path: root/service/src/vdev.c

/*
 * \brief 虚拟设备管理器
 * 
 * 虚拟设备，用于驱动与应用进行数据通信。驱动程序将从物理设备
 * 收取的报文，经过交换逻辑送入虚拟设备中，从而分发到各应用进行处理。
 * 
 * \author Lu Qiuwen<[email protected]>
 * \date 2017-02-17
 * 
 */


#include <sys/queue.h>
#include <rte_ring.h>
#include <rte_malloc.h>
#include <rte_string_fns.h>
#include <assert.h>
#include <common.h>
#include <vnode.h>
#include <sc_common.h>
#include <sc_vdev.h>
#include <MESA_prof_load.h>
#include <ctrlmsg.h>
#include <ctrlmsg_define.h>
#include <sc_app.h>
#include <sc_mrb.h>
#include <arpa/inet.h>


#ifndef VDEV_DEFAULT_SZ_BUFFER
#define VDEV_DEFAULT_SZ_BUFFER		32
#endif

#ifndef VDEV_DEFAULT_SZ_TUNNEL
#define VDEV_DEFAULT_SZ_TUNNEL		512
#endif

/* =========================================================================================== */

int vdev_iterate(struct vdev_main * v_main, struct vdev ** vdev_result, unsigned int * next)
{
	if (*next >= v_main->vdev_max_idx) return -ENOENT;
	else *vdev_result = &v_main->_vdev_array[(*next)++]->vdev;
	return 0;
}

struct vdev * vdev_lookup(struct vdev_main * v_main, const char * symbol)
{
	struct vdev * vdev = NULL; unsigned int next = 0;
	while(vdev_iterate(v_main, &vdev, &next) >= 0)
	{
		if (strncmp(vdev->symbol, symbol, sizeof(vdev->symbol)) != 0) continue;
		return vdev;
	}

	return NULL;
}

// 设置虚设备的IP地址
int vdev_set_inaddr(struct vdev * vdev, struct in_addr in_addr, struct in_addr in_mask,
	struct in_addr gateway)
{
	vdev->in_addr = in_addr;
	vdev->in_mask = in_mask;
	vdev->in_gateway = gateway;
	return 0;
}

// 读取虚设备的IP地址
int vdev_get_inaddr(struct vdev * vdev, struct in_addr * in_addr, struct in_addr * in_mask,
	struct in_addr * gateway)
{
	*in_addr = vdev->in_addr;
	*in_mask = vdev->in_mask;
	*gateway = vdev->in_gateway;
	return 0;
}

int vdev_set_ether_addr(struct vdev * vdev, struct rte_ether_addr * ether_addr)
{
    vdev->ether_addr = *ether_addr;
    return 0;
}

// 设置虚设备的MTU
int vdev_set_mtu(struct vdev * vdev, unsigned int mtu)
{
	vdev->mtu = mtu;
	return 0;
}

// 读取虚设备的MTU
int vdev_get_mtu(struct vdev * vdev)
{
	return vdev->mtu;
}

// 启动设备
int vdev_enable(struct vdev * vdev)
{
    vdev->enable = 1;
	return 0;
}

// 禁用设备
int vdev_disable(struct vdev * vdev)
{
    vdev->enable = 0;
	return 0;
}

/* Datapath Functions */
int vdev_dispatch(struct vdev * vdev, queue_id_t qid, struct rte_mbuf * pkts[],
	unsigned int nr_pkts, int flags)
{
	struct _vdev * _vdev = container_of(vdev, struct _vdev, vdev);
	return _vdev->dispatch(_vdev, qid, pkts, nr_pkts, flags);
}

int vdev_collect(struct vdev * vdev, queue_id_t qid, struct rte_mbuf * pkts[],
	unsigned int nr_pkts, int flags)
{
	struct _vdev * _vdev = container_of(vdev, struct _vdev, vdev);
	return _vdev->collect(_vdev, qid, pkts, nr_pkts, flags);
}

int vdev_idle_poll(struct vdev * vdev, queue_id_t qid)
{
	struct _vdev * _vdev = container_of(vdev, struct _vdev, vdev);
	return _vdev->idle_pool(_vdev, qid);
}

