[PATCH v3 0/3] drm/etnaviv: Trivial mmu map and ummap cleanups

[Lucas Stach]

Am Freitag, dem 15.11.2024 um 20:32 +0800 schrieb Sui Jingfeng:
> The 'sg->offset' denotes the offset into a page in bytes, but under drm
> subsystem, there has NO drivers that etnaviv can contact that actually
> touch the 'offset' data member of SG anymore. This means that all DMA
> addresses that sg_dma_address() gives us will be PAGE_SIZE aligned, in
> other words, sg->offset will always equal to 0.
> 
> But if 'sg->offset != 0' really could happens, then the current implement
> might be not correct. Previous commits[1] fix the 'sg->offset == 0' cases
> effectively, below is a simple illustration.
> 
>        One CPU page       Another one CPU page
>   +----+----+----+----+   +----+----+----+----+
>   ||||||              |   ||||||              |
>   +----+----+----+----+   +----+----+----+----+
>   ^    ^                  ^    ^
>   |    |                  |    |
>   |    | .----------------'    |
>   |    | |    .----------------'
>   |    | |    |
>   +----+ +----+ +----+
>   |||||| |||||| |    |  GPU pages, each one is SZ_4K
>   +----+ +----+ +----+
>             Correct implementation.
> 
> --------------------------------------------------------------
> 
>        One CPU page       Another one CPU page
>   +----+----+----+----+   +----+----+----+----+
>   |///////////////////|   ||||||              |
>   +----+----+----+----+   +----+----+----+----+
>   ^                   ^   ^    ^
>   |                   |   |    |
>   |      .------------|---'    |
>   |      |    .-------|--------'
>   |      |    |       |
>   |      +----+       |
>   |      ||||||       |
>   |      +----+       |
>   |       IOVA        |  GPUVA range collision if use 'sg_dma_len(sg)'
>   +----+ +----+-------+  directly to map. Because 'sg_dma_len(sg)' is
>   |////|/|////////////|  frequently larger than SZ_4K.
>   +----+ +----+-------+
>             Wrong implementation.
> 
> If we map the address range with respect to the size of the backing memory,
> it will occupy GPUVA ranges that doesn't belong to. Which results in GPUVA
> range collision for different buffers.
> 
> Patch 0001 of this series give a fix, patch 0002 and 0003
> do trivial cleanup which eliminates unnecessary overheads.

Thanks, applied to etnaviv/next.

Regards,
Lucas

> v2 -> v3
> 	* Reword and improve commit message
> v1 -> v2
> 	* Reword and fix typos and mistakes
> 
> v1 Link: https://patchwork.freedesktop.org/series/140589/
> 
> Sui Jingfeng (3):
>   drm/etnaviv: Drop the offset in page manipulation
>   drm/etnaviv: Fix the debug log  of the etnaviv_iommu_map()
>   drm/etnaviv: Improve VA, PA, SIZE alignment checking
> 
>  drivers/gpu/drm/etnaviv/etnaviv_mmu.c | 26 +++++++++++---------------
>  1 file changed, 11 insertions(+), 15 deletions(-)
>