virtio_pci_modern: use completion instead of busy loop to wait on admin cmd result

Currently, the code waits in a busy loop on every admin virtqueue issued
command to get a reply. That prevents callers from issuing multiple
commands in parallel.

To overcome this limitation, introduce a virtqueue event callback for
admin virtqueue. For every issued command, use completion mechanism
to wait on a reply. In the event callback, trigger the completion
is done for every incoming reply.

Alongside with that, introduce a spin lock to protect the admin
virtqueue operations.

Signed-off-by: Jiri Pirko <jiri@nvidia.com>
Message-Id: <20240716113552.80599-13-jiri@resnulli.us>
Signed-off-by: Michael S. Tsirkin <mst@redhat.com>

drivers/virtio/virtio_pci_common.c

@@ -395,6 +395,8 @@ static int vp_find_vqs_msix(struct virtio_device *vdev, unsigned int nvqs,
 			if (vqi->name && vqi->callback)
 				++nvectors;
 		}
+		if (avq_num && vector_policy == VP_VQ_VECTOR_POLICY_EACH)
+			++nvectors;
 	} else {
 		/* Second best: one for change, shared for all vqs. */
 		nvectors = 2;
@@ -425,9 +427,9 @@ static int vp_find_vqs_msix(struct virtio_device *vdev, unsigned int nvqs,
 	if (!avq_num)
 		return 0;
 	sprintf(avq->name, "avq.%u", avq->vq_index);
-	vq = vp_find_one_vq_msix(vdev, avq->vq_index, NULL, avq->name, false,
-				 true, &allocated_vectors, vector_policy,
-				 &vp_dev->admin_vq.info);
+	vq = vp_find_one_vq_msix(vdev, avq->vq_index, vp_modern_avq_done,
+				 avq->name, false, true, &allocated_vectors,
+				 vector_policy, &vp_dev->admin_vq.info);
 	if (IS_ERR(vq)) {
 		err = PTR_ERR(vq);
 		goto error_find;
@@ -486,8 +488,9 @@ static int vp_find_vqs_intx(struct virtio_device *vdev, unsigned int nvqs,
 	if (!avq_num)
 		return 0;
 	sprintf(avq->name, "avq.%u", avq->vq_index);
-	vq = vp_setup_vq(vdev, queue_idx++, NULL, avq->name, false,
-			 VIRTIO_MSI_NO_VECTOR, &vp_dev->admin_vq.info);
+	vq = vp_setup_vq(vdev, queue_idx++, vp_modern_avq_done, avq->name,
+			 false, VIRTIO_MSI_NO_VECTOR,
+			 &vp_dev->admin_vq.info);
 	if (IS_ERR(vq)) {
 		err = PTR_ERR(vq);
 		goto out_del_vqs;

drivers/virtio/virtio_pci_common.h

@@ -47,6 +47,8 @@ struct virtio_pci_admin_vq {
 	struct virtio_pci_vq_info *info;
 	/* serializing admin commands execution. */
 	struct mutex cmd_lock;
+	/* Protects virtqueue access. */
+	spinlock_t lock;
 	u64 supported_cmds;
 	/* Name of the admin queue: avq.$vq_index. */
 	char name[10];
@@ -178,6 +180,7 @@ struct virtio_device *virtio_pci_vf_get_pf_dev(struct pci_dev *pdev);
 #define VIRTIO_ADMIN_CMD_BITMAP 0
 #endif
 
+void vp_modern_avq_done(struct virtqueue *vq);
 int vp_modern_admin_cmd_exec(struct virtio_device *vdev,
 			     struct virtio_admin_cmd *cmd);
 

drivers/virtio/virtio_pci_modern.c

@@ -53,6 +53,23 @@ static bool vp_is_avq(struct virtio_device *vdev, unsigned int index)
 	return index == vp_dev->admin_vq.vq_index;
 }
 
+void vp_modern_avq_done(struct virtqueue *vq)
+{
+	struct virtio_pci_device *vp_dev = to_vp_device(vq->vdev);
+	struct virtio_pci_admin_vq *admin_vq = &vp_dev->admin_vq;
+	struct virtio_admin_cmd *cmd;
+	unsigned long flags;
+	unsigned int len;
+
+	spin_lock_irqsave(&admin_vq->lock, flags);
+	do {
+		virtqueue_disable_cb(vq);
+		while ((cmd = virtqueue_get_buf(vq, &len)))
+			complete(&cmd->completion);
+	} while (!virtqueue_enable_cb(vq));
+	spin_unlock_irqrestore(&admin_vq->lock, flags);
+}
+
 static int virtqueue_exec_admin_cmd(struct virtio_pci_admin_vq *admin_vq,
 				    u16 opcode,
 				    struct scatterlist **sgs,
@@ -61,7 +78,8 @@ static int virtqueue_exec_admin_cmd(struct virtio_pci_admin_vq *admin_vq,
 				    struct virtio_admin_cmd *cmd)
 {
 	struct virtqueue *vq;
-	int ret, len;
+	unsigned long flags;
+	int ret;
 
