On Fri, Apr 26, 2024 at 3:53 PM Dmitry Baryshkov <[email protected]> wrote: > > On Fri, 26 Apr 2024 at 17:05, Connor Abbott <[email protected]> wrote: > > > > On Fri, Apr 26, 2024 at 2:31 PM Dmitry Baryshkov > > <[email protected]> wrote: > > > > > > On Fri, 26 Apr 2024 at 15:35, Connor Abbott <[email protected]> wrote: > > > > > > > > On Fri, Apr 26, 2024 at 12:02 AM Dmitry Baryshkov > > > > <[email protected]> wrote: > > > > > > > > > > On Thu, 25 Apr 2024 at 16:44, Connor Abbott <[email protected]> > > > > > wrote: > > > > > > > > > > > > On all Qualcomm platforms with a7xx GPUs, qcom_scm provides a > > > > > > method to > > > > > > initialize cx_mem. Copy this from downstream (minus BCL which we > > > > > > currently don't support). On a750, this includes a new "fuse" > > > > > > register > > > > > > which can be used by qcom_scm to fuse off certain features like > > > > > > raytracing in software. The fuse is default off, and is initialized > > > > > > by > > > > > > calling the method. Afterwards we have to read it to find out which > > > > > > features were enabled. > > > > > > > > > > > > Signed-off-by: Connor Abbott <[email protected]> > > > > > > --- > > > > > > drivers/gpu/drm/msm/adreno/a6xx_gpu.c | 89 > > > > > > ++++++++++++++++++++++++- > > > > > > drivers/gpu/drm/msm/adreno/adreno_gpu.h | 2 + > > > > > > 2 files changed, 90 insertions(+), 1 deletion(-) > > > > > > > > > > > > diff --git a/drivers/gpu/drm/msm/adreno/a6xx_gpu.c > > > > > > b/drivers/gpu/drm/msm/adreno/a6xx_gpu.c > > > > > > index cf0b1de1c071..fb2722574ae5 100644 > > > > > > --- a/drivers/gpu/drm/msm/adreno/a6xx_gpu.c > > > > > > +++ b/drivers/gpu/drm/msm/adreno/a6xx_gpu.c > > > > > > @@ -10,6 +10,7 @@ > > > > > > > > > > > > #include <linux/bitfield.h> > > > > > > #include <linux/devfreq.h> > > > > > > +#include <linux/firmware/qcom/qcom_scm.h> > > > > > > #include <linux/pm_domain.h> > > > > > > #include <linux/soc/qcom/llcc-qcom.h> > > > > > > > > > > > > @@ -1686,7 +1687,8 @@ static int a6xx_zap_shader_init(struct > > > > > > msm_gpu *gpu) > > > > > > A6XX_RBBM_INT_0_MASK_RBBM_HANG_DETECT | \ > > > > > > A6XX_RBBM_INT_0_MASK_UCHE_OOB_ACCESS | \ > > > > > > A6XX_RBBM_INT_0_MASK_UCHE_TRAP_INTR | \ > > > > > > - A6XX_RBBM_INT_0_MASK_TSBWRITEERROR) > > > > > > + A6XX_RBBM_INT_0_MASK_TSBWRITEERROR | \ > > > > > > + A6XX_RBBM_INT_0_MASK_SWFUSEVIOLATION) > > > > > > > > > > > > #define A7XX_APRIV_MASK (A6XX_CP_APRIV_CNTL_ICACHE | \ > > > > > > A6XX_CP_APRIV_CNTL_RBFETCH | \ > > > > > > @@ -2356,6 +2358,26 @@ static void a6xx_fault_detect_irq(struct > > > > > > msm_gpu *gpu) > > > > > > kthread_queue_work(gpu->worker, &gpu->recover_work); > > > > > > } > > > > > > > > > > > > +static void a7xx_sw_fuse_violation_irq(struct msm_gpu *gpu) > > > > > > +{ > > > > > > + u32 status; > > > > > > + > > > > > > + status = gpu_read(gpu, REG_A7XX_RBBM_SW_FUSE_INT_STATUS); > > > > > > + gpu_write(gpu, REG_A7XX_RBBM_SW_FUSE_INT_MASK, 0); > > > > > > + > > > > > > + dev_err_ratelimited(&gpu->pdev->dev, "SW fuse violation > > > > > > status=%8.8x\n", status); > > > > > > + > > > > > > + /* Ignore FASTBLEND violations, because the HW will > > > > > > silently fall back > > > > > > + * to legacy blending. > > > > > > + */ > > > > > > + if (status & (A7XX_CX_MISC_SW_FUSE_VALUE_RAYTRACING | > > > > > > + A7XX_CX_MISC_SW_FUSE_VALUE_LPAC)) { > > > > > > + del_timer(&gpu->hangcheck_timer); > > > > > > + > > > > > > + kthread_queue_work(gpu->worker, &gpu->recover_work); > > > > > > + } > > > > > > +} > > > > > > + > > > > > > static irqreturn_t a6xx_irq(struct msm_gpu *gpu) > > > > > > { > > > > > > struct msm_drm_private *priv = gpu->dev->dev_private; > > > > > > @@ -2384,6 +2406,9 @@ static irqreturn_t a6xx_irq(struct msm_gpu > > > > > > *gpu) > > > > > > if (status & A6XX_RBBM_INT_0_MASK_UCHE_OOB_ACCESS) > > > > > > dev_err_ratelimited(&gpu->pdev->dev, "UCHE | Out of > > > > > > bounds access\n"); > > > > > > > > > > > > + if (status & A6XX_RBBM_INT_0_MASK_SWFUSEVIOLATION) > > > > > > + a7xx_sw_fuse_violation_irq(gpu); > > > > > > + > > > > > > if (status & A6XX_RBBM_INT_0_MASK_CP_CACHE_FLUSH_TS) > > > > > > msm_gpu_retire(gpu); > > > > > > > > > > > > @@ -2525,6 +2550,60 @@ static void a6xx_llc_slices_init(struct > > > > > > platform_device *pdev, > > > > > > a6xx_gpu->llc_mmio = ERR_PTR(-EINVAL); > > > > > > } > > > > > > > > > > > > +static int a7xx_cx_mem_init(struct a6xx_gpu *a6xx_gpu) > > > > > > +{ > > > > > > + struct adreno_gpu *adreno_gpu = &a6xx_gpu->base; > > > > > > + struct msm_gpu *gpu = &adreno_gpu->base; > > > > > > + u32 gpu_req = QCOM_SCM_GPU_ALWAYS_EN_REQ; > > > > > > + u32 fuse_val; > > > > > > + int ret; > > > > > > + > > > > > > + if (adreno_is_a740(adreno_gpu)) { > > > > > > + /* Raytracing is always enabled on a740 */ > > > > > > + adreno_gpu->has_ray_tracing = true; > > > > > > + } > > > > > > + > > > > > > + if (!qcom_scm_is_available()) { > > > > > > + /* Assume that if qcom scm isn't available, that > > > > > > whatever > > > > > > + * replacement allows writing the fuse register > > > > > > ourselves. > > > > > > + * Users of alternative firmware need to make sure > > > > > > this > > > > > > + * register is writeable or indicate that it's not > > > > > > somehow. > > > > > > + * Print a warning because if you mess this up > > > > > > you're about to > > > > > > + * crash horribly. > > > > > > + */ > > > > > > + if (adreno_is_a750(adreno_gpu)) { > > > > > > + dev_warn_once(gpu->dev->dev, > > > > > > + "SCM is not available, poking fuse > > > > > > register\n"); > > > > > > + a6xx_llc_write(a6xx_gpu, > > > > > > REG_A7XX_CX_MISC_SW_FUSE_VALUE, > > > > > > + > > > > > > A7XX_CX_MISC_SW_FUSE_VALUE_RAYTRACING | > > > > > > + > > > > > > A7XX_CX_MISC_SW_FUSE_VALUE_FASTBLEND | > > > > > > + A7XX_CX_MISC_SW_FUSE_VALUE_LPAC); > > > > > > + adreno_gpu->has_ray_tracing = true; > > > > > > + } > > > > > > + > > > > > > + return 0; > > > > > > + } > > > > > > + > > > > > > + if (adreno_is_a750(adreno_gpu)) > > > > > > > > > > Most of the function is under the if (adreno_is_a750) conditions. Can > > > > > we invert the logic and add a single block of if(adreno_is_a750) and > > > > > then place all the code underneath? > > > > > > > > You mean to duplicate the qcom_scm_is_available check and qcom_scm_ > > > > > > > > > > > > > > > + gpu_req |= QCOM_SCM_GPU_TSENSE_EN_REQ; > > > > > > + > > > > > > + ret = qcom_scm_gpu_init_regs(gpu_req); > > > > > > + if (ret) > > > > > > + return ret; > > > > > > + > > > > > > + /* On a750 raytracing may be disabled by the firmware, find > > > > > > out whether > > > > > > + * that's the case. The scm call above sets the fuse > > > > > > register. > > > > > > + */ > > > > > > + if (adreno_is_a750(adreno_gpu)) { > > > > > > + fuse_val = a6xx_llc_read(a6xx_gpu, > > > > > > REG_A7XX_CX_MISC_SW_FUSE_VALUE); > > > > > > > > > > This register isn't accessible with the current sm8650.dtsi. Since DT > > > > > and driver are going through different trees, please add safety guards > > > > > here, so that the driver doesn't crash if used with older dtsi > > > > > > > > I don't see how this is an issue. msm-next is currently based on 6.9, > > > > which doesn't have the GPU defined in sm8650.dtsi. AFAIK patches 1 and > > > > 2 will have to go through the linux-arm-msm tree, which will have to > > > > be merged into msm-next before this patch lands there, so there will > > > > never be any breakage. > > > > > > linux-arm-msm isn't going to be merged into msm-next. If we do not ask > > > for ack for the fix to go through msm-next, they will get these > > > patches in parallel. > > > > I'm not familiar with how complicated cross-tree changes like this get > > merged, but why would we merge these in parallel given that this patch > > depends on the previous patch that introduces > > qcom_scm_gpu_init_regs(), and that would (I assume?) normally go > > through the same tree as patch 1? Even if patch 1 gets merged in > > parallel in linux-arm-msm, in what scenario would we have a broken > > boot? You won't have a devicetree with a working sm8650 GPU and > > drm/msm with raytracing until linux-arm-msm is merged into msm-next at > > which point patch 1 will have landed somehow. > > arch/arm64/qcom/dts and drivers/firmware/qcom are two separate trees. > So yes, this needs a lot of coordination.
> > > > > > > > > Another option is to get dtsi fix into 6.9 and delay the raytracing > > > until 6.10-rc which doesn't make a lot of sense from my POV). > > > > > > > > > > > > (not to mention that dts is considered to be an ABI and newer kernels > > > > > are supposed not to break with older DT files). > > > > > > > > That policy only applies to released kernels, so that's irrelevant here. > > > > > > It applies to all kernels, the reason being pretty simple: git-bisect > > > should not be broken. > > > > As I wrote above, this is not an issue. The point I was making is that > > mixing and matching dtb's from one unmerged subsystem tree and a > > kernel from another isn't supported AFAIK, and that's the only > > scenario where this could break. > > And it can happen if somebody running a bisect ends up in the branch > with these patches in, but with the dtsi bits not being picked up. That wouldn't be possible unless we merged the "bad" commit introducing the GPU node to sm8650.dtsi into msm-next but not the fix. So yeah, it's going to require a lot of careful cooperation but it should be possible to avoid that happening. Connor
