Add host buffer & bug fix (#469)

* support add host buffer message
* TDR should not move context out of partition as it doesn't know if update partition is required or not

---------

Signed-off-by: Min Ma <[email protected]>

Author: mamin506

src/driver/amdxdna/aie2_ctx_runqueue.c

@@ -392,7 +392,6 @@ static void part_ctx_start(struct aie2_partition *part, struct amdxdna_ctx *ctx)
 
 static void part_ctx_stop_wait(struct amdxdna_ctx *ctx, bool wait)
 {
-	struct aie2_partition *part;
 	struct amdxdna_dev *xdna;
 	struct aie2_ctx_rq *rq;
 
@@ -404,8 +403,21 @@ static void part_ctx_stop_wait(struct amdxdna_ctx *ctx, bool wait)
 		return;
 	}
 	aie2_ctx_disconnect(ctx, wait);
+
 	list_move_tail(&ctx->entry, &rq->disconn_list);
 	ctx->priv->status = CTX_STATE_DISCONNECTED;
+}
+
+static void part_ctx_stop(struct amdxdna_ctx *ctx)
+{
+	struct aie2_partition *part;
+	struct amdxdna_dev *xdna;
+	struct aie2_ctx_rq *rq;
+
+	xdna = ctx->client->xdna;
+	rq = &xdna->dev_handle->ctx_rq;
+	part_ctx_stop_wait(ctx, true);
+
 	part = ctx->priv->part;
 	if (part) {
 		part->hwctx_cnt--;
@@ -418,11 +430,6 @@ static void part_ctx_stop_wait(struct amdxdna_ctx *ctx, bool wait)
 	XDNA_DBG(xdna, "%s disconnected", ctx->name);
 }
 
-static void part_ctx_stop(struct amdxdna_ctx *ctx)
-{
-	part_ctx_stop_wait(ctx, true);
-}
-
 static void rq_ctx_cancel(struct aie2_ctx_rq *rq, struct amdxdna_ctx *ctx)
 {
 	struct aie2_partition *part;
@@ -452,8 +459,10 @@ static void part_block_all_ctx(struct aie2_partition *part)
 {
 	struct amdxdna_ctx *ctx;
 
-	list_for_each_entry(ctx, &part->conn_list, entry)
+	list_for_each_entry(ctx, &part->conn_list, entry) {
+		XDNA_DBG(ctx->client->xdna, "%s set block", ctx->name);
 		ctx->priv->should_block = true;
+	}
 }
 
 static void part_cleanup(struct aie2_partition *part)
@@ -731,7 +740,6 @@ static bool handle_busy_ctxs(struct aie2_ctx_rq *rq)
 	bool active = false;
 	int i;
 
-	rq->paused = true;
 	for (i = 0; i < rq->num_parts; i++) {
 		part = &rq->parts[i];
 		if (!part->hwctx_cnt)
@@ -762,6 +770,7 @@ static void rq_parts_work(struct work_struct *work)
 		goto out;
 
 	/* Partition expanding or trimming is needed */
+	rq->paused = true;
 	if (handle_busy_ctxs(rq)) {
 		XDNA_DBG(xdna, "Wait for disconneting active contexts");
 		goto out;
@@ -873,6 +882,8 @@ void aie2_rq_stop_all(struct aie2_ctx_rq *rq)
 			down_write(&ctx->priv->io_sem);
 			XDNA_DBG(xdna, "%s @[%d, %d] stop", ctx->name,
 				 part->start_col, part->end_col);
+			ctx->priv->should_block = true;
+			ctx->priv->force_yield = true;
 			part_ctx_stop_wait(ctx, false);
 			up_write(&ctx->priv->io_sem);
 		}
@@ -1101,7 +1112,7 @@ void aie2_rq_del(struct aie2_ctx_rq *rq, struct amdxdna_ctx *ctx)
 
 	if (wait_update_parts && wait_parts)
 		wait_for_completion_killable(&ctx->priv->parts_work_comp);
-	cancel_work_sync(&ctx->yield_work);
+	flush_work(&ctx->yield_work);
 	XDNA_DBG(xdna, "%s deleted, status %d priority %d",
 		 ctx->name, ctx->priv->status, ctx->priv->priority);
 }

