]> git.hungrycats.org Git - linux/commitdiff
KVM: guest_memfd: take the invalidate lock when unbinding a dying file
authorShivank Garg <shivankg@amd.com>
Tue, 1 Sep 2026 07:20:38 +0000 (07:20 +0000)
committerGreg Kroah-Hartman <gregkh@linuxfoundation.org>
Fri, 6 Nov 2026 17:10:49 +0000 (12:10 -0500)
commit 0ee2c883b62deb74d8c43007e8049dc006d3a2fb upstream.

kvm_gmem_unbind() takes mapping->invalidate_lock only if the
get_file_active() succeeds. When the guest_memfd file is dying, its
reference count is already zero so get_file_active fails and the binding
is removed without holding the invalidate lock.

kvm_gmem_invalidate_{start,end}() checks f->bindings independently to
decide whether to begin or end KVM MMU invalidation. So, the bindings
must therefore remain stable between the two calls. Otherwise, unbind
can remove a binding after start increments mmu_invalidate_in_progress,
but before end finds the binding and decrements it.
Example, unbind race with memory failure:

  CPU 0: memory failure               CPU 1: memslot delete
  ----------------------------------  ---------------------------
                                      (guest_memfd file is dying)
  kvm_gmem_error_folio()
    kvm_gmem_invalidate_start()
      finds binding
      mmu_invalidate_in_progress++
                                      kvm_gmem_unbind()
                                        get_file_active() fails
                                        removes binding
    kvm_gmem_invalidate_end()
      no binding found
      counter stays elevated

mmu_invalidate_retry() then returns 1 forever, so guest page faults
retry without ever installing a mapping and the guest hangs.

Remove the distinction between live and dying files in the unbind path.
Always use slot->gmem.file and take the invalidate lock when it's non-NULL.
Normal unbind callers hold slots_lock, which prevents release from clearing
the pointer or freeing the file state until unbind completes. Final VM
teardown can only see a NULL pointer because guest_memfd pins the KVM until
release has cleared all bindings.

Fixes: ae431059e75d ("KVM: guest_memfd: Remove bindings on memslot deletion when gmem is dying")
Cc: stable@vger.kernel.org
Reported-by: Sashiko <sashiko-bot@kernel.org>
Closes: https://lore.kernel.org/all/20260728092027.225CF1F000E9@smtp.kernel.org
Suggested-by: Sean Christopherson <seanjc@google.com>
Signed-off-by: Shivank Garg <shivankg@amd.com>
Reviewed-by: David Hildenbrand (Arm) <david@kernel.org>
Link: https://patch.msgid.link/20260901-gmem-unbind-fix-v2-1-12febe1d2216@amd.com
Signed-off-by: Sean Christopherson <seanjc@google.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
virt/kvm/guest_memfd.c

index 7fb8295f424b69780fa9ced65b6465219d1a51d5..dcb0cd16c4aa475b287b691f5385714374168902 100644 (file)
@@ -686,48 +686,40 @@ int kvm_gmem_commit_memory_region(struct kvm *kvm, struct kvm_memory_slot *slot)
        return r;
 }
 
-static void __kvm_gmem_unbind(struct kvm_memory_slot *slot, struct gmem_file *f)
+void kvm_gmem_unbind(struct kvm_memory_slot *slot)
 {
+       struct file *file = slot->gmem.file;
        unsigned long start = slot->gmem.pgoff;
        unsigned long end = start + slot->npages;
+       struct gmem_file *f;
 
-       xa_store_range(&f->bindings, start, end - 1, NULL, GFP_KERNEL);
-
-       /*
-        * synchronize_srcu(&kvm->srcu) ensured that kvm_gmem_get_pfn()
-        * cannot see this memslot.
-        */
-       WRITE_ONCE(slot->gmem.file, NULL);
-}
-
-void kvm_gmem_unbind(struct kvm_memory_slot *slot)
-{
        /*
         * Nothing to do if the underlying file was _already_ closed, as
         * kvm_gmem_release() invalidates and nullifies all bindings.
         */
-       if (!slot->gmem.file)
+       if (!file)
                return;
 
-       CLASS(gmem_get_file, file)(slot);
-
        /*
         * However, if the file is _being_ closed, then the bindings need to be
         * removed as kvm_gmem_release() might not run until after the memslot
-        * is freed.  Note, modifying the bindings is safe even though the file
-        * is dying as kvm_gmem_release() nullifies slot->gmem.file under
-        * slots_lock, and only puts its reference to KVM after destroying all
-        * bindings.  I.e. reaching this point means kvm_gmem_release() hasn't
-        * yet destroyed the bindings or freed the gmem_file, and can't do so
-        * until the caller drops slots_lock.
+        * is freed.  Modifying the bindings is safe even if the file is dying
+        * as kvm_gmem_release() nullifies slot->gmem.file under slots_lock,
+        * and only puts its reference to KVM after destroying all bindings.
+        * I.e. reaching this point means kvm_gmem_release() hasn't destroyed
+        * the bindings or freed the gmem_file and can't do so until the caller
+        * drops slots_lock, so there's no need to verify the file is live.
         */
-       if (!file) {
-               __kvm_gmem_unbind(slot, slot->gmem.file->private_data);
-               return;
-       }
+       f = file->private_data;
 
        filemap_invalidate_lock(file->f_mapping);
-       __kvm_gmem_unbind(slot, file->private_data);
+       xa_store_range(&f->bindings, start, end - 1, NULL, GFP_KERNEL);
+
+       /*
+        * synchronize_srcu(&kvm->srcu) ensured that kvm_gmem_get_pfn()
+        * cannot see this memslot.
+        */
+       WRITE_ONCE(slot->gmem.file, NULL);
        filemap_invalidate_unlock(file->f_mapping);
 }