Linux kernel 7.3 is now in its active merge window, bringing massive performance optimizations to memory management and marking sched_ext as feature complete. The release adds significant hardware support for AMD's UALink infrastructure and CXL Type-2 accelerators, while also providing initial mainline boot support for Apple M3 Pro/Max/Ultra SoCs. Rust-in-kernel officially gains PowerPC support, and the NTFS3 filesystem receives a major security hardening alongside Alternative Data Streams. As a likely Long Term Support candidate, this cycle also highlights the community's shifting development practices, including Linus Torvalds using Gemini AI to debug a one-character Intel Xe GPU bug.

Linux Kernel 7.3 Merge Window: Performance Boost, AMD UALink, and Apple M3 Support

Linux kernel 7.3 is officially in its active merge window as of mid-August 2026, right on the heels of the 7.2 release on August 17th. This cycle is shaping up to be a massive one for AMD hardware, memory management tweaks, and the maturation of sched_ext. It also brings you a bit of drama involving Linus Torvalds, a one-character bug fix, and a lot of help from Google's Gemini AI.

The 7.3 cycle is likely this year's Long Term Support candidate, which means whatever lands here will stick around for a while. Andrew Morton shipped a particularly heavy MM pull request with 1,250 patches added to the memory management subsystem. That's a notable jump from the 920 patches in the prior cycle. You're going to see performance wins for memory-constrained workloads and some serious work on the scheduling front.

Memory and Scheduling Wins

The locking contention in zsmalloc has been tackled by Wenchao Hao of Xiaomi. The result is up to a 1.83x speedup on a Raspberry Pi 4B under concurrent unmapping workloads. A 20-core Intel system saw a 1.4x improvement as well. This matters for Android devices and servers running heavy Zswap or zRAM usage. Keep in mind that memory pressure is only going to get more common as workloads get heavier.

Meanwhile, ZTE engineers found a severe performance bug in rmap_walk_ksm(). Applications could freeze for hundreds of milliseconds under memory pressure. The fix drops worst-case lock hold times from 705ms down to a 1.44ms average. If you run JVM runtimes, Go runtimes, or database engines like MySQL and PostgreSQL that create tens of thousands of VMAs sharing a single anon_vma, this is the fix you've been waiting for.

The sched_ext framework is now feature complete in 7.3. The headline here is hierarchical sub-scheduling, which lets a root BPF scheduler hand a cgroup subtree over to a nested sub-scheduler with revocable CPU grants. Sub-schedulers now own all scheduling decisions on those CPUs. Arena objects also cross the kernel/BPF boundary as typed pointer arguments via the new arena argument support in BPF tree. Rescue execution for starved tasks now runs on a small bandwidth budget instead of waiting for watchdog ejection.

The x86 memory management code landed patches specifically targeting latency-sensitive workloads. Gaming and hybrid CPUs get complementary improvements alongside the scheduler changes. If you're running latency-sensitive applications on x86 hardware, you'll notice the difference.

Hardware Enablement and AMD Infrastructure

Apple M3 Pro, M3 Max, and M3 Ultra SoCs finally get initial Device Tree support for the mainline kernel. This lays the groundwork for Asahi Linux, though GPU acceleration and many other features remain unsupported by both mainline and the downstream project. However, at the same time, this is a critical foundation for future development. Apple M1 Pro/Max/Ultra SoCs get a small win too, with roughly 1 watt of power savings during s2idle sleep states thanks to improved low-power state handling.

AMD gets a big push with the UALink infrastructure. A massive 95-patch series adds UALink support for AMDGPU, which is an open standard for connecting up to 1,024 accelerators in a single domain. Connected GPUs can access remote memory directly over UALink without copying, unlike RDMA. The memory uses Network Physical Address space with dedicated page tables and remote TLB shootdowns for eviction notifications. Applications use opaque handles to represent this remote memory.

The first in-tree CXL Type-2 accelerator driver landed for the AMD Solarflare SFC9100-family network interfaces. This was enabled via a new _SFC_CXL_ Kconfig option, building on basic Type-2 support added in 7.2. Not cheap, but it opens the door for in-place memory expansion for high-performance computing.

Rust in the kernel gets a new architecture. PowerPC is now supported for both PPC32BE and PPC64LE, joining ARM32, ARM64, LoongArch, RISC-V, UM, and x86_64. The code was initially marked experimental but has been updated to maintained status following adjustment patches.

NTFS3 gets a security overhaul. A 21-commit, 922-line update fixes slab-out-of-bounds writes, out-of-bounds reads of INDEX_ROOT during reparse and objid init, information leaks in ntfs_rename(), and integer overflows in MFT cluster validation. It also adds basic support for NTFS Alternative Data Streams, which parses the colon delimiter in filenames and handles UTF-16 name conversion.

ext4 gets parallel DIO writes unlocked for better performance when safe. The ext4_mb_prefetch() optimization for fallocate() avoids unnecessary work by checking conditions before prefetching. btrfs removes free space v1 as the default since the free space tree v2 has been the mkfs default since kernel 5.15. Backward compatibility is maintained for existing v1 filesystems.

A new af_alg_restrict sysctl and whitelist allow administrators to restrict which algorithms can be used via the AF_ALG socket interface. This tightens security around cryptographic operations exposed through userspace.

Linus Torvalds found and fixed a bug in the Intel Xe GPU driver himself. The bug was a subtle VRAM offset calculation error in get_flat_ccs_offset() that caused black screens on Battlemage G21 GPUs. The fix was literally one character: changing round_up() to round_down(). Torvalds used Gemini AI as his "tireless helper" throughout the debugging process, though the AI repeatedly insisted the bug was impossible and unsolvable. It took 24 patches adding debug code and 18 kernel boots from startup to narrow down the issue. The commit message itself was written by the AI.

The community is also discussing the impact of AI-generated code on development. ARM64 development reportedly suffered during the 7.2 cycle due to "AI/LLM patch craziness," with no real features landing for AArch64 in that cycle. That gap is being closed in 7.3. Andrew Morton turned to Google's Gemini AI to help write his verbose patch summaries given the unusually heavy patch volume.

Linux 7.3 is expected to land around October 2026 with the first release candidate coming soon. Head here to follow the development progress and check out the specific patches if you're interested in the technical details.