Sorry missed your previous mail

2014-03-25 16:19 GMT+01:00 Balaji T K <balaj...@ti.com>:
>>>
>>>>>>
>>>>>> This seems wrong to me.
>>>>>>
>>>>>> What if the host is suspended (runtime PM wise) and you want to
>>>>>> disable SDIO irq.
>>>>>> Won't SDIO irq be kept enabled in this case?
>
>
> First of all, In runtime suspended state is a corner case where sdio-irq
> got triggered while .runtime_suspend hook was getting executed
> in multi-processor system.

Just to make sure I got you right

2 cores,

1. device was idle for xx ms, transition to runtime suspend
   initiated, host spinlock is aquired
2. meanwhile an sdio-irq is detected but not yet served
3. 2nd core starts process mmc irq, decodes it as an sdio_irq, but
   not yet processed due to spinlock held by runtime_suspend
4. device irqs are masked, clocks are gated, no more register
   access possible, spinlock is released
5. sdio irq gets handled, and register access in
   omap_hsmmc_enable_sdio_irq fails with SIGBUS


- Sdio irq is not a dedicated irq line, but is encoded in irqstat. So yoiu need
  to read irqstat to know it's an sdio_irq. This is racy as well
- pm_runtime_suspend clears the IRQSTAT register, so the vulnerable
  window is very small.

The single core situation, I think:
1. Runtime_suspend holds the spinock irqsave so irqstat is not
    fetched early by the irq handler
2. irqstat is cleared by pm_runtime_suspend
3. after pm_runtime hook is complete, a pending the irq is handled
   delaying the gating of the clocks, so irqstat is read safely
4 there is never an sdio_irq after pm_suspend ran

In the multi-core case runtime_suspend / irq handler can run in parallel
- irqstat is read without spinlock, so it probably happen before
  gating the clocks
-- if irqstat happens to be already cleared, nothing happens
-- if sdio_irq shall be handled, a spin_lock is taken, which is held by
   runtime_suspend, clocks are probably gated soon after releasing it.

Your solution will probably work, but relying on reading irqstat before
clocks are gated is not really kosher, I think we should abort
runtime_suspend if irqstat is non-nulll and/or indicates an sdio_irq.

multi-core... N#r%gl
--
To unsubscribe from this list: send the line "unsubscribe linux-omap" in
the body of a message to majord...@vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html

Reply via email to