[PATCH] drm/ttm: ioremap buffer according to TTM mem caching setting

Christian König ckoenig.leichtzumerken at gmail.com
Wed Mar 3 10:37:22 UTC 2021


Hi Thomas,

Am 03.03.21 um 09:49 schrieb Thomas Zimmermann:
> Hi
>
> Am 01.03.21 um 23:43 schrieb Oak Zeng:
>> If tbo.mem.bus.caching is cached, buffer is intended to be mapped
>> as cached from CPU. Map it with ioremap_cache.
>
> Just a question for my understanding: This is on-device memory? 
> Accessing device memory is usually slow. If that memory can be mapped 
> with CPU caching enabled, access will roughly be as fast as for system 
> memory?

There is still a penalty associated with accessing it from the CPU, but 
it is much faster (both lower latency as well as throughput) as 
traditional device memory accessed over PCIe.

Regards,
Christian.

>
> Best regards
> Thomas
>
>>
>> This wasn't necessary before as device memory was never mapped
>> as cached from CPU side. It becomes necessary for aldebaran as
>> device memory is mapped cached from CPU.
>>
>> Signed-off-by: Oak Zeng <Oak.Zeng at amd.com>
>> Reviewed-by: Christian Konig <Christian.Koenig at amd.com>
>> ---
>>   drivers/gpu/drm/ttm/ttm_bo_util.c | 8 ++++++++
>>   1 file changed, 8 insertions(+)
>>
>> diff --git a/drivers/gpu/drm/ttm/ttm_bo_util.c 
>> b/drivers/gpu/drm/ttm/ttm_bo_util.c
>> index 031e581..8c65a13 100644
>> --- a/drivers/gpu/drm/ttm/ttm_bo_util.c
>> +++ b/drivers/gpu/drm/ttm/ttm_bo_util.c
>> @@ -91,6 +91,8 @@ static int ttm_resource_ioremap(struct ttm_device 
>> *bdev,
>>             if (mem->bus.caching == ttm_write_combined)
>>               addr = ioremap_wc(mem->bus.offset, bus_size);
>> +        else if (mem->bus.caching == ttm_cached)
>> +            addr = ioremap_cache(mem->bus.offset, bus_size);
>>           else
>>               addr = ioremap(mem->bus.offset, bus_size);
>>           if (!addr) {
>> @@ -372,6 +374,9 @@ static int ttm_bo_ioremap(struct 
>> ttm_buffer_object *bo,
>>           if (mem->bus.caching == ttm_write_combined)
>>               map->virtual = ioremap_wc(bo->mem.bus.offset + offset,
>>                             size);
>> +        else if (mem->bus.caching == ttm_cached)
>> +            map->virtual = ioremap_cache(bo->mem.bus.offset + offset,
>> +                          size);
>>           else
>>               map->virtual = ioremap(bo->mem.bus.offset + offset,
>>                              size);
>> @@ -490,6 +495,9 @@ int ttm_bo_vmap(struct ttm_buffer_object *bo, 
>> struct dma_buf_map *map)
>>           else if (mem->bus.caching == ttm_write_combined)
>>               vaddr_iomem = ioremap_wc(mem->bus.offset,
>>                            bo->base.size);
>> +        else if (mem->bus.caching == ttm_cached)
>> +            vaddr_iomem = ioremap_cache(mem->bus.offset,
>> +                          bo->base.size);
>>           else
>>               vaddr_iomem = ioremap(mem->bus.offset, bo->base.size);
>>
>
>
> _______________________________________________
> amd-gfx mailing list
> amd-gfx at lists.freedesktop.org
> https://lists.freedesktop.org/mailman/listinfo/amd-gfx

-------------- next part --------------
An HTML attachment was scrubbed...
URL: <https://lists.freedesktop.org/archives/amd-gfx/attachments/20210303/9d807770/attachment-0001.htm>


More information about the amd-gfx mailing list