Linux 7.2.5 catches the Blackwell driver up and hardens ksmbd
Stable kernels 7.2.5 and 6.18.51, released September 11, 2026, fix video output for Blackwell GPUs in nouveau and close a memory leak and a use-after-free in ksmbd. Update first if you expose SMB shares or run a Blackwell card.
September 11, 2026. Greg Kroah-Hartman signed off on two stable Linux kernels: 7.2.5 and 6.18.51. Each carries roughly 550 patches. September 2026. The bulk of the volume sits in the nouveau driver, which finally catches up to Blackwell cards, and in hardening for ksmbd, the kernel’s SMB server. Why it matters: beneath the weekly routine, this cycle shows a stable kernel now maintained with explicit AI assistance — and two bug families that hit real production machines.
Two trees, one cadence
7.2.5 builds on the August 17, 2026 mainline base; 6.18.51 continues the series that launched in December 2025 with its fifty-first incremental patch. Both carry the same Greg Kroah-Hartman signature, stable maintainer, and follow the usual weekly rhythm. Nothing dramatic in the calendar — which is exactly what makes the content instructive.
The 6.18 branch has still not been officially labeled LTS, but fifty-one stable patches in nine months tell a different story: the community already treats it as a long-term line. Enterprise distributions will push 6.18.51 through their usual QA cycles, while rolling distributions will fold 7.2.5 into their next refresh. For an operator, the choice between the two is less about content — nearly identical — than about their distribution’s support policy.
The Blackwell driver finally catches the registers
The largest patch cluster targets nouveau, the free GPU driver. The GB20x chips in the Blackwell generation had changed their register layout, and the driver had not kept up: “head timing” interrupts were still routed to legacy vectors instead of the new low-latency message-based ones. The direct consequence: vblank synchronization and sequence queries broke.
This is not a cosmetic bug. Vblank synchronization is what aligns rendering to the display refresh; without it, compositors show tearing or wait on events that never arrive. The same cycle fixes HDMI vendor infoframes and register stride mismatches on the same hardware. For anyone debugging display output on a Blackwell workstation with the free driver, this is the patch that moves the screen from “unstable” to “usable.”
The detail that captures the nature of the effort sits elsewhere: one of this cycle’s nouveau patches credits Claude for helping trace a NULL pointer dereference in the UVMM code. This is no longer an anecdote — AI is becoming a named contributor in the bug trail.
ksmbd: two flaws that count
ksmbd, the kernel’s built-in SMB server, gets two security fixes that alone justify the update for estates that use it.
The first: ksmbd was not zeroing the padding on pipe compound reads, leaking up to seven bytes of uninitialized heap memory. Seven bytes is small — but it is kernel memory, and an information leak, however tiny, can chip away at ASLR or confirm the presence of a structure. The fix forces that padding to zero before returning it to the client.
The second is more severe: a use-after-free in the oplock break path was fixed. Oplocks (opportunistic locks) are the mechanism by which an SMB server delegates caching to a client to speed up access; breaking an oplock is a routine operation, and a use-after-free on that path opens the door to potential exploitation. Both flaws hit the same component — ksmbd — which is only active on machines that expose SMB shares: precisely the profile of small self-hosted file servers.
The rest: memory, BPF, and DAMON
Behind those two headlines, the cycle ranges wide across memory management. The HugeTLB vmemmap rollback was overwriting head-page metadata during unmap attempts, and xarray node accounting was charged to the wrong memory cgroup during page-cache splits. These are quiet bugs, but of the kind that silently corrupt cgroup billing or page state — exactly the defect that surfaces as an inexplicable OOM-kill three months later.
BPF receives a preemption-disable fix meant to prevent buffer corruption in stack ID lookups. DAMON, the data access monitoring subsystem, gets a round of race condition and infinite loop fixes from SJ Park. Calm maintenance, no public exploit — but it is what keeps servers alive.
AI, from experiment to routine
The most structural fact in this cycle is not a bug. Several commits explicitly credit AI assistance: a fix in the ab8500_fg power supply driver thanks “Assisted-by: Codex:gpt-5.6” for catching a use-after-free, and a nouveau fix thanks Claude. The same pattern already appeared in the June stable release, in the DRM tree.
Read this signal for what it is. This is no longer a maintainer experimenting: AI assistance is becoming a standard credit line in stable patches. The benefit is real — one more use-after-free caught before it reaches a distribution. The risk is real too: an AI-generated or AI-suggested patch goes through the same human review as any other, and the explicit credit exists precisely to keep that chain legible. It is the right practice, generalized.
What this means for distributions
The version number you see does not tell the whole story. An LTS or enterprise distribution — Red Hat, Ubuntu LTS, Debian — almost never jumps from one kernel to the next across stable releases: it backports the useful fixes into its own kernel. In practice, the two ksmbd fixes in this cycle are likely to be backported quickly because they are security-relevant, while the Blackwell catch-up will wait for the distribution’s next major kernel bump.
Rolling distributions — Arch, openSUSE Tumbleweed, Fedora Rawhide derivatives — fold 7.2.5 in almost immediately. On such a distribution, uname -r will show the new version after your next system update; on a fixed-kernel distribution, check the vendor’s security advisory instead, which will reference the ksmbd fixes by commit identifier rather than a version number. It is the distinction that stops you from believing a machine is “up to date” because it carries a recent kernel — or the reverse.
The 6.18 line also deserves a specific note. It predates the 7.x renaming that followed 6.18 in the kernel’s numbering, and it has been receiving steady maintenance since December 2025 without yet carrying the formal LTS label. For operators, that gap between “treated like LTS” and “labeled LTS” matters: it means the branch is stable in practice, but you should not assume the multi-year support window that a formally designated long-term line would guarantee. If you plan to standardize on 6.18, verify your distribution’s stated support horizon rather than relying on the upstream label.
Verdict
If you expose SMB shares through ksmbd, update to 7.2.5 or 6.18.51 without waiting for the next cycle: the memory leak and the use-after-free hit a network-facing component, and they are the only fixes in this batch that warrant urgency. If you run a Blackwell card with the free driver, this cycle repairs your video output — go through your distribution’s update rather than a manual backport. For everything else, follow your usual update cadence: the memory, BPF, and DAMON work is the “integrate cleanly” kind, not the “deploy urgently” kind. And take away the deeper signal: the stable kernel is now maintained, in part, by AI-assisted patches — a chain to watch, not to fear.