Skip to contentSkip to conversation

drm/virtio: notify virtqueues without holding spinlock

PATCH In review Gerd Hoffmann 2 participants 7 messages Aug 13, 2019 – Aug 27, 2019

Subsystems (4)
lore
Share & tools
lore

Gerd Hoffmann

author maintainer
#0 lore raw

cover drm/virtio: notify virtqueues without holding spinlock

Gerd Hoffmann (2):
drm/virtio: cleanup queue functions drm/virtio: notify virtqueues without holding spinlock

 drivers/gpu/drm/virtio/virtgpu_vq.c | 54 ++++++++++++++---------------
 1 file changed, 27 insertions(+), 27 deletions(-)
Signature · 6 lines
2.18.1

_______________________________________________
dri-devel mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/dri-devel
author maintainer
#1 lore raw

PATCH 1/2 drm/virtio: cleanup queue functions

Make the queue functions return void, none of
the call sites checks the return value.

---
 drivers/gpu/drm/virtio/virtgpu_vq.c | 41 ++++++++++-------------------
 1 file changed, 14 insertions(+), 27 deletions(-)
drivers/gpu/drm/virtio/virtgpu_vq.c +14 -27
index 7ac20490e1b4..ca91e83ffaef 100644
@@ -252,8 +252,8 @@ void virtio_gpu_dequeue_cursor_func(struct work_struct *work)
 	wake_up(&vgdev->cursorq.ack_queue);
 }
 
-static int virtio_gpu_queue_ctrl_buffer_locked(struct virtio_gpu_device *vgdev,
-					       struct virtio_gpu_vbuffer *vbuf)
+static void virtio_gpu_queue_ctrl_buffer_locked(struct virtio_gpu_device *vgdev,
+						struct virtio_gpu_vbuffer *vbuf)
 		__releases(&vgdev->ctrlq.qlock)
 		__acquires(&vgdev->ctrlq.qlock)
 {
@@ -263,7 +263,7 @@ static int virtio_gpu_queue_ctrl_buffer_locked(struct virtio_gpu_device *vgdev,
 	int ret;
 
 	if (!vgdev->vqs_ready)
-		return -ENODEV;
+		return;
 
 	sg_init_one(&vcmd, vbuf->buf, vbuf->size);
 	sgs[outcnt + incnt] = &vcmd;
@@ -294,30 +294,22 @@ static int virtio_gpu_queue_ctrl_buffer_locked(struct virtio_gpu_device *vgdev,
 
 		virtqueue_kick(vq);
 	}
-
-	if (!ret)
-		ret = vq->num_free;
-	return ret;
 }
 
-static int virtio_gpu_queue_ctrl_buffer(struct virtio_gpu_device *vgdev,
-					struct virtio_gpu_vbuffer *vbuf)
+static void virtio_gpu_queue_ctrl_buffer(struct virtio_gpu_device *vgdev,
+					 struct virtio_gpu_vbuffer *vbuf)
 {
-	int rc;
-
 	spin_lock(&vgdev->ctrlq.qlock);
-	rc = virtio_gpu_queue_ctrl_buffer_locked(vgdev, vbuf);
+	virtio_gpu_queue_ctrl_buffer_locked(vgdev, vbuf);
 	spin_unlock(&vgdev->ctrlq.qlock);
-	return rc;
 }
 
