Re: [PATCH 2/2] KVM: guest_memfd: Remove RCU-protected attribute from slot->gmem.file

[Paolo Bonzini]

On 11/4/24 09:43, Yan Zhao wrote:
> Remove the RCU-protected attribute from slot->gmem.file. No need to use RCU
> primitives rcu_assign_pointer()/synchronize_rcu() to update this pointer.
>
> - slot->gmem.file is updated in 3 places:
>    kvm_gmem_bind(), kvm_gmem_unbind(), kvm_gmem_release().
>    All of them are protected by kvm->slots_lock.
>
> - slot->gmem.file is read in 2 paths:
>    (1) kvm_gmem_populate
>          kvm_gmem_get_file
>          __kvm_gmem_get_pfn
>
>    (2) kvm_gmem_get_pfn
>           kvm_gmem_get_file
>           __kvm_gmem_get_pfn
>
>    Path (1) kvm_gmem_populate() requires holding kvm->slots_lock, so
>    slot->gmem.file is protected by the kvm->slots_lock in this path.
>
>    Path (2) kvm_gmem_get_pfn() does not require holding kvm->slots_lock.
>    However, it's also not guarded by rcu_read_lock() and rcu_read_unlock().
>    So synchronize_rcu() in kvm_gmem_unbind()/kvm_gmem_release() actually
>    will not wait for the readers in kvm_gmem_get_pfn() due to lack of RCU
>    read-side critical section.
>
>    The path (2) kvm_gmem_get_pfn() is safe without RCU protection because:
>    a) kvm_gmem_bind() is called on a new memslot, before the memslot is
>       visible to kvm_gmem_get_pfn().
>    b) kvm->srcu ensures that kvm_gmem_unbind() and freeing of a memslot
>       occur after the memslot is no longer visible to kvm_gmem_get_pfn().
>    c) get_file_active() ensures that kvm_gmem_get_pfn() will not access the
>       stale file if kvm_gmem_release() sets it to NULL.  This is because if
>       kvm_gmem_release() occurs before kvm_gmem_get_pfn(), get_file_active()
>       will return NULL; if get_file_active() does not return NULL,
>       kvm_gmem_release() should not occur until after kvm_gmem_get_pfn()
>       releases the file reference.

Thanks for the analysis, I added some notes:

diff --git a/include/linux/kvm_host.h b/include/linux/kvm_host.h
index 4ec2564c0d0f..c788d0bd952a 100644
--- a/include/linux/kvm_host.h
+++ b/include/linux/kvm_host.h
@@ -602,6 +602,11 @@ struct kvm_memory_slot {
 
 #ifdef CONFIG_KVM_PRIVATE_MEM
 	struct {
+		/*
+		 * Writes protected by kvm->slots_lock.  Acquiring a
+		 * reference via kvm_gmem_get_file() is protected by
+		 * either kvm->slots_lock or kvm->srcu.
+		 */
 		struct file *file;
 		pgoff_t pgoff;
 	} gmem;
diff --git a/virt/kvm/guest_memfd.c b/virt/kvm/guest_memfd.c
index 9d9bf3d033bd..411ff7224caa 100644
--- a/virt/kvm/guest_memfd.c
+++ b/virt/kvm/guest_memfd.c
@@ -261,6 +261,12 @@ static int kvm_gmem_release(struct inode *inode, struct file *file)
 	 * dereferencing the slot for existing bindings needs to be protected
 	 * against memslot updates, specifically so that unbind doesn't race
 	 * and free the memslot (kvm_gmem_get_file() will return NULL).
+	 *
+	 * Since .release is called only when the reference count is zero,
+	 * after which file_ref_get() and get_file_active() fail,
+	 * kvm_gmem_get_pfn() cannot be using the file concurrently.
+	 * file_ref_put() provides a full barrier, and get_file_active() the
+	 * matching acquire barrier.
 	 */
 	mutex_lock(&kvm->slots_lock);
 
@@ -508,8 +514,8 @@ int kvm_gmem_bind(struct kvm *kvm, struct kvm_memory_slot *slot,
 
 	/*
 	 * memslots of flag KVM_MEM_GUEST_MEMFD are immutable to change, so
-	 * kvm_gmem_bind() must occur on a new memslot.
-	 * Readers are guaranteed to see this new file.
+	 * kvm_gmem_bind() must occur on a new memslot.  Because the memslot
+	 * is not visible yet, kvm_gmem_get_pfn() is guaranteed to see the file.
 	 */
 	WRITE_ONCE(slot->gmem.file, file);
 	slot->gmem.pgoff = start;
@@ -547,6 +554,11 @@ void kvm_gmem_unbind(struct kvm_memory_slot *slot)
 
 	filemap_invalidate_lock(file->f_mapping);
 	xa_store_range(&gmem->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);
 
Queued to kvm-coco-queue.

Paolo