RE: [PATCH] x86/hyper-v: guard against cpu mask changes in hyperv_flush_tlb_others()

[Michael Kelley <mikelley@microsoft.com>]

From: תומר אבוטבול <tomer432100@gmail.com>  Sent: Friday, August 6, 2021 11:03 AM

> Attaching the patches Michael asked for debugging 
> 1) Print the cpumask when < num_possible_cpus():
> diff --git a/arch/x86/hyperv/mmu.c b/arch/x86/hyperv/mmu.c
> index e666f7eaf32d..620f656d6195 100644
> --- a/arch/x86/hyperv/mmu.c
> +++ b/arch/x86/hyperv/mmu.c
> @@ -60,6 +60,7 @@ static void hyperv_flush_tlb_others(const struct cpumask *cpus,
>         struct hv_tlb_flush *flush;
>         u64 status = U64_MAX;
>         unsigned long flags;
> +       unsigned int cpu_last;
>
>        trace_hyperv_mmu_flush_tlb_others(cpus, info);
>
> @@ -68,6 +69,11 @@ static void hyperv_flush_tlb_others(const struct cpumask *cpus,
>
>        local_irq_save(flags);
>
> +       cpu_last = cpumask_last(cpus);
> +       if (cpu_last > num_possible_cpus()) {

I think this should be ">=" since cpus are numbered starting at zero.
In your VM with 64 CPUs, having CPU #64 in the list would be error.

> +               pr_emerg("ERROR_HYPERV: cpu_last=%*pbl", cpumask_pr_args(cpus));
> +       }
> +
>        /*
>         * Only check the mask _after_ interrupt has been disabled to avoid the
>         * mask changing under our feet.
>
> 2) disable the Hyper-V specific flush routines:
> diff --git a/arch/x86/hyperv/mmu.c b/arch/x86/hyperv/mmu.c
> index e666f7eaf32d..8e77cc84775a 100644
> --- a/arch/x86/hyperv/mmu.c
> +++ b/arch/x86/hyperv/mmu.c
> @@ -235,6 +235,7 @@ static u64 hyperv_flush_tlb_others_ex(const struct cpumask *cpus,
>
> void hyperv_setup_mmu_ops(void)
> {
> +  return;
>        if (!(ms_hyperv.hints & HV_X64_REMOTE_TLB_FLUSH_RECOMMENDED))
>                return;

Otherwise, this code looks good to me and matches what I had in mind.

Note that the function native_flush_tlb_others() is used when the Hyper-V specific
flush function is disabled per patch #2 above, or when hv_cpu_to_vp_index() returns
VP_INVALID.  In a quick glance through the code, it appears that native_flush_tlb_others()
will work even if there's a non-existent CPU in the cpumask that is passed as an argument.
So perhaps an immediate workaround is Patch #2 above.  

Perhaps hyperv_flush_tlb_others() should be made equally tolerant of a non-existent
CPU being in the list. But if you are willing, I'm still interested in the results of an
experiment with just Patch #1.  I'm curious about what the CPU list looks like when
it has a non-existent CPU.  Is it complete garbage, or is there just one non-existent
CPU?

The other curiosity is that I haven't seen this Linux panic reported by other users,
and I think it would have come to our attention if it were happening with any frequency.
You see the problem fairly regularly.  So I'm wondering what the difference is.

Michael