[PATCH] KVM: prevent read from desc->shadow_ptes[-1] in rmap_desc_remove_entry()

[PATCH] KVM: prevent read from desc->shadow_ptes[-1] in rmap_desc_remove_entry()

[Roel Kluin]

prevent read from desc->shadow_ptes[-1]

Signed-off-by: Roel Kluin <roel.kluin@gmail.com>
---
If in rmap_remove() (bottom) we do:

while (desc) {
        for (i = 0; i < RMAP_EXT && desc->shadow_ptes[i]; ++i)
                if (desc->shadow_ptes[i] == spte) {
                        rmap_desc_remove_entry(rmapp,
                                               desc, i,
                                               prev_desc);
                        return;
                }
        prev_desc = desc;
        desc = desc->more;
}

If in the first iteration esc->shadow_ptes[0] == spte, then we
call rmap_desc_remove_entry() with i == 0, and then we read in
the last iteration from desc->shadow_ptes[-1].

I found this by code analysis.

diff --git a/arch/x86/kvm/mmu.c b/arch/x86/kvm/mmu.c
index 0ef5bb2..e1b2e46 100644
--- a/arch/x86/kvm/mmu.c
+++ b/arch/x86/kvm/mmu.c
@@ -541,7 +541,7 @@ static void rmap_desc_remove_entry(unsigned long *rmapp,
 {
 	int j;
 
-	for (j = RMAP_EXT - 1; !desc->shadow_ptes[j] && j > i; --j)
+	for (j = RMAP_EXT - 1; j > i && !desc->shadow_ptes[j]; --j)
 		;
 	desc->shadow_ptes[i] = desc->shadow_ptes[j];
 	desc->shadow_ptes[j] = NULL;

Re: [PATCH] KVM: prevent read from desc->shadow_ptes[-1] in rmap_desc_remove_entry()

[Avi Kivity]

On 08/29/2009 03:13 PM, Roel Kluin wrote:
> prevent read from desc->shadow_ptes[-1]
>
> Signed-off-by: Roel Kluin<roel.kluin@gmail.com>
> ---
> If in rmap_remove() (bottom) we do:
>
> while (desc) {
>          for (i = 0; i<  RMAP_EXT&&  desc->shadow_ptes[i]; ++i)
>                  if (desc->shadow_ptes[i] == spte) {
>                          rmap_desc_remove_entry(rmapp,
>                                                 desc, i,
>                                                 prev_desc);
>                          return;
>                  }
>          prev_desc = desc;
>          desc = desc->more;
> }
>
> If in the first iteration esc->shadow_ptes[0] == spte, then we
> call rmap_desc_remove_entry() with i == 0, and then we read in
> the last iteration from desc->shadow_ptes[-1].
>
> I found this by code analysis.
>
> diff --git a/arch/x86/kvm/mmu.c b/arch/x86/kvm/mmu.c
> index 0ef5bb2..e1b2e46 100644
> --- a/arch/x86/kvm/mmu.c
> +++ b/arch/x86/kvm/mmu.c
> @@ -541,7 +541,7 @@ static void rmap_desc_remove_entry(unsigned long *rmapp,
>   {
>   	int j;
>
> -	for (j = RMAP_EXT - 1; !desc->shadow_ptes[j]&&  j>  i; --j)
> +	for (j = RMAP_EXT - 1; j>  i&&  !desc->shadow_ptes[j]; --j)
>   		;
>   	desc->shadow_ptes[i] = desc->shadow_ptes[j];
>   	desc->shadow_ptes[j] = NULL;
>    

There is never a case when j < 0.  The loop finds the highest j that has 
a non-NULL shadow_pte[j], and that is guaranteed to exist.

I think the j > i test is redundant.  Not sure though.

-- 
I have a truly marvellous patch that fixes the bug which this
signature is too narrow to contain.