[media] V4L: improve the BKL replacement heuristic

The BKL replacement mutex had some serious performance side-effects on
V4L drivers. It is replaced by a better heuristic that works around the
worst of the side-effects.

Read the v4l2-dev.c comments for the whole sorry story. This is a
temporary measure only until we can convert all v4l drivers to use
unlocked_ioctl.

Signed-off-by: Hans Verkuil <hverkuil@xs4all.nl>
Signed-off-by: Mauro Carvalho Chehab <mchehab@redhat.com>

2 files changed, 29 insertions, 3 deletions

diff --git a/drivers/media/video/v4l2-dev.c b/drivers/media/video/v4l2-dev.c
index bfd392e2436d..6b64fd607b20 100644
--- a/drivers/media/video/v4l2-dev.c
+++ b/drivers/media/video/v4l2-dev.c
@@ -245,13 +245,38 @@ static long v4l2_ioctl(struct file *filp, unsigned int cmd, unsigned long arg)
                 if (vdev->lock)
                         mutex_unlock(vdev->lock);
         } else if (vdev->fops->ioctl) {
-                /* TODO: convert all drivers to unlocked_ioctl */
+                /* This code path is a replacement for the BKL. It is a major
+                 * hack but it will have to do for those drivers that are not
+                 * yet converted to use unlocked_ioctl.
+                 *
+                 * There are two options: if the driver implements struct
+                 * v4l2_device, then the lock defined there is used to
+                 * serialize the ioctls. Otherwise the v4l2 core lock defined
+                 * below is used. This lock is really bad since it serializes
+                 * completely independent devices.
+                 *
+                 * Both variants suffer from the same problem: if the driver
+                 * sleeps, then it blocks all ioctls since the lock is still
+                 * held. This is very common for VIDIOC_DQBUF since that
+                 * normally waits for a frame to arrive. As a result any other
+                 * ioctl calls will proceed very, very slowly since each call
+                 * will have to wait for the VIDIOC_QBUF to finish. Things that
+                 * should take 0.01s may now take 10-20 seconds.
+                 *
+                 * The workaround is to *not* take the lock for VIDIOC_DQBUF.
+                 * This actually works OK for videobuf-based drivers, since
+                 * videobuf will take its own internal lock.
+                 */
                 static DEFINE_MUTEX(v4l2_ioctl_mutex);
+                struct mutex *m = vdev->v4l2_dev ?
+                        &vdev->v4l2_dev->ioctl_lock : &v4l2_ioctl_mutex;
 
-                mutex_lock(&v4l2_ioctl_mutex);
+                if (cmd != VIDIOC_DQBUF && mutex_lock_interruptible(m))
+                        return -ERESTARTSYS;
                 if (video_is_registered(vdev))
                         ret = vdev->fops->ioctl(filp, cmd, arg);
-                mutex_unlock(&v4l2_ioctl_mutex);
+                if (cmd != VIDIOC_DQBUF)
+                        mutex_unlock(m);
         } else
                 ret = -ENOTTY;
 
diff --git a/drivers/media/video/v4l2-device.c b/drivers/media/video/v4l2-device.c
index 0b08f96b74a5..7fe6f92af480 100644
--- a/drivers/media/video/v4l2-device.c
+++ b/drivers/media/video/v4l2-device.c
@@ -35,6 +35,7 @@ int v4l2_device_register(struct device *dev, struct v4l2_device *v4l2_dev)
 
         INIT_LIST_HEAD(&v4l2_dev->subdevs);
         spin_lock_init(&v4l2_dev->lock);
+        mutex_init(&v4l2_dev->ioctl_lock);
         v4l2_dev->dev = dev;
         if (dev == NULL) {
                 /* If dev == NULL, then name must be filled in by the caller */