Re: [PATCH v4 12/23] KVM: Destroy memslots immediately after mmu_notifiers are unregistered

[Ritesh Harjani (IBM) <ritesh.list@gmail.com>]

Sean Christopherson <seanjc@google.com> writes:

> +Gautam and PPC lists/maintainers
>
> On Fri, Oct 30, 2026, sashiko-bot@kernel.org wrote:
>> Thank you for your contribution! Sashiko AI review found 1 potential issue(s) to consider:
>> - [Medium] KVM: PPC: Book3S HV: rmap_nested memory leak in kvmhv_release_all_nested()
>> --
>> 
>> commit 9c25a1b4a76e8526eb8a2b11e1b9e63893b0de80
>> Author: Sean Christopherson <seanjc@google.com>
>> 
>> KVM: Destroy memslots immediately after mmu_notifiers are unregistered
>> 
>> This commit replaces active memslots with an empty memslot set and frees all
>> memslots immediately after unregistering the MMU notifier in kvm_destroy_vm(),
>> prior to calling kvm_arch_destroy_vm().
>> 
>> > diff --git a/virt/kvm/kvm_main.c b/virt/kvm/kvm_main.c
>> > index fa548b12228ff..2eda12a52bf4b 100644
>> > --- a/virt/kvm/kvm_main.c
>> > +++ b/virt/kvm/kvm_main.c
>> 
>> [ ... ]
>> 
>> > @@ -1310,12 +1329,10 @@ static void kvm_destroy_vm(struct kvm *kvm)
>> >  		kvm->mn_active_invalidate_count = 0;
>> >  	else
>> >  		WARN_ON(kvm->mmu_invalidate_in_progress);
>> > +	kvm_destroy_memslots(kvm);
>> > +
>> >  	kvm_arch_destroy_vm(kvm);
>> >  	kvm_destroy_devices(kvm);
>> > -	for (i = 0; i < kvm_arch_nr_memslot_as_ids(kvm); i++) {
>> > -		kvm_free_memslots(kvm, &kvm->__memslots[i][0]);
>> > -		kvm_free_memslots(kvm, &kvm->__memslots[i][1]);
>> > -	}
>> 
>> [Severity: Medium]
>> Does hoisting kvm_destroy_memslots() above kvm_arch_destroy_vm() leak
>> struct rmap_nested allocations in slot->arch.rmap on PowerPC Book3S HV when
>> destroying a VM with nested virtualization enabled?
>
> I don't think so?  Patch 3[*], "KVM: PPC: Flush/zap all memslots on kvm_arch_flush_shadow_all()",
> should free the rmaps in response to kvm_arch_flush_shadow_all().   From what I
> can tell, the page tables have to be in radix mode, and so the rmaps are freed via:
>
>   kvmppc_radix_flush_memslot()
>   |
>   -> kvmppc_unmap_pte()
>      |
>      -> kvmhv_remove_nest_rmap_range()
>         |
>         -> kvmhv_remove_nest_rmap_list()
>

That's correct. Maybe, I think Sashiko didn't consider patch-3 of the series
during review of this patch-12.

> And if that doesn't work, it means memslot deletion on an active VM is completely
> broken.
>
> However, it means the memslot code in kvmhv_release_all_nested() is
> dead code,

Not really. I do see a caller of this, for when we switch from radix
to HPT. So we still need it.

INCOMING CALLS
- f kvmhv_release_all_nested
  - f kvmppc_switch_mmu_to_hpt    [arch/powerpc/kvm/book3s_hv.c:5591]
    + f kvmppc_alloc_reset_hpt    [arch/powerpc/kvm/book3s_64_mmu_hv.c:144]
    + f kvmhv_configure_mmu       [arch/powerpc/kvm/book3s_hv.c:6330]
  + f kvmppc_core_destroy_vm_hv   [arch/powerpc/kvm/book3s_hv.c:5930]



So FWIW patch-12 as such looks good to me from PowerPC perspective,
since those rmap entries will get freed up as part of the
mmu_notifier_unregister() (after patch-3). Agreed that Sashiko
comment does not look correct.

