Re: sendfile compatibility

[Alan Hourihane]

On Thu, 2011-01-27 at 11:00 -0700, Eric Blake wrote:
> On 01/27/2011 02:20 AM, Alan Hourihane wrote:
> > Hello,
> > 
> > I'm wondering if gnulib is scheduled to provide sendfile compatibility
> > using lseek/read/write ?
> 
> Interesting.  The compile-time ease-of-use aspect makes it seem like a
> function worth porting into gnulib.  For Linux, it opens up the door to
> providing the read/write fallback in user space for the case where the
> kernel lacks support (that is, the kernel requires mmap-like input and
> socket output, but the gnulib version can support much more).
> Obviously, the gnulib replacement wouldn't have the efficiency on
> non-Linux platforms, but the mere usability aspect of the interface
> makes it seem worthwhile to me (you can write more compact code if you
> have this interface present, and don't have to worry about writing the
> read/write fallback yourself).
> 
> Would you like to take a shot at proposing an implementation for the
> module?  The gnulib "schedule" is basically that things are added when
> someone who cares about using something takes the time to submit it;
> while this interface sounds nice, it's not my current itch, so I
> probably won't be adding it myself (but I don't mind reviewing if
> someone else puts in the work).

O.k. sounds good. I'll take a stab and send out for review.

> > I know sendfile is a kernel routine, and I'm unsure of whether the file
> > pointer needs restoration after the operation or not in a user-space
> > replacement.
> 
> Hence, a good module proposal would include a testsuite that validates
> such corner-case behavior on the kernel implementation, and documents
> the fallback semantics as well (for example, be sure to test
> regular->regular, regular->pipe, pipe->regular, pipe->pipe).

Sure. I'll get a test suite for it too.

Thanks Eric.

Alan.