Build2 publishers2 min readPublished
Large test patches inflate the selftest share of Linux 7.3-rc5's diffstat
Linus Torvalds said selftests make up a quarter of the Linux 7.3-rc5 patch by line count, a split he called unusual. A few very large test patches produce that share, so the size of the rc alone gives teams planning for 7.3 little reason to wait.
The Engineer · Build desk
What happened
- Torvalds wrote that the single biggest patch in rc5 is, he thinks, the arm64 KVM ITS table test.
- Most of the rc5 driver changes were networking, with DRM graphics drivers a close second.
- The rc5 merges include more Cache Aware Scheduling fixes, support for new mice and other HID devices, and a revert of Logitech Holt Receiver HID++ support over issues.
- Linux 7.3 is due for release in October and will power Ubuntu 26.10 and other autumn distributions.
Compiled by The EngineerSomething wrong?How this is made
Why it matters
- contradiction Phoronix ties the kernel's busy pace to AI agents while Torvalds sees no pattern beyond size, so the selftest quarter cannot be read as a tally of agent-found bugs in rc5.
- decision Teams qualifying 7.3 should weight testing toward KVM, scheduler and BPF paths, where most of rc5's change outside drivers and tests landed.
- cost Anyone who validated Logitech receivers against HID++ support in an earlier 7.3 candidate has to retest on rc5, because that support has been reverted.
- exposure Ubuntu 26.10 users inherit whatever the remaining 7.3 candidates fix, or miss, in the KVM and scheduler code still being patched.
Torvalds called the selftest figure "admittedly a bit misleading" in the same paragraph that reported it [3]. The quarter is a share of the diffstat, the summary that weighs each patch by how many lines it changes [2]. He explained that "most of the driver patches are rather small, and some of the selftest patches are much bigger" [3]. A handful of big test patches can hold a large share of the lines without making up a large share of the commits [3][5]. Counting commits, he wrote, "it skews quite differently" [5].
The line shares still show where the code outside the tests sits. Drivers were under a third of the diff [2]. Add the selftest quarter, and more than 40 percent is left for everything else [6]. Torvalds put most of that remainder in architecture fixes, "largely kvm-related, some perf", and "a fair amount of core kernel (both scheduler and bpf)" [7]. Networking, core VFS and some XFS fixes make up the rest [7].
According to Phoronix, kernel development is as busy as ever, "driven in large part by AI/LLM agents continuing to uncover new issues" in a codebase of nearly 41 million lines [10]. The attribution is Phoronix's own, and it covers kernel development in general [10]. The announcement text Phoronix published does not mention AI agents [15]. Torvalds explained the size more briefly: "I don't think there's any real pattern to it other than 'big'." [11]
On a fleet without KVM hosts, rc5 is an even weaker case for extra soak time. The architecture fixes Torvalds listed are largely KVM-related [7]. His rating of the diffstat was "Not *bad* unusual, just not what I typically see." [12] He added that "nothing looks particularly alarming" and ended with "Please keep testing" [11][14].
What to watch
- The rc6 and rc7 announcements: whether Torvalds again calls the release large, and whether the selftest share drops once the big KVM test patches are in.
- Whether Logitech Holt Receiver HID++ support returns before the October release or moves to a later kernel.
- Further Cache Aware Scheduling fixes in rc6 or later, a sign the scheduler work is still settling.