Re: [PATCH 1/7] KVM: Fix out of range accesses to memslots

[Peter Xu]

On Fri, Mar 20, 2020 at 03:40:41PM -0700, Sean Christopherson wrote:
> On Fri, Mar 20, 2020 at 06:17:08PM -0400, Peter Xu wrote:
> > On Fri, Mar 20, 2020 at 01:55:40PM -0700, Sean Christopherson wrote:
> > > Reset the LRU slot if it becomes invalid when deleting a memslot to fix
> > > an out-of-bounds/use-after-free access when searching through memslots.
> > > 
> > > Explicitly check for there being no used slots in search_memslots(), and
> > > in the caller of s390's approximation variant.
> > > 
> > > Fixes: 36947254e5f9 ("KVM: Dynamically size memslot array based on number of used slots")
> > > Reported-by: Qian Cai <cai@xxxxxx>
> > > Cc: Peter Xu <peterx@xxxxxxxxxx>
> > > Signed-off-by: Sean Christopherson <sean.j.christopherson@xxxxxxxxx>
> > > ---
> > >  arch/s390/kvm/kvm-s390.c | 3 +++
> > >  include/linux/kvm_host.h | 3 +++
> > >  virt/kvm/kvm_main.c      | 3 +++
> > >  3 files changed, 9 insertions(+)
> > > 
> > > diff --git a/arch/s390/kvm/kvm-s390.c b/arch/s390/kvm/kvm-s390.c
> > > index 807ed6d722dd..cb15fdda1fee 100644
> > > --- a/arch/s390/kvm/kvm-s390.c
> > > +++ b/arch/s390/kvm/kvm-s390.c
> > > @@ -2002,6 +2002,9 @@ static int kvm_s390_get_cmma(struct kvm *kvm, struct kvm_s390_cmma_log *args,
> > >  	struct kvm_memslots *slots = kvm_memslots(kvm);
> > >  	struct kvm_memory_slot *ms;
> > >  
> > > +	if (unlikely(!slots->used_slots))
> > > +		return 0;
> > > +
> > >  	cur_gfn = kvm_s390_next_dirty_cmma(slots, args->start_gfn);
> > >  	ms = gfn_to_memslot(kvm, cur_gfn);
> > >  	args->count = 0;
> > > diff --git a/include/linux/kvm_host.h b/include/linux/kvm_host.h
> > > index 35bc52e187a2..b19dee4ed7d9 100644
> > > --- a/include/linux/kvm_host.h
> > > +++ b/include/linux/kvm_host.h
> > > @@ -1032,6 +1032,9 @@ search_memslots(struct kvm_memslots *slots, gfn_t gfn)
> > >  	int slot = atomic_read(&slots->lru_slot);
> > >  	struct kvm_memory_slot *memslots = slots->memslots;
> > >  
> > > +	if (unlikely(!slots->used_slots))
> > > +		return NULL;
> > > +
> > >  	if (gfn >= memslots[slot].base_gfn &&
> > >  	    gfn < memslots[slot].base_gfn + memslots[slot].npages)
> > >  		return &memslots[slot];
> > > diff --git a/virt/kvm/kvm_main.c b/virt/kvm/kvm_main.c
> > > index 28eae681859f..f744bc603c53 100644
> > > --- a/virt/kvm/kvm_main.c
> > > +++ b/virt/kvm/kvm_main.c
> > > @@ -882,6 +882,9 @@ static inline void kvm_memslot_delete(struct kvm_memslots *slots,
> > >  
> > >  	slots->used_slots--;
> > >  
> > > +	if (atomic_read(&slots->lru_slot) >= slots->used_slots)
> > > +		atomic_set(&slots->lru_slot, 0);
> > 
> > Nit: could we drop the atomic ops?  The "slots" is still only used in
> > the current thread before the rcu assignment, so iirc it's fine (and
> > RCU assignment should have a mem barrier when needed anyway).
> 
> No, atomic_t wraps an int inside a struct to prevent direct usage, e.g.
> 
> virt/kvm/kvm_main.c: In function âkvm_memslot_deleteâ:
> virt/kvm/kvm_main.c:885:22: error: invalid operands to binary >= (have âatomic_t {aka struct <anonymous>}â and âintâ)
>   if (slots->lru_slot >= slots->used_slots)
>                       ^
> virt/kvm/kvm_main.c:886:19: error: incompatible types when assigning to type âatomic_t {aka struct <anonymous>}â from type âintâ
>    slots->lru_slot = 0;
> 
> 
> Buy yeah, the compiler barrier associated with atomic_read() isn't
> necessary.

Oops, sorry. I was obviously thinking about QEMU's atomic helpers.

> 
> > I thought resetting lru_slot to zero would be good enough when
> > duplicating the slots... however if we want to do finer grained reset,
> > maybe even better to reset also those invalidated ones (since we know
> > this information)?
> > 
> >    	if (slots->lru_slot >= slots->id_to_index[memslot->id])
> >    		slots->lru_slot = 0;
> 
> I'd prefer to go with the more obviously correct version.  This code will
> rarely trigger, e.g. larger slots have lower indices and are more likely to
> be the LRU (by virtue of being the biggest), and dynamic memslot deletion
> is usually for relatively small ranges that are being remapped by the guest.

IMHO the more obvious correct version could be unconditionally setting
lru_slot to something invalid (e.g. -1) in kvm_dup_memslots(), then in
the two search_memslots() skip the cache if lru_slot is invalid.
That's IMHO easier to understand than the "!slots->used_slots" checks.
But I've no strong opinion.

Thanks,

-- 
Peter Xu