io_uring (Axboe) is a pair of shared memory rings between user space and the kernel: you push submission queue entries (SQEs) for reads, writes, accepts, timeouts, and more; the kernel pushes completion queue entries (CQEs). One syscall (io_uring_enter) can harvest many completions. Compared with epoll plus a thread pool doing blocking read, you avoid thread handoff. Compared with POSIX AIO, the API actually works for files and sockets in one model.
The rings
Map SQ and CQ. Fill an SQE with opcode, fd, buffer. Advance the tail. Completions include a result code. Registered buffers and files skip per-I/O bookkeeping. SQPOLL is a kernel thread that polls the SQ so you can submit without syscalls in the hottest path — at the cost of a pinned core and a bigger attack surface. Most servers should start without SQPOLL.
user: SQE read fd 7 buf A → ring
kernel: DMA / filesystem → CQE res=4096
user: process CQE, recycle bufferBuffer lifetime is yours. Completing into a buffer you already reused is memory corruption. Libraries (liburing) help. Languages without a mature wrapper should not reimplement the rings casually.
When it wins
High IOPS small I/O, lots of sockets, NVMe. Node and Java will not magically use it until the runtime does (Netty, latest JDKs, and some HTTP servers have paths). A Spring app on Tomcat blocking JDBC will not get io_uring's benefits from a sysctl.
File I/O: buffered vs O_DIRECT still matters. io_uring is not a substitute for a missing index.
Security and kernels
io_uring had a series of CVEs; some shops disable it (sysctl / seccomp). Production: current kernel, limited opcodes if you sandbox, no untrusted user access to the uring fd. Privilege: creating a ring is a capability story on some policies.
Read Axboe's PDF for the design and man io_uring for opcodes. Then measure iouring vs epoll on your syscall profile. If read is not in the profile, you needed a different optimization. If it is, and you control the runtime, the rings are the current Linux answer.