by Akhil P Oommen

[permalink] [raw]

Subject: Re: [PATCH v5 06/15] drm/msm/a6xx: Introduce GMU wrapper support

On Fri, Mar 31, 2023 at 01:25:20AM +0200, Konrad Dybcio wrote:
> Some (particularly SMD_RPM, a.k.a non-RPMh) SoCs implement A6XX GPUs
> but don't implement the associated GMUs. This is due to the fact that
> the GMU directly pokes at RPMh. Sadly, this means we have to take care
> of enabling & scaling power rails, clocks and bandwidth ourselves.
>
> Reuse existing Adreno-common code and modify the deeply-GMU-infused
> A6XX code to facilitate these GPUs. This involves if-ing out lots
> of GMU callbacks and introducing a new type of GMU - GMU wrapper (it's
> the actual name that Qualcomm uses in their downstream kernels).
>
> This is essentially a register region which is convenient to model
> as a device. We'll use it for managing the GDSCs. The register
> layout matches the actual GMU_CX/GX regions on the "real GMU" devices
> and lets us reuse quite a bit of gmu_read/write/rmw calls.

Commenting here after going through rest of the patch...

Only convenience I see with modeling a dummy gmu is that we can reuse gmu read/write routines which I think would be less that 10 instances. If we just add a gmu_wrapper region to gpu node, wouldn't that help to create a clean separation between gmu-supported vs gmu-wrapper/no-gmu architectures? Also, creating a dummy gmu device in device tree doesn't sound right to me.

>
> Signed-off-by: Konrad Dybcio <[email protected]>
> ---
> drivers/gpu/drm/msm/adreno/a6xx_gmu.c | 72 +++++++-
> drivers/gpu/drm/msm/adreno/a6xx_gpu.c | 254 +++++++++++++++++++++++++---
> drivers/gpu/drm/msm/adreno/a6xx_gpu.h | 1 +
> drivers/gpu/drm/msm/adreno/a6xx_gpu_state.c | 14 +-
> drivers/gpu/drm/msm/adreno/adreno_gpu.c | 8 +-
> drivers/gpu/drm/msm/adreno/adreno_gpu.h | 6 +
> 6 files changed, 317 insertions(+), 38 deletions(-)
>
> diff --git a/drivers/gpu/drm/msm/adreno/a6xx_gmu.c b/drivers/gpu/drm/msm/adreno/a6xx_gmu.c
> index 1514b3ed0fcf..c6001e82e03d 100644
> --- a/drivers/gpu/drm/msm/adreno/a6xx_gmu.c
> +++ b/drivers/gpu/drm/msm/adreno/a6xx_gmu.c
> @@ -1474,6 +1474,7 @@ static int a6xx_gmu_get_irq(struct a6xx_gmu *gmu, struct platform_device *pdev,
>
> void a6xx_gmu_remove(struct a6xx_gpu *a6xx_gpu)
> {
> + struct adreno_gpu *adreno_gpu = &a6xx_gpu->base;
> struct a6xx_gmu *gmu = &a6xx_gpu->gmu;
> struct platform_device *pdev = to_platform_device(gmu->dev);
>
> @@ -1499,10 +1500,12 @@ void a6xx_gmu_remove(struct a6xx_gpu *a6xx_gpu)
> gmu->mmio = NULL;
> gmu->rscc = NULL;
>
> - a6xx_gmu_memory_free(gmu);
> + if (!adreno_has_gmu_wrapper(adreno_gpu)) {
> + a6xx_gmu_memory_free(gmu);
>
> - free_irq(gmu->gmu_irq, gmu);
> - free_irq(gmu->hfi_irq, gmu);
> + free_irq(gmu->gmu_irq, gmu);
> + free_irq(gmu->hfi_irq, gmu);
> + }
>
> /* Drop reference taken in of_find_device_by_node */
> put_device(gmu->dev);
> @@ -1521,6 +1524,69 @@ static int cxpd_notifier_cb(struct notifier_block *nb,
> return 0;
> }
>
> +int a6xx_gmu_wrapper_init(struct a6xx_gpu *a6xx_gpu, struct device_node *node)
> +{
> + struct platform_device *pdev = of_find_device_by_node(node);
> + struct a6xx_gmu *gmu = &a6xx_gpu->gmu;
> + int ret;
> +
> + if (!pdev)
> + return -ENODEV;
> +
> + gmu->dev = &pdev->dev;
> +
> + of_dma_configure(gmu->dev, node, true);
If GMU is dummy, why should we configure dma?
> +
> + pm_runtime_enable(gmu->dev);
> +
> + /* Mark legacy for manual SPTPRAC control */
> + gmu->legacy = true;
> +
> + /* Map the GMU registers */
> + gmu->mmio = a6xx_gmu_get_mmio(pdev, "gmu");
> + if (IS_ERR(gmu->mmio)) {
> + ret = PTR_ERR(gmu->mmio);
> + goto err_mmio;
> + }
> +
> + gmu->cxpd = dev_pm_domain_attach_by_name(gmu->dev, "cx");
> + if (IS_ERR(gmu->cxpd)) {
> + ret = PTR_ERR(gmu->cxpd);
> + goto err_mmio;
> + }
> +
> + if (!device_link_add(gmu->dev, gmu->cxpd, DL_FLAG_PM_RUNTIME)) {
> + ret = -ENODEV;
> + goto detach_cxpd;
> + }
> +
> + init_completion(&gmu->pd_gate);
> + complete_all(&gmu->pd_gate);
> + gmu->pd_nb.notifier_call = cxpd_notifier_cb;
> +
> + /* Get a link to the GX power domain to reset the GPU */
> + gmu->gxpd = dev_pm_domain_attach_by_name(gmu->dev, "gx");
> + if (IS_ERR(gmu->gxpd)) {
> + ret = PTR_ERR(gmu->gxpd);
> + goto err_mmio;
> + }
> +
> + gmu->initialized = true;
> +
> + return 0;
> +
> +detach_cxpd:
> + dev_pm_domain_detach(gmu->cxpd, false);
> +
> +err_mmio:
> + iounmap(gmu->mmio);
> +
> + /* Drop reference taken in of_find_device_by_node */
> + put_device(gmu->dev);
> +
> + return ret;
> +}
> +
> int a6xx_gmu_init(struct a6xx_gpu *a6xx_gpu, struct device_node *node)
> {
> struct adreno_gpu *adreno_gpu = &a6xx_gpu->base;
> diff --git a/drivers/gpu/drm/msm/adreno/a6xx_gpu.c b/drivers/gpu/drm/msm/adreno/a6xx_gpu.c
> index ae0a90b2834f..a7ecb0a87e98 100644
> --- a/drivers/gpu/drm/msm/adreno/a6xx_gpu.c
> +++ b/drivers/gpu/drm/msm/adreno/a6xx_gpu.c
> @@ -20,9 +20,11 @@ static inline bool _a6xx_check_idle(struct msm_gpu *gpu)
> struct adreno_gpu *adreno_gpu = to_adreno_gpu(gpu);
> struct a6xx_gpu *a6xx_gpu = to_a6xx_gpu(adreno_gpu);
>
> - /* Check that the GMU is idle */
> - if (!a6xx_gmu_isidle(&a6xx_gpu->gmu))
> - return false;
> + if (!adreno_has_gmu_wrapper(adreno_gpu)) {
> + /* Check that the GMU is idle */
> + if (!a6xx_gmu_isidle(&a6xx_gpu->gmu))
> + return false;
> + }
>
> /* Check tha the CX master is idle */
> if (gpu_read(gpu, REG_A6XX_RBBM_STATUS) &
> @@ -612,13 +614,15 @@ static void a6xx_set_hwcg(struct msm_gpu *gpu, bool state)
> return;
>
> /* Disable SP clock before programming HWCG registers */
> - gmu_rmw(gmu, REG_A6XX_GPU_GMU_GX_SPTPRAC_CLOCK_CONTROL, 1, 0);
> + if (!adreno_has_gmu_wrapper(adreno_gpu))
> + gmu_rmw(gmu, REG_A6XX_GPU_GMU_GX_SPTPRAC_CLOCK_CONTROL, 1, 0);
>
> for (i = 0; (reg = &adreno_gpu->info->hwcg[i], reg->offset); i++)
> gpu_write(gpu, reg->offset, state ? reg->value : 0);
>
> /* Enable SP clock */
> - gmu_rmw(gmu, REG_A6XX_GPU_GMU_GX_SPTPRAC_CLOCK_CONTROL, 0, 1);
> + if (!adreno_has_gmu_wrapper(adreno_gpu))
> + gmu_rmw(gmu, REG_A6XX_GPU_GMU_GX_SPTPRAC_CLOCK_CONTROL, 0, 1);
>
> gpu_write(gpu, REG_A6XX_RBBM_CLOCK_CNTL, state ? clock_cntl_on : 0);
> }
> @@ -1002,10 +1006,13 @@ static int hw_init(struct msm_gpu *gpu)
> {
> struct adreno_gpu *adreno_gpu = to_adreno_gpu(gpu);
> struct a6xx_gpu *a6xx_gpu = to_a6xx_gpu(adreno_gpu);
> + struct a6xx_gmu *gmu = &a6xx_gpu->gmu;
> int ret;
>
> - /* Make sure the GMU keeps the GPU on while we set it up */
> - a6xx_gmu_set_oob(&a6xx_gpu->gmu, GMU_OOB_GPU_SET);
> + if (!adreno_has_gmu_wrapper(adreno_gpu)) {
> + /* Make sure the GMU keeps the GPU on while we set it up */
> + a6xx_gmu_set_oob(&a6xx_gpu->gmu, GMU_OOB_GPU_SET);
> + }
>
> /* Clear GBIF halt in case GX domain was not collapsed */
> if (a6xx_has_gbif(adreno_gpu))
> @@ -1128,6 +1135,17 @@ static int hw_init(struct msm_gpu *gpu)
> 0x3f0243f0);
> }
>
> + if (adreno_has_gmu_wrapper(adreno_gpu)) {
> + /* Do it here, as GMU wrapper only inits the GMU for memory reservation etc. */
> +
> + /* Set up the CX GMU counter 0 to count busy ticks */
> + gmu_write(gmu, REG_A6XX_GPU_GMU_AO_GPU_CX_BUSY_MASK, 0xff000000);
> +
> + /* Enable power counter 0 */
> + gmu_rmw(gmu, REG_A6XX_GMU_CX_GMU_POWER_COUNTER_SELECT_0, 0xff, BIT(5));
> + gmu_write(gmu, REG_A6XX_GMU_CX_GMU_POWER_COUNTER_ENABLE, 1);
> + }
> +
> /* Protect registers from the CP */
> a6xx_set_cp_protect(gpu);
>
> @@ -1236,6 +1254,8 @@ static int hw_init(struct msm_gpu *gpu)
> }
>
> out:
> + if (adreno_has_gmu_wrapper(adreno_gpu))
> + return ret;
> /*
> * Tell the GMU that we are done touching the GPU and it can start power
> * management
> @@ -1270,6 +1290,9 @@ static void a6xx_dump(struct msm_gpu *gpu)
> adreno_dump(gpu);
> }
>
> +#define GBIF_GX_HALT_MASK BIT(0)
> +#define GBIF_CLIENT_HALT_MASK BIT(0)
> +#define GBIF_ARB_HALT_MASK BIT(1)
> #define VBIF_RESET_ACK_TIMEOUT 100
> #define VBIF_RESET_ACK_MASK 0x00f0
>
> @@ -1302,7 +1325,8 @@ static void a6xx_recover(struct msm_gpu *gpu)
> * Turn off keep alive that might have been enabled by the hang
> * interrupt
> */
> - gmu_write(&a6xx_gpu->gmu, REG_A6XX_GMU_GMU_PWR_COL_KEEPALIVE, 0);
> + if (!adreno_has_gmu_wrapper(adreno_gpu))
> + gmu_write(&a6xx_gpu->gmu, REG_A6XX_GMU_GMU_PWR_COL_KEEPALIVE, 0);

