diff mbox

[2/2] drm/radeon: fix bank tiling parameters on SI

Message ID 1343735331-4922-2-git-send-email-deathsimple@vodafone.de (mailing list archive)
State New, archived
Headers show

Commit Message

Christian König July 31, 2012, 11:48 a.m. UTC
The sixteen bank case wasn't handled here, leading to GPU
crashes because of userspace miscalculation.

Signed-off-by: Christian König <deathsimple@vodafone.de>
Cc: stable@vger.kernel.org
---
 drivers/gpu/drm/radeon/si.c |   16 ++++++++++++----
 1 file changed, 12 insertions(+), 4 deletions(-)

Comments

Alex Deucher July 31, 2012, 2:02 p.m. UTC | #1
On Tue, Jul 31, 2012 at 7:48 AM, Christian König
<deathsimple@vodafone.de> wrote:
> The sixteen bank case wasn't handled here, leading to GPU
> crashes because of userspace miscalculation.

You mean a GPU hang or a segfault in userspace?  IIRC, from the hw
perspective numbers of banks over 8 are considered 8 for tiling.

Alex

>
> Signed-off-by: Christian König <deathsimple@vodafone.de>
> Cc: stable@vger.kernel.org
> ---
>  drivers/gpu/drm/radeon/si.c |   16 ++++++++++++----
>  1 file changed, 12 insertions(+), 4 deletions(-)
>
> diff --git a/drivers/gpu/drm/radeon/si.c b/drivers/gpu/drm/radeon/si.c
> index 0b02792..0182fc8 100644
> --- a/drivers/gpu/drm/radeon/si.c
> +++ b/drivers/gpu/drm/radeon/si.c
> @@ -1639,11 +1639,19 @@ static void si_gpu_init(struct radeon_device *rdev)
>                 /* XXX what about 12? */
>                 rdev->config.si.tile_config |= (3 << 0);
>                 break;
> -       }
> -       if ((mc_arb_ramcfg & NOOFBANK_MASK) >> NOOFBANK_SHIFT)
> -               rdev->config.si.tile_config |= 1 << 4;
> -       else
> +       }
> +       switch ((mc_arb_ramcfg & NOOFBANK_MASK) >> NOOFBANK_SHIFT) {
> +       case 0: /* four banks */
>                 rdev->config.si.tile_config |= 0 << 4;
> +               break;
> +       case 1: /* eight banks */
> +               rdev->config.si.tile_config |= 1 << 4;
> +               break;
> +       case 2: /* sixteen banks */
> +       default:
> +               rdev->config.si.tile_config |= 2 << 4;
> +               break;
> +       }
>         rdev->config.si.tile_config |=
>                 ((gb_addr_config & PIPE_INTERLEAVE_SIZE_MASK) >> PIPE_INTERLEAVE_SIZE_SHIFT) << 8;
>         rdev->config.si.tile_config |=
> --
> 1.7.9.5
>
> _______________________________________________
> dri-devel mailing list
> dri-devel@lists.freedesktop.org
> http://lists.freedesktop.org/mailman/listinfo/dri-devel
Christian König July 31, 2012, 2:21 p.m. UTC | #2
On 31.07.2012 16:02, Alex Deucher wrote:
> On Tue, Jul 31, 2012 at 7:48 AM, Christian König
> <deathsimple@vodafone.de> wrote:
>> The sixteen bank case wasn't handled here, leading to GPU
>> crashes because of userspace miscalculation.
> You mean a GPU hang or a segfault in userspace?  IIRC, from the hw
> perspective numbers of banks over 8 are considered 8 for tiling.
Well it was a GPU segfault :) The GPU was accessing memory regions that 
weren't VM mapped.

As far as i could figure it out it wasn't happy with the alignment of 
the color buffer. The number of banks we used for that calculation 
seemed to be different from what the kernel programmed into the tiling 
config registers. So I tried changing it from 8 to 16 and hurray it 
started working (ok, more or less currently digging into the next problem).

It is possible that this is just masquerading another bug, but as far as 
I can see it is the most logical explanation.

Why are 16 banks equal to 8 banks on the hw side?

