drm/amdgpu: load balance VCN3 decode as well v8

Add VCN3 IB parsing to figure out to which instance we can send the
stream for decode.

v2: remove VCN instance limit as well, fix amdgpu_cs_find_mapping,
    check supported formats instead of unsupported.
v3: fix typo and error handling
v4: make sure the message BO is CPU accessible
v5: fix addr calculation once more
v6: only check message buffers
v7: fix constant and use defines
v8: fix create msg calculation
Signed-off-by: Christian König <christian.koenig@amd.com>
Reviewed-by: Sonny Jiang <sonny.jiang@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>

drivers/gpu/drm/amd/amdgpu/vcn_v3_0.c

@@ -50,6 +50,9 @@
 #define VCN_INSTANCES_SIENNA_CICHLID				2
 #define DEC_SW_RING_ENABLED					FALSE
 
+#define RDECODE_MSG_CREATE					0x00000000
+#define RDECODE_MESSAGE_CREATE					0x00000001
+
 static int amdgpu_ih_clientid_vcns[] = {
 	SOC15_IH_CLIENTID_VCN,
 	SOC15_IH_CLIENTID_VCN1
@@ -208,8 +211,6 @@ static int vcn_v3_0_sw_init(void *handle)
 		} else {
 			ring->doorbell_index = (adev->doorbell_index.vcn.vcn_ring0_1 << 1) + 8 * i;
 		}
-		if (adev->asic_type == CHIP_SIENNA_CICHLID && i != 0)
-			ring->no_scheduler = true;
 		sprintf(ring->name, "vcn_dec_%d", i);
 		r = amdgpu_ring_init(adev, ring, 512, &adev->vcn.inst[i].irq, 0,
 				     AMDGPU_RING_PRIO_DEFAULT,
@@ -1847,6 +1848,132 @@ static const struct amdgpu_ring_funcs vcn_v3_0_dec_sw_ring_vm_funcs = {
 	.emit_reg_write_reg_wait = amdgpu_ring_emit_reg_write_reg_wait_helper,
 };
 
+static int vcn_v3_0_limit_sched(struct amdgpu_cs_parser *p)
+{
+	struct drm_gpu_scheduler **scheds;
+
+	/* The create msg must be in the first IB submitted */
+	if (atomic_read(&p->entity->fence_seq))
+		return -EINVAL;
+
+	scheds = p->adev->gpu_sched[AMDGPU_HW_IP_VCN_DEC]
+		[AMDGPU_RING_PRIO_DEFAULT].sched;
+	drm_sched_entity_modify_sched(p->entity, scheds, 1);
+	return 0;
+}
+
+static int vcn_v3_0_dec_msg(struct amdgpu_cs_parser *p, uint64_t addr)
+{
+	struct ttm_operation_ctx ctx = { false, false };
+	struct amdgpu_bo_va_mapping *map;
+	uint32_t *msg, num_buffers;
+	struct amdgpu_bo *bo;
+	uint64_t start, end;
+	unsigned int i;
+	void * ptr;
+	int r;
+
+	addr &= AMDGPU_GMC_HOLE_MASK;
+	r = amdgpu_cs_find_mapping(p, addr, &bo, &map);
+	if (r) {
+		DRM_ERROR("Can't find BO for addr 0x%08Lx\n", addr);
+		return r;
+	}
+
+	start = map->start * AMDGPU_GPU_PAGE_SIZE;
+	end = (map->last + 1) * AMDGPU_GPU_PAGE_SIZE;
+	if (addr & 0x7) {
+		DRM_ERROR("VCN messages must be 8 byte aligned!\n");
+		return -EINVAL;
+	}
+
+	bo->flags |= AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED;
+	amdgpu_bo_placement_from_domain(bo, bo->allowed_domains);
+	r = ttm_bo_validate(&bo->tbo, &bo->placement, &ctx);
+	if (r) {
+		DRM_ERROR("Failed validating the VCN message BO (%d)!\n", r);
+		return r;
+	}
+
+	r = amdgpu_bo_kmap(bo, &ptr);
+	if (r) {
+		DRM_ERROR("Failed mapping the VCN message (%d)!\n", r);
+		return r;
+	}
+
+	msg = ptr + addr - start;
+
+	/* Check length */
+	if (msg[1] > end - addr) {
+		r = -EINVAL;
+		goto out;
+	}
+
+	if (msg[3] != RDECODE_MSG_CREATE)
+		goto out;
+
+	num_buffers = msg[2];
+	for (i = 0, msg = &msg[6]; i < num_buffers; ++i, msg += 4) {
+		uint32_t offset, size, *create;
+
+		if (msg[0] != RDECODE_MESSAGE_CREATE)
+			continue;
+
+		offset = msg[1];
+		size = msg[2];
+
+		if (offset + size > end) {
+			r = -EINVAL;
+			goto out;
+		}
+
+		create = ptr + addr + offset - start;
+
+		/* H246, HEVC and VP9 can run on any instance */
+		if (create[0] == 0x7 || create[0] == 0x10 || create[0] == 0x11)
+			continue;
+
+		r = vcn_v3_0_limit_sched(p);
+		if (r)
+			goto out;
+	}
+
+out:
+	amdgpu_bo_kunmap(bo);
+	return r;
+}
+
+static int vcn_v3_0_ring_patch_cs_in_place(struct amdgpu_cs_parser *p,
+					   uint32_t ib_idx)
+{
+	struct amdgpu_ring *ring = to_amdgpu_ring(p->entity->rq->sched);
+	struct amdgpu_ib *ib = &p->job->ibs[ib_idx];
+	uint32_t msg_lo = 0, msg_hi = 0;
+	unsigned i;
+	int r;
+
+	/* The first instance can decode anything */
+	if (!ring->me)
+		return 0;
+
+	for (i = 0; i < ib->length_dw; i += 2) {
+		uint32_t reg = amdgpu_get_ib_value(p, ib_idx, i);
+		uint32_t val = amdgpu_get_ib_value(p, ib_idx, i + 1);
+
+		if (reg == PACKET0(p->adev->vcn.internal.data0, 0)) {
+			msg_lo = val;
+		} else if (reg == PACKET0(p->adev->vcn.internal.data1, 0)) {
+			msg_hi = val;
+		} else if (reg == PACKET0(p->adev->vcn.internal.cmd, 0) &&
+			   val == 0) {
+			r = vcn_v3_0_dec_msg(p, ((u64)msg_hi) << 32 | msg_lo);
+			if (r)
+				return r;
+		}
+	}
+	return 0;
+}
+
 static const struct amdgpu_ring_funcs vcn_v3_0_dec_ring_vm_funcs = {
 	.type = AMDGPU_RING_TYPE_VCN_DEC,
 	.align_mask = 0xf,
@@ -1854,6 +1981,7 @@ static const struct amdgpu_ring_funcs vcn_v3_0_dec_ring_vm_funcs = {
 	.get_rptr = vcn_v3_0_dec_ring_get_rptr,
 	.get_wptr = vcn_v3_0_dec_ring_get_wptr,
 	.set_wptr = vcn_v3_0_dec_ring_set_wptr,
+	.patch_cs_in_place = vcn_v3_0_ring_patch_cs_in_place,
 	.emit_frame_size =
 		SOC15_FLUSH_GPU_TLB_NUM_WREG * 6 +
 		SOC15_FLUSH_GPU_TLB_NUM_REG_WAIT * 8 +