Re: [PATCH RT 00/02] SLOB optimizations

[Eric Dumazet <dada1@cosmosbay.com>]

Ingo Molnar a écrit :
> * Eric Dumazet <dada1@cosmosbay.com> wrote:
> 
>>> in any case, on sane platforms (i386, x86_64) an irq-disable is 
>>> well-optimized in hardware, and is just as fast as a preempt_disable().
>> I'm afraid its not the case on current hardware.
>>
>> The irq enable/disable pair count for more than 50% the cpu time spent 
>> in kmem_cache_alloc()/kmem_cache_free()/kfree()
> 
> because you are not using NMI based profiling?
> 
>> oprofile results on a dual Opteron 246 :
>>
>> You can see the high profile numbers right after cli and popf(sti) 
>> instructions, popf being VERY expensive.
> 
> that's just the profiling interrupt hitting them. You should not analyze 
> irq-safe code with a non-NMI profiling interrupt.
> 

I'm using oprofile on Opteron, and AFAIK it's NMI based.

# grep NMI /proc/interrupts ; sleep 1 ; grep NMI /proc/interrupts
NMI:  391352095 2867983903
NMI:  391359678 2867998498

thats 7583 and 14595 NMI / second on cpu0 and cpu1 respectivly in this sample.

> CLI/STI is extremely fast. (In fact in the -rt tree i'm using them 
> within mutexes instead of preempt_enable()/preempt_disable(), because 
> they are faster and generate less register side-effect.)
> 
> 	Ingo
> 
>