Perhaps, we can move this register write to the begining of a6xx_gmu_force_off(). That looks cleaner to me.

>
> pm_runtime_dont_use_autosuspend(&gpu->pdev->dev);
>
> @@ -1332,6 +1356,32 @@ static void a6xx_recover(struct msm_gpu *gpu)
>
> dev_pm_genpd_remove_notifier(gmu->cxpd);
>
> + /* Software-reset the GPU */
This is not soft reset. We are quiescenting gpu to ddr traffic before we
pull the regulator
> + if (adreno_has_gmu_wrapper(adreno_gpu)) {
> + /* Halt the GX side of GBIF */
> + gpu_write(gpu, REG_A6XX_RBBM_GBIF_HALT, GBIF_GX_HALT_MASK);
> + spin_until(gpu_read(gpu, REG_A6XX_RBBM_GBIF_HALT_ACK) &
> + GBIF_GX_HALT_MASK);
> +
> + /* Halt new client requests on GBIF */
> + gpu_write(gpu, REG_A6XX_GBIF_HALT, GBIF_CLIENT_HALT_MASK);
> + spin_until((gpu_read(gpu, REG_A6XX_GBIF_HALT_ACK) &
> + (GBIF_CLIENT_HALT_MASK)) == GBIF_CLIENT_HALT_MASK);
> +
> + /* Halt all AXI requests on GBIF */
> + gpu_write(gpu, REG_A6XX_GBIF_HALT, GBIF_ARB_HALT_MASK);
> + spin_until((gpu_read(gpu, REG_A6XX_GBIF_HALT_ACK) &
> + (GBIF_ARB_HALT_MASK)) == GBIF_ARB_HALT_MASK);
> +
> + /* Clear the halts */
> + gpu_write(gpu, REG_A6XX_GBIF_HALT, 0);
> +
> + gpu_write(gpu, REG_A6XX_RBBM_GBIF_HALT, 0);
> +
> + /* This *really* needs to go through before we do anything else! */
> + mb();
> + }

This sequence should be before we collapse the regulators and clocks. Can't we handle this similar
to the gmu ops in the pm_suspend cb? Also, We can reuse a6xx_bus_clear_pending_transactions() for this sequence.

> +
> pm_runtime_use_autosuspend(&gpu->pdev->dev);
>
> if (active_submits)
> @@ -1516,7 +1566,8 @@ static void a6xx_fault_detect_irq(struct msm_gpu *gpu)
> * Force the GPU to stay on until after we finish
> * collecting information
> */
> - gmu_write(&a6xx_gpu->gmu, REG_A6XX_GMU_GMU_PWR_COL_KEEPALIVE, 1);
> + if (!adreno_has_gmu_wrapper(adreno_gpu))
> + gmu_write(&a6xx_gpu->gmu, REG_A6XX_GMU_GMU_PWR_COL_KEEPALIVE, 1);

It would be more readable if we wrap this into a subroutine.

>
> DRM_DEV_ERROR(&gpu->pdev->dev,
> "gpu fault ring %d fence %x status %8.8X rb %4.4x/%4.4x ib1 %16.16llX/%4.4x ib2 %16.16llX/%4.4x\n",
> @@ -1677,7 +1728,7 @@ static void a6xx_llc_slices_init(struct platform_device *pdev,
> a6xx_gpu->llc_mmio = ERR_PTR(-EINVAL);
> }
>
> -static int a6xx_pm_resume(struct msm_gpu *gpu)
> +static int a6xx_gmu_pm_resume(struct msm_gpu *gpu)
> {
> struct adreno_gpu *adreno_gpu = to_adreno_gpu(gpu);
> struct a6xx_gpu *a6xx_gpu = to_a6xx_gpu(adreno_gpu);
> @@ -1697,10 +1748,61 @@ static int a6xx_pm_resume(struct msm_gpu *gpu)
>
> a6xx_llc_activate(a6xx_gpu);
>
> - return 0;
> + return ret;
Isn't ret uninitialized here?
> }
>
> -static int a6xx_pm_suspend(struct msm_gpu *gpu)
> +static int a6xx_pm_resume(struct msm_gpu *gpu)
> +{
> + struct adreno_gpu *adreno_gpu = to_adreno_gpu(gpu);
> + struct a6xx_gpu *a6xx_gpu = to_a6xx_gpu(adreno_gpu);
> + struct a6xx_gmu *gmu = &a6xx_gpu->gmu;
> + unsigned long freq = 0;
> + struct dev_pm_opp *opp;
> + int ret;
> +
> + gpu->needs_hw_init = true;
> +
> + trace_msm_gpu_resume(0);
> +
> + mutex_lock(&a6xx_gpu->gmu.lock);
I don't think we need gmu lock when there is no gmu. Btw, this lock
doesn't seem to be initialized in the current patch.
> +
> + pm_runtime_resume_and_get(gmu->dev);
> + pm_runtime_resume_and_get(gmu->gxpd);
> +
> + /* Set the core clock, having VDD scaling in mind */
> + ret = dev_pm_opp_set_rate(&gpu->pdev->dev, gpu->fast_rate);
> + if (ret)
> + goto err_core_clk;

From power perspective, it is pretty bad if we resume with Fmax. Are you
doing this because you see janks with the current set of GPUs without
GMU?
> +
> + ret = clk_bulk_prepare_enable(gpu->nr_clocks, gpu->grp_clks);
> + if (ret)
> + goto err_bulk_clk;
> +
> + ret = clk_prepare_enable(gpu->ebi1_clk);

Which clk exactly is this? I didn't any patch adding this one.

