::type to distinguish data and tree ref - Use proper union to store data/tree ref specific structures. - Use separate functions to fill data/tree ref data, with a common generic function to fill common bytenr/num_bytes members. All parameters will find its place in btrfs_ref, and an extra member, @real_root, inspired by ref-verify code, is newly introduced for later qgroup code, to record which tree is triggered by this extent modification. This patch doesn't touch any code, but provides the basis for further refactoring. Signed-off-by: Qu Wenruo <wqu@suse.com> Reviewed-by: David Sterba <dsterba@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2019-04-04 14:45:29 +08:00
btrfs: delayed-ref: Introduce better documented delayed ref structures Current delayed ref interface has several problems: - Longer and longer parameter lists bytenr num_bytes parent ---------- so far so good ref_root owner offset ---------- I don't feel good now - Different interpretation of the same parameter Above @owner for data ref is inode number (u64), while for tree ref, it's level (int). They are even in different size range. For level we only need 0 ~ 8, while for ino it's BTRFS_FIRST_FREE_OBJECTID ~ BTRFS_LAST_FREE_OBJECTID. And @offset doesn't even make sense for tree ref. Such parameter reuse may look clever as an hidden union, but it destroys code readability. To solve both problems, we introduce a new structure, btrfs_ref to solve them: - Structure instead of long parameter list This makes later expansion easier, and is better documented. - Use btrfs_ref::type to distinguish data and tree ref - Use proper union to store data/tree ref specific structures. - Use separate functions to fill data/tree ref data, with a common generic function to fill common bytenr/num_bytes members. All parameters will find its place in btrfs_ref, and an extra member, @real_root, inspired by ref-verify code, is newly introduced for later qgroup code, to record which tree is triggered by this extent modification. This patch doesn't touch any code, but provides the basis for further refactoring. Signed-off-by: Qu Wenruo <wqu@suse.com> Reviewed-by: David Sterba <dsterba@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2019-04-04 14:45:29 +08:00
btrfs: delayed-ref: Introduce better documented delayed ref structures Current delayed ref interface has several problems: - Longer and longer parameter lists bytenr num_bytes parent ---------- so far so good ref_root owner offset ---------- I don't feel good now - Different interpretation of the same parameter Above @owner for data ref is inode number (u64), while for tree ref, it's level (int). They are even in different size range. For level we only need 0 ~ 8, while for ino it's BTRFS_FIRST_FREE_OBJECTID ~ BTRFS_LAST_FREE_OBJECTID. And @offset doesn't even make sense for tree ref. Such parameter reuse may look clever as an hidden union, but it destroys code readability. To solve both problems, we introduce a new structure, btrfs_ref to solve them: - Structure instead of long parameter list This makes later expansion easier, and is better documented. - Use btrfs_ref::type to distinguish data and tree ref - Use proper union to store data/tree ref specific structures. - Use separate functions to fill data/tree ref data, with a common generic function to fill common bytenr/num_bytes members. All parameters will find its place in btrfs_ref, and an extra member, @real_root, inspired by ref-verify code, is newly introduced for later qgroup code, to record which tree is triggered by this extent modification. This patch doesn't touch any code, but provides the basis for further refactoring. Signed-off-by: Qu Wenruo <wqu@suse.com> Reviewed-by: David Sterba <dsterba@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2019-04-04 14:45:29 +08:00
btrfs: stop doing excessive space reservation for csum deletion Currently when reserving space for deleting the csum items for a data extent, when adding or updating a delayed ref head, we determine how many leaves of csum items we can have and then pass that number to the helper btrfs_calc_delayed_ref_bytes(). This helper is used for calculating space for all tree modifications we need when running delayed references, however the amount of space it computes is excessive for deleting csum items because: 1) It uses btrfs_calc_insert_metadata_size() which is excessive because we only need to delete csum items from the csum tree, we don't need to insert any items, so btrfs_calc_metadata_size() is all we need (as it computes space needed to delete an item); 2) If the free space tree is enabled, it doubles the amount of space, which is pointless for csum deletion since we don't need to touch the free space tree or any other tree other than the csum tree. So improve on this by tracking how many csum deletions we have and using a new helper to calculate space for csum deletions (just a wrapper around btrfs_calc_metadata_size() with a comment). This reduces the amount of space we need to reserve for csum deletions by a factor of 4, and it helps reduce the number of times we have to block space reservations and have the reclaim task enter the space flushing algorithm (flush delayed items, flush delayed refs, etc) in order to satisfy tickets. For example this results in a total time decrease when unlinking (or truncating) files with many extents, as we end up having to block on space metadata reservations less often. Example test: $ cat test.sh #!