io_uring和epoll的区别

（1）epoll设置事件完成之后，以后只要不修改或删除事件，就可以一直等待IO事件触发。即事件驱动机制。

epoll对事件的管理使用的是红黑树。

（2）io_uring有两个队列，SQ和CQ，io_uring_submite之后，事件提交在SQ等待，事件达到后交给CQ，应用程序调用io_uring_peek_batch_cqe从CQ取出后，会调用io_uring_cq_advance将事件触发销毁，因此要想一直可以等待事件，

需要从CQ取出后再次把事件加入SQ中

。即异步机制。

io_uring对事件的管理采用两个队列：SQ（submition queue）和CQ（completion queue）。

io_uring 与epoll性能比较

安装rust_echo_bench测试工具

（1）安装cargo

ubuntu

sudo apt-get install cargo

centos

sudo yum install cargo

（2）下载rust_echo_bench

git clone https://github.com/haraldh/rust_echo_bench.git

（3）编译rust_echo_bench

cd rust_echo_bench
cargo run --release -- --help

过程

    Updating crates.io index
  Downloaded unicode-width v0.1.9
  Downloaded getopts v0.2.21
  Downloaded 2 crates (35.2 KB) in 1.58s
   Compiling unicode-width v0.1.9
   Compiling getopts v0.2.21
   Compiling echo_bench v0.2.0 (/home/user/rust_echo_bench)
    Finished release [optimized] target(s) in 4m 05s
     Running `target/release/echo_bench --help`
Echo benchmark.

Usage:
  target/release/echo_bench [ -a <address> ] [ -l <length> ] [ -c <number> ] [ -t <duration> ]
  target/release/echo_bench (-h | --help)
  target/release/echo_bench --version

Options:
    -h, --help          Print this help.
    -a, --address <address>
                        Target echo server address. Default: 127.0.0.1:12345
    -l, --length <length>
                        Test message length. Default: 512
    -t, --duration <duration>
                        Test duration in seconds. Default: 60
    -c, --number <number>
                        Test connection number. Default: 50

（3）运行

cargo run --release -- --address "192.168.7.233:9999" --number 1000 --duration 60 --length 512

将相关参数修改即可

测试比较

io_uring结果：

    Finished release [optimized] target(s) in 0.00s
     Running `target/release/echo_bench --address '192.168.7.233:9999' --number 1000 --duration 60 --length 512`
Benchmarking: 192.168.7.235:9999
1000 clients, running 512 bytes, 60 sec.

Speed: 8836 request/sec, 8836 response/sec
Requests: 530176
Responses: 530176

epoll测试前，需要先调整ulimit大小：ulimit -n 1024567。

epoll结果：

    Finished release [optimized] target(s) in 0.01s
     Running `target/release/echo_bench --address '192.168.7.233:8888' --number 1000 --duration 60 --length 512`
Benchmarking: 192.168.7.233:8888
1000 clients, running 512 bytes, 60 sec.

Speed: 7908 request/sec, 7908 response/sec
Requests: 474517
Responses: 474516

注意，此测试结果仅仅是某次数据或访问的比较，为自己编写的代码提供一种测试方案，不作为说明io_uring与epoll/poll/select的性能高低。

小结：io_uring在性能上不比reactor高多少，io_uring不一定会完全替代epoll，未来是io_uring与epoll并存，只是网络IO事件的处理方案多了一个选择。

实现封装io_uring用户态文件读写接口

io_uring提供三个系统调用接口：io_uring_submit()、io_uring_enter()、io_uring_register()。不使用liburing情况下，需要自己实现用户层的接口。io_uring除了可以处理网络IO，也可以处理磁盘IO事件。

系统调用

系统调用函数是syscall。调用syscall函数时，会触发一个0x80的中断。每个系统调用都有系统调用号，存放在sys_call_table数组中。

流程：调用syscall，触发0x80中断，将系统调用号赋给eax寄存器，参数赋给ebx,ecx等寄存器，然后执行system_call。

函数原型：

#define _GNU_SOURCE         /* See feature_test_macros(7) */
#include <unistd.h>
#include <sys/syscall.h>   /* For SYS_xxx definitions */

long syscall(long number, ...);

