Question

我正在编写RDMA（InfiniBand）内核模块。

到目前为止，我已成功创建保护域，发送和接收队列的完成队列。

但是每当我尝试通过调用ib_create_qp来创建队列对时，它都无法创建队列对。我写的代码如下所示：

#include <linux/kernel.h>
#include <linux/init.h>
#include <linux/module.h>
#include <linux/list.h>
#include <linux/module.h>
#include <linux/err.h>
#include "myClient.h"


struct workqueue_struct *myClient_workqueue;
struct ib_sa_client myClient_sa_client;
/*
static void myClient_add_one(struct ib_device *device);
static void myClient_remove_one(struct ib_device *device);
*/

struct ib_pd *mypd;
struct ib_cq *myrcvcq;
struct ib_cq *myClientsendcq;
struct ib_qp *myClientqp;

void myClient_ib_recvcompletion(struct ib_cq *cq)
{
    printk("A user-specified callback that is invoked when a completion event occurs on the CQ.\n");
}


void myClient_ib_sendcompletion(struct ib_cq *cq)
{
        printk("A user-specified callback that is invoked when a completion event occurs on the CQ.\n");
}
static void my_qp_event_handler(struct ib_event *myqpAsyncEvent, void *anyPointer)
{
        printk(KERN_INFO "Dummy affiliated asynchronous event occured function called \n");
}


static void myClient_add_one(struct ib_device *device)
{
    union ib_gid tmp_gid;
    int ret;
    int hcaport = 1;
    int result = -ENOMEM;
    u16 port1Pkey;
    struct ib_port_attr attr;

        ret = ib_query_port(device,hcaport,&attr);
        printk("ib query port result %d  \n", ret);

//  Creating the Protection Domain for RDMA
    mypd = ib_alloc_pd(device);

    if(IS_ERR(mypd)){
        printk(KERN_INFO "Failed to allocate PD\n");
        return;
    }
    else{
        printk(KERN_INFO "1Successfully allocated the PD\n");
        pdset = true;
    }

//  Creating the receive completion queue for RDMA
    myrcvcq = ib_create_cq(device,myClient_ib_recvcompletion,NULL,NULL,myClient_recvq_size,0);
        if(IS_ERR(myrcvcq)){
                pr_err("%s:%d error code for receive cq%d\n", __func__, __LINE__, PTR_ERR(myrcvcq));
                //printk("Error creating QP: %d \n",PTR_ERR(myClientqp));
        }
    else{
        printk("Recieve CQ successfully created in address: %x \n",myrcvcq);
    }

//  Creating the send completion queue for RDMA
    myClientsendcq = ib_create_cq(device,myClient_ib_sendcompletion, NULL, NULL,myClient_sendq_size,0 );
        if(IS_ERR(myClientsendcq)){
                pr_err("%s:%d scqerror code for send cq%d\n", __func__, __LINE__, PTR_ERR(myClientsendcq));
                //printk("Error creating QP: %d \n",PTR_ERR(myClientqp));
        }
        else{
                printk("1Send CQ successfully created in address: %x \n",myClientsendcq);
        }

//  Creating the queue pair
//      Creating the queue pair

        struct ib_qp_init_attr init_qpattr;

        memset(&init_qpattr,0,sizeof(init_qpattr));
        init_qpattr.event_handler = myClient_qp_event_handler;
        init_qpattr.cap.max_send_wr = 2;
        init_qpattr.cap.max_recv_wr = 2;
        init_qpattr.cap.max_recv_sge = 1;
        init_qpattr.cap.max_send_sge = 1;
        init_qpattr.sq_sig_type = IB_SIGNAL_ALL_WR;
        init_qpattr.qp_type = IB_QPT_UD;
        init_qpattr.send_cq = myClientsendcq;
        init_qpattr.recv_cq = myrcvcq;

        myClientqp = ib_create_qp(mypd,&init_qpattr);

        if(IS_ERR(myClientqp)){
                pr_err("%s:%d error code %d\n", __func__, __LINE__, PTR_ERR(myClientqp));
                //printk("Error creating QP: %d \n",PTR_ERR(myClientqp));
        }
        else{
                printk(KERN_INFO "1The queue pair is successfully created \n");
                qpcreated = true;
        }



