Re: Faulty MTRR setups in Toshiba laptops crash video drivers

[Robert Hancock <hancockrwd@gmail.com>]

On Tue, Nov 3, 2009 at 7:08 AM, Mehmet Giritli <mehmet@giritli.eu> wrote:
> On Mon, 2009-11-02 at 18:06 -0600, Robert Hancock wrote:
>> On 11/02/2009 10:28 AM, Mehmet Giritli wrote:
>> > Hello all,
>> >
>> > On some laptops using ATI 3650HD radeon chipsets and 4GB ram (and may be
>> > others as well), it is impossible to start any xserver due to (faulty?)
>> > mtrr setups on the machine. The computer locks up or just gives an
>> > unresponsive black screen. There are various drivers for radeon cards
>> > and the crash occurs in all of these drivers in the very same way!
>> >
>> > This bug seems to occur mostly on Toshiba laptops with ATI graphic
>> > cards. For some people, CONFIG_MTRR_SANITIZER seemed to cure the issue
>> > but made no difference for many others like me.
>> >
>> > RadeonHD people seem to have problems in understanding the nature of
>> > this bug. Please have a look at the threads:
>> >
>> > http://lists.opensuse.org/radeonhd/2009-05/msg00040.html
>> >
>> > and,
>> >
>> > http://bugs.freedesktop.org/show_bug.cgi?id=20645
>> >
>> > for problem reports and dev responses. RadeonHD devs suggest that some
>> > laptops have weird mtrr setups and this must be fixed in the kernel code
>> > and not the driver code.
>>
>> Unfortunately I don't see enough details to be able to tell what lead
>> them to this conclusion. Can you post the output of:
>>
>> lspci -vv
>> cat /proc/mtrr
>
> lspci -vv
> http://bugzilla.kernel.org/attachment.cgi?id=23634
>
> /proc/mtrr when 4GB installed
> http://bugzilla.kernel.org/attachment.cgi?id=23632
>
> /proc/mtrr when 2GB installed
> http://bugzilla.kernel.org/attachment.cgi?id=23633
>
> There is also a bug report that I opened a while back here:
> http://bugzilla.kernel.org/show_bug.cgi?id=14054

So it appears that your machine has an MTRR entry which marks the
first 16MB of the video card's video RAM BAR area as "write-through",
which seems quite clearly wrong. However, all that I would expect to
see this do would be to cause X to be unable to map the BAR as
write-combining (which you can see from dmesg is indeed occurring)
causing a loss of performance. And the same MTRR is there in the 2GB
case as in the 4GB case. So I don't think the question of why it works
with 2GB and not with 4GB is anything to do with MTRRs..