Re: [patch] Real-Time Preemption, -RT-2.6.10-rc2-mm3-V0.7.31-19

[Ingo Molnar <mingo@elte.hu>]

[-- Attachment #1: Type: text/plain, Size: 1925 bytes --]


* Ingo Molnar <mingo@elte.hu> wrote:

> i think i found the bug - it's an upstream ACPI bug. Does the patch
> below (or the -31-19 kernel, which i've just uploaded) fix the xruns?
>
> the upstream ACPI bug is this: we check need_resched() _before_
> disabling preemption. This opens up the following scenario:
> 
>  swapper:	!need_resched()
>  [IRQ context]
> 		wakes up a task
> 		marks idle task as need-resched
> 
>  swapper:	acpi_processor_idle(); // sleeps until next irq
> 
> instant 1msec latency introduced...
>
> normally default_idle() is safe because it re-checks need_resched with
> interrupts disabled before it truly halts the CPU. But
> acpi_processor_idle() doesnt seem to be doing this! Your trace clearly
> shows a missed preemption due to ACPI. I'm wondering why no-one has
> triggered this before, it's a really bad bug that should be fixed in
> 2.6.10.

i've attached the particular trace from Rui's collection of traces which
nicely shows the ACPI race in action. Here are the relevant events:

 swapper-0     08000000 0.840ms (+0.000ms): preempt_schedule (cpu_idle)
 swapper-0     ........ 0.913ms (+0.000ms): ->    IRQ 5-3052 [ 0000001d 0000008c ]: wake_up_process
 swapper-0     08000001 0.913ms (+0.920ms): acpi_processor_idle (cpu_idle)
 swapper-0     ........ 1.834ms (+0.000ms): common_interrupt:  [ e0022266 00000000 00000000 ]
 swapper-0     88010001 1.834ms (+0.000ms): direct_timer_interrupt (do_IRQ)
   IRQ 5-3052  80000002 1.848ms (+0.000ms): __switch_to (__sched_text_start)

i.e. at timestamp 0.840ms cpu_idle() got preempted (it was after the
need_resched() check but before the rcu_read_lock() which disables
preemption), then at 0.913ms a task was woken up by the hardirq context,
but despite this we called into acpi_processor_idle() which slept
dutifully until the next timer interrupt hit us. Then did only the woken
up task schedule. (which caused an xrun in Rui's setup.)

	Ingo

[-- Attachment #2: xruntrace1-2.6.10-rc2-mm3-RT-V0.7.31-15-20041201143614.trc.gz --]
[-- Type: application/x-gzip, Size: 30870 bytes --]