}
static void myClient_remove_one(struct ib_device *device)
{
}

static struct ib_client my_client = {
        .name   = "myRDMAclient",
        .add    = myClient_add_one,
        .remove = myClient_remove_one
};


static int __init myRDMAclient_init(void)
{
    int ret;

    ret = ib_register_client(&my_client);
    if(ret){
        //printk(KERN_ALERT "KERN_ERR Failed to register IB client\n");
        goto err_sa;
    }
    printk(KERN_ALERT "lKERN_INFO Successfully registered myRDMAclient module \n");
    return 0;

err_sa:


    return ret;
}


module_init(myRDMAclient_init);

此处所有查询都有效，但ib_create_qp(mypd,&init_qpattr);无法创建队列对。

更新：在创建队列对之前注册内存。但仍然显示ib_create_qp无效参数错误（错误代码-22）

#include <linux/kernel.h>
#include <linux/init.h>
#include <linux/module.h>
#include <linux/list.h>
#include <linux/module.h>
#include <linux/err.h>
#include "myClient.h"


struct workqueue_struct *myClient_workqueue;
struct ib_sa_client myClient_sa_client;
/*
static void myClient_add_one(struct ib_device *device);
static void myClient_remove_one(struct ib_device *device);
*/

struct ib_pd *mypd;
struct ib_cq *myrcvcq;
struct ib_cq *myClientsendcq;
struct ib_qp *myClientqp;
struct ib_mr *mymr;

void myClient_ib_recvcompletion(struct ib_cq *cq)
{
    printk("A user-specified callback that is invoked when a completion event occurs on the CQ.\n");
}


void myClient_ib_sendcompletion(struct ib_cq *cq)
{
        printk("A user-specified callback that is invoked when a completion event occurs on the CQ.\n");
}
static void my_qp_event_handler(struct ib_event *myqpAsyncEvent, void *anyPointer)
{
        printk(KERN_INFO "Dummy affiliated asynchronous event occured function called \n");
}


static void myClient_add_one(struct ib_device *device)
{
    union ib_gid tmp_gid;
    int ret;
    int hcaport = 1;
    int result = -ENOMEM;
    u16 port1Pkey;
    struct ib_port_attr attr;

        ret = ib_query_port(device,hcaport,&attr);
        printk("ib query port result %d  \n", ret);

//  Creating the Protection Domain for RDMA
    mypd = ib_alloc_pd(device);

    if(IS_ERR(mypd)){
        printk(KERN_INFO "Failed to allocate PD\n");
        return;
    }
    else{
        printk(KERN_INFO "1Successfully allocated the PD\n");
        pdset = true;
    }
// Registering Memory
    mymr = ib_get_dma_mr(mypd,IB_ACCESS_LOCAL_WRITE | IB_ACCESS_REMOTE_READ| IB_ACCESS_REMOTE_WRITE);
    if(IS_ERR(mymr)){
            printk("failed to register memory :( %d \n",PTR_ERR(mymr));
    }else{
            printk(KERN_INFO "Successfully registered memory region :) \n");
    }
// End Registering Memory
//  Creating the receive completion queue for RDMA
    myrcvcq = ib_create_cq(device,myClient_ib_recvcompletion,NULL,NULL,myClient_recvq_size,0);
        if(IS_ERR(myrcvcq)){
                pr_err("%s:%d error code for receive cq%d\n", __func__, __LINE__, PTR_ERR(myrcvcq));
                //printk("Error creating QP: %d \n",PTR_ERR(myClientqp));
        }
    else{
        printk("Recieve CQ successfully created in address: %x \n",myrcvcq);
    }

//  Creating the send completion queue for RDMA
    myClientsendcq = ib_create_cq(device,myClient_ib_sendcompletion, NULL, NULL,myClient_sendq_size,0 );
        if(IS_ERR(myClientsendcq)){
                pr_err("%s:%d scqerror code for send cq%d\n", __func__, __LINE__, PTR_ERR(myClientsendcq));
                //printk("Error creating QP: %d \n",PTR_ERR(myClientqp));
        }
        else{
                printk("1Send CQ successfully created in address: %x \n",myClientsendcq);
        }