/bin/bash DEV=/dev/nullb0 MNT=/mnt/test umount $DEV &> /dev/null mkfs.btrfs -f $DEV # Use compression to quickly create files with a lot of extents # (each with a size of 128K). mount -o compress=lzo $DEV $MNT # 100G gives at least 983040 extents with a size of 128K. xfs_io -f -c "pwrite -S 0xab -b 1M 0 120G" $MNT/foobar # Flush all delalloc and clear all metadata from memory. umount $MNT mount -o compress=lzo $DEV $MNT start=$(date +%s%N) rm -f $MNT/foobar end=$(date +%s%N) dur=$(( (end - start) / 1000000 )) echo "rm took $dur milliseconds" umount $MNT Before this change rm took: 7504 milliseconds After this change rm took: 6574 milliseconds (-12.4%) Reviewed-by: Josef Bacik <josef@toxicpanda.com> Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2023-09-08 18:20:37 +01:00
btrfs: delayed-ref: Introduce better documented delayed ref structures Current delayed ref interface has several problems: - Longer and longer parameter lists bytenr num_bytes parent ---------- so far so good ref_root owner offset ---------- I don't feel good now - Different interpretation of the same parameter Above @owner for data ref is inode number (u64), while for tree ref, it's level (int). They are even in different size range. For level we only need 0 ~ 8, while for ino it's BTRFS_FIRST_FREE_OBJECTID ~ BTRFS_LAST_FREE_OBJECTID. And @offset doesn't even make sense for tree ref. Such parameter reuse may look clever as an hidden union, but it destroys code readability. To solve both problems, we introduce a new structure, btrfs_ref to solve them: - Structure instead of long parameter list This makes later expansion easier, and is better documented. - Use btrfs_ref::type to distinguish data and tree ref - Use proper union to store data/tree ref specific structures. - Use separate functions to fill data/tree ref data, with a common generic function to fill common bytenr/num_bytes members. All parameters will find its place in btrfs_ref, and an extra member, @real_root, inspired by ref-verify code, is newly introduced for later qgroup code, to record which tree is triggered by this extent modification. This patch doesn't touch any code, but provides the basis for further refactoring. Signed-off-by: Qu Wenruo <wqu@suse.com> Reviewed-by: David Sterba <dsterba@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2019-04-04 14:45:29 +08:00
Btrfs: do extent allocation and reference count updates in the background The extent allocation tree maintains a reference count and full back reference information for every extent allocated in the filesystem. For subvolume and snapshot trees, every time a block goes through COW, the new copy of the block adds a reference on every block it points to. If a btree node points to 150 leaves, then the COW code needs to go and add backrefs on 150 different extents, which might be spread all over the extent allocation tree. These updates currently happen during btrfs_cow_block, and most COWs happen during btrfs_search_slot. btrfs_search_slot has locks held on both the parent and the node we are COWing, and so we really want to avoid IO during the COW if we can. This commit adds an rbtree of pending reference count updates and extent allocations. The tree is ordered by byte number of the extent and byte number of the parent for the back reference. The tree allows us to: 1) Modify back references in something close to disk order, reducing seeks 2) Significantly reduce the number of modifications made as block pointers are balanced around 3) Do all of the extent insertion and back reference modifications outside of the performance critical btrfs_search_slot code. #3 has the added benefit of greatly reducing the btrfs stack footprint. The extent allocation tree modifications are done without the deep (and somewhat recursive) call chains used in the past. These delayed back reference updates must be done before the transaction commits, and so the rbtree is tied to the transaction. Throttling is implemented to help keep the queue of backrefs at a reasonable size. Since there was a similar mechanism in place for the extent tree extents, that is removed and replaced by the delayed reference tree. Yan Zheng <yan.zheng@oracle.com> helped review and fixup this code. Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-03-13 10:10:06 -04:00
