Debian 11107 Ubuntu 7263 Published by

Kernel maintainer Steven Barrett released Liquorix Linux Kernel 7.3-5, adding more Project-C sleep-credit fixes to the custom PDS scheduler path so wake-up tasks get priority back instead of starved. Liquorix swaps mainline's default EEVDF scheduler for either the PDS or BMQ engine and layers low-latency knobs like a 1000 Hz tick rate, preemptible RCU, Kyber or BFQ disk I/O, and TCP BBR3 networking onto a hardened 7.3.4 base. 



Liquorix 7.3-5 brings back the scheduler tinkering it tried to put to rest

Enthusiast kernel maintainer Steven Barrett has released a new version of the Liquorix Linux Kernel with various fixes to Project-C.

Liquorix Linux Kernel 7.3-5 is now available. The full changelog reads: "apply more fixes / optimizations to Project-C," followed by two sched/alt entries.

Both land in the custom PDS code path. Both revolve around Project-C's "sleep credit" accounting: when a task sleeps, the scheduler tracks how long and hands back priority credit on wake-up, so frequently-woken interactive tasks don't get starved. It's a subtle move. In practice it means the kernel tries harder not to preempt something about to wake.

Liquorix

What Project-C is actually about

Liquorix swaps mainline's default CFS (now branded EEVDF) for one of two custom schedulers, and that single decision is most of the reason the kernel feels different.

The default is PDS, tracing back to Con Kolivas's MuQSS work. It uses a skiplist structure that prioritizes interactive tasks aggressively. There's also BMQ, an older alternative with its own CK-patchset lineage. On top of either, Liquorix stacks a battery of low-latency knobs.

The scheduler tick rate runs at 1000 Hz versus the ~250–300 Hz you typically get. Preemption is fully preemptible with hard preemption, RCU is preemptible tree-based, and disk I/O defaults to Kyber or BFQ. Networking uses TCP BBR3 instead of CUBIC.

Project-C is the running theme across releases since mid-2026. It's described in changelogs as an "ongoing overhaul" and "tightened optimizations." The goal is shaving microseconds off wake latency by keeping related tasks close together.

A producer thread and its consumer should land on SMT sibling cores that share L2 cache. Separating them forces cold-cache reloads that spike latency. The work targets cache-line alignment, SMT-sibling awareness, lock contention, and wake placement.

The sleep credit system is the centerpiece. It decays in step with time-slice consumption, caps at the current nice level, and as of 7.2-16 can be toggled from the kernel command line rather than forcing a recompile.

July and August ballooned with hundreds of sched/alt tweaks. September saw patches reverted as regressions. Late October brought rapid-fire builds. The 7.3-5 build is simply the latest chapter.

Riding 7.3.4

Liquorix rides whatever stable upstream release is current, so the value of a given build depends heavily on what landed in mainline.

Linux 7.3 shipped October 18, 2026, from Linus Torvalds after a two-week merge window and seven release candidates. Several changes land squarely in Liquorix's wheelhouse. There's new VRAM management for low-end GPUs, driven by Valve's Natalie Vock, letting a GPU evict unprotected buffers when short on video memory so overcommitted games keep running instead of downgrading. sched_ext finished out, letting individual cgroups run their own BPF-based schedulers. Btrfs picked up roughly 5× faster fsync in some scenarios. Older Radeon GPUs can now recover from hangs by resetting just the GFX block.

Liquorix merged 7.3.3 for its baseline at 7.3-1, then moved to the security-focused 7.3.4 stable release for 7.3-2 onward. That 7.3.4 fold-in includes two Xen virtual-networking security advisories, an AMD Zen 5 TSA_FILL CPU mitigation, and a cluster of AppArmor hardening and filesystem correctness fixes.

So the resumed Project-C patches in 7.3-5 sit directly on top of a hardened networking and virtualization stack. Which is arguably the point.

The trade-off, laid bare

Barrett and the community alike are explicit about what Liquorix gives up.

The wins are snappier pointer and input response, smoother and more consistent frame times, and lower tail latency, fewer microsecond stalls that otherwise become audio dropouts, gaming hitches, or video-call freezes. The documented cost is higher power use and somewhat lower raw throughput. It drains laptop batteries faster. It's a poor fit for servers, which want throughput and efficiency, and for heavily multithreaded batch workloads.

The "Zen Interactive Tuning" profile makes these trade-offs deliberately. A config diff shows the block scheduler moving from mq-deadline to kyber or bfq, watermark boost factor dropping from 1.5 to 0, the scheduling timeslice halved from 4ms to 2ms, split-lock detection switched off, and CPU-frequency thresholds trimmed so cores ramp faster.

What you should know if you install it

This is an AMD64-only kernel, and that's a deliberate constraint, not an oversight.

Secure Boot isn't supported out of the box, so you'll likely disable it in BIOS. The good news: your stock distro kernel stays installed as a GRUB fallback, so a bad boot can usually be recovered by selecting the previous entry. Barrett repeatedly advises backing up and verifying hardware before switching boot entries.

If you're on Fedora, note that the Liquorix kernel becomes and stays the default boot entry over stock updates unless you set UPDATEDEFAULT=no in /etc/sysconfig/kernel. And for real-time workloads, this isn't your kernel — a proper RT kernel or Zen's real-time options will serve you better.

If you want to try it, the install one-liner across supported distros is:

curl -s 'https://liquorix.net/install-liquorix.sh' | sudo bash

Binary packages ship for Debian (Stable, Testing, Unstable), Ubuntu (now officially hosted by liquorix.net), Fedora, and Arch. Bu

Head here to the GitHub release announcement.