基于NETLINK的内核与用户空间共享内存的实现[转]

基于NETLINK的内核与用户空间共享内存的实现

author:bripengandre Email:bripengandre@126.com
原文转自：http://blog.chinaunix.net/u3/94771/showart_1945422.html

一、前言
前些日子，开发中用到了netlink来实现内核与用户空间共享内存，写点笔记与大家分享。因为我对这块也不了解，写出来的东西一定存在很多错误，请大家批评指正~
内核与用户空间共享内存的关键是，用户空间必须得知共享内存的起始地址，这就要求内核空间应该有一种通信机制来通知用户空间。已经有Godbach版主等人用proc文件系统实现了（可以google '共享内存 内核 用户空间'）,很显然任何内核空间与用户空间的通信方法都可资利用。本文主要讲基于NETLINK机制的实现。

二、NETLINK简介
netlink在linux的内核与用户空间通信中用得很多（但具体例子我举不出，因为我不清楚~~请google之），其最大优势是接口与网络编程中的socket相似，且内核要主动发信息给用户空间很方便。
但通过实践，我发现netlink通信机制的最大弊病在于其在各内核版本中接口变化太大，让人难以适从（可从后文列出的源码中的kernel_receive的声明窥一斑）。
既然涉及到内核与用户空间两个空间，就应该在两个空间各有一套接口。用户空间的接口很简单，与一般的socket接口相似，内核空间则稍先复杂，但简单的应用只需简单地了解即可：首先也是建立描述符，建立描述符时会注册一个回调函数（源码中的kernel_receive即是），然后当用户空间有消息发过来时，我们的函数将被调用，显然在这个函数里我们可做相应的处理；当内核要主动发消息给用户进程时，直接调用一个类send函数即可（netlink_unicast系列函数）。当然这个过程中，有很多结构体变量需要填充。具体用法请google，我差不多忘光了~。

三、基于netlink的共享内存
这里的共享内存是指内核与用户空间，而不是通常的用户进程间的。
大概流程如下。
内核：__get_free__pages分配连续的物理内存页（貌似返回的其实是虚拟地址）-->SetPageReserved每一页（每一页都需这个操作，参见源码）-->如果用户空间通过netlink要求获取共享内存的起始物理地址，将__get_free__pages返回的地址__pa下发给用户空间。
用户空间：open "/dev/shm"（一个读写物理内存的设备，具体请google"linux读写物理内存"）-->发netlink消息给内核，得到共享内存的起始物理地址-->mmap上步得到的物理地址。

四、源码说明
正如二中提到的，netlink接口在各版本中接口变化很大，本人懒惰及时间紧，只实验了比较新的内核2.6.25，源码如需移植到老版本上，需要一些改动，敬请原谅。
另外，由于本源码是从一个比较大的程序里抠出来的，所以命名什么的可能有点怪异~
源码包括shm_k.c（内核模块）和用户空间程序（shm_u.c）两部分。shm_k.c模块的工作是：分配8KB内存供共享，并把前几十个字节置为“hello, use share memory with netlink"字样。shm_u.c工作是：读取共享内存的前几十个字节，将内容输出在stdout上。
特别说明：该程序只适用于2.6.25左右的新版本！用__get_free_pages分配连续内存时，宜先用get_order获取页数，然后需将各页都SetPageReserved下，同样地，释放内存时，需要对每一页调用ClearPageReserved。
我成功用该程序分配了4MB共享内存，运行还比较稳定。因为linux内核的默认设置，一般情况下用get_free_pages只能分配到4MB内存左右，如需增大，可能需改相应的参数并重新编译内核。

五、内核源码

1、common.h（内核与用户空间都用到的头文件）
#ifndef _COMMON_H_
#define _COMMON_H_
/* protocol type */
#define SHM_NETLINK 30
/* message type */
#define SHM_GET_SHM_INFO 1
/* you can add othe message type here */
#define SHM_WITH_NETLINK "hello, use share memory with netlink"

typedef struct _nlk_msg
{
union _data
{
struct _shm_info
{
uint32_t mem_addr;
uint32_t mem_size;
}shm_info;

/* you can add other content here */
}data;
}nlk_msg_t;

