[PATCH v6 0/6] Allwinner H6 Mali GPU support

[Robin Murphy]

On 31/05/2019 13:04, Tomeu Vizoso wrote:
> On Wed, 29 May 2019 at 19:38, Robin Murphy <robin.murphy at arm.com> wrote:
>>
>> On 29/05/2019 16:09, Tomeu Vizoso wrote:
>>> On Tue, 21 May 2019 at 18:11, Clément Péron <peron.clem at gmail.com> wrote:
>>>>
>>> [snip]
>>>> [  345.204813] panfrost 1800000.gpu: mmu irq status=1
>>>> [  345.209617] panfrost 1800000.gpu: Unhandled Page fault in AS0 at VA
>>>> 0x0000000002400400
>>>
>>>   From what I can see here, 0x0000000002400400 points to the first byte
>>> of the first submitted job descriptor.
>>>
>>> So mapping buffers for the GPU doesn't seem to be working at all on
>>> 64-bit T-760.
>>>
>>> Steven, Robin, do you have any idea of why this could be?
>>
>> I tried rolling back to the old panfrost/nondrm shim, and it works fine
>> with kbase, and I also found that T-820 falls over in the exact same
>> manner, so the fact that it seemed to be common to the smaller 33-bit
>> designs rather than anything to do with the other
>> job_descriptor_size/v4/v5 complication turned out to be telling.
> 
> Is this complication something you can explain? I don't know what v4
> and v5 are meant here.

I was alluding to BASE_HW_FEATURE_V4, which I believe refers to the 
Midgard architecture version - the older versions implemented by T6xx 
and T720 seem to be collectively treated as "v4", while T760 and T8xx 
would effectively be "v5".

>> [ as an aside, are 64-bit jobs actually known not to work on v4 GPUs, or
>> is it just that nobody's yet observed a 64-bit blob driving one? ]
> 
> I'm looking right now at getting Panfrost working on T720 with 64-bit
> descriptors, with the ultimate goal of making Panfrost
> 64-bit-descriptor only so we can have a single build of Mesa in
> distros.

Cool, I'll keep an eye out, and hope that it might be enough for T620 on 
Juno, too :)

>> Long story short, it appears that 'Mali LPAE' is also lacking the start
>> level notion of VMSA, and expects a full 4-level table even for <40 bits
>> when level 0 effectively redundant. Thus walking the 3-level table that
>> io-pgtable comes back with ends up going wildly wrong. The hack below
>> seems to do the job for me; if Clément can confirm (on T-720 you'll
>> still need the userspace hack to force 32-bit jobs as well) then I think
>> I'll cook up a proper refactoring of the allocator to put things right.
> 
> Mmaps seem to work with this patch, thanks.
> 
> The main complication I'm facing right now seems to be that the SFBD
> descriptor on T720 seems to be different from the one we already had
> (tested on T6xx?).

OK - with the 32-bit hack pointed to up-thread, a quick kmscube test 
gave me the impression that T720 works fine, but on closer inspection 
some parts of glmark2 do seem to go a bit wonky (although I suspect at 
least some of it is just down to the FPGA setup being both very slow and 
lacking in memory bandwidth), and the "nv12-1img" mode of kmscube turns 
out to render in some delightfully wrong colours.

I'll try to get a 'proper' version of the io-pgtable patch posted soon.

Thanks,
Robin.

> 
> Cheers,
> 
> Tomeu
> 
>> Robin.
>>
>>
>> ----->8-----
>> diff --git a/drivers/iommu/io-pgtable-arm.c b/drivers/iommu/io-pgtable-arm.c
>> index 546968d8a349..f29da6e8dc08 100644
>> --- a/drivers/iommu/io-pgtable-arm.c
>> +++ b/drivers/iommu/io-pgtable-arm.c
>> @@ -1023,12 +1023,14 @@ arm_mali_lpae_alloc_pgtable(struct
>> io_pgtable_cfg *cfg, void *cookie)
>>          iop = arm_64_lpae_alloc_pgtable_s1(cfg, cookie);
>>          if (iop) {
>>                  u64 mair, ttbr;
>> +               struct arm_lpae_io_pgtable *data = io_pgtable_ops_to_data(&iop->ops);
>>
>> +               data->levels = 4;
>>                  /* Copy values as union fields overlap */
>>                  mair = cfg->arm_lpae_s1_cfg.mair[0];
>>                  ttbr = cfg->arm_lpae_s1_cfg.ttbr[0];
>>
>> _______________________________________________
>> dri-devel mailing list
>> dri-devel at lists.freedesktop.org
>> https://lists.freedesktop.org/mailman/listinfo/dri-devel