Re: [PATCH v2 3/3] x86: Support local_flush_tlb_kernel_range

[Seth Jennings <sjenning@linux.vnet.ibm.com>]

On 06/15/2012 11:35 AM, Dan Magenheimer wrote:

>> From: Seth Jennings [mailto:sjenning@linux.vnet.ibm.com]
>> Sent: Friday, June 15, 2012 9:13 AM
>> To: Peter Zijlstra
>> Cc: Minchan Kim; Greg Kroah-Hartman; Nitin Gupta; Dan Magenheimer; linux-kernel@vger.kernel.org;
>> linux-mm@kvack.org; Thomas Gleixner; Ingo Molnar; Tejun Heo; David Howells; x86@kernel.org; Nick
>> Piggin
>> Subject: Re: [PATCH v2 3/3] x86: Support local_flush_tlb_kernel_range
>>
>> On 05/17/2012 09:51 AM, Peter Zijlstra wrote:
>>
>>> On Thu, 2012-05-17 at 17:11 +0900, Minchan Kim wrote:
>>>>> +++ b/arch/x86/include/asm/tlbflush.h
>>>>> @@ -172,4 +172,16 @@ static inline void flush_tlb_kernel_range(unsigned long start,
>>>>>       flush_tlb_all();
>>>>>  }
>>>>>
>>>>> +static inline void local_flush_tlb_kernel_range(unsigned long start,
>>>>> +             unsigned long end)
>>>>> +{
>>>>> +     if (cpu_has_invlpg) {
>>>>> +             while (start < end) {
>>>>> +                     __flush_tlb_single(start);
>>>>> +                     start += PAGE_SIZE;
>>>>> +             }
>>>>> +     } else
>>>>> +             local_flush_tlb();
>>>>> +}
>>>
>>> It would be much better if you wait for Alex Shi's patch to mature.
>>> doing the invlpg thing for ranges is not an unconditional win.
>>
>> From what I can tell Alex's patches have stalled.  The last post was v6
>> on 5/17 and there wasn't a single reply to them afaict.
>>
>> According to Alex's investigation of this "tipping point", it seems that
>> a good generic value is 8.  In other words, on most x86 hardware, it is
>> cheaper to flush up to 8 tlb entries one by one rather than doing a
>> complete flush.
>>
>> So we can do something like:
>>
>>      if (cpu_has_invlpg && (end - start)/PAGE_SIZE <= 8) {
>>              while (start < end) {
>>
>> Would this be acceptable?
> 
> Hey Seth, Nitin --
> 
> After more work digging around zsmalloc and zbud, I really think
> this TLB flushing, as well as the "page pair mapping" code can be
> completely eliminated IFF zsmalloc is limited to items PAGE_SIZE or
> less.


To add to what Nitin just sent, without the page mapping, zsmalloc and
the late xvmalloc have the same issue.  Say you have a whole class of
objects that are 3/4 of a page.  Without the mapping, you can't cross
non-contiguous page boundaries and you'll have 25% fragmentation in the
memory pool.  This is the whole point of zsmalloc.

--
Seth

--
To unsubscribe, send a message with 'unsubscribe linux-mm' in
the body to majordomo@kvack.org.  For more info on Linux MM,
see: http://www.linux-mm.org/ .
Don't email: <a href=mailto:"dont@kvack.org"> email@kvack.org </a>