Christian.

>
> Alex
>
>> Signed-off-by: Christian König <deathsimple@vodafone.de>
>> Cc: stable@vger.kernel.org
>> ---
>>   drivers/gpu/drm/radeon/si.c |   16 ++++++++++++----
>>   1 file changed, 12 insertions(+), 4 deletions(-)
>>
>> diff --git a/drivers/gpu/drm/radeon/si.c b/drivers/gpu/drm/radeon/si.c
>> index 0b02792..0182fc8 100644
>> --- a/drivers/gpu/drm/radeon/si.c
>> +++ b/drivers/gpu/drm/radeon/si.c
>> @@ -1639,11 +1639,19 @@ static void si_gpu_init(struct radeon_device *rdev)
>>                  /* XXX what about 12? */
>>                  rdev->config.si.tile_config |= (3 << 0);
>>                  break;
>> -       }
>> -       if ((mc_arb_ramcfg & NOOFBANK_MASK) >> NOOFBANK_SHIFT)
>> -               rdev->config.si.tile_config |= 1 << 4;
>> -       else
>> +       }
>> +       switch ((mc_arb_ramcfg & NOOFBANK_MASK) >> NOOFBANK_SHIFT) {
>> +       case 0: /* four banks */
>>                  rdev->config.si.tile_config |= 0 << 4;
>> +               break;
>> +       case 1: /* eight banks */
>> +               rdev->config.si.tile_config |= 1 << 4;
>> +               break;
>> +       case 2: /* sixteen banks */
>> +       default:
>> +               rdev->config.si.tile_config |= 2 << 4;
>> +               break;
>> +       }
>>          rdev->config.si.tile_config |=
>>                  ((gb_addr_config & PIPE_INTERLEAVE_SIZE_MASK) >> PIPE_INTERLEAVE_SIZE_SHIFT) << 8;
>>          rdev->config.si.tile_config |=
>> --
>> 1.7.9.5
>>
>> _______________________________________________
>> dri-devel mailing list
>> dri-devel@lists.freedesktop.org
>> http://lists.freedesktop.org/mailman/listinfo/dri-devel
Alex Deucher July 31, 2012, 2:42 p.m. UTC | #3
On Tue, Jul 31, 2012 at 10:21 AM, Christian König
<deathsimple@vodafone.de> wrote:
> On 31.07.2012 16:02, Alex Deucher wrote:
>>
>> On Tue, Jul 31, 2012 at 7:48 AM, Christian König
>> <deathsimple@vodafone.de> wrote:
>>>
>>> The sixteen bank case wasn't handled here, leading to GPU
>>> crashes because of userspace miscalculation.
>>
>> You mean a GPU hang or a segfault in userspace?  IIRC, from the hw
>> perspective numbers of banks over 8 are considered 8 for tiling.
>
> Well it was a GPU segfault :) The GPU was accessing memory regions that
> weren't VM mapped.
>
> As far as i could figure it out it wasn't happy with the alignment of the
> color buffer. The number of banks we used for that calculation seemed to be
> different from what the kernel programmed into the tiling config registers.
> So I tried changing it from 8 to 16 and hurray it started working (ok, more
> or less currently digging into the next problem).
>
> It is possible that this is just masquerading another bug, but as far as I
> can see it is the most logical explanation.

Seems reasonable.

Reviewed-by: Alex Deucher <alexander.deucher@amd.com>

>
> Why are 16 banks equal to 8 banks on the hw side?

I was talking to one of the hw guys about some issues with tiling.
Certain 6xx or 7xx chips report more than 8 banks in the MC config,
but the tiling configuration on them only supports 4 and 8 banks.
IIRC, he said number of banks above 8 should be treated as 8 at least
for 6xx-9xx.  Although, looking at it again it looks like maybe we
want 16 bank support for evergreen/cayman as well after all.  It
shouldn't hurt other than possibly over allocating a bit.

