Re: [PATCHv3 02/13] xread: poll on non blocking fds

[Jeff King <peff@peff.net>]

On Tue, Sep 22, 2015 at 12:45:51PM -0700, Junio C Hamano wrote:

> Jacob Keller <jacob.keller@gmail.com> writes:
> 
> > I don't think this patch actually changes behavior as it stands now. I
> > think Junio's suggestion does. Personally, I'd prefer some sort of
> > warning when you use xread and get EAGAIN or EWOULDBLOCK. I'd rather
> > see it somehow warn so that we can find the bug (since we really
> > really shouldn't be calling xread with a blocking socket, especially
> > if we have xread_noblock or similar as in this series.
> 
> One caveat is that the caller may not know in the first place.
> 
> The last time I checked the existing callers of xread(), there were
> a few that read from a file descriptor they did not open themselves
> (e.g. unpack-objects that read from standard input).  The invoker of
> these processes is free to do O_NONBLOCK their input stream for
> whatever reason.

Yeah. I do not think this is a bug at all; the user might have their
reasons for handing off an O_NONBLOCK pipe. If we take xread() to mean
"try to read from fd until we get a real error, some data, or an EOF",
then it is perfectly reasonable to replace spinning on read() (which we
do now) with a poll() for efficiency. The caller (and the user) does not
have to care, and should not notice; the outcome will be the same.

-Peff