On Tue, May 20, 2025 at 2:25 PM Dave Airlie <airl...@gmail.com> wrote: > > On Sat, 17 May 2025 at 02:20, Rob Clark <robdcl...@gmail.com> wrote: > > > > On Fri, May 16, 2025 at 2:01 AM Danilo Krummrich <d...@kernel.org> wrote: > > > > > > On Thu, May 15, 2025 at 02:57:46PM -0700, Rob Clark wrote: > > > > On Thu, May 15, 2025 at 10:55 AM Danilo Krummrich <d...@kernel.org> > > > > wrote: > > > > > Anyways, I don't agree with that. Even if you can tweak your driver > > > > > to not run > > > > > into trouble with this, we can't introduce a mode that violates > > > > > GOUVM's internal > > > > > lifetimes and subsequently fix it up with WARN_ON() or BUG_ON(). > > > > > > > > > > I still don't see a real technical reason why msm can't be reworked > > > > > to follow > > > > > those lifetime rules. > > > > > > > > The basic issue is that (a) it would be really awkward to have two > > > > side-by-side VM/VMA management/tracking systems. But in legacy mode, > > > > we have the opposite direction of reference holding. (But at the same > > > > time, don't need/use most of the features of gpuvm.) > > > > > > Ok, let's try to move this forward; I see three options (in order of > > > descending > > > preference): > > > > > > 1) Rework the legacy code to properly work with GPUVM. > > > 2) Don't use GPUVM for the legacy mode. > > > . > > > . > > > . > > > 3) Get an ACK from Dave / Sima to implement those workarounds for MSM in > > > GPUVM. > > > > > > If you go for 3), the code introduced by those two patches should be > > > guarded > > > with a flag that makes it very clear that this is a workaround > > > specifically > > > for MSM legacy mode and does not give any guarantees in terms of > > > correctness > > > regarding lifetimes etc., e.g. DRM_GPUVM_MSM_LEGACY_QUIRK. > > > > I'm not even sure how #2 would work, other than just copy/pasta all of > > drm_gpuvm into msm, which doesn't really seem great. > > > > As for #1, even if I could get it to work, it would still be a lot > > more mmu map/unmap (like on every pageflip, vs the current state that > > the vma is kept around until the object is freed). For the > > non-VM_BIND world, there are advantages to the BO holding the ref to > > the VMA, rather than the other way around. Even at just a modest > > single layer 1080p the map takes ~.2ms and unmap ~.3ms (plus the unmap > > costs a tlbinv). So from that standpoint, #3 is the superior option. > > > > Before we get to #3, I'll need a bit more info here on why you have to > map/unmap the VMA on every pageflip.
Previously we'd keep the VMA hanging around until the GEM obj is freed. But that can't work if the VMA (via the VM_BO) is holding a reference to the GEM obj. I was kinda thinking about keeping the VMA around until the handle is closed.. but that doesn't cover the dma-buf case (ie. when you re-import the dma-buf fd each frame.. I know android does this, unsure about other wsi's). > But actually I think 2 is the best option, I think in nouveau this is > where we ended up, we didn't modify the old submission paths at all > and kept the old bo/vm lifetimes. > > We just added completely new bind/exec ioctls and you can only use one > method once you've opened an fd. hmm, but that means tracking VMAs against a single BO differently.. which.. at least seems ugly.. BR, -R
