Re: BTRFS file clone support for cp

[Jim Meyering]

Joel Becker wrote:

> On Wed, Jul 29, 2009 at 07:14:37PM +0100, Pádraig Brady wrote:
>> Chris Mason wrote:
>> > On Wed, Jul 29, 2009 at 03:14:49PM +0100, Pádraig Brady wrote:
>> >>
>> >> We may need to play around with fallocate()
>> >> if we want to get back to the original
>> >> cp semantics of actually allocating space
>> >> on the file system for the new file.
>> >
>> > Well, best to just use the original cp code.  I was talking with
>> > Giuseppe about this as well, I think we should the option to do regular
>> > cp via a flag.
>>
>> Right. Well we can turn off this cloning by doing --sparse={never,always}
>> but that has side effects. If we need an option then maybe we should have
>> it turn on cloning rather than restore default cp behaviour?
>> The side effects I thought of earlier, of COW without corresponding 
>> allocation
>> were possible fragmentation on write or unexpected/mishandled ENOSPC.
>> Also for endangered mechanical disks, subsequent processing could
>> be slowed as the head seeks between the old and new data to be copied.
>> Perhaps these are a small price to pay, especially considering that
>> solid state disks will only be affected by the write()=ENOSPC issue.
>>
>> At the moment we have these linking options:
>>
>> cp -l, --link #for hardlinks
>> cp -s, --symbolic-link #for symlinks
>>
>> So perhaps we should support:
>>
>> cp --link={soft,hard,cow}
>> for symlink(), link() and reflink() respectively?
>> I.E. link to the name, inode or extents respectively.
>
>       I've cooked up 'ln -r' for reflinks, which works for ln(1) but
> not for cp(1).

Thanks.  I haven't looked, but after reading about the reflink syscall
[http://lwn.net/Articles/332802/] had come to the same conclusion:
this feature belongs with ln rather than with cp.

Besides, putting the new behavior on a new option avoids
the current semantic change we would otherwise induce in cp.