Re: [PATCH] x86: {reverve,free}_memtype() take a physical address

[Rene Herman <rene.herman@keyaccess.nl>]

On 22-08-08 00:57, Pallipadi, Venkatesh wrote:
> 
>> -----Original Message-----
>> From: Rene Herman [mailto:rene.herman@keyaccess.nl]
>> Sent: Thursday, August 21, 2008 3:27 PM
>> To: Pallipadi, Venkatesh
>> Cc: Ingo Molnar; Li, Shaohua; Dave Airlie; Yinghai Lu; Andreas
>> Herrmann; Arjan van de Ven; Linux Kernel; Siddha, Suresh B;
>> Thomas Gleixner; H. Peter Anvin; Dave Jones
>> Subject: Re: [PATCH] x86: {reverve,free}_memtype() take a
>> physical address
>>
>> On 22-08-08 00:16, Pallipadi, Venkatesh wrote:
>>
>>> Yes. Noticed that too and sent a patch here for x86/tip.
>>>
>>> http://www.ussg.iu.edu/hypermail/linux/kernel/0808.2/2270.html
>>>
>>> It is not very critical as it sounds as only set_memory_uc sets PAT
>>> bits for RAM pages. Most other users (devmem mmap, ioramep, pci)
>>> set PAT bits on the reserved memory. And there will not be conflicts
>>> across RAM and reserveed regions. Regardless, this was a stupid
>>> bug that we had missed earlier.
>> And unfortunately I don't think the above fully fixes it for
>> AGP. __pa()
>> gets the real physical address and the memtypes should be on the GART
>> remapped physical addresses it seems.
>>
> 
> Page being marked here as uncached is the page got from alloc_page().
> We are not really marking GART physical address as uncacheable. And
> that page returned from alloc_page is what we are tracking with
> reserve and free.
> IOW, the tracking is only to keep CPU accesses consistent across
> different va->pa and va across different CPUs and has nothing to do
> with GART physical address here.

Okay, if you say so... it _used_ to be before this array change to AGP 
that the GART addresses were in the memtype list, but I'll take your 
word for that being okay.

Rene