Re: [PATCHv2 1/5] arm64/entry-common: push the judgement of nmi ahead

From: Pingfan Liu
Date: Fri Oct 08 2021 - 23:49:57 EST


On Fri, Oct 08, 2021 at 06:25:13PM +0100, Mark Rutland wrote:
> On Fri, Oct 08, 2021 at 10:55:04PM +0800, Pingfan Liu wrote:
> > On Fri, Oct 08, 2021 at 12:01:25PM +0800, Pingfan Liu wrote:
> > > Sorry that I missed this message and I am just back from a long
> > > festival.
> > >
> > > Adding Paul for RCU guidance.
> > >
> > > On Thu, Sep 30, 2021 at 02:32:57PM +0100, Mark Rutland wrote:
> > > > On Sat, Sep 25, 2021 at 11:39:55PM +0800, Pingfan Liu wrote:
> > > > > On Fri, Sep 24, 2021 at 06:53:06PM +0100, Mark Rutland wrote:
> > > > > > On Fri, Sep 24, 2021 at 09:28:33PM +0800, Pingfan Liu wrote:
> > > > > > > In enter_el1_irq_or_nmi(), it can be the case which NMI interrupts an
> > > > > > > irq, which makes the condition !interrupts_enabled(regs) fail to detect
> > > > > > > the NMI. This will cause a mistaken account for irq.
> > > > > >
> > > > > Sorry about the confusing word "account", it should be "lockdep/rcu/.."
> > > > >
> > > > > > Can you please explain this in more detail? It's not clear which
> > > > > > specific case you mean when you say "NMI interrupts an irq", as that
> > > > > > could mean a number of distinct scenarios.
> > > > > >
> > > > > > AFAICT, if we're in an IRQ handler (with NMIs unmasked), and an NMI
> > > > > > causes a new exception we'll do the right thing. So either I'm missing a
> > > > > > subtlety or you're describing a different scenario..
> > > > > >
> > > > > > Note that the entry code is only trying to distinguish between:
> > > > > >
> > > > > > a) This exception is *definitely* an NMI (because regular interrupts
> > > > > > were masked).
> > > > > >
> > > > > > b) This exception is *either* and IRQ or an NMI (and this *cannot* be
> > > > > > distinguished until we acknowledge the interrupt), so we treat it as
> > > > > > an IRQ for now.
> > > > > >
> > > > > b) is the aim.
> > > > >
> > > > > At the entry, enter_el1_irq_or_nmi() -> enter_from_kernel_mode()->rcu_irq_enter()/rcu_irq_enter_check_tick() etc.
> > > > > While at irqchip level, gic_handle_irq()->gic_handle_nmi()->nmi_enter(),
> > > > > which does not call rcu_irq_enter_check_tick(). So it is not proper to
> > > > > "treat it as an IRQ for now"
> > > >
> > > > I'm struggling to understand the problem here. What is "not proper", and
> > > > why?
> > > >
> > > > Do you think there's a correctness problem, or that we're doing more
> > > > work than necessary?
> > > >
> > > I had thought it just did redundant accounting. But after revisiting RCU
> > > code, I think it confronts a real bug.
> > >
> > > > If you could give a specific example of a problem, it would really help.
> > > >
> > > Refer to rcu_nmi_enter(), which can be called by
> > > enter_from_kernel_mode():
> > >
> > > ||noinstr void rcu_nmi_enter(void)
> > > ||{
> > > || ...
> > > || if (rcu_dynticks_curr_cpu_in_eqs()) {
> > > ||
> > > || if (!in_nmi())
> > > || rcu_dynticks_task_exit();
> > > ||
> > > || // RCU is not watching here ...
> > > || rcu_dynticks_eqs_exit();
> > > || // ... but is watching here.
> > > ||
> > > || if (!in_nmi()) {
> > > || instrumentation_begin();
> > > || rcu_cleanup_after_idle();
> > > || instrumentation_end();
> > > || }
> > > ||
> > > || instrumentation_begin();
> > > || // instrumentation for the noinstr rcu_dynticks_curr_cpu_in_eqs()
> > > || instrument_atomic_read(&rdp->dynticks, sizeof(rdp->dynticks));
> > > || // instrumentation for the noinstr rcu_dynticks_eqs_exit()
> > > || instrument_atomic_write(&rdp->dynticks, sizeof(rdp->dynticks));
> > > ||
> > > || incby = 1;
> > > || } else if (!in_nmi()) {
> > > || instrumentation_begin();
> > > || rcu_irq_enter_check_tick();
> > > || } else {
> > > || instrumentation_begin();
> > > || }
> > > || ...
> > > ||}
> > >
> >
> > Forget to supplement the context for understanding the case:
> > On arm64, at present, a pNMI (akin to NMI) may call rcu_nmi_enter()
> > without calling "__preempt_count_add(NMI_OFFSET + HARDIRQ_OFFSET);".
> > As a result it can be mistaken as an normal interrupt in
> > rcu_nmi_enter().
>
> I appreciate that there's a window where we treat the pNMI like an IRQ,
> but that's by design, and we account for this in gic_handle_irq() and
> gic_handle_nmi() where we "upgrade" to NMI context with
> nmi_enter()..nmi_exit().
>
> The idea is that we have two cases:
>
> 1) If we take a pNMI from a context where IRQs were masked, we know it
> must be a pNMI, and perform the NMI entry immediately to avoid
> reentrancy problems.
>
> I think we're all happy with this case.
>
Right.