> + if (ret)
> + goto err_mem_clk;
> +
> + /* If anything goes south, tear the GPU down piece by piece.. */
> + if (ret) {
> +err_mem_clk:
> + clk_bulk_disable_unprepare(gpu->nr_clocks, gpu->grp_clks);
> +err_bulk_clk:
> + opp = dev_pm_opp_find_freq_ceil(&gpu->pdev->dev, &freq);
> + dev_pm_opp_put(opp);
> + dev_pm_opp_set_rate(&gpu->pdev->dev, 0);
> +err_core_clk:
> + pm_runtime_put(gmu->gxpd);
> + pm_runtime_put(gmu->dev);
> + }
> + mutex_unlock(&a6xx_gpu->gmu.lock);
> +
> + if (!ret)
> + msm_devfreq_resume(gpu);
> +
> + return ret;
> +}
> +
> +static int a6xx_gmu_pm_suspend(struct msm_gpu *gpu)
> {
> struct adreno_gpu *adreno_gpu = to_adreno_gpu(gpu);
> struct a6xx_gpu *a6xx_gpu = to_a6xx_gpu(adreno_gpu);
> @@ -1727,11 +1829,62 @@ static int a6xx_pm_suspend(struct msm_gpu *gpu)
> return 0;
> }
>
> +static int a6xx_pm_suspend(struct msm_gpu *gpu)
> +{
> + struct adreno_gpu *adreno_gpu = to_adreno_gpu(gpu);
> + struct a6xx_gpu *a6xx_gpu = to_a6xx_gpu(adreno_gpu);
> + struct a6xx_gmu *gmu = &a6xx_gpu->gmu;
> + unsigned long freq = 0;
> + struct dev_pm_opp *opp;
> + int i, ret;
> +
> + trace_msm_gpu_suspend(0);
> +
> + opp = dev_pm_opp_find_freq_ceil(&gpu->pdev->dev, &freq);
> + dev_pm_opp_put(opp);
> +
> + msm_devfreq_suspend(gpu);
> +
> + mutex_lock(&a6xx_gpu->gmu.lock);
> +
> + clk_disable_unprepare(gpu->ebi1_clk);
> +
> + clk_bulk_disable_unprepare(gpu->nr_clocks, gpu->grp_clks);
> +
> + /* Set frequency to the minimum supported level (no 27MHz on A6xx!) */
> + ret = dev_pm_opp_set_rate(&gpu->pdev->dev, freq);
> + if (ret)
> + goto err;

Why exactly are we doing this here?

> +
> + pm_runtime_put_sync(gmu->gxpd);
> + pm_runtime_put_sync(gmu->dev);
> +
> + mutex_unlock(&a6xx_gpu->gmu.lock);
gmu lock??
> +
> + if (a6xx_gpu->shadow_bo)
> + for (i = 0; i < gpu->nr_rings; i++)
> + a6xx_gpu->shadow[i] = 0;
> +
> + gpu->suspend_count++;
> +
> + return 0;
> +
> +err:
> + mutex_unlock(&a6xx_gpu->gmu.lock);
> +
> + return ret;
> +}
> +
> static int a6xx_get_timestamp(struct msm_gpu *gpu, uint64_t *value)
> {
> struct adreno_gpu *adreno_gpu = to_adreno_gpu(gpu);
> struct a6xx_gpu *a6xx_gpu = to_a6xx_gpu(adreno_gpu);
>
> + if (adreno_has_gmu_wrapper(adreno_gpu)) {
> + *value = gpu_read64(gpu, REG_A6XX_CP_ALWAYS_ON_COUNTER);
> + return 0;
> + }
> +

I feel overusing gmu_wrapper check reduces readability and also
increases the mental gymnastic required while modifying the driver. This chunk
clearly fit better as a separate ".get_timestamp" op.

> mutex_lock(&a6xx_gpu->gmu.lock);
>
> /* Force the GPU power on so we can read this register */
> @@ -1769,7 +1922,8 @@ static void a6xx_destroy(struct msm_gpu *gpu)
> drm_gem_object_put(a6xx_gpu->shadow_bo);
> }
>
> - a6xx_llc_slices_destroy(a6xx_gpu);
> + if (!adreno_has_gmu_wrapper(adreno_gpu))
> + a6xx_llc_slices_destroy(a6xx_gpu);
>
> mutex_lock(&a6xx_gpu->gmu.lock);
> a6xx_gmu_remove(a6xx_gpu);
> @@ -2009,8 +2163,8 @@ static const struct adreno_gpu_funcs funcs = {
> .get_param = adreno_get_param,
> .set_param = adreno_set_param,
> .hw_init = a6xx_hw_init,
> - .pm_suspend = a6xx_pm_suspend,
> - .pm_resume = a6xx_pm_resume,
> + .pm_suspend = a6xx_gmu_pm_suspend,
> + .pm_resume = a6xx_gmu_pm_resume,
> .recover = a6xx_recover,
> .submit = a6xx_submit,
> .active_ring = a6xx_active_ring,
> @@ -2034,6 +2188,34 @@ static const struct adreno_gpu_funcs funcs = {
> .get_timestamp = a6xx_get_timestamp,
> };
>
> +static const struct adreno_gpu_funcs funcs_gmuwrapper = {
> + .base = {
> + .get_param = adreno_get_param,
> + .set_param = adreno_set_param,
> + .hw_init = a6xx_hw_init,
> + .pm_suspend = a6xx_pm_suspend,
> + .pm_resume = a6xx_pm_resume,
> + .recover = a6xx_recover,
> + .submit = a6xx_submit,
> + .active_ring = a6xx_active_ring,
> + .irq = a6xx_irq,
> + .destroy = a6xx_destroy,
> +#if defined(CONFIG_DRM_MSM_GPU_STATE)
> + .show = a6xx_show,
> +#endif
> + .gpu_busy = a6xx_gpu_busy,
> +#if defined(CONFIG_DRM_MSM_GPU_STATE)
> + .gpu_state_get = a6xx_gpu_state_get,
> + .gpu_state_put = a6xx_gpu_state_put,
> +#endif
> + .create_address_space = a6xx_create_address_space,
> + .create_private_address_space = a6xx_create_private_address_space,
> + .get_rptr = a6xx_get_rptr,
> + .progress = a6xx_progress,
> + },
> + .get_timestamp = a6xx_get_timestamp,
> +};
> +
> struct msm_gpu *a6xx_gpu_init(struct drm_device *dev)
> {
> struct msm_drm_private *priv = dev->dev_private;
> @@ -2055,18 +2237,36 @@ struct msm_gpu *a6xx_gpu_init(struct drm_device *dev)
>
> adreno_gpu->registers = NULL;
>
> + /* Check if there is a GMU phandle and set it up */
> + node = of_parse_phandle(pdev->dev.of_node, "qcom,gmu", 0);
> + /* FIXME: How do we gracefully handle this? */
> + BUG_ON(!node);
Would creating a gmu component help to gracefully handle this?
> +
> + adreno_gpu->gmu_is_wrapper = of_device_is_compatible(node, "qcom,adreno-gmu-wrapper");
> +
> /*
> * We need to know the platform type before calling into adreno_gpu_init
> * so that the hw_apriv flag can be correctly set. Snoop into the info
> * and grab the revision number
> */
> info = adreno_info(config->rev);
> -
> - if (info && (info->revn == 650 || info->revn == 660 ||
> - adreno_cmp_rev(ADRENO_REV(6, 3, 5, ANY_ID), info->rev)))
> + if (!info)
> + return ERR_PTR(-EINVAL);
> +
> + /* Assign these early so that we can use the is_aXYZ helpers */
> + /* Numeric revision IDs (e.g. 630) */
> + adreno_gpu->revn = info->revn;
> + /* New-style ADRENO_REV()-only */
> + adreno_gpu->rev = info->rev;
> + /* Quirk data */
> + adreno_gpu->info = info;
> +
> + if (adreno_is_a650(adreno_gpu) || adreno_is_a660_family(adreno_gpu))
> adreno_gpu->base.hw_apriv = true;
>
> - a6xx_llc_slices_init(pdev, a6xx_gpu);
> + /* No LLCC on non-RPMh (and by extension, non-GMU) SoCs */

Assuming this comment is based on the correlation you see with the
current set GPUs, can we update it to "No LLCC on non-RPMh (and by extension, non-GMU) SoCs so far"?

I think the llcc driver currently doesn't handle well if you call
get_slice api when the platform doesn't have an llc. It returns
EPROBE_DEFER instead of ENODEV or something. So this check to
skip is fine for now, I think.

> + if (!adreno_has_gmu_wrapper(adreno_gpu))
can we move this check inside a6xx_llc_slices_init()? Same for the
remove part.

