Linux 7.3 Ships With 18,765 Commits, a Filesystem That Fails Everything and Per-Cgroup BPF Schedulers
Software / news
Linux 7.3 Ships With 18,765 Commits, a Filesystem That Fails Everything and Per-Cgroup BPF Schedulers
Linus Torvalds released the kernel on Oct. 18 after a final week he called uneventful, with the biggest figures coming from storage and memory fixes.

Linus Torvalds released Linux 7.3 on Sunday, Oct. 18, with 18,765 non-merge commits from 3,021 developers, 807 of whom were first-time kernel contributors, according to LWN's release note. The 7.4 merge window opened the next day.
Torvalds said "nothing very special seems to have happened this last week," so the final release arrived on schedule. LWN describes the commit count as a record.
The new interfaces: failfs, fchroot and BPF-backed binfmt_misc
The KernelNewbies changelog lists failfs as a pseudo-filesystem that is the counterpart of nullfs. Its page says "nothing is supported here," so every operation that reaches it fails with EOPNOTSUPP, the kernel's "operation not supported" error. A process can use it to drop its filesystem state.
The new fchroot(2) system call changes a process's root directory through a file descriptor rather than a path name. KernelNewbies says that avoids a time-of-check-to-time-of-use race, where the path changes between the check and the use.
binfmt_misc, the mechanism that picks an interpreter for a binary, can now back its handlers with BPF programs, so matching and interpreter selection are computed on each exec. KernelNewbies cites Nix and relocatable binaries as the motivation.

Scheduler: each cgroup can run its own BPF scheduler
sched_ext, the framework for writing CPU schedulers as BPF programs, gains cgroup sub-schedulers. Each cgroup can run its own scheduler, and a parent can delegate per-CPU controls such as enqueueing, preemption and CPU frequency to a child, per KernelNewbies. LWN lists the completion of the sched_ext subscheduler feature among the release's headline items.
Separately, Peter Zijlstra's patches move cgroup scheduling onto a single runqueue. The new cgroup_mode setting takes up, max, concur or tasks, and the default, concur, is the most precise and the most expensive. It's FOSS reported frame-rate gains in its testing on an old Sandy Bridge machine with a Radeon RX 580.
The performance numbers, and where they disagree
It's FOSS wrote its figures from the merge window in August, before any release candidate, so they describe what was merged rather than a measured 7.3. KernelNewbies repeats most of them.
| Change | Reported figure | Source |
|---|---|---|
| ext4 small direct I/O | 1.92 million to 2.19 million IOPS | It's FOSS, KernelNewbies |
| Btrfs direct I/O | about 50% to about 95% of theoretical maximum | It's FOSS, KernelNewbies |
| zsmalloc zs_free() | up to 1.83x | KernelNewbies; It's FOSS reports up to 1.83x on a Raspberry Pi 4B |
| SMP cross-CPU calls, slowest requests (ByteDance, DPDK fleet) | about 17 ms to about 1.5 ms | It's FOSS |
The two sources disagree on one number. It's FOSS says a KSM reverse-mapping lock stall fixed by ZTE engineers fell from about 700 ms to under 2 ms in the worst case. KernelNewbies puts the starting point at over 500 ms and the end at under 2 ms.
- Before1.92 million IOPS
- After2.19 million IOPS
Source: It's FOSS and KernelNewbies, accessed 2026-10-19
Neither source says which hardware produced the ext4 figure beyond "fast NVMe," and neither says the figure was reproduced on the released 7.3 kernel.
Smaller changes that will reach desktops and file servers
The NTFS driver gains read support for Windows System Compression, with XPRESS and LZX decompressors, and ntfs3 gains basic alternate data stream support. The in-kernel SMB server, ksmbd, adds opt-in Apple Time Machine support, per-share SMB3 encryption and SMB Direct RDMA encryption. userfaultfd gains read-write protection for tracking a virtual machine's working set, with the cover letter linked from KernelNewbies pointing to Kirill Shutemov.
It's FOSS also reports that RISC-V adds eleven newer ISA extensions, and that Torvalds spent 18 kernel boots and 24 debug patches on an Intel Xe driver bug, with AI help.
The three sources describe no security fix as a headline item, which is a contrast with Microsoft's October Patch Tuesday, where one flaw was exploited. Embedded gateways such as those in the Broadcom Wi-Fi report run their own kernel versions, so mainline releases do not reach them directly.
The merge window for Linux 7.4 opened Oct. 19, the day after the release, and Torvalds's first release candidate is the next fixed date.
Sources
More in Software
- 01KiviDB Claims 21.2 Million GETs a Second, but Its Own Table Has Dragonfly Ahead Without PipeliningThe Rust, AGPL-licensed Redis-compatible store has 208 GitHub stars, no named authors and a benchmark run by its own company.
- 02Google Research's Harness2 Has 10,700 Stars and Three Commits, and Publishes No Absolute ScoresThe recursive agent-harness framework reports gains of up to 11.8 percentage points, but its README and arXiv page give no baselines, no cost per run and 16 open bugs.
- 03Deno Gets 12 Months of Releases and Deploy Gets Six After Cloudflare Deal, and Two Forks AppearFeature pull requests are being closed unmerged, and the community fork called Ende wants to drop the Node compatibility that Ryan Dahl now calls a constraint.
- 04VoiceStudio Has 75,300 Stars and an AGPL Licence, but Its OmniVoice Model Is CC-BY-NCThe README tells users that a paid app licence does not cover the model weights, and a CLA thread asks contributors for commercial relicensing rights.