KVM forbids usage of tagged userspace addresses for memslots. It is done
by checking if the address stays the same after untagging.
It is works fine for ARM TBI, but it the check gets racy for LAM. TBI
enabling happens per-thread, so nobody can enable tagging for the thread
while the memslot gets added.
LAM gets enabled per-process. If it gets enabled after the
untagged_addr() check, but before access_ok() check the kernel can
wrongly allow tagged userspace_addr.
Use mmap lock to protect against parallel LAM enabling.
Signed-off-by: Kirill A. Shutemov <kirill.shutemov@xxxxxxxxxxxxxxx>
Reported-by: Rick Edgecombe <rick.p.edgecombe@xxxxxxxxx>
Cc: Marc Zyngier <maz@xxxxxxxxxx>
---
virt/kvm/kvm_main.c | 12 +++++++++++-
1 file changed, 11 insertions(+), 1 deletion(-)
diff --git a/virt/kvm/kvm_main.c b/virt/kvm/kvm_main.c
index 8c86b06b35da..833742c21c91 100644
--- a/virt/kvm/kvm_main.c
+++ b/virt/kvm/kvm_main.c
@@ -1943,12 +1943,22 @@ int __kvm_set_memory_region(struct kvm *kvm,
return -EINVAL;
if (mem->guest_phys_addr & (PAGE_SIZE - 1))
return -EINVAL;
+
+ /* Serialize against tagging enabling */
+ if (mmap_read_lock_killable(kvm->mm))
+ return -EINTR;
+
/* We can read the guest memory with __xxx_user() later on. */
if ((mem->userspace_addr & (PAGE_SIZE - 1)) ||
(mem->userspace_addr != untagged_addr(kvm->mm, mem->userspace_addr)) ||
!access_ok((void __user *)(unsigned long)mem->userspace_addr,
- mem->memory_size))
+ mem->memory_size)) {
+ mmap_read_unlock(kvm->mm);
return -EINVAL;
+ }
+
+ mmap_read_unlock(kvm->mm);
+
if (as_id >= KVM_ADDRESS_SPACE_NUM || id >= KVM_MEM_SLOTS_NUM)
return -EINVAL;
if (mem->guest_phys_addr + mem->memory_size < mem->guest_phys_addr)
--
2.38.0
