Linux 7.3-rc7 arrived six hours late because Linus Torvalds lost track of which day it was at the maintainers' summit — and half of the patch is test code for mount fixes. Read that as the strongest possible signal: the final release is one quiet week away.

Six hours late, by the boss's own admission

The final release candidate of Linux 7.3 went out on Sunday, October 12, 2026 — confirmable on kernel.org's front page, where mainline reads 7.3-rc7. It landed about six hours behind the usual Sunday rhythm, and Torvalds explained why in his announcement with characteristic bluntness: travel for the kernel maintainer summit and Linux Plumbers Conference left him "mentally off by a whole day," convinced the entire time it was Saturday until dinner, when it struck him that he owed the world a release candidate. He got it out on the right day. "If barely."

This is the same maintainer who once called himself "a grade A nincompoop" for a botched system upgrade mid-merge-window. The self-deprecation is a feature, not a bug: it keeps the release process honest in public, and it keeps everyone else's expectations calibrated.

The strangest diffstat of the cycle

What makes rc7 interesting isn't the delay — it's the shape of the patch. Torvalds called the diffstat "slightly strange": roughly half of rc7's code lives in selftests. The driver is a batch of mount-related bug fixes from VFS maintainer Christian Brauner. The fixes themselves were small; the tests proving them correct were voluminous.

Strip out tools/testing/ and the picture snaps back to normal: about half drivers (DRM graphics and networking), the rest scattered across the tree, with the VFS layer slightly prominent thanks to those mount fixes, plus the usual core networking and architecture fixes.

For an operator, a test-heavy late release candidate is the best news of the cycle. An rc that arrives carrying its own proof is one you can trust to land on time. Torvalds said as much: nothing in rc7 looks like a reason to delay, so the final 7.3 stays set for next weekend — October 18.

What else came along

Beyond the mount work, rc7 carries an AMD TLBI erratum workaround, two newly supported game controllers in the XPad driver, and the continuing background hum of small fixes surfaced by AI-assisted code review — the "new AI normal" this cycle is known for. None of it is scary; all of it is late-cycle tightening doing its job.

Your one-week checklist

The final release is seven days out. If you run infrastructure, this is the calm week — use it.

1. Test your fleet's edge cases in a VM, not on hardware. rc7 is a release candidate: stable enough to validate, not to deploy. Boot it in a VM, build your DKMS modules and out-of-tree drivers against it, and confirm your monitoring still reports sanely. One afternoon now beats one incident later.

2. Read your distro's drumbeat. Rolling distributions typically package a new stable kernel within days of release; point-release distros backport or wait for the next scheduled rebase. Know which camp yours is in so 7.3's arrival doesn't ambush your change calendar.

3. If you don't test candidates, take the boring option instead. The stable branches didn't sit still either: 7.2.10 and longterm 6.18.56 landed on October 11 with security and correctness fixes. If your policy is "stable only," that's your upgrade — no need to watch the rc at all.

4. Skim the mount fixes if you juggle namespaces. Containers live and die by mount semantics. Brauner's fixes plus their selftests are the part of rc7 most likely to matter to anyone running heavy container or sandbox workloads. The shortlog is your friend.

The 7.3 cycle will be remembered for the AI-assisted fix flood, and for Linus misplacing a day at a conference. But rc7's real legacy is quieter: a release candidate that proved itself in tests before asking anyone to trust it. That's the whole game.

Sources: kernel.org (7.3-rc7 on the official front page, released 2026-10-12) · Phoronix — "Linux 7.3-rc7 Released With Linux 7.3 Hopefully Out Next Weekend" · Neowin — "Linux 7.3-rc7 patch packs a slightly strange diffstat with half its code in selftests"