AMD’s “driver problem” complaints are one thing the corporate is properly conscious of, and it’s definitely making an lively effort to fight them in fascinating and revolutionary methods, like utilizing a brand new AI software, for instance. Surprisingly, maybe, it has been Nvidia that has been the one coping with extra driver-related in addition to different issues with its present gen RTX 50 sequence playing cards.

So whereas Nvidia continues coping with issues on Home windows, an engineer from there helped resolve an AMD driver bug on Linux. A latest patch notice reveals this because it says:

Repair a efficiency regression on AMD iGPU and dGPU drivers, associated to the unintended activation of DMA bounce buffers that regressed recreation efficiency if KASLR disturbed issues simply sufficient.

Apparently, this problem, because it seems, was the consequence of a earlier “dangerous commit” by an Nvidia engineer. It was capable of be pinned down due to the trouble by Bert Karwatzki, who observed issues when taking part in Stellaris on Steam. He wrote:

Utilizing linux next-20250307 to play the sport stellaris by way of steam I observed that loading the sport will get sluggish with the progress bar getting caught at 100%. On this scenario mouse and keyboard inputs do not work correctly anymore. Switching to a VT and killing stellaris considerably fixes the scenario although in a single occasion the touchpad didn’t work after that. I bisected this between v6.14-rc5 and next-20250307 and bought this as the primary dangerous commit

….

Reverting commit 7ffb791423c7 in next-20250307 fixes the problem for me. The OS is debian sid (final up to date 20250309) and that is the {hardware} is an MSI Alpha 15 Laptop computer

Upon investigating additional, it was understood that the bug was a results of a problem with the kernel deal with area structure randomization (KASLR) function when it was disabled (nokaslr), which results in a DMA (direct reminiscence entry) addressing error. For these questioning, KASLR is a safety function that helps load the kernel to a random location in reminiscence and is supposed for reminiscence security.

The patch notes say:

As Bert Karwatzki reported, the next latest commit causes a efficiency regression on AMD iGPU and dGPU techniques:

7ffb791423c7 (“x86/kaslr: Cut back KASLR entropy on most x86 techniques”)

It uncovered a bug with nokaslr and zone machine interplay. The basis reason behind the bug is that, the GPU driver registers a zone machine personal reminiscence area. When KASLR is disabled or the above commit is utilized, the direct_map_physmem_end is ready to a lot greater than 10 TiB usually to the 64TiB deal with.

When zone machine personal reminiscence is added to the system by way of add_pages(), it bumps up the max_pfn to the identical worth. This causes dma_addressing_limited() to return true, because the machine can not deal with reminiscence all the way in which as much as max_pfn.

This prompted a regression for video games performed on the iGPU, because it resulted within the DMA32 zone getting used for GPU allocations.

You may learn them in full on the supply LKML (Linux Kernel Mailing Listing) hyperlinks under.

Supply: LKML (link1, link2, link3)