//  Creating the queue pair
//      Creating the queue pair

        struct ib_qp_init_attr init_qpattr;

        memset(&init_qpattr,0,sizeof(init_qpattr));
        init_qpattr.event_handler = myClient_qp_event_handler;
        init_qpattr.cap.max_send_wr = 2;
        init_qpattr.cap.max_recv_wr = 2;
        init_qpattr.cap.max_recv_sge = 1;
        init_qpattr.cap.max_send_sge = 1;
        init_qpattr.sq_sig_type = IB_SIGNAL_ALL_WR;
        init_qpattr.qp_type = IB_QPT_UD;
        init_qpattr.send_cq = myClientsendcq;
        init_qpattr.recv_cq = myrcvcq;

        myClientqp = ib_create_qp(mypd,&init_qpattr);

        if(IS_ERR(myClientqp)){
                pr_err("%s:%d error code %d\n", __func__, __LINE__, PTR_ERR(myClientqp));
                //printk("Error creating QP: %d \n",PTR_ERR(myClientqp));
        }
        else{
                printk(KERN_INFO "1The queue pair is successfully created \n");
                qpcreated = true;
        }



}
static void myClient_remove_one(struct ib_device *device)
{
}

static struct ib_client my_client = {
        .name   = "myRDMAclient",
        .add    = myClient_add_one,
        .remove = myClient_remove_one
};


static int __init myRDMAclient_init(void)
{
    int ret;

    ret = ib_register_client(&my_client);
    if(ret){
        //printk(KERN_ALERT "KERN_ERR Failed to register IB client\n");
        goto err_sa;
    }
    printk(KERN_ALERT "lKERN_INFO Successfully registered myRDMAclient module \n");
    return 0;

err_sa:


    return ret;
}


module_init(myRDMAclient_init);

Answer 1

<强>更新：

根据以下评论中的讨论，我猜你在当前版本的基础上安装了Mellanox OFED驱动程序。看看Mellanox OFED内核驱动程序的3.1-1.0.3源代码，我看到他们通过添加一些字段来改变struct ib_qp_init_attr的布局。我很确定您的问题是您正在针对原始SLE 3.0.76-0.11内核头文件构建模块，因此传递给create QP函数的init_qpattr结构没有您设置的值在正确的地方。

我不知道您是如何安装新的树外驱动程序的，所以我无法准确地告诉您如何正确构建模块，但您可以尝试添加类似

的内容

    init_qpattr.qpg_type = 0;

到你设置结构的地方。（我知道你memset已经将整个事情归零了，但是这将确保你构建的标题具有结构的新qpg_type成员。我认为这是一个新的字段添加了OFED不在您的原始内核头文件中，因此如果您的模块编译，那么您正在构建正确的标头）

OLD ANSWER：

所以我怀疑你遇到了与创建这么小的QP（max_send_wr == max_recv_wr == 2和max_send_sge == max_recv_sge == 1）相关的mlx4驱动程序中的错误。我设法找到你正在使用的3.0.76-0.11内核的源代码，不幸的是我没有看到任何明显的错误。

您可以尝试帮助调试此内容

加载模块参数debug_level=1到mlx4_core模块。使用驱动程序初始化的所有输出更新您的问题（关于“最大CQE：”的一堆行等.mlx4驱动程序中有相当多的逻辑，它依赖于fimrware在初始化期间返回的参数，并且此输出将允许我们看看那些是什么。
就此而言，值得检查一下你的HCA固件是否是最新的 - 使用更新的固件你可能会得到更好的结果（尽管驱动程序应该可以工作，你可能会遇到未经测试的驱动程序代码中的错误，因为缺少触发不同代码路径的固件功能。）
尝试更新代码以增加这些参数。您可以尝试将max_send_sge和max_recv_sge增加到2并将max_send_wr和max_recv_wr增加到32或128.（尝试单独增加或组合使用）
如果你知道如何启用函数跟踪器（This LWN article是有用的;我假设旧的SLES内核具有所有必需的功能），然后启用对mlx4_ib和mlx4_core模块的跟踪，然后加载模块会很好。如果您使用跟踪更新问题，那么我们可以查看创建QP操作失败的位置 - 例如，它是set_rq_size()失败，是set_kernel_sq_size()还是失败？

Answer 2

我想你忘了注册记忆区域。在创建QP之前需要执行的操作是：

创建保护域
注册记忆区域
创建完成队列

然后才创建QP。

我不知道你正在使用什么设备和免费软件，但是在Mellanox IB lib中它是：

char mr_buffer[REGION_SIZE];
//mypd its your protection domain that you allocated 
struct ibv_mr *mr = ibv_reg_mr(mypd , mr_buffer, REGION_SIZE, 0);
if (!mr) {
    //ERROR MSG
}

无法使用ib_create_qp

2 个答案: