[PATCH RFC v1 0/3] Support live migration with mlx5_vdpa

[PATCH RFC v1 0/3] Support live migration with mlx5_vdpa

[Eli Cohen]

Hi all,
I am sending this patch set with adressing some of the comments.

Please note that I will be on vacation for the next three weeks so any
comments will be addressed after I return.

Eugenio and myself were able to actually see live migation working over
mlx5_vdpa so that's a good sign too :-)

Following patchset supports the suspend callback to suspend VQs to allow
sshdow VQs to kick in. It also adds address space support also required
for live migration. We use two groups, one for the control virtqueue and
one for the data virtqueues.

Finally there is a patch that disables VLAN support since it is not
supported currently.

The aim of this patchset is to enable development/testing of live
migration over a real hardware device.

The below two patches are also required to be applied but I did not
include them in the series since they were sent separately.

https://lore.kernel.org/all/20220613075958.511064-1-elic@nvidia.com/T/
https://lore.kernel.org/all/20220613075958.511064-2-elic@nvidia.com/T/

v0 -> v1:
1. Fix a few typos
2. handle memory map updates for all groups with a given asid.

Eli Cohen (3):
  vdpa/mlx5: Implement susupend virtqueue callback
  vdpa/mlx5: Support different address spaces for control and data
  vdpa/mlx5: Disable VLAN support to support live migration

 drivers/vdpa/mlx5/core/mlx5_vdpa.h |  11 ++
 drivers/vdpa/mlx5/net/mlx5_vnet.c  | 157 ++++++++++++++++++++++++++---
 include/linux/mlx5/mlx5_ifc_vdpa.h |   8 ++
 3 files changed, 163 insertions(+), 13 deletions(-)

-- 
2.35.1

[PATCH RFC v1 1/3] vdpa/mlx5: Implement susupend virtqueue callback

[Eli Cohen]

Implement the suspend callback allowing to suspend the virtqueues so
they stop processing descriptors. This is required to allow to query a
consistent state of the virtqueue while live migration is taking place.

Signed-off-by: Eli Cohen <elic@nvidia.com>
---
 drivers/vdpa/mlx5/net/mlx5_vnet.c  | 68 +++++++++++++++++++++++++++++-
 include/linux/mlx5/mlx5_ifc_vdpa.h |  8 ++++
 2 files changed, 75 insertions(+), 1 deletion(-)

diff --git a/drivers/vdpa/mlx5/net/mlx5_vnet.c b/drivers/vdpa/mlx5/net/mlx5_vnet.c
index fb0b23e71383..ea4bc8a0cd25 100644
--- a/drivers/vdpa/mlx5/net/mlx5_vnet.c
+++ b/drivers/vdpa/mlx5/net/mlx5_vnet.c
@@ -895,6 +895,7 @@ static int create_virtqueue(struct mlx5_vdpa_net *ndev, struct mlx5_vdpa_virtque
 	if (err)
 		goto err_cmd;
 
+	mvq->fw_state = MLX5_VIRTIO_NET_Q_OBJECT_STATE_INIT;
 	kfree(in);
 	mvq->virtq_id = MLX5_GET(general_obj_out_cmd_hdr, out, obj_id);
 
@@ -922,6 +923,7 @@ static void destroy_virtqueue(struct mlx5_vdpa_net *ndev, struct mlx5_vdpa_virtq
 		mlx5_vdpa_warn(&ndev->mvdev, "destroy virtqueue 0x%x\n", mvq->virtq_id);
 		return;
 	}
+	mvq->fw_state = MLX5_VIRTIO_NET_Q_OBJECT_NONE;
 	umems_destroy(ndev, mvq);
 }
 
@@ -1121,6 +1123,20 @@ static int query_virtqueue(struct mlx5_vdpa_net *ndev, struct mlx5_vdpa_virtqueu
 	return err;
 }
 
+static bool is_valid_state_change(int oldstate, int newstate)
+{
+	switch (oldstate) {
+	case MLX5_VIRTIO_NET_Q_OBJECT_STATE_INIT:
+		return newstate == MLX5_VIRTIO_NET_Q_OBJECT_STATE_RDY;
+	case MLX5_VIRTIO_NET_Q_OBJECT_STATE_RDY:
+		return newstate == MLX5_VIRTIO_NET_Q_OBJECT_STATE_SUSPEND;
+	case MLX5_VIRTIO_NET_Q_OBJECT_STATE_SUSPEND:
+	case MLX5_VIRTIO_NET_Q_OBJECT_STATE_ERR:
+	default:
+		return false;
+	}
+}
+
 static int modify_virtqueue(struct mlx5_vdpa_net *ndev, struct mlx5_vdpa_virtqueue *mvq, int state)
 {
 	int inlen = MLX5_ST_SZ_BYTES(modify_virtio_net_q_in);
@@ -1130,6 +1146,12 @@ static int modify_virtqueue(struct mlx5_vdpa_net *ndev, struct mlx5_vdpa_virtque
 	void *in;
 	int err;
 
+	if (mvq->fw_state == MLX5_VIRTIO_NET_Q_OBJECT_NONE)
+		return 0;
+
+	if (!is_valid_state_change(mvq->fw_state, state))
+		return -EINVAL;
+
 	in = kzalloc(inlen, GFP_KERNEL);
 	if (!in)
 		return -ENOMEM;
@@ -1991,6 +2013,7 @@ static void mlx5_vdpa_set_vq_ready(struct vdpa_device *vdev, u16 idx, bool ready
 	struct mlx5_vdpa_dev *mvdev = to_mvdev(vdev);
 	struct mlx5_vdpa_net *ndev = to_mlx5_vdpa_ndev(mvdev);
 	struct mlx5_vdpa_virtqueue *mvq;
+	int err;
 
 	if (!mvdev->actual_features)
 		return;
@@ -2004,8 +2027,16 @@ static void mlx5_vdpa_set_vq_ready(struct vdpa_device *vdev, u16 idx, bool ready
 	}
 
 	mvq = &ndev->vqs[idx];
-	if (!ready)
+	if (!ready) {
 		suspend_vq(ndev, mvq);
+	} else {
+		err = modify_virtqueue(ndev, mvq, MLX5_VIRTIO_NET_Q_OBJECT_STATE_RDY);
+		if (err) {
+			mlx5_vdpa_warn(mvdev, "modify VQ %d to ready failed (%d)\n", idx, err);
+			ready = false;
+		}
+	}
+
 
 	mvq->ready = ready;
 }
@@ -2732,6 +2763,39 @@ static int mlx5_vdpa_get_vendor_vq_stats(struct vdpa_device *vdev, u16 idx,
 	return err;
 }
 
+static void mlx5_vdpa_cvq_suspend(struct mlx5_vdpa_dev *mvdev, bool suspend)
+{
+	struct mlx5_control_vq *cvq;
+
+	if (!(mvdev->actual_features & BIT_ULL(VIRTIO_NET_F_CTRL_VQ)))
+		return;
+
+	cvq = &mvdev->cvq;
+	cvq->ready = !suspend;
+}
+
+static int mlx5_vdpa_suspend(struct vdpa_device *vdev, bool suspend)
+{
+	struct mlx5_vdpa_dev *mvdev = to_mvdev(vdev);
+	struct mlx5_vdpa_net *ndev = to_mlx5_vdpa_ndev(mvdev);
+	struct mlx5_vdpa_virtqueue *mvq;
+	int i;
+
+	if (!suspend) {
+		mlx5_vdpa_warn(mvdev, "Resume of virtqueues is not supported\n");
+		return -EOPNOTSUPP;
+	}
+
+	down_write(&ndev->reslock);
+	for (i = 0; i < ndev->cur_num_vqs; i++) {
+		mvq = &ndev->vqs[i];
+		suspend_vq(ndev, mvq);
+	}
+	mlx5_vdpa_cvq_suspend(mvdev, suspend);
+	up_write(&ndev->reslock);
+	return 0;
+}
+
 static const struct vdpa_config_ops mlx5_vdpa_ops = {
 	.set_vq_address = mlx5_vdpa_set_vq_address,
 	.set_vq_num = mlx5_vdpa_set_vq_num,
@@ -2762,6 +2826,7 @@ static const struct vdpa_config_ops mlx5_vdpa_ops = {
 	.get_generation = mlx5_vdpa_get_generation,
 	.set_map = mlx5_vdpa_set_map,
 	.free = mlx5_vdpa_free,
+	.suspend = mlx5_vdpa_suspend,
 };
 
