[0/7] KVM: Fix memslot use-after-free bug

Message

Sean Christopherson March 20, 2020, 8:55 p.m. UTC

Fix a bug introduced by dynamic memslot allocation where the LRU slot can
become invalid and lead to a out-of-bounds/use-after-free scenario.

The patch is different that what Qian has already tested, but I was able
to reproduce the bad behavior by enhancing the set memory region selftest,
i.e. I'm relatively confident the bug is fixed.

Patches 2-6 are a variety of selftest cleanup, with the aforementioned
set memory region enhancement coming in patch 7.

Note, I couldn't get the selftest to fail outright or with KASAN, but was
able to hit a WARN_ON an invalid slot 100% of the time (without the fix,
obviously).

Regarding s390, I played around a bit with merging gfn_to_memslot_approx()
into search_memslots().  Code wise it's trivial since they're basically
identical, but doing so increases the code footprint of search_memslots()
on x86 by 30 bytes, so I ended up abandoning the effort.

Sean Christopherson (7):
  KVM: Fix out of range accesses to memslots
  KVM: selftests: Fix cosmetic copy-paste error in vm_mem_region_move()
  KVM: selftests: Take vcpu pointer instead of id in vm_vcpu_rm()
  KVM: selftests: Add helpers to consolidate open coded list operations
  KVM: selftests: Add util to delete memory region
  KVM: selftests: Expose the primary memslot number to tests
  KVM: selftests: Add "delete" testcase to set_memory_region_test

 arch/s390/kvm/kvm-s390.c                      |   3 +
 include/linux/kvm_host.h                      |   3 +
 .../testing/selftests/kvm/include/kvm_util.h  |   3 +
 tools/testing/selftests/kvm/lib/kvm_util.c    | 139 ++++++++++--------
 .../kvm/x86_64/set_memory_region_test.c       | 122 +++++++++++++--
 virt/kvm/kvm_main.c                           |   3 +
 6 files changed, 201 insertions(+), 72 deletions(-)

Comments

Paolo Bonzini March 24, 2020, 11:30 a.m. UTC | #1

On 20/03/20 21:55, Sean Christopherson wrote:
> Fix a bug introduced by dynamic memslot allocation where the LRU slot can
> become invalid and lead to a out-of-bounds/use-after-free scenario.
> 
> The patch is different that what Qian has already tested, but I was able
> to reproduce the bad behavior by enhancing the set memory region selftest,
> i.e. I'm relatively confident the bug is fixed.
> 
> Patches 2-6 are a variety of selftest cleanup, with the aforementioned
> set memory region enhancement coming in patch 7.
> 
> Note, I couldn't get the selftest to fail outright or with KASAN, but was
> able to hit a WARN_ON an invalid slot 100% of the time (without the fix,
> obviously).
> 
> Regarding s390, I played around a bit with merging gfn_to_memslot_approx()
> into search_memslots().  Code wise it's trivial since they're basically
> identical, but doing so increases the code footprint of search_memslots()
> on x86 by 30 bytes, so I ended up abandoning the effort.
> 
> Sean Christopherson (7):
>   KVM: Fix out of range accesses to memslots
>   KVM: selftests: Fix cosmetic copy-paste error in vm_mem_region_move()
>   KVM: selftests: Take vcpu pointer instead of id in vm_vcpu_rm()
>   KVM: selftests: Add helpers to consolidate open coded list operations
>   KVM: selftests: Add util to delete memory region
>   KVM: selftests: Expose the primary memslot number to tests
>   KVM: selftests: Add "delete" testcase to set_memory_region_test
> 
>  arch/s390/kvm/kvm-s390.c                      |   3 +
>  include/linux/kvm_host.h                      |   3 +
>  .../testing/selftests/kvm/include/kvm_util.h  |   3 +
>  tools/testing/selftests/kvm/lib/kvm_util.c    | 139 ++++++++++--------
>  .../kvm/x86_64/set_memory_region_test.c       | 122 +++++++++++++--
>  virt/kvm/kvm_main.c                           |   3 +
>  6 files changed, 201 insertions(+), 72 deletions(-)
> 

Queued patches 1-3, thanks.

Paolo