btrfs_clone_files and bind mounts
Subsystems (4)
What is the problem with cloning files between different (vfs)mounts of the same filesystem?
Quoted 3 lines · Hristo Venev
What is the problem with cloning files between different (vfs)mounts of the same filesystem?
The "problem" is not really a problem, but rather a well-imposed restriction:
Quoted 1 line
From http://man7.org/linux/man-pages/man2/ioctl_ficlonerange.2.html
"Both files must reside within the same filesystem."
And as a matter of fact this is enforced in the generic vfs_clone_file_range. Of course if we were to do this across filesystem then we'd have all the problems associated with not being able to ensure atomicity of operations.
Quoted 12 lines · Nikolay Borisov
Quoted 4 lines · Hristo Venev
What is the problem with cloning files between different (vfs)mounts of
the same filesystem?The "problem" is not really a problem, but rather a well-imposed restriction:
From http://man7.org/linux/man-pages/man2/ioctl_ficlonerange.2.html
Quoted 8 lines
"Both files must reside within the same filesystem."
And as a matter of fact this is enforced in the generic
vfs_clone_file_range. Of course if we were to do this across filesystem
then we'd have all the problems associated with not being able to ensure
atomicity of operations.
My question was about doing this within the same filesystem. In my case (and I think it's relatively common), subvolumes of the same filesystem are mounted separately, and I can't think of a good reason why FICLONERANGE can't be made to work (it works if the subvolumes are accessed within the same mount point).
[CCing Amir since he touched precisely the code responsible for this semantics]
Quoted 22 lines · Hristo Venev
Quoted 12 lines · Nikolay Borisov
Quoted 4 lines · Hristo Venev
What is the problem with cloning files between different (vfs)mounts of
the same filesystem?The "problem" is not really a problem, but rather a well-imposed restriction:
From http://man7.org/linux/man-pages/man2/ioctl_ficlonerange.2.html
Quoted 8 lines
"Both files must reside within the same filesystem."
And as a matter of fact this is enforced in the generic
vfs_clone_file_range. Of course if we were to do this across filesystem
then we'd have all the problems associated with not being able to ensure
atomicity of operations.
Quoted 6 lines
My question was about doing this within the same filesystem. In my case (and I think it's relatively common), subvolumes of the same filesystem are mounted separately, and I can't think of a good reason why FICLONERANGE can't be made to work (it works if the subvolumes are accessed within the same mount point).
So Amir,
In 913b86e92e1f ("vfs: allow vfs_clone_file_range() across mount points") you did relax the requirements of vfs_clone_file_range by moving the mount point check into ioctl_file_clone. In btrfs we can have a situation where multiple subvolumes, belonging to the same FS are at different mountpoint I.e. :
mount /dev/vdc /media/scratch/
echo "test file" >> /media/scratch/foo
btrfs subvolume create /media/scratch/subvol1
btrfs subvolume list /media/scratch/
ID 257 gen 7 top level 5 path subvol1
mount -osubvolid=257 /dev/vdc /media/subvol-mount/
cp --reflink=always /media/scratch/foo /media/subvol-mount/foo-reflink
cp: failed to clone '/media/subvol-mount/foo-reflink' from '/media/scratch/foo': Invalid cross-device link
Shouldn't reflinking be allowed in this case? ftrace shows that we are failing "if (src_file.file->f_path.mnt != dst_file->f_path.mnt)" in ioctl_file_clone
In your commit you state:
FICLONE/FICLONERANGE ioctls enforce that src and dest files are on
the same mount.
Whereas the man page says they should be on the same filesystem (not necessarily on the same mount?)
Quoted 2 lines · Nikolay Borisov
[CCing Amir since he touched precisely the code responsible for this semantics]
[CCing fsdevel and NFS list since there is already a long standing attempt
to relax the copy_file_range() API for cross server copy]
Quoted 36 lines · Hristo Venev
Quoted 22 lines · Nikolay Borisov
Quoted 12 lines · Hristo Venev
Quoted 4 lines
What is the problem with cloning files between different (vfs)mounts of
the same filesystem?The "problem" is not really a problem, but rather a well-imposed restriction:
From http://man7.org/linux/man-pages/man2/ioctl_ficlonerange.2.html
Quoted 8 lines
"Both files must reside within the same filesystem."
And as a matter of fact this is enforced in the generic
vfs_clone_file_range. Of course if we were to do this across filesystem
then we'd have all the problems associated with not being able to ensure
atomicity of operations.Quoted 6 lines
My question was about doing this within the same filesystem. In my case (and I think it's relatively common), subvolumes of the same filesystem are mounted separately, and I can't think of a good reason why FICLONERANGE can't be made to work (it works if the subvolumes are accessed within the same mount point).
So Amir,
In
913b86e92e1f("vfs: allowvfs_clone_file_range()across mount points") you did relax the requirements ofvfs_clone_file_rangeby moving the mount point check intoioctl_file_clone.
[...]
Quoted 7 lines
Shouldn't reflinking be allowed in this case? ftrace shows that we are failing "if (
src_file.file->f_path.mnt!=dst_file->f_path.mnt)" inioctl_file_clone
Quoted 7 lines
In your commit you state:
FICLONE/FICLONERANGE ioctls enforce that src and dest files are on
the same mount.Whereas the man page says they should be on the same filesystem (not necessarily on the same mount?)
See, all I did was avoid the flames involved with changing application behavior, because I had no reason to change it.
Not that the man page is what matters, its what applications expect that matter, but the man page also says this open for interpretation phrase:
EXDEV dest_fd and src_fd are not on the same *mounted* filesystem.
NFS folks have been trying to change the behavior of copy_file_range() for over two years now, we little success as far as I can tell:
https://marc.info/?l=linux-fsdevel&m=149971144627690&w=2
Semi-interesting point - it looks like this commit has changed behavior of clone file range on CIFS:04b38d601239 vfs: pull btrfs clone API to vfs layer
I don't see that CIFS has the "same mount" restriction before the generalized VFS API, but btrfs had the restriction in-place and it still exists in the btrfs implementation btrfs_clone_files().
So I think you will have no way around this other than proposing a new ioctl FICLONERANGE2 with flags argument to opt-in for the new required behavior. I you do that please CC linux-api and expect the flames.
Thanks,
Amir.