src/driver/amdxdna/aie2_message.c

@@ -360,19 +360,30 @@ int aie2_destroy_context(struct amdxdna_dev_hdl *ndev, struct amdxdna_ctx *ctx)
 
 int aie2_map_host_buf(struct amdxdna_dev_hdl *ndev, u32 context_id, u64 addr, u64 size)
 {
-	DECLARE_AIE2_MSG(map_host_buffer, MSG_OP_MAP_HOST_BUFFER);
+	DECLARE_AIE2_MSG(host_buffer, MSG_OP_MAP_HOST_BUFFER);
 	struct amdxdna_dev *xdna = ndev->xdna;
+	size_t chunk_size;
 	int ret;
 
-	req.context_id = context_id;
-	req.buf_addr = addr;
-	req.buf_size = size;
-	ret = aie2_send_mgmt_msg_wait(ndev, &msg);
-	if (ret)
-		return ret;
+	chunk_size = xdna->dev_info->dev_mem_size;
+	WARN_ON(!is_power_of_2(chunk_size));
+	WARN_ON(!IS_ALIGNED(size, chunk_size));
+	do {
+		req.context_id = context_id;
+		req.buf_addr = addr;
+		req.buf_size = chunk_size;
+		ret = aie2_send_mgmt_msg_wait(ndev, &msg);
+		if (ret)
+			return ret;
+
+		addr += chunk_size;
+		size -= chunk_size;
+		XDNA_DBG(xdna, "hwctx %d map host buf addr 0x%llx size 0x%lx",
+			 context_id, addr, chunk_size);
 
-	XDNA_DBG(xdna, "hwctx %d map host buf addr 0x%llx size 0x%llx",
-		 context_id, addr, size);
+		/* Change opcode if there are more than one chunk */
+		msg.opcode = MSG_OP_ADD_HOST_BUFFER;
+	} while (size);
 
 	return 0;
 }

src/driver/amdxdna/aie2_msg_priv.h

@@ -45,6 +45,7 @@ enum aie2_msg_opcode {
 	MSG_OP_START_EVENT_TRACE           = 0x10F,
 	MSG_OP_STOP_EVENT_TRACE            = 0x110,
 	MSG_OP_UPDATE_PROPERTY             = 0x113,
+	MSG_OP_ADD_HOST_BUFFER             = 0x114,
 	MSG_OP_MAX_DRV_OPCODE,
 	MSG_OP_GET_PROTOCOL_VERSION        = 0x301,
 	MSG_OP_MAX_OPCODE
@@ -105,13 +106,14 @@ struct assign_mgmt_pasid_resp {
 	enum aie2_msg_status	status;
 } __packed;
 
-struct map_host_buffer_req {
+/* For MSG_OP_MAP_HOST_BUFFER and MSG_OP_ADD_HOST_BUFFER */
+struct host_buffer_req {
 	u32		context_id;
 	u64		buf_addr;
 	u64		buf_size;
 } __packed;
 
-struct map_host_buffer_resp {
+struct host_buffer_resp {
 	enum aie2_msg_status	status;
 } __packed;
 

src/driver/amdxdna/amdxdna_gem.c

@@ -772,8 +772,9 @@ amdxdna_drm_create_dev_heap_bo(struct drm_device *dev,
 	struct amdxdna_gem_obj *abo;
 	int ret;
 
-	if (args->size > xdna->dev_info->dev_mem_size) {
-		XDNA_ERR(xdna, "Invalid dev heap size 0x%llx, limit 0x%lx",
+	WARN_ON(!is_power_of_2(xdna->dev_info->dev_mem_size));
+	if (!IS_ALIGNED(args->size, xdna->dev_info->dev_mem_size)) {
+		XDNA_ERR(xdna, "The dev heap size 0x%llx is not multiple of 0x%lx",
 			 args->size, xdna->dev_info->dev_mem_size);
 		return ERR_PTR(-EINVAL);
 	}