bug#68504: [PATCH] Add copy-on-write support to scm_copy_file.

Hi,

Tomas Volf <[email protected]> skribis:

> On modern file-systems (BTRFS, ZFS) it is possible to copy a file using
> copy-on-write method.  For large files it has the advantage of being
> much faster and saving disk space (since identical extents are not
> duplicated).  This feature is stable and for example coreutils' `cp'
> does use it automatically (see --reflink).
>
> This commit adds support for this feature into our
> copy-file (scm_copy_file) procedure.  Same as `cp', it defaults to
> 'auto, meaning the copy-on-write is attempted, and in case of failure
> the regular copy is performed.
>
> No tests are provided, because the behavior depends on the system,
> underlying file-system and its configuration.  That makes it challenging
> to write a test for it.  Manual testing was performed instead:
>
>     $ btrfs filesystem du /tmp/cow*
>          Total   Exclusive  Set shared  Filename
>       36.00KiB    36.00KiB       0.00B  /tmp/cow
>
>     $ cat cow-test.scm
>     (copy-file "/tmp/cow" "/tmp/cow-unspecified")
>     (copy-file "/tmp/cow" "/tmp/cow-always" #:copy-on-write 'always)
>     (copy-file "/tmp/cow" "/tmp/cow-auto" #:copy-on-write 'auto)
>     (copy-file "/tmp/cow" "/tmp/cow-never" #:copy-on-write 'never)
>     (copy-file "/tmp/cow" "/dev/shm/cow-unspecified")
>     (copy-file "/tmp/cow" "/dev/shm/cow-auto" #:copy-on-write 'auto)
>     (copy-file "/tmp/cow" "/dev/shm/cow-never" #:copy-on-write 'never)
>     $ ./meta/guile -s cow-test.scm
>
>     $ btrfs filesystem du /tmp/cow*
>          Total   Exclusive  Set shared  Filename
>       36.00KiB       0.00B    36.00KiB  /tmp/cow
>       36.00KiB       0.00B    36.00KiB  /tmp/cow-always
>       36.00KiB       0.00B    36.00KiB  /tmp/cow-auto
>       36.00KiB    36.00KiB       0.00B  /tmp/cow-never
>       36.00KiB       0.00B    36.00KiB  /tmp/cow-unspecified
>
>     $ sha1sum /tmp/cow* /dev/shm/cow*
>     4c665f87b5dc2e7d26279c4b48968d085e1ace32  /tmp/cow
>     4c665f87b5dc2e7d26279c4b48968d085e1ace32  /tmp/cow-always
>     4c665f87b5dc2e7d26279c4b48968d085e1ace32  /tmp/cow-auto
>     4c665f87b5dc2e7d26279c4b48968d085e1ace32  /tmp/cow-never
>     4c665f87b5dc2e7d26279c4b48968d085e1ace32  /tmp/cow-unspecified
>     4c665f87b5dc2e7d26279c4b48968d085e1ace32  /dev/shm/cow-auto
>     4c665f87b5dc2e7d26279c4b48968d085e1ace32  /dev/shm/cow-never
>     4c665f87b5dc2e7d26279c4b48968d085e1ace32  /dev/shm/cow-unspecified
>
> This commit also adds to new failure modes for (copy-file).
>
> Failure to copy-on-write when 'always was passed in:
>
>     scheme@(guile-user)> (copy-file "/tmp/cow" "/dev/shm/cow" #:copy-on-write 'always)
>     ice-9/boot-9.scm:1676:22: In procedure raise-exception:
>     In procedure copy-file: copy-on-write failed: Invalid cross-device link
>
> Passing in invalid value for the #:copy-on-write keyword argument:
>
>     scheme@(guile-user)> (copy-file "/tmp/cow" "/dev/shm/cow" #:copy-on-write 'nevr)
>     ice-9/boot-9.scm:1676:22: In procedure raise-exception:
>     In procedure copy-file: invalid value for #:copy-on-write: nevr
>
> * NEWS: Add note for copy-file supporting copy-on-write.
> * configure.ac: Check for linux/fs.h.
> * doc/ref/posix.texi (File System)[copy-file]: Document the new
> signature.
> * libguile/filesys.c (clone_file): New function cloning a file using
> FICLONE, if supported.
> (k_copy_on_write): New keyword.
> (sym_always, sym_auto, sym_never): New symbols.
> (scm_copy_file): New #:copy-on-write keyword argument.  Attempt
> copy-on-write copy by default.
> * libguile/filesys.h: Update signature for scm_copy_file.

The patch looks great (and very useful) to me, modulo one issue:

> -SCM_API SCM scm_copy_file (SCM oldfile, SCM newfile);
> +SCM_API SCM scm_copy_file (SCM oldfile, SCM newfile, SCM rest);

Since this is a public interface, we cannot change this function’s
signature during the 3.0 stable series.

Thus, I would suggest keeping the public ‘scm_copy_file’ unchanged and
internally having a three-argument variant.  The Scheme-level
‘copy-file’ would map to that three-argument variant.  (See how
‘scm_pipe’ and ‘scm accept’ as examples.)

Could you send an updated patch?

BTW, copyright assignment to the FSF is now optional but encouraged.
Please see
<https://lists.gnu.org/archive/html/guile-devel/2022-10/msg00008.html>.

Thanks,
Ludo’.