Fixes break
software too.
Critical fixes and security patches can’t wait. But roughly 1 in 30 breaks something new.* Now a year of fixes is landing in weeks. See what each fix breaks before it lands.
Eight patches land in eight files of a repository, half from upstream and half written in-house. The first four change nothing here. Then open connections climb, dip for a moment, climb faster and reach their limit. New connections are refused. Patch 5 breaks a cleanup assumption in lib/reaper.c, a file unchanged since 2011 and outside every patch.
Eight patches land. One starts a leak. Each is reasonable on its own. Watch the open connections.
Open connections
against a limit of 4,096
Our estimate from 97,933 Linux kernel runtime fixes: 3.4% broke something new that needed a fix of its own within a year.
Reviewed, tested, shipped. It still broke something.
In November 2024 the Linux kernel merged a fix for a high-severity security hole. Two reviewers approved it, a third developer tested it, and it worked.
It also broke something nobody was checking: a file sealed against writing could no longer be mapped read-only. The break shipped in Linux 6.12. A user reported it seventeen days after the merge, and the repair reached the kernel eight weeks after the fix did.
Read the full write-upWhy it broke: the fix moved a safety check earlier
The mapping is marked read-only, then the seal is checked. Allowed.
The seal is checked first, before the mapping is marked read-only. Refused.
17 daysuntil a user reported it
Our approach: understand a fix’s impact instead of just trusting it.
We ran fix verification on that fix with only what a reviewer had on the day it merged: the patch and the code before it. It flagged the same break in under six minutes of model and test time.
- 01
It works out what else must hold.
It traces the hidden, long-range protocols your fix touches. In C and systems code, those live in convention, not in types.
- 02
It goes looking for the break.
It writes tests that exercise each of those protocols, then runs them on the code before and after the fix.
- 03
It hands you a result, not an opinion.
“This might affect sealed files” invites a debate. A failing test, with the passing run on the old code beside it, is something you can act on.