[RFC PATCH 04/14] pipe: Add O_NOTIFICATION_PIPE [ver #2]

[David Howells]

Andy Lutomirski <luto at kernel.org> wrote:

> I can open a normal pipe from userspace (with pipe() or pipe2()), and
> I can have two threads.  One thread writes to the pipe with write().
> The other thread writes with splice().  Everything works fine.

Yes.  Every operation you do on a pipe from userspace is serialised with the
pipe mutex - and both ends share the same pipe.

> What's special about notifications?

The post_notification() cannot take the pipe mutex.  It has to be callable
from softirq context.  Linus's idea is that when you're actually altering the
ring pointers you should hold the wake-queue spinlock, and post_notification()
holds the wake queue spinlock for the duration of the operation.

This means that post_notification() can be writing to the pipe whilst a
userspace-invoked operation is holding the pipe mutex and is also doing
something to the ring.

David