lkml.org 
[lkml]   [2008]   [Apr]   [26]   [last100]   RSS Feed
Views: [more markup]  [less markup]  [headers]  [forward] 
 
Messages in this thread
/
SubjectRE: [git pull] x86 PAT changes
DateSat, 26 Apr 2008 09:54:23 -0700
From"Pallipadi, Venkatesh" <>

>-----Original Message-----
>From: linux-kernel-owner@vger.kernel.org 
>[mailto:linux-kernel-owner@vger.kernel.org] On Behalf Of Ingo Molnar
>Sent: Saturday, April 26, 2008 1:56 AM
>To: Linus Torvalds
>Cc: H. Peter Anvin; linux-kernel@vger.kernel.org; Andrew 
>Morton; Thomas Gleixner
>Subject: Re: [git pull] x86 PAT changes
>
>
>* Linus Torvalds <torvalds@linux-foundation.org> wrote:
>
>> On Fri, 25 Apr 2008, H. Peter Anvin wrote:
>> > 
>> > The problem is that that can create cached/uncached aliases, which 
>> > can cause some processors to lock up (especially AMD is known to 
>> > have a lot of errata in this area.)
>> 
>> Umm.. I don't think you understand. Right now, NONPROMISC_DEVMEM 
>> doesn't just disable mmap() on /dev/mem, it disables totally regular 
>> reads and writes too. That seems pretty damn excessive.
>> 
>> If it was just mmap(), I don't think it would matter much. I don't 
>> think we traditionally even supported mmap() on real RAM 
>(because the 
>> page counting would get confused), and that actually got supported 
>> only thanks to VM changes that made it possible.
>> 
>> But read/write has always been supported, and shouldn't cause any 
>> cached/uncached aliases!
>
>You are right, there should be no architectural need to make PAT 
>dependent on nonpromisc-devmem, and thus the patch below 
>should be safe.
>

Agreed that NONPROMISC_DEVMEM is not really needed for read/write. But,
we will still need it for /dev/mem.

The problem with /dev/mem maps of RAM is situation like this:
1) drivers does vmalloc(), followed by set_memory_uc.
2) User does a /dev/mem map of that vmalloced physical address. User
will get a UC mapping for /dev/mem.
3) driver changes the memory to set_memory_wb and frees the memory.
4) user mapping for this address is still UC which will lead to
aliasing.

Read/write is ok, as they will just use __va for RAM to access and that
will always be consistent.

Thanks,
Venki


\
 
 \ /
  Last update: 2008-04-26 18:55    [from the cache]
©2003-2008