Alex

>
> Christian.
>
>
>>
>> Alex
>>
>>> Signed-off-by: Christian König <deathsimple@vodafone.de>
>>> Cc: stable@vger.kernel.org
>>> ---
>>>   drivers/gpu/drm/radeon/si.c |   16 ++++++++++++----
>>>   1 file changed, 12 insertions(+), 4 deletions(-)
>>>
>>> diff --git a/drivers/gpu/drm/radeon/si.c b/drivers/gpu/drm/radeon/si.c
>>> index 0b02792..0182fc8 100644
>>> --- a/drivers/gpu/drm/radeon/si.c
>>> +++ b/drivers/gpu/drm/radeon/si.c
>>> @@ -1639,11 +1639,19 @@ static void si_gpu_init(struct radeon_device
>>> *rdev)
>>>                  /* XXX what about 12? */
>>>                  rdev->config.si.tile_config |= (3 << 0);
>>>                  break;
>>> -       }
>>> -       if ((mc_arb_ramcfg & NOOFBANK_MASK) >> NOOFBANK_SHIFT)
>>> -               rdev->config.si.tile_config |= 1 << 4;
>>> -       else
>>> +       }
>>> +       switch ((mc_arb_ramcfg & NOOFBANK_MASK) >> NOOFBANK_SHIFT) {
>>> +       case 0: /* four banks */
>>>                  rdev->config.si.tile_config |= 0 << 4;
>>> +               break;
>>> +       case 1: /* eight banks */
>>> +               rdev->config.si.tile_config |= 1 << 4;
>>> +               break;
>>> +       case 2: /* sixteen banks */
>>> +       default:
>>> +               rdev->config.si.tile_config |= 2 << 4;
>>> +               break;
>>> +       }
>>>          rdev->config.si.tile_config |=
>>>                  ((gb_addr_config & PIPE_INTERLEAVE_SIZE_MASK) >>
>>> PIPE_INTERLEAVE_SIZE_SHIFT) << 8;
>>>          rdev->config.si.tile_config |=
>>> --
>>> 1.7.9.5
>>>
>>> _______________________________________________
>>> dri-devel mailing list
>>> dri-devel@lists.freedesktop.org
>>> http://lists.freedesktop.org/mailman/listinfo/dri-devel
>
>
Christian König July 31, 2012, 3:21 p.m. UTC | #4
On 31.07.2012 16:42, Alex Deucher wrote:
> On Tue, Jul 31, 2012 at 10:21 AM, Christian König
> <deathsimple@vodafone.de> wrote:
>> On 31.07.2012 16:02, Alex Deucher wrote:
>>> On Tue, Jul 31, 2012 at 7:48 AM, Christian König
>>> <deathsimple@vodafone.de> wrote:
>>>> The sixteen bank case wasn't handled here, leading to GPU
>>>> crashes because of userspace miscalculation.
>>> You mean a GPU hang or a segfault in userspace?  IIRC, from the hw
>>> perspective numbers of banks over 8 are considered 8 for tiling.
>> Well it was a GPU segfault :) The GPU was accessing memory regions that
>> weren't VM mapped.
>>
>> As far as i could figure it out it wasn't happy with the alignment of the
>> color buffer. The number of banks we used for that calculation seemed to be
>> different from what the kernel programmed into the tiling config registers.
>> So I tried changing it from 8 to 16 and hurray it started working (ok, more
>> or less currently digging into the next problem).
>>
>> It is possible that this is just masquerading another bug, but as far as I
>> can see it is the most logical explanation.
> Seems reasonable.
>
> Reviewed-by: Alex Deucher <alexander.deucher@amd.com>
>
>> Why are 16 banks equal to 8 banks on the hw side?
> I was talking to one of the hw guys about some issues with tiling.
> Certain 6xx or 7xx chips report more than 8 banks in the MC config,
> but the tiling configuration on them only supports 4 and 8 banks.
> IIRC, he said number of banks above 8 should be treated as 8 at least
> for 6xx-9xx.  Although, looking at it again it looks like maybe we
> want 16 bank support for evergreen/cayman as well after all.  It
> shouldn't hurt other than possibly over allocating a bit.
The alternative is to reduce the number of banks in the gb_tile_mode* regs.

