Re: [PATCH v2 8/8] drm/amdgpu: start using kernel doorbell bo

[Shashank Sharma <shashank.sharma@amd.com>]

On 14/02/2023 19:40, Christian König wrote:
> Am 14.02.23 um 17:15 schrieb Shashank Sharma:
>> From: Shashank Sharma <contactshashanksharma@gmail.com>
>>
>> This patch does the following:
>>
>> - Adds new variables like mman.doorbell_bo_size/gpu_addr/cpu_addr.
>>    The cpu_addr ptr will be used now for doorbell read/write from
>>    doorbell BAR.
>> - Adjusts the existing code to use kernel doorbell BO's size and its
>>    cpu_address.
>>
>> Cc: Alex Deucher <alexander.deucher@amd.com>
>> Cc: Christian Koenig <christian.koenig@amd.com>
>> Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
>> Signed-off-by: Shashank Sharma <shashank.sharma@amd.com>
>
> Maybe squash this one together with the previous patch.

I just split it from the last patch in this series, thought it was too 
scattered and might not be

easy to review :D


>
> But see below.
>
>> ---
>>   drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.c   |  5 ++-
>>   drivers/gpu/drm/amd/amdgpu/amdgpu_device.c   | 33 +++++++++-----------
>>   drivers/gpu/drm/amd/amdgpu/amdgpu_doorbell.h |  1 -
>>   3 files changed, 16 insertions(+), 23 deletions(-)
>>
>> diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.c 
>> b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.c
>> index 0493c64e9d0a..87f486f522ae 100644
>> --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.c
>> +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.c
>> @@ -109,11 +109,10 @@ static void amdgpu_doorbell_get_kfd_info(struct 
>> amdgpu_device *adev,
>>           *aperture_base = adev->doorbell.doorbell_aper_base;
>>           *aperture_size = 0;
>>           *start_offset = 0;
>> -    } else if (adev->doorbell.doorbell_aper_size > 
>> adev->doorbell.num_doorbells *
>> -                        sizeof(u32)) {
>> +    } else if (adev->doorbell.doorbell_aper_size > 
>> adev->mman.doorbell_kernel_bo_size) {
>>           *aperture_base = adev->doorbell.doorbell_aper_base;
>>           *aperture_size = adev->doorbell.doorbell_aper_size;
>> -        *start_offset = adev->doorbell.num_doorbells * sizeof(u32);
>> +        *start_offset = adev->mman.doorbell_kernel_bo_size;
>>       } else {
>>           *aperture_base = 0;
>>           *aperture_size = 0;
>> diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c 
>> b/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c
>> index 43c1b67c2778..fde199434579 100644
>> --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c
>> +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c
>> @@ -596,8 +596,8 @@ u32 amdgpu_mm_rdoorbell(struct amdgpu_device 
>> *adev, u32 index)
>>       if (amdgpu_device_skip_hw_access(adev))
>>           return 0;
>>   -    if (index < adev->doorbell.num_doorbells) {
>> -        return readl(adev->mman.doorbell_aper_base_kaddr + index);
>> +    if (index < adev->mman.num_kernel_doorbells) {
>> +        return readl(adev->mman.doorbell_cpu_addr + index);
>>       } else {
>>           DRM_ERROR("reading beyond doorbell aperture: 0x%08x!\n", 
>> index);
>>           return 0;
>> @@ -619,8 +619,8 @@ void amdgpu_mm_wdoorbell(struct amdgpu_device 
>> *adev, u32 index, u32 v)
>>       if (amdgpu_device_skip_hw_access(adev))
>>           return;
>>   -    if (index < adev->doorbell.num_doorbells) {
>> -        writel(v, adev->mman.doorbell_aper_base_kaddr + index);
>> +    if (index < adev->mman.num_kernel_doorbells) {
>> +        writel(v, adev->mman.doorbell_cpu_addr + index);
>>       } else {
>>           DRM_ERROR("writing beyond doorbell aperture: 0x%08x!\n", 
>> index);
>>       }
>> @@ -640,8 +640,8 @@ u64 amdgpu_mm_rdoorbell64(struct amdgpu_device 
>> *adev, u32 index)
>>       if (amdgpu_device_skip_hw_access(adev))
>>           return 0;
