Re: [PATCH -tip 4/5] kprobes/x86: Use text_poke_smp_batch

[Masami Hiramatsu <mhiramat@redhat.com>]

Mathieu Desnoyers wrote:
> * Masami Hiramatsu (mhiramat@redhat.com) wrote:
>> Mathieu Desnoyers wrote:
>>> * Masami Hiramatsu (mhiramat@redhat.com) wrote:
>>>> Mathieu Desnoyers wrote:
>>>>> * Masami Hiramatsu (mhiramat@redhat.com) wrote:
>>>>>> Use text_poke_smp_batch() in optimization path for reducing
>>>>>> the number of stop_machine() issues.
>>>>>>
>>>>>> Signed-off-by: Masami Hiramatsu <mhiramat@redhat.com>
>>>>>> Cc: Ananth N Mavinakayanahalli <ananth@in.ibm.com>
>>>>>> Cc: Ingo Molnar <mingo@elte.hu>
>>>>>> Cc: Jim Keniston <jkenisto@us.ibm.com>
>>>>>> Cc: Jason Baron <jbaron@redhat.com>
>>>>>> Cc: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
>>>>>> ---
>>>>>>
>>>>>>  arch/x86/kernel/kprobes.c |   37 ++++++++++++++++++++++++++++++-------
>>>>>>  include/linux/kprobes.h   |    2 +-
>>>>>>  kernel/kprobes.c          |   13 +------------
>>>>>>  3 files changed, 32 insertions(+), 20 deletions(-)
>>>>>>
>>>>>> diff --git a/arch/x86/kernel/kprobes.c b/arch/x86/kernel/kprobes.c
>>>>>> index 345a4b1..63a5c24 100644
>>>>>> --- a/arch/x86/kernel/kprobes.c
>>>>>> +++ b/arch/x86/kernel/kprobes.c
>>>>>> @@ -1385,10 +1385,14 @@ int __kprobes arch_prepare_optimized_kprobe(struct optimized_kprobe *op)
>>>>>>  	return 0;
>>>>>>  }
>>>>>>  
>>>>>> -/* Replace a breakpoint (int3) with a relative jump.  */
>>>>>> -int __kprobes arch_optimize_kprobe(struct optimized_kprobe *op)
>>>>>> +#define MAX_OPTIMIZE_PROBES 256
>>>>>
>>>>> So what kind of interrupt latency does a 256-probes batch generate on the
>>>>> system ?  Are we talking about a few milliseconds, a few seconds ?
>>>>
>>>> From my experiment on kvm/4cpu, it took about 3 seconds in average.
>>>
>>> That's 3 seconds for multiple calls to stop_machine(). So we can expect
>>> latencies in the area of few microseconds for each call, right ?
>>
>> Theoretically yes.
>> But if we register more than 1000 probes at once, it's hard to do
>> anything except optimizing a while(more than 10 sec), because
>> it stops machine so frequently.
>>
>>>> With this patch, it went down to 30ms. (x100 faster :))
>>>
>>> This is beefing up the latency from few microseconds to 30ms. It sounds like a
>>> regression rather than a gain to me.
>>
>> If it is not acceptable, I can add a knob for control how many probes
>> optimize/unoptimize at once. Anyway, it is expectable latency (after
>> registering/unregistering probes) and it will be small if we put a few probes.
>> (30ms is the worst case)
>> And if you want, it can be disabled by sysctl.
> 
> I think we are starting to see the stop_machine() approach is really limiting
> our ability to do even relatively small amount of work without hurting
> responsiveness significantly.
> 
> What's the current showstopper with the breakpoint-bypass-ipi approach that
> solves this issue properly and makes this batching approach unnecessary ?

We still do not have any official answer from chip vendors.
As you know, basic implementation has been done.

Thank you,

-- 
Masami Hiramatsu
e-mail: mhiramat@redhat.com