使用示例：

#define _GNU_SOURCE
#include <unistd.h>
#include <sys/syscall.h>
#include <sys/types.h>
#include <signal.h>

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

    tid = syscall(SYS_gettid);
    tid = syscall(SYS_tgkill, getpid(), tid, SIGHUP);
}

内存映射mmap

应用层和内核直接数据交互，可以通过映射内存块的方式。

函数原型：

 #include <sys/mman.h>

void *mmap(void *addr, size_t length, int prot, int flags,
           int fd, off_t offset);
int munmap(void *addr, size_t length);

使用示例：

#include <sys/mman.h>
#include <sys/stat.h>
#include <fcntl.h>
#include <stdio.h>
#include <stdlib.h>
#include <unistd.h>

#define handle_error(msg) \
    do { perror(msg); exit(EXIT_FAILURE); } while (0)

int
main(int argc, char *argv[])
{
    char *addr;
    int fd;
    struct stat sb;
    off_t offset, pa_offset;
    size_t length;
    ssize_t s;

    if (argc < 3 || argc > 4) {
        fprintf(stderr, "%s file offset [length]\n", argv[0]);
        exit(EXIT_FAILURE);
    }

    fd = open(argv[1], O_RDONLY);
    if (fd == -1)
        handle_error("open");

    if (fstat(fd, &sb) == -1)           /* To obtain file size */
        handle_error("fstat");

    offset = atoi(argv[2]);
    pa_offset = offset & ~(sysconf(_SC_PAGE_SIZE) - 1);
        /* offset for mmap() must be page aligned */

    if (offset >= sb.st_size) {
        fprintf(stderr, "offset is past end of file\n");
        exit(EXIT_FAILURE);
    }

    if (argc == 4) {
        length = atoi(argv[3]);
        if (offset + length > sb.st_size)
            length = sb.st_size - offset;
                /* Can't display bytes past end of file */

    } else {    /* No length arg ==> display to end of file */
        length = sb.st_size - offset;
    }

    addr = mmap(NULL, length + offset - pa_offset, PROT_READ,
                MAP_PRIVATE, fd, pa_offset);
    if (addr == MAP_FAILED)
        handle_error("mmap");

    s = write(STDOUT_FILENO, addr + offset - pa_offset, length);
    if (s != length) {
        if (s == -1)
            handle_error("write");

        fprintf(stderr, "partial write");
        exit(EXIT_FAILURE);
    }

    exit(EXIT_SUCCESS);
}

SQ_RING、CQ_RING、SQES关系

SQ_RING和CQ_RIeNG的各项偏移指向SQES，真正存储数据的是SQES。

代码实现示例

#include <stdio.h>
#include <stdlib.h>
#include <sys/stat.h>
#include <sys/ioctl.h>
#include <sys/syscall.h>
#include <sys/mman.h>
#include <sys/uio.h>
#include <linux/fs.h>
#include <fcntl.h>
#include <unistd.h>
#include <string.h>

#include <linux/io_uring.h>

#define URING_QUEUE_DEPTH	1024
#define BLOCK_SZ			1024

// 定义结构体，将内核数据mmap到用户层
struct app_io_sq_ring {
	unsigned *head;
	unsigned *tail;

	unsigned *ring_mask;
	unsigned *ring_entries;

	unsigned *flags;
	unsigned *array;
};

struct app_io_cq_ring {
	unsigned *head;
	unsigned *tail;

	unsigned *ring_mask;
	unsigned *ring_entries;

	struct io_uring_cqe *cqes;
};

// 返回
struct submitter {
	int ring_fd;

	struct app_io_sq_ring sq_ring;
	struct app_io_cq_ring cq_ring;

	struct io_uring_sqe *sqes;
};

// file info
struct file_info {
	off_t file_sz;
	struct iovec iovecs[];
};

int io_uring_setup(unsigned entries, struct io_uring_params *p)
{
	// io_uring.c
	// SYSCALL_DEFINE2(io_uring_setup,u32,entries,struct io_uring_params __user *,params)
	// 内核代码 io_uring_setup
	return (int)syscall(__NR_io_uring_setup, entries, p);
	
