Re: [PATCH 3/5] drm: msm: scale DDR BW along with GPU frequency
From: Sharat Masetty
Date: Wed Apr 01 2020 - 08:30:27 EST
On 3/31/2020 10:56 PM, Jordan Crouse wrote:
On Tue, Mar 31, 2020 at 01:25:51PM +0530, Sharat Masetty wrote:
This patch adds support to parse the OPP tables attached the GPU device,
the main opp table and the DDR bandwidth opp table. Additionally, vote
for the GPU->DDR bandwidth when setting the GPU frequency by querying
the linked DDR BW opp to the GPU opp.
Signed-off-by: Sharat Masetty <smasetty@xxxxxxxxxxxxxx>
---
drivers/gpu/drm/msm/adreno/a6xx_gmu.c | 41 ++++++++++++++++++++++++++----
drivers/gpu/drm/msm/adreno/adreno_gpu.c | 44 +++++++++++++++++++++++++++++----
drivers/gpu/drm/msm/msm_gpu.h | 9 +++++++
3 files changed, 84 insertions(+), 10 deletions(-)
diff --git a/drivers/gpu/drm/msm/adreno/a6xx_gmu.c b/drivers/gpu/drm/msm/adreno/a6xx_gmu.c
index 748cd37..489d9b6 100644
--- a/drivers/gpu/drm/msm/adreno/a6xx_gmu.c
+++ b/drivers/gpu/drm/msm/adreno/a6xx_gmu.c
@@ -100,6 +100,40 @@ bool a6xx_gmu_gx_is_on(struct a6xx_gmu *gmu)
A6XX_GMU_SPTPRAC_PWR_CLK_STATUS_GX_HM_CLK_OFF));
}
+void a6xx_gmu_set_icc_vote(struct msm_gpu *gpu, unsigned long gpu_freq)
+{
+ struct dev_pm_opp *gpu_opp, *ddr_opp;
+ struct opp_table **tables = gpu->opp_tables;
+ unsigned long peak_bw;
+
+ if (!gpu->opp_tables[GPU_DDR_OPP_TABLE_INDEX])
+ goto done;
+
+ gpu_opp = dev_pm_opp_find_freq_exact(&gpu->pdev->dev, gpu_freq, true);
+ if (IS_ERR_OR_NULL(gpu_opp))
+ goto done;
+
+ ddr_opp = dev_pm_opp_xlate_required_opp(tables[GPU_OPP_TABLE_INDEX],
+ tables[GPU_DDR_OPP_TABLE_INDEX],
+ gpu_opp);
+ dev_pm_opp_put(gpu_opp);
+
+ if (IS_ERR_OR_NULL(ddr_opp))
+ goto done;
I think that the final approach is still up in the air but either way we're
going to pull the bandwidth from an OPP, its just a question of which OPP.
+ peak_bw = dev_pm_opp_get_bw(ddr_opp, NULL);
+ dev_pm_opp_put(ddr_opp);
+
+ icc_set_bw(gpu->icc_path, 0, peak_bw);
+ return;
+done:
+ /*
+ * If there is a problem, for now leave it at max so that the
+ * performance is nominal.
+ */
+ icc_set_bw(gpu->icc_path, 0, MBps_to_icc(7216));
+}
+
static void __a6xx_gmu_set_freq(struct a6xx_gmu *gmu, int index)
{
struct a6xx_gpu *a6xx_gpu = container_of(gmu, struct a6xx_gpu, gmu);
@@ -128,11 +162,8 @@ static void __a6xx_gmu_set_freq(struct a6xx_gmu *gmu, int index)
gmu->freq = gmu->gpu_freqs[index];
- /*
- * Eventually we will want to scale the path vote with the frequency but
- * for now leave it at max so that the performance is nominal.
- */
- icc_set_bw(gpu->icc_path, 0, MBps_to_icc(7216));
+ if (gpu->icc_path)
+ a6xx_gmu_set_icc_vote(gpu, gmu->freq);
This function is annoying because we call it from two different spots, but it
feels wasteful that devfreq gives us an OPP pointer and we go out of our way to
not use it only to search for it again in the set_icc_vote function. I think
maybe we should pass the OPP through from msm_gpu.c. We could have a helper
function to pull the initial opp in a6xx_gmu_resume to make it clean.
Yes Jordan, it makes sense. I did think about this too, but may be I
was a bit too lazy to change the existing plumbing :)
I will take care of this in the next iteration.
