Re: [PATCH v5 1/2] ring-buffer: Introducing ring-buffer mapping functions

[Steven Rostedt]

On Wed, 2 Aug 2023 13:30:56 +0100
Vincent Donnefort <vdonnefort@xxxxxxxxxx> wrote:

> On Wed, Aug 02, 2023 at 07:45:26AM -0400, Steven Rostedt wrote:
> > On Tue, 1 Aug 2023 13:26:03 -0400
> > Steven Rostedt <rostedt@xxxxxxxxxxx> wrote:
> >   
> > > > +
> > > > +	if (READ_ONCE(cpu_buffer->mapped)) {
> > > > +		/* Ensure the meta_page is ready */
> > > > +		smp_rmb();
> > > > +		WRITE_ONCE(cpu_buffer->meta_page->pages_touched,
> > > > +			   local_read(&cpu_buffer->pages_touched));
> > > > +	}    
> > > 
> > > I was thinking instead of doing this in the semi fast path, put this logic
> > > into the rb_wakeup_waiters() code. That is, if a task is mapped, we call
> > > the irq_work() to do this for us. It could even do more, like handle
> > > blocked mapped waiters.  
> > 
> > I was thinking how to implement this, and I worry that it may cause an irq
> > storm. Let's keep this (and the other locations) as is, where we do the
> > updates in place. Then we can look at seeing if it is possible to do it in
> > a delayed fashion another time.  
> 
> I actually looking at this. How about:
> 
> On the userspace side, a simple poll:
> 
>   static void wait_entries(int fd)
>   {
>           struct pollfd pollfd = {
>                   .fd     = fd,
>                   .events = POLLIN,
>           };
>   
>           if (poll(&pollfd, 1, -1) == -1)
>                   pdie("poll");
>   }
> 
> And on the kernel side, just a function to update the "writer fields" of the
> meta-page:
> 
>    static void rb_wake_up_waiters(struct irq_work *work)
>    {
>           struct rb_irq_work *rbwork = container_of(work, struct rb_irq_work, work);
>   +       struct ring_buffer_per_cpu *cpu_buffer =
>   +               container_of(rbwork, struct ring_buffer_per_cpu, irq_work);
>   +
>   +       rb_update_meta_page(cpu_buffer);
>    
>           wake_up_all(&rbwork->waiters);
> 
> That would rate limit the number of updates to the meta-page without any irq storm?
> 

Is poll an issue? It requires user space to do a system call to see if
there's more data? But I guess that's not too much of an issue, as it needs
to do the ioctl to get the reader page.

We could also add an option to the ioctl to block, or have the ioctl honor
the NON_BLOCK flags of the fd?

-- Steve