Re: [PATCH for-next 3/6] xen/sched: Fix build when NR_CPUS == 1

[Connor Davis <connojdavis@gmail.com>]

On Fri, Feb 26, 2021 at 09:31:02AM +0100, Jan Beulich wrote:
> On 26.02.2021 04:08, Connor Davis wrote:
> > On Thu, Feb 25, 2021 at 04:50:02PM +0100, Jan Beulich wrote:
> >> On 25.02.2021 16:24, Connor Davis wrote:
> >>> Return from cpu_schedule_up when either cpu is 0 or
> >>> NR_CPUS == 1. This fixes the following:
> >>>
> >>> core.c: In function 'cpu_schedule_up':
> >>> core.c:2769:19: error: array subscript 1 is above array bounds
> >>> of 'struct vcpu *[1]' [-Werror=array-bounds]
> >>>  2769 |     if ( idle_vcpu[cpu] == NULL )
> >>>       |
> >>>
> 
> Ah yes, at -O2 I can observe the warning on e.g.
> 
> extern int array[N];
> 
> int test(unsigned i) {
> 	if(i == N - 1)
> 		return 0;
> 	return array[i];
> }
> 
> when N=1. No warning appears when N=2 or higher, yet if it is
> sensible to emit for N=1 then it would imo be similarly
> sensible to emit in other cases. The only difference is that
> when N=1, there's no i for which the array access would ever
> be valid, while e.g. for N=2 there's exactly one such i.
> 
> I've tried an x86 build with NR_CPUS=1, and this hits the case
> you found and a 2nd one, where behavior is even more puzzling.
> For the case you've found I'd like to suggest as alternative
> 
> @@ -2769,6 +2769,12 @@ static int cpu_schedule_up(unsigned int
>      if ( cpu == 0 )
>          return 0;
>  
> +    /*
> +     * Guard in particular also against the compiler suspecting out-of-bounds
> +     * array accesses below when NR_CPUS=1.
> +     */
> +    BUG_ON(cpu >= NR_CPUS);
> +

Yeah I like this better than my approach.

>      if ( idle_vcpu[cpu] == NULL )
>          vcpu_create(idle_vcpu[0]->domain, cpu);
>      else
> 
> To fix the x86 build in this regard we'd additionally need
> something along the lines of
> 
> --- unstable.orig/xen/arch/x86/genapic/x2apic.c
> +++ unstable/xen/arch/x86/genapic/x2apic.c
> @@ -54,7 +54,17 @@ static void init_apic_ldr_x2apic_cluster
>      per_cpu(cluster_cpus, this_cpu) = cluster_cpus_spare;
>      for_each_online_cpu ( cpu )
>      {
> -        if (this_cpu == cpu || x2apic_cluster(this_cpu) != x2apic_cluster(cpu))
> +        if ( this_cpu == cpu )
> +            continue;
> +        /*
> +         * Guard in particular against the compiler suspecting out-of-bounds
> +         * array accesses below when NR_CPUS=1 (oddly enough with gcc 10 it
> +         * is the 1st of these alone which actually helps, not the 2nd, nor
> +         * are both required together there).
> +         */
> +        BUG_ON(this_cpu >= NR_CPUS);
> +        BUG_ON(cpu >= NR_CPUS);
> +        if ( x2apic_cluster(this_cpu) != x2apic_cluster(cpu) )
>              continue;
>          per_cpu(cluster_cpus, this_cpu) = per_cpu(cluster_cpus, cpu);
>          break;
> 
> but the comment points out how strangely the compiler behaves here.
> Even flipping around the two sides of the != doesn't change its
> behavior. It is perhaps relevant to note here that there's no
> special casing of smp_processor_id() in the NR_CPUS=1 case, so the
> compiler can't infer this_cpu == 0.
> 
> Once we've settled on how to change common/sched/core.c I guess
> I'll then adjust the x86-specific change accordingly and submit as
> a separate fix (or I could of course also bundle both changes then).

Feel free to bundle both.

    Connor