-static int virtio_gpu_queue_fenced_ctrl_buffer(struct virtio_gpu_device *vgdev,
-					       struct virtio_gpu_vbuffer *vbuf,
-					       struct virtio_gpu_ctrl_hdr *hdr,
-					       struct virtio_gpu_fence *fence)
+static void virtio_gpu_queue_fenced_ctrl_buffer(struct virtio_gpu_device *vgdev,
+						struct virtio_gpu_vbuffer *vbuf,
+						struct virtio_gpu_ctrl_hdr *hdr,
+						struct virtio_gpu_fence *fence)
 {
 	struct virtqueue *vq = vgdev->ctrlq.vq;
-	int rc;
 
 again:
 	spin_lock(&vgdev->ctrlq.qlock);
@@ -338,13 +330,12 @@ static int virtio_gpu_queue_fenced_ctrl_buffer(struct virtio_gpu_device *vgdev,
 
 	if (fence)
 		virtio_gpu_fence_emit(vgdev, hdr, fence);
-	rc = virtio_gpu_queue_ctrl_buffer_locked(vgdev, vbuf);
+	virtio_gpu_queue_ctrl_buffer_locked(vgdev, vbuf);
 	spin_unlock(&vgdev->ctrlq.qlock);
-	return rc;
 }
 
-static int virtio_gpu_queue_cursor(struct virtio_gpu_device *vgdev,
-				   struct virtio_gpu_vbuffer *vbuf)
+static void virtio_gpu_queue_cursor(struct virtio_gpu_device *vgdev,
+				    struct virtio_gpu_vbuffer *vbuf)
 {
 	struct virtqueue *vq = vgdev->cursorq.vq;
 	struct scatterlist *sgs[1], ccmd;
@@ -352,7 +343,7 @@ static int virtio_gpu_queue_cursor(struct virtio_gpu_device *vgdev,
 	int outcnt;
 
 	if (!vgdev->vqs_ready)
-		return -ENODEV;
+		return;
 
 	sg_init_one(&ccmd, vbuf->buf, vbuf->size);
 	sgs[0] = &ccmd;
@@ -374,10 +365,6 @@ static int virtio_gpu_queue_cursor(struct virtio_gpu_device *vgdev,
 	}
 
 	spin_unlock(&vgdev->cursorq.qlock);
-
-	if (!ret)
-		ret = vq->num_free;
-	return ret;
 }
 
 /* just create gem objects for userspace and long lived objects,
Signature · 1 line
2.18.1
author maintainer
#2 lore raw

PATCH 1/2 drm/virtio: cleanup queue functions

In reply to Gerd Hoffmann

Make the queue functions return void, none of
the call sites checks the return value.

---
 drivers/gpu/drm/virtio/virtgpu_vq.c | 41 ++++++++++-------------------
 1 file changed, 14 insertions(+), 27 deletions(-)
drivers/gpu/drm/virtio/virtgpu_vq.c +14 -27
index 7ac20490e1b4..ca91e83ffaef 100644
@@ -252,8 +252,8 @@ void virtio_gpu_dequeue_cursor_func(struct work_struct *work)
 	wake_up(&vgdev->cursorq.ack_queue);
 }
 
-static int virtio_gpu_queue_ctrl_buffer_locked(struct virtio_gpu_device *vgdev,
-					       struct virtio_gpu_vbuffer *vbuf)
+static void virtio_gpu_queue_ctrl_buffer_locked(struct virtio_gpu_device *vgdev,
+						struct virtio_gpu_vbuffer *vbuf)
 		__releases(&vgdev->ctrlq.qlock)
 		__acquires(&vgdev->ctrlq.qlock)
 {
@@ -263,7 +263,7 @@ static int virtio_gpu_queue_ctrl_buffer_locked(struct virtio_gpu_device *vgdev,
 	int ret;
 
 	if (!vgdev->vqs_ready)
-		return -ENODEV;
+		return;
 
 	sg_init_one(&vcmd, vbuf->buf, vbuf->size);
 	sgs[outcnt + incnt] = &vcmd;
@@ -294,30 +294,22 @@ static int virtio_gpu_queue_ctrl_buffer_locked(struct virtio_gpu_device *vgdev,
 
 		virtqueue_kick(vq);
 	}
-
-	if (!ret)
-		ret = vq->num_free;
-	return ret;
 }
 
-static int virtio_gpu_queue_ctrl_buffer(struct virtio_gpu_device *vgdev,
-					struct virtio_gpu_vbuffer *vbuf)
+static void virtio_gpu_queue_ctrl_buffer(struct virtio_gpu_device *vgdev,
+					 struct virtio_gpu_vbuffer *vbuf)
 {
-	int rc;
-
 	spin_lock(&vgdev->ctrlq.qlock);
-	rc = virtio_gpu_queue_ctrl_buffer_locked(vgdev, vbuf);
+	virtio_gpu_queue_ctrl_buffer_locked(vgdev, vbuf);
 	spin_unlock(&vgdev->ctrlq.qlock);
-	return rc;
 }
 