Btrfs: Mixed back reference (FORWARD ROLLING FORMAT CHANGE) This commit introduces a new kind of back reference for btrfs metadata. Once a filesystem has been mounted with this commit, IT WILL NO LONGER BE MOUNTABLE BY OLDER KERNELS. When a tree block in subvolume tree is cow'd, the reference counts of all extents it points to are increased by one. At transaction commit time, the old root of the subvolume is recorded in a "dead root" data structure, and the btree it points to is later walked, dropping reference counts and freeing any blocks where the reference count goes to 0. The increments done during cow and decrements done after commit cancel out, and the walk is a very expensive way to go about freeing the blocks that are no longer referenced by the new btree root. This commit reduces the transaction overhead by avoiding the need for dead root records. When a non-shared tree block is cow'd, we free the old block at once, and the new block inherits old block's references. When a tree block with reference count > 1 is cow'd, we increase the reference counts of all extents the new block points to by one, and decrease the old block's reference count by one. This dead tree avoidance code removes the need to modify the reference counts of lower level extents when a non-shared tree block is cow'd. But we still need to update back ref for all pointers in the block. This is because the location of the block is recorded in the back ref item. We can solve this by introducing a new type of back ref. The new back ref provides information about pointer's key, level and in which tree the pointer lives. This information allow us to find the pointer by searching the tree. The shortcoming of the new back ref is that it only works for pointers in tree blocks referenced by their owner trees. This is mostly a problem for snapshots, where resolving one of these fuzzy back references would be O(number_of_snapshots) and quite slow. The solution used here is to use the fuzzy back references in the common case where a given tree block is only referenced by one root, and use the full back references when multiple roots have a reference on a given block. This commit adds per subvolume red-black tree to keep trace of cached inodes. The red-black tree helps the balancing code to find cached inodes whose inode numbers within a given range. This commit improves the balancing code by introducing several data structures to keep the state of balancing. The most important one is the back ref cache. It caches how the upper level tree blocks are referenced. This greatly reduce the overhead of checking back ref. The improved balancing code scales significantly better with a large number of snapshots. This is a very large commit and was written in a number of pieces. But, they depend heavily on the disk format change and were squashed together to make sure git bisect didn't end up in a bad state wrt space balancing or the format change. Signed-off-by: Yan Zheng <zheng.yan@oracle.com> Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-06-10 10:45:14 -04:00
Btrfs: allow delayed refs to be merged Daniel Blueman reported a bug with fio+balance on a ramdisk setup. Basically what happens is the balance relocates a tree block which will drop the implicit refs for all of its children and adds a full backref. Once the block is relocated we have to add the implicit refs back, so when we cow the block again we add the implicit refs for its children back. The problem comes when the original drop ref doesn't get run before we add the implicit refs back. The delayed ref stuff will specifically prefer ADD operations over DROP to keep us from freeing up an extent that will have references to it, so we try to add the implicit ref before it is actually removed and we panic. This worked fine before because the add would have just canceled the drop out and we would have been fine. But the backref walking work needs to be able to freeze the delayed ref stuff in time so we have this ever increasing sequence number that gets attached to all new delayed ref updates which makes us not merge refs and we run into this issue. So to fix this we need to merge delayed refs. So everytime we run a clustered ref we need to try and merge all of its delayed refs. The backref walking stuff locks the delayed ref head before processing, so if we have it locked we are safe to merge any refs inside of the sequence number. If there is no sequence number we can merge all refs. Doing this not only fixes our bug but keeps the delayed ref code from adding and removing useless refs and batching together multiple refs into one search instead of one search per delayed ref, which will really help our commit times. I ran this with Daniels test and 276 and I haven't seen any problems. Thanks, Reported-by: Daniel J Blueman <daniel@quora.org> Signed-off-by: Josef Bacik <jbacik@fusionio.com>
2012-08-07 16:00:32 -04:00
