Re: [RFC][PATCH] irq_work

[Peter Zijlstra <peterz@infradead.org>]

On Thu, 2010-06-24 at 14:38 +0200, Andi Kleen wrote:
>  The sleepable 
> soft irq would have avoided that (that's not a show stopper) 

I'm still not convinced sleepable softirq is a workable thing.

Softirqs:
  A) are non-preemptible
  B) are per-cpu because of A
  C) can be ran from ksoftirqd context
  D) generic kernel infrastructure with identical semantics on all archs

If you were to make something like a sleepable softirq, you'd loose A
(per definition), B (sleepable implies migratable to cpus_allowed) and
possibly D (unless you want to touch all architectures).

Now from your 'requirements':

> I have one case that needs to sleep (but only when interrupting user code)

> TIF works for user space, but it's a bit ugly because it requires adding
> more data to the task_struct because CPUs can change.

Which I read as:

 1) needs to run in the task context of the task that got 'interrupted'
 2) needs to stay on the cpu it got interrupted on.

So C is out of the window too, at which point there's nothing resembling
softirqs left.

To boot, x86_64 runs softirqs from the hardirq stack:

/* Call softirq on interrupt stack. Interrupts are off. */
ENTRY(call_softirq)
        CFI_STARTPROC
        push %rbp
        CFI_ADJUST_CFA_OFFSET   8
        CFI_REL_OFFSET rbp,0
        mov  %rsp,%rbp
        CFI_DEF_CFA_REGISTER rbp
        incl PER_CPU_VAR(irq_count)
        cmove PER_CPU_VAR(irq_stack_ptr),%rsp
        push  %rbp                      # backlink for old unwinder
        call __do_softirq
        leaveq
        CFI_DEF_CFA_REGISTER    rsp
        CFI_ADJUST_CFA_OFFSET   -8
        decl PER_CPU_VAR(irq_count)
        ret 
        CFI_ENDPROC
END(call_softirq) 

Also, -rt has something that could be considered sleepable softirqs,
although we call them preemptible softirqs. It runs all softirqs from
cpu bound kthreads, which again doesn't match your requirements.

So no, I don't think your idea of sleepable softirqs is sound.