#endif /* _COMMON_H_ */

2、shm_k.c（内核模块）
#include <linux/init.h>
#include <linux/module.h>
#include <linux/version.h>
#include <linux/types.h>
#include <linux/skbuff.h>
#include <linux/netlink.h>
#include <net/sock.h>
#include <linux/spinlock.h>
#include "common.h"
#define SHM_TEST_DEBUG
#ifdef SHM_TEST_DEBUG
#define SHM_DBG(args...) printk(KERN_DEBUG "SHM_TEST: " args)
#else
#define SHM_DBG(args...)
#endif
#define SHM_ERR(args...) printk(KERN_ERR "SHM_TEST: " args)

static struct _glb_para
{
struct _shm_para
{
uint32_t mem_addr; /* memory starting address */
uint32_t mem_size; /* memory size */
uint32_t page_cnt; /* memory page count*/
uint16_t order;
uint8_t mem_init_flag; /* 0, init failed; 1, init successful */
}shm_para;

struct sock *nlfd; /* netlink descriptor */
uint32_t pid; /* user-space process's pid */
rwlock_t lock;
}glb_para;

static void init_glb_para(void);
static int init_netlink(void);
static void kernel_receive(struct sk_buff* __skb);
static int nlk_get_mem_addr(struct nlmsghdr *pnhdr);
static void clean_netlink(void);
static int init_shm(void);
static void clean_shm(void);
static int __init init_shm_test(void);
static void clean_shm_test(void);

static void init_glb_para(void)
{
memset(&glb_para, 0, sizeof(glb_para));
}

static int init_netlink(void)
{
rwlock_init(&glb_para.lock);
SHM_DBG("linux version:%08x/n", LINUX_VERSION_CODE);
#if(LINUX_VERSION_CODE < KERNEL_VERSION(2,6,18))
glb_para.nlfd = netlink_kernel_create(SHM_NETLINK, kernel_receive);
#elif(LINUX_VERSION_CODE < KERNEL_VERSION(2,6,24))
glb_para.nlfd = netlink_kernel_create(SHM_NETLINK, 0, kernel_receive, THIS_MODULE));
#else
glb_para.nlfd = netlink_kernel_create(&init_net, SHM_NETLINK, 0, kernel_receive, NULL, THIS_MODULE);
#endif

if(glb_para.nlfd == NULL)
{
SHM_ERR("init_netlink::netlink_kernel_create error/n");
return (-1);
}

return (0);
}

static void kernel_receive(struct sk_buff* __skb)
{
struct sk_buff *skb;
struct nlmsghdr *nlh = NULL;
int invalid;

SHM_DBG("begin kernel_receive/n");
skb = skb_get(__skb);
invalid = 0;
if(skb->len >= sizeof(struct nlmsghdr))
{
nlh = (struct nlmsghdr *)skb->data;
if((nlh->nlmsg_len >= sizeof(struct nlmsghdr))
&& (skb->len >= nlh->nlmsg_len))
{
switch(nlh->nlmsg_type)
{
case SHM_GET_SHM_INFO:
SHM_DBG("receiv TA_GET_SHM_INFO/n");
nlk_get_mem_addr(nlh);
break;
default:
break;
}
}
}
kfree_skb(skb);
}

static int nlk_get_mem_addr(struct nlmsghdr *pnhdr)
{
int ret, size;
unsigned char *old_tail;
struct sk_buff *skb;
struct nlmsghdr *nlh;
struct _nlk_msg *p;

glb_para.pid = pnhdr->nlmsg_pid; /* get the user-space process's pid */

size = NLMSG_SPACE(sizeof(struct _nlk_msg)); /* compute the needed memory size */
if( (skb = alloc_skb(size, GFP_ATOMIC)) == NULL) /* allocate memory */
{
SHM_DBG("nlk_hello_test::alloc_skb error./n");
return (-1);
}

old_tail = skb->tail;
nlh = NLMSG_PUT(skb, 0, 0, SHM_GET_SHM_INFO, size-sizeof(struct nlmsghdr)); /* put netlink message structure into memory */

p = NLMSG_DATA(nlh); /* get netlink message body pointer */
p->data.shm_info.mem_addr = __pa(glb_para.shm_para.mem_addr); /* __pa:convert virtual address to physical address, which needed by /dev/mem */
p->data.shm_info.mem_size = glb_para.shm_para.mem_size;

nlh->nlmsg_len = skb->tail - old_tail;
NETLINK_CB(skb).pid = 0; /* from kernel */
NETLINK_CB(skb).dst_group = 0;
read_lock_bh(&glb_para.lock);
ret = netlink_unicast(glb_para.nlfd, skb, glb_para.pid, MSG_DONTWAIT); /* send message to user-space process */
read_unlock_bh(&glb_para.lock);
SHM_DBG("nlk_get_mem_addr ok./n");
return (ret);

