用户消息传递服务

简介

在多 OS 混合部署场景下，Linux 与实时操作系统（RTOS）通过 RPMsg 协议进行跨核通信。默认配置下，RPMsg 使用 512 字节固定大小的消息缓冲区进行数据传输。当传输较大数据时，会触发以下问题：

需要将数据分片为多个512字节的消息包
每个消息包的传输都会产生IPI（处理器间中断）
频繁的IPI导致系统吞吐量下降和通信延迟增加

因此，当前 MICA 提供了 rpmsg-umt (user message transfer)服务，实现了高效通信服务：

预先分配物理连续的共享内存区域
用户数据写入到共享内存区域
通过 RPMsg 单次传输共享内存的元数据（起始地址+大小）

对比如下：

传统方式： [Data] -> RPMsg Buffer -> IPI -> RPMsg Buffer -> [Data]

rpmsg-umt： [Data] -> Shared Memory -> (IPI + Metadata) -> Shared Memory -> [Data]

UMT 服务支持 baremetal 和 hetero 两种部署底座，通过 ped_type 参数进行区分。

使用方法

UMT 提供了两类接口：上下文接口和一次性接口。

上下文接口

上下文接口适用于需要多次通信的场景，创建一次上下文后可以重复使用，减少资源分配开销。

创建上下文

/**
 * @brief Create UMT communication context
 *
 * Allocates and initializes resources.
 *
 * @param target_instance Target instance ID (only 0 supported)
 * @param ped_type Pedestal type: MCS_KM_PED_BAREMETAL or MCS_KM_PED_RISCV
 * @return Context handle on success, NULL on failure
 * @note Call umt_context_destroy when done
 */
extern umt_context_t* umt_context_create(int target_instance, enum mcs_km_pedestal_type ped_type);

销毁上下文

/**
 * @brief Destroy UMT communication context
 *
 * @param ctx Context handle
 * @note Safe if ctx is NULL; releases any held lock before destroy
 */
extern void umt_context_destroy(umt_context_t *ctx);

发送数据

/**
 * @brief Send data using context (no reply wait)
 *
 * @param ctx Context handle
 * @param offset Offset in umt shared memory (bytes), 0 ~ (OPENAMP_SHM_COPY_SIZE - data_len)
 * @param data Data to send
 * @param data_len Data length
 * @return 0 on success, negative errno on failure (e.g. -EINVAL for invalid ctx, offset or data_len)
 * @note Lock is acquired/released internally
 */
extern int send_data_with_umt_context(umt_context_t *ctx, int offset, void *data, int data_len);

接收数据（阻塞）

/**
 * @brief Receive data using context (block until data arrives or timeout)
 *
 * @param ctx Context handle
 * @param rcv_data Receive buffer
 * @param rcv_data_len Output: received length
 * @param timeout_ms Timeout in ms; 0 = wait forever
 * @return 0 on success, negative errno on failure (e.g. -EINVAL invalid ctx, -ETIMEDOUT on timeout,
 *         -ENODATA, -ENODEV, -EFAULT from read path)
 * @note Lock released during wait, re-acquired after
 * @note Do not use the same context for both blocking receive and callback; use one mode per context.
 */
extern int receive_data_with_umt_context(umt_context_t *ctx, void *rcv_data, int *rcv_data_len, int timeout_ms);

注册接收回调

/**
 * @brief Register receive callback (library runs an internal thread that waits for data and calls callback)
 *
 * One callback per context. Callback and blocking receive_data_with_umt_context must not be used on the same context.
 *
 * @param ctx Context handle
 * @param callback  Called with (data, data_len, priv); data is valid only during the call
 * @param priv  Opaque pointer passed to callback (e.g. application context)
 * @return 0 on success, negative errno on failure (e.g. -EINVAL if ctx/callback NULL or already registered,
 *         -ENOMEM, -EAGAIN if thread create fails)
 */
extern int umt_register_rcv_cb(umt_context_t *ctx, umt_rcv_cb_t callback, void *priv);

/**
 * @brief Unregister receive callback and stop the internal receive thread
 *
 * @param ctx Context handle
 * @return 0 on success, negative errno on failure (e.g. -EINVAL if ctx is NULL, or no callback registered)
 */
extern int umt_unregister_rcv_cb(umt_context_t *ctx);

一次性接口

一次性接口适用于简单的单次通信场景，内部会自动创建和销毁上下文。

发送数据

/**
 * @brief One-shot send to RTOS (creates/destroys context internally; offset 0)
 *
 * @param data Data to send
 * @param data_len Data length
 * @param target_instance Target instance ID (must be 0)
 * @param ped_type MCS_KM_PED_BAREMETAL or MCS_KM_PED_RISCV
 * @return 0 on success, negative errno on failure (e.g. -EINVAL if ctx or offset invalid)
 */
extern int send_data_to_rtos(void *data, int data_len, int target_instance, enum mcs_km_pedestal_type ped_type);

接收数据

/**
 * @brief One-shot receive from RTOS (creates/destroys context internally)
 *
 * @param rcv_data Receive buffer
 * @param rcv_data_len Output: received length
 * @param target_instance Target instance ID (must be 0)
 * @param timeout_ms Timeout in ms; 0 = wait forever
 * @param ped_type MCS_KM_PED_BAREMETAL or MCS_KM_PED_RISCV
 * @return 0 on success, negative errno on failure (e.g. -ENOMEM, -ETIMEDOUT, -EINVAL, -ENODATA, -EFAULT)
 */
int receive_data_from_rtos(void *rcv_data, int *rcv_data_len, int target_instance, int timeout_ms, enum mcs_km_pedestal_type ped_type);

发送并等待回复（旧接口）

/**
 * @brief One-shot send and wait for reply (legacy API; uses BAREMETAL)
 *
 * Use send_data_to_rtos_and_wait_rcv_ped to specify pedestal.
 *
 * @param data Data to send
 * @param data_len Data length
 * @param target_instance Target instance ID (must be 0)
 * @param rcv_data Receive buffer
 * @param rcv_data_len Output: received length
 * @return 0 on success, negative errno on failure (e.g. -ENOMEM, -EINVAL, -ETIMEDOUT)
 */
extern int send_data_to_rtos_and_wait_rcv(void *data, int data_len, int target_instance, void *rcv_data, int *rcv_data_len);

/**
 * @brief One-shot send and wait for reply with specified pedestal type
 *
 * @param data Data to send
 * @param data_len Data length
 * @param target_instance Target instance ID (must be 0)
 * @param rcv_data Receive buffer
 * @param rcv_data_len Output: received length
 * @param ped_type MCS_KM_PED_BAREMETAL or MCS_KM_PED_RISCV
 * @return 0 on success, negative errno on failure (e.g. -ENOMEM, -EINVAL, -ETIMEDOUT)
 */
extern int send_data_to_rtos_and_wait_rcv_ped(void *data, int data_len, int target_instance, void *rcv_data, int *rcv_data_len, enum mcs_km_pedestal_type ped_type);

Note

对于 baremetal 底座，使用 MCS_KM_PED_BAREMETAL 作为 ped_type 参数
对于 hetero 底座（RISC-V），使用 MCS_KM_PED_RISCV 作为 ped_type 参数

Linux 端示例可以参考 test_umt 。

RTOS 端实现可以参考 UniProton PR454，PR457 。

性能指标

测试环境：

KunPeng 920
openEuler Embedded 25.03
UniProton

测试结果：

数据量	rpmsg-umt耗时
1MB	5.14ms