Linux I/O Models and io_uring Performance Overview

Linux I/O models are classified into native I/O, kernel bypass, etc. Native I/O includes synchronous and asynchronous operations. Synchronous I/O blocks the calling process until the operation completes, while asynchronous I/O allows the process to continue.

Blocking refers to the process state when a system call puts the process to sleep until the requested event occurs.

Common file operation APIs:

#include <unistd.h>
#include <sys/types.h>
#include <sys/stat.h>
#include <fcntl.h>

//  返回值:成功返回新分配的文件描述符，出错返回-1并设置errno
int open(const char *pathname, int flags);
int open(const char *pathname, int flags, mode_t mode);

// 返回值:成功返回0，出错返回-1并设置errno
int close(int fd);

// 返回值:成功返回读取的字节数，出错返回-1并设置errno，如果在调read之前已到达文件末尾，则这次read返回0
ssize_t read(int fd, void *buf, size_t count);

// 返回值:成功返回写入的字节数，出错返回-1并设置errno
ssize_t write(int fd, const void *buf, size_t count);

Typical open flags include O_CREAT, O_EXCL, O_TRUNC, O_RDONLY, O_WRONLY, O_RDWR, O_APPEND, O_NONBLOCK, O_SYNC, O_DIRECT.

Mmap maps a file into the process address space, allowing direct memory access to file data.

// 成功执行时，mmap()返回被映射区的指针。失败时，mmap()返回MAP_FAILED[其值为(void *)-1]
void *mmap(void *start, size_t length, int prot, int flags, int fd, off_t offset);

// 成功执行时，munmap()返回0。失败时，munmap返回-1，error返回标志和mmap一致；
int munmap(void *addr, size_t len);

// 将内存中的修改同步到磁盘
int msync(void *addr, size_t len, int flags);

Direct I/O (O_DIRECT) bypasses the kernel page cache, reducing data copies.

sendfile transfers data from a file descriptor to a socket without copying to user space.

splice connects two file descriptors, allowing data movement without user‑space copies; it is a superset of sendfile.

io_uring is a modern Linux asynchronous I/O interface that supports both storage and network I/O. It uses submission (SQ) and completion (CQ) queues.

Key io_uring system calls:

int io_uring_setup(u32 entries, struct io_uring_params *p);
int io_uring_register(unsigned int fd, unsigned int opcode, void *arg, unsigned int nr_args);
int io_uring_enter(unsigned int fd, unsigned int to_submit, unsigned int min_complete, unsigned int flags, sigset_t *sig);

io_uring operates in three modes: interrupt‑driven (default), polling, and kernel‑polling. Polling reduces latency at the cost of CPU usage; kernel‑polling uses a kernel thread to submit and reap I/O without user‑space syscalls.

liburing provides a convenient wrapper around io_uring. Example initialization, submission, and completion handling are shown in the source code.

Performance comparisons show that io_uring outperforms traditional libaio in polling mode and approaches the performance of SPDK at high queue depths, while its advantage over libaio in non‑polling mode is modest.

Conclusion: io_uring offers a high‑performance, flexible I/O solution for Linux, though it requires recent kernel versions.