-Akhil

> + a6xx_llc_slices_init(pdev, a6xx_gpu);
>
> ret = a6xx_set_supported_hw(&pdev->dev, config->rev);
> if (ret) {
> @@ -2074,7 +2274,10 @@ struct msm_gpu *a6xx_gpu_init(struct drm_device *dev)
> return ERR_PTR(ret);
> }
>
> - ret = adreno_gpu_init(dev, pdev, adreno_gpu, &funcs, 1);
> + if (adreno_has_gmu_wrapper(adreno_gpu))
> + ret = adreno_gpu_init(dev, pdev, adreno_gpu, &funcs_gmuwrapper, 1);
> + else
> + ret = adreno_gpu_init(dev, pdev, adreno_gpu, &funcs, 1);
> if (ret) {
> a6xx_destroy(&(a6xx_gpu->base.base));
> return ERR_PTR(ret);
> @@ -2087,13 +2290,10 @@ struct msm_gpu *a6xx_gpu_init(struct drm_device *dev)
> if (adreno_is_a618(adreno_gpu) || adreno_is_7c3(adreno_gpu))
> priv->gpu_clamp_to_idle = true;
>
> - /* Check if there is a GMU phandle and set it up */
> - node = of_parse_phandle(pdev->dev.of_node, "qcom,gmu", 0);
> -
> - /* FIXME: How do we gracefully handle this? */
> - BUG_ON(!node);
> -
> - ret = a6xx_gmu_init(a6xx_gpu, node);
> + if (adreno_has_gmu_wrapper(adreno_gpu))
> + ret = a6xx_gmu_wrapper_init(a6xx_gpu, node);
> + else
> + ret = a6xx_gmu_init(a6xx_gpu, node);
> of_node_put(node);
> if (ret) {
> a6xx_destroy(&(a6xx_gpu->base.base));
> diff --git a/drivers/gpu/drm/msm/adreno/a6xx_gpu.h b/drivers/gpu/drm/msm/adreno/a6xx_gpu.h
> index eea2e60ce3b7..51a7656072fa 100644
> --- a/drivers/gpu/drm/msm/adreno/a6xx_gpu.h
> +++ b/drivers/gpu/drm/msm/adreno/a6xx_gpu.h
> @@ -76,6 +76,7 @@ int a6xx_gmu_set_oob(struct a6xx_gmu *gmu, enum a6xx_gmu_oob_state state);
> void a6xx_gmu_clear_oob(struct a6xx_gmu *gmu, enum a6xx_gmu_oob_state state);
>
> int a6xx_gmu_init(struct a6xx_gpu *a6xx_gpu, struct device_node *node);
> +int a6xx_gmu_wrapper_init(struct a6xx_gpu *a6xx_gpu, struct device_node *node);
> void a6xx_gmu_remove(struct a6xx_gpu *a6xx_gpu);
>
> void a6xx_gmu_set_freq(struct msm_gpu *gpu, struct dev_pm_opp *opp,
> diff --git a/drivers/gpu/drm/msm/adreno/a6xx_gpu_state.c b/drivers/gpu/drm/msm/adreno/a6xx_gpu_state.c
> index 30ecdff363e7..4e5d650578c6 100644
> --- a/drivers/gpu/drm/msm/adreno/a6xx_gpu_state.c
> +++ b/drivers/gpu/drm/msm/adreno/a6xx_gpu_state.c
> @@ -1041,16 +1041,18 @@ struct msm_gpu_state *a6xx_gpu_state_get(struct msm_gpu *gpu)
> /* Get the generic state from the adreno core */
> adreno_gpu_state_get(gpu, &a6xx_state->base);
>
> - a6xx_get_gmu_registers(gpu, a6xx_state);
> + if (!adreno_has_gmu_wrapper(adreno_gpu)) {
> + a6xx_get_gmu_registers(gpu, a6xx_state);
>
> - a6xx_state->gmu_log = a6xx_snapshot_gmu_bo(a6xx_state, &a6xx_gpu->gmu.log);
> - a6xx_state->gmu_hfi = a6xx_snapshot_gmu_bo(a6xx_state, &a6xx_gpu->gmu.hfi);
> - a6xx_state->gmu_debug = a6xx_snapshot_gmu_bo(a6xx_state, &a6xx_gpu->gmu.debug);
> + a6xx_state->gmu_log = a6xx_snapshot_gmu_bo(a6xx_state, &a6xx_gpu->gmu.log);
> + a6xx_state->gmu_hfi = a6xx_snapshot_gmu_bo(a6xx_state, &a6xx_gpu->gmu.hfi);
> + a6xx_state->gmu_debug = a6xx_snapshot_gmu_bo(a6xx_state, &a6xx_gpu->gmu.debug);
>
> - a6xx_snapshot_gmu_hfi_history(gpu, a6xx_state);
> + a6xx_snapshot_gmu_hfi_history(gpu, a6xx_state);
> + }
>
> /* If GX isn't on the rest of the data isn't going to be accessible */
> - if (!a6xx_gmu_gx_is_on(&a6xx_gpu->gmu))
> + if (!adreno_has_gmu_wrapper(adreno_gpu) && !a6xx_gmu_gx_is_on(&a6xx_gpu->gmu))
> return &a6xx_state->base;
>
> /* Get the banks of indexed registers */
> diff --git a/drivers/gpu/drm/msm/adreno/adreno_gpu.c b/drivers/gpu/drm/msm/adreno/adreno_gpu.c
> index 84f25122afba..e6216b4169be 100644
> --- a/drivers/gpu/drm/msm/adreno/adreno_gpu.c
> +++ b/drivers/gpu/drm/msm/adreno/adreno_gpu.c
> @@ -468,6 +468,10 @@ int adreno_load_fw(struct adreno_gpu *adreno_gpu)
> if (!adreno_gpu->info->fw[i])
> continue;
>
> + /* Skip loading GMU firwmare with GMU Wrapper */
> + if (adreno_has_gmu_wrapper(adreno_gpu) && i == ADRENO_FW_GMU)
> + continue;
> +
> /* Skip if the firmware has already been loaded */
> if (adreno_gpu->fw[i])
> continue;
> @@ -1021,8 +1025,8 @@ int adreno_gpu_init(struct drm_device *drm, struct platform_device *pdev,
> u32 speedbin;
> int ret;
>
> - /* Only handle the core clock when GMU is not in use */
> - if (config->rev.core < 6) {
> + /* Only handle the core clock when GMU is not in use (or is absent). */
> + if (adreno_has_gmu_wrapper(adreno_gpu) || config->rev.core < 6) {
> /*
> * This can only be done before devm_pm_opp_of_add_table(), or
> * dev_pm_opp_set_config() will WARN_ON()
> diff --git a/drivers/gpu/drm/msm/adreno/adreno_gpu.h b/drivers/gpu/drm/msm/adreno/adreno_gpu.h
> index b4f9b1343d63..2c0f0ef094cb 100644
> --- a/drivers/gpu/drm/msm/adreno/adreno_gpu.h
> +++ b/drivers/gpu/drm/msm/adreno/adreno_gpu.h
> @@ -115,6 +115,7 @@ struct adreno_gpu {
> * code (a3xx_gpu.c) and stored in this common location.
> */
> const unsigned int *reg_offsets;
> + bool gmu_is_wrapper;
> };
> #define to_adreno_gpu(x) container_of(x, struct adreno_gpu, base)
>
> @@ -145,6 +146,11 @@ struct adreno_platform_config {
>
> bool adreno_cmp_rev(struct adreno_rev rev1, struct adreno_rev rev2);
>
> +static inline bool adreno_has_gmu_wrapper(struct adreno_gpu *gpu)
> +{
> + return gpu->gmu_is_wrapper;
> +}
> +
> static inline bool adreno_is_a2xx(struct adreno_gpu *gpu)
> {
> return (gpu->revn < 300);
>
> --
> 2.40.0
>

2023-05-01 21:26:19

by Konrad Dybcio

[permalink] [raw]

Subject: Re: [PATCH v5 06/15] drm/msm/a6xx: Introduce GMU wrapper support

On 1.05.2023 21:40, Akhil P Oommen wrote:
> On Fri, Mar 31, 2023 at 01:25:20AM +0200, Konrad Dybcio wrote:
>> Some (particularly SMD_RPM, a.k.a non-RPMh) SoCs implement A6XX GPUs
>> but don't implement the associated GMUs. This is due to the fact that
>> the GMU directly pokes at RPMh. Sadly, this means we have to take care
>> of enabling & scaling power rails, clocks and bandwidth ourselves.
>>
>> Reuse existing Adreno-common code and modify the deeply-GMU-infused
>> A6XX code to facilitate these GPUs. This involves if-ing out lots
>> of GMU callbacks and introducing a new type of GMU - GMU wrapper (it's
>> the actual name that Qualcomm uses in their downstream kernels).
>>
>> This is essentially a register region which is convenient to model
>> as a device. We'll use it for managing the GDSCs. The register
>> layout matches the actual GMU_CX/GX regions on the "real GMU" devices
>> and lets us reuse quite a bit of gmu_read/write/rmw calls.
>
> Commenting here after going through rest of the patch...
>
> Only convenience I see with modeling a dummy gmu is that we can reuse gmu read/write routines which I think would be less that 10 instances. If we just add a gmu_wrapper region to gpu node, wouldn't that help to create a clean separation between gmu-supported vs gmu-wrapper/no-gmu architectures?
It makes managing CX/GX easy with more or less the existing GMU code.

Also, creating a dummy gmu device in device tree doesn't sound right to me.
The GMU_CX region is also reserved with the actual GMU, even though
it has seemingly very little to do with the Cortex-M3 core, so I'd
say this is only fair..

>
>
>>
>> Signed-off-by: Konrad Dybcio <[email protected]>
>> ---
>> drivers/gpu/drm/msm/adreno/a6xx_gmu.c | 72 +++++++-
>> drivers/gpu/drm/msm/adreno/a6xx_gpu.c | 254 +++++++++++++++++++++++++---
>> drivers/gpu/drm/msm/adreno/a6xx_gpu.h | 1 +
>> drivers/gpu/drm/msm/adreno/a6xx_gpu_state.c | 14 +-
>> drivers/gpu/drm/msm/adreno/adreno_gpu.c | 8 +-
>> drivers/gpu/drm/msm/adreno/adreno_gpu.h | 6 +
>> 6 files changed, 317 insertions(+), 38 deletions(-)
>>
>> diff --git a/drivers/gpu/drm/msm/adreno/a6xx_gmu.c b/drivers/gpu/drm/msm/adreno/a6xx_gmu.c
>> index 1514b3ed0fcf..c6001e82e03d 100644
>> --- a/drivers/gpu/drm/msm/adreno/a6xx_gmu.c
>> +++ b/drivers/gpu/drm/msm/adreno/a6xx_gmu.c
>> @@ -1474,6 +1474,7 @@ static int a6xx_gmu_get_irq(struct a6xx_gmu *gmu, struct platform_device *pdev,
>>
>> void a6xx_gmu_remove(struct a6xx_gpu *a6xx_gpu)
>> {
>> + struct adreno_gpu *adreno_gpu = &a6xx_gpu->base;
>> struct a6xx_gmu *gmu = &a6xx_gpu->gmu;
>> struct platform_device *pdev = to_platform_device(gmu->dev);
>>
>> @@ -1499,10 +1500,12 @@ void a6xx_gmu_remove(struct a6xx_gpu *a6xx_gpu)
>> gmu->mmio = NULL;
>> gmu->rscc = NULL;
>>
>> - a6xx_gmu_memory_free(gmu);
>> + if (!adreno_has_gmu_wrapper(adreno_gpu)) {
>> + a6xx_gmu_memory_free(gmu);
>>
>> - free_irq(gmu->gmu_irq, gmu);
>> - free_irq(gmu->hfi_irq, gmu);
>> + free_irq(gmu->gmu_irq, gmu);
>> + free_irq(gmu->hfi_irq, gmu);
>> + }
>>
>> /* Drop reference taken in of_find_device_by_node */
>> put_device(gmu->dev);
>> @@ -1521,6 +1524,69 @@ static int cxpd_notifier_cb(struct notifier_block *nb,
>> return 0;
>> }
>>
>> +int a6xx_gmu_wrapper_init(struct a6xx_gpu *a6xx_gpu, struct device_node *node)
>> +{
>> + struct platform_device *pdev = of_find_device_by_node(node);
>> + struct a6xx_gmu *gmu = &a6xx_gpu->gmu;
>> + int ret;
>> +
>> + if (!pdev)
>> + return -ENODEV;
>> +
>> + gmu->dev = &pdev->dev;
>> +
>> + of_dma_configure(gmu->dev, node, true);
> If GMU is dummy, why should we configure dma?
We're still accessing registers through a device with an OF node
that's not picked up by any initcalls, but only consumed manually.

>> +
>> + pm_runtime_enable(gmu->dev);
>> +
>> + /* Mark legacy for manual SPTPRAC control */
>> + gmu->legacy = true;
>> +
>> + /* Map the GMU registers */
>> + gmu->mmio = a6xx_gmu_get_mmio(pdev, "gmu");
>> + if (IS_ERR(gmu->mmio)) {
>> + ret = PTR_ERR(gmu->mmio);
>> + goto err_mmio;
>> + }
>> +
>> + gmu->cxpd = dev_pm_domain_attach_by_name(gmu->dev, "cx");
>> + if (IS_ERR(gmu->cxpd)) {
>> + ret = PTR_ERR(gmu->cxpd);
>> + goto err_mmio;
>> + }
>> +
>> + if (!device_link_add(gmu->dev, gmu->cxpd, DL_FLAG_PM_RUNTIME)) {
>> + ret = -ENODEV;
>> + goto detach_cxpd;
>> + }
>> +
>> + init_completion(&gmu->pd_gate);
>> + complete_all(&gmu->pd_gate);
>> + gmu->pd_nb.notifier_call = cxpd_notifier_cb;
>> +
>> + /* Get a link to the GX power domain to reset the GPU */
>> + gmu->gxpd = dev_pm_domain_attach_by_name(gmu->dev, "gx");
>> + if (IS_ERR(gmu->gxpd)) {
>> + ret = PTR_ERR(gmu->gxpd);
>> + goto err_mmio;
>> + }
>> +
>> + gmu->initialized = true;
>> +
>> + return 0;
>> +
>> +detach_cxpd:
>> + dev_pm_domain_detach(gmu->cxpd, false);
>> +
>> +err_mmio:
>> + iounmap(gmu->mmio);
>> +
>> + /* Drop reference taken in of_find_device_by_node */
>> + put_device(gmu->dev);
>> +
>> + return ret;
>> +}
>> +
>> int a6xx_gmu_init(struct a6xx_gpu *a6xx_gpu, struct device_node *node)
>> {
>> struct adreno_gpu *adreno_gpu = &a6xx_gpu->base;
>> diff --git a/drivers/gpu/drm/msm/adreno/a6xx_gpu.c b/drivers/gpu/drm/msm/adreno/a6xx_gpu.c
>> index ae0a90b2834f..a7ecb0a87e98 100644
>> --- a/drivers/gpu/drm/msm/adreno/a6xx_gpu.c
>> +++ b/drivers/gpu/drm/msm/adreno/a6xx_gpu.c
>> @@ -20,9 +20,11 @@ static inline bool _a6xx_check_idle(struct msm_gpu *gpu)
>> struct adreno_gpu *adreno_gpu = to_adreno_gpu(gpu);
>> struct a6xx_gpu *a6xx_gpu = to_a6xx_gpu(adreno_gpu);
>>
>> - /* Check that the GMU is idle */
>> - if (!a6xx_gmu_isidle(&a6xx_gpu->gmu))
>> - return false;
>> + if (!adreno_has_gmu_wrapper(adreno_gpu)) {
>> + /* Check that the GMU is idle */
>> + if (!a6xx_gmu_isidle(&a6xx_gpu->gmu))
>> + return false;
>> + }
>>
>> /* Check tha the CX master is idle */
>> if (gpu_read(gpu, REG_A6XX_RBBM_STATUS) &
>> @@ -612,13 +614,15 @@ static void a6xx_set_hwcg(struct msm_gpu *gpu, bool state)
>> return;
>>
>> /* Disable SP clock before programming HWCG registers */
>> - gmu_rmw(gmu, REG_A6XX_GPU_GMU_GX_SPTPRAC_CLOCK_CONTROL, 1, 0);
>> + if (!adreno_has_gmu_wrapper(adreno_gpu))
>> + gmu_rmw(gmu, REG_A6XX_GPU_GMU_GX_SPTPRAC_CLOCK_CONTROL, 1, 0);
>>
>> for (i = 0; (reg = &adreno_gpu->info->hwcg[i], reg->offset); i++)
>> gpu_write(gpu, reg->offset, state ? reg->value : 0);
>>
>> /* Enable SP clock */
>> - gmu_rmw(gmu, REG_A6XX_GPU_GMU_GX_SPTPRAC_CLOCK_CONTROL, 0, 1);
>> + if (!adreno_has_gmu_wrapper(adreno_gpu))
>> + gmu_rmw(gmu, REG_A6XX_GPU_GMU_GX_SPTPRAC_CLOCK_CONTROL, 0, 1);
>>
>> gpu_write(gpu, REG_A6XX_RBBM_CLOCK_CNTL, state ? clock_cntl_on : 0);
>> }
>> @@ -1002,10 +1006,13 @@ static int hw_init(struct msm_gpu *gpu)
>> {
>> struct adreno_gpu *adreno_gpu = to_adreno_gpu(gpu);
>> struct a6xx_gpu *a6xx_gpu = to_a6xx_gpu(adreno_gpu);
>> + struct a6xx_gmu *gmu = &a6xx_gpu->gmu;
>> int ret;
>>
>> - /* Make sure the GMU keeps the GPU on while we set it up */
>> - a6xx_gmu_set_oob(&a6xx_gpu->gmu, GMU_OOB_GPU_SET);
>> + if (!adreno_has_gmu_wrapper(adreno_gpu)) {
>> + /* Make sure the GMU keeps the GPU on while we set it up */
>> + a6xx_gmu_set_oob(&a6xx_gpu->gmu, GMU_OOB_GPU_SET);
>> + }
>>
>> /* Clear GBIF halt in case GX domain was not collapsed */
>> if (a6xx_has_gbif(adreno_gpu))
>> @@ -1128,6 +1135,17 @@ static int hw_init(struct msm_gpu *gpu)
>> 0x3f0243f0);
>> }
>>
>> + if (adreno_has_gmu_wrapper(adreno_gpu)) {
>> + /* Do it here, as GMU wrapper only inits the GMU for memory reservation etc. */
>> +
>> + /* Set up the CX GMU counter 0 to count busy ticks */
>> + gmu_write(gmu, REG_A6XX_GPU_GMU_AO_GPU_CX_BUSY_MASK, 0xff000000);
>> +
>> + /* Enable power counter 0 */
>> + gmu_rmw(gmu, REG_A6XX_GMU_CX_GMU_POWER_COUNTER_SELECT_0, 0xff, BIT(5));
>> + gmu_write(gmu, REG_A6XX_GMU_CX_GMU_POWER_COUNTER_ENABLE, 1);
>> + }
>> +
>> /* Protect registers from the CP */
>> a6xx_set_cp_protect(gpu);
>>
>> @@ -1236,6 +1254,8 @@ static int hw_init(struct msm_gpu *gpu)
>> }
>>
>> out:
>> + if (adreno_has_gmu_wrapper(adreno_gpu))
>> + return ret;
>> /*
>> * Tell the GMU that we are done touching the GPU and it can start power
>> * management
>> @@ -1270,6 +1290,9 @@ static void a6xx_dump(struct msm_gpu *gpu)
>> adreno_dump(gpu);
>> }
>>
>> +#define GBIF_GX_HALT_MASK BIT(0)
>> +#define GBIF_CLIENT_HALT_MASK BIT(0)
>> +#define GBIF_ARB_HALT_MASK BIT(1)
>> #define VBIF_RESET_ACK_TIMEOUT 100
>> #define VBIF_RESET_ACK_MASK 0x00f0
>>
>> @@ -1302,7 +1325,8 @@ static void a6xx_recover(struct msm_gpu *gpu)
>> * Turn off keep alive that might have been enabled by the hang
>> * interrupt
>> */
>> - gmu_write(&a6xx_gpu->gmu, REG_A6XX_GMU_GMU_PWR_COL_KEEPALIVE, 0);
>> + if (!adreno_has_gmu_wrapper(adreno_gpu))
>> + gmu_write(&a6xx_gpu->gmu, REG_A6XX_GMU_GMU_PWR_COL_KEEPALIVE, 0);
>
> Perhaps, we can move this register write to the begining of a6xx_gmu_force_off(). That looks cleaner to me.
Okay that sounds like it could work!

