Re: [PATCH] cpuidle: Do not reinstall the idle handler while paused

From: Rafael J. Wysocki (Intel)

Date: Fri Nov 20 2026 - 13:23:03 EST


On Tue, Nov 17, 2026 at 8:01 PM Wondertan via B4 Relay
<devnull+hlibwondertan.gmail.com@xxxxxxxxxx> wrote:
>
> From: Wondertan <hlibwondertan@xxxxxxxxx>
>
> The PM core pauses cpuidle before taking secondary CPUs offline during
> system-wide transitions and resumes it after bringing them back online
> (pm_sleep_disable_secondary_cpus()/pm_sleep_enable_secondary_cpus()), so
> that CPUs do not enter cpuidle-managed idle states in that window.
>
> However, with acpi_idle, acpi_soft_cpu_online() calls
> acpi_processor_hotplug() for every CPU brought back online, and that
> function brackets its update with cpuidle_pause_and_lock() and
> cpuidle_resume_and_unlock(). The latter reinstalls the idle handler
> unconditionally, so cpuidle is effectively resumed as soon as the first
> secondary CPU is back online. The remaining CPUs are then brought up,
> and cache_aps_init() runs stop_machine(), with idle CPUs entering deep
> ACPI C-states. cpuidle_register_device() and cpuidle_switch_governor()
> can undo the pause in the same way.
>
> On a Framework Laptop 16 (AMD Ryzen AI 9 HX 370, BIOS 04.01), whose
> _CST exposes C2/C3 with the SystemIO entry method, this intermittently
> hangs hibernation after the image has been created: a CPU idling in
> io_idle() does not respond to an IPI and another CPU soft-locks waiting
> for it, either in smp_call_function_many_cond() (text poking from
> sched_cpu_activate() while an SMT sibling is onlined) or in
> multi_cpu_stop() (the MTRR rendezvous in cache_aps_init()). The image is
> never written and the next boot reports "PM: Image not found (code -22)":
>
> watchdog: BUG: soft lockup - CPU#12 stuck for 22s! [cpuhp/12:89]
> RIP: 0010:smp_call_function_many_cond+0x147/0x520
> on_each_cpu_cond_mask+0x24/0x40
> smp_text_poke_batch_finish+0xe2/0x380
> arch_jump_label_transform_apply+0x1a/0x30
> static_key_slow_inc_cpuslocked+0x4f/0xa0
> sched_cpu_activate+0x13c/0x150
> ...
> thaw_secondary_cpus.cold+0x5e/0xd2
> hibernation_snapshot+0x231/0x530
> NMI backtrace for cpu 0 skipped: idling at io_idle+0x3/0x30
>
> watchdog: BUG: soft lockup - CPU#3 stuck for 22s! [migration/3:37]
> Stopper: multi_cpu_stop+0x0/0x140 <- __stop_cpus.constprop.0+0x6d/0xb0
> (23 CPUs in multi_cpu_stop)
> NMI backtrace for cpu 2 skipped: idling at io_idle+0x3/0x30
>
> Online CPUs can only be in io_idle() via cpuidle, so cpuidle was not
> paused at that point. Tracing power:cpu_idle across the transition on
> 6.18.59 confirms it: while CPUs 1-23 were being brought online (between
> the first and the last CPU_ON "end" events, about 35 ms), idle CPUs
> entered C2 263 times and C3 305 times on all 24 CPUs, the first C3 entry
> coming 0.6 ms after CPU1 was onlined. The failure rate was about one in
> four hibernations, higher on battery; disabling C2/C3 via sysfs around
> hibernation made it go away.
>
> Fix this by recording that cpuidle has been paused by cpuidle_pause()
> and not reinstalling the idle handler until cpuidle_resume() is called.
> All callers of cpuidle_install_idle_handler() hold cpuidle_lock, which
> also protects the new flag.
>
> With this change on top of 6.18.62 and 7.3.4, the same trace across a
> real hibernation and resume with C2/C3 enabled shows no C2/C3 entries
> in that window (only C1), hibernation completes, and C2/C3 are used
> again once the transition is over.
>
> The root cause analysis, the fix, the tracing used to verify it and
> this changelog were produced with an AI coding assistant (Claude Code,
> model claude-opus-5-5), working from pstore dumps, ftrace captures and
> the kernel sources. All builds and tests were run on the affected
> machine as described above.
>
> Fixes: 8651f97bd951 ("PM / cpuidle: System resume hang fix with cpuidle")
> Cc: stable@xxxxxxxxxxxxxxx
> Signed-off-by: Wondertan <hlibwondertan@xxxxxxxxx>
> ---
> Tested only on the one affected machine (AMD, acpi_idle, BIOS 04.01
> and 04.02): built and run-tested on 6.18.62 and 7.3.4 (Arch linux-lts
> and linux configs, x86_64). drivers/cpuidle/cpuidle.c in 7.3.4 is
> identical to mainline apart from this change. On mainline
> (141b164bbe18) drivers/cpuidle/ was only built (W=1, no warnings), not
> booted. Not tested with intel_idle or other cpuidle drivers.
>
> Earlier report of the hangs, before the root cause was known:
> "[REGRESSION?] hibernation soft lockup in migration/0 after image
> creation", linux-pm, 2026-07-09.
> ---
> drivers/cpuidle/cpuidle.c | 10 +++++++++-
> 1 file changed, 9 insertions(+), 1 deletion(-)
>
> diff --git a/drivers/cpuidle/cpuidle.c b/drivers/cpuidle/cpuidle.c
> index 2d2f40a2c..3e26b1760 100644
> --- a/drivers/cpuidle/cpuidle.c
> +++ b/drivers/cpuidle/cpuidle.c
> @@ -39,6 +39,7 @@ LIST_HEAD(cpuidle_detected_devices);
> static int enabled_devices;
> static int off __read_mostly;
> static int initialized __read_mostly;
> +static bool paused; /* Protected by cpuidle_lock */
>
> int cpuidle_disabled(void)
> {
> @@ -465,7 +466,12 @@ __cpuidle u64 cpuidle_poll_time(struct cpuidle_driver *drv,
> */
> void cpuidle_install_idle_handler(void)
> {
> - if (enabled_devices) {
> + /*
> + * Do not reinstall the idle handler while cpuidle is paused for a
> + * system transition, even if a nested cpuidle_resume_and_unlock()
> + * (e.g. from the ACPI processor CPU online callback) asks for it.
> + */
> + if (enabled_devices && !paused) {
> /* Make sure all changes finished before we switch to new idle */
> smp_wmb();
> initialized = 1;
> @@ -515,6 +521,7 @@ EXPORT_SYMBOL_GPL(cpuidle_resume_and_unlock);
> void cpuidle_pause(void)
> {
> mutex_lock(&cpuidle_lock);
> + paused = true;
> cpuidle_uninstall_idle_handler();
> mutex_unlock(&cpuidle_lock);
> }
> @@ -523,6 +530,7 @@ void cpuidle_pause(void)
> void cpuidle_resume(void)
> {
> mutex_lock(&cpuidle_lock);
> + paused = false;
> cpuidle_install_idle_handler();
> mutex_unlock(&cpuidle_lock);
> }
>
> ---

Applied as 7.5 material, thanks!