gpiolib: rework the locking mechanism for lineevent kfifo
authorBartosz Golaszewski <bgolaszewski@baylibre.com>
Wed, 27 Nov 2019 11:19:21 +0000 (12:19 +0100)
committerBartosz Golaszewski <bgolaszewski@baylibre.com>
Wed, 12 Feb 2020 11:05:47 +0000 (12:05 +0100)
The read_lock mutex is supposed to prevent collisions between reading
and writing to the line event kfifo but it's actually only taken when
the events are being read from it.

Drop the mutex entirely and reuse the spinlock made available to us in
the waitqueue struct. Take the lock whenever the fifo is modified or
inspected. Drop the call to kfifo_to_user() and instead first extract
the new element from kfifo when the lock is taken and only then pass
it on to the user after the spinlock is released.

Signed-off-by: Bartosz Golaszewski <bgolaszewski@baylibre.com>
Reviewed-by: Andy Shevchenko <andriy.shevchenko@linux.intel.com>
drivers/gpio/gpiolib.c

index 753283486037435f33351e99c67c13e13fe04103..43d98309e725691e767a6698954bb56144bd3d4c 100644 (file)
@@ -787,8 +787,6 @@ out_free_lh:
  * @irq: the interrupt that trigger in response to events on this GPIO
  * @wait: wait queue that handles blocking reads of events
  * @events: KFIFO for the GPIO events
- * @read_lock: mutex lock to protect reads from colliding with adding
- * new events to the FIFO
  * @timestamp: cache for the timestamp storing it between hardirq
  * and IRQ thread, used to bring the timestamp close to the actual
  * event
@@ -801,7 +799,6 @@ struct lineevent_state {
        int irq;
        wait_queue_head_t wait;
        DECLARE_KFIFO(events, struct gpioevent_data, 16);
-       struct mutex read_lock;
        u64 timestamp;
 };
 
@@ -817,7 +814,7 @@ static __poll_t lineevent_poll(struct file *filep,
 
        poll_wait(filep, &le->wait, wait);
 
-       if (!kfifo_is_empty(&le->events))
+       if (!kfifo_is_empty_spinlocked_noirqsave(&le->events, &le->wait.lock))
                events = EPOLLIN | EPOLLRDNORM;
 
        return events;
@@ -830,43 +827,52 @@ static ssize_t lineevent_read(struct file *filep,
                              loff_t *f_ps)
 {
        struct lineevent_state *le = filep->private_data;
-       unsigned int copied;
+       struct gpioevent_data event;
+       ssize_t bytes_read = 0;
        int ret;
 
-       if (count < sizeof(struct gpioevent_data))
+       if (count < sizeof(event))
                return -EINVAL;
 
        do {
+               spin_lock(&le->wait.lock);
                if (kfifo_is_empty(&le->events)) {
-                       if (filep->f_flags & O_NONBLOCK)
+                       if (bytes_read) {
+                               spin_unlock(&le->wait.lock);
+                               return bytes_read;
+                       }
+
+                       if (filep->f_flags & O_NONBLOCK) {
+                               spin_unlock(&le->wait.lock);
                                return -EAGAIN;
+                       }
 
-                       ret = wait_event_interruptible(le->wait,
+                       ret = wait_event_interruptible_locked(le->wait,
                                        !kfifo_is_empty(&le->events));
-                       if (ret)
+                       if (ret) {
+                               spin_unlock(&le->wait.lock);
                                return ret;
+                       }
                }
 
-               if (mutex_lock_interruptible(&le->read_lock))
-                       return -ERESTARTSYS;
-               ret = kfifo_to_user(&le->events, buf, count, &copied);
-               mutex_unlock(&le->read_lock);
-
-               if (ret)
-                       return ret;
-
-               /*
-                * If we couldn't read anything from the fifo (a different
-                * thread might have been faster) we either return -EAGAIN if
-                * the file descriptor is non-blocking, otherwise we go back to
-                * sleep and wait for more data to arrive.
-                */
-               if (copied == 0 && (filep->f_flags & O_NONBLOCK))
-                       return -EAGAIN;
+               ret = kfifo_out(&le->events, &event, 1);
+               spin_unlock(&le->wait.lock);
+               if (ret != 1) {
+                       /*
+                        * This should never happen - we were holding the lock
+                        * from the moment we learned the fifo is no longer
+                        * empty until now.
+                        */
+                       ret = -EIO;
+                       break;
+               }
 
-       } while (copied == 0);
+               if (copy_to_user(buf + bytes_read, &event, sizeof(event)))
+                       return -EFAULT;
+               bytes_read += sizeof(event);
+       } while (count >= bytes_read + sizeof(event));
 
-       return copied;
+       return bytes_read;
 }
 
 static int lineevent_release(struct inode *inode, struct file *filep)
@@ -968,7 +974,8 @@ static irqreturn_t lineevent_irq_thread(int irq, void *p)
                return IRQ_NONE;
        }
 
-       ret = kfifo_put(&le->events, ge);
+       ret = kfifo_in_spinlocked_noirqsave(&le->events, &ge,
+                                           1, &le->wait.lock);
        if (ret)
                wake_up_poll(&le->wait, EPOLLIN);
 
@@ -1083,7 +1090,6 @@ static int lineevent_create(struct gpio_device *gdev, void __user *ip)
 
        INIT_KFIFO(le->events);
        init_waitqueue_head(&le->wait);
-       mutex_init(&le->read_lock);
 
        /* Request a thread to read the events */
        ret = request_threaded_irq(le->irq,