Re: Realtime Preemption, 2.6.12, Beginners Guide?

From: Ingo Molnar
Date: Wed Jul 06 2005 - 20:31:05 EST



* Alistair John Strachan <s0348365@xxxxxxxxxxxx> wrote:

> > I'm beginning to understand the issue, and I see why you think the proposed
> > patch fixes it. I'll compile and boot V0.7.51-05 now.
>
> Indeed, this seems to have fixed it.
>
> ( softirq-timer/0-3 |#0): new 8 us maximum-latency wakeup.
> ( softirq-timer/0-3 |#0): new 9 us maximum-latency wakeup.
> ( softirq-timer/0-3 |#0): new 9 us maximum-latency wakeup.
> ( softirq-timer/0-3 |#0): new 9 us maximum-latency wakeup.
> ( softirq-timer/0-3 |#0): new 10 us maximum-latency wakeup.
> ( softirq-timer/0-3 |#0): new 14 us maximum-latency wakeup.

great! Do the softlockup warnings still occur?

> Find attached another trace (only 33us this time).

the main latency comes from here:

> <...>-3485 0Dnh2 13us : enqueue_task (__schedule)
> <...>-3485 0Dnh2 14us+: trace_array (__schedule)
> <...>-3485 0Dnh2 18us : trace_array <softirq--3> (69 6e)
> <...>-3485 0Dnh2 18us : trace_array <<...>-3485> (76 78)
> <...>-3485 0Dnh2 20us+: trace_array (__schedule)
> softirq--3 0Dnh2 28us+: __switch_to (__schedule)

trace_array() can be quite expensive (it generates a trace entry of
every runnable task, with interrupts and preemption disabled). It is
disabled if RT_DEADLOCK_DETECT is disabled. For pure wakeup latency
tracing, the most optimal combination of options is:

# CONFIG_DETECT_SOFTLOCKUP is not set
# CONFIG_SCHEDSTATS is not set
# CONFIG_DEBUG_SLAB is not set
# CONFIG_DEBUG_PREEMPT is not set
# CONFIG_DEBUG_SPINLOCK_SLEEP is not set
CONFIG_WAKEUP_TIMING=y
# CONFIG_CRITICAL_PREEMPT_TIMING is not set
# CONFIG_CRITICAL_IRQSOFF_TIMING is not set
CONFIG_LATENCY_TIMING=y
CONFIG_LATENCY_TRACE=y
CONFIG_MCOUNT=y
# CONFIG_RT_DEADLOCK_DETECT is not set
# CONFIG_DEBUG_RT_LOCKING_MODE is not set
# CONFIG_DEBUG_KOBJECT is not set
CONFIG_DEBUG_BUGVERBOSE=y
# CONFIG_DEBUG_INFO is not set
# CONFIG_DEBUG_FS is not set
CONFIG_FRAME_POINTER=y
CONFIG_EARLY_PRINTK=y
# CONFIG_DEBUG_STACKOVERFLOW is not set
# CONFIG_KPROBES is not set
# CONFIG_DEBUG_STACK_USAGE is not set
# CONFIG_DEBUG_PAGEALLOC is not set

such a kernel will still be able to generate /proc/latency_trace traces,
but has much lower runtime overhead than your current kernel. (But you
should probably keep all debugging enabled until all of the current
problems have been resolved.)

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/