[PATCH] riscv: Always inline on_thread_stack()

[Karl Mehltretter]

With CONFIG_IRQ_STACKS=y, GCC may outline on_thread_stack().
CONFIG_KCOV_INSTRUMENT_ALL instruments it, and do_irq() calls it before
irq_enter_rcu() increments the hardirq count.

Normal KCOV records this interrupt coverage in the task's buffer, making
results timing-dependent.

With CONFIG_KCOV_SELFTEST=y, current->kcov_area remains NULL. In QEMU, a
timer interrupt called the helper before hardirq accounting. KCOV treated
it as task coverage and dereferenced NULL. Fault handling recursed to stack
overflow and stopped boot.

On a BeagleV Ahead, the unpatched kernel did not finish the selftest within
120 seconds. Three patched boots completed it and reached init.

Mark the helper __always_inline to keep it in noinstr do_irq(), following
commit 1894a4030582 ("x86: Always inline on_thread_stack() and
current_top_of_stack()"), which fixed the same out-of-line call from x86
noinstr code.

Fixes: 163e76cc6ef4 ("riscv: stack: Support HAVE_IRQ_EXIT_ON_IRQ_STACK")
Cc: stable at vger.kernel.org
Assisted-by: LLM
Signed-off-by: Karl Mehltretter <kmehltretter at gmail.com>
---
Tested with GCC 15.2.0 and QEMU 10.2.1 on v7.3-rc7. Both images
used byte-identical configs and v7.3-rc7 rebases of these seven pending
patches (canonical commit IDs shown):

  3a25097bf845 kcov: use unsigned int for kcov_start() mode parameter
  e39a9c820c37 kcov: add kcov_pause()/kcov_resume() helpers
  ae2662563c3e hrtimer: pause KCOV during deferred rearm
  554af946139e sched: pause KCOV in __schedule()
  7f608b777204 sched: pause KCOV in try_to_wake_up()
  2ba73ec2a246 sched: pause KCOV in wake_up_new_task()
  5ee777070fac context_tracking: exclude from KCOV instrumentation

                              plain inline  __always_inline
KCOV selftest started                  yes              yes
KCOV selftest finished                  no              yes
Reached /init                           no              yes

The plain-inline image stopped after printing "kcov: running self test"
and reached the 45-second limit. The patched image reached /init in two
60-second runs. Those runs timed out after the marker because the test
/init deliberately loops.

The first common patch is a type prerequisite. The remaining pause and
exclusion changes suppress separate KCOV callbacks that otherwise obscure
this test. CONFIG_RISCV_ISA_V_PREEMPTIVE was disabled to avoid a separate
vector-entry callback. The only A/B source difference was this patch.

The same images were tested on a BeagleV Ahead (TH1520, four C910 CPUs).
The plain-inline image did not finish the selftest or reach /init during
120 seconds after the start message. It reported RCU stalls, and CPU 1
did not answer NMI requests. No panic or stack-overflow message was
printed. All three patched runs completed the selftest in about 0.301
seconds and reached /init.

 arch/riscv/include/asm/stacktrace.h | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/arch/riscv/include/asm/stacktrace.h b/arch/riscv/include/asm/stacktrace.h
index b1495a7e06ce..c7c991452e11 100644
--- a/arch/riscv/include/asm/stacktrace.h
+++ b/arch/riscv/include/asm/stacktrace.h
@@ -16,7 +16,7 @@ extern void notrace walk_stackframe(struct task_struct *task, struct pt_regs *re
 extern void dump_backtrace(struct pt_regs *regs, struct task_struct *task,
 			   const char *loglvl);
 
-static inline bool on_thread_stack(void)
+static __always_inline bool on_thread_stack(void)
 {
 	return !(((unsigned long)(current->stack) ^ current_stack_pointer) & ~(THREAD_SIZE - 1));
 }

base-commit: 24a19b953e2a3d0ea5a8c4951a0ab8f32720c907
-- 
2.53.0