 	vq = admin_vq->info->vq;
 	if (!vq)
@@ -72,21 +90,33 @@ static int virtqueue_exec_admin_cmd(struct virtio_pci_admin_vq *admin_vq,
 	    !((1ULL << opcode) & admin_vq->supported_cmds))
 		return -EOPNOTSUPP;
 
-	ret = virtqueue_add_sgs(vq, sgs, out_num, in_num, cmd, GFP_KERNEL);
-	if (ret < 0)
-		return -EIO;
-
-	if (unlikely(!virtqueue_kick(vq)))
-		return -EIO;
-
-	while (!virtqueue_get_buf(vq, &len) &&
-	       !virtqueue_is_broken(vq))
-		cpu_relax();
+	init_completion(&cmd->completion);
 
+again:
 	if (virtqueue_is_broken(vq))
 		return -EIO;
 
-	return 0;
+	spin_lock_irqsave(&admin_vq->lock, flags);
+	ret = virtqueue_add_sgs(vq, sgs, out_num, in_num, cmd, GFP_KERNEL);
+	if (ret < 0) {
+		if (ret == -ENOSPC) {
+			spin_unlock_irqrestore(&admin_vq->lock, flags);
+			cpu_relax();
+			goto again;
+		}
+		goto unlock_err;
+	}
+	if (!virtqueue_kick(vq))
+		goto unlock_err;
+	spin_unlock_irqrestore(&admin_vq->lock, flags);
+
+	wait_for_completion(&cmd->completion);
+
+	return cmd->ret;
+
+unlock_err:
+	spin_unlock_irqrestore(&admin_vq->lock, flags);
+	return -EIO;
 }
 
 int vp_modern_admin_cmd_exec(struct virtio_device *vdev,
@@ -209,6 +239,25 @@ static void vp_modern_avq_activate(struct virtio_device *vdev)
 	virtio_pci_admin_cmd_list_init(vdev);
 }
 
+static void vp_modern_avq_cleanup(struct virtio_device *vdev)
+{
+	struct virtio_pci_device *vp_dev = to_vp_device(vdev);
+	struct virtio_admin_cmd *cmd;
+	struct virtqueue *vq;
+
+	if (!virtio_has_feature(vdev, VIRTIO_F_ADMIN_VQ))
+		return;
+
+	vq = vp_dev->vqs[vp_dev->admin_vq.vq_index]->vq;
+	if (!vq)
+		return;
+
+	while ((cmd = virtqueue_detach_unused_buf(vq))) {
+		cmd->ret = -EIO;
+		complete(&cmd->completion);
+	}
+}
+
 static void vp_transport_features(struct virtio_device *vdev, u64 features)
 {
 	struct virtio_pci_device *vp_dev = to_vp_device(vdev);
@@ -403,6 +452,8 @@ static void vp_reset(struct virtio_device *vdev)
 	while (vp_modern_get_status(mdev))
 		msleep(1);
 
+	vp_modern_avq_cleanup(vdev);
+
 	/* Flush pending VQ/configuration callbacks. */
 	vp_synchronize_vectors(vdev);
 }
@@ -785,6 +836,7 @@ int virtio_pci_modern_probe(struct virtio_pci_device *vp_dev)
 	vp_dev->isr = mdev->isr;
 	vp_dev->vdev.id = mdev->id;
 
+	spin_lock_init(&vp_dev->admin_vq.lock);
 	mutex_init(&vp_dev->admin_vq.cmd_lock);
 	return 0;
 }

include/linux/virtio.h

@@ -10,6 +10,7 @@
 #include <linux/mod_devicetable.h>
 #include <linux/gfp.h>
 #include <linux/dma-mapping.h>
+#include <linux/completion.h>
 
 /**
  * struct virtqueue - a queue to register buffers for sending or receiving.
@@ -109,6 +110,8 @@ struct virtio_admin_cmd {
 	__le64 group_member_id;
 	struct scatterlist *data_sg;
 	struct scatterlist *result_sg;
+	struct completion completion;
+	int ret;
 };
 
 /**