[PATCH 0/3] sched/preempt: Check notrace preempt-count updates

From: Karl Mehltretter

Date: Mon Nov 16 2026 - 02:54:47 EST


With CONFIG_DEBUG_PREEMPT, the notrace preemption macros bypass the
bounds checks used by ordinary operations. Add checked noinstr updates
for their existing callers and share validation with
preempt_count_add() and preempt_count_sub().

Patch 1 forces the arm64 raw updates called by the new helpers inline.
Patch 2 does the same for preempt_count(), which the helpers use, for
should_resched(), reached by the debug notrace enable path, and for the
existing non-debug __preempt_count_dec_and_test() path. The corresponding
x86, s390 and asm-generic operations are already __always_inline.

Patch 1 is extracted from Hongyan Xia's entry RFC. Song Liu also posted
notrace annotations for the same arm64 helpers:

https://lore.kernel.org/r/20261023011049.1164625-10-song@xxxxxxxxxx

Patch 3 disables branch profiling for core.o when DEBUG_PREEMPT and
TRACE_BRANCH_PROFILING are both enabled. DEBUG_LOCKS_WARN_ON() evaluates
its condition before entering its instrumentation region, so profiling
that condition would call ftrace_likely_update() from noinstr text. The
scheduler Makefile already uses this per-object exclusion for
build_policy.o and build_utility.o.

The cost is loss of branch-profile data for the rest of core.o in that
configuration. Scheduler tracepoints and function tracing are unchanged.
A dedicated non-profiled object is the fallback if this object-wide cost
is considered too broad.

This series is independently useful. It also supplies the checked count
update planned for follow-up IRQ-entry work; it does not change IRQ entry,
trace-recursion lifetime or TRACE_CTX_TRANSITION.

This addresses Peter's suggestion here:

https://lore.kernel.org/r/20261029093219.GH1779021@xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx

The series is based on v7.4-rc2.

Testing:

- Clang 19.1.7 x86-64: a W=1 defconfig-derived build with
DEBUG_PREEMPT, annotated branch profiling, DEBUG_ENTRY and noinstr
validation passed. A nested-KVM boot passed the built-in and external
module checks and all five tracer workloads. A DEBUG_PREEMPT=n control
retained branch profiling for core.o and emitted no new helper symbols.
- GCC 14.2 arm64: a W=1 PREEMPT_RT build and allmodconfig passed. The
latter's nVHE object had no undefined references to the new helpers.
A native Axion boot passed the built-in and module checks and all five
tracers.
- GCC 14.2 and 16.2 ARC, C-SKY and MIPS: performance, size, all-branches
and forced-noinline core.o inspections passed without prohibited
noinstr relocations or branch-profile records. Full baseline and patched
kernels plus the external debug module built. Matching QEMU boots reached
the expected balanced and invalid-count markers on ARC and C-SKY UP and
MIPS UP/SMP. The patched kernels diagnosed every invalid operation;
rejected decrements retained their counts. An ARC DEBUG_PREEMPT=n
control retained core.o profiling without references to the new helpers.

The MIPS UP baseline and patched positive boots emitted the same three
existing smp_processor_id() diagnostics; the SMP boots emitted none.

I used an LLM to help extract and review the changes, write the test
harnesses and prepare the commit messages.

Hongyan Xia (1):
arm64: Always inline raw preempt count updates

Karl Mehltretter (2):
arm64: Always inline preempt count and reschedule checks
sched/preempt: Check preempt counts in notrace helpers

arch/arm64/include/asm/preempt.h | 10 ++++----
include/linux/preempt.h | 22 ++++++++++++++----
kernel/sched/Makefile | 3 +++
kernel/sched/core.c | 40 ++++++++++++++++++++++++++++----
4 files changed, 62 insertions(+), 13 deletions(-)


base-commit: 7475ea8b58616b33c04afb976cf0fbaad9dd63fb