Re: [PATCH] notifier: Make atomic_notifiers use raw_spinlock

From: Peter Zijlstra
Date: Mon Nov 23 2020 - 09:14:57 EST


On Sun, Nov 22, 2020 at 08:19:04PM +0000, Valentin Schneider wrote:
> Booting a recent PREEMPT_RT kernel (v5.10-rc3-rt7-rebase) on my arm64 Juno
> leads to the idle task blocking on an RT sleeping spinlock down some
> notifier path:
>
> [ 1.809101] BUG: scheduling while atomic: swapper/5/0/0x00000002
> [ 1.809116] Modules linked in:
> [ 1.809123] Preemption disabled at:
> [ 1.809125] secondary_start_kernel (arch/arm64/kernel/smp.c:227)
> [ 1.809146] CPU: 5 PID: 0 Comm: swapper/5 Tainted: G W 5.10.0-rc3-rt7 #168
> [ 1.809153] Hardware name: ARM Juno development board (r0) (DT)
> [ 1.809158] Call trace:
> [ 1.809160] dump_backtrace (arch/arm64/kernel/stacktrace.c:100 (discriminator 1))
> [ 1.809170] show_stack (arch/arm64/kernel/stacktrace.c:198)
> [ 1.809178] dump_stack (lib/dump_stack.c:122)
> [ 1.809188] __schedule_bug (kernel/sched/core.c:4886)
> [ 1.809197] __schedule (./arch/arm64/include/asm/preempt.h:18 kernel/sched/core.c:4913 kernel/sched/core.c:5040)
> [ 1.809204] preempt_schedule_lock (kernel/sched/core.c:5365 (discriminator 1))
> [ 1.809210] rt_spin_lock_slowlock_locked (kernel/locking/rtmutex.c:1072)
> [ 1.809217] rt_spin_lock_slowlock (kernel/locking/rtmutex.c:1110)
> [ 1.809224] rt_spin_lock (./include/linux/rcupdate.h:647 kernel/locking/rtmutex.c:1139)
> [ 1.809231] atomic_notifier_call_chain_robust (kernel/notifier.c:71 kernel/notifier.c:118 kernel/notifier.c:186)
> [ 1.809240] cpu_pm_enter (kernel/cpu_pm.c:39 kernel/cpu_pm.c:93)
> [ 1.809249] psci_enter_idle_state (drivers/cpuidle/cpuidle-psci.c:52 drivers/cpuidle/cpuidle-psci.c:129)
> [ 1.809258] cpuidle_enter_state (drivers/cpuidle/cpuidle.c:238)
> [ 1.809267] cpuidle_enter (drivers/cpuidle/cpuidle.c:353)
> [ 1.809275] do_idle (kernel/sched/idle.c:132 kernel/sched/idle.c:213 kernel/sched/idle.c:273)
> [ 1.809282] cpu_startup_entry (kernel/sched/idle.c:368 (discriminator 1))
> [ 1.809288] secondary_start_kernel (arch/arm64/kernel/smp.c:273)
>
> Two points worth noting:
>
> 1) That this is conceptually the same issue as pointed out in:
> 313c8c16ee62 ("PM / CPU: replace raw_notifier with atomic_notifier")
> 2) Only the _robust() variant of atomic_notifier callchains suffer from
> this
>
> AFAICT only the cpu_pm_notifier_chain really needs to be changed, but
> singling it out would mean introducing a new (truly) non-blocking API. At
> the same time, callers that are fine with any blocking within the call
> chain should use blocking notifiers, so patching up all atomic_notifier's
> doesn't seem *too* crazy to me.

How long are these notifier chains?, and all this pcs_enter_idle_state()
is still horribly broken vs RCU, witness the RCU_NONIDLE() there and the
rcu_irq_enter_irqson() in the pm_notifier code.

That said, we're running these notifiers from the idle path with IRQs
disabled, so taking that spinlock isn't going to make it worse..