FreeBSD 5.0 Ships with SMPng, the Start of Fine-Grained Kernel Locking

FreeBSD 5.0-RELEASE was announced on January 19, 2003 with the foundational architecture of SMPng, FreeBSD’s effort to make the kernel execute more work concurrently on multiprocessor systems. It was a major shipping milestone, but not the day the “Giant” kernel lock disappeared. FreeBSD’s own release material warned that 5.0 contained large amounts of new code, had not yet received production exposure comparable to 4-STABLE, and could regress in stability, performance, or functionality.

Multiprocessor support existed before SMPng

FreeBSD 4.x could run on symmetric multiprocessor machines, so SMPng did not introduce the mere ability to boot a second CPU. The limiting design was a coarse mutex known as Giant. Much kernel execution had to acquire this shared lock, which prevented two processors from entering protected kernel paths simultaneously even when their operations touched unrelated data.

That approach was a practical way to make a historically uniprocessor kernel safe enough for early SMP. It was also a scalability ceiling. Adding processors did not guarantee proportional throughput when system calls, networking, virtual memory, storage, or other kernel-heavy work repeatedly converged on one lock.

SMPng changed the locking model

SMPng began in June 2000, according to the project’s historical SMP page and John Baldwin’s contemporary paper on locking a multithreaded kernel. The work moved FreeBSD toward fine-grained synchronization: protect particular data structures or subsystems with their own locks, define who owns them, and allow independent operations to proceed on different CPUs.

The project involved more than replacing one mutex with many. FreeBSD needed synchronization primitives with documented semantics, interrupt handling that could coexist with threaded execution, changes to scheduling and process structures, lock-order discipline, and tools for finding races and deadlocks. Release notes also credit code obtained from BSD/OS 5.0 as part of the SMP foundation integrated into FreeBSD 5.0.

FreeBSD’s SMP history summarizes the basic architecture delivered in 5.0: new memory allocators and synchronization primitives, interrupt threads, and removal of Giant from scheduling and common inter-process communication paths. Those changes established where later subsystem work could attach; they did not make every driver and kernel service immediately MPSAFE.

Giant still covered substantial code

Fine-grained locking is incremental because each protected object has invariants that must remain valid under concurrency. Removing Giant from a subsystem requires identifying shared state, selecting lock boundaries, handling sleeping and interrupt contexts, avoiding lock-order cycles, and measuring whether added synchronization actually improves the workload. A hurried conversion can trade visible serialization for rare corruption or deadlock.

For that reason, FreeBSD 5.0 deliberately retained Giant around code that had not yet been made safe. Later development removed it from more of virtual memory, VFS, UFS, networking, and drivers. The useful way to read the 5.0 milestone is “the new architecture shipped and broad conversion could proceed,” not “the kernel became lock-free” or “SMP scaling was finished.” Fine-grained locking still uses locks; it scopes them so unrelated work need not serialize on a single global gate.

The project explicitly advised conservative users to wait

The 5.0 announcement called the release the culmination of nearly three years of development, but SMPng was only one of several ambitious changes. FreeBSD 5.0 also shipped UFS2, GEOM, TrustedBSD MAC, Kernel Scheduled Entities, new platform work, and extensive toolchain and userland updates. Attributing the entire development period to one locking objective understates the breadth—and the integration risk—of the release.

The official Early Adopter’s Guide was unusually direct. It said the new technologies had not been widely tested in production, warned of regressions, and recommended that more conservative users continue with the 4.X series. That was not an admission that SMPng had failed. It was release engineering distinguishing an innovation release from the older branch with a longer operational record.

This caution also explains why anecdotes about administrators waiting for later 5.x or 6.x releases should not be presented as if 5.0 itself promised immediate production replacement. The project’s primary documents already define the intended audience: adopters who valued the new architecture and could tolerate investigation, compatibility work, and rapid follow-up changes.

Why fine-grained locking was essential

Processor design was shifting toward systems where performance growth increasingly came from additional cores rather than ever-faster single cores. A kernel whose critical work passed through one mutex would leave a growing portion of that hardware underused. SMPng supplied the synchronization framework and engineering direction needed to reduce that bottleneck subsystem by subsystem.

The work also changed how FreeBSD developers reasoned about kernel code. Lock ownership, sleepability, interrupt interaction, and ordering became explicit design properties. The Architecture Handbook’s locking chapter reflects that lasting discipline: mutexes, shared/exclusive locks, reader locks, condition variables, and atomic operations each solve different synchronization problems and have constraints that code must respect.

The accurate legacy of FreeBSD 5.0

FreeBSD 5.0 made the new SMP architecture available in a release on January 19, 2003. It proved that the project had moved beyond treating Giant as the permanent kernel-wide concurrency model, while its documentation accurately acknowledged the unfinished conversion and the risks of a development-heavy branch.

That combination is the important historical result. SMPng was not a single patch and 5.0 was not a finish line. It was a multi-release restructuring that allowed FreeBSD to keep removing serialization as multiprocessor hardware became ordinary, with the 5.0 release serving as the first public, integrated foundation.

Related:

Sources: