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

Reply via email to