-static int virtio_gpu_queue_fenced_ctrl_buffer(struct virtio_gpu_device *vgdev,
-					       struct virtio_gpu_vbuffer *vbuf,
-					       struct virtio_gpu_ctrl_hdr *hdr,
-					       struct virtio_gpu_fence *fence)
+static void virtio_gpu_queue_fenced_ctrl_buffer(struct virtio_gpu_device *vgdev,
+						struct virtio_gpu_vbuffer *vbuf,
+						struct virtio_gpu_ctrl_hdr *hdr,
+						struct virtio_gpu_fence *fence)
 {
 	struct virtqueue *vq = vgdev->ctrlq.vq;
-	int rc;
 
 again:
 	spin_lock(&vgdev->ctrlq.qlock);
@@ -338,13 +330,12 @@ static int virtio_gpu_queue_fenced_ctrl_buffer(struct virtio_gpu_device *vgdev,
 
 	if (fence)
 		virtio_gpu_fence_emit(vgdev, hdr, fence);
-	rc = virtio_gpu_queue_ctrl_buffer_locked(vgdev, vbuf);
+	virtio_gpu_queue_ctrl_buffer_locked(vgdev, vbuf);
 	spin_unlock(&vgdev->ctrlq.qlock);
-	return rc;
 }
 
-static int virtio_gpu_queue_cursor(struct virtio_gpu_device *vgdev,
-				   struct virtio_gpu_vbuffer *vbuf)
+static void virtio_gpu_queue_cursor(struct virtio_gpu_device *vgdev,
+				    struct virtio_gpu_vbuffer *vbuf)
 {
 	struct virtqueue *vq = vgdev->cursorq.vq;
 	struct scatterlist *sgs[1], ccmd;
@@ -352,7 +343,7 @@ static int virtio_gpu_queue_cursor(struct virtio_gpu_device *vgdev,
 	int outcnt;
 
 	if (!vgdev->vqs_ready)
-		return -ENODEV;
+		return;
 
 	sg_init_one(&ccmd, vbuf->buf, vbuf->size);
 	sgs[0] = &ccmd;
@@ -374,10 +365,6 @@ static int virtio_gpu_queue_cursor(struct virtio_gpu_device *vgdev,
 	}
 
 	spin_unlock(&vgdev->cursorq.qlock);
-
-	if (!ret)
-		ret = vq->num_free;
-	return ret;
 }
 
 /* just create gem objects for userspace and long lived objects,
Signature · 1 line
2.18.1
author maintainer
#3 lore raw

PATCH 2/2 drm/virtio: notify virtqueues without holding spinlock

In reply to Gerd Hoffmann

Split virtqueue_kick() call into virtqueue_kick_prepare(), which requires serialization, and virtqueue_notify(), which does not. Move the virtqueue_notify() call out of the critical section protected by the queue lock. This avoids triggering a vmexit while holding the lock and thereby fixes a rather bad spinlock contention.

