Due to the limitation of btrfs_cross_ref_exist(), run_delalloc_nocow() can still fall back to CoW even only (unrelated) part of the preallocated extent is shared.
This makes the follow case to do unnecessary CoW: # xfs_io -f -c "falloc 0 2M" $mnt/file # xfs_io -c "pwrite 0 1M" $mnt/file # xfs_io -c "reflink $mnt/file 1M 4M 1M" $mnt/file # sync The pwrite will still be CoWed, since at writeback time, the preallocated extent is already shared, btrfs_cross_ref_exist() will return 1 and make run_delalloc_nocow() fall back to cow_file_range(). This is definitely an overkilling workaround, but this should be the simplest way without further screwing up already complex NOCOW routine. Signed-off-by: Qu Wenruo <[email protected]> --- fs/btrfs/ctree.h | 1 + fs/btrfs/file.c | 4 ++-- fs/btrfs/ioctl.c | 21 +++++++++++++++++++++ 3 files changed, 24 insertions(+), 2 deletions(-) diff --git a/fs/btrfs/ctree.h b/fs/btrfs/ctree.h index 53af9f5253f4..ddacc41ff124 100644 --- a/fs/btrfs/ctree.h +++ b/fs/btrfs/ctree.h @@ -3228,6 +3228,7 @@ int btrfs_add_inode_defrag(struct btrfs_trans_handle *trans, struct btrfs_inode *inode); int btrfs_run_defrag_inodes(struct btrfs_fs_info *fs_info); void btrfs_cleanup_defrag_inodes(struct btrfs_fs_info *fs_info); +int btrfs_start_ordered_ops(struct inode *inode, loff_t start, loff_t end); int btrfs_sync_file(struct file *file, loff_t start, loff_t end, int datasync); void btrfs_drop_extent_cache(struct btrfs_inode *inode, u64 start, u64 end, int skip_pinned); diff --git a/fs/btrfs/file.c b/fs/btrfs/file.c index 2be00e873e92..118bfd019c6c 100644 --- a/fs/btrfs/file.c +++ b/fs/btrfs/file.c @@ -1999,7 +1999,7 @@ int btrfs_release_file(struct inode *inode, struct file *filp) return 0; } -static int start_ordered_ops(struct inode *inode, loff_t start, loff_t end) +int btrfs_start_ordered_ops(struct inode *inode, loff_t start, loff_t end) { int ret; struct blk_plug plug; @@ -2056,7 +2056,7 @@ int btrfs_sync_file(struct file *file, loff_t start, loff_t end, int datasync) * multi-task, and make the performance up. See * btrfs_wait_ordered_range for an explanation of the ASYNC check. */ - ret = start_ordered_ops(inode, start, end); + ret = btrfs_start_ordered_ops(inode, start, end); if (ret) goto out; diff --git a/fs/btrfs/ioctl.c b/fs/btrfs/ioctl.c index 63600dc2ac4c..866979f530bc 100644 --- a/fs/btrfs/ioctl.c +++ b/fs/btrfs/ioctl.c @@ -4266,6 +4266,27 @@ static noinline int btrfs_clone_files(struct file *file, struct file *file_src, goto out_unlock; } + /* + * btrfs_cross_ref_exist() only does check at extent level, + * we could cause unexpected NOCOW write to be COWed. + * E.g.: + * falloc 0 2M file1 + * pwrite 0 1M file1 (at this point it should go NOCOW) + * reflink src=file1 srcoff=1M dst=file1 dstoff=4M len=1M + * sync + * + * In above case, due to the preallocated extent is shared + * the data at 0~1M can't go NOCOW. + * + * So flush the whole src inode to avoid any unneeded CoW. + */ + ret = btrfs_start_ordered_ops(src, 0, -1); + if (ret < 0) + goto out_unlock; + ret = btrfs_wait_ordered_range(src, 0, -1); + if (ret < 0) + goto out_unlock; + /* * Lock the target range too. Right after we replace the file extent * items in the fs tree (which now point to the cloned data), we might -- 2.18.0
