On Mon, Nov 15, 2021 at 08:18:53AM +0100, Thomas Hellström wrote:
On 11/14/21 22:19, Linus Torvalds wrote:
On Sun, Nov 14, 2021 at 1:00 PM Dave Airlie airlied@gmail.com wrote:
i915 will no longer be x86-64 only in theory, since Intel now produces PCIe graphics cards using the same hw designs.
Well, at least in my tree, it still has the "depends on X86", along with several other x86-only things (like "select INTEL_GTT", which is also x86-only)
Yeah it's work in progress and 12+ years of x86 pile up high&deep aren't easy to fix. So please assume the depends on X86 is gone already (and select INTEL_GTT is optional, it's only needed for igfx older than about 8 years or so), but we can't do that yet because all the build boots would run out of air screaming so much :-)
So by the time that non-x86 theory becomes reality, hopefully the i915 people will also have figured out how to do the cache flushing properly.
And hopefully that "do it properly" ends up being simply that the particular configuration that ends up being portable simply doesn't need to do it at all and can statically just not build it, sidestepping the issue entirely.
Fingers crossed.
For non-x86 / discrete graphics, plan is only coherent mappings, although the "Just not build it" part hasn't been properly figured out yet I guess. But point taken.
Yeah for non-x86 it'll be standard dma-api all the way down. For x86 that ship sailed long ago, but also I'm not clear on why we added a new cache flush primitive for that instead of just continuing to use clflush like we've done since forever. For x86 integrated gpu we _know_ which cpu is there, clflush will exist. So any new horrors shouldn't be needed, not sure why a wbinvd_on_all_cpus slipped in. -Daniel