> 2) If we take a pNMI from a context where IRQs were unmasked, we don't know
> whether the trigger was a pNMI/IRQ until we read from the GIC, and
> since we *could* have taken an IRQ, this is equivalent to taking a
> spurious IRQ, and while handling that, taking the NMI, e.g.
>
> < run with IRQs unmasked >
> ~~~ take IRQ ~~~
> < enter IRQ >
> ~~~ take NMI exception ~~~
> < enter NMI >
> < handle NMI >
> < exit NMI >
> ~~~ return from NMI exception ~~~
> < handle IRQ / spurious / do-nothing >
> < exit IRQ >
> ~~~ return from IRQ exception ~~~
> < continue running with IRQs unmasked >
>
Yes, here I am on the same page. (I think I used a wrong example in
previous email, which caused the confusion)

> ... except that we don't do the HW NMI exception entry/exit, just all
> the necessary SW accounting.
>
A little but important thing: local_irq_save() etc can not disable pNMI.

>
> Note that case (2) can *never* nest within itself or within case (1).
>
> Do you have a specific example of something that goes wrong with the
> above? e.g. something that's inconsistent with that rationale?
>
Please see the following comment.

> > And this may cause the following issue:
> > > There is 3 pieces of code put under the
> > > protection of if (!in_nmi()). At least the last one
> > > "rcu_irq_enter_check_tick()" can trigger a hard lock up bug. Because it
> > > is supposed to hold a spin lock with irqoff by
> > > "raw_spin_lock_rcu_node(rdp->mynode)", but pNMI can breach it. The same
> > > scenario in rcu_nmi_exit()->rcu_prepare_for_idle().

Sorry that this should be an wrong example, since here it takes the case (1).

Concentrating on the spin lock rcu_node->lock, there are two operators:
raw_spin_lock_rcu_node()
raw_spin_trylock_rcu_node()

Then suppose the scenario for deadlock:
note_gp_changes() in non-irq-context
{
local_irq_save(flags);
...
raw_spin_trylock_rcu_node(rnp) // hold lock
needwake = __note_gp_changes(rnp, rdp); ------\
raw_spin_unlock_irqrestore_rcu_node(rnp, flags); \
} \
\---> pNMI breaks in due to local_irq_save() can not disable it.
rcu_irq_enter() without __preempt_count_add(NMI_OFFSET + HARDIRQ_OFFSET)
->rcu_nmi_enter()
{
else if (!in_nmi())
rcu_irq_enter_check_tick()
->__rcu_irq_enter_check_tick()
{
...
raw_spin_lock_rcu_node(rdp->mynode);

//Oops deadlock!
}
}


> > >
> > > As for the first two "if (!in_nmi())", I have no idea of why, except
> > > breaching spin_lock_irq() by NMI. Hope Paul can give some guide.
>
> That code (in enter_from_kernel_mode()) only runs in case 2, where it
> cannot be nested within a pNMI, so I struggle to see how this can
> deadlock. It it can, then I would expect the general case of a pNMI
> nesting within and IRQ would be broken?
>
Sorry again for the previous misleading wrong example. Hope my new
example can help.

> Can you give a concrete example of a sequence that would lockup?
> Currently I can't see how that's possible.
>

It seems the RCU subsystem has a strict semantic on NMI and normal
interrupt. Besides the deadlock example, there may be other supprise to
confront with (will trace it on another mail with Paul)

Thanks,

Pingfan