Thanks for looking into this.

-ritesh


> so this as follow-up to patch 3?  Or maybe folded in?
>
> diff --git a/arch/powerpc/kvm/book3s_hv_nested.c b/arch/powerpc/kvm/book3s_hv_nested.c
> index a6ff42d7666c..27985210f972 100644
> --- a/arch/powerpc/kvm/book3s_hv_nested.c
> +++ b/arch/powerpc/kvm/book3s_hv_nested.c
> @@ -802,8 +802,6 @@ void kvmhv_release_all_nested(struct kvm *kvm)
>  	int lpid;
>  	struct kvm_nested_guest *gp;
>  	struct kvm_nested_guest *freelist = NULL;
> -	struct kvm_memory_slot *memslot;
> -	int srcu_idx, bkt;
>  
>  	spin_lock(&kvm->mmu_lock);
>  	idr_for_each_entry(&kvm->arch.kvm_nested_guest_idr, gp, lpid) {
> @@ -820,11 +818,6 @@ void kvmhv_release_all_nested(struct kvm *kvm)
>  		freelist = gp->next;
>  		kvmhv_release_nested(gp);
>  	}
> -
> -	srcu_idx = srcu_read_lock(&kvm->srcu);
> -	kvm_for_each_memslot(memslot, bkt, kvm_memslots(kvm))
> -		kvmhv_free_memslot_nest_rmap(memslot);
> -	srcu_read_unlock(&kvm->srcu, srcu_idx);
>  }
>  
>  /* caller must hold gp->tlb_lock */
> 			
>
> [*] https://lore.kernel.org/all/20261030212112.2598640-4-seanjc@google.com
>
>> In virt/kvm/kvm_main.c, kvm_destroy_memslots() replaces kvm->memslots[i]
>> with kvm_empty_memslots and frees each memslot via
>> kvm_free_memslot() -> kvm_arch_free_memslot() -> kvmppc_core_free_memslot_hv()
>> in arch/powerpc/kvm/book3s_hv.c:
>> 
>>     static void kvmppc_core_free_memslot_hv(struct kvm_memory_slot *slot)
>>     {
>>             vfree(slot->arch.rmap);
>>             slot->arch.rmap = NULL;
>>     }
>> 
>> which frees the slot->arch.rmap array without freeing the linked
>> struct rmap_nested entries.
>> 
>> Previously, those entries were freed before memslot destruction via
>> kvm_arch_destroy_vm() -> kvmppc_core_destroy_vm_hv() ->
>> kvmhv_release_all_nested() in arch/powerpc/kvm/book3s_hv_nested.c:
>> 
>>     srcu_idx = srcu_read_lock(&kvm->srcu);
>>     kvm_for_each_memslot(memslot, bkt, kvm_memslots(kvm))
>>             kvmhv_free_memslot_nest_rmap(memslot);
>>     srcu_read_unlock(&kvm->srcu, srcu_idx);
>> 
>> With kvm_destroy_memslots() running before kvm_arch_destroy_vm(),
>> kvm_memslots(kvm) now returns kvm_empty_memslots and the memslots have
>> already been freed when kvmhv_release_all_nested() runs, so
>> kvmhv_free_memslot_nest_rmap() is never called. Any struct rmap_nested
>> entries remaining in slot->arch.rmap at teardown (for example, when
>> KVMPPC_SECURE_INIT_DONE causes kvmppc_radix_flush_memslot() to return early,
>> or when 4KB PTEs were replaced by a huge page in kvmppc_unmap_free_pte())
>> will be leaked.
>> 
>> -- 
>> Sashiko AI review · https://sashiko.dev/#/patchset/20261030212112.2598640-1-seanjc@google.com?part=12
>>