Re: lockdep and threaded IRQs (was: ...)

[Ingo Molnar]

* Peter Zijlstra <a.p.zijlstra@xxxxxxxxx> wrote:

> On Mon, 2009-03-02 at 18:37 -0800, David Brownell wrote:
> > No.  But I did get a non-response that didn't include any
> > explanation, and relied totally on unfounded assertions
> > combined with the presumption that someday IRQF_DISABLED
> > will be forced on in all drivers.
> 
> Enabling IRQs in hardirq context is BAD because:
> 
>  - IRQ handler nesting leads to stack overflow
>  - It gives the false impression its OK for IRQ handlers to be slow,
>    it is _NOT_, as you still generate horrible preemption latency.
> 
> Therefore IRQF_DISABLED _will_ be forced on everybody some day 
> soon, and I'll provide an IRQF_ENABLED for use by broken 
> hardware only (and make a TAINT flag for that too).

Basically the problem why !IRQF_DISABLED is bad that if there 
are enough interrupt handlers we can get nesting like this:

  <irq 20>
  <handler runs with irqs enabled>
      <irq 21>
      <handler runs with irqs enabled>
          <irq 22>
          <handler runs with irqs enabled>
              <irq 23>
              <handler runs with irqs enabled>
                  <irq 24>
                  <handler runs with irqs enabled>

Suppose each handler gets interruped while it already used up 
1000 bytes of the stack (conservative estimation - often it's 
more) - the above sequence is already 5000 bytes into the stack.

There is no protection against stack overflow there and such 
bugs can be _very_ hard to trigger and find. If there's a 
sufficient number of devices and a high enough load it can 
trigger spuriously.

Yes, in a few limited embedded environments where you dont have 
more than 3-4 IRQ sources you might decide that it's safe to do 
(or you might decide that you dont care). Also, there's a few 
legacy pieces of hardware with either very long hw access 
latencies or too short buffers. Plus there might be any number 
of other hw factors - or architecture details (such as the use 
of separate per IRQ stacks) that limit IRQ handler parallelism 
in practice.

So we'll have the quirk flag for the weird cases - but these are 
the exceptions that strengthen the general rule. The concept of 
enabling interrupts in a hardirq handler is a no-no on a general 
purpose kernel and no modern driver should make use of it.

I hope this explains why lockdep never supported this case.

	Ingo
--
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/