---
 drivers/gpu/drm/virtio/virtgpu_vq.c | 25 +++++++++++++++++++------
 1 file changed, 19 insertions(+), 6 deletions(-)
drivers/gpu/drm/virtio/virtgpu_vq.c +19 -6
index ca91e83ffaef..e41c96143342 100644
@@ -252,7 +252,7 @@ void virtio_gpu_dequeue_cursor_func(struct work_struct *work)
 	wake_up(&vgdev->cursorq.ack_queue);
 }
 
-static void virtio_gpu_queue_ctrl_buffer_locked(struct virtio_gpu_device *vgdev,
+static bool virtio_gpu_queue_ctrl_buffer_locked(struct virtio_gpu_device *vgdev,
 						struct virtio_gpu_vbuffer *vbuf)
 		__releases(&vgdev->ctrlq.qlock)
 		__acquires(&vgdev->ctrlq.qlock)
@@ -260,10 +260,11 @@ static void virtio_gpu_queue_ctrl_buffer_locked(struct virtio_gpu_device *vgdev,
 	struct virtqueue *vq = vgdev->ctrlq.vq;
 	struct scatterlist *sgs[3], vcmd, vout, vresp;
 	int outcnt = 0, incnt = 0;
+	bool notify = false;
 	int ret;
 
 	if (!vgdev->vqs_ready)
-		return;
+		return notify;
 
 	sg_init_one(&vcmd, vbuf->buf, vbuf->size);
 	sgs[outcnt + incnt] = &vcmd;
@@ -292,16 +293,21 @@ static void virtio_gpu_queue_ctrl_buffer_locked(struct virtio_gpu_device *vgdev,
 		trace_virtio_gpu_cmd_queue(vq,
 			(struct virtio_gpu_ctrl_hdr *)vbuf->buf);
 
-		virtqueue_kick(vq);
+		notify = virtqueue_kick_prepare(vq);
 	}
+	return notify;
 }
 
 static void virtio_gpu_queue_ctrl_buffer(struct virtio_gpu_device *vgdev,
 					 struct virtio_gpu_vbuffer *vbuf)
 {
+	bool notify;
+
 	spin_lock(&vgdev->ctrlq.qlock);
-	virtio_gpu_queue_ctrl_buffer_locked(vgdev, vbuf);
+	notify = virtio_gpu_queue_ctrl_buffer_locked(vgdev, vbuf);
 	spin_unlock(&vgdev->ctrlq.qlock);
+	if (notify)
+		virtqueue_notify(vgdev->ctrlq.vq);
 }
 
 static void virtio_gpu_queue_fenced_ctrl_buffer(struct virtio_gpu_device *vgdev,
@@ -310,6 +316,7 @@ static void virtio_gpu_queue_fenced_ctrl_buffer(struct virtio_gpu_device *vgdev,
 						struct virtio_gpu_fence *fence)
 {
 	struct virtqueue *vq = vgdev->ctrlq.vq;
+	bool notify;
 
 again:
 	spin_lock(&vgdev->ctrlq.qlock);
@@ -330,8 +337,10 @@ static void virtio_gpu_queue_fenced_ctrl_buffer(struct virtio_gpu_device *vgdev,
 
 	if (fence)
 		virtio_gpu_fence_emit(vgdev, hdr, fence);
-	virtio_gpu_queue_ctrl_buffer_locked(vgdev, vbuf);
+	notify = virtio_gpu_queue_ctrl_buffer_locked(vgdev, vbuf);
 	spin_unlock(&vgdev->ctrlq.qlock);
+	if (notify)
+		virtqueue_notify(vgdev->ctrlq.vq);
 }
 
 static void virtio_gpu_queue_cursor(struct virtio_gpu_device *vgdev,
@@ -339,6 +348,7 @@ static void virtio_gpu_queue_cursor(struct virtio_gpu_device *vgdev,
 {
 	struct virtqueue *vq = vgdev->cursorq.vq;
 	struct scatterlist *sgs[1], ccmd;
+	bool notify;
 	int ret;
 	int outcnt;
 
@@ -361,10 +371,13 @@ static void virtio_gpu_queue_cursor(struct virtio_gpu_device *vgdev,
 		trace_virtio_gpu_cmd_queue(vq,
 			(struct virtio_gpu_ctrl_hdr *)vbuf->buf);
 
