Snapshots have been merged! 9 months of work and 3k lines of new code, finally
released. Some highlights:
- btrfs style subvolumes & snapshots interface
- snapshots are writeable
- highly scalable: number of snapshots is limited only by your disk space
- highly space efficient: no internal fragmentation issues
Design doc here: https://bcachefs.org/Snapshots/
The core functionality is complete - snapshot creation and deletion works, fsck
changes are done (most of the complexity was in making fsck work without
O(number of snapshots) performance - tricky). Everything else is a todo item:
- still need to export different st_dev for files in different subvolumes
(we'll never allocate a new inode with an inode number that collides with an
inode inother subvolume - but snapshots will naturally result in colliding
inode numbers)
- need to hide dirents that point to snapshots when inside snapshots...
- snapshot creation is not atomic w.r.t. page cache, we do sync_fs() but don't
block buffered writes
- other niggling page cache stuff - need to walk page cache and mark blocks as
no longer reserved on snapshot creation
- we no longer have quota support, since old style quotas interact badly with
snapshots
- we need per subvolume disk space accounting before i can implement btrfs
style subvolume quotas
- all the things I neglected to think of yet, and all the bugs I haven't found
yet
Go wild, please try and break it.
On 9/27/21 3:49 AM, Kent Overstreet wrote:
> Snapshots have been merged! 9 months of work and 3k lines of new code, finally
> released. Some highlights:
>
> - btrfs style subvolumes & snapshots interface
> - snapshots are writeable
> - highly scalable: number of snapshots is limited only by your disk space
> - highly space efficient: no internal fragmentation issues
>
> Design doc here: https://bcachefs.org/Snapshots/
>
> The core functionality is complete - snapshot creation and deletion works, fsck
> changes are done (most of the complexity was in making fsck work without
> O(number of snapshots) performance - tricky). Everything else is a todo item:
>
> - still need to export different st_dev for files in different subvolumes
> (we'll never allocate a new inode with an inode number that collides with an
> inode inother subvolume - but snapshots will naturally result in colliding
> inode numbers)
With my limited high level view on it - shouldn't you discuss with Neil
about a solution and to avoid going the btrfs route for colliding inode
numbers?
> On 9/27/21 3:49 AM, Kent Overstreet wrote:
> > Snapshots have been merged! 9 months of work and 3k lines of new code,
> > finally released. Some highlights:
> >
> > - btrfs style subvolumes & snapshots interface
> > - snapshots are writeable
> > - highly scalable: number of snapshots is limited only by your disk
> > space
> > - highly space efficient: no internal fragmentation issues
> >
> > Design doc here: https://bcachefs.org/Snapshots/
> >
> > The core functionality is complete - snapshot creation and deletion
> > works, fsck changes are done (most of the complexity was in making
> > fsck work without O(number of snapshots) performance - tricky). Everything
> else is a todo item:
> >
> > - still need to export different st_dev for files in different subvolumes
> > (we'll never allocate a new inode with an inode number that collides with an
> > inode inother subvolume - but snapshots will naturally result in colliding
> > inode numbers)
>
> With my limited high level view on it - shouldn't you discuss with Neil about a
> solution and to avoid going the btrfs route for colliding inode numbers?
I was going to ask that also having been watching the btrfs subvolume saga. As maintainer of the Ganesha user space NFS server I have an interest in this also though we haven't had anyone talk about bcachefs yet.
Frank