Re: some accept-process-output races fixed; Tramp FIXMEs

[Paul Eggert]

On 1/16/19 11:07 AM, Stefan Monnier wrote:
> I think we should make sure that
>
>      (if (not (process-live-p p))
>          (progn
>            (accept-process-output p)
>            ;; Here we should be sure that we processed all the output
>
> whereas IIUC currently we need
>
>      (if (not (process-live-p p))
>          (progn
>            (while (accept-process-output p))
>            ;; Here we are presumably sure that we processed all the output

Both those approaches should work now. In my earlier email I was worried 
about the slightly-different case where P exits just before 
(accept-process-output P DELAY) is called. In this case 
accept-process-output might not slurp all the data from P even if P has 
already exited. However, in the above scenarios, where Emacs "knows" 
that P has exited before (accept-process-output P) is called with no 
DELAY, no data should go missing on any sane POSIX platform (though 
admittedly I can't cite chapter and verse on this and I suspect POSIX 
does not strictly require it).

However, I'm still puzzled by Tramp's use of (while (or 
(accept-process-output p 0.1) (process-live-p p))) to get all P's 
output. Why not use (while (accept-process-output p)) instead, which is 
simpler and avoids the race condition I mentioned? Maybe something to do 
with timers? I'm at a loss there.