}
void a6xx_gmu_set_freq(struct msm_gpu *gpu, unsigned long freq)
diff --git a/drivers/gpu/drm/msm/adreno/adreno_gpu.c b/drivers/gpu/drm/msm/adreno/adreno_gpu.c
index 2d13694..bbbcc7a 100644
--- a/drivers/gpu/drm/msm/adreno/adreno_gpu.c
+++ b/drivers/gpu/drm/msm/adreno/adreno_gpu.c
@@ -882,7 +882,7 @@ static int adreno_get_pwrlevels(struct device *dev,
{
unsigned long freq = ULONG_MAX;
struct dev_pm_opp *opp;
- int ret;
+ int ret, i;
gpu->fast_rate = 0;
@@ -890,9 +890,29 @@ static int adreno_get_pwrlevels(struct device *dev,
if (!of_find_property(dev->of_node, "operating-points-v2", NULL))
ret = adreno_get_legacy_pwrlevels(dev);
else {
- ret = dev_pm_opp_of_add_table(dev);
- if (ret)
- DRM_DEV_ERROR(dev, "Unable to set the OPP table\n");
+ int count = of_count_phandle_with_args(dev->of_node,
+ "operating-points-v2", NULL);
+
+ count = min(count, GPU_DDR_OPP_TABLE_INDEX + 1);
+ count = max(count, 1);
+
+ for (i = 0; i < count; i++) {
+ ret = dev_pm_opp_of_add_table_indexed(dev, i);
+ if (ret) {
+ DRM_DEV_ERROR(dev, "Add OPP table %d: failed %d\n",
+ i, ret);
+ goto err;
+ }
+
+ gpu->opp_tables[i] =
+ dev_pm_opp_get_opp_table_indexed(dev, i);
+ if (!gpu->opp_tables[i]) {
+ DRM_DEV_ERROR(dev, "Get OPP table failed index %d\n",
+ i);
+ ret = -EINVAL;
+ goto err;
+ }
+ }
}
if (!ret) {
@@ -919,12 +939,24 @@ static int adreno_get_pwrlevels(struct device *dev,
gpu->icc_path = NULL;
return 0;
+err:
+ for (; i >= 0; i--) {
+ if (gpu->opp_tables[i]) {
+ dev_pm_opp_put_opp_table(gpu->opp_tables[i]);
+ gpu->opp_tables[i] = NULL;
+ }
+ }
+
+ dev_pm_opp_remove_table(dev);
+ return ret;
}
int adreno_gpu_init(struct drm_device *drm, struct platform_device *pdev,
struct adreno_gpu *adreno_gpu,
const struct adreno_gpu_funcs *funcs, int nr_rings)
{
+ int ret = 0;
+
struct adreno_platform_config *config = pdev->dev.platform_data;
struct msm_gpu_config adreno_gpu_config = { 0 };
struct msm_gpu *gpu = &adreno_gpu->base;
@@ -945,7 +977,9 @@ int adreno_gpu_init(struct drm_device *drm, struct platform_device *pdev,
adreno_gpu_config.nr_rings = nr_rings;
- adreno_get_pwrlevels(&pdev->dev, gpu);
+ ret = adreno_get_pwrlevels(&pdev->dev, gpu);
+ if (ret)
+ return ret;
pm_runtime_set_autosuspend_delay(&pdev->dev,
adreno_gpu->info->inactive_period);
diff --git a/drivers/gpu/drm/msm/msm_gpu.h b/drivers/gpu/drm/msm/msm_gpu.h
index ab8f0f9c..5b98b48 100644
--- a/drivers/gpu/drm/msm/msm_gpu.h
+++ b/drivers/gpu/drm/msm/msm_gpu.h
@@ -66,6 +66,12 @@ struct msm_gpu_funcs {
void (*gpu_set_freq)(struct msm_gpu *gpu, unsigned long freq);
};
+/* opp table indices */
+enum {
+ GPU_OPP_TABLE_INDEX,
+ GPU_DDR_OPP_TABLE_INDEX,
+};
+
struct msm_gpu {
const char *name;
struct drm_device *dev;
@@ -113,6 +119,9 @@ struct msm_gpu {
struct icc_path *icc_path;
+ /* gpu/ddr opp tables */
+ struct opp_table *opp_tables[2];
You don't need an array here. We're not going to have that many tables.
struct opp_table *gpu_opp_table;
struct opp_table *bw_opp_table;
Is sufficient and we don't need an enum.
+
/* Hang and Inactivity Detection:
*/
#define DRM_MSM_INACTIVE_PERIOD 66 /* in ms (roughly four frames) */
--
2.7.4
Jordan