Re: [PATCH] turn off writable page tables

[Andrew Theurer <habanero@us.ibm.com>]

>> fork a quite linear from small number to large number of dirty pages.
>> Below are the min and max:
>>
>>          1280 pages    128000 pages
>> wtpt:     813 usec      37552 usec
>> emulate: 3279 usec     283879 usec
>>     
>
> Good, at least that suggests that the code works for the usage it was
> intended for. 
>
>   
>> So, in a -perfect-world- this works great.  Problem is most workloads
>> don't appear to have a vast percentage of entries that need to be
>> updated.   I'll go ahead and  expand this test to find out what the
>> threshold is to break even.  I'll also see if we can implement a
>>     
> batched
>   
>> call in fork to update the parent -I hope this will show just as good
>> performance even when most entries need modification and even better
>> performance over wtpt with a low number of entries modified.
>>     
>
> With license to make more invasive changes to core Linux mm it certainly
> should be possible to optimize this specific case with a batched update
> fairly easily. You could even go further an implement a 'make all PTEs
> in pagetable RO' hypercall, possibly including a copy to the child. This
> could potentially work better than current 'late pin', at least the
> validation would be incremental rather than in one big hit at the end. 
>
> Ian
>   
FWIW, I found the threshold for emulate vs wtpt.  I ran the fork test 
with a set number of pages dirtied such that we had x number of PTEs per 
pte_page.

writable-pt
-----------
#pte usec
002 5242
004 5251
006 5373
008 5519
010 5873

emulate
--------
#pte usec
002 4922
004 5265
006 6074
008 6991
010 7806
012 5988

So, the threshold appears to be around 4 PTEs/page.  I was a little 
shocked at first how low this number is, but considering the near 
identical performance with the various workloads, this make sense.  All 
of the workloads had the vast majority of writable pages flushed with 
just 2 PTEs/page changed and a handful with more PTEs/page changed.  It 
would not surprise me if the overall average was around 4 PTEs/page.

I am having a hard time finding any "enterprise" workloads which have a 
lot of PTEs/page right before fork.  If anyone can point me to some, 
that would be great.

I will look into batching next, but I am curious if simply using a 
hypercall in stead of write fault + emulate will make any difference at 
all.  I'll try that first, then implement the batched update. 

Eventually a hypercall which does more would be nice, but I guess we'll 
have to convince the Linux maintainers it's a good idea.

-Andrew