Re: [PATCH] Add copy_file_range system call

[Rich Felker <dalias@libc.org>]

On Mon, Aug 19, 2019 at 08:56:35PM -0400, Rich Felker wrote:
> On Mon, Aug 19, 2019 at 11:41:14PM +0000, Árni Dagur wrote:
> > This patch was based on commit 53147f9, which added splice and vmsplice.
> > ---
> >  The function signature in the glibc manpage specifies `loff_t` instead
> >  of `off_t`, for both `copy_file_range` and `splice`. In musl, however, 
> >  the function signature for `splice` specifies `off_t`, so I did the
> >  same here. I'm not an experienced C programmer, so that may have been
> >  wrong.
> 
> I think this looks ok. Regarding loff_t vs off_t, loff_t is the
> kernel's API type for functions that take a 64-bit offset
> unconditionally rather than glibc providing 32-bit off_t and 64-bit
> off_t versions of them. This is gratuitous for musl where off_t is
> always 64-bit. We provide loff_t as a macro that expands to off_t, but
> even if it were a typedef the types woule be the same type, so it's
> fine to use off_t here, and I think it's the cleanest and most
> consistent with what we're doing elsewhere even if it's not textually
> the same as the man page.
> 
> >  include/unistd.h            | 1 +
> >  src/linux/copy_file_range.c | 8 ++++++++
> >  2 files changed, 9 insertions(+)
> >  create mode 100644 src/linux/copy_file_range.c
> > 
> > diff --git a/include/unistd.h b/include/unistd.h
> > index 9485da7a..00cc7042 100644
> > --- a/include/unistd.h
> > +++ b/include/unistd.h
> > @@ -188,6 +188,7 @@ char *get_current_dir_name(void);
> >  int syncfs(int);
> >  int euidaccess(const char *, int);
> >  int eaccess(const char *, int);
> > +ssize_t copy_file_range(int fd_in, off_t *off_in, int fd_out, off_t *off_out, size_t len, unsigned flags);
> >  #endif
> 
> Is there a reason for the choice to put it in unistd.h? Similar
> functions seem to have gone in fcntl.h. unistd.h does not make the
> loff_t type available which could be problematic to callers using it,
> since they really should (for API compatibility) be declaring the
> objects whose addresses they pass as loff_t.
> 
> If glibc does it here and exposes loff_t in unistd.h we might need to
> consider doing that too with _GNU_SOURCE.

OK I went and looked at what glibc did (glibc commit
bad7a0c81f501fbbcc79af9eaa4b8254441c4a1f) and they define the function
with arguments having type __off64_t and declare it in unistd.h. So I
think the expectation is that you use off_t with it (or off64_t if
doing the LFS dance on glibc with _FILE_OFFSET_BITS==32), loff_t is
not needed as part of the API to invoke it, and your patch looks fine.

Rich