Re: Feedback for fastselect and one-copy-pipe

[Manfred Spraul]

From: <[e-mail]>
> Hello!
>
> > * davem's patch breaks apps that assume that write(,PIPE_BUF) after
> > poll(POLLOUT) never blocks, even for blocking pipes.
>
> Pardon, but PIPE_BUF <= PAGE_SIZE yet, so that fears have no reasons.
>

The difference is the =

> <<<<< davem's patch
> +       if (count >= PAGE_SIZE &&
>                       ^^
> +           !(filp->f_flags & O_NONBLOCK)) {
> <<<<<<< my patch
> +  if (count > PIPE_BUF && chars == PIPE_SIZE &&
                     ^
> +      (!(filp->f_flags & O_NONBLOCK))) {
> <<<<<<<

davem used >=, I used >. All other differences between our patches are
code cleanups.

Just try this on i386: (PIPE_BUF is defined to 4096 on i386 - I really
don't understand why, but now it's too late to reverse it back to 512)

<<<<
char buf[PIPE_BUF];
void main()
{
    int pipes[2];
    pipe(pipes);
    write(pipes[1],buf,sizeof(buf));
}
<<<<<<<

It returns immediately on all unix platforms I tested, including all
linux versions, except with davem's patch.
It's not guaranteed in sus or posix, but I'm reluctant to change it.

--
    Manfred