 static int query_mtu(struct mlx5_core_dev *mdev, u16 *mtu)
@@ -2827,6 +2892,7 @@ static void init_mvqs(struct mlx5_vdpa_net *ndev)
 		mvq->index = i;
 		mvq->ndev = ndev;
 		mvq->fwqp.fw = true;
+		mvq->fw_state = MLX5_VIRTIO_NET_Q_OBJECT_NONE;
 	}
 	for (; i < ndev->mvdev.max_vqs; i++) {
 		mvq = &ndev->vqs[i];
diff --git a/include/linux/mlx5/mlx5_ifc_vdpa.h b/include/linux/mlx5/mlx5_ifc_vdpa.h
index 4414ed5b6ed2..9becdc3fa503 100644
--- a/include/linux/mlx5/mlx5_ifc_vdpa.h
+++ b/include/linux/mlx5/mlx5_ifc_vdpa.h
@@ -150,6 +150,14 @@ enum {
 	MLX5_VIRTIO_NET_Q_OBJECT_STATE_ERR      = 0x3,
 };
 
+/* This indicates that the object was not created or has already
+ * been desroyed. It is very safe to assume that this object will never
+ * have so many states
+ */
+enum {
+	MLX5_VIRTIO_NET_Q_OBJECT_NONE = 0xffffffff
+};
+
 enum {
 	MLX5_RQTC_LIST_Q_TYPE_RQ            = 0x0,
 	MLX5_RQTC_LIST_Q_TYPE_VIRTIO_NET_Q  = 0x1,
-- 
2.35.1

[PATCH RFC v1 2/3] vdpa/mlx5: Support different address spaces for control and data

[Eli Cohen]

Partition virtqueues to two different address spaces: one for control
virtqueue which is implemented in software, and one for data virtqueues.

Based-on: <20220526124338.36247-1-eperezma@redhat.com>
Signed-off-by: Eli Cohen <elic@nvidia.com>
---
 drivers/vdpa/mlx5/core/mlx5_vdpa.h | 11 ++++
 drivers/vdpa/mlx5/net/mlx5_vnet.c  | 88 ++++++++++++++++++++++++++----
 2 files changed, 88 insertions(+), 11 deletions(-)

diff --git a/drivers/vdpa/mlx5/core/mlx5_vdpa.h b/drivers/vdpa/mlx5/core/mlx5_vdpa.h
index 44104093163b..6af9fdbb86b7 100644
--- a/drivers/vdpa/mlx5/core/mlx5_vdpa.h
+++ b/drivers/vdpa/mlx5/core/mlx5_vdpa.h
@@ -70,6 +70,16 @@ struct mlx5_vdpa_wq_ent {
 	struct mlx5_vdpa_dev *mvdev;
 };
 
