Re: [PATCH] fdutimensat: add an atflag parameter

[Eric Blake]

On 09/17/2010 12:14 PM, Paul Eggert wrote:
> Thanks for adding that functionality to gnulib.  I prepared a patch
> to have GNU tar use it, but found one more problem.  fdutimensat
> uses this signature:
>
> fdutimensat (int dir, char const *file, int fd, [ other args ] )
>
> whereas the similar function in GNU tar uses the following
> order instead.
>
> fdutimensat (int fd, int dir, char const *file, [ other args ] )
>
> The GNU tar convention is clearer, since it reflects the
> semantics better: try FD first, and use that if it works;
> otherwise fall back on DIR+FILE.  Could we please switch to
> this order, while we're changing the API anyway?

Hmm; I'm reluctant to make the change.  Look at the history of 
utimensat.h: originally, we declared futimens (fd, name, times), but 
that conflicted with the POSIX naming, so it was renamed gl_futimens 
(fd, name, times).  But when it came time to extend it to the *at 
calling conventions, all of our macros for quickly implementing openat 
on top of open, mkdirat on top of mkdir, and so forth, all expected that 
the wrapped function have name be first.  So I added fdutimens (name, 
fd, times), at which point I rewrote gl_futimens to be a thin wrapper 
around fdutimens that swapped arguments, and added utimensat using the 
*at implementation that had fdutimens with name first.

Now you're proposing making fdutimensat have the opposite signature from 
fdutimens (fd before name, instead of after), not to mention 
inconsistent with all the other *at functions (fd before dir, instead of 
dir first).  I'm leaning towards keeping the signature we have, and 
change fchmodat and friends to have the same style as the current 
fdutimensat (dir first, because it is a *at function, then name/fd in 
that order, because that is the order in which it is easiest to wrap a 
non-at function to turn it in an at function).

--
Eric Blake   address@hidden    +1-801-349-2682
Libvirt virtualization library http://libvirt.org