Re: [PATCH 1/2] pipe: change pipe_write() to never add a zero-sized buffer

[Oleg Nesterov]

On 02/10, Linus Torvalds wrote:
>
> On Mon, 10 Feb 2025 at 09:22, Oleg Nesterov <oleg@xxxxxxxxxx> wrote:
> >
> > +               int avail = PAGE_SIZE - offset;
> >
> > -               if ((buf->flags & PIPE_BUF_FLAG_CAN_MERGE) &&
> > -                   offset + chars <= PAGE_SIZE) {
> > +               if (avail && (buf->flags & PIPE_BUF_FLAG_CAN_MERGE)) {
> >                         ret = pipe_buf_confirm(pipe, buf);
> >                         if (ret)
> >                                 goto out;
> >
> > +                       chars = min_t(ssize_t, chars, avail);
>
> If I read this correctly, this patch is horribly broken.
>
> You can't do partial writes. Pipes have one very core atomicity
> guarantee: from the man-pages:
>
>    PIPE_BUF
>        POSIX.1 says that writes of less than PIPE_BUF bytes must be
>        atomic:

Ah, I didn't know!

Thanks for your explanation and the quick NACK!

> Maybe we should add a comment somewhere about this.

Agreed. It would certainly help the ignorant readers like me ;)

So I guess that the "goto again" logic in sender/receiver in
tools/perf/bench/sched-messaging.c is currently pointless,
DATASIZE == 100 < PIPE_BUF.

Thanks.

Oleg.