Re: [RFC GIT PULL] softirq: Consolidation and stack overrun fix

[Frederic Weisbecker]

2013/9/20 Linus Torvalds <torvalds@xxxxxxxxxxxxxxxxxxxx>:
> On Fri, Sep 20, 2013 at 9:26 AM, Frederic Weisbecker <fweisbec@xxxxxxxxx> wrote:
>>
>> Now just for clarity, what do we then do with inline sofirq executions: on local_bh_enable()
>> for example, or explicit calls to do_softirq() other than irq exit?
>
> If we do a softirq because it was pending and we did a
> "local_bh_enable()" in normal code, we need a new stack. The
> "local_bh_enable()" may be pretty deep in the callchain on a normal
> process stack, so I think it would be safest to switch to a separate
> stack for softirq handling.

Right.

>
> So you have a few different cases:
>
>  - irq_exit(). The irq stack is by definition empty (assuming
> itq_exit() is done on the irq stack), so doing softirq in that context
> should be fine. However, that assumes that if we get *another*
> interrupt, then we'll switch stacks again, so this does mean that we
> need two irq stacks. No, irq's don't nest, but if we run softirq on
> the first irq stack, the other irq *can* nest that softirq.

Well, most archs don't define __ARCH_IRQ_EXIT_IRQS_DISABLED. It doesn't
even mean that the majority of them actually run irq_exit() with irqs enabled in
practice. But there may be thoretically some where hardirqs can nest
without even the
help of softirqs.

So it's quite possible to run softirqs on a hardirq stack that is not empty.

Now certainly what needs to be fixed then is archs that don't have
__ARCH_IRQ_EXIT_IRQS_DISABLED
or archs that have any other significant opportunity to nest interrupt.

>
>  - process context doing local_bh_enable, and a bh became pending
> while it was disabled. See above: this needs a stack switch. Which
> stack to use is open, again assuming that a hardirq coming in will
> switch to yet another stack.

Right. Now if we do like Thomas suggested, we can have a common irq
stack that is big enough for hard and softirqs. After all there should
never be more than two or three nesting irq contexts:
hardirq->softirq->hardirq, softirq->hardirq, ...

At least if we put aside the unsane archs that can nest irqs somehow.
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/