On Mon, Sep 26, 2016 at 9:32 PM, Ard Biesheuvel
<[email protected]> wrote:
> Some subdevices (i.e., fb/nv50.c and fb/gf100.c) map a scratch page using
> dma_map_page() way before the TTM layer has had a chance to set the DMA
> mask. This may prevent the driver from loading at all on platforms whose
> system memory is not covered by the default DMA mask of 32-bit (i.e., when
> all RAM is above 4 GB).
>
> So set a preliminary DMA mask right after constructing the PCI device, and
> base it on the .dma_bits member of the MMU subdevice, which is what the TTM
> layer will base the DMA mask on as well.
>
> Signed-off-by: Ard Biesheuvel <[email protected]>
> ---
> drivers/gpu/drm/nouveau/nouveau_drm.c | 11 +++++++++++
> 1 file changed, 11 insertions(+)
>
> diff --git a/drivers/gpu/drm/nouveau/nouveau_drm.c b/drivers/gpu/drm/nouveau/nouveau_drm.c
> index 652ab111dd74..e61e9a0adb51 100644
> --- a/drivers/gpu/drm/nouveau/nouveau_drm.c
> +++ b/drivers/gpu/drm/nouveau/nouveau_drm.c
> @@ -361,6 +361,17 @@ static int nouveau_drm_probe(struct pci_dev *pdev,
>
> pci_set_master(pdev);
>
> + /*
> + * Set a preliminary DMA mask based on the .dma_bits member of the
> + * MMU subdevice. This allows other subdevices to create DMA mappings
> + * in their init() function, which are called before the TTM layer sets
> + * the DMA mask definitively.
> + * This is necessary for platforms where the default DMA mask of 32
> + * does not cover any system memory, i.e., when all RAM is > 4 GB.
> + */
> + dma_set_mask_and_coherent(device->dev,
> + DMA_BIT_MASK(device->mmu->dma_bits));
I would just move this to nvkm_device_pci_new() so that it perfectly
mirrors the same call done in nvkm_device_tegra_new(), which was done
for the same purpose. Otherwise, looks good to me.
On 3 October 2016 at 06:39, Alexandre Courbot <[email protected]> wrote:
> On Mon, Sep 26, 2016 at 9:32 PM, Ard Biesheuvel
> <[email protected]> wrote:
>> Some subdevices (i.e., fb/nv50.c and fb/gf100.c) map a scratch page using
>> dma_map_page() way before the TTM layer has had a chance to set the DMA
>> mask. This may prevent the driver from loading at all on platforms whose
>> system memory is not covered by the default DMA mask of 32-bit (i.e., when
>> all RAM is above 4 GB).
>>
>> So set a preliminary DMA mask right after constructing the PCI device, and
>> base it on the .dma_bits member of the MMU subdevice, which is what the TTM
>> layer will base the DMA mask on as well.
>>
>> Signed-off-by: Ard Biesheuvel <[email protected]>
>> ---
>> drivers/gpu/drm/nouveau/nouveau_drm.c | 11 +++++++++++
>> 1 file changed, 11 insertions(+)
>>
>> diff --git a/drivers/gpu/drm/nouveau/nouveau_drm.c b/drivers/gpu/drm/nouveau/nouveau_drm.c
>> index 652ab111dd74..e61e9a0adb51 100644
>> --- a/drivers/gpu/drm/nouveau/nouveau_drm.c
>> +++ b/drivers/gpu/drm/nouveau/nouveau_drm.c
>> @@ -361,6 +361,17 @@ static int nouveau_drm_probe(struct pci_dev *pdev,
>>
>> pci_set_master(pdev);
>>
>> + /*
>> + * Set a preliminary DMA mask based on the .dma_bits member of the
>> + * MMU subdevice. This allows other subdevices to create DMA mappings
>> + * in their init() function, which are called before the TTM layer sets
>> + * the DMA mask definitively.
>> + * This is necessary for platforms where the default DMA mask of 32
>> + * does not cover any system memory, i.e., when all RAM is > 4 GB.
>> + */
>> + dma_set_mask_and_coherent(device->dev,
>> + DMA_BIT_MASK(device->mmu->dma_bits));
>
> I would just move this to nvkm_device_pci_new() so that it perfectly
> mirrors the same call done in nvkm_device_tegra_new(), which was done
> for the same purpose. Otherwise, looks good to me.
OK, will do that.