Re: [PATCH 4/8] x86: add support for remotely triggering the block softirq

[Ingo Molnar <mingo@elte.hu>]

* Jens Axboe <jens.axboe@oracle.com> wrote:

> > pick up the threaded softirq patches from -rt, those move all 
> > softirqs processing into kernel threads. I'd suggest to extend those 
> > via wakeup-from-remote functionality - it fits the construct quite 
> > naturally. You should also be able to directly observe any 
> > performance impact of threaded softirq handlers. (and if you find 
> > any, let me know so that we can make it faster :-)
> 
> I was just considering that, since I knew -rt moved the softirqs into 
> threads. I'll look into it, but may not post anything until after my 
> vacation.

we should more seriously investigate kernel thread scalability for 
another reason as well: besides -rt, any generic async IO facility we 
pick up will likely heavily rely on them. Kernel thread scheduling is 
quite a bit lighter than user task scheduling [no TLB flushes, etc.] - 
and if it is still not good enough we could probably accelerate them 
some more. (and everyone will benefit) irq-context softirqs on the other 
hand are quite rigid and bring in many atomicity assumptions so they are 
not as natural to program for.

	Ingo