Re: [PATCH 2/4] KVM: x86/mmu: Document the "rules" for using host_pfn_mapping_level()

[Mingwei Zhang]

On Fri, Jul 15, 2022, Sean Christopherson wrote:
> Add a comment to document how host_pfn_mapping_level() can be used safely,
> as the line between safe and dangerous is quite thin.  E.g. if KVM were
> to ever support in-place promotion to create huge pages, consuming the
> level is safe if the caller holds mmu_lock and checks that there's an
> existing _leaf_ SPTE, but unsafe if the caller only checks that there's a
> non-leaf SPTE.
> 
> Opportunistically tweak the existing comments to explicitly document why
> KVM needs to use READ_ONCE().
> 
> No functional change intended.
> 
> Signed-off-by: Sean Christopherson <seanjc@xxxxxxxxxx>
> ---
>  arch/x86/kvm/mmu/mmu.c | 42 +++++++++++++++++++++++++++++++++++-------
>  1 file changed, 35 insertions(+), 7 deletions(-)
> 
> diff --git a/arch/x86/kvm/mmu/mmu.c b/arch/x86/kvm/mmu/mmu.c
> index bebff1d5acd4..d5b644f3e003 100644
> --- a/arch/x86/kvm/mmu/mmu.c
> +++ b/arch/x86/kvm/mmu/mmu.c
> @@ -2919,6 +2919,31 @@ static void direct_pte_prefetch(struct kvm_vcpu *vcpu, u64 *sptep)
>  	__direct_pte_prefetch(vcpu, sp, sptep);
>  }
>  
> +/*
> + * Lookup the mapping level for @gfn in the current mm.
> + *
> + * WARNING!  Use of host_pfn_mapping_level() requires the caller and the end
> + * consumer to be tied into KVM's handlers for MMU notifier events!
Since calling this function won't cause kernel crash now, I guess we can
remove the warning sign here, but keep the remaining statement since it
is necessary.
> + *
> + * There are several ways to safely use this helper:
> + *
> + * - Check mmu_notifier_retry_hva() after grabbing the mapping level, before
> + *   consuming it.  In this case, mmu_lock doesn't need to be held during the
> + *   lookup, but it does need to be held while checking the MMU notifier.

but it does need to be held while checking the MMU notifier and
consuming the result.
> + *
> + * - Hold mmu_lock AND ensure there is no in-progress MMU notifier invalidation
> + *   event for the hva.  This can be done by explicit checking the MMU notifier
> + *   or by ensuring that KVM already has a valid mapping that covers the hva.

Yes, more specifically, "mmu notifier sequence counter".
> + *
> + * - Do not use the result to install new mappings, e.g. use the host mapping
> + *   level only to decide whether or not to zap an entry.  In this case, it's
> + *   not required to hold mmu_lock (though it's highly likely the caller will
> + *   want to hold mmu_lock anyways, e.g. to modify SPTEs).
> + *
> + * Note!  The lookup can still race with modifications to host page tables, but
> + * the above "rules" ensure KVM will not _consume_ the result of the walk if a
> + * race with the primary MMU occurs.
> + */
>  static int host_pfn_mapping_level(struct kvm *kvm, gfn_t gfn,
>  				  const struct kvm_memory_slot *slot)
>  {
> @@ -2941,16 +2966,19 @@ static int host_pfn_mapping_level(struct kvm *kvm, gfn_t gfn,
>  	hva = __gfn_to_hva_memslot(slot, gfn);
>  
>  	/*
> -	 * Lookup the mapping level in the current mm.  The information
> -	 * may become stale soon, but it is safe to use as long as
> -	 * 1) mmu_notifier_retry was checked after taking mmu_lock, and
> -	 * 2) mmu_lock is taken now.
> -	 *
> -	 * We still need to disable IRQs to prevent concurrent tear down
> -	 * of page tables.
> +	 * Disable IRQs to prevent concurrent tear down of host page tables,
> +	 * e.g. if the primary MMU promotes a P*D to a huge page and then frees
> +	 * the original page table.
>  	 */
>  	local_irq_save(flags);
>  
> +	/*
> +	 * Read each entry once.  As above, a non-leaf entry can be promoted to
> +	 * a huge page _during_ this walk.  Re-reading the entry could send the
> +	 * walk into the weeks, e.g. p*d_large() returns false (sees the old
> +	 * value) and then p*d_offset() walks into the target huge page instead
> +	 * of the old page table (sees the new value).
> +	 */
>  	pgd = READ_ONCE(*pgd_offset(kvm->mm, hva));
>  	if (pgd_none(pgd))
>  		goto out;
> -- 
> 2.37.0.170.g444d1eabd0-goog
>