-		virtqueue_kick(vq);
+		notify = virtqueue_kick_prepare(vq);
 	}
 
 	spin_unlock(&vgdev->cursorq.qlock);
+
+	if (notify)
+		virtqueue_notify(vq);
 }
 
 /* just create gem objects for userspace and long lived objects,
Signature · 1 line
2.18.1
reviewer Reviewed-by on v1 2/2
#4 lore raw
Quoted 9 lines · Gerd Hoffmann

Split virtqueue_kick() call into virtqueue_kick_prepare(), which requires serialization, and virtqueue_notify(), which does not. Move the virtqueue_notify() call out of the critical section protected by the queue lock. This avoids triggering a vmexit while holding the lock and thereby fixes a rather bad spinlock contention.

Series is

Quoted 100 lines
---
 drivers/gpu/drm/virtio/virtgpu_vq.c | 25 +++++++++++++++++++------
 1 file changed, 19 insertions(+), 6 deletions(-)
drivers/gpu/drm/virtio/virtgpu_vq.c +19 -7
index ca91e83ffaef..e41c96143342 100644
@@ -252,7 +252,7 @@ void virtio_gpu_dequeue_cursor_func(struct work_struct *work)
        wake_up(&vgdev->cursorq.ack_queue);
 }
 
-static void virtio_gpu_queue_ctrl_buffer_locked(struct virtio_gpu_device *vgdev,
+static bool virtio_gpu_queue_ctrl_buffer_locked(struct virtio_gpu_device *vgdev,
                                                struct virtio_gpu_vbuffer *vbuf)
                __releases(&vgdev->ctrlq.qlock)
                __acquires(&vgdev->ctrlq.qlock)
@@ -260,10 +260,11 @@ static void virtio_gpu_queue_ctrl_buffer_locked(struct virtio_gpu_device *vgdev,
        struct virtqueue *vq = vgdev->ctrlq.vq;
        struct scatterlist *sgs[3], vcmd, vout, vresp;
        int outcnt = 0, incnt = 0;
+       bool notify = false;
        int ret;
 
        if (!vgdev->vqs_ready)
-               return;
+               return notify;
 
        sg_init_one(&vcmd, vbuf->buf, vbuf->size);
        sgs[outcnt + incnt] = &vcmd;
@@ -292,16 +293,21 @@ static void virtio_gpu_queue_ctrl_buffer_locked(struct virtio_gpu_device *vgdev,
                trace_virtio_gpu_cmd_queue(vq,
                        (struct virtio_gpu_ctrl_hdr *)vbuf->buf);
 
-               virtqueue_kick(vq);
+               notify = virtqueue_kick_prepare(vq);
        }
+       return notify;
 }
 
 static void virtio_gpu_queue_ctrl_buffer(struct virtio_gpu_device *vgdev,
                                         struct virtio_gpu_vbuffer *vbuf)
 {
+       bool notify;
+
        spin_lock(&vgdev->ctrlq.qlock);
-       virtio_gpu_queue_ctrl_buffer_locked(vgdev, vbuf);
+       notify = virtio_gpu_queue_ctrl_buffer_locked(vgdev, vbuf);
        spin_unlock(&vgdev->ctrlq.qlock);
+       if (notify)
+               virtqueue_notify(vgdev->ctrlq.vq);
 }
 
 static void virtio_gpu_queue_fenced_ctrl_buffer(struct virtio_gpu_device *vgdev,
@@ -310,6 +316,7 @@ static void virtio_gpu_queue_fenced_ctrl_buffer(struct virtio_gpu_device *vgdev,
                                                struct virtio_gpu_fence *fence)
 {
        struct virtqueue *vq = vgdev->ctrlq.vq;
+       bool notify;
 
 again:
        spin_lock(&vgdev->ctrlq.qlock);
@@ -330,8 +337,10 @@ static void virtio_gpu_queue_fenced_ctrl_buffer(struct virtio_gpu_device *vgdev,
 
