Merge "msm: vidc: Add support to prefetch & drain memory"

msm/vidc/msm_smem.c

@@ -9,7 +9,6 @@
 #include "msm_vidc_debug.h"
 #include "msm_vidc_resources.h"
 
-
 static int msm_dma_get_device_address(struct dma_buf *dbuf, unsigned long align,
 	dma_addr_t *iova, unsigned long *buffer_size,
 	unsigned long flags, enum hal_buffer buffer_type,
@@ -599,3 +598,73 @@ struct context_bank_info *msm_smem_get_context_bank(u32 session_type,
 	return match;
 }
 
+int msm_smem_memory_prefetch(struct msm_vidc_inst *inst)
+{
+	int i, rc = 0;
+	struct memory_regions *vidc_regions = NULL;
+	struct ion_prefetch_region ion_region[MEMORY_REGIONS_MAX];
+
+	if (!inst) {
+		d_vpr_e("%s: invalid parameters\n", __func__);
+		return -EINVAL;
+	}
+
+	vidc_regions = &inst->regions;
+	if (vidc_regions->num_regions > MEMORY_REGIONS_MAX) {
+		s_vpr_e(inst->sid, "%s: invalid num_regions %d, max %d\n",
+			__func__, vidc_regions->num_regions,
+			MEMORY_REGIONS_MAX);
+		return -EINVAL;
+	}
+
+	memset(ion_region, 0, sizeof(ion_region));
+	for (i = 0; i < vidc_regions->num_regions; i++) {
+		ion_region[i].size = vidc_regions->region[i].size;
+		ion_region[i].vmid = vidc_regions->region[i].vmid;
+	}
+
+	rc = msm_ion_heap_prefetch(ION_SECURE_HEAP_ID, ion_region,
+		vidc_regions->num_regions);
+	if (rc)
+		s_vpr_e(inst->sid, "%s: prefetch failed, ret: %d\n",
+			__func__, rc);
+	else
+		s_vpr_l(inst->sid, "%s: prefetch succeeded\n", __func__);
+
+	return rc;
+}
+
+int msm_smem_memory_drain(struct msm_vidc_inst *inst)
+{
+	int i, rc = 0;
+	struct memory_regions *vidc_regions = NULL;
+	struct ion_prefetch_region ion_region[MEMORY_REGIONS_MAX];
+
+	if (!inst) {
+		d_vpr_e("%s: invalid parameters\n", __func__);
+		return -EINVAL;
+	}
+
+	vidc_regions = &inst->regions;
+	if (vidc_regions->num_regions > MEMORY_REGIONS_MAX) {
+		s_vpr_e(inst->sid, "%s: invalid num_regions %d, max %d\n",
+			__func__, vidc_regions->num_regions,
+			MEMORY_REGIONS_MAX);
+		return -EINVAL;
+	}
+
+	memset(ion_region, 0, sizeof(ion_region));
+	for (i = 0; i < vidc_regions->num_regions; i++) {
+		ion_region[i].size = vidc_regions->region[i].size;
+		ion_region[i].vmid = vidc_regions->region[i].vmid;
+	}
+
+	rc = msm_ion_heap_drain(ION_SECURE_HEAP_ID, ion_region,
+		vidc_regions->num_regions);
+	if (rc)
+		s_vpr_e(inst->sid, "%s: drain failed, ret: %d\n", __func__, rc);
+	else
+		s_vpr_l(inst->sid, "%s: drain succeeded\n", __func__);
+
+	return rc;
+}

msm/vidc/msm_vdec.c

@@ -883,6 +883,7 @@ int msm_vdec_s_ctrl(struct msm_vidc_inst *inst, struct v4l2_ctrl *ctrl)
 				__func__);
 			return -ENOTSUPP;
 		}
+		msm_comm_memory_prefetch(inst);
 		break;
 	case V4L2_CID_MPEG_VIDC_VIDEO_EXTRADATA:
 		if (ctrl->val == EXTRADATA_NONE)

msm/vidc/msm_vidc.c

@@ -1415,6 +1415,8 @@ static const struct v4l2_ctrl_ops msm_vidc_ctrl_ops = {
 static struct msm_vidc_inst_smem_ops  msm_vidc_smem_ops = {
 	.smem_map_dma_buf = msm_smem_map_dma_buf,
 	.smem_unmap_dma_buf = msm_smem_unmap_dma_buf,
+	.smem_prefetch = msm_smem_memory_prefetch,
+	.smem_drain = msm_smem_memory_drain,
 };
 
 void *msm_vidc_open(int core_id, int session_type)
@@ -1756,6 +1758,7 @@ int msm_vidc_close(void *instance)
 	}
 
 	msm_comm_session_clean(inst);
+	msm_comm_memory_drain(inst);
 
 	kref_put(&inst->kref, close_helper);
 	return 0;

msm/vidc/msm_vidc.h

@@ -13,6 +13,7 @@
 
 #define HAL_BUFFER_MAX 0xe
 #define CVP_FRAME_RATE_MAX (60)
+#define MEMORY_REGIONS_MAX 30
 
 enum smem_type {
 	SMEM_DMA = 1,
@@ -75,6 +76,19 @@ enum smem_cache_ops {
 	SMEM_CACHE_CLEAN_INVALIDATE,
 };
 
