RFC: Allow block drivers to poll for I/O instead of sleeping

[Ingo Molnar]

* Linus Torvalds <torvalds at linux-foundation.org> wrote:

> On Sun, Jun 23, 2013 at 12:09 AM, Ingo Molnar <mingo at kernel.org> wrote:
> >
> > The spinning approach you add has the disadvantage of actively wasting 
> > CPU time, which could be used to run other tasks. In general it's much 
> > better to make sure the completion IRQs are rate-limited and just 
> > schedule. This (combined with a metric ton of fine details) is what 
> > the networking code does in essence, and they have no trouble reaching 
> > very high throughput.
> 
> It's not about throughput - it's about latency. Don't ever confuse the 
> two, they have almost nothing in common. Networking very very seldom has 
> the kind of "submit and wait for immediate result" issues that disk 
> reads do.

Yeah, indeed that's true, the dd measurement Matthew did issued IO at a 
rate of one sector at a time and waiting for every sector to complete:

    dd if=/dev/nvme0n1 of=/dev/null iflag=direct bs=512 count=1000000

So my suggestions about batching and IRQ rate control are immaterial...

> That said, I dislike the patch intensely. I do not think it's at all a 
> good idea to look at "need_resched" to say "I can spin now". You're 
> still wasting CPU cycles.
> 
> So Willy, please do *not* mix this up with the scheduler, or at least 
> not "need_resched". Instead, maybe we should introduce a notion of "if 
> we are switching to the idle thread, let's see if we can try to do some 
> IO synchronously".
> 
> You could try to do that either *in* the idle thread (which would take 
> the context switch overhead - maybe negating some of the advantages), or 
> alternatively hook into the scheduler idle logic before actually doing 
> the switch.
> 
> But anything that starts polling when there are other runnable processes 
> to be done sounds really debatable. Even if it's "only" 5us or so. 
> There's a lot of real work that could be done in 5us.

I'm wondering, how will this scheme work if the IO completion latency is a 
lot more than the 5 usecs in the testcase? What if it takes 20 usecs or 
100 usecs or more?

Will we still burn our CPU time, wasting power, inflating this CPU's load 
which keeps other CPUs from balancing tasks over to this CPU, etc?

In the 5 usecs case it looks beneficial to do. In the longer-latency cases 
I'm not so sure.

Thanks,

	Ingo