>
>>
>> pm_runtime_dont_use_autosuspend(&gpu->pdev->dev);
>>
>> @@ -1332,6 +1356,32 @@ static void a6xx_recover(struct msm_gpu *gpu)
>>
>> dev_pm_genpd_remove_notifier(gmu->cxpd);
>>
>> + /* Software-reset the GPU */
> This is not soft reset. We are quiescenting gpu to ddr traffic before we
> pull the regulator
Right, the logic for this comment got tangled up in the "Add A610
support" commit, but it should have been part of this one.

>> + if (adreno_has_gmu_wrapper(adreno_gpu)) {
>> + /* Halt the GX side of GBIF */
>> + gpu_write(gpu, REG_A6XX_RBBM_GBIF_HALT, GBIF_GX_HALT_MASK);
>> + spin_until(gpu_read(gpu, REG_A6XX_RBBM_GBIF_HALT_ACK) &
>> + GBIF_GX_HALT_MASK);
>> +
>> + /* Halt new client requests on GBIF */
>> + gpu_write(gpu, REG_A6XX_GBIF_HALT, GBIF_CLIENT_HALT_MASK);
>> + spin_until((gpu_read(gpu, REG_A6XX_GBIF_HALT_ACK) &
>> + (GBIF_CLIENT_HALT_MASK)) == GBIF_CLIENT_HALT_MASK);
>> +
>> + /* Halt all AXI requests on GBIF */
>> + gpu_write(gpu, REG_A6XX_GBIF_HALT, GBIF_ARB_HALT_MASK);
>> + spin_until((gpu_read(gpu, REG_A6XX_GBIF_HALT_ACK) &
>> + (GBIF_ARB_HALT_MASK)) == GBIF_ARB_HALT_MASK);
>> +
>> + /* Clear the halts */
>> + gpu_write(gpu, REG_A6XX_GBIF_HALT, 0);
>> +
>> + gpu_write(gpu, REG_A6XX_RBBM_GBIF_HALT, 0);
>> +
>> + /* This *really* needs to go through before we do anything else! */
>> + mb();
>> + }
>
> This sequence should be before we collapse the regulators and clocks.
Hm, that makes sense, draining the registers on disabled hardware isn't
very effective..

Can't we handle this similar to the gmu ops in the pm_suspend cb?
We totally can!

Also, We can reuse a6xx_bus_clear_pending_transactions() for this sequence.
Good point! Should we also clear the GBIF halt there?

>
>> +
>> pm_runtime_use_autosuspend(&gpu->pdev->dev);
>>
>> if (active_submits)
>> @@ -1516,7 +1566,8 @@ static void a6xx_fault_detect_irq(struct msm_gpu *gpu)
>> * Force the GPU to stay on until after we finish
>> * collecting information
>> */
>> - gmu_write(&a6xx_gpu->gmu, REG_A6XX_GMU_GMU_PWR_COL_KEEPALIVE, 1);
>> + if (!adreno_has_gmu_wrapper(adreno_gpu))
>> + gmu_write(&a6xx_gpu->gmu, REG_A6XX_GMU_GMU_PWR_COL_KEEPALIVE, 1);
>
> It would be more readable if we wrap this into a subroutine.
Ack

>
>>
>> DRM_DEV_ERROR(&gpu->pdev->dev,
>> "gpu fault ring %d fence %x status %8.8X rb %4.4x/%4.4x ib1 %16.16llX/%4.4x ib2 %16.16llX/%4.4x\n",
>> @@ -1677,7 +1728,7 @@ static void a6xx_llc_slices_init(struct platform_device *pdev,
>> a6xx_gpu->llc_mmio = ERR_PTR(-EINVAL);
>> }
>>
>> -static int a6xx_pm_resume(struct msm_gpu *gpu)
>> +static int a6xx_gmu_pm_resume(struct msm_gpu *gpu)
>> {
>> struct adreno_gpu *adreno_gpu = to_adreno_gpu(gpu);
>> struct a6xx_gpu *a6xx_gpu = to_a6xx_gpu(adreno_gpu);
>> @@ -1697,10 +1748,61 @@ static int a6xx_pm_resume(struct msm_gpu *gpu)
>>
>> a6xx_llc_activate(a6xx_gpu);
>>
>> - return 0;
>> + return ret;
> Isn't ret uninitialized here?
ret = a6xx_gmu_resume(a6xx_gpu);

>> }
>>
>> -static int a6xx_pm_suspend(struct msm_gpu *gpu)
>> +static int a6xx_pm_resume(struct msm_gpu *gpu)
>> +{
>> + struct adreno_gpu *adreno_gpu = to_adreno_gpu(gpu);
>> + struct a6xx_gpu *a6xx_gpu = to_a6xx_gpu(adreno_gpu);
>> + struct a6xx_gmu *gmu = &a6xx_gpu->gmu;
>> + unsigned long freq = 0;
>> + struct dev_pm_opp *opp;
>> + int ret;
>> +
>> + gpu->needs_hw_init = true;
>> +
>> + trace_msm_gpu_resume(0);
>> +
>> + mutex_lock(&a6xx_gpu->gmu.lock);
> I don't think we need gmu lock when there is no gmu.
I'd say locking dvfs calls and some register accesses still makes
sense.

Btw, this lock doesn't seem to be initialized in the current patch.
Ouch, I distinctly remember having a patch for this before sending
v1.. Wonder if/how many silent issues that might have caused..

Anyway, this patch by Dmitry should take care of it in parallel:

https://lore.kernel.org/linux-arm-msm/[email protected]/

>> +
>> + pm_runtime_resume_and_get(gmu->dev);
>> + pm_runtime_resume_and_get(gmu->gxpd);
>> +
>> + /* Set the core clock, having VDD scaling in mind */
>> + ret = dev_pm_opp_set_rate(&gpu->pdev->dev, gpu->fast_rate);
>> + if (ret)
>> + goto err_core_clk;
>
> From power perspective, it is pretty bad if we resume with Fmax. Are you
> doing this because you see janks with the current set of GPUs without
> GMU?
This is what the Adreno-common code (msm_gpu.c : enable_clk()) does.
I suppose scaling CX/GX to turbo/turbo_l1 on every GPU open is a rather
undesirable thing, but we can discuss this separately as it will also
impact performance in tiny workloads where devfreq may not boost in time.