I should also note that I have encountered a couple of more problems 
with SI and 16 banks, they go away if I reduce the banks to 8. Not sure 
what the reason is but it is possible that our surface_manager can't 
handle 16 banks yet.

So be careful if you enable that for cayman and evergreen.

Cheers,
Christian.

>
> Alex
>
>> Christian.
>>
>>
>>> Alex
>>>
>>>> Signed-off-by: Christian König <deathsimple@vodafone.de>
>>>> Cc: stable@vger.kernel.org
>>>> ---
>>>>    drivers/gpu/drm/radeon/si.c |   16 ++++++++++++----
>>>>    1 file changed, 12 insertions(+), 4 deletions(-)
>>>>
>>>> diff --git a/drivers/gpu/drm/radeon/si.c b/drivers/gpu/drm/radeon/si.c
>>>> index 0b02792..0182fc8 100644
>>>> --- a/drivers/gpu/drm/radeon/si.c
>>>> +++ b/drivers/gpu/drm/radeon/si.c
>>>> @@ -1639,11 +1639,19 @@ static void si_gpu_init(struct radeon_device
>>>> *rdev)
>>>>                   /* XXX what about 12? */
>>>>                   rdev->config.si.tile_config |= (3 << 0);
>>>>                   break;
>>>> -       }
>>>> -       if ((mc_arb_ramcfg & NOOFBANK_MASK) >> NOOFBANK_SHIFT)
>>>> -               rdev->config.si.tile_config |= 1 << 4;
>>>> -       else
>>>> +       }
>>>> +       switch ((mc_arb_ramcfg & NOOFBANK_MASK) >> NOOFBANK_SHIFT) {
>>>> +       case 0: /* four banks */
>>>>                   rdev->config.si.tile_config |= 0 << 4;
>>>> +               break;
>>>> +       case 1: /* eight banks */
>>>> +               rdev->config.si.tile_config |= 1 << 4;
>>>> +               break;
>>>> +       case 2: /* sixteen banks */
>>>> +       default:
>>>> +               rdev->config.si.tile_config |= 2 << 4;
>>>> +               break;
>>>> +       }
>>>>           rdev->config.si.tile_config |=
>>>>                   ((gb_addr_config & PIPE_INTERLEAVE_SIZE_MASK) >>
>>>> PIPE_INTERLEAVE_SIZE_SHIFT) << 8;
>>>>           rdev->config.si.tile_config |=
>>>> --
>>>> 1.7.9.5
>>>>
>>>> _______________________________________________
>>>> dri-devel mailing list
>>>> dri-devel@lists.freedesktop.org
>>>> http://lists.freedesktop.org/mailman/listinfo/dri-devel
>>
diff mbox

Patch

diff --git a/drivers/gpu/drm/radeon/si.c b/drivers/gpu/drm/radeon/si.c
index 0b02792..0182fc8 100644
--- a/drivers/gpu/drm/radeon/si.c
+++ b/drivers/gpu/drm/radeon/si.c
@@ -1639,11 +1639,19 @@  static void si_gpu_init(struct radeon_device *rdev)
 		/* XXX what about 12? */
 		rdev->config.si.tile_config |= (3 << 0);
 		break;
-	}
-	if ((mc_arb_ramcfg & NOOFBANK_MASK) >> NOOFBANK_SHIFT)
-		rdev->config.si.tile_config |= 1 << 4;
-	else
+	}	
+	switch ((mc_arb_ramcfg & NOOFBANK_MASK) >> NOOFBANK_SHIFT) {
+	case 0: /* four banks */
 		rdev->config.si.tile_config |= 0 << 4;
+		break;
+	case 1: /* eight banks */
+		rdev->config.si.tile_config |= 1 << 4;
+		break;
+	case 2: /* sixteen banks */
+	default:
+		rdev->config.si.tile_config |= 2 << 4;
+		break;
+	}
 	rdev->config.si.tile_config |=
 		((gb_addr_config & PIPE_INTERLEAVE_SIZE_MASK) >> PIPE_INTERLEAVE_SIZE_SHIFT) << 8;
 	rdev->config.si.tile_config |=