ipc/mqueue.c: update/document memory barriers
[sfrench/cifs-2.6.git] / ipc / mqueue.c
index 270456530f6a53002dcb90933272c625d4ced1e2..49a05ba3000dbab9f6821e3deb3a5223c29a528a 100644 (file)
@@ -63,6 +63,66 @@ struct posix_msg_tree_node {
        int                     priority;
 };
 
+/*
+ * Locking:
+ *
+ * Accesses to a message queue are synchronized by acquiring info->lock.
+ *
+ * There are two notable exceptions:
+ * - The actual wakeup of a sleeping task is performed using the wake_q
+ *   framework. info->lock is already released when wake_up_q is called.
+ * - The exit codepaths after sleeping check ext_wait_queue->state without
+ *   any locks. If it is STATE_READY, then the syscall is completed without
+ *   acquiring info->lock.
+ *
+ * MQ_BARRIER:
+ * To achieve proper release/acquire memory barrier pairing, the state is set to
+ * STATE_READY with smp_store_release(), and it is read with READ_ONCE followed
+ * by smp_acquire__after_ctrl_dep(). In addition, wake_q_add_safe() is used.
+ *
+ * This prevents the following races:
+ *
+ * 1) With the simple wake_q_add(), the task could be gone already before
+ *    the increase of the reference happens
+ * Thread A
+ *                             Thread B
+ * WRITE_ONCE(wait.state, STATE_NONE);
+ * schedule_hrtimeout()
+ *                             wake_q_add(A)
+ *                             if (cmpxchg()) // success
+ *                                ->state = STATE_READY (reordered)
+ * <timeout returns>
+ * if (wait.state == STATE_READY) return;
+ * sysret to user space
+ * sys_exit()
+ *                             get_task_struct() // UaF
+ *
+ * Solution: Use wake_q_add_safe() and perform the get_task_struct() before
+ * the smp_store_release() that does ->state = STATE_READY.
+ *
+ * 2) Without proper _release/_acquire barriers, the woken up task
+ *    could read stale data
+ *
+ * Thread A
+ *                             Thread B
+ * do_mq_timedreceive
+ * WRITE_ONCE(wait.state, STATE_NONE);
+ * schedule_hrtimeout()
+ *                             state = STATE_READY;
+ * <timeout returns>
+ * if (wait.state == STATE_READY) return;
+ * msg_ptr = wait.msg;         // Access to stale data!
+ *                             receiver->msg = message; (reordered)
+ *
+ * Solution: use _release and _acquire barriers.
+ *
+ * 3) There is intentionally no barrier when setting current->state
+ *    to TASK_INTERRUPTIBLE: spin_unlock(&info->lock) provides the
+ *    release memory barrier, and the wakeup is triggered when holding
+ *    info->lock, i.e. spin_lock(&info->lock) provided a pairing
+ *    acquire memory barrier.
+ */
+
 struct ext_wait_queue {                /* queue of sleeping tasks */
        struct task_struct *task;
        struct list_head list;
@@ -646,18 +706,23 @@ static int wq_sleep(struct mqueue_inode_info *info, int sr,
        wq_add(info, sr, ewp);
 
        for (;;) {
+               /* memory barrier not required, we hold info->lock */
                __set_current_state(TASK_INTERRUPTIBLE);
 
                spin_unlock(&info->lock);
                time = schedule_hrtimeout_range_clock(timeout, 0,
                        HRTIMER_MODE_ABS, CLOCK_REALTIME);
 
-               if (ewp->state == STATE_READY) {
+               if (READ_ONCE(ewp->state) == STATE_READY) {
+                       /* see MQ_BARRIER for purpose/pairing */
+                       smp_acquire__after_ctrl_dep();
                        retval = 0;
                        goto out;
                }
                spin_lock(&info->lock);
-               if (ewp->state == STATE_READY) {
+
+               /* we hold info->lock, so no memory barrier required */
+               if (READ_ONCE(ewp->state) == STATE_READY) {
                        retval = 0;
                        goto out_unlock;
                }
@@ -923,16 +988,11 @@ static inline void __pipelined_op(struct wake_q_head *wake_q,
                                  struct ext_wait_queue *this)
 {
        list_del(&this->list);
-       wake_q_add(wake_q, this->task);
-       /*
-        * Rely on the implicit cmpxchg barrier from wake_q_add such
-        * that we can ensure that updating receiver->state is the last
-        * write operation: As once set, the receiver can continue,
-        * and if we don't have the reference count from the wake_q,
-        * yet, at that point we can later have a use-after-free
-        * condition and bogus wakeup.
-        */
-       this->state = STATE_READY;
+       get_task_struct(this->task);
+
+       /* see MQ_BARRIER for purpose/pairing */
+       smp_store_release(&this->state, STATE_READY);
+       wake_q_add_safe(wake_q, this->task);
 }
 
 /* pipelined_send() - send a message directly to the task waiting in
@@ -1049,7 +1109,9 @@ static int do_mq_timedsend(mqd_t mqdes, const char __user *u_msg_ptr,
                } else {
                        wait.task = current;
                        wait.msg = (void *) msg_ptr;
-                       wait.state = STATE_NONE;
+
+                       /* memory barrier not required, we hold info->lock */
+                       WRITE_ONCE(wait.state, STATE_NONE);
                        ret = wq_sleep(info, SEND, timeout, &wait);
                        /*
                         * wq_sleep must be called with info->lock held, and
@@ -1152,7 +1214,9 @@ static int do_mq_timedreceive(mqd_t mqdes, char __user *u_msg_ptr,
                        ret = -EAGAIN;
                } else {
                        wait.task = current;
-                       wait.state = STATE_NONE;
+
+                       /* memory barrier not required, we hold info->lock */
+                       WRITE_ONCE(wait.state, STATE_NONE);
                        ret = wq_sleep(info, RECV, timeout, &wait);
                        msg_ptr = wait.msg;
                }