        if (fence)
                virtio_gpu_fence_emit(vgdev, hdr, fence);
-       virtio_gpu_queue_ctrl_buffer_locked(vgdev, vbuf);
+       notify = virtio_gpu_queue_ctrl_buffer_locked(vgdev, vbuf);
        spin_unlock(&vgdev->ctrlq.qlock);
+       if (notify)
+               virtqueue_notify(vgdev->ctrlq.vq);
 }
 
 static void virtio_gpu_queue_cursor(struct virtio_gpu_device *vgdev,
@@ -339,6 +348,7 @@ static void virtio_gpu_queue_cursor(struct virtio_gpu_device *vgdev,
 {
        struct virtqueue *vq = vgdev->cursorq.vq;
        struct scatterlist *sgs[1], ccmd;
+       bool notify;
        int ret;
        int outcnt;
 
@@ -361,10 +371,13 @@ static void virtio_gpu_queue_cursor(struct virtio_gpu_device *vgdev,
                trace_virtio_gpu_cmd_queue(vq,
                        (struct virtio_gpu_ctrl_hdr *)vbuf->buf);
 
-               virtqueue_kick(vq);
+               notify = virtqueue_kick_prepare(vq);
        }
 
        spin_unlock(&vgdev->cursorq.qlock);
+
+       if (notify)
+               virtqueue_notify(vq);
 }
 
 /* just create gem objects for userspace and long lived objects,
--

2.18.1

reviewer Reviewed-by on v1 2/2
#5 lore raw
In reply to Gerd Hoffmann
Quoted 9 lines · Gerd Hoffmann

Split virtqueue_kick() call into virtqueue_kick_prepare(), which requires serialization, and virtqueue_notify(), which does not. Move the virtqueue_notify() call out of the critical section protected by the queue lock. This avoids triggering a vmexit while holding the lock and thereby fixes a rather bad spinlock contention.

Series is

Quoted 100 lines
---
 drivers/gpu/drm/virtio/virtgpu_vq.c | 25 +++++++++++++++++++------
 1 file changed, 19 insertions(+), 6 deletions(-)
drivers/gpu/drm/virtio/virtgpu_vq.c +19 -7
index ca91e83ffaef..e41c96143342 100644
@@ -252,7 +252,7 @@ void virtio_gpu_dequeue_cursor_func(struct work_struct *work)
        wake_up(&vgdev->cursorq.ack_queue);
 }
 
-static void virtio_gpu_queue_ctrl_buffer_locked(struct virtio_gpu_device *vgdev,
+static bool virtio_gpu_queue_ctrl_buffer_locked(struct virtio_gpu_device *vgdev,
                                                struct virtio_gpu_vbuffer *vbuf)
                __releases(&vgdev->ctrlq.qlock)
                __acquires(&vgdev->ctrlq.qlock)
@@ -260,10 +260,11 @@ static void virtio_gpu_queue_ctrl_buffer_locked(struct virtio_gpu_device *vgdev,
        struct virtqueue *vq = vgdev->ctrlq.vq;
        struct scatterlist *sgs[3], vcmd, vout, vresp;
        int outcnt = 0, incnt = 0;
+       bool notify = false;
        int ret;
 
        if (!vgdev->vqs_ready)
-               return;
+               return notify;
 
        sg_init_one(&vcmd, vbuf->buf, vbuf->size);
        sgs[outcnt + incnt] = &vcmd;
@@ -292,16 +293,21 @@ static void virtio_gpu_queue_ctrl_buffer_locked(struct virtio_gpu_device *vgdev,
                trace_virtio_gpu_cmd_queue(vq,
                        (struct virtio_gpu_ctrl_hdr *)vbuf->buf);
 
