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 07:35 PM, Rusty Russell wrote:

> On Thu, 24 May 2012 16:48:37 +0800, Alex Shi <alex.shi@intel.com> wrote:
>> On 05/24/2012 04:42 PM, Peter Zijlstra wrote:
>>
>>> On Thu, 2012-05-24 at 16:32 +0800, Alex Shi wrote:
>>>> So, I use cpumask_t in stack.
>>>
>>> cpumask_t is 512 bytes with NR_CPUS=4096, that's generally considered
>>> too big to be on stack.
>>>
>>> A number of people spend a lot of time removing cpumask_t from stacks a
>>> while ago, I'm very sure they'll not be happy if you're going to add it
>>> back.
>>
>>
>> In my testing, allocate a cpumask_var_t is more worse than cpumask_t.
>> So, another choice is using percpu pre-allocatd cpumask for this, but I
>> am wondering if it is acceptable. What's suggestion for this point?
> 
> Thanks for the ping Peter!
> 
> Please don't use cpus_ operations: they're deprecated.  Use cpumask_.
> Similarly, avoid cpumask_t.


Thanks, Rusty and Peter!

> 
> And yes, if you configure for thousands of CPUs, it's not free!  If it's
> a significant, you will want to use a per-cpu cpumask_var_t.


I see.

> 
> My other thought: your patch seems optimal as far as avoiding IPIs goes,
> but I wonder how often it folds down to a single CPU?  That case is
> easier to fast-path without using a new cpumask.


It will be quite often if threads number of user level APP is more than
a half of LCPUs.
It needs a new cpumask because we can not remove SMT bit on
mm->cpu_vm_mask_var directly.

> 
> Cheers,
> Rusty.