[PATCH v3 2/6] module/dups: Fix use-after-free in kmod_dup_req lifetime handling

The kmod dups code uses RCU to ensure that a kmod_dup_req instance is freed
only after it is no longer referenced. When releasing an instance, the
kmod_dup_request_delete() function removes the kmod_dup_req from the
dup_kmod_reqs list, waits via synchronize_rcu() and finally frees it.
However, this doesn't work correctly because parallel users referencing the
instance in kmod_dup_request_exists_wait() don't enter an RCU read-side
critical section. This can result in a use-after-free.

The kmod_dup_request_exists_wait() function may need to hold a valid
reference to a kmod_dup_req instance across a blocking wait until the
corresponding modprobe command completes. This makes it unsuitable for RCU.

Fix the issue by changing the lifecycle management of kmod_dup_req to use
reference counting.

Fixes: 8660484ed1cf ("module: add debugging auto-load duplicate module support")
Reviewed-by: Aaron Tomlin <[email protected]>
Signed-off-by: Petr Pavlu <[email protected]>
---
 kernel/module/dups.c | 56 +++++++++++++++++++++++++++++++++++++---------------
 1 file changed, 40 insertions(+), 16 deletions(-)

diff --git a/kernel/module/dups.c b/kernel/module/dups.c
index 45080f451e5c..db7377229703 100644
--- a/kernel/module/dups.c
+++ b/kernel/module/dups.c
@@ -30,6 +30,7 @@
 #include <linux/ptrace.h>
 #include <linux/async.h>
 #include <linux/uaccess.h>
+#include <linux/refcount.h>
 
 #include "internal.h"
 
@@ -38,13 +39,12 @@
 static bool enable_dups_trace = IS_ENABLED(CONFIG_MODULE_DEBUG_AUTOLOAD_DUPS_TRACE);
 module_param(enable_dups_trace, bool_enable_only, 0644);
 
-/*
- * Protects dup_kmod_reqs list, adds / removals with RCU.
- */
+/* A mutex-protected list of active kmod requests. */
 static DEFINE_MUTEX(kmod_dup_mutex);
 static LIST_HEAD(dup_kmod_reqs);
 
 struct kmod_dup_req {
+	refcount_t refcount;
 	struct list_head list;
 	char name[MODULE_NAME_LEN];
 	struct completion first_req_done;
@@ -52,12 +52,24 @@ struct kmod_dup_req {
 	int dup_ret;
 };
 
+static void get_kmod_req(struct kmod_dup_req *kmod_req)
+{
+	refcount_inc(&kmod_req->refcount);
+}
+
+static void put_kmod_req(struct kmod_dup_req *kmod_req)
+{
+	if (refcount_dec_and_test(&kmod_req->refcount))
+		kfree(kmod_req);
+}
+
 static struct kmod_dup_req *kmod_dup_request_lookup(char *module_name)
 {
 	struct kmod_dup_req *kmod_req;
 
-	list_for_each_entry_rcu(kmod_req, &dup_kmod_reqs, list,
-				lockdep_is_held(&kmod_dup_mutex)) {
+	lockdep_assert_held(&kmod_dup_mutex);
+
+	list_for_each_entry(kmod_req, &dup_kmod_reqs, list) {
 		if (strlen(kmod_req->name) == strlen(module_name) &&
 		    !memcmp(kmod_req->name, module_name, strlen(module_name))) {
 			return kmod_req;
@@ -86,10 +98,10 @@ static void kmod_dup_request_delete(struct work_struct *work)
 	 * just returning 0.
 	 */
 	mutex_lock(&kmod_dup_mutex);
-	list_del_rcu(&kmod_req->list);
-	synchronize_rcu();
+	list_del(&kmod_req->list);
 	mutex_unlock(&kmod_dup_mutex);
-	kfree(kmod_req);
+
+	put_kmod_req(kmod_req);
 }
 
 bool kmod_dup_request_exists_wait(char *module_name, bool wait, int *dup_ret)
@@ -105,6 +117,7 @@ bool kmod_dup_request_exists_wait(char *module_name, bool wait, int *dup_ret)
 	if (!new_kmod_req)
 		return false;
 
+	refcount_set(&new_kmod_req->refcount, 1);
 	strscpy(new_kmod_req->name, module_name);
 	INIT_DELAYED_WORK(&new_kmod_req->delete_work, kmod_dup_request_delete);
 	init_completion(&new_kmod_req->first_req_done);
@@ -136,10 +149,12 @@ bool kmod_dup_request_exists_wait(char *module_name, bool wait, int *dup_ret)
 		 * keep tab on duplicates later.
 		 */
 		pr_debug("New request_module() for %s\n", module_name);
-		list_add_rcu(&new_kmod_req->list, &dup_kmod_reqs);
+		list_add(&new_kmod_req->list, &dup_kmod_reqs);
 		mutex_unlock(&kmod_dup_mutex);
 		return false;
 	}
+
+	get_kmod_req(kmod_req);
 	mutex_unlock(&kmod_dup_mutex);
 
 	/* We are dealing with a duplicate request now */
@@ -169,7 +184,7 @@ bool kmod_dup_request_exists_wait(char *module_name, bool wait, int *dup_ret)
 		 * calls bail out right away.
 		 */
 		*dup_ret = 0;
-		return true;
+		goto out;
 	}
 
 	/*
@@ -184,12 +199,14 @@ bool kmod_dup_request_exists_wait(char *module_name, bool wait, int *dup_ret)
 					TASK_KILLABLE);
 	if (ret) {
 		*dup_ret = ret;
-		return true;
+		goto out;
 	}
 
 	/* Now the duplicate request has the same exact return value as the first request */
 	*dup_ret = kmod_req->dup_ret;
 
+out:
+	put_kmod_req(kmod_req);
 	return true;
 }
 
@@ -199,15 +216,25 @@ void kmod_dup_request_announce(char *module_name, int ret)
 
 	mutex_lock(&kmod_dup_mutex);
 
+	/*
+	 * Look for a kmod_dup_req previously added in
+	 * kmod_dup_request_exists_wait(). Note that a request_module_nowait()
+	 * without its own kmod_dup_req entry can announce a result of
+	 * a concurrent request_module() call.
+	 */
 	kmod_req = kmod_dup_request_lookup(module_name);
-	if (!kmod_req)
-		goto out;
+	if (!kmod_req || completion_done(&kmod_req->first_req_done)) {
+		mutex_unlock(&kmod_dup_mutex);
+		return;
+	}
 
 	kmod_req->dup_ret = ret;
 
 	/* Inform all duplicate waiters to check the return value. */
 	complete_all(&kmod_req->first_req_done);
 
+	mutex_unlock(&kmod_dup_mutex);
+
 	/*
 	 * Now that we have allowed prior request_module() calls to go on
 	 * with life, let's schedule deleting this entry. We don't have
@@ -216,7 +243,4 @@ void kmod_dup_request_announce(char *module_name, int ret)
 	 * possible abuses of vmalloc() incurred by finit_module() thrashing.
 	 */
 	queue_delayed_work(system_dfl_wq, &kmod_req->delete_work, 60 * HZ);
-
-out:
-	mutex_unlock(&kmod_dup_mutex);
 }

-- 
2.55.0