+struct memory_regions {
+	u32 num_regions;
+	struct {
+		u64 size;
+		u32 vmid;
+	} region[MEMORY_REGIONS_MAX];
+};
+
+enum memory_ops {
+	MEMORY_PREFETCH = 1,
+	MEMORY_DRAIN,
+};
+
 enum core_id {
 	MSM_VIDC_CORE_VENUS = 0,
 	MSM_VIDC_CORE_Q6,

msm/vidc/msm_vidc_common.c

@@ -7351,3 +7351,94 @@ int msm_comm_fetch_framerate(struct msm_vidc_inst *inst,
 	}
 	return rc;
 }
+
+static int msm_comm_memory_regions_prepare(struct msm_vidc_inst *inst)
+{
+	u32 i = 0;
+	struct msm_vidc_platform_resources *res;
+
+	if (!inst || !inst->core) {
+		d_vpr_e("%s: invalid parameters\n", __func__);
+		return -EINVAL;
+	}
+	res = &inst->core->resources;
+
+	inst->regions.num_regions = res->prefetch_non_pix_buf_count +
+		res->prefetch_pix_buf_count;
+	if (inst->regions.num_regions > MEMORY_REGIONS_MAX) {
+		s_vpr_e(inst->sid, "%s: invalid num_regions: %d, max: %d\n",
+			__func__, inst->regions.num_regions,
+			MEMORY_REGIONS_MAX);
+		return -EINVAL;
+	}
+
+	s_vpr_h(inst->sid,
+		"%s: preparing %d nonpixel memory regions of %ld bytes each and %d pixel memory regions of %ld bytes each\n",
+		__func__, res->prefetch_non_pix_buf_count,
+		res->prefetch_non_pix_buf_size, res->prefetch_pix_buf_count,
+		res->prefetch_pix_buf_size);
+
+	for (i = 0; i < res->prefetch_non_pix_buf_count; i++) {
+		inst->regions.region[i].size = res->prefetch_non_pix_buf_size;
+		inst->regions.region[i].vmid = ION_FLAG_CP_NON_PIXEL;
+	}
+
+	for (i = res->prefetch_non_pix_buf_count;
+		i < inst->regions.num_regions; i++) {
+		inst->regions.region[i].size = res->prefetch_pix_buf_size;
+		inst->regions.region[i].vmid = ION_FLAG_CP_PIXEL;
+	}
+
+	return 0;
+}
+
+int msm_comm_memory_prefetch(struct msm_vidc_inst *inst)
+{
+	int rc = 0;
+
+	if (!inst || !inst->smem_ops) {
+		d_vpr_e("%s: invalid parameters\n", __func__);
+		return -EINVAL;
+	}
+
+	if (inst->memory_ops & MEMORY_PREFETCH) {
+		s_vpr_h(inst->sid, "%s: prefetch done already\n", __func__);
+		return 0;
+	}
+
+	rc = msm_comm_memory_regions_prepare(inst);
+	if (rc)
+		return rc;
+
+	if (inst->regions.num_regions == 0)
+		return 0;
+
+	rc = inst->smem_ops->smem_prefetch(inst);
+	if (rc)
+		return rc;
+
+	inst->memory_ops |= MEMORY_PREFETCH;
+
+	return rc;
+}
+
+int msm_comm_memory_drain(struct msm_vidc_inst *inst)
+{
+	int rc = 0;
+
+	if (!inst || !inst->smem_ops) {
+		d_vpr_e("%s: invalid parameters\n", __func__);
+		return -EINVAL;
+	}
+
+	if (!(inst->memory_ops & MEMORY_PREFETCH))
+		return 0;
+
+	rc = inst->smem_ops->smem_drain(inst);
+	if (rc)
+		return rc;
+
+	inst->memory_ops &= ~MEMORY_PREFETCH;
+
+	return rc;
+}

