Re: [PATCH v2] firmware: arm_sdei: Fix sleep from invalid context BUG

[Pierre Gondois]

Hello James,

On 2/10/23 18:42, James Morse wrote:
> Hi Pierre,
>
> Sorry its taken so long for me to catch up with this ...

No worries.

> On 18/10/2022 14:04, Pierre Gondois wrote:
> > On an Ampere Altra,
> > Running a preemp_rt kernel based on v5.19-rc3-rt5 on an
> > Ampere Altra triggers:
> > [   15.683141] BUG: sleeping function called from invalid context at kernel/locking/spinlock_rt.c:46
> > [   15.683154] in_atomic(): 0, irqs_disabled(): 128, non_block: 0, pid: 24, name: cpuhp/0
> > [   15.683157] preempt_count: 0, expected: 0
> > [   15.683159] RCU nest depth: 0, expected: 0
> > [   15.683163] 3 locks held by cpuhp/0/24:
> > [   15.683167]  #0: ffffda30217c70d0 (cpu_hotplug_lock){++++}-{0:0}, at: cpuhp_thread_fun+0x5c/0x248
> > [   15.683201]  #1: ffffda30217c7120 (cpuhp_state-up){+.+.}-{0:0}, at: cpuhp_thread_fun+0x5c/0x248
> > [   15.683205]  #2: ffffda3021c711f0 (sdei_list_lock){....}-{3:3}, at: sdei_cpuhp_up+0x3c/0x130
> > [   15.683224] irq event stamp: 36
> > [   15.683226] hardirqs last  enabled at (35): [<ffffda301e85b7bc>] finish_task_switch+0xb4/0x2b0
> > [   15.683236] hardirqs last disabled at (36): [<ffffda301e812fec>] cpuhp_thread_fun+0x21c/0x248
> > [   15.683238] softirqs last  enabled at (0): [<ffffda301e80b184>] copy_process+0x63c/0x1ac0
> > [   15.683245] softirqs last disabled at (0): [<0000000000000000>] 0x0
> > [   15.683258] CPU: 0 PID: 24 Comm: cpuhp/0 Not tainted 5.19.0-rc3-rt5-[...]
> > [   15.683265] Hardware name: WIWYNN Mt.Jade Server System B81.03001.0005/Mt.Jade Motherboard, BIOS 1.08.20220218 (SCP: 1.08.20220218) 2022/02/18
> > [   15.683268] Call trace:
> > [   15.683271]  dump_backtrace+0x114/0x120
> > [   15.683277]  show_stack+0x20/0x70
> > [   15.683279]  dump_stack_lvl+0x9c/0xd8
> > [   15.683288]  dump_stack+0x18/0x34
> > [   15.683289]  __might_resched+0x188/0x228
> > [   15.683292]  rt_spin_lock+0x70/0x120
> > [   15.683301]  sdei_cpuhp_up+0x3c/0x130
> > [   15.683303]  cpuhp_invoke_callback+0x250/0xf08
> > [   15.683305]  cpuhp_thread_fun+0x120/0x248
> > [   15.683308]  smpboot_thread_fn+0x280/0x320
> > [   15.683315]  kthread+0x130/0x140
> > [   15.683321]  ret_from_fork+0x10/0x20
>
> > sdei_cpuhp_up() is called in the STARTING hotplug section,
> > which runs whith interrupts disabled. Move CPUHP_AP_ARM_SDEI_
> > state to the _ONLINE section to execute the cpuhp cb with
> > preemption enabled.
>
> The background to this is SDEI got its own cpuhp slot because 'perf NMI' support
> was one of the use-cases, but this got superseded by pNMI. Without an interaction with
> perf, the slot doesn't need to be that early.
>
>
> > Some SDEI calls (e.g. SDEI_1_0_FN_SDEI_PE_MASK) take actions on the
> > calling CPU. It is checked that preemption is disabled for them.
> > _ONLINE cpuhp cb are executed in the 'per CPU hotplug thread'.
> > Preemption is enabled in those threads, but their cpumask is limited
> > to 1 CPU.
>
> > Move 'WARN_ON_ONCE(preemptible())' statements so that SDEI cpuhp cb
> > don't trigger them. This means that no check will be done for some
> > cases, e.g. sdei_mask_local_cpu() invocations.
>
> (these things are documentation anyway)
>
>
> > Also add a check for the SDEI_1_0_FN_SDEI_PRIVATE_RESET SDEI call
> > which acts on the calling CPU.
>
> > diff --git a/drivers/firmware/arm_sdei.c b/drivers/firmware/arm_sdei.c
> > index 1e1a51510e83..9b03e164a37a 100644
> > --- a/drivers/firmware/arm_sdei.c
> > +++ b/drivers/firmware/arm_sdei.c
>
> > @@ -401,6 +399,8 @@ int sdei_event_enable(u32 event_num)
> >   	int err = -EINVAL;
> >   	struct sdei_event *event;
> >   
> > +	WARN_ON_ONCE(preemptible());
> > +
> >   	mutex_lock(&sdei_events_lock);
> >   	event = sdei_event_find(event_num);
> >   	if (!event) {
>
> This doesn't look right. How can this code take a mutex if its in a non-preemptable context?
>
> > @@ -492,6 +490,7 @@ int sdei_event_unregister(u32 event_num)
> >   	struct sdei_event *event;
> >   
> >   	WARN_ON(in_nmi());
> > +	WARN_ON_ONCE(preemptible());
> >   
> >   	mutex_lock(&sdei_events_lock);
> >   	event = sdei_event_find(event_num);
>
> Same again.
>
> > @@ -576,6 +573,7 @@ int sdei_event_register(u32 event_num, sdei_event_callback *cb, void *arg)
> >   	struct sdei_event *event;
> >   
> >   	WARN_ON(in_nmi());
> > +	WARN_ON_ONCE(preemptible());
> >   
> >   	mutex_lock(&sdei_events_lock);
> >   	if (sdei_event_find(event_num)) {
>
> Same again.
>
> I think you've copied these here because they called _local_event_unregister(), but they
> did that via on_each_cpu(), which is what made _local_event_unregister() non-preemtable.

Yes right, these 3 warnings don't really make sense. They will be removed.

> You can just remove these three, the original warnings were mostly documentation, and to
> catch myself out. (before RT moved the goal posts).
>
> With that:
> Reviewed-by: James Morse <james.morse@xxxxxxx>
>
> (could you trim the [timestamps] out of the commit log?)

Yes sure.

> > @@ -765,7 +765,7 @@ static int sdei_device_freeze(struct device *dev)
> >   	int err;
> >   
> >   	/* unregister private events */
> > -	cpuhp_remove_state(CPUHP_AP_ARM_SDEI_STARTING);
> > +	cpuhp_remove_state(CPUHP_AP_ARM_SDEI_ONLINE);
>
> Is there any mileage in making this CPUHP_AP_ONLINE_DYN ?
> Perf really was the only reason that this needed to happen in any particular order.

Ok I will do that.

Thanks for the review,
Regards,
Pierre

> Thanks,
>
> James