[PATCH 1/7] drm/amdgpu: fix VM PD addr shift

[Christian König]

Am 28.11.2017 um 03:42 schrieb Chunming Zhou:
>
>
> On 2017年11月28日 00:02, Christian König wrote:
>> The block size only affects the leave nodes, everything else is fixed.
> This is not true, block size affects every level entries, see the 
> register explains for VM_CONTEXTx_CNTL.PAGE_TABLE_BLOCK_SIZE:
> "LOG2(number of 2MB logical address ranges) pointed to by a PDE0 page 
> text directory entry. The native page size for the component ptes 
> comes from the PDE0.block_fragment_size field or 
> BASE_ADDR.block_fragment_size field (for a flat page table). The 
> PAGE_TABLE_BLOCK_SIZE field only has an effect on address calculations 
> for >= 2 level page tables however. A flat page table will utilize as 
> many ptes as are required to represent the logical address between 
> START_ADDR and END_ADDR, not limited by PAGE_TABLE_BLOCK_SIZE."

The description of the register is incorrect or at least misleading. 
I've fallen into the same trap as well.

Take a look at the VM documentation, there is a small footnote that 
intermediate page directories are always 512 entries in size.

I've confirmed that by configuring a Vega10 into 3 level page directory 
instead of the usual 4 and then using a block size of 18 for the page 
tables.

Regards,
Christian.

>
> Regards,
> David Zhou
>>
>> Signed-off-by: Christian König <christian.koenig at amd.com>
>> ---
>>   drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c | 28 
>> +++++++++++++++++++++++-----
>>   1 file changed, 23 insertions(+), 5 deletions(-)
>>
>> diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c 
>> b/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c
>> index 122379dfc7d8..f1e541e9b514 100644
>> --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c
>> +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c
>> @@ -139,6 +139,24 @@ struct amdgpu_prt_cb {
>>   };
>>     /**
>> + * amdgpu_vm_level_shift - return the addr shift for each level
>> + *
>> + * @adev: amdgpu_device pointer
>> + *
>> + * Returns the number of bits the pfn needs to be right shifted for 
>> a level.
>> + */
>> +static unsigned amdgpu_vm_level_shift(struct amdgpu_device *adev,
>> +                      unsigned level)
>> +{
>> +    if (level != adev->vm_manager.num_level)
>> +        return 9 * (adev->vm_manager.num_level - level - 1) +
>> +            adev->vm_manager.block_size;
>> +    else
>> +        /* For the page tables on the leaves */
>> +        return 0;
>> +}
>> +
>> +/**
>>    * amdgpu_vm_num_entries - return the number of entries in a PD/PT
>>    *
>>    * @adev: amdgpu_device pointer
>> @@ -288,8 +306,7 @@ static int amdgpu_vm_alloc_levels(struct 
>> amdgpu_device *adev,
>>                     uint64_t saddr, uint64_t eaddr,
>>                     unsigned level)
>>   {
>> -    unsigned shift = (adev->vm_manager.num_level - level) *
>> -        adev->vm_manager.block_size;
>> +    unsigned shift = amdgpu_vm_level_shift(adev, level);
>>       unsigned pt_idx, from, to;
>>       int r;
>>       u64 flags;
>> @@ -1302,18 +1319,19 @@ void amdgpu_vm_get_entry(struct 
>> amdgpu_pte_update_params *p, uint64_t addr,
>>                struct amdgpu_vm_pt **entry,
>>                struct amdgpu_vm_pt **parent)
>>   {
>> -    unsigned idx, level = p->adev->vm_manager.num_level;
>> +    unsigned level = 0;
>>         *parent = NULL;
>>       *entry = &p->vm->root;
>>       while ((*entry)->entries) {
>> -        idx = addr >> (p->adev->vm_manager.block_size * level--);
>> +        unsigned idx = addr >> amdgpu_vm_level_shift(p->adev, level++);
>> +
>>           idx %= amdgpu_bo_size((*entry)->base.bo) / 8;
>>           *parent = *entry;
>>           *entry = &(*entry)->entries[idx];
>>       }
>>   -    if (level)
>> +    if (level != p->adev->vm_manager.num_level)
>>           *entry = NULL;
>>   }
>