Linux 7.3 is the biggest kernel ever and delivers faster storage and smarter scheduling

Wow. That's one big Linux kernel release! Released on October 18, the 7.3 kernel set development records with 18,765 non-merge commits from 3,021 developers, including 807 first-time contributors.

Why was it so big? Well, it's not because it came with a horde of new features. No, two different forces drove its growth. First, there was an enormous number of AMD graphics hardware-definition headers. Then, and this shouldn't surprise anyone, it came filled with AI-related fixes and cleanup.

As for the first, Torvalds has explained earlier on the Linux Kernel Mailing List (LKML), "This is yet another one of those releases with a big AMD GPU register dump,” he wrote. “This time it’s the AMD DCN6 register headers. They’re huge. As usual.”

Part of the reason is that these headers encode hardware definitions as C preprocessor definitions. Combined with the accompanying code for the new graphics generation, they accounted for about one-third of the entire 7.3-rc1 patch—not one-third of the kernel itself.

Torvalds added: “But if you just mask out that big chunk, things look pretty normal, and this merge window didn’t feel that different from most of them despite being large."

As for the AI part, the networking subsystem alone, maintainer Jakub Kicinski reported, had merged 632 patches into the fixes stream and 648 into the next-release development stream. He estimated that one-third to one-half of the latter “seem like AI-driven low-priority fixes, cleanups and clarifications.”

“We are completely overwhelmed, of course,” Kicinski added. He said the maintainers were themselves using multiple frontier models to assist with patch review, while cautioning that “the LLMs can only do so much.”

Still, when all was said and done, 7.3 arrived on schedule. Linus Torvalds wrote, "It's Sunday afternoon, and this week I somehow kept my wits together long enough to keep track of the day of the week, so the release isn't randomly delayed. Will wonders never cease?" Given he'd just spent the week before at the Linux Plumbers Conference in Prague, color me impressed. I was there too, and I'm still not sure what day it is.

Beyond those headline numbers are improvements that matter to Linux desktops, storage servers, virtual-machine (VM) hosts, and container deployments.

Faster storage, with workload-specific gains

Some of Linux 7.3’s most immediately useful work is in its storage stack. Ext4, the most popular general-use file system, gains better parallel direct I/O writes and faster space allocation through fallocate(). The developers also fixed a buffered-write performance regression. A simplified path for small direct-I/O operations increased ext4 performance from 1.92 million to 2.19 million I/O operations per second in one reported sample workload. That is a specific benchmark result, not a promise that every ext4 installation will become proportionally faster. Still, that's a decent kick-in-the-pants for file reads and writes.

Btrfs also got some love with several improvements. For example, Btrfs has had r direct-I/O performance restored. Its developers also removed an unnecessary delay, improving latency and throughput for logging workloads on devices operating in non-SSD mode. Another optimization skips redundant hole detection during full fsync() operations for suitable files. Reported gains of roughly fivefold apply to those particular test cases, not Btrfs performance generally.

Filesystem in Userspace (FUSE) also saw performance improvements. Servers using io_uring can provide buffer pools, reducing memory requirements and giving the kernel more flexibility in assigning request buffers. A new zero-copy option lets a suitably privileged server access client pages or page-cache folios without an intermediate data copy.

For databases and other applications accessing raw block devices, Linux 7.3 extends RWF_DONTCACHE support to those devices. Applications can choose whether to retain data in the page cache per request, rather than treating direct I/O as an all-or-nothing policy.

More flexible CPU scheduling

Linux 7.3 completes support for subschedulers within sched_ext, the framework for implementing task schedulers with BPF. Individual control groups, or cgroups, can have their own BPF-based scheduling policies, allowing different groups of processes to make different decisions about when and where their tasks run.

That is distinct from changes to the kernel’s fair scheduler. Linux 7.3 revises how CPU time is distributed among cgroups and flattens task selection into a single runqueue, addressing problems involving latency, overhead, and scheduling behavior. It also improves scheduling latency for tasks with shorter time slices.

For organizations mixing different workloads on the same systems, these changes will improve existing scheduling behavior and expand the scope for specialized scheduling policies. However, Linux hasn't replaced its default scheduler, Completely Fair Scheduler (CFS), with BPF.

Better game frames per second when graphics memory runs short

Desktop users and gamers with memory-constrained graphics cards may benefit from changes to VRAM management, particularly in AMD’s graphics stack. That would be me for one. Previously, an application arriving after another program had consumed much of the available VRAM could end up allocating from system memory, even when its configured memory protection should have given it a stronger claim to graphics memory. Linux 7.3 allows the kernel to evict unprotected buffers to make room for protected allocations, using the device-memory cgroup controller.

The change makes those protections more effective under contention. That doesn't guarantee higher frame rates in every game. Darn it! But it can help by changing how the kernel allocates scarce graphics memory when applications are competing for fast memory.

The new kernel also comes with further AMD Zen 6 platform preparations, initial support for Apple’s M3 Pro, Max, and Ultra, and Rust support extending to PowerPC. Don't get too excited. The Apple additions are still very much works in progress.

A filesystem designed to refuse access

One of the release’s more unusual additions is failfs, a pseudofilesystem whose purpose is to make operations fail. Yes, you read that right: Fail. Paired with the new fchroot() system call, it provides a building block for restricting how a process resolves filesystem paths.

Sandboxing is the point of failfs. It enables a program to give up ordinary filesystem access while retaining access through explicitly held file descriptors. Its deliberate failures enforce a boundary. It's not a way to simulate broken storage.

A process can move its root into failfs, causing absolute-path lookups and absolute symbolic links to fail. Moving its working directory there as well leaves it dependent on explicit directory file descriptors for path resolution. Instead of relying on a conventional, visible directory tree, an application can anchor its remaining filesystem operations to directory handles it already holds.

The new fchroot() call is also useful independently: it changes a process’s root using a file descriptor, avoiding the need to resolve a pathname again after the caller has already obtained a handle to the target. These are new primitives for sandbox developers, not an automatic sandbox applied to every Linux application.

VM and file server Improvements

VM hosts gain userfaultfd support for tracking a guest’s memory working set. A virtual-machine manager can use this to tell the difference between actively used pages and cold ones. This helps it decide what memory to move to slower tiers or storage.

The in-kernel Server Message Block (SMB) server, ksmbd, adds opt-in support for Apple Time Machine extensions. This allows appropriately configured shares to serve as macOS backup destinations. That's been a nuisance for me for quite some time, so I welcome it. Other SMB improvements include per-share SMB3 encryption and encryption for SMB Direct over RDMA. Keep in mind that you'll only see these improvements if you use the kernel SMB server; they're not in the current version of the popular Samba SMB server.

So, 7.3 arrives with several notable improvements inside its new Hulk-sized kernel. Looking ahead. 7.4 bodes to be even larger as AI continues to find and help fix new bugs. With AI now used to write new code, even Torvalds uses Claude to write code, and Christian F.K. Schaller, Red Hat's director of software engineering, found a way to make it much easier to reverse Windows hardware drivers to Linux. I foresee ever larger kernels for at least the next year or two.

For your amusement

Irony or coincidence? You decide.

Leave Granola and get up to 12 months free of Wispr Flow Notetaker + Dictation

If you have paid time left on an individual Granola plan, we'll match it with a Wispr Flow subscription that includes Notetaker and dictation, and add bonus time, up to 12 months total. Sign in or create a Wispr account and submit proof of your plan to check eligibility.

Try Notetaker Free

Update your email preferences or unsubscribe here

© 2026 Vaughan-Nichols & Associates

134 Flint St.
Asheville, North Carolina 28801, United States