Linux 7.2 Ships: Cache-Aware Scheduling, MGLRU Gains, and GPU Power Management

Linux 7.2 was tagged and released this week, on schedule. This cycle was one of the busiest ever, second only to 6.7, and the maintainers call it the "new normal" — the last three or four cycles have all been heavy on fixes.

The headline features include cache-aware scheduling, improvements to MGLRU, the introduction of sub-schedulers for sched_ext, and automatic creation of multi-size transparent hugepages. For the full picture, LWN's part 1 and part 2 merge-window summaries are worth a read.

Igalia's contributions

Igalia had a solid share of work in this cycle, headlined by the DRM scheduler fair policy, which is present but opt-in due to a last-minute regression report during rc7. The fair policy was supposed to become the default, improving GPU sharing between multiple clients and protecting light interactive workloads from being starved by heavy ones. The fix is already known and early testing looks promising, so expect it to be re-enabled in a future release.

On the power front, Igalia landed runtime power management for the Raspberry Pi 4 and 5 GPUs. Previously, the V3D driver kept the GPU clock enabled from probe until driver unload, meaning an idle GPU still drew power. With runtime PM, the GPU is powered only when it's actually processing work, and the clock can be disabled while idle. The team also fixed two long-standing bugs in the Raspberry Pi 3 GPU driver that had been causing random GPU hangs and system crashes for RetroPie users — the root cause was in how tile memory was handled when the GPU ran out of space mid-frame. Each graphics job now only writes to its own memory area, and reused memory is properly cleared before reuse.

In sched_ext, Igalia improved observability for debugging custom schedulers. When a custom scheduler hits a runtime error (e.g., failing to schedule a task for over 30 seconds), the kernel ejects it and falls back to the default scheduler. The kernel dumps the status of each CPU, but on high-core systems these dumps could be truncated due to buffer size limits. The fix prioritizes the exit CPU — the one that triggered the error — so it's dumped first, and its CPU ID is surfaced directly to BPF schedulers and userspace tools.

They also helped design and land a fix for a 14-year-old futex bug affecting the robust list mechanism, which could cause data corruption in edge cases. And in general bugfixing, they fixed a long-standing race condition in the ueagle-atm driver between kernfs create and remove operations during device probe/disconnect, and improved the correctness of the inline-assembly memcmp() used during x86 boot to prevent subtle bugs from compiler optimization and instruction reordering.

Finally, part of the HDMI 2.1 Fixed Rate Link (FRL) support in this release traces back to Rodrigo Siqueira's work at AMD, which sat outside the main tree for years before being incorporated. It gives the amdgpu driver the ability to communicate with HDMI 2.1 monitors, primarily the FRL implementation.