>> + + ret = clk_bulk_prepare_enable(gpu->nr_clocks, gpu->grp_clks); +
>> if (ret) + goto err_bulk_clk; + + ret =
>> clk_prepare_enable(gpu->ebi1_clk);
>
> Which clk exactly is this? I didn't any patch adding this one.
msm_gpu.c : msm_gpu_init()

On a note, this dates back to

7198e6b03155 ("drm/msm: add a3xx gpu support")

Perhaps we could improve some of these ancient names..
>
>> + if (ret)
>> + goto err_mem_clk;
>> +
>> + /* If anything goes south, tear the GPU down piece by piece.. */
>> + if (ret) {
>> +err_mem_clk:
>> + clk_bulk_disable_unprepare(gpu->nr_clocks, gpu->grp_clks);
>> +err_bulk_clk:
>> + opp = dev_pm_opp_find_freq_ceil(&gpu->pdev->dev, &freq);
>> + dev_pm_opp_put(opp);
>> + dev_pm_opp_set_rate(&gpu->pdev->dev, 0);
>> +err_core_clk:
>> + pm_runtime_put(gmu->gxpd);
>> + pm_runtime_put(gmu->dev);
>> + }
>> + mutex_unlock(&a6xx_gpu->gmu.lock);
>> +
>> + if (!ret)
>> + msm_devfreq_resume(gpu);
>> +
>> + return ret;
>> +}
>> +
>> +static int a6xx_gmu_pm_suspend(struct msm_gpu *gpu)
>> {
>> struct adreno_gpu *adreno_gpu = to_adreno_gpu(gpu);
>> struct a6xx_gpu *a6xx_gpu = to_a6xx_gpu(adreno_gpu);
>> @@ -1727,11 +1829,62 @@ static int a6xx_pm_suspend(struct msm_gpu *gpu)
>> return 0;
>> }
>>
>> +static int a6xx_pm_suspend(struct msm_gpu *gpu)
>> +{
>> + struct adreno_gpu *adreno_gpu = to_adreno_gpu(gpu);
>> + struct a6xx_gpu *a6xx_gpu = to_a6xx_gpu(adreno_gpu);
>> + struct a6xx_gmu *gmu = &a6xx_gpu->gmu;
>> + unsigned long freq = 0;
>> + struct dev_pm_opp *opp;
>> + int i, ret;
>> +
>> + trace_msm_gpu_suspend(0);
>> +
>> + opp = dev_pm_opp_find_freq_ceil(&gpu->pdev->dev, &freq);
>> + dev_pm_opp_put(opp);
>> +
>> + msm_devfreq_suspend(gpu);
>> +
>> + mutex_lock(&a6xx_gpu->gmu.lock);
>> +
>> + clk_disable_unprepare(gpu->ebi1_clk);
>> +
>> + clk_bulk_disable_unprepare(gpu->nr_clocks, gpu->grp_clks);
>> +
>> + /* Set frequency to the minimum supported level (no 27MHz on A6xx!) */
>> + ret = dev_pm_opp_set_rate(&gpu->pdev->dev, freq);
>> + if (ret)
>> + goto err;
>
> Why exactly are we doing this here?
Hm.. I think this is not quite right.. I must've got confused and thought
that we should be setting F_MIN on a6xx targets, but we should actually be
able to call set_rate with freq=0. I do however recall the latter
causing issues..

I suppose I'll solve that improper register draining sequence and try
sorting this out, sounds a lot like these 2 may be correlated..

>
>> +
>> + pm_runtime_put_sync(gmu->gxpd);
>> + pm_runtime_put_sync(gmu->dev);
>> +
>> + mutex_unlock(&a6xx_gpu->gmu.lock);
> gmu lock??
I can send a separate series that:

1. moves this to a6xx_gpu
2. renames it to dvfs_lock

so that things would make more sense, especially with the aforementioned
patch by Dmitry that moves it to a6xx gpu init, anyway.

preferably after this one, so that I don't have to resolve 1000 small
conflicts, though..
>> +
>> + if (a6xx_gpu->shadow_bo)
>> + for (i = 0; i < gpu->nr_rings; i++)
>> + a6xx_gpu->shadow[i] = 0;
>> +
>> + gpu->suspend_count++;
>> +
>> + return 0;
>> +
>> +err:
>> + mutex_unlock(&a6xx_gpu->gmu.lock);
>> +
>> + return ret;
>> +}
>> +
>> static int a6xx_get_timestamp(struct msm_gpu *gpu, uint64_t *value)
>> {
>> struct adreno_gpu *adreno_gpu = to_adreno_gpu(gpu);
>> struct a6xx_gpu *a6xx_gpu = to_a6xx_gpu(adreno_gpu);
>>
>> + if (adreno_has_gmu_wrapper(adreno_gpu)) {
>> + *value = gpu_read64(gpu, REG_A6XX_CP_ALWAYS_ON_COUNTER);
>> + return 0;
>> + }
>> +
>
> I feel overusing gmu_wrapper check reduces readability and also
> increases the mental gymnastic required while modifying the driver. This chunk
> clearly fit better as a separate ".get_timestamp" op.
Fair

>
>> mutex_lock(&a6xx_gpu->gmu.lock);
>>
>> /* Force the GPU power on so we can read this register */
>> @@ -1769,7 +1922,8 @@ static void a6xx_destroy(struct msm_gpu *gpu)
>> drm_gem_object_put(a6xx_gpu->shadow_bo);
>> }
>>
>> - a6xx_llc_slices_destroy(a6xx_gpu);
>> + if (!adreno_has_gmu_wrapper(adreno_gpu))
>> + a6xx_llc_slices_destroy(a6xx_gpu);
>>
>> mutex_lock(&a6xx_gpu->gmu.lock);
>> a6xx_gmu_remove(a6xx_gpu);
>> @@ -2009,8 +2163,8 @@ static const struct adreno_gpu_funcs funcs = {
>> .get_param = adreno_get_param,
>> .set_param = adreno_set_param,
>> .hw_init = a6xx_hw_init,
>> - .pm_suspend = a6xx_pm_suspend,
>> - .pm_resume = a6xx_pm_resume,
>> + .pm_suspend = a6xx_gmu_pm_suspend,
>> + .pm_resume = a6xx_gmu_pm_resume,
>> .recover = a6xx_recover,
>> .submit = a6xx_submit,
>> .active_ring = a6xx_active_ring,
>> @@ -2034,6 +2188,34 @@ static const struct adreno_gpu_funcs funcs = {
>> .get_timestamp = a6xx_get_timestamp,
>> };
>>
>> +static const struct adreno_gpu_funcs funcs_gmuwrapper = {
>> + .base = {
>> + .get_param = adreno_get_param,
>> + .set_param = adreno_set_param,
>> + .hw_init = a6xx_hw_init,
>> + .pm_suspend = a6xx_pm_suspend,
>> + .pm_resume = a6xx_pm_resume,
>> + .recover = a6xx_recover,
>> + .submit = a6xx_submit,
>> + .active_ring = a6xx_active_ring,
>> + .irq = a6xx_irq,
>> + .destroy = a6xx_destroy,
>> +#if defined(CONFIG_DRM_MSM_GPU_STATE)
>> + .show = a6xx_show,
>> +#endif
>> + .gpu_busy = a6xx_gpu_busy,
>> +#if defined(CONFIG_DRM_MSM_GPU_STATE)
>> + .gpu_state_get = a6xx_gpu_state_get,
>> + .gpu_state_put = a6xx_gpu_state_put,
>> +#endif
>> + .create_address_space = a6xx_create_address_space,
>> + .create_private_address_space = a6xx_create_private_address_space,
>> + .get_rptr = a6xx_get_rptr,
>> + .progress = a6xx_progress,
>> + },
>> + .get_timestamp = a6xx_get_timestamp,
>> +};
>> +
>> struct msm_gpu *a6xx_gpu_init(struct drm_device *dev)
>> {
>> struct msm_drm_private *priv = dev->dev_private;
>> @@ -2055,18 +2237,36 @@ struct msm_gpu *a6xx_gpu_init(struct drm_device *dev)
>>
>> adreno_gpu->registers = NULL;
>>
>> + /* Check if there is a GMU phandle and set it up */
>> + node = of_parse_phandle(pdev->dev.of_node, "qcom,gmu", 0);
>> + /* FIXME: How do we gracefully handle this? */
>> + BUG_ON(!node);
> Would creating a gmu component help to gracefully handle this?
if !node
return -EINVAL

this original comment was a bit exaggerative here, I think..

>> +
>> + adreno_gpu->gmu_is_wrapper = of_device_is_compatible(node, "qcom,adreno-gmu-wrapper");
>> +
>> /*
>> * We need to know the platform type before calling into adreno_gpu_init
>> * so that the hw_apriv flag can be correctly set. Snoop into the info
>> * and grab the revision number
>> */
>> info = adreno_info(config->rev);
>> -
>> - if (info && (info->revn == 650 || info->revn == 660 ||
>> - adreno_cmp_rev(ADRENO_REV(6, 3, 5, ANY_ID), info->rev)))
>> + if (!info)
>> + return ERR_PTR(-EINVAL);
>> +
>> + /* Assign these early so that we can use the is_aXYZ helpers */
>> + /* Numeric revision IDs (e.g. 630) */
>> + adreno_gpu->revn = info->revn;
>> + /* New-style ADRENO_REV()-only */
>> + adreno_gpu->rev = info->rev;
>> + /* Quirk data */
>> + adreno_gpu->info = info;
>> +
>> + if (adreno_is_a650(adreno_gpu) || adreno_is_a660_family(adreno_gpu))
>> adreno_gpu->base.hw_apriv = true;
>>
>> - a6xx_llc_slices_init(pdev, a6xx_gpu);
>> + /* No LLCC on non-RPMh (and by extension, non-GMU) SoCs */
>
> Assuming this comment is based on the correlation you see with the
> current set GPUs, can we update it to "No LLCC on non-RPMh (and by extension, non-GMU) SoCs so far"?
I can, but keep in mind:

- The GMU clearly calls into RPMh via cmd-db entries (e.g. bw scaling)

- LLCC could be theoretically retrofitted and integrated into SoCs with
SMD RPM (and by extention BIMC), but IIUC the additional separation that
DDRSS (its parent, found on SoCs with RPMh) has is sorta.. necessary..
otherwise a new bus would need to be jammed in, connecting GNoC, BIMC
and all others, resulting in a very overcomplicated connection mesh

>
> I think the llcc driver currently doesn't handle well if you call
> get_slice api when the platform doesn't have an llc. It returns
> EPROBE_DEFER instead of ENODEV or something. So this check to
> skip is fine for now, I think.
Yeah it looks quite messy..

