Re: [Bug #11382] e1000e: 2.6.27-rc1 corrupts EEPROM/NVM

["Dave Airlie" <airlied-Re5JQEeQqe8AvxtiuMwx3w@public.gmane.org>]

On Wed, Sep 24, 2008 at 2:12 PM, David Miller <davem-fT/PcQaiUtIeIZ0/mPfg9Q@public.gmane.org> wrote:
> From: Jiri Kosina <jkosina-AlSwsSmVLrQ@public.gmane.org>
> Date: Wed, 24 Sep 2008 00:19:00 +0200 (CEST)
>
>> On Tue, 23 Sep 2008, Jeff Kirsher wrote:
>>
>> > >> I don't think OpenSUSE was shipping any of the GEM bits.
>> > > Good data point, can someone confirm this?  Also, what X server version
>> > > is the effected OpenSUSE shipping?
>> > OpenSuSE 11 ships x server version 7.3.
>>
>> Opensuse 11 is fine.
>>
>> The problem can be reproduced [not only] on opensuse 11.1 beta1, which has
>>
>>       xorg-x11-7.4-1.6.x86_64.rpm
>
> I did some snooping around, and while doing so I noticed that the PCI
> mmap code for x86 doesn't do one bit of range checking on the size, or
> any other aspect of the request, wrt. the MMIO regions actually mapped
> in the BARs of the PCI device.
>
> Yikes!
>
> It just does a reserve_memtype() on the address range, and says "ok".
>
> So if, for example, the X server tries to mmap() more than an MMIO bar
> actually maps, the kernel lets the user do this.
>
> It would be very interesting to add the appropriate checks to
> pci_mmap_page_range() in arch/x86/pci/i386.c, anyone who wants to do
> this can use the code in arch/sparc64/kernel/pci.c:
> __pci_mmap_make_offset() as a guide, and see what happens.
>
> If the MMIO space regions of the video cards sit right before the
> E1000E ones on the effected systems, that would pretty much
> convince me that this is the kind of problem we are having here.
>
> This also reminds me that there was that whole set of issues that
> had to get worked out wrt. write-caching of mappings on x86.
>

I'm still dubious about this, wouldn't we see other wierdass side
effects if X was trashing the BARs on other devices?

I think tglx is on the right path, same problem as e1000, code is
stupid, it can reenter the nvram read/write code from irq
context, and pwn itself.

Dave.