Re: [PATCH v5 9/9] btrfs: btrfs_copy_file_range() only supports reflinks

On Fri, Oct 16, 2015 at 05:25:44AM -0700, Christoph Hellwig wrote:

On Fri, Oct 16, 2015 at 07:49:19AM -0400, Chris Mason wrote:

quoted

quoted

Yes, that would be my preference.  I'd also like to understand what
exactly btrfs does in fallocate.

For which part?  The answer changes based on how many references there
are to a given fallocated region.

Both cases.  With btrfs allocating new block on every write how do you
avoid that ENOSPC?  Is there a unassigned block preallocation that's
made persistent in some way?

So:

fallocate 1g -> foo

reflink foo foo2

We've now implicitly doubled the size of the fallocate, but at reflink
time btrfs doesn't account for the doubling.  It's actually much
better in this case to just use a hole because neither foo or foo2 can
use the preallocated space until the 1g is fully unshared.

When we're doing writes, it'll check the preallocated extents for extra
refs and force COW if any exist.  So writes into a preallocated region
can enospc.

-chris

On Fri, Oct 16, 2015 at 09:19:50AM -0400, Chris Mason wrote:

On Fri, Oct 16, 2015 at 05:25:44AM -0700, Christoph Hellwig wrote:

quoted

On Fri, Oct 16, 2015 at 07:49:19AM -0400, Chris Mason wrote:

quoted

quoted

Yes, that would be my preference.  I'd also like to understand what
exactly btrfs does in fallocate.

For which part?  The answer changes based on how many references there
are to a given fallocated region.

Both cases.  With btrfs allocating new block on every write how do you
avoid that ENOSPC?  Is there a unassigned block preallocation that's
made persistent in some way?

So:

fallocate 1g -> foo

reflink foo foo2

We've now implicitly doubled the size of the fallocate, but at reflink

No, I don't think it implies that at all. the posix_fallocate()
"future writes will succeed" guarantee only applies to foo, not to
/copies/ such as foo2. At it's core, reflink is just an optimised
file copy mechanism - the resultant copy should have the same
behaviour as a file copied by read/write. Copies done by physically
copying data do not duplicate fallocate() regions or guarantees from
the source file to the destination file.

time btrfs doesn't account for the doubling.  It's actually much
better in this case to just use a hole because neither foo or foo2 can
use the preallocated space until the 1g is fully unshared.

Right - this implies unwritten extents should not be shared by
reflink, instead either skipped (i.e. leave as a hole in foo2 as you
suggest) or duplicated so that the next write to the region of foo2
will also succeed. I'd suggest that COPY_FALLOC (or whatever it'll
get called) implies the latter behaviour, the default behaviour
being the former...

When we're doing writes, it'll check the preallocated extents for extra
refs and force COW if any exist.  So writes into a preallocated region
can enospc.

This really seems like an btrfs interpretation/implementation
issue, not a problem for reflink in general.

Cheers,

Dave.
-- 
Dave Chinner
david@fromorbit.com

On Sat, Oct 17, 2015 at 08:44:35AM +1100, Dave Chinner wrote:

quoted

When we're doing writes, it'll check the preallocated extents for extra
refs and force COW if any exist.  So writes into a preallocated region
can enospc.

This really seems like an btrfs interpretation/implementation
issue, not a problem for reflink in general.

Right, now matter how we do it there are tradeoffs, and this one seemed
the least surprising to me.  I don't think it's a big problem at all.

Automatically replacing preallocated extents with holes during clone
seems like a better compromise though (at least for btrfs anyway).

-chris