>
>> + if (!adreno_has_gmu_wrapper(adreno_gpu))
> can we move this check inside a6xx_llc_slices_init()? Same for the
> remove part.
Yep, this will make it a lot cleaner!

Thanks for the review!

Konrad
>
> -Akhil
>
>> + a6xx_llc_slices_init(pdev, a6xx_gpu);
>>
>> ret = a6xx_set_supported_hw(&pdev->dev, config->rev);
>> if (ret) {
>> @@ -2074,7 +2274,10 @@ struct msm_gpu *a6xx_gpu_init(struct drm_device *dev)
>> return ERR_PTR(ret);
>> }
>>
>> - ret = adreno_gpu_init(dev, pdev, adreno_gpu, &funcs, 1);
>> + if (adreno_has_gmu_wrapper(adreno_gpu))
>> + ret = adreno_gpu_init(dev, pdev, adreno_gpu, &funcs_gmuwrapper, 1);
>> + else
>> + ret = adreno_gpu_init(dev, pdev, adreno_gpu, &funcs, 1);
>> if (ret) {
>> a6xx_destroy(&(a6xx_gpu->base.base));
>> return ERR_PTR(ret);
>> @@ -2087,13 +2290,10 @@ struct msm_gpu *a6xx_gpu_init(struct drm_device *dev)
>> if (adreno_is_a618(adreno_gpu) || adreno_is_7c3(adreno_gpu))
>> priv->gpu_clamp_to_idle = true;
>>
>> - /* Check if there is a GMU phandle and set it up */
>> - node = of_parse_phandle(pdev->dev.of_node, "qcom,gmu", 0);
>> -
>> - /* FIXME: How do we gracefully handle this? */
>> - BUG_ON(!node);
>> -
>> - ret = a6xx_gmu_init(a6xx_gpu, node);
>> + if (adreno_has_gmu_wrapper(adreno_gpu))
>> + ret = a6xx_gmu_wrapper_init(a6xx_gpu, node);
>> + else
>> + ret = a6xx_gmu_init(a6xx_gpu, node);
>> of_node_put(node);
>> if (ret) {
>> a6xx_destroy(&(a6xx_gpu->base.base));
>> diff --git a/drivers/gpu/drm/msm/adreno/a6xx_gpu.h b/drivers/gpu/drm/msm/adreno/a6xx_gpu.h
>> index eea2e60ce3b7..51a7656072fa 100644
>> --- a/drivers/gpu/drm/msm/adreno/a6xx_gpu.h
>> +++ b/drivers/gpu/drm/msm/adreno/a6xx_gpu.h
>> @@ -76,6 +76,7 @@ int a6xx_gmu_set_oob(struct a6xx_gmu *gmu, enum a6xx_gmu_oob_state state);
>> void a6xx_gmu_clear_oob(struct a6xx_gmu *gmu, enum a6xx_gmu_oob_state state);
>>
>> int a6xx_gmu_init(struct a6xx_gpu *a6xx_gpu, struct device_node *node);
>> +int a6xx_gmu_wrapper_init(struct a6xx_gpu *a6xx_gpu, struct device_node *node);
>> void a6xx_gmu_remove(struct a6xx_gpu *a6xx_gpu);
>>
>> void a6xx_gmu_set_freq(struct msm_gpu *gpu, struct dev_pm_opp *opp,
>> diff --git a/drivers/gpu/drm/msm/adreno/a6xx_gpu_state.c b/drivers/gpu/drm/msm/adreno/a6xx_gpu_state.c
>> index 30ecdff363e7..4e5d650578c6 100644
>> --- a/drivers/gpu/drm/msm/adreno/a6xx_gpu_state.c
>> +++ b/drivers/gpu/drm/msm/adreno/a6xx_gpu_state.c
>> @@ -1041,16 +1041,18 @@ struct msm_gpu_state *a6xx_gpu_state_get(struct msm_gpu *gpu)
>> /* Get the generic state from the adreno core */
>> adreno_gpu_state_get(gpu, &a6xx_state->base);
>>
>> - a6xx_get_gmu_registers(gpu, a6xx_state);
>> + if (!adreno_has_gmu_wrapper(adreno_gpu)) {
>> + a6xx_get_gmu_registers(gpu, a6xx_state);
>>
>> - a6xx_state->gmu_log = a6xx_snapshot_gmu_bo(a6xx_state, &a6xx_gpu->gmu.log);
>> - a6xx_state->gmu_hfi = a6xx_snapshot_gmu_bo(a6xx_state, &a6xx_gpu->gmu.hfi);
>> - a6xx_state->gmu_debug = a6xx_snapshot_gmu_bo(a6xx_state, &a6xx_gpu->gmu.debug);
>> + a6xx_state->gmu_log = a6xx_snapshot_gmu_bo(a6xx_state, &a6xx_gpu->gmu.log);
>> + a6xx_state->gmu_hfi = a6xx_snapshot_gmu_bo(a6xx_state, &a6xx_gpu->gmu.hfi);
>> + a6xx_state->gmu_debug = a6xx_snapshot_gmu_bo(a6xx_state, &a6xx_gpu->gmu.debug);
>>
>> - a6xx_snapshot_gmu_hfi_history(gpu, a6xx_state);
>> + a6xx_snapshot_gmu_hfi_history(gpu, a6xx_state);
>> + }
>>
>> /* If GX isn't on the rest of the data isn't going to be accessible */
>> - if (!a6xx_gmu_gx_is_on(&a6xx_gpu->gmu))
>> + if (!adreno_has_gmu_wrapper(adreno_gpu) && !a6xx_gmu_gx_is_on(&a6xx_gpu->gmu))
>> return &a6xx_state->base;
>>
>> /* Get the banks of indexed registers */
>> diff --git a/drivers/gpu/drm/msm/adreno/adreno_gpu.c b/drivers/gpu/drm/msm/adreno/adreno_gpu.c
>> index 84f25122afba..e6216b4169be 100644
>> --- a/drivers/gpu/drm/msm/adreno/adreno_gpu.c
>> +++ b/drivers/gpu/drm/msm/adreno/adreno_gpu.c
>> @@ -468,6 +468,10 @@ int adreno_load_fw(struct adreno_gpu *adreno_gpu)
>> if (!adreno_gpu->info->fw[i])
>> continue;
>>
>> + /* Skip loading GMU firwmare with GMU Wrapper */
>> + if (adreno_has_gmu_wrapper(adreno_gpu) && i == ADRENO_FW_GMU)
>> + continue;
>> +
>> /* Skip if the firmware has already been loaded */
>> if (adreno_gpu->fw[i])
>> continue;
>> @@ -1021,8 +1025,8 @@ int adreno_gpu_init(struct drm_device *drm, struct platform_device *pdev,
>> u32 speedbin;
>> int ret;
>>
>> - /* Only handle the core clock when GMU is not in use */
>> - if (config->rev.core < 6) {
>> + /* Only handle the core clock when GMU is not in use (or is absent). */
>> + if (adreno_has_gmu_wrapper(adreno_gpu) || config->rev.core < 6) {
>> /*
>> * This can only be done before devm_pm_opp_of_add_table(), or
>> * dev_pm_opp_set_config() will WARN_ON()
>> diff --git a/drivers/gpu/drm/msm/adreno/adreno_gpu.h b/drivers/gpu/drm/msm/adreno/adreno_gpu.h
>> index b4f9b1343d63..2c0f0ef094cb 100644
>> --- a/drivers/gpu/drm/msm/adreno/adreno_gpu.h
>> +++ b/drivers/gpu/drm/msm/adreno/adreno_gpu.h
>> @@ -115,6 +115,7 @@ struct adreno_gpu {
>> * code (a3xx_gpu.c) and stored in this common location.
>> */
>> const unsigned int *reg_offsets;
>> + bool gmu_is_wrapper;
>> };
>> #define to_adreno_gpu(x) container_of(x, struct adreno_gpu, base)
>>
>> @@ -145,6 +146,11 @@ struct adreno_platform_config {
>>
>> bool adreno_cmp_rev(struct adreno_rev rev1, struct adreno_rev rev2);
>>
>> +static inline bool adreno_has_gmu_wrapper(struct adreno_gpu *gpu)
>> +{
>> + return gpu->gmu_is_wrapper;
>> +}
>> +
>> static inline bool adreno_is_a2xx(struct adreno_gpu *gpu)
>> {
>> return (gpu->revn < 300);
>>
>> --
>> 2.40.0
>>