msm/vidc/msm_vidc_common.h

@@ -339,4 +339,6 @@ void msm_comm_release_timestamps(struct msm_vidc_inst *inst);
 u32 msm_comm_get_max_framerate(struct msm_vidc_inst *inst);
 u32 msm_comm_calc_framerate(struct msm_vidc_inst *inst,	u64 timestamp_us,
 	u64 prev_ts);
+int msm_comm_memory_prefetch(struct msm_vidc_inst *inst);
+int msm_comm_memory_drain(struct msm_vidc_inst *inst);
 #endif

msm/vidc/msm_vidc_internal.h

@@ -485,6 +485,8 @@ struct msm_vidc_inst_smem_ops {
 		struct msm_smem *smem);
 	int (*smem_unmap_dma_buf)(struct msm_vidc_inst *inst,
 		struct msm_smem *smem);
+	int (*smem_prefetch)(struct msm_vidc_inst *inst);
+	int (*smem_drain)(struct msm_vidc_inst *inst);
 };
 
 struct msm_vidc_inst {
@@ -552,6 +554,8 @@ struct msm_vidc_inst {
 	struct batch_mode batch;
 	struct delayed_work batch_work;
 	struct msm_vidc_inst_smem_ops *smem_ops;
+	enum memory_ops memory_ops;
+	struct memory_regions regions;
 	int (*buffer_size_calculators)(struct msm_vidc_inst *inst);
 	bool all_intra;
 	bool is_perf_eligible_session;
@@ -612,6 +616,8 @@ void msm_smem_put_dma_buf(void *dma_buf, u32 sid);
 int msm_smem_cache_operations(struct dma_buf *dbuf,
 	enum smem_cache_ops cache_op, unsigned long offset,
 	unsigned long size, u32 sid);
+int msm_smem_memory_prefetch(struct msm_vidc_inst *inst);
+int msm_smem_memory_drain(struct msm_vidc_inst *inst);
 void msm_vidc_fw_unload_handler(struct work_struct *work);
 void msm_vidc_ssr_handler(struct work_struct *work);
 /*

msm/vidc/msm_vidc_platform.c

@@ -431,6 +431,33 @@ static struct msm_vidc_common_data lahaina_common_data[] = {
 		.key = "qcom,avsync-window-size",
 		.value = 40,
 	},
+	{
+		.key = "qcom,prefetch_non_pix_buf_count",
+		.value = 1,
+	},
+	{
+		.key = "qcom,prefetch_non_pix_buf_size",
+		/*
+		 * Internal buffer size is calculated for secure decode session
+		 * of resolution 4k (4096x2160)
+		 * Internal buf size = calculate_scratch_size() +
+		 *	calculate_scratch1_size() + calculate_persist1_size()
+		 * Take maximum between VP9 10bit, HEVC 10bit, AVC, MPEG2 secure
+		 * decoder sessions
+		 */
+		.value = 209715200,
+	},
+	{
+		.key = "qcom,prefetch_pix_buf_count",
+		.value = 18,
+	},
+	{
+		.key = "qcom,prefetch_pix_buf_size",
+		/*
+		 * Calculated by VENUS_BUFFER_SIZE for 4096x2160 UBWC
+		 */
+		.value = 13434880,
+	},
 };
 
 static struct msm_vidc_common_data bengal_common_data_v0[] = {

msm/vidc/msm_vidc_res_parse.c

@@ -775,6 +775,15 @@ int read_platform_resources_from_drv_data(
 	res->max_secure_inst_count = find_key_value(platform_data,
 			"qcom,max-secure-instances");
 
+	res->prefetch_pix_buf_count = find_key_value(platform_data,
+			"qcom,prefetch_pix_buf_count");
+	res->prefetch_pix_buf_size = find_key_value(platform_data,
+			"qcom,prefetch_pix_buf_size");
+	res->prefetch_non_pix_buf_count = find_key_value(platform_data,
+			"qcom,prefetch_non_pix_buf_count");
+	res->prefetch_non_pix_buf_size = find_key_value(platform_data,
+			"qcom,prefetch_non_pix_buf_size");
+
 	res->slave_side_cp = find_key_value(platform_data,
 			"qcom,slave-side-cp");
 	res->thermal_mitigable = find_key_value(platform_data,

msm/vidc/msm_vidc_resources.h

@@ -168,6 +168,10 @@ struct msm_vidc_platform_resources {
 	uint32_t pm_qos_latency_us;
 	uint32_t max_inst_count;
 	uint32_t max_secure_inst_count;
+	uint32_t prefetch_pix_buf_count;
+	uint32_t prefetch_pix_buf_size;
+	uint32_t prefetch_non_pix_buf_count;
+	uint32_t prefetch_non_pix_buf_size;
 	int msm_vidc_hw_rsp_timeout;
 	int msm_vidc_firmware_unload_delay;
 	uint32_t msm_vidc_pwr_collapse_delay;