rockchip64-6.14: Set dma mask to 64 bit

[SuperKali]

Description

This patch sets the DMA mask to 64-bit for Rockchip DRM drivers to properly handle physical addresses above 4GB. The VOP MMU can translate upper 4GB physical addresses to iova below 4GB, resolving "swiotlb buffer full" warnings seen on systems with >4GB RAM.

Source: https://source.mnt.re/reform/reform-debian-packages/-/blob/487191f7ad5c48850c22b3fe8d4d9187f21945e5/linux/patches6.11/rk3588-mnt-reform2/5200-drm-rockchip-Set-dma-mask-to-64-bit.patch

Motivation: Fixes graphical subsystem warnings that occur when handling large memory allocations on Rockchip devices with >4GB DDR.

How Has This Been Tested?

• Verified on RK3588-based SBC with 8GB RAM
• Confirmed disappearance of swiotlb warnings via dmesg monitoring

Checklist:

• My code follows the style guidelines of this project
• I have performed a self-review of my own code
• My changes generate no new warnings

[Tonymac32]

Please test on RK3399 and/or RK3328. I assume RK3566/8 will behave like RK3588

[SuperKali]

Please test on RK3399

Unfortunately I don't possess any board with those soc

[paolosabatino]

Perhaps today I can give a shot to rk3328/rk3399 boards

[paolosabatino]

Quick check on OrangePi 4 LTS (rk3399) and a rk3318 tvbox shows no harm: both present basic DRM console, XFCE works in both composite and non-composite modes, OpenGL works fine (tested quake) and dmesg is clean.

The patch seems ok to me

[lanefu]

My unscientific testing on glmark2 on pinebook pro (rk3399) shows a bit of a regression.

before1: 397
before2: 400
after1: 381
after2: 386

[paolosabatino]

My unscientific testing on glmark2 on pinebook pro (rk3399) shows a bit of a regression.

before1: 397 before2: 400 after1: 381 after2: 386

Weird, I felt "something" while watching Quake demo going on... Anyway I guess the patch as-is cannot be mainlined in the kernel because IMHO it would break 32bit, wouldn't it? A nicer way to do that would be enable only for those SoCs which can address more than 4gb of DRAM.

This is only for edge kernel though, so there is time for refinement and testing.

[paolosabatino]

Did a benchmark with glmark2 on rk3399, no regression on my side:

• 360 points with kernel 6.12 and no patch
• 366 points with kernel 6.14 and this patch applied