freenode

← digests

ipset double-free, KASLR filter hole, and deadline accounting fixes

Kernel & Low-Level2026-11-20

Kernel security work led the day: an ipset double-free that can escalate from unprivileged namespaces, a tracing filter check that closes a KASLR bypass, and a stable CVE for a pagemap use-after-free. Scheduler deadline accounting, MGLRU promotion, and device-trust binding also advanced, while an MTD maintainer asked a submitter to stop flooding the list with generated patches.

ipset excludes add-path GC during resize

A netfilter ipset patch excludes add-path garbage collection while a set is being resized, fixing a double-free. The failure can be used for local privilege escalation from unprivileged namespaces. Anyone exposing ipset to user namespaces should treat the fix as a privilege boundary, not only a crash fix.

Tracing .function filter requires kallsyms permission

A tracing patch requires kallsyms permission before the .function filter modifier is accepted. The check closes a KASLR bypass reachable through a perf ioctl when the paranoid setting is left at its default of 2.

Stable pin of mm_users in do_pagemap_scan

CVE-2026-99477 is a stable backport that pins mm_users in do_pagemap_scan() under fs/proc/task_mmu. Without the pin, a scan can race a teardown of hugetlb VMAs and hit a use-after-free.

SCHED_DEADLINE bandwidth accounting fixes and selftests

A resend of a 17-patch series (v4, after v3) fixes 14 SCHED_DEADLINE bugs, including crashes, mutex leaks, bandwidth accounting errors, and wrong admission. The series also adds debug output and selftests. Deadline admission depends on those counters, so the bugs affect whether the scheduler accepts work it can actually meet.

MGLRU frequency-guided promotion reaches v5

A 17-patch v5 series adds frequency-guided promotion to MGLRU, fixes metrics and PSI accounting, and drops a page flag. Andrew Morton called the work really good and offered a Tested-by. The series is the main memory-reclaim proposal drawing maintainer attention in this batch.

Driver core authorizes device trust before binding

A 31-patch v3 series adds a device trust policy and requires authorization before a driver binds, covering the driver core plus PCI, USB, and IOMMU. The change moves a trust decision to the point of binding rather than leaving it implicit in each bus.

MTD maintainer objects to a flood of generated patches

A 58-patch v5 series proposed retaining NAND chip and controller state across removal. The maintainer publicly called out the submitter for flooding the list with hundreds of LLM-generated MTD patches and asked them to stop. The dispute is about review load on the MTD list, not only the removal-state change.