This article explains the Linux kernel's Read‑Copy‑Update (RCU) mechanism, detailing its lock‑free read path, update workflow, grace period handling, performance benefits, limitations, real‑world applications, and essential API functions with practical code examples.
The article compares spin_lock, rwlock, seqlock and RCU for protecting a shared kernel resource in a multi‑core embedded system, explains how RCU enables lock‑free reads, discusses the challenges of safe deletion, and outlines the RCU read‑copy‑update workflow.
DGraph, a C++ recommendation engine launched in 2022 for 得物, combines an index layer with a service layer, uses lock‑free RCU structures, a custom mmap‑based D‑Allocator, RoaringBitmap invert indexes, and a multi‑operator fusion scheduler to achieve high‑performance, eventually consistent, scalable recommendations.
Linux kernel synchronization requires protecting shared mutable state from concurrent access using primitives such as spinlocks, mutexes, read‑write locks, or lock‑less techniques like RCU, which copies data and waits for a grace period, each offering distinct performance, latency, and complexity trade‑offs.
The Linux kernel’s page‑cache indexing migrated from the classic radix‑tree structure—using shift‑based multi‑way nodes, RCU‑protected insert, lookup, and delete operations—to the newer XArray API introduced in 4.20, which retains the radix layout while offering automatic resizing, built‑in locking, richer marking, and a cleaner, safer interface.
The article introduces HotRing, a hotspot‑aware in‑memory key‑value store from Alibaba's Tair team, explains the extreme traffic challenges of Alibaba's Double‑11 peak, details the design of ordered‑ring hashing, dynamic hotspot detection, lock‑free concurrency and rehashing, and shows how these innovations achieve up to 2.58× higher throughput than existing KV systems.
This article explains the design principles of Linux kernel RCU, details the implementation of rcu_read_lock, describes how interrupt handlers update shared data safely, and clarifies when old pointers can be reclaimed using call_rcu or synchronize_rcu.
