Chinese Engineers Boost Linux 7.3; Linus Uses AI to Fix One‑Char Bug

Chinese Engineers Lead Major Linux 7.3 Optimizations

The Linux 7.3 merge window opened immediately after the 7.2 release, and three Chinese‑origin contributions dominate the performance work.

1. ByteDance reduces SMP call latency

The smp_call_function*() family disables preemption while waiting for IPI responses. ByteDance engineer Chuyi Zhou observed on a 16‑core machine that pure IPI wait could inflate scheduling latency to ~16 ms, fatal for real‑time frameworks like DPDK. Their patch records the target CPU in each task’s CPU mask, re‑enables preemption before the wait, and cuts the worst‑case latency of high‑priority DPDK tasks from ~17 ms to ~1.5 ms – a roughly 90 % reduction. The change passed ten rounds of review before merging.

2. ZTE eliminates a 705 ms lock in KSM

ZTE engineer Xu Xin identified a severe performance issue in the reverse‑mapping function rmap_walk_ksm(), where under memory pressure an internal lock could be held for up to 705 ms, causing applications to freeze for hundreds of milliseconds. The lock is triggered by common patterns such as JVM/Go mmap‑based heap allocation and database page reclamation, which split a large VMA into tens of thousands of shared anon_vmas. After fixing the lock, a Java workload that previously held the lock for 228 ms saw the worst lock hold drop to 1.44 ms (average) or 1.67 ms (peak), a ~490× speed‑up.

3. Xiaomi improves zsmalloc lock contention

Xiaomi engineer Wenchao Hao targeted the zsmalloc allocator’s zs_free() lock, a bottleneck for compressed memory (zRAM, Zswap) on Android and low‑end devices. Benchmarks on a Raspberry Pi 4B showed a 1.83× speed‑up in concurrent memory release, while a 20‑core Intel server achieved a 1.4× improvement, making memory‑constrained devices feel noticeably snappier.

Linus Torvalds Uses Gemini AI to Find a One‑Character Bug

Although Linus Torvalds has not written kernel code for about 20 years, he personally fixed a bug in the Intel Xe driver during this merge window, assisted by Gemini AI. The defect lay in get_flat_ccs_offset(), where an incorrect VRAM offset calculation marked already‑used memory as free, causing the Battlemage G21 graphics boot to loop and black‑screen. The fix required changing a single character: replacing round_up() with round_down(). Linus described the debugging process as “hellish,” noting that the AI repeatedly claimed the issue was unsolvable before finally adding debugging code that led to the fix after 24 temporary patches and 18 kernel reboots.

Other Notable Additions in Linux 7.3

sched_ext : BPF scheduler framework graduates, allowing root schedulers to delegate cgroup sub‑trees to nested schedulers with pluggable CPU quotas.

AMD UALink : 95 patches enable direct connection of up to 1 024 accelerators within a single domain, laying groundwork for AI‑cluster interconnects.

Apple M3 Pro/Max/Ultra : Initial device‑tree support merged, aiding Asahi Linux on Apple silicon.

Rust on PowerPC : Rust support extended to PPC32BE and PPC64LE, moving from experimental to maintained.

NTFS3 hardening : 21 commits (922 lines) fix slab over‑/under‑flows, information leaks, integer overflows, and add ADS parsing.

Filesystem throughput boosts : Btrfs now uses IOmap bounce buffers, raising throughput from ~50 % to 95 %; EXT4 enables parallel direct writes, pushing NVMe IOPS from 1.92 M to 2.19 M.

Legacy hardware improvements : Peter Zijlstra compresses cgroup scheduling to a single run‑queue, stabilizing frame rates on Sandy Bridge + RX 580; Ricardo Neri fixes Intel hybrid‑core scheduling issues.

Why Linux 7.3 May Become This Year’s LTS

Typically only one LTS kernel is released per year; 7.3 is the leading candidate. If selected, all the performance patches, AMD UALink, Rust on PowerPC, and other enhancements will receive long‑term maintenance across major distributions.

Ubuntu Adoption Timeline

Ubuntu 26.10, scheduled for 15 Oct, ships with the 7.2 kernel, so 7.3 will likely arrive via later HWE updates. Users can try 7.3 by waiting for the rc1 (expected 30 Aug) and installing via the Mainline PPA, compiling from source, or waiting for the final release in late October.

Conclusion

The Linux 7.3 merge window demonstrates both the growing impact of Chinese developers on core kernel performance and the emerging role of AI assistants in low‑level debugging, while delivering a suite of substantial new features that position the release as a strong long‑term support candidate.