-               virtqueue_kick(vq);
+               notify = virtqueue_kick_prepare(vq);
        }
+       return notify;
 }
 
 static void virtio_gpu_queue_ctrl_buffer(struct virtio_gpu_device *vgdev,
                                         struct virtio_gpu_vbuffer *vbuf)
 {
+       bool notify;
+
        spin_lock(&vgdev->ctrlq.qlock);
-       virtio_gpu_queue_ctrl_buffer_locked(vgdev, vbuf);
+       notify = virtio_gpu_queue_ctrl_buffer_locked(vgdev, vbuf);
        spin_unlock(&vgdev->ctrlq.qlock);
+       if (notify)
+               virtqueue_notify(vgdev->ctrlq.vq);
 }
 
 static void virtio_gpu_queue_fenced_ctrl_buffer(struct virtio_gpu_device *vgdev,
@@ -310,6 +316,7 @@ static void virtio_gpu_queue_fenced_ctrl_buffer(struct virtio_gpu_device *vgdev,
                                                struct virtio_gpu_fence *fence)
 {
        struct virtqueue *vq = vgdev->ctrlq.vq;
+       bool notify;
 
 again:
        spin_lock(&vgdev->ctrlq.qlock);
@@ -330,8 +337,10 @@ static void virtio_gpu_queue_fenced_ctrl_buffer(struct virtio_gpu_device *vgdev,
 
        if (fence)
                virtio_gpu_fence_emit(vgdev, hdr, fence);
-       virtio_gpu_queue_ctrl_buffer_locked(vgdev, vbuf);
+       notify = virtio_gpu_queue_ctrl_buffer_locked(vgdev, vbuf);
        spin_unlock(&vgdev->ctrlq.qlock);
+       if (notify)
+               virtqueue_notify(vgdev->ctrlq.vq);
 }
 
 static void virtio_gpu_queue_cursor(struct virtio_gpu_device *vgdev,
@@ -339,6 +348,7 @@ static void virtio_gpu_queue_cursor(struct virtio_gpu_device *vgdev,
 {
        struct virtqueue *vq = vgdev->cursorq.vq;
        struct scatterlist *sgs[1], ccmd;
+       bool notify;
        int ret;
        int outcnt;
 
@@ -361,10 +371,13 @@ static void virtio_gpu_queue_cursor(struct virtio_gpu_device *vgdev,
                trace_virtio_gpu_cmd_queue(vq,
                        (struct virtio_gpu_ctrl_hdr *)vbuf->buf);
 
-               virtqueue_kick(vq);
+               notify = virtqueue_kick_prepare(vq);
        }
 
        spin_unlock(&vgdev->cursorq.qlock);
+
+       if (notify)
+               virtqueue_notify(vq);
 }
 
 /* just create gem objects for userspace and long lived objects,
--

2.18.1

author maintainer
#6 lore raw

PATCH 2/2 drm/virtio: notify virtqueues without holding spinlock

In reply to Gerd Hoffmann

Split virtqueue_kick() call into virtqueue_kick_prepare(), which requires serialization, and virtqueue_notify(), which does not. Move the virtqueue_notify() call out of the critical section protected by the queue lock. This avoids triggering a vmexit while holding the lock and thereby fixes a rather bad spinlock contention.

---
 drivers/gpu/drm/virtio/virtgpu_vq.c | 25 +++++++++++++++++++------
 1 file changed, 19 insertions(+), 6 deletions(-)
drivers/gpu/drm/virtio/virtgpu_vq.c +19 -6
index ca91e83ffaef..e41c96143342 100644
@@ -252,7 +252,7 @@ void virtio_gpu_dequeue_cursor_func(struct work_struct *work)
 	wake_up(&vgdev->cursorq.ack_queue);
 }
 