nlmsg_failure:
SHM_DBG("nlmsg_failure/n");
if(skb)
{
kfree_skb(skb);
}
return (-1);
}

static void clean_netlink(void)
{
if(glb_para.nlfd != NULL)
{
sock_release(glb_para.nlfd->sk_socket);
}
}

static int init_shm(void)
{
int i;
char *p;
uint32_t page_addr;

glb_para.shm_para.order = get_order(1024*8); /* allocate 8kB */
glb_para.shm_para.mem_addr = __get_free_pages(GFP_KERNEL, glb_para.shm_para.order);
if(glb_para.shm_para.mem_addr == 0)
{
SHM_ERR("init_mem_pool::__get_free_pages error./n");
glb_para.shm_para.mem_init_flag = 0;
return (-1);
}
else
{
glb_para.shm_para.page_cnt = (1<<glb_para.shm_para.order);
glb_para.shm_para.mem_size = glb_para.shm_para.page_cnt*PAGE_SIZE;
glb_para.shm_para.mem_init_flag = 1;
page_addr = glb_para.shm_para.mem_addr;
SHM_DBG("size=%08x, page_cnt=%d/n", glb_para.shm_para.mem_size, glb_para.shm_para.page_cnt);
for(i = 0; i < glb_para.shm_para.page_cnt; i++)
{
SetPageReserved(virt_to_page(page_addr)); /* reserved for used */
page_addr += PAGE_SIZE;
}

p = (char *)glb_para.shm_para.mem_addr;
strcpy(p, SHM_WITH_NETLINK); /* write */
SHM_DBG("__get_free_pages ok./n");
}

return (0);
}

static void clean_shm(void)
{
int i;
uint32_t page_addr;

if(glb_para.shm_para.mem_init_flag == 1)
{
page_addr = glb_para.shm_para.mem_addr;
for(i = 0; i < glb_para.shm_para.page_cnt; i++)
{
ClearPageReserved(virt_to_page(page_addr));
page_addr += PAGE_SIZE;
}
free_pages(glb_para.shm_para.mem_addr, glb_para.shm_para.order);
}
}

static int __init init_shm_test(void)
{
init_glb_para();
if(init_netlink() < 0)
{
SHM_ERR("init_shm_test::init_netlink error./n");
return (-1);
}
SHM_DBG("init_netlink ok./n");

if(init_shm() < 0)
{
SHM_ERR("init_shm_test::init_mem_pool error./n");
clean_shm_test();
return (-1);
}
SHM_DBG("init_mem_pool ok./n");

return (0);
}

static void clean_shm_test(void)
{
clean_shm();
clean_netlink();

SHM_DBG("ta_exit ok./n");
}
module_init(init_shm_test);
module_exit(clean_shm_test);
MODULE_LICENSE("GPL");
MODULE_AUTHOR("bripengandre (bripengandre@126.com)");
MODULE_DESCRIPTION("Memory Share between user-space and kernel-space with netlink.");

3、shm_u.c(用户进程)
#include <stdio.h>
#include <sys/socket.h>
#include <arpa/inet.h>
#include <linux/netlink.h>
#include <stdlib.h>
#include <string.h>
#include <errno.h>
#include <sys/types.h>
#include <unistd.h>
#include <fcntl.h>
#include <sys/mman.h>
#include "common.h"

/* netlink */
#define MAX_SEND_BUF_SIZE 2500
#define MAX_RECV_BUF_SIZE 2500
#define SHM_TEST_DEBUG
#ifdef SHM_TEST_DEBUG
#define SHM_DBG(args...) fprintf(stderr, "SHM_TEST: " args)
#else
#define SHM_DBG(args...)
#endif
#define SHM_ERR(args...) fprintf(stderr, "SHM_TEST: " args)
struct _glb_para
{
struct _shm_para
{
uint32_t mem_addr;
uint32_t mem_size;
}shm_para;

int nlk_fd;
char send_buf[MAX_SEND_BUF_SIZE];
char recv_buf[MAX_RECV_BUF_SIZE];
}glb_para;

static void init_glb_para(void);
static int create_nlk_connect(void);
static int nlk_get_shm_info(void);
static int init_mem_pool(void);