}

int io_uring_enter(int ring_fd, unsigned int to_submit,
	unsigned int min_complete, unsigned int flags)
{
	return (int)syscall(__NR_io_uring_enter, ring_fd, to_submit, min_complete, flags, NULL, 0);
}

int app_setup_uring(struct submitter *s)
{
	struct io_uring_params p;
	memset(&p, 0, sizeof(p));

	// 创建 CQ和SQ 内存空间
	s->ring_fd = io_uring_setup(URING_QUEUE_DEPTH, &p);
	if (s->ring_fd < 0)
		return -1;

	int sring_sz = p.sq_off.array + p.sq_entries * sizeof(unsigned);
	int cring_sz = p.cq_off.cqes + p.cq_entries * sizeof(struct io_uring_cqe);

	// 单映射，判断cq和sq是否公用一块内存
	if (p.features & IORING_FEAT_SINGLE_MMAP)
	{
		if (cring_sz > sring_sz)
			sring_sz = cring_sz;
		cring_sz = sring_sz;
	}

	// mmap sq空间
	void *sq_ptr = mmap(0, sring_sz, PROT_READ | PROT_WRITE,
		MAP_SHARED | MAP_POPULATE, s->ring_fd, IORING_OFF_SQ_RING);
	if (sq_ptr == MAP_FAILED)
		return -1;

	// 单映射，判断cq和sq是否公用一块内存
	void *cq_ptr = NULL;
	if (p.features & IORING_FEAT_SINGLE_MMAP)
	{
		cq_ptr = sq_ptr;
	}
	else
	{
		// mmap cq空间
		cq_ptr = mmap(0, sring_sz, PROT_READ | PROT_WRITE,
			MAP_SHARED | MAP_POPULATE, s->ring_fd, IORING_OFF_CQ_RING);
		if (cq_ptr == MAP_FAILED)
			return -1;
	}

	struct app_io_sq_ring *sring = &s->sq_ring;
	struct app_io_cq_ring *cring = &s->cq_ring;

	// sq赋值
	sring->head = sq_ptr + p.sq_off.head;
	sring->tail = sq_ptr + p.sq_off.tail;
	sring->ring_mask= sq_ptr + p.sq_off.ring_mask;
	sring->ring_entries = sq_ptr + p.sq_off.ring_entries;
	sring->flags = sq_ptr + p.sq_off.flags;
	sring->array = sq_ptr + p.sq_off.array;
	// 将sqes映射到用户空间
	s->sqes = mmap(0, p.sq_entries * sizeof(struct io_uring_sqe),PROT_READ | PROT_WRITE,
		MAP_SHARED | MAP_POPULATE, s->ring_fd, IORING_OFF_SQES);
	if (s->sqes == MAP_FAILED)
		return -1;

	// cq赋值
	cring->head = cq_ptr + p.cq_off.head;
	cring->tail = cq_ptr + p.cq_off.tail;
	cring->ring_mask = cq_ptr + p.cq_off.ring_mask;
	cring->ring_entries = cq_ptr + p.cq_off.ring_entries;
	cring->cqes = sq_ptr + p.cq_off.cqes;

	return 0;
}

// get file size
off_t get_file_size(int fd)
{
	struct stat st;
	if (fstat(fd, &st) < 0)
		return -1;
	if (S_ISBLK(st.st_mode))
	{
		unsigned long long bytes;
		if (ioctl(fd, BLKGETSIZE64, &bytes) != 0)
			return -1;
		return bytes;
	}
	else if (S_ISREG(st.st_mode))
		return st.st_size;

	return -1;
}

void output_to_console(char *buf,int len)
{
	while (len--)
	{
		fputc(*buf++, stdout);
	}
}

void read_from_cq(struct submitter *s)
{
	struct file_info *fi;
	struct app_io_cq_ring *cring = &s->cq_ring;
	struct io_uring_cqe *cqe;

