unnecessary tlb flush in mprotect

unnecessary tlb flush in mprotect

[Alex Shi]

when mprotect just change prots of non-present pages, current mprotect
still do the tlb flush in check_protection().
but according to 4.10.2.3 Intel SDM V3A (
www.intel.com/Assets/ja_JP/PDF/manual/253668.pdf ) at that time, TLB has
no this lines for this page. So, tlb flush is just waste time. (for cr3
rewrite, flush all tlb, or invlpg, like a 'nop' in intel cpu)

Do we need to add the pte_present similar check here to prevent the
unnecessary tlb flushing? I mean, are there real case in word, User like
to change page prots before assign a physical page to it?



Alex Shi

Re: unnecessary tlb flush in mprotect

[Alex Shi]

On 05/21/2012 04:30 PM, Alex Shi wrote:

> when mprotect just change prots of non-present pages, current mprotect
> still do the tlb flush in check_protection().
> but according to 4.10.2.3 Intel SDM V3A (
> www.intel.com/Assets/ja_JP/PDF/manual/253668.pdf ) at that time, TLB has
> no this lines for this page. So, tlb flush is just waste time. (for cr3
> rewrite, flush all tlb, or invlpg, like a 'nop' in intel cpu)
> 
> Do we need to add the pte_present similar check here to prevent the
> unnecessary tlb flushing? I mean, are there real case in word, User like
> to change page prots before assign a physical page to it?
> 


Any comments from linux-mm emailing list?

> 
> 
> Alex Shi

Re: unnecessary tlb flush in mprotect

[Andi Kleen]

On Tue, May 22, 2012 at 09:08:47AM +0800, Alex Shi wrote:
> On 05/21/2012 04:30 PM, Alex Shi wrote:
> 
> > when mprotect just change prots of non-present pages, current mprotect
> > still do the tlb flush in check_protection().
> > but according to 4.10.2.3 Intel SDM V3A (
> > www.intel.com/Assets/ja_JP/PDF/manual/253668.pdf ) at that time, TLB has
> > no this lines for this page. So, tlb flush is just waste time. (for cr3
> > rewrite, flush all tlb, or invlpg, like a 'nop' in intel cpu)
> > 
> > Do we need to add the pte_present similar check here to prevent the
> > unnecessary tlb flushing? I mean, are there real case in word, User like
> > to change page prots before assign a physical page to it?
> > 
> 
> 
> Any comments from linux-mm emailing list?

I would be careful with TLB optimizations if it's not a clear performance
wins. A lot of these interactions are tricky and it's very easy to break
things in subtle and hard to debug ways.

-Andi
-- 
ak@linux.intel.com -- Speaking for myself only

Re: unnecessary tlb flush in mprotect

[Alex Shi]

On 05/23/2012 12:37 AM, Andi Kleen wrote:

> On Tue, May 22, 2012 at 09:08:47AM +0800, Alex Shi wrote:
>> On 05/21/2012 04:30 PM, Alex Shi wrote:
>>
>>> when mprotect just change prots of non-present pages, current mprotect
>>> still do the tlb flush in check_protection().
>>> but according to 4.10.2.3 Intel SDM V3A (
>>> www.intel.com/Assets/ja_JP/PDF/manual/253668.pdf ) at that time, TLB has
>>> no this lines for this page. So, tlb flush is just waste time. (for cr3
>>> rewrite, flush all tlb, or invlpg, like a 'nop' in intel cpu)
>>>
>>> Do we need to add the pte_present similar check here to prevent the
>>> unnecessary tlb flushing? I mean, are there real case in word, User like
>>> to change page prots before assign a physical page to it?
>>>
>>
>>
>> Any comments from linux-mm emailing list?
> 
> I would be careful with TLB optimizations if it's not a clear performance
> wins. A lot of these interactions are tricky and it's very easy to break
> things in subtle and hard to debug ways.
> 
> -Andi


Sure, functionality is fundamental.

Re: unnecessary tlb flush in mprotect

[Hugh Dickins]

On Wed, 23 May 2012, Alex Shi wrote:
> On 05/23/2012 12:37 AM, Andi Kleen wrote:
> > On Tue, May 22, 2012 at 09:08:47AM +0800, Alex Shi wrote:
> >> On 05/21/2012 04:30 PM, Alex Shi wrote:
> >>
> >>> when mprotect just change prots of non-present pages, current mprotect
> >>> still do the tlb flush in check_protection().
> >>> but according to 4.10.2.3 Intel SDM V3A (
> >>> www.intel.com/Assets/ja_JP/PDF/manual/253668.pdf ) at that time, TLB has
> >>> no this lines for this page. So, tlb flush is just waste time. (for cr3
> >>> rewrite, flush all tlb, or invlpg, like a 'nop' in intel cpu)

Yes.

> >>>
> >>> Do we need to add the pte_present similar check here to prevent the
> >>> unnecessary tlb flushing? I mean, are there real case in word, User like
> >>> to change page prots before assign a physical page to it?

We have always simply assumed that mprotecting an entirely empty range
is too uncommon to bother with the slight complication of optimizing it.
Our assumption might be very wrong.

Hugh

> >>>
> >>
> >>
> >> Any comments from linux-mm emailing list?
> > 
> > I would be careful with TLB optimizations if it's not a clear performance
> > wins. A lot of these interactions are tricky and it's very easy to break
> > things in subtle and hard to debug ways.
> > 
> > -Andi
> 
> 
> Sure, functionality is fundamental.