-static void virtio_gpu_queue_ctrl_buffer_locked(struct virtio_gpu_device *vgdev,
+static bool virtio_gpu_queue_ctrl_buffer_locked(struct virtio_gpu_device *vgdev,
 						struct virtio_gpu_vbuffer *vbuf)
 		__releases(&vgdev->ctrlq.qlock)
 		__acquires(&vgdev->ctrlq.qlock)
@@ -260,10 +260,11 @@ static void virtio_gpu_queue_ctrl_buffer_locked(struct virtio_gpu_device *vgdev,
 	struct virtqueue *vq = vgdev->ctrlq.vq;
 	struct scatterlist *sgs[3], vcmd, vout, vresp;
 	int outcnt = 0, incnt = 0;
+	bool notify = false;
 	int ret;
 
 	if (!vgdev->vqs_ready)
-		return;
+		return notify;
 
 	sg_init_one(&vcmd, vbuf->buf, vbuf->size);
 	sgs[outcnt + incnt] = &vcmd;
@@ -292,16 +293,21 @@ static void virtio_gpu_queue_ctrl_buffer_locked(struct virtio_gpu_device *vgdev,
 		trace_virtio_gpu_cmd_queue(vq,
 			(struct virtio_gpu_ctrl_hdr *)vbuf->buf);
 
-		virtqueue_kick(vq);
+		notify = virtqueue_kick_prepare(vq);
 	}
+	return notify;
 }
 
 static void virtio_gpu_queue_ctrl_buffer(struct virtio_gpu_device *vgdev,
 					 struct virtio_gpu_vbuffer *vbuf)
 {
+	bool notify;
+
 	spin_lock(&vgdev->ctrlq.qlock);
-	virtio_gpu_queue_ctrl_buffer_locked(vgdev, vbuf);
+	notify = virtio_gpu_queue_ctrl_buffer_locked(vgdev, vbuf);
 	spin_unlock(&vgdev->ctrlq.qlock);
+	if (notify)
+		virtqueue_notify(vgdev->ctrlq.vq);
 }
 
 static void virtio_gpu_queue_fenced_ctrl_buffer(struct virtio_gpu_device *vgdev,
@@ -310,6 +316,7 @@ static void virtio_gpu_queue_fenced_ctrl_buffer(struct virtio_gpu_device *vgdev,
 						struct virtio_gpu_fence *fence)
 {
 	struct virtqueue *vq = vgdev->ctrlq.vq;
+	bool notify;
 
 again:
 	spin_lock(&vgdev->ctrlq.qlock);
@@ -330,8 +337,10 @@ static void virtio_gpu_queue_fenced_ctrl_buffer(struct virtio_gpu_device *vgdev,
 
 	if (fence)
 		virtio_gpu_fence_emit(vgdev, hdr, fence);
-	virtio_gpu_queue_ctrl_buffer_locked(vgdev, vbuf);
+	notify = virtio_gpu_queue_ctrl_buffer_locked(vgdev, vbuf);
 	spin_unlock(&vgdev->ctrlq.qlock);
+	if (notify)
+		virtqueue_notify(vgdev->ctrlq.vq);
 }
 
 static void virtio_gpu_queue_cursor(struct virtio_gpu_device *vgdev,
@@ -339,6 +348,7 @@ static void virtio_gpu_queue_cursor(struct virtio_gpu_device *vgdev,
 {
 	struct virtqueue *vq = vgdev->cursorq.vq;
 	struct scatterlist *sgs[1], ccmd;
+	bool notify;
 	int ret;
 	int outcnt;
 
@@ -361,10 +371,13 @@ static void virtio_gpu_queue_cursor(struct virtio_gpu_device *vgdev,
 		trace_virtio_gpu_cmd_queue(vq,
 			(struct virtio_gpu_ctrl_hdr *)vbuf->buf);
 
-		virtqueue_kick(vq);
+		notify = virtqueue_kick_prepare(vq);
 	}
 
 	spin_unlock(&vgdev->cursorq.qlock);
+
+	if (notify)
+		virtqueue_notify(vq);
 }
 
 /* just create gem objects for userspace and long lived objects,
Signature · 1 line
2.18.1