int vdev_stats_get(struct vdev * vdev, struct vdev_stat_info * stat_info)
{
    struct _vdev * _vdev = container_of(vdev, struct _vdev, vdev);
    return _vdev->stats_get(_vdev, stat_info);
}

void vdev_stats_last_save(struct vdev * vdev, struct vdev_stat_info * stat_info_last)
{
    struct _vdev * _vdev = container_of(vdev, struct _vdev, vdev);
    _vdev->stat_info_last = *stat_info_last;
}

void vdev_stats_last_get(struct vdev * vdev, struct vdev_stat_info * stat_info_last)
{
    struct _vdev * _vdev = container_of(vdev, struct _vdev, vdev);
    *stat_info_last = _vdev->stat_info_last;
}


/* 虚设备用户配置查询 */
int vdev_usercfg_query_vdev_list(struct sc_main * sc, const char * str_vdev_type,
		char vdevsyms[MR_SYMBOL_MAX][MR_VDEV_MAX], unsigned int * nr_vdevsyms)
{
	char str_direct_devices[MR_STRING_MAX] = { 0 };
	int ret = MESA_load_profile_string_nodef(sc->local_cfgfile, "device", str_vdev_type,
		str_direct_devices, sizeof(str_direct_devices));

	if (ret < 0)
	{	
		*nr_vdevsyms = 0;
		return RT_SUCCESS;
	}
	
	char * str_vdev_symbol[MR_TOKENS_MAX];
	int nr_str_tokens = rte_strsplit(str_direct_devices, sizeof(str_direct_devices),
		str_vdev_symbol, MR_TOKENS_MAX, ',');

	for (int i = 0; i < nr_str_tokens; i++)
	{
		MR_VERIFY(i < MR_VDEV_MAX);
		snprintf(vdevsyms[i], sizeof(vdevsyms[i]), "%s", str_vdev_symbol[i]);
	}

	*nr_vdevsyms = nr_str_tokens;
	return RT_SUCCESS;
}

int vdev_usercfg_query_vdev_in_addr(struct sc_main * sc, const char * devsym,
    struct in_addr * addr, struct in_addr * mask, struct in_addr * gateway)
{
    // 生成配置文件节名称，例如[device:mr0]
    char str_section[MR_SYMBOL_MAX];
    snprintf(str_section, sizeof(str_section), "device:%s", devsym);

    // 读IP地址（必要）
    char str_in_addr[MR_SYMBOL_MAX];
    if (MESA_load_profile_string_nodef(sc->local_cfgfile, str_section, "in_addr",
        str_in_addr, sizeof(str_in_addr)) < 0) return 0;

    // 地址转换
    int ret = inet_pton(AF_INET, str_in_addr, addr);
    if (ret < 0)
    {
        MR_CFGERR_INVALID_FORMAT(sc->local_cfgfile, str_section, "in_addr");
        return RT_ERR;
    }

    // 读掩码（必要）
    char str_in_mask[MR_SYMBOL_MAX];
    if (MESA_load_profile_string_nodef(sc->local_cfgfile, str_section, "in_mask",
        str_in_mask, sizeof(str_in_mask)) < 0) return RT_ERR;

    // 掩码转换
    ret = inet_pton(AF_INET, str_in_mask, mask);
    if (ret < 0)
    {
        MR_CFGERR_INVALID_FORMAT(sc->local_cfgfile, str_section, "in_mask");
        return RT_ERR;
    }

    // 读默认网关（可选项）
    char str_gateway[MR_SYMBOL_MAX];
    if (MESA_load_profile_string_nodef(sc->local_cfgfile, str_section, "gateway",
        str_gateway, sizeof(str_gateway)) >= 0)
    {
        ret = inet_pton(AF_INET, str_gateway, gateway);
        if (ret < 0)
        {
            MR_CFGERR_INVALID_FORMAT(sc->local_cfgfile, str_section, "gateway");
            return RT_ERR;
        }
    }

    return RT_SUCCESS;
}

int vdev_usercfg_query_vdev_setup(struct sc_main * sc, const char * devsym,
	unsigned int * sz_tunnel, unsigned int * sz_buffer)
{
	unsigned int default_sz_tunnel;
	unsigned int default_sz_buffer;

