[PATCH] MIPS: Keep TLB cache hot while flushing

[PATCH] MIPS: Keep TLB cache hot while flushing

[Hillf Danton]

Hi David,

If we only flush the TLB of the given huge page, the TLB cache remains hot for
the relevant mm as it is, and less will be refilled after flush, huge or not.

As always all comments and ideas welcome.

Thanks

Signed-off-by: Hillf Danton <dhillf@gmail.com>
---

--- a/arch/mips/include/asm/hugetlb.h	Sat May 14 15:21:01 2011
+++ b/arch/mips/include/asm/hugetlb.h	Fri Oct 28 22:08:05 2011
@@ -70,7 +70,7 @@ static inline pte_t huge_ptep_get_and_cl
 static inline void huge_ptep_clear_flush(struct vm_area_struct *vma,
 					 unsigned long addr, pte_t *ptep)
 {
-	flush_tlb_mm(vma->vm_mm);
+	flush_tlb_page(vma, addr & huge_page_mask(hstate_vma(vma)));
 }

 static inline int huge_pte_none(pte_t pte)

Re: [PATCH] MIPS: Keep TLB cache hot while flushing

[David Daney]

On 10/28/2011 07:15 AM, Hillf Danton wrote:
> Hi David,
>
> If we only flush the TLB of the given huge page, the TLB cache remains hot for
> the relevant mm as it is, and less will be refilled after flush, huge or not.
>
> As always all comments and ideas welcome.
>

I haven't tested it, but it looks correct.  When I wrote the original 
flush_tlb_mm(), I was in a hurry and was more concerned about 
maintaining TLB consistency, rather than performance.

> Thanks
>
> Signed-off-by: Hillf Danton<dhillf@gmail.com>

Acked-by: David Daney <david.daney@cavium.com>

> ---
>
> --- a/arch/mips/include/asm/hugetlb.h	Sat May 14 15:21:01 2011
> +++ b/arch/mips/include/asm/hugetlb.h	Fri Oct 28 22:08:05 2011
> @@ -70,7 +70,7 @@ static inline pte_t huge_ptep_get_and_cl
>   static inline void huge_ptep_clear_flush(struct vm_area_struct *vma,
>   					 unsigned long addr, pte_t *ptep)
>   {
> -	flush_tlb_mm(vma->vm_mm);
> +	flush_tlb_page(vma, addr&  huge_page_mask(hstate_vma(vma)));
>   }
>
>   static inline int huge_pte_none(pte_t pte)
>

Re: [PATCH] MIPS: Keep TLB cache hot while flushing

[Ralf Baechle]

On Fri, Oct 28, 2011 at 09:16:29AM -0700, David Daney wrote:

> >If we only flush the TLB of the given huge page, the TLB cache remains hot for
> >the relevant mm as it is, and less will be refilled after flush, huge or not.
> >
> >As always all comments and ideas welcome.
> >
> 
> I haven't tested it, but it looks correct.  When I wrote the
> original flush_tlb_mm(), I was in a hurry and was more concerned
> about maintaining TLB consistency, rather than performance.

Yes, looks sane and like an obvious performance improvment.  but as always
with TLB stuff I'm paranoid so I've applied this to my 3.3 queue for it to
be tested well.

Hillf, do you have any benchmark numbers for this?

Thanks folks!

  Ralf

Re: [PATCH] MIPS: Keep TLB cache hot while flushing

[Hillf Danton]

On Fri, Nov 4, 2011 at 10:56 PM, Ralf Baechle <ralf@linux-mips.org> wrote:
> On Fri, Oct 28, 2011 at 09:16:29AM -0700, David Daney wrote:
>
>> >If we only flush the TLB of the given huge page, the TLB cache remains hot for
>> >the relevant mm as it is, and less will be refilled after flush, huge or not.
>> >
>> >As always all comments and ideas welcome.
>> >
>>
>> I haven't tested it, but it looks correct.  When I wrote the
>> original flush_tlb_mm(), I was in a hurry and was more concerned
>> about maintaining TLB consistency, rather than performance.
>
> Yes, looks sane and like an obvious performance improvment.  but as always
> with TLB stuff I'm paranoid so I've applied this to my 3.3 queue for it to
> be tested well.
>

Thanks Ralf and David:)

> Hillf, do you have any benchmark numbers for this?
>

Simple application related to libhugetlbfs works as it was after this patch
applied. Unlike x86, MIPS box is rare resource, and I have to ask David, the
author of MIPS huge TLB, to benchmark this patch, and to test GUP on Cavium
chips as well.

Btw, the patch for THP is prepared, and I am chasing box for test. Would you
please, David, take a look at it before posting?

Thanks

Hillf