Soft-RoCE内核实现

参考http://hustcat.github.io/linux-soft-roce-implementation/

  内核在4.9实现的Soft-RoCE实现了RoCEv2.

队列初始化

libRXE (user space library)

ibv_create_qp
|--- rxe_create_qp
    |--- ibv_cmd_create_qp

• ibv_create_qp

LATEST_SYMVER_FUNC(ibv_create_qp, 1_1, "IBVERBS_1.1",
           struct ibv_qp *,
           struct ibv_pd *pd,
           struct ibv_qp_init_attr *qp_init_attr)
{
    struct ibv_qp *qp = pd->context->ops.create_qp(pd, qp_init_attr); ///rxe_ctx_ops
///..
}

• rxe_create_qp

static struct ibv_qp *rxe_create_qp(struct ibv_pd *pd,
                    struct ibv_qp_init_attr *attr)
{
    struct ibv_create_qp cmd;
    struct rxe_create_qp_resp resp;
    struct rxe_qp *qp;
    int ret;
////..
    ret = ibv_cmd_create_qp(pd, &qp->ibv_qp, attr, &cmd, sizeof cmd,
                &resp.ibv_resp, sizeof resp); /// ibv_create_qp CMD, to kernel
///...
    qp->sq.max_sge = attr->cap.max_send_sge;
    qp->sq.max_inline = attr->cap.max_inline_data;
    qp->sq.queue = mmap(NULL, resp.sq_mi.size, PROT_READ | PROT_WRITE,
                MAP_SHARED,
                pd->context->cmd_fd, resp.sq_mi.offset); ///mmap，参考rxe_mmap

ibv_context->cmd_fd指向对应的ibv_device，由ibv_open_device返回。

ibv_cmd_create_qp会通过ibv_context->cmd_fd给内核发送IB_USER_VERBS_CMD_CREATE_QP命令，参考libiverbs@ibv_cmd_create_qp.

对应的内核write函数为ib_uverbs_write:

///drivers/infiniband/core/uverbs_main.c
static const struct file_operations uverbs_fops = {
    .owner   = THIS_MODULE,
    .write   = ib_uverbs_write,
    .open    = ib_uverbs_open,
    .release = ib_uverbs_close,
    .llseek  = no_llseek,
};

• ibv_open_device

///libibverbs/device.c
LATEST_SYMVER_FUNC(ibv_open_device, 1_1, "IBVERBS_1.1",
           struct ibv_context *,
           struct ibv_device *device)
{
    struct verbs_device *verbs_device = verbs_get_device(device);
    char *devpath;
    int cmd_fd, ret;
    struct ibv_context *context;
    struct verbs_context *context_ex;

    if (asprintf(&devpath, "/dev/infiniband/%s", device->dev_name) < 0)
        return NULL;

    /*
     * We'll only be doing writes, but we need O_RDWR in case the
     * provider needs to mmap() the file.
     */
    cmd_fd = open(devpath, O_RDWR | O_CLOEXEC); /// /dev/infiniband/uverbs0
    free(devpath);

    if (cmd_fd < 0)
        return NULL;

    if (!verbs_device->ops->init_context) {
        context = verbs_device->ops->alloc_context(device, cmd_fd); ///rxe_alloc_context, rxe_dev_ops
        if (!context)
            goto err;
    }
///...
    context->device = device;
    context->cmd_fd = cmd_fd;
    pthread_mutex_init(&context->mutex, NULL);

    ibverbs_device_hold(device);

    return context;
///...
}

kernel (rdma_rxe module)

• ib_uverbs_create_qp

IB_USER_VERBS_CMD_CREATE_QP的处理函数为函数ib_uverbs_create_qp.

ib_uverbs_write
|--- ib_uverbs_create_qp
     |--- create_qp
          |--- ib_device->create_qp
               |--- rxe_create_qp

create_qp调用ib_device->create_qp,对于RXE， 为函数rxe_create_qp， 参考rxe_register_device.

• rxe_create_qp

rxe_create_qp
|--- rxe_qp_from_init
     |--- rxe_qp_init_req

rxe_qp_from_init完成发送队列和接收队列的初始化。

• rxe_qp_init_req

rxe_qp_init_req主要做以下一些事情：

创建对应的UDP socket

调用rxe_queue_init完成发送队列的初始化.

初始化对应的tasklet

• rxe_queue_init

rxe_queue_init给队列分配内存空间：

struct rxe_queue *rxe_queue_init(struct rxe_dev *rxe,
                 int *num_elem,
                 unsigned int elem_size)
{
    struct rxe_queue *q;
    size_t buf_size;
    unsigned int num_slots;
///...
    buf_size = sizeof(struct rxe_queue_buf) + num_slots * elem_size;

    q->buf = vmalloc_user(buf_size);
///...
}

rxe_queue->buf指向的内存缓冲区，由rxe_mmap映射到用户空间，队列的element对应数据结构struct rxe_send_wqe.

libiverbs API调用ibv_post_send时，会将对应的struct rxe_send_wqe加入到该队列，参考rdma-core@post_one_send.