	MESA_load_profile_uint_def(sc->local_cfgfile, "device", "sz_tunnel",
		&default_sz_tunnel, VDEV_DEFAULT_SZ_TUNNEL);
	MESA_load_profile_uint_def(sc->local_cfgfile, "device", "sz_buffer",
		&default_sz_buffer, VDEV_DEFAULT_SZ_BUFFER);

	char str_dev_section[MR_SYMBOL_MAX];
	snprintf(str_dev_section, sizeof(str_dev_section), "device:%s", devsym);

	MESA_load_profile_uint_def(sc->local_cfgfile, str_dev_section, "sz_tunnel",
		sz_tunnel, default_sz_tunnel);
	MESA_load_profile_uint_def(sc->local_cfgfile, str_dev_section, "sz_buffer",
		sz_buffer, default_sz_buffer);
	
	return RT_SUCCESS;
}

int vdev_usercfg_query_mempool(struct sc_main * sc, const char * devsym,
    char mempool_direct_pool[MR_SYMBOL_MAX], char mempool_indirect_pool[MR_SYMBOL_MAX])
{
	char str_dev_section[MR_SYMBOL_MAX] = { 0 };
    snprintf(str_dev_section, sizeof(str_dev_section), "device:%s", devsym);

    /* 没有定义直接内存池名称，使用设备名 */
    MESA_load_profile_string_def(sc->local_cfgfile, str_dev_section, "direct-pool",
        mempool_direct_pool, MR_SYMBOL_MAX, devsym);

    /* 没有定义间接内存池名称，使用直接内存池名称 */
    MESA_load_profile_string_def(sc->local_cfgfile, str_dev_section, "indirect-pool",
        mempool_indirect_pool, MR_SYMBOL_MAX, mempool_direct_pool);

    return RT_SUCCESS;
}

/* 处理应用打开设备的请求。为应用创建请求设备的VDI结构 */
static int __vdev_open_request_handler(struct ctrlmsg_handler * ct_hand, struct ctrlmsg_conn * ct_conn,
    struct ctrl_msg_header * msg, void * arg)
{
    struct app * app_object = (struct app *)ct_conn->pme;
    struct vdev_main_pme * vdev_main_pme = (struct vdev_main_pme *)app_object->pme_vdev_main;
    struct vdev_main * vdev_main = (struct vdev_main *)arg;

    struct ctrl_msg_vdev_open_request * msg_req = (struct ctrl_msg_vdev_open_request *)msg;

    /* 构造应答头部 */
    struct ctrl_msg_vdev_open_response msg_rep;
    memset(&msg_rep, 0, sizeof(msg_rep));

    ctrl_msg_header_construct(&msg_rep.msg_header, sizeof(msg_rep), 
        CTRL_MSG_TYPE_RESPONSE, (char *)msg->msg_topic);
    
    /* 查找应用请求打开的VDEV的句柄，检查是否存在 */
    struct vdev * vdev = vdev_lookup(vdev_main, (char *)msg_req->devsym);
    if (vdev == NULL)
    {
        MR_WARNING("Wrong Request: from %s, device %s is not existed", app_object->symbol, (char *)msg_req->devsym);
        ctrl_msg_error_construct(&msg_rep.msg_err, ENODEV, "device %s is not existed.", (char *)msg_req->devsym);
        ctrlmsg_msg_send(ct_hand, ct_conn, &msg_rep.msg_header);
        return 0;
    }

    struct _vdev * _vdev = container_of(vdev, struct _vdev, vdev);

    /* 申请VDI */
    struct vdev_instance * vdi = _vdev->vdi_create(_vdev, app_object->symbol, 
        msg_req->nr_rxstream, msg_req->nr_txstream);

    if (vdi == NULL)
    {
        MR_ERROR("Create vdev instance %s for app %s failed. ", (char *)msg_req->devsym, app_object->symbol);
        ctrl_msg_error_construct(&msg_rep.msg_err, ENOMEM, "Create vdev instance failed.");
        ctrlmsg_msg_send(ct_hand, ct_conn, &msg_rep.msg_header);
        return 0;
    }

