Re: Linux/x86's _PAGE_SWP_SOFT_DIRTY definition

[David Vrabel <david.vrabel@citrix.com>]

On 21/08/13 12:58, Konrad Rzeszutek Wilk wrote:
> David Vrabel <david.vrabel@citrix.com> wrote:
>> On 21/08/13 08:42, Jan Beulich wrote:
>>> All,
>>>
>>> was anyone of you involved in the recent (rc5->rc6) changes here?
>>> I'm asking because this new definition conflicts with _PAGE_PAT,
>>> which is unused only for native Linux (and I continue to not really
>>> understand their motivation to restrict themselves to just the four
>>> most trivial memory types).
>>
>> I was not aware of it and that just looks broken -- not just Xen but it
>> looks like it wouldn't work with (transparent) huge pages either.
>>
>> The soft dirty tracking was introduced (in 3.11-rc1) by 0f8975ec4 (mm:
>> soft-dirty bits for user memory changes tracking) and the problematic
>> patch adding the conflicting PTE bit is 179ef71cb (mm: save soft-dirty
>> bits on swapped pages).
>>
>> David
> 
> I am going to be in meetings most of today. David or Jan  would you
> be OK emailing the folks who came up with the patch and the
> committeer to mention that it causes a regression?
> 
> And naturally test it first with a upstream kernel?  I presume the
> regressions is in the form of pages of WB becoming WC and suddenly
> applications failing oddly?

It's not clear how a test case to show the regression can be reliably
produced in a limited time.  The failures from using WC instead of WB
will be pretty subtle.

David