Re: [PATCH v7 11/16] irqchip/gic-v3: Add support for ACPI's disabled but 'online capable' CPUs

[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

 



On Thu, 25 Apr 2024 13:31:50 +0100
Jonathan Cameron <Jonathan.Cameron@xxxxxxxxxx> wrote:

> On Wed, 24 Apr 2024 16:33:22 +0100
> Marc Zyngier <maz@xxxxxxxxxx> wrote:
> 
> > On Wed, 24 Apr 2024 13:54:38 +0100,
> > Jonathan Cameron <Jonathan.Cameron@xxxxxxxxxx> wrote:  
> > > 
> > > On Tue, 23 Apr 2024 13:01:21 +0100
> > > Marc Zyngier <maz@xxxxxxxxxx> wrote:
> > >     
> > > > On Mon, 22 Apr 2024 11:40:20 +0100,
> > > > Jonathan Cameron <Jonathan.Cameron@xxxxxxxxxx> wrote:    
> > > > > 
> > > > > On Thu, 18 Apr 2024 14:54:07 +0100
> > > > > Jonathan Cameron <Jonathan.Cameron@xxxxxxxxxx> wrote:    
> > 
> > [...]
> >   
> > > > >       
> > > > > > +	/*
> > > > > > +	 * Capable but disabled CPUs can be brought online later. What about
> > > > > > +	 * the redistributor? ACPI doesn't want to say!
> > > > > > +	 * Virtual hotplug systems can use the MADT's "always-on" GICR entries.
> > > > > > +	 * Otherwise, prevent such CPUs from being brought online.
> > > > > > +	 */
> > > > > > +	if (!(gicc->flags & ACPI_MADT_ENABLED)) {
> > > > > > +		pr_warn_once("CPU %u's redistributor is inaccessible: this CPU can't be brought online\n", cpu);
> > > > > > +		set_cpu_present(cpu, false);
> > > > > > +		set_cpu_possible(cpu, false);
> > > > > > +		return 0;
> > > > > > +	}      
> > > > 
> > > > It seems dangerous to clear those this late in the game, given how
> > > > disconnected from the architecture code this is. Are we sure that
> > > > nothing has sampled these cpumasks beforehand?    
> > > 
> > > Hi Marc,
> > > 
> > > Any firmware that does this is being considered as buggy already
> > > but given it is firmware and the spec doesn't say much about this,
> > > there is always the possibility.    
> > 
> > There is no shortage of broken firmware out there, and I expect this
> > trend to progress.
> >   
> > > Not much happens between the point where these are setup and
> > > the point where the the gic inits and this code runs, but even if careful
> > > review showed it was fine today, it will be fragile to future changes.
> > > 
> > > I'm not sure there is a huge disadvantage for such broken firmware in
> > > clearing these masks from the point of view of what is used throughout
> > > the rest of the kernel. Here I think we are just looking to prevent the CPU
> > > being onlined later.    
> > 
> > I totally agree on the goal, I simply question the way you get to it.
> >   
> > > 
> > > We could add a set_cpu_broken() with appropriate mask.
> > > Given this is very arm64 specific I'm not sure Rafael will be keen on
> > > us checking such a mask in the generic ACPI code, but we could check it in
> > > arch_register_cpu() and just not register the cpu if it matches.
> > > That will cover the vCPU hotplug case.
> > >
> > > Does that sounds sensible, or would you prefer something else?    
> > 
> > 
> > Such a 'broken_rdists' mask is exactly what I have in mind, just
> > keeping it private to the GIC driver, and not expose it anywhere else.
> > You can then fail the hotplug event early, and avoid changing the
> > global masks from within the GIC driver. At least, we don't mess with
> > the internals of the kernel, and the CPU is properly marked as dead
> > (that mechanism should already work).
> > 
> > I'd expect the handling side to look like this (will not compile, but
> > you'll get the idea):  
> Hi Marc,
> 
> In general this looks good - but...
> 
> I haven't gotten to the bottom of why yet (and it might be a side
> effect of how I hacked the test by lying in minimal fashion and
> just frigging the MADT read functions) but the hotplug flow is only getting
> as far as calling __cpu_up() before it seems to enter an infinite loop.
> That is it never gets far enough to fail this test.
> 
> Getting stuck in a psci cpu_on call.  I'm guessing something that
> we didn't get to in the earlier gicv3 calls before bailing out is blocking that?
> Looks like it gets to
> SMCCC smc
> and is never seen again.
> 
> Any ideas on where to look?  The one advantage so far of the higher level
> approach is we never tried the hotplug callbacks at all so avoided hitting
> that call.  One (little bit horrible) solution that might avoid this would 
> be to add another cpuhp state very early on and fail at that stage.
> I'm not keen on doing that without a better explanation than I have so far!

Whilst it still doesn't work I suspect I'm loosing ability to print to the console
between that point and somewhat later and real problem is elsewhere.

Jonathan

> 
> Thanks,
> 
> J
> 
>  
> > diff --git a/drivers/irqchip/irq-gic-v3.c b/drivers/irqchip/irq-gic-v3.c
> > index 6fb276504bcc..e8f02bfd0e21 100644
> > --- a/drivers/irqchip/irq-gic-v3.c
> > +++ b/drivers/irqchip/irq-gic-v3.c
> > @@ -1009,6 +1009,9 @@ static int __gic_populate_rdist(struct redist_region *region, void __iomem *ptr)
> >  	u64 typer;
> >  	u32 aff;
> >  
> > +	if (cpumask_test_cpu(smp_processor_id(), &broken_rdists))
> > +		return 1;
> > +
> >  	/*
> >  	 * Convert affinity to a 32bit value that can be matched to
> >  	 * GICR_TYPER bits [63:32].
> > @@ -1260,14 +1263,15 @@ static int gic_dist_supports_lpis(void)
> >  		!gicv3_nolpi);
> >  }
> >  
> > -static void gic_cpu_init(void)
> > +static int gic_cpu_init(void)
> >  {
> >  	void __iomem *rbase;
> > -	int i;
> > +	int ret, i;
> >  
> >  	/* Register ourselves with the rest of the world */
> > -	if (gic_populate_rdist())
> > -		return;
> > +	ret = gic_populate_rdist();
> > +	if (ret)
> > +		return ret;
> >  
> >  	gic_enable_redist(true);
> >  
> > @@ -1286,6 +1290,8 @@ static void gic_cpu_init(void)
> >  
> >  	/* initialise system registers */
> >  	gic_cpu_sys_reg_init();
> > +
> > +	return 0;
> >  }
> >  
> >  #ifdef CONFIG_SMP
> > @@ -1295,7 +1301,11 @@ static void gic_cpu_init(void)
> >  
> >  static int gic_starting_cpu(unsigned int cpu)
> >  {
> > -	gic_cpu_init();
> > +	int ret;
> > +
> > +	ret = gic_cpu_init();
> > +	if (ret)
> > +		return ret;
> >  
> >  	if (gic_dist_supports_lpis())
> >  		its_cpu_init();
> > 
> > But the question is: do you rely on these masks having been
> > "corrected" anywhere else?
> > 
> > Thanks,
> > 
> > 	M.
> >   
> 
> 
> _______________________________________________
> linux-arm-kernel mailing list
> linux-arm-kernel@xxxxxxxxxxxxxxxxxxx
> http://lists.infradead.org/mailman/listinfo/linux-arm-kernel





[Index of Archives]     [Linux Kernel]     [Kernel Newbies]     [x86 Platform Driver]     [Netdev]     [Linux Wireless]     [Netfilter]     [Bugtraq]     [Linux Filesystems]     [Yosemite Discussion]     [MIPS Linux]     [ARM Linux]     [Linux Security]     [Linux RAID]     [Samba]     [Device Mapper]

  Powered by Linux