int main(int argc ,char *argv[])
{
char *p;

init_glb_para();
if(create_nlk_connect() < 0)
{
SHM_ERR("main::create_nlk_connect error./n");
return (1);
}

if(nlk_get_shm_info() < 0)
{
SHM_ERR("main::nlk_get_shm_info error./n");
return (1);
}

init_mem_pool();
/* printf the first 30 bytes */
p = (char *)glb_para.shm_para.mem_addr;
p[strlen(SHM_WITH_NETLINK)] = '/0';
printf("the first 30 bytes of shm are: %s/n", p);

return (0);
}

static void init_glb_para(void)
{
memset(&glb_para, 0, sizeof(glb_para));
}

static int create_nlk_connect(void)
{
int sockfd;
struct sockaddr_nl local;

sockfd = socket(PF_NETLINK, SOCK_RAW, SHM_NETLINK);
if(sockfd < 0)
{
SHM_ERR("create_nlk_connect::socket error:%s/n", strerror(errno));
return (-1);
}
memset(&local, 0, sizeof(local));
local.nl_family = AF_NETLINK;
local.nl_pid = getpid();
local.nl_groups = 0;
if(bind(sockfd, (struct sockaddr*)&local, sizeof(local)) != 0)
{
SHM_ERR("create_nlk_connect::bind error: %s/n", strerror(errno));
return -1;
}

glb_para.nlk_fd = sockfd;

return (sockfd);
}

static int nlk_get_shm_info(void)
{
struct nlmsghdr *nlh;
struct _nlk_msg *p;
struct sockaddr_nl kpeer;
int recv_len, kpeerlen;

memset(&kpeer, 0, sizeof(kpeer));
kpeer.nl_family = AF_NETLINK;
kpeer.nl_pid = 0;
kpeer.nl_groups = 0;

memset(glb_para.send_buf, 0, sizeof(glb_para.send_buf));
nlh = (struct nlmsghdr *)glb_para.send_buf;
nlh->nlmsg_len = NLMSG_SPACE(0);
nlh->nlmsg_flags = 0;
nlh->nlmsg_type = SHM_GET_SHM_INFO;
nlh->nlmsg_pid = getpid();
sendto(glb_para.nlk_fd, nlh, nlh->nlmsg_len, 0, (struct sockaddr*)&kpeer, sizeof(kpeer));
memset(glb_para.send_buf, 0, sizeof(glb_para.send_buf));
kpeerlen = sizeof(struct sockaddr_nl);
recv_len = recvfrom(glb_para.nlk_fd, glb_para.recv_buf, sizeof(glb_para.recv_buf), 0, (struct sockaddr*)&kpeer, &kpeerlen);
p = NLMSG_DATA((struct nlmsghdr *) glb_para.recv_buf);
SHM_DBG("%d, errno=%d.%s, %08x, %08x/n", recv_len, errno, strerror(errno), p->data.shm_info.mem_addr, p->data.shm_info.mem_size);
glb_para.shm_para.mem_addr = p->data.shm_info.mem_addr;
glb_para.shm_para.mem_size = p->data.shm_info.mem_size;

return (0);
}

static int init_mem_pool(void)
{
int map_fd;
void *map_addr;

map_fd = open("/dev/mem", O_RDWR);
if(map_fd < 0)
{
SHM_ERR("init_mem_pool::open %s error: %s/n", "/dev/mem", strerror(errno));
return (-1);
}

map_addr = mmap(0, glb_para.shm_para.mem_size, PROT_READ|PROT_WRITE, MAP_SHARED, map_fd, glb_para.shm_para.mem_addr);
if(map_addr == NULL)
{
SHM_ERR("init_mem_pool::mmap error: %s/n", strerror(errno));
return (-1);
}
glb_para.shm_para.mem_addr = (uint32_t)map_addr;
return (0);
}

4、Makefile
#PREFIX = powerpc-e300c3-linux-gnu-
CC ?= $(PREFIX)gcc
KERNELDIR ?= /lib/modules/`uname -r`/build

all: modules app
obj-m:= shm_k.o
module-objs := shm_k.c
modules:
make -C $(KERNELDIR) M=`pwd` modules
app: shm_u.o
$(CC) -o shm_u shm_u.c
clean:
rm -rf *.o Module.symvers modules.order shm_u shm_k.ko shm_k.mod.c .tmp_versions .shm_k.*