Nouveau - Jan 2018 - Problems getting nouveau to work with either Geforce GT710 or Geforce 9800GT on ARM Cortex-A9

On Thu, Jan 25, 2018 at 8:47 AM, Luís Mendes <luis.p.mendes at gmail.com>
wrote:
> Sorry for alI these individual emails, but I think is more readable
> this way, by having each independent information in a separate email.
>
> I also have these additional cards that I can try: Gefoce GT 1030 and
> Geforce GTX 1050 Ti.
>
> I have tried Geforce GT 1030 and it is working nicely but without
> hardware acceleration. Without the hardware acceleration I am able to
> get into X (Ubuntu Mate 17.10) and run glmark2.
> dmesg follows in attachment.
This is a GP108 for which you need not-yet-upstream code to get acceleration.
>
> I believe this indicates there is no issue with cards initialization,
> but is rather something to do with the hardware acceleration code.
Someone associated with (or working at?) RH was investigating this for
some server-type ARM board with PCIe slots. (Was it you? Rob Clark,
CC'd, will remember who it was, and may be able to point you in an
appropriate direction.) I think that there were various odd issues
with coherency which remained only lightly diagnosed. Note that all
your boots instantaneously print "GPU lockup, have a nice day".
Everything after that is largely irrelevant.

This happens when it tries to execute some very basic things and sees
that there is no progress. This means sticking a command buffer on the
DMA ring, letting the GPU execute it, and making sure that the fence
is hit as seen by the CPU. I would recommend tracking down why the
accel doesn't work (i.e. you get that message), and not worry about
little details like "image appears on screen".

Unfortunately the initialization logic in nouveau is a giant pile of
spaghetti (with various arguably good reasons, but it doesn't change
the end result), so tracking down exactly what fails in the GPU init
will be a bit of an undertaking. You should join #nouveau on
irc.freenode.net and get some additional advice there.

Good luck,

  -ilia

On Thu, Jan 25, 2018 at 8:57 AM, Ilia Mirkin <imirkin at alum.mit.edu>
wrote:
> On Thu, Jan 25, 2018 at 8:47 AM, Luís Mendes <luis.p.mendes at
gmail.com> wrote:
>> Sorry for alI these individual emails, but I think is more readable
>> this way, by having each independent information in a separate email.
>>
>> I also have these additional cards that I can try: Gefoce GT 1030 and
>> Geforce GTX 1050 Ti.
>>
>> I have tried Geforce GT 1030 and it is working nicely but without
>> hardware acceleration. Without the hardware acceleration I am able to
>> get into X (Ubuntu Mate 17.10) and run glmark2.
>> dmesg follows in attachment.
>
> This is a GP108 for which you need not-yet-upstream code to get
acceleration.
>
>>
>> I believe this indicates there is no issue with cards initialization,
>> but is rather something to do with the hardware acceleration code.
>
> Someone associated with (or working at?) RH was investigating this for
> some server-type ARM board with PCIe slots. (Was it you? Rob Clark,
> CC'd, will remember who it was, and may be able to point you in an
> appropriate direction.) I think that there were various odd issues
> with coherency which remained only lightly diagnosed. Note that all
> your boots instantaneously print "GPU lockup, have a nice day".
> Everything after that is largely irrelevant.
well, I don't actually have any ARM board w/ PCIe yet.. although I
would expect that drm_cflush* being no-ops on arm/arm64 isn't likely
to help things.

BR,
-R
> This happens when it tries to execute some very basic things and sees
> that there is no progress. This means sticking a command buffer on the
> DMA ring, letting the GPU execute it, and making sure that the fence
> is hit as seen by the CPU. I would recommend tracking down why the
> accel doesn't work (i.e. you get that message), and not worry about
> little details like "image appears on screen".
>
> Unfortunately the initialization logic in nouveau is a giant pile of
> spaghetti (with various arguably good reasons, but it doesn't change
> the end result), so tracking down exactly what fails in the GPU init
> will be a bit of an undertaking. You should join #nouveau on
> irc.freenode.net and get some additional advice there.
>
> Good luck,
>
>   -ilia

Hi Rob, Ilia,

Thanks for your information.

I'll try to have a look into the nouveau code... maybe I can find why
the dma fence doesn't complete. In the amdgpu and radeon drivers I am
being able to get further with the ARM platform, may be I can compare
how they're handling the dma transfers, despite I am getting gpu
lockups after a while of playing videos on ARM with AMD cards too, but
at least it is going a lot further.

Can you point me the files or functions in the nouveau code, where the
initialization tests dma transfers are started? I can try to find
them, but this could save me some time.

Thanks,
Luís

On Thu, Jan 25, 2018 at 2:47 PM, Rob Clark <robdclark at gmail.com>
wrote:
> On Thu, Jan 25, 2018 at 8:57 AM, Ilia Mirkin <imirkin at
alum.mit.edu> wrote:
>> On Thu, Jan 25, 2018 at 8:47 AM, Luís Mendes <luis.p.mendes at
gmail.com> wrote:
>>> Sorry for alI these individual emails, but I think is more readable
>>> this way, by having each independent information in a separate
email.
>>>
>>> I also have these additional cards that I can try: Gefoce GT 1030
and
>>> Geforce GTX 1050 Ti.
>>>
>>> I have tried Geforce GT 1030 and it is working nicely but without
>>> hardware acceleration. Without the hardware acceleration I am able
to
>>> get into X (Ubuntu Mate 17.10) and run glmark2.
>>> dmesg follows in attachment.
>>
>> This is a GP108 for which you need not-yet-upstream code to get
acceleration.
>>
>>>
>>> I believe this indicates there is no issue with cards
initialization,
>>> but is rather something to do with the hardware acceleration code.
>>
>> Someone associated with (or working at?) RH was investigating this for
>> some server-type ARM board with PCIe slots. (Was it you? Rob Clark,
>> CC'd, will remember who it was, and may be able to point you in an
>> appropriate direction.) I think that there were various odd issues
>> with coherency which remained only lightly diagnosed. Note that all
>> your boots instantaneously print "GPU lockup, have a nice
day".
>> Everything after that is largely irrelevant.
>
> well, I don't actually have any ARM board w/ PCIe yet.. although I
> would expect that drm_cflush* being no-ops on arm/arm64 isn't likely
> to help things.
>
> BR,
> -R
>
>> This happens when it tries to execute some very basic things and sees
>> that there is no progress. This means sticking a command buffer on the
>> DMA ring, letting the GPU execute it, and making sure that the fence
>> is hit as seen by the CPU. I would recommend tracking down why the
>> accel doesn't work (i.e. you get that message), and not worry about
>> little details like "image appears on screen".
>>
>> Unfortunately the initialization logic in nouveau is a giant pile of
>> spaghetti (with various arguably good reasons, but it doesn't
change
>> the end result), so tracking down exactly what fails in the GPU init
>> will be a bit of an undertaking. You should join #nouveau on
>> irc.freenode.net and get some additional advice there.
>>
>> Good luck,
>>
>>   -ilia