Btrfs: do extent allocation and reference count updates in the background The extent allocation tree maintains a reference count and full back reference information for every extent allocated in the filesystem. For subvolume and snapshot trees, every time a block goes through COW, the new copy of the block adds a reference on every block it points to. If a btree node points to 150 leaves, then the COW code needs to go and add backrefs on 150 different extents, which might be spread all over the extent allocation tree. These updates currently happen during btrfs_cow_block, and most COWs happen during btrfs_search_slot. btrfs_search_slot has locks held on both the parent and the node we are COWing, and so we really want to avoid IO during the COW if we can. This commit adds an rbtree of pending reference count updates and extent allocations. The tree is ordered by byte number of the extent and byte number of the parent for the back reference. The tree allows us to: 1) Modify back references in something close to disk order, reducing seeks 2) Significantly reduce the number of modifications made as block pointers are balanced around 3) Do all of the extent insertion and back reference modifications outside of the performance critical btrfs_search_slot code. #3 has the added benefit of greatly reducing the btrfs stack footprint. The extent allocation tree modifications are done without the deep (and somewhat recursive) call chains used in the past. These delayed back reference updates must be done before the transaction commits, and so the rbtree is tied to the transaction. Throttling is implemented to help keep the queue of backrefs at a reasonable size. Since there was a similar mechanism in place for the extent tree extents, that is removed and replaced by the delayed reference tree. Yan Zheng <yan.zheng@oracle.com> helped review and fixup this code. Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-03-13 10:10:06 -04:00
btrfs: stop doing excessive space reservation for csum deletion Currently when reserving space for deleting the csum items for a data extent, when adding or updating a delayed ref head, we determine how many leaves of csum items we can have and then pass that number to the helper btrfs_calc_delayed_ref_bytes(). This helper is used for calculating space for all tree modifications we need when running delayed references, however the amount of space it computes is excessive for deleting csum items because: 1) It uses btrfs_calc_insert_metadata_size() which is excessive because we only need to delete csum items from the csum tree, we don't need to insert any items, so btrfs_calc_metadata_size() is all we need (as it computes space needed to delete an item); 2) If the free space tree is enabled, it doubles the amount of space, which is pointless for csum deletion since we don't need to touch the free space tree or any other tree other than the csum tree. So improve on this by tracking how many csum deletions we have and using a new helper to calculate space for csum deletions (just a wrapper around btrfs_calc_metadata_size() with a comment). This reduces the amount of space we need to reserve for csum deletions by a factor of 4, and it helps reduce the number of times we have to block space reservations and have the reclaim task enter the space flushing algorithm (flush delayed items, flush delayed refs, etc) in order to satisfy tickets. For example this results in a total time decrease when unlinking (or truncating) files with many extents, as we end up having to block on space metadata reservations less often. Example test: $ cat test.sh #!/bin/bash DEV=/dev/nullb0 MNT=/mnt/test umount $DEV &> /dev/null mkfs.btrfs -f $DEV # Use compression to quickly create files with a lot of extents # (each with a size of 128K). mount -o compress=lzo $DEV $MNT # 100G gives at least 983040 extents with a size of 128K. xfs_io -f -c "pwrite -S 0xab -b 1M 0 120G" $MNT/foobar # Flush all delalloc and clear all metadata from memory. umount $MNT mount -o compress=lzo $DEV $MNT start=$(date +%s%N) rm -f $MNT/foobar end=$(date +%s%N) dur=$(( (end - start) / 1000000 )) echo "rm took $dur milliseconds" umount $MNT Before this change rm took: 7504 milliseconds After this change rm took: 6574 milliseconds (-12.4%) Reviewed-by: Josef Bacik <josef@toxicpanda.com> Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2023-09-08 18:20:37 +01:00
btrfs: check delayed refs when we're checking if a ref exists In the patch 78c52d9eb6b7 ("btrfs: check for refs on snapshot delete resume") I added some code to handle file systems that had been corrupted by a bug that incorrectly skipped updating the drop progress key while dropping a snapshot. This code would check to see if we had already deleted our reference for a child block, and skip the deletion if we had already. Unfortunately there is a bug, as the check would only check the on-disk references. I made an incorrect assumption that blocks in an already deleted snapshot that was having the deletion resume on mount wouldn't be modified. If we have 2 pending deleted snapshots that share blocks, we can easily modify the rules for a block. Take the following