Re: Query regarding pseudo nmi support on GIC V3 and request_nmi()

[Marc Zyngier]

On Fri, 08 May 2020 14:34:10 +0100,
Neeraj Upadhyay <neeraju@xxxxxxxxxxxxxx> wrote:
> 
> Hi Marc,
> 
> On 5/8/2020 6:23 PM, Marc Zyngier wrote:
> > On Fri, 8 May 2020 18:09:00 +0530
> > Neeraj Upadhyay <neeraju@xxxxxxxxxxxxxx> wrote:
> > 
> >> Hi Marc,
> >> 
> >> On 5/8/2020 5:57 PM, Marc Zyngier wrote:
> >>> On Fri, 8 May 2020 16:36:42 +0530
> >>> Neeraj Upadhyay <neeraju@xxxxxxxxxxxxxx> wrote:
> >>>    
> >>>> Hi Marc,
> >>>> 
> >>>> On 5/8/2020 4:15 PM, Marc Zyngier wrote:
> >>>>> On Thu, 07 May 2020 17:06:19 +0100,
> >>>>> Neeraj Upadhyay <neeraju@xxxxxxxxxxxxxx> wrote:
> >>>>>> 
> >>>>>> Hi,
> >>>>>> 
> >>>>>> I have one query regarding pseudo NMI support on GIC v3; from what I
> >>>>>> could understand, GIC v3 supports pseudo NMI setup for SPIs and PPIs.
> >>>>>> However the request_nmi() in irq framework requires NMI to be per cpu
> >>>>>> interrupt source (it checks for IRQF_PERCPU). Can you please help
> >>>>>> understand this part, how SPIs can be configured as NMIs, if there is
> >>>>>> a per cpu interrupt source restriction?
> >>>>> 
> >>>>> Let me answer your question by another question: what is the semantic
> >>>>> of a NMI if you can't associate it with a particular CPU?
> >>>>>    >>
> >>>> I was actually thinking of a use case, where, we have a watchdog
> >>>> interrupt (which is a SPI), which is used for detecting software
> >>>> hangs and cause device reset; If that interrupt's current cpu
> >>>> affinity is on a core, where interrupts are disabled, we won't be
> >>>> able to serve it; so, we need to group that interrupt as an fiq;
> >>> 
> >>> Linux doesn't use Group-0 interrupts, as they are strictly secure
> >>> (unless your SoC doesn't have EL3, which I doubt).
> >> 
> >> Yes, we handle that watchdog interrupt as a Group-0 interrupt, which
> >> is handled as fiq in EL3.
> >> 
> >>>    
> >>>> I was thinking, if its feasible to mark that interrupt as pseudo
> >>>> NMI and route it to EL1 as irq. However, looks like that is not the
> >>>> semantic of a NMI and we would need something like pseudo NMI ipi
> >>>> for this.
> >>> 
> >>> Sending a NMI IPI from another NMI handler? Even once I've added
> >>> these, there is no way this will work for that particular scenario.
> >>> Just look at the restrictions we impose on NMIs.
> >>>    
> >> 
> >> Sending a pseudo NMI IPI (to EL1) from fiq handler (which runs in
> >> EL3); I will check, but do you think, that might not work?
> > 
> > How do you know, from EL3, what to write in memory so that the NMI
> > handler knows what you want to do? Are you going to parse the S1 page
> > tables? Hard-code the behaviour of some random Linux version in your
> > legendary non-updatable firmware? This isn't an acceptable behaviour.
> > 
> 
> Ok, I understand;
> 
> Initial thought was to use watchdog SPI as pseudo NMI; however, as
> pseudo NMIs are only per CPU sources, we were exploring the
> possibility of using an unused ipi (using the work which is done in
> [1] and [2]  for SGIs) as pseudo NMI, which EL3 sends to EL1, on
> receiving watchdog fiq. The pseudo NMI handler would collect required
> debug information, to help indentify the lockup cause. We weren't
> thinking of communicating any information from EL3 fiq handler to
> EL1.

What if the operating system running at EL1/EL2 is *not* Linux?

> 
> However, from this discussion, I realize that calling irq handler from
> fiq handler, would not be possible. So, the approach looks flawed.
> 
> I believe, allowing a non-percpu pseudo NMI is not acceptable to community?

No, I really don't want to entertain this idea, because the semantics
are way too loosely defined and you'd end up with everyone wanting
something mildly different.

> > An IPI is between two CPUs used by the same SW entitiy. What runs at
> > EL3 is completely alien to Linux, and so is Linux to EL3. If you want
> > to IPI, send Group-0 IPIs that are private to the firmware.
> > 
> 
> Ok got it; however, I wonder what's the use case of sending
> SGI to EL1, from secure world, using ICC_ASGI1R. I thought it
> allowed communication between EL1 and EL3; but, looks like I
> understood in wrong.

There is what the GIC architecture can do, and there is what is
sensible for Linux. The GIC allows IPIs from S-to-NS as well as the
opposite. This doesn't make it a good idea (it actually is a terrible
idea, and I really hope that future versions of the architecture will
simply kill the feature).

The idea was that you'd make SGIs an first class ABI between S and
NS. Given that the two are developed separately and that nobody ever
standardised what the SGI numbers mean, this idea is completely dead.

> 
> > If you want to inject NMI-type exceptions into EL1, you can always try
> > SDEI (did I actually write this? Help!). But given your use case below,
> > that wouldn't work either.
> > 
> 
> Ok.
> 
> >>> Frankly, if all you need to do is to reset the SoC, use EL3
> >>> firmware. That is what it is for.
> >>>    
> >> 
> >> Before triggering SoC reset, we want to collect certain  EL1 debug
> >> information like stack trace for CPUs and other debug information.
> > 
> > Frankly, if you are going to reset the SoC because EL1/EL2 has gone
> > bust, how can you trust it to do anything sensible when injecting an
> > interrupt?. Once you take a SPI at EL3, gather whatever state you want
> > from EL3. Do not involve EL1 at all.
> > 
> > 	M.
> > 
> 
> Agree that it might not work for all cases. But, for the cases like,
> some kernel code is stuck after disabling local irqs; pseudo NMI might
> still be able to run and capture stack and other debug info, to help
> detect the cause of lockups.

And for that we'll have pseudo-NMI IPIs, initiated from the kernel
itself as part of the normal debugging infrastructure. It is the EL3
initiated IPI to EL1 that I strongly oppose against. Not to mention
that if the kernel locks up with PSTATE.I set (which still happens on
exception entry), the pseudo-NMI won't work either.

As I said, you only have two options: either implement everything in
EL3 (and the NS OS doesn't need to know anything at all), or use SDEI
as the architected way to inject an exception into the NS world (and
Linux already supports it).

	M.

-- 
Without deviation from the norm, progress is not possible.