	unsigned head = *cring->head;
	while (1)
	{
		if (head == *cring->tail)
			break;

		cqe = &cring->cqes[head&*s->cq_ring.ring_mask];
		fi = (struct file_info*)cqe->user_data;

		if (cqe->res < 0)
			fprintf(stderr, "error: %d\n", cqe->res);

		int blocks = fi->file_sz / BLOCK_SZ;
		if (fi->file_sz%BLOCK_SZ)
			blocks++;

		int i = 0;
		while (++i < blocks)
		{
			output_to_console(fi->iovecs[i].iov_base, fi->iovecs[i].iov_len);
			printf("################################ i : %d, blocks : %d\n",i, blocks);
		}
		head++;
		printf("head : %d,tail : %d, blocks: %d\n", head, *cring->tail,blocks);
	}

	*cring->head = head;

	printf("read form cq end!\n");
}

int submit_to_sq(char *file_path,struct submitter *s)
{
	int filefd = open(file_path, O_RDONLY);
	if (filefd < 0)
		return -1;
	
	off_t filesz = get_file_size(filefd);
	if (filesz < 0) {
		close(filefd);
		return -1;
	}
		

	struct app_io_sq_ring *sring = &s->sq_ring;

	off_t bytes_remaining = filesz;

	int blocks = filesz / BLOCK_SZ;
	if (filesz%BLOCK_SZ)
		blocks++;

	struct file_info *fi = malloc(sizeof(struct file_info) + sizeof(struct iovec)*blocks);
	if (!fi)
	{
		close(filefd);
		return -2;
	}
		

	fi->file_sz = filesz;

	unsigned current_block=0;
	while (bytes_remaining)
	{
		off_t bytes_to_read = bytes_remaining;
		if (bytes_to_read > BLOCK_SZ)
			bytes_to_read = BLOCK_SZ;

		fi->iovecs[current_block].iov_len = bytes_to_read;

		void *buf;
		if (posix_memalign(&buf, BLOCK_SZ, BLOCK_SZ))
		{
			close(filefd);
			return 1;
		}

		fi->iovecs[current_block].iov_base = buf;

		current_block++;

		bytes_remaining -= bytes_to_read;
	}
	unsigned next_tail = 0, tail = 0, index = 0;

	next_tail = tail = *sring->tail;
	next_tail++;
	index = tail &*s->sq_ring.ring_mask;

	struct io_uring_sqe *sqe = &s->sqes[index];
	sqe->fd = filefd;
	sqe->flags = 0;
	sqe->opcode = IORING_OP_READV;
	sqe->addr = (unsigned long)fi->iovecs;
	sqe->len = blocks;
	sqe->off = 0;
	sqe->user_data = (unsigned long long)fi;

	sring->array[index] = index;
	tail = next_tail;

	if (*sring->tail != tail)
		*sring->tail = tail;

	int ret = io_uring_enter(s->ring_fd, 1, 1, IORING_ENTER_GETEVENTS);
	if (ret < 0)
	{
		close(filefd);
		return -1;
	}
		
	close(filefd);
	return 0;

}

int main(int argc, char *argv[])
{
	struct submitter *s = malloc(sizeof(struct submitter));
	if (s == NULL)
	{
		perror("malloc fail");
		return -1;
	}
	memset(s, 0, sizeof(struct submitter));

	if (app_setup_uring(s))
		return 1;

	int i = 1;
	for (i = 1; i < argc; i++)
	{
		if (submit_to_sq(argv[i], s))
			return 1;

		read_from_cq(s);
	}

	return 0;
}

总结

io_uring比epoll好的点是io_uring使用共享内存，不仅仅共享了IO事件，需要的变量也通过共享内存共享到用户空间，像SQ和CQ队列。io_uring不仅可以处理网络IO，也可以处理磁盘IO。

封装io_uring处理磁盘IO事件：

（1）io_uring_setup，创建SQ和CQ以及SQES，mmap io_uring_cqe / io_uring_sqe / SQES到用户空间。

（2）submit 文件IO到SQ

（3）从CQ中读取文件IO