>>   -    if (index < adev->doorbell.num_doorbells) {
>> -        return atomic64_read((atomic64_t 
>> *)(adev->mman.doorbell_aper_base_kaddr + index));
>> +    if (index < adev->mman.num_kernel_doorbells) {
>> +        return atomic64_read((atomic64_t 
>> *)(adev->mman.doorbell_cpu_addr + index));
>>       } else {
>>           DRM_ERROR("reading beyond doorbell aperture: 0x%08x!\n", 
>> index);
>>           return 0;
>> @@ -663,8 +663,8 @@ void amdgpu_mm_wdoorbell64(struct amdgpu_device 
>> *adev, u32 index, u64 v)
>>       if (amdgpu_device_skip_hw_access(adev))
>>           return;
>>   -    if (index < adev->doorbell.num_doorbells) {
>> -        atomic64_set((atomic64_t 
>> *)(adev->mman.doorbell_aper_base_kaddr + index), v);
>> +    if (index < adev->mman.num_kernel_doorbells) {
>> +        atomic64_set((atomic64_t *)(adev->mman.doorbell_cpu_addr + 
>> index), v);
>>       } else {
>>           DRM_ERROR("writing beyond doorbell aperture: 0x%08x!\n", 
>> index);
>>       }
>> @@ -1037,7 +1037,7 @@ static int amdgpu_device_doorbell_init(struct 
>> amdgpu_device *adev)
>>       if (adev->asic_type < CHIP_BONAIRE) {
>>           adev->doorbell.doorbell_aper_base = 0;
>>           adev->doorbell.doorbell_aper_size = 0;
>> -        adev->doorbell.num_doorbells = 0;
>> +        adev->mman.num_kernel_doorbells = 0;
>>           adev->mman.doorbell_aper_base_kaddr = NULL;
>>           return 0;
>>       }
>> @@ -1052,13 +1052,13 @@ static int amdgpu_device_doorbell_init(struct 
>> amdgpu_device *adev)
>>       adev->doorbell.doorbell_aper_size = 
>> pci_resource_len(adev->pdev, 2);
>>         if (adev->enable_mes) {
>> -        adev->doorbell.num_doorbells =
>> +        adev->mman.num_kernel_doorbells =
>>               adev->doorbell.doorbell_aper_size / sizeof(u32);
>>       } else {
>> -        adev->doorbell.num_doorbells =
>> +        adev->mman.num_kernel_doorbells =
>>               min_t(u32, adev->doorbell.doorbell_aper_size / 
>> sizeof(u32),
>>                     adev->doorbell_index.max_assignment+1);
>> -        if (adev->doorbell.num_doorbells == 0)
>> +        if (adev->mman.num_kernel_doorbells == 0)
>>               return -EINVAL;
>>             /* For Vega, reserve and map two pages on doorbell BAR 
>> since SDMA
>> @@ -1068,15 +1068,10 @@ static int amdgpu_device_doorbell_init(struct 
>> amdgpu_device *adev)
>>            * the max num_doorbells should + 1 page (0x400 in dword)
>>            */
>>           if (adev->asic_type >= CHIP_VEGA10)
>> -            adev->doorbell.num_doorbells += 0x400;
>> +            adev->mman.num_kernel_doorbells += 0x400;
>>       }
>>   -    adev->mman.doorbell_aper_base_kaddr = 
>> ioremap(adev->doorbell.doorbell_aper_base,
>> -                              adev->doorbell.num_doorbells *
>> -                              sizeof(u32));
>> -    if (adev->mman.doorbell_aper_base_kaddr == NULL)
>> -        return -ENOMEM;
>> -
>> +    adev->mman.doorbell_kernel_bo_size = 
>> adev->mman.num_kernel_doorbells * sizeof(u32);
>
> I would just keep the kernel_bo_size around and make the 
> num_kernel_doorbells a local variable.
>
Noted,

- Shashank

> Christian.
>
>>       return 0;
>>   }
>>   diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_doorbell.h 
>> b/drivers/gpu/drm/amd/amdgpu/amdgpu_doorbell.h
>> index 526b6b4a86dd..7bdff4f926ad 100644
>> --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_doorbell.h
>> +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_doorbell.h
>> @@ -28,7 +28,6 @@ struct amdgpu_doorbell {
>>       /* doorbell mmio */
>>       resource_size_t        doorbell_aper_base;
>>       resource_size_t        doorbell_aper_size;
>> -    u32            num_doorbells;    /* Number of doorbells actually 
>> reserved for amdgpu. */
>>   };
>>     /* Reserved doorbells for amdgpu (including multimedia).
>