[PATCH] kunit: Make kunit_remove_resource() idempotent

[PATCH] kunit: Make kunit_remove_resource() idempotent

[David Gow]

The kunit_remove_resource() function is used to unlink a resource from
the list of resources in the test, making it no longer show up in
kunit_find_resource().

However, this could lead to a race condition if two threads called
kunit_remove_resource() on the same resource at the same time: the
resource would be removed from the list twice (causing a crash at the
second list_del()), and the refcount for the resource would be
decremented twice (instead of once, for the reference held by the
resource list).

Fix both problems, the first by using list_del_init(), and the second by
checking if the resource has already been removed using list_empty(),
and only decrementing its refcount if it has not.

Also add a KUnit test for the kunit_remove_resource() function which
tests this behaviour.

Reported-by: Daniel Latypov <dlatypov@google.com>
Signed-off-by: David Gow <davidgow@google.com>
---
 lib/kunit/kunit-test.c | 35 +++++++++++++++++++++++++++++++++++
 lib/kunit/test.c       |  8 ++++++--
 2 files changed, 41 insertions(+), 2 deletions(-)

diff --git a/lib/kunit/kunit-test.c b/lib/kunit/kunit-test.c
index 555601d17f79..9005034558aa 100644
--- a/lib/kunit/kunit-test.c
+++ b/lib/kunit/kunit-test.c
@@ -190,6 +190,40 @@ static void kunit_resource_test_destroy_resource(struct kunit *test)
 	KUNIT_EXPECT_TRUE(test, list_empty(&ctx->test.resources));
 }
 
+static void kunit_resource_test_remove_resource(struct kunit *test)
+{
+	struct kunit_test_resource_context *ctx = test->priv;
+	struct kunit_resource *res = kunit_alloc_and_get_resource(
+			&ctx->test,
+			fake_resource_init,
+			fake_resource_free,
+			GFP_KERNEL,
+			ctx);
+
+	/* The resource is in the list */
+	KUNIT_EXPECT_FALSE(test, list_empty(&ctx->test.resources));
+
+	/* Remove the resource. The pointer is still valid, but it can't be
+	 * found.
+	 */
+	kunit_remove_resource(test, res);
+	KUNIT_EXPECT_TRUE(test, list_empty(&ctx->test.resources));
+	/* We haven't been freed yet. */
+	KUNIT_EXPECT_TRUE(test, ctx->is_resource_initialized);
+
+	/* Removing the resource multiple times is valid. */
+	kunit_remove_resource(test, res);
+	KUNIT_EXPECT_TRUE(test, list_empty(&ctx->test.resources));
+	/* Despite having been removed twice (from only one reference), the
+	 * resource still has not been freed.
+	 */
+	KUNIT_EXPECT_TRUE(test, ctx->is_resource_initialized);
+
+	/* Free the resource. */
+	kunit_put_resource(res);
+	KUNIT_EXPECT_FALSE(test, ctx->is_resource_initialized);
+}
+
 static void kunit_resource_test_cleanup_resources(struct kunit *test)
 {
 	int i;
@@ -387,6 +421,7 @@ static struct kunit_case kunit_resource_test_cases[] = {
 	KUNIT_CASE(kunit_resource_test_init_resources),
 	KUNIT_CASE(kunit_resource_test_alloc_resource),
 	KUNIT_CASE(kunit_resource_test_destroy_resource),
+	KUNIT_CASE(kunit_resource_test_remove_resource),
 	KUNIT_CASE(kunit_resource_test_cleanup_resources),
 	KUNIT_CASE(kunit_resource_test_proper_free_ordering),
 	KUNIT_CASE(kunit_resource_test_static),
diff --git a/lib/kunit/test.c b/lib/kunit/test.c
index 3bca3bf5c15b..8411cdfe40a8 100644
--- a/lib/kunit/test.c
+++ b/lib/kunit/test.c
@@ -680,11 +680,15 @@ EXPORT_SYMBOL_GPL(kunit_alloc_and_get_resource);
 void kunit_remove_resource(struct kunit *test, struct kunit_resource *res)
 {
 	unsigned long flags;
+	bool was_linked;
 
 	spin_lock_irqsave(&test->lock, flags);
-	list_del(&res->node);
+	was_linked = !list_empty(&res->node);
+	list_del_init(&res->node);
 	spin_unlock_irqrestore(&test->lock, flags);
-	kunit_put_resource(res);
+
+	if (was_linked)
+		kunit_put_resource(res);
 }
 EXPORT_SYMBOL_GPL(kunit_remove_resource);
 
-- 
2.35.1.894.gb6a874cedc-goog

Re: [PATCH] kunit: Make kunit_remove_resource() idempotent

[Brendan Higgins]

On Fri, Mar 18, 2022 at 2:50 AM David Gow <davidgow@google.com> wrote:
>
> The kunit_remove_resource() function is used to unlink a resource from
> the list of resources in the test, making it no longer show up in
> kunit_find_resource().
>
> However, this could lead to a race condition if two threads called
> kunit_remove_resource() on the same resource at the same time: the
> resource would be removed from the list twice (causing a crash at the
> second list_del()), and the refcount for the resource would be
> decremented twice (instead of once, for the reference held by the
> resource list).
>
> Fix both problems, the first by using list_del_init(), and the second by
> checking if the resource has already been removed using list_empty(),
> and only decrementing its refcount if it has not.
>
> Also add a KUnit test for the kunit_remove_resource() function which
> tests this behaviour.
>
> Reported-by: Daniel Latypov <dlatypov@google.com>
> Signed-off-by: David Gow <davidgow@google.com>

Reviewed-by: Brendan Higgins <brendanhiggins@google.com>