+enum {
+	MLX5_VDPA_DATAVQ_GROUP,
+	MLX5_VDPA_CVQ_GROUP,
+	MLX5_VDPA_NUMVQ_GROUPS
+};
+
+enum {
+	MLX5_VDPA_NUM_AS = MLX5_VDPA_NUMVQ_GROUPS
+};
+
 struct mlx5_vdpa_dev {
 	struct vdpa_device vdev;
 	struct mlx5_core_dev *mdev;
@@ -85,6 +95,7 @@ struct mlx5_vdpa_dev {
 	struct mlx5_vdpa_mr mr;
 	struct mlx5_control_vq cvq;
 	struct workqueue_struct *wq;
+	unsigned int group2asid[MLX5_VDPA_NUMVQ_GROUPS];
 };
 
 int mlx5_vdpa_alloc_pd(struct mlx5_vdpa_dev *dev, u32 *pdn, u16 uid);
diff --git a/drivers/vdpa/mlx5/net/mlx5_vnet.c b/drivers/vdpa/mlx5/net/mlx5_vnet.c
index ea4bc8a0cd25..cb637dee2ea8 100644
--- a/drivers/vdpa/mlx5/net/mlx5_vnet.c
+++ b/drivers/vdpa/mlx5/net/mlx5_vnet.c
@@ -2125,9 +2125,14 @@ static u32 mlx5_vdpa_get_vq_align(struct vdpa_device *vdev)
 	return PAGE_SIZE;
 }
 
-static u32 mlx5_vdpa_get_vq_group(struct vdpa_device *vdpa, u16 idx)
+static u32 mlx5_vdpa_get_vq_group(struct vdpa_device *vdev, u16 idx)
 {
-	return 0;
+	struct mlx5_vdpa_dev *mvdev = to_mvdev(vdev);
+
+	if (is_ctrl_vq_idx(mvdev, idx))
+		return MLX5_VDPA_CVQ_GROUP;
+
+	return MLX5_VDPA_DATAVQ_GROUP;
 }
 
 enum { MLX5_VIRTIO_NET_F_GUEST_CSUM = 1 << 9,
@@ -2541,6 +2546,15 @@ static void mlx5_vdpa_set_status(struct vdpa_device *vdev, u8 status)
 	up_write(&ndev->reslock);
 }
 
+static void init_group_to_asid_map(struct mlx5_vdpa_dev *mvdev)
+{
+	int i;
+
+	/* default mapping all groups are mapped to asid 0 */
+	for (i = 0; i < MLX5_VDPA_NUMVQ_GROUPS; i++)
+		mvdev->group2asid[i] = 0;
+}
+
 static int mlx5_vdpa_reset(struct vdpa_device *vdev)
 {
 	struct mlx5_vdpa_dev *mvdev = to_mvdev(vdev);
@@ -2559,7 +2573,9 @@ static int mlx5_vdpa_reset(struct vdpa_device *vdev)
 	ndev->mvdev.cvq.completed_desc = 0;
 	memset(ndev->event_cbs, 0, sizeof(*ndev->event_cbs) * (mvdev->max_vqs + 1));
 	ndev->mvdev.actual_features = 0;
+	init_group_to_asid_map(mvdev);
 	++mvdev->generation;
+
 	if (MLX5_CAP_GEN(mvdev->mdev, umem_uid_0)) {
 		if (mlx5_vdpa_create_mr(mvdev, NULL))
 			mlx5_vdpa_warn(mvdev, "create MR failed\n");
@@ -2597,26 +2613,63 @@ static u32 mlx5_vdpa_get_generation(struct vdpa_device *vdev)
 	return mvdev->generation;
 }
 
