Re: [PATCH v7 2/8] x86/flush_tlb: try flush_tlb_single one by one in flush_tlb_range

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

On 05/24/2012 04:12 PM, Jan Beulich wrote:

>>>> On 24.05.12 at 08:41, Alex Shi <alex.shi@intel.com> wrote:
>> On 05/23/2012 10:51 PM, Jan Beulich wrote:
>>> Unless there is an implicit assumption that 'start' and 'end' are on
>>> the same page (which I doubt, as then it would be pointless to
>>> add 'end' here), this one is definitely wrong - you'd either have
>>> to issue multiple MMUEXT_INVLPG_MULTI-s, or you'd have to
>>> also use MMUEXT_TLB_FLUSH_MULTI for the multi-page case.
>>
>> Thanks comments!
>> So, the following change should be more safe for PV?
>>
>> -       if (va == TLB_FLUSH_ALL) {
>> -               args->op.cmd = MMUEXT_TLB_FLUSH_MULTI;
>> -       } else {
>> -               args->op.cmd = MMUEXT_INVLPG_MULTI;
>> -               args->op.arg1.linear_addr = va;
>> -       }
>> +       args->op.cmd = MMUEXT_TLB_FLUSH_MULTI;
> 
> This would be safe ...
> 
>> +       if (start != TLB_FLUSH_ALL)
>> +               args->op.arg1.linear_addr = start;
> 
> ... and then this superfluous, but it'd result in an unconditional
> full TLB flush. When start and end (or perhaps end-1, assuming
> end is not inclusive) are on the same page, MMUEXT_INVLPG_MULTI
> should be used; MMUEXT_TLB_FLUSH_MULTI might need to be
> used in all other cases, unless you want to split multi-page, non-
> global invalidations into multiple MMUEXT_INVLPG_MULTI-s (which
> would appear to be what the whole patch aims at).


        args->op.cmd = MMUEXT_TLB_FLUSH_MULTI;
-       if (start != TLB_FLUSH_ALL)
+       if (start != TLB_FLUSH_ALL && (end - start) < PAGE_SIZE) {
+               args->op.cmd = MMUEXT_INVLPG_MULTI;
                args->op.arg1.linear_addr = start;
+       }

So, above it correct code for xen?
As to the xen optimisation of flush range, it is may better to be done
in a separate patch.

> 
> Perhaps the abstraction layer needs to be changed instead:
> Have the low level routines (Xen, UV, native) just deal with
> single pages, and do the splitting in common code (using the
> TLB size metrics).
> 
> But then again these metrics will become stale after a
> migration (not only on Xen, but in all virtualization scenarios), so
> some additional aspects will need to be taken care of anyway.


Sure, thanks for your care!