example subvolume a exists, and subvolume b is a snapshot of subvolume a. They share references to block 1. Block 1 will have 2 full references, one for subvolume a and one for subvolume b, and it belongs to subvolume a (btrfs_header_owner(block 1) == subvolume a). When deleting subvolume a, we will drop our full reference for block 1, and because we are the owner we will drop our full reference for all of block 1's children, convert block 1 to FULL BACKREF, and add a shared reference to all of block 1's children. Then we will start the snapshot deletion of subvolume b. We look up the extent info for block 1, which checks delayed refs and tells us that FULL BACKREF is set, so sets parent to the bytenr of block 1. However because this is a resumed snapshot deletion, we call into check_ref_exists(). Because check_ref_exists() only looks at the disk, it doesn't find the shared backref for the child of block 1, and thus returns 0 and we skip deleting the reference for the child of block 1 and continue. This orphans the child of block 1. The fix is to lookup the delayed refs, similar to what we do in btrfs_lookup_extent_info(). However we only care about whether the reference exists or not. If we fail to find our reference on disk, go look up the bytenr in the delayed refs, and if it exists look for an existing ref in the delayed ref head. If that exists then we know we can delete the reference safely and carry on. If it doesn't exist we know we have to skip over this block. This bug has existed since I introduced this fix, however requires having multiple deleted snapshots pending when we unmount. We noticed this in production because our shutdown path stops the container on the system, which deletes a bunch of subvolumes, and then reboots the box. This gives us plenty of opportunities to hit this issue. Looking at the history we've seen this occasionally in production, but we had a big spike recently thanks to faster machines getting jobs with multiple subvolumes in the job. Chris Mason wrote a reproducer which does the following mount /dev/nvme4n1 /btrfs btrfs subvol create /btrfs/s1 simoop -E -f 4k -n 200000 -z /btrfs/s1 while(true) ; do btrfs subvol snap /btrfs/s1 /btrfs/s2 simoop -f 4k -n 200000 -r 10 -z /btrfs/s2 btrfs subvol snap /btrfs/s2 /btrfs/s3 btrfs balance start -dusage=80 /btrfs btrfs subvol del /btrfs/s2 /btrfs/s3 umount /btrfs btrfsck /dev/nvme4n1 || exit 1 mount /dev/nvme4n1 /btrfs done On the second loop this would fail consistently, with my patch it has been running for hours and hasn't failed. I also used dm-log-writes to capture the state of the failure so I could debug the problem. Using the existing failure case to test my patch validated that it fixes the problem. Fixes: 78c52d9eb6b7 ("btrfs: check for refs on snapshot delete resume") CC: stable@vger.kernel.org # 5.4+ Reviewed-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: Josef Bacik <josef@toxicpanda.com> Signed-off-by: David Sterba <dsterba@suse.com>
2024-04-11 16:41:20 -04:00
Btrfs: do extent allocation and reference count updates in the background The extent allocation tree maintains a reference count and full back reference information for every extent allocated in the filesystem. For subvolume and snapshot trees, every time a block goes through COW, the new copy of the block adds a reference on every block it points to. If a btree node points to 150 leaves, then the COW code needs to go and add backrefs on 150 different extents, which might be spread all over the extent allocation tree. These updates currently happen during btrfs_cow_block, and most COWs happen during btrfs_search_slot. btrfs_search_slot has locks held on both the parent and the node we are COWing, and so we really want to avoid IO during the COW if we can. This commit adds an rbtree of pending reference count updates and extent allocations. The tree is ordered by byte number of the extent and byte number of the parent for the back reference. The tree allows us to: 1) Modify back references in something close to disk order, reducing seeks 2) Significantly reduce the number of modifications made as block pointers are balanced around 3) Do all of the extent insertion and back reference modifications outside of the performance critical btrfs_search_slot code. #3 has the added benefit of greatly reducing the btrfs stack footprint. The extent allocation tree modifications are done without the deep (and somewhat recursive) call chains used in the past. These delayed back reference updates must be done before the transaction commits, and so the rbtree is tied to the transaction. Throttling is implemented to help keep the queue of backrefs at a reasonable size. Since there was a similar mechanism in place for the extent tree extents, that is removed and replaced by the delayed reference tree. Yan Zheng <yan.zheng@oracle.com> helped review and fixup this code. Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-03-13 10:10:06 -04:00