-static int mlx5_vdpa_set_map(struct vdpa_device *vdev, unsigned int asid,
-			     struct vhost_iotlb *iotlb)
+static int set_map_control(struct mlx5_vdpa_dev *mvdev, struct vhost_iotlb *iotlb)
+{
+	u64 start = 0ULL, last = 0ULL - 1;
+	struct vhost_iotlb_map *map;
+	int err = 0;
+
+	spin_lock(&mvdev->cvq.iommu_lock);
+	vhost_iotlb_reset(mvdev->cvq.iotlb);
+
+	for (map = vhost_iotlb_itree_first(iotlb, start, last); map;
+	     map = vhost_iotlb_itree_next(map, start, last)) {
+		err = vhost_iotlb_add_range(mvdev->cvq.iotlb, map->start,
+					    map->last, map->addr, map->perm);
+		if (err)
+			goto out;
+	}
+
+out:
+	spin_unlock(&mvdev->cvq.iommu_lock);
+	return err;
+}
+
+static int set_map_data(struct mlx5_vdpa_dev *mvdev, struct vhost_iotlb *iotlb)
 {
-	struct mlx5_vdpa_dev *mvdev = to_mvdev(vdev);
-	struct mlx5_vdpa_net *ndev = to_mlx5_vdpa_ndev(mvdev);
 	bool change_map;
 	int err;
 
-	down_write(&ndev->reslock);
-
 	err = mlx5_vdpa_handle_set_map(mvdev, iotlb, &change_map);
 	if (err) {
 		mlx5_vdpa_warn(mvdev, "set map failed(%d)\n", err);
-		goto err;
+		return err;
 	}
 
 	if (change_map)
 		err = mlx5_vdpa_change_map(mvdev, iotlb);
 
-err:
+	return err;
+}
+
+static int mlx5_vdpa_set_map(struct vdpa_device *vdev, unsigned int asid,
+			     struct vhost_iotlb *iotlb)
+{
+	struct mlx5_vdpa_dev *mvdev = to_mvdev(vdev);
+	struct mlx5_vdpa_net *ndev = to_mlx5_vdpa_ndev(mvdev);
+	int err;
+
+	down_write(&ndev->reslock);
+	if (mvdev->group2asid[MLX5_VDPA_DATAVQ_GROUP] == asid) {
+		err = set_map_data(mvdev, iotlb);
+		if (err)
+			goto out;
+	}
+
+	if (mvdev->group2asid[MLX5_VDPA_CVQ_GROUP] == asid)
+		err = set_map_control(mvdev, iotlb);
+
+out:
 	up_write(&ndev->reslock);
 	return err;
 }
@@ -2796,6 +2849,18 @@ static int mlx5_vdpa_suspend(struct vdpa_device *vdev, bool suspend)
 	return 0;
 }
 
+static int mlx5_set_group_asid(struct vdpa_device *vdev, u32 group,
+			       unsigned int asid)
+{
+	struct mlx5_vdpa_dev *mvdev = to_mvdev(vdev);
+
+	if (group >= MLX5_VDPA_NUMVQ_GROUPS)
+		return -EINVAL;
+
+	mvdev->group2asid[group] = asid;
+	return 0;
+}
+
 static const struct vdpa_config_ops mlx5_vdpa_ops = {
 	.set_vq_address = mlx5_vdpa_set_vq_address,
 	.set_vq_num = mlx5_vdpa_set_vq_num,
@@ -2825,6 +2890,7 @@ static const struct vdpa_config_ops mlx5_vdpa_ops = {
 	.set_config = mlx5_vdpa_set_config,
 	.get_generation = mlx5_vdpa_get_generation,
 	.set_map = mlx5_vdpa_set_map,
+	.set_group_asid = mlx5_set_group_asid,
 	.free = mlx5_vdpa_free,
 	.suspend = mlx5_vdpa_suspend,
 };
@@ -3047,7 +3113,7 @@ static int mlx5_vdpa_dev_add(struct vdpa_mgmt_dev *v_mdev, const char *name,
 	}
 
 	ndev = vdpa_alloc_device(struct mlx5_vdpa_net, mvdev.vdev, mdev->device, &mlx5_vdpa_ops,
-				 1, 1, name, false);
+				 MLX5_VDPA_NUMVQ_GROUPS, MLX5_VDPA_NUM_AS, name, false);
 	if (IS