What is io_uring zero-copy networking (IORING_OP_SEND_ZC) in the modern Linux kernel?

Introduced in Linux kernel 6.0 by Jens Axboe and Pavel Begunkov, IORING_OP_SEND_ZC (and later IORING_OP_SENDMSG_ZC) brings true zero-copy network transmission to the asynchronous io_uring subsystem. In traditional POSIX networking (send(), write(), sendto()), the kernel must synchronously copy data bytes from userspace memory into kernel socket network sk_buffs. For 40GbE, 100GbE, and 200GbE network interfaces, copying data across the memory bus saturates CPU cache lines and consumes upwards of 60% of total CPU cycles. With io_uring zero-copy, userspace memory pages are pinned and directly mapped for Network Interface Card (NIC) Direct Memory Access (DMA), transmitting packets straight from user buffers without a single memory copy.

Why does IORING_OP_SEND_ZC generate two Completion Queue Entries (CQEs)?

In traditional synchronous send(), once the function returns, the user is free to overwrite the buffer because the kernel has already made its private copy. In zero-copy transmission, the NIC hardware DMA engine reads directly from user memory asynchronously. To prevent userspace from modifying the buffer before the NIC has physically completed reading the packet, io_uring issues two CQEs: (1) Initial CQE with the IORING_CQE_F_MORE flag set: Indicates that the kernel has accepted and queued the network transmission request; (2) Second Notification CQE (without F_MORE): Minted only after the network driver receives hardware transmission completion from the NIC, signaling to userspace that the buffer is now safely reusable.

How do Registered Buffers and Registered File Descriptors optimize io_uring throughput?

Under high packet rates (millions of packets per second), repeated page table lookups and file descriptor table locking introduce significant CPU overhead: (1) Registered Buffers (io_uring_register_buffers): Userspace pre-registers an array of memory buffers with the kernel. The kernel pins the virtual memory pages and maps them into kernel space once at startup. Subsequent I/O operations reference buffer indexes (IORING_OP_WRITE_FIXED / IORING_OP_SEND_ZC) rather than raw pointers, eliminating page locking overhead; (2) Registered Files (IORING_REGISTER_FILES): Pre-registers socket file descriptors into a fixed array, allowing system calls to bypass the process file descriptor table spinlock and reference sockets via direct array offsets.

When should high-performance network applications choose io_uring SEND_ZC over AF_XDP or DPDK?

AF_XDP and DPDK provide raw kernel bypass by taking complete control of the network device queue, achieving maximum theoretical line rate (tens of millions of PPS). However, DPDK bypasses the Linux TCP/IP stack entirely, requiring complex userspace TCP implementations, and AF_XDP requires specialized raw packet framing. In contrast, io_uring SEND_ZC works natively with standard Linux TCP/IP sockets (supporting TLS via kTLS, standard routing tables, iptables/nftables, and socket options). It delivers 80% to 90% of kernel-bypass speed while retaining 100% of standard Linux operating system network stack compatibility.

Linux io_uring Zero-Copy Network I/O Studio: Kernel Bypass & Registered Buffers

Verified Specification | Standardized Formulas | Instant Precision

Secure & Private (Zero Data Retention) Free Access • No Sign-Up

Linux io_uring Zero-Copy Network I/O Studio

Benchmark 100GbE line-rate network transmission: compare traditional POSIX send() memory copies against io_uring IORING_OP_SEND_ZC hardware DMA pipelines and registered fixed buffers.

io_uring Zero-Copy Active Registered Fixed Buffers

94.8 Gbps

Network Line-Rate Throughput

7,850,000 PPS

Packets Per Second (Transmission)

0.0 GB / s (Zero-Copy)

Memory Bus Copy Bandwidth

112 Cycles

CPU Cost Per Transmitted Packet

1. Select Network Transmission Architecture

Packet Buffer Payload Size

Registered Buffer Pool Strategy

Submission Queue Polling (SQPOLL)

2. io_uring Submission & Completion Rings

Submission Queue (SQ Ring) SQE Depth: 128 / 512

Opcode: IORING_OP_SEND_ZC • Flags: IOSQE_FIXED_FILE • BufIdx: #4

Kernel Async Worker / SQPOLL Thread CPU Core #3 Pinning

Direct DMA mapping to Intel E810 / Mellanox ConnectX-6 NIC descriptor queue.

Completion Queue (CQ Ring) - Dual CQE CQE_F_MORE Active

CQE #1: Accepted & queued. CQE #2: Hardware DMA Tx complete & buffer freed.

3. Memory Bus Data Path

Zero-Copy Direct DMA

Linux io_uring Zero-Copy Network I/O Studio

1. Select Network Transmission Architecture

2. io_uring Submission & Completion Rings

3. Memory Bus Data Path

4. Production liburing C Implementation

🔍 Companion Workbenches & Verification Tools

Linux Kernel Slab Allocator & Free List Hardening Studio | Heap Exploitation Mitigations

Linux io_uring Zero-Copy Network Transmit Studio | IORING_OP_SEND_ZC & splice()

Database Sharding, Distributed Transactions & 2PC/Saga Pattern Studio

Linux Permissions & Chmod Calculator

Why Smart Users Are Switching to This Free Toolkit

The Browser Feature Most People Never Use

Stop Paying for Software You Can Get Free

File Conversion Tricks That Save Hours Every Week

This Free Tool Replaced 5 Paid Apps for Our Team

What Your Browser Can Do Without Any Extensions