Re: Pending splice(file -> FIFO) excludes all other FIFO operations forever (was: ... always blocks read(FIFO), regardless of O_NONBLOCK on read side?)

[Christian Brauner]

On Thu, Jul 06, 2023 at 02:56:45PM -0700, Linus Torvalds wrote:
> On Mon, 26 Jun 2023 at 09:14, Ahelenia Ziemiańska
> <nabijaczleweli@xxxxxxxxxxxxxxxxxx> wrote:
> >
> > And even if that was a working work-around, the fundamental problem of
> > ./spl>fifo excluding all other access to fifo is quite unfortunate too.
> 
> So I already sent an earlier broken version of this patch to Ahelenia
> and Christian, but here's an actually slightly tested version with the
> obvious bugs fixed.
> 
> The keyword here being "obvious". It's probably broken in many
> non-obvious ways, but I'm sending it out in case anybody wants to play
> around with it.
> 
> It boots for me. It even changes behavior of programs that splice()
> and used to keep the pipe lock over the IO and no longer do.
> 
> But it might do unspeakable things to your pets, so caveat emptor. It
> "feels" right to me. But it's a rather quick hack, and really needs
> more eyes and more thought. I mention O_NDELAY in the (preliminary)
> commit message, but there are probably other issues that need thinking
> about.

Forgot to say, fwiw, I've been running this through the LTP splice,
pipe, and ipc tests without issues. A hanging reader can be signaled
away cleanly with this.