• rxe_mmap

/**
 * rxe_mmap - create a new mmap region
 * @context: the IB user context of the process making the mmap() call
 * @vma: the VMA to be initialized
 * Return zero if the mmap is OK. Otherwise, return an errno.
 */
int rxe_mmap(struct ib_ucontext *context, struct vm_area_struct *vma)
{
    struct rxe_dev *rxe = to_rdev(context->device);
    unsigned long offset = vma->vm_pgoff << PAGE_SHIFT;
    unsigned long size = vma->vm_end - vma->vm_start;
    struct rxe_mmap_info *ip, *pp;
///...
found_it:
    list_del_init(&ip->pending_mmaps);
    spin_unlock_bh(&rxe->pending_lock);

    ret = remap_vmalloc_range(vma, ip->obj, 0);
    if (ret) {
        pr_err("rxe: err %d from remap_vmalloc_range\n", ret);
        goto done;
    }

    vma->vm_ops = &rxe_vm_ops;
    vma->vm_private_data = ip;
    rxe_vma_open(vma);
///...
}

发送数据

libRXE

rxe_post_send会将struct ibv_send_wr转成struct rxe_send_wqe，并加入到发送队列rxe_qp->rq，然后通过cmd_fd给RXE内核模块发送IB_USER_VERBS_CMD_POST_SEND命令：

///providers/rxe/rxe.c
/* this API does not make a distinction between
   restartable and non-restartable errors */
static int rxe_post_send(struct ibv_qp *ibqp,
             struct ibv_send_wr *wr_list,
             struct ibv_send_wr **bad_wr)
{
    int rc = 0;
    int err;
    struct rxe_qp *qp = to_rqp(ibqp);/// ibv_qp -> rxe_qp
    struct rxe_wq *sq = &qp->sq;

    if (!bad_wr)
        return EINVAL;

    *bad_wr = NULL;

    if (!sq || !wr_list || !sq->queue)
        return EINVAL;

    pthread_spin_lock(&sq->lock);

    while (wr_list) {
        rc = post_one_send(qp, sq, wr_list); /// ibv_send_wr -> rxe_send_wqe, enqueue
        if (rc) {
            *bad_wr = wr_list;
            break;
        }

        wr_list = wr_list->next;
    }

    pthread_spin_unlock(&sq->lock);

    err =  post_send_db(ibqp); /// IB_USER_VERBS_CMD_POST_SEND cmd
    return err ? err : rc;
}

kernel

处理的IB_USER_VERBS_CMD_POST_SEND的函数为ib_uverbs_post_send:

ib_uverbs_post_send -> ib_device->post_send -> rxe_post_send -> rxe_requester -> ip_local_out。

• rxe_post_send

static int rxe_post_send(struct ib_qp *ibqp, struct ib_send_wr *wr,
             struct ib_send_wr **bad_wr)
{
    int err = 0;
    struct rxe_qp *qp = to_rqp(ibqp); ///ib_qp -> rxe_qp
///...
    /*
     * Must sched in case of GSI QP because ib_send_mad() hold irq lock,
     * and the requester call ip_local_out_sk() that takes spin_lock_bh.
     */
    must_sched = (qp_type(qp) == IB_QPT_GSI) ||
            (queue_count(qp->sq.queue) > 1);

    rxe_run_task(&qp->req.task, must_sched); /// to rxe_requester

    return err;
}

• rxe_requester

rxe_requester从rxe_qp队列取出rxe_send_wqe，生成对应的skb_buff，然后下发给对应的rxe_dev设备：

///sw/rxe/rxe_req.c
int rxe_requester(void *arg)
{
    struct rxe_qp *qp = (struct rxe_qp *)arg;
    struct rxe_pkt_info pkt;
    struct sk_buff *skb;
    struct rxe_send_wqe *wqe;
///...
    wqe = req_next_wqe(qp); /// get rxe_send_wqe
///...
    /// rxe_send_wqe -> skb
    skb = init_req_packet(qp, wqe, opcode, payload, &pkt);
///...
    ret = rxe_xmit_packet(to_rdev(qp->ibqp.device), qp, &pkt, skb);
///...
}

static inline int rxe_xmit_packet(struct rxe_dev *rxe, struct rxe_qp *qp,
                  struct rxe_pkt_info *pkt, struct sk_buff *skb)
{
///...
    if (pkt->mask & RXE_LOOPBACK_MASK) {
        memcpy(SKB_TO_PKT(skb), pkt, sizeof(*pkt));
        err = rxe->ifc_ops->loopback(skb);
    } else {
        err = rxe->ifc_ops->send(rxe, pkt, skb);/// ifc_ops->send, send
    }
///...
}

ifc_ops->send最后会调用ip_local_out，从对应的物理NIC发送出去。

Refs

• SOFT-RoCE RDMA TRANSPORT IN A SOFTWARE IMPLEMENTATION
• THE LINUX SOFTROCE DRIVER