[PATCH v3 02/15] drm/rockchip: Set dma mask to 64 bit
Robin Murphy
robin.murphy at arm.com
Wed Oct 16 17:38:23 UTC 2024
On 2024-09-20 9:20 am, Andy Yan wrote:
> From: Andy Yan <andy.yan at rock-chips.com>
>
> The vop mmu support translate physical address upper 4 GB to iova
> below 4 GB. So set dma mask to 64 bit to indicate we support address
>> 4GB.
>
> This can avoid warnging message like this on some boards with DDR
>> 4 GB:
>
> rockchip-drm display-subsystem: swiotlb buffer is full (sz: 266240 bytes), total 32768 (slots), used 130 (slots)
> rockchip-drm display-subsystem: swiotlb buffer is full (sz: 266240 bytes), total 32768 (slots), used 0 (slots)
> rockchip-drm display-subsystem: swiotlb buffer is full (sz: 266240 bytes), total 32768 (slots), used 130 (slots)
> rockchip-drm display-subsystem: swiotlb buffer is full (sz: 266240 bytes), total 32768 (slots), used 130 (slots)
> rockchip-drm display-subsystem: swiotlb buffer is full (sz: 266240 bytes), total 32768 (slots), used 0 (slots)
There are several things wrong with this...
AFAICS the VOP itself still only supports 32-bit addresses, so the VOP
driver should only be setting a 32-bit DMA mask. The IOMMUs support
either 32-bit or 40-bit addresses, and the IOMMU driver does set its DMA
mask appropriately. None of those numbers is 64, so that's clearly
suspicious already. Plus it would seem the claim of the IOMMU being able
to address >4GB isn't strictly true for RK3288 (which does supposedly
support 8GB of RAM).
Furthermore, the "display-subsystem" doesn't even exist - it does not
represent any actual DMA-capable hardware, so it should not have a DMA
mask, and it should not be used for DMA API operations. Buffers for the
VOP should be DMA-mapped for the VOP device itself. At the very least,
the rockchip_gem_alloc_dma() path is clearly broken otherwise (I guess
this patch possibly *would* make that brokenness apparent).
> Signed-off-by: Andy Yan <andy.yan at rock-chips.com>
> Tested-by: Derek Foreman <derek.foreman at collabora.com>
> ---
>
> (no changes since v1)
>
> drivers/gpu/drm/rockchip/rockchip_drm_drv.c | 4 +++-
> 1 file changed, 3 insertions(+), 1 deletion(-)
>
> diff --git a/drivers/gpu/drm/rockchip/rockchip_drm_drv.c b/drivers/gpu/drm/rockchip/rockchip_drm_drv.c
> index 04ef7a2c3833..8bc2ff3b04bb 100644
> --- a/drivers/gpu/drm/rockchip/rockchip_drm_drv.c
> +++ b/drivers/gpu/drm/rockchip/rockchip_drm_drv.c
> @@ -445,7 +445,9 @@ static int rockchip_drm_platform_probe(struct platform_device *pdev)
> return ret;
> }
>
> - return 0;
> + ret = dma_coerce_mask_and_coherent(dev, DMA_BIT_MASK(64));
Finally as a general thing, please don't misuse
dma_coerce_mask_and_coherent() in platform drivers, just use normal
dma_set_mask_and_coherent(). The platform bus code has been initialising
the dev->dma_mask pointer for years now, drivers should not be messing
with it any more.
Thanks,
Robin.
> +
> + return ret;
> }
>
> static void rockchip_drm_platform_remove(struct platform_device *pdev)
More information about the dri-devel
mailing list