    /* VDI申请成功 */
    msg_rep.ptr_vdi = (uint64_t)vdi;
    msg_rep.nr_rxstream = msg_req->nr_rxstream;
    msg_rep.nr_txstream = msg_req->nr_txstream;
    snprintf((char *)msg_rep.devsym, sizeof(msg_rep.devsym), "%s", vdev->symbol);
    ctrlmsg_msg_send(ct_hand, ct_conn, &msg_rep.msg_header);

    TAILQ_INSERT_TAIL(&vdev_main_pme->vdi_list, vdi, next);
    return 0;
}

static void __vdev_app_register_handler(struct app_main * app_main, struct app * app, void * arg)
{
    app->pme_vdev_main = ZMALLOC(sizeof(struct vdev_main_pme));
    MR_VERIFY_MALLOC(app->pme_vdev_main);

    struct vdev_main_pme * vdev_main_pme = (struct vdev_main_pme *)app->pme_vdev_main;
    TAILQ_INIT(&vdev_main_pme->vdi_list);
    return;
}

static void __vdev_app_unregister_handler(struct app_main * app_main, struct app * app, void * arg)
{
    struct vdev_instance * vdi;
    struct vdev_main_pme * vdev_main_pme = (struct vdev_main_pme *)app->pme_vdev_main;

    if (vdev_main_pme == NULL) return;

    while (!TAILQ_EMPTY(&vdev_main_pme->vdi_list))
    {
        vdi = TAILQ_FIRST(&vdev_main_pme->vdi_list);
        TAILQ_REMOVE(&vdev_main_pme->vdi_list, vdi, next);

        struct _vdev * _vdev = container_of(vdi->vdev, struct _vdev, vdev);
        _vdev->vdi_destory(vdi);
    }

    FREE(vdev_main_pme);
    return;
}

/* 根据配置文件创建回环虚设备 */
static int vdev_loop_usercfg_setup(struct sc_main * sc)
{
    /* 加载虚设备配置，虚设备和物理设备一一对应 */
    char vdevsyms[MR_SYMBOL_MAX][MR_VDEV_MAX] = { { 0 } };
    unsigned int nr_vdevsyms = 0;
    vdev_usercfg_query_vdev_list(sc, "loop", vdevsyms, &nr_vdevsyms);

    for (int i = 0; i < nr_vdevsyms; i++)
    {
        const char * vdevsym = vdevsyms[i];
        struct rte_mempool * direct_pool = NULL;
        struct rte_mempool * indirect_pool = NULL;

		char sym_direct_mempool[MR_SYMBOL_MAX] = { 0 };
		char sym_indirect_mempool[MR_SYMBOL_MAX] = { 0 };
        vdev_usercfg_query_mempool(sc, vdevsym, sym_direct_mempool, sym_indirect_mempool);

        /* 查询Indirect Pool，使用与物理设备对应的Indirect Pool */
        direct_pool = mrb_direct_mempool_locate(sc->mrb_pool_main, sym_direct_mempool, 0, 0);
        if (direct_pool == NULL)
        {
            MR_ERROR("Direct mempool %s for device %s is not existed. ",
                sym_direct_mempool, vdevsym); goto err;
        }

        indirect_pool = mrb_indirect_mempool_locate(sc->mrb_pool_main, sym_indirect_mempool, 0, 0);
        if (indirect_pool == NULL)
        {
            MR_ERROR("Indirect mempool %s for device %s is not existed. ",
                sym_indirect_mempool, vdevsym); goto err;
        }

        /* 查询队列缓冲区大小 */
        unsigned int sz_tunnel, sz_buffer;
        vdev_usercfg_query_vdev_setup(sc, vdevsym, &sz_tunnel, &sz_buffer);

        /* 创建回环虚设备 */
        int ret = vdev_loop_create(sc->vdev_main, vdevsym, sz_tunnel, sz_buffer, 
            direct_pool, indirect_pool);

        if (ret < 0) goto err;

        struct vdev * vdev = vdev_lookup(sc->vdev_main, vdevsym);
        MR_VERIFY_2(vdev != NULL, "vdev_lookup() returns NULL.");
    }

    return RT_SUCCESS;

err:
    MR_ERROR("Loop devices defined by user setup failed. ");
    return RT_ERR;
}

static int __vdev_data_setup_one_device(struct sc_main * sc, const char * vdevsym,
    unsigned int nr_rxstream, unsigned int nr_txstream)
{
    struct rte_mempool * direct_pool = NULL;
    struct rte_mempool * indirect_pool = NULL;

	char sym_direct_mempool[MR_SYMBOL_MAX] = { 0 };
	char sym_indirect_mempool[MR_SYMBOL_MAX] = { 0 };
    vdev_usercfg_query_mempool(sc, vdevsym, sym_direct_mempool, sym_indirect_mempool);

    /* 查询Indirect Pool，使用与物理设备对应的Indirect Pool */
    direct_pool = mrb_direct_mempool_locate(sc->mrb_pool_main, sym_direct_mempool, 0, 0);
    if (direct_pool == NULL)
    {
        MR_ERROR("Direct mempool %s for virtual device %s is not existed. ",
            sym_direct_mempool, vdevsym); goto err;
    }

    indirect_pool = mrb_indirect_mempool_locate(sc->mrb_pool_main, sym_indirect_mempool, 0, 0);
    if (indirect_pool == NULL)
    {
        MR_ERROR("Indirect mempool %s for virtual device %s is not existed. ",
            sym_indirect_mempool, vdevsym); goto err;
    }

    /* 查询队列缓冲区大小 */
    unsigned int sz_tunnel, sz_buffer;
    vdev_usercfg_query_vdev_setup(sc, vdevsym, &sz_tunnel, &sz_buffer);

    /* 创建对应的虚拟设备 */
    int ret = vdev_data_create(sc->vdev_main, vdevsym, nr_rxstream, nr_txstream,
        sz_tunnel, sz_buffer, direct_pool, indirect_pool);

    if (ret < 0) goto err;

    struct vdev * vdev = vdev_lookup(sc->vdev_main, vdevsym);
    MR_VERIFY_2(vdev != NULL, "vdev_lookup() returns NULL.");

    /* 读IP地址、掩码、网关设置 */
    struct in_addr addr = { 0 };
    struct in_addr mask = { 0 };
    struct in_addr gateway = { 0 };
    vdev_usercfg_query_vdev_in_addr(sc, vdevsym, &addr, &mask, &gateway);
    vdev_set_inaddr(vdev, addr, mask, gateway);
    vdev_enable(vdev);

    /* 查询物理设备，如果有对应的物理设备，复制对应的MAC地址、MTU等参数 */
    struct phydev * phydev = phydev_lookup(sc->phydev_main, vdevsym);
    if (phydev == NULL) return RT_SUCCESS;

    vdev_set_ether_addr(vdev, &phydev->ether_addr);
    vdev_set_mtu(vdev, phydev->mtu);

    return RT_SUCCESS;

err:
    return RT_ERR;
}

/* 根据配置文件创建虚设备 */
static int vdev_data_usercfg_setup(struct sc_main * sc)
{
    /* 加载虚设备配置，虚设备和物理设备一一对应 */
    char vdevsyms[MR_SYMBOL_MAX][MR_VDEV_MAX] = { { 0 } };
    unsigned int nr_vdevsyms = 0;
    vdev_usercfg_query_vdev_list(sc, "device", vdevsyms, &nr_vdevsyms);

    /* 计算队列数 */
    unsigned int nr_rxstream = sc->nr_serv_thread;
    unsigned int nr_txstream = sc->nr_serv_thread;

    for (int i = 0; i < nr_vdevsyms; i++)
    {
        const char * vdevsym = vdevsyms[i];
        int ret = __vdev_data_setup_one_device(sc, vdevsym, nr_rxstream, nr_txstream);
        if (ret < 0) goto err;
    }

    /* 原始设备套接字 */
    memset(vdevsyms, 0, sizeof(vdevsyms));
    vdev_usercfg_query_vdev_list(sc, "rawsocket", vdevsyms, &nr_vdevsyms);

    for (int i = 0; i < nr_vdevsyms; i++)
    {
        const char * vdevsym = vdevsyms[i];
        int ret = __vdev_data_setup_one_device(sc, vdevsym, nr_rxstream, nr_txstream);
        if (ret < 0) goto err;
    }

    /* Virtio */
    memset(vdevsyms, 0, sizeof(vdevsyms));
    vdev_usercfg_query_vdev_list(sc, "virtio_user", vdevsyms, &nr_vdevsyms);

    for (int i = 0; i < nr_vdevsyms; i++)
    {
        const char * vdevsym = vdevsyms[i];
        int ret = __vdev_data_setup_one_device(sc, vdevsym, nr_rxstream, nr_txstream);
        if (ret < 0) goto err;
    }

    return RT_SUCCESS;

err:
    MR_ERROR("Virtual devices defined by user setup failed. ");
    return RT_ERR;
}

int vdev_main_init(struct sc_main * sc)
{
	sc->vdev_main = ZMALLOC(sizeof(struct vdev_main));
	MR_VERIFY_MALLOC(sc->vdev_main);

	/* 加载公共参数配置 */
	MESA_load_profile_uint_def(sc->local_cfgfile, "device", "default_sz_tunnel",
		&sc->vdev_main->sz_tunnel, VDEV_DEFAULT_SZ_TUNNEL);
	MESA_load_profile_uint_def(sc->local_cfgfile, "device", "default_sz_buffer",
		&sc->vdev_main->sz_buffer, VDEV_DEFAULT_SZ_BUFFER);

    /* 数据分发虚拟设备创建 */
    if (vdev_data_usercfg_setup(sc) == RT_ERR)
        return RT_ERR;

    /* 本地回环虚拟设备创建 */
    if (vdev_loop_usercfg_setup(sc) == RT_ERR)
        return RT_ERR;

    /* 注册消息处理函数 */
    ctrlmsg_msg_reciver_register(sc->ctrlmsg_handler, CTRLMSG_TOPIC_VDEV_OPEN,
        CTRL_MSG_TYPE_REQUEST, __vdev_open_request_handler, sc->vdev_main);

    /* 注册应用注册、反注册处理函数，主要用于释放资源 */
    app_event_handler_register(sc->app_main, APP_EV_TYPE_REGISTER, 
        __vdev_app_register_handler, sc->vdev_main);
    app_event_handler_register(sc->app_main, APP_EV_TYPE_UNREGISTER,
        __vdev_app_unregister_handler, sc->vdev_main);
    	
	return RT_SUCCESS;
}

int vdev_dump(struct sc_main * sc)
{
    MR_INFO(" ");
    for (int i = 0; i < sc->vdev_main->vdev_max_idx; i++)
    {
        struct _vdev * _vdev_iter = sc->vdev_main->_vdev_array[i];
        MR_INFO("Virtual Device %s:", _vdev_iter->vdev.symbol);
        MR_INFO("   RX thread count                : %d", _vdev_iter->nr_rxstream);
        MR_INFO("   TX thread count                : %d", _vdev_iter->nr_txstream);
        MR_INFO("   Packet deliver ring size       : %d", _vdev_iter->sz_tunnel);
        MR_INFO("   Packet deliver buffer size     : %d", _vdev_iter->sz_buffer);

        /* IP地址没有配置，忽略IP类信息的输出 */
        if (_vdev_iter->vdev.in_addr.s_addr != 0)
        {
            MR_INFO("   IP Address                     : %s", inet_ntoa(_vdev_iter->vdev.in_addr));
            MR_INFO("   Subnet mask                    : %s", inet_ntoa(_vdev_iter->vdev.in_mask));
            MR_INFO("   Gateway                        : %s", inet_ntoa(_vdev_iter->vdev.in_gateway));
        }

        MR_INFO(" ");
    }

    return 0;
}