Re: [PATCH v7 8/8] x86/tlb: just do tlb flush on one of siblings of SMT

[Alex Shi <alex.shi@intel.com>]

On 05/24/2012 01:09 AM, Peter Zijlstra wrote:

> On Wed, 2012-05-23 at 16:05 +0100, Jan Beulich wrote:
>>>>> On 23.05.12 at 16:15, Alex Shi <alex.shi@intel.com> wrote:
>>> +	/* doing flush on both siblings of SMT is just wasting time */
>>> +	cpumask_copy(&flush_mask, cpumask);
>>> +	if (likely(smp_num_siblings > 1)) {
>>> +		rand = jiffies;
>>> +		/* See "Numerical Recipes in C", second edition, p. 284 */
>>> +		rand = rand * 1664525L + 1013904223L;
>>> +		rand &= 0x1;
>>> +
>>> +		for_each_cpu(cpu, &flush_mask) {
>>> +			sblmask = cpu_sibling_mask(cpu);
>>> +			if (cpumask_subset(sblmask, &flush_mask)) {
>>> +				if (rand == 0)
>>> +					cpu_clear(cpu, flush_mask);
>>> +				else
>>> +					cpu_clear(cpumask_next(cpu, sblmask),
>>> +								flush_mask);
>>> +			}
>>> +		}
>>> +	}
>>> +
>>
>> There is no comment or anything else indicating that this is
>> suitable for dual-thread CPUs only - when there are more than
>> 2 threads per core, the intended effect won't be achieved.
> 
> Why would that be? Won't higher thread count still share the same
> resources just more so?
> 
>>  I'd
>> recommend making the logic generic from the beginning, but if
>> that doesn't seem feasible to you, at least a comment stating
>> the limitation should be added imo.


Sure. but just want to know how many commercial x86 CPU uses >2 SMTs?
Write a short, quick function to do random selection in SMT is quite
complicate considering cpumask maybe just contain random number SMT
siblings in a core.

> 
> My objection to the whole lot is that its looks mightily expensive on
> large machines, cpumask operations aren't cheap when you've got 4k cpus
> etc..
> 
> Also, you very much cannot put cpumask_t on stack.


Sure, and do you has related data for this?

I just measured the cost of this function on my Romely EP(32 LCPUs) with
cpumask_t and NR_CPUS = 32/256/512/4096, the cost are similar with
256/512/4096 and that increased about 20% time cost from 32.

I also tried to use cpumask_var_t and alloc it in heap(use
CPUMASK_OFFSTACK), actually, it cost same time with cpumask_t in stack.
But, the allocation bring another big cost. So, I use cpumask_t in stack.
The performance gain data in commit log is getting with NR_CPUS = 256.