Re: [RFC][PATCH 0/1] kmemleak: Fix false positive with alias

[Hiroshi DOYU <Hiroshi.DOYU@nokia.com>]

Hi Catalin,

From: ext Catalin Marinas <catalin.marinas@arm.com>
Subject: Re: [RFC][PATCH 0/1] kmemleak: Fix false positive with alias
Date: Fri, 17 Sep 2010 18:18:47 +0200

> On Tue, 2010-08-10 at 18:49 +0300, Hiroshi DOYU wrote:
>> Now there's not much difference with the attached patch, a new version
>> of alias.
>> 
>> / # modprobe kmemleak-special-test use_alias=0
>> / # time echo scan > /sys/kernel/debug/kmemleak
>> real    0m 2.30s
>> user    0m 0.00s
>> sys     0m 2.30s
>> 
>> / # modprobe kmemleak-special-test use_alias=1
>> / # time echo scan > /sys/kernel/debug/kmemleak
>> real    0m 3.91s
>> user    0m 0.00s
>> sys     0m 3.91s
> 
> So to understand - the first case is memory scanning without any aliases
> configured. The second case is the alias scanning using a separate
> prio_tree. The impact seems to be quite big.
> 
> But I wouldn't complicate the code with the callback mechanism,
> especially when loadable modules are considered. Is the pointer
> conversion always linear? Maybe we can just add an offset to the
> scan_area structure that is used for conversion rather than a callback.
>
> Another advantage of the linear offset would be that we can avoid the
> call for removing the conversion.
> 
> Is this feasible for your needs?

The formula is:

  new_value = virt_to_phys(original address) | each attributes;

Attribute bits must be ingored. So the conversion is:

  new_value &= ~each attributes;
  original address = phys_to_virt(new_value);

Could adding an offset to the scan_area solve this case?

> No point really in making it too
> generic if the simple offset would (hopefully) do.

I guess other iommu pagetable may be same?