Linux Kernel 7.2: Cache Scheduling and NOVA Rust Driver

Linux 7.2 is two to three weeks from stable. RC6 landed August 2 as the largest release candidate by commit count in years — Linus Torvalds called it “huge” and pressed forward anyway. Stable arrives August 16 or 23 depending on whether one more release candidate is needed. If you run databases, ship containers, or touch NVIDIA hardware, here is what is actually changing.

The Performance Numbers First

Cache-aware scheduling is the headline feature. The kernel now co-locates related tasks within the same Last Level Cache domain rather than scattering them across CPU cores. On AMD Zen 5 systems — and any CPU with a non-uniform LLC topology — this shows real gains for PostgreSQL, Valkey, and network-heavy workloads. Nothing synthetic; these are production-workload numbers.

MGLRU improvements compound this. MongoDB throughput is up 30 to 100 percent in benchmarks. Container shutdown time collapses from 92–138 milliseconds to 5–8 milliseconds — roughly an 18x improvement in the best case. Reading /proc/filesystems in a multi-process environment gets 444 percent faster, which matters if you use libselinux or anything that polls /proc frequently.

The pipe implementation also got an overhaul: 6–28 percent throughput improvement under normal conditions, up to 48 percent under heavy memory pressure. These numbers add up on busy servers.

strncpy() Is Gone. Finally.

After six years and over 360 patches, strncpy() has been completely eliminated from the Linux kernel. The replacement — strscpy() and strscpy_pad() — handles null-termination correctly and does not silently truncate in the ways that made strncpy() a reliable source of buffer overflow vulnerabilities.

This is worth pausing on. Six years. 360-plus patches across dozens of subsystems, maintainers, and driver trees. No single person did this; it was coordinated persistence across the entire kernel community, with AI-assisted tools helping surface the remaining occurrences near the end. The result is a kernel that eliminated an entire category of bug at the source level. That is how you pay off technical debt at scale — not in a weekend sprint, but as a long-running parallel effort that eventually closes.

NVIDIA’s NOVA Driver Is in Rust, and It Is in Mainline

The NOVA GPU driver is the most significant Rust development in the kernel this cycle. It is an open-source NVIDIA driver written entirely in Rust — a two-tier architecture with nova-core handling hardware initialization and drm/nova managing the DRM interface. It lives in drivers/gpu/nova-core/ and drivers/gpu/drm/nova/ in the mainline tree.

Linux 7.2 adds GPUVM immediate mode abstraction and Higher-Ranked Lifetime Types for Rust device drivers. Target hardware: Hopper (GeForce RTX 30xx) and Blackwell (RTX 40xx/50xx), with full coverage of everything from the RTX 20xx Turing generation forward. The driver is early stage — do not expect it to replace the proprietary driver for production workloads yet — but it is architecturally serious and actively developed.

On the AMD side, HDMI 2.1 FRL support lands in the AMDGPU driver, enabling 4K at 120Hz and higher refresh rates on compatible Radeon hardware. It ships disabled by default pending compliance certification, but the code is in mainline.

The AI Patch Problem Torvalds Will Not Quite Name

Here is the uncomfortable story behind RC6’s size. Developer Tony Luck noted that approximately 29 percent of recent code changes carry “Fixes:” tags — a sign that something upstream introduced a problem that needed correcting. Torvalds acknowledged AI bots are contributing to this pattern, calling it the “new normal” while describing the volume as “unnerving.”

The dynamic is this: LLM-assisted tools are finding real bugs in the kernel faster than humans alone could. That is useful. The problem is that the patches those tools generate are often low quality — correct enough to pass automated checks, but requiring significant human review time to verify, clean up, and merge safely. The result is a higher fix throughput that comes at the cost of maintainer bandwidth.

Kernel maintainers are not paid by the patch. Review time is a finite resource. When 29 percent of commits are fixing something recently introduced, that signals either a code quality regression or a review pipeline under strain — and neither is good. Torvalds is proceeding, but the concern is real and the trajectory is worth watching.

Everything Else That Matters

• USB4STREAM lands, giving any machine with a USB4 or Thunderbolt port a /dev/tbstreamX device that accepts standard read()/write() calls. Machine-to-machine data transfer at 40–80 Gbps — no network stack, no NFS configuration required.
• AF_ALG is deprecated. The cryptographic socket interface carries a massive attack surface and is being wound down in 7.2 with zero-copy and hardware offloading removed. If your code calls AF_ALG directly, migrate to OpenSSL or libgcrypt now. The programs still using it — iwd, libkcapi, cryptosetup — are already being migrated.
• openat2 gains the OPENAT2_REGULAR flag. Security-sensitive programs can now guarantee they only open regular files, blocking attacks that substitute device nodes or FIFOs in place of expected files.
• Apple M3 gets early mainline boot support — serial console only, not practical for end users yet, but M3 Pro/Max/Ultra work is targeting 7.3.
• AMD ISP4 driver finally enables webcams on HP ZBook Ultra G1a and similar Ryzen laptops. Intel TDX gets runtime security module updates without reboots.

What to Do Before Stable Ships

If you use AF_ALG directly, start the migration now. If you run MongoDB or PostgreSQL on AMD Zen 5, plan a kernel upgrade for late August — the performance case is strong. If you build drivers or security tooling, the openat2 OPENAT2_REGULAR flag and the full Linux 7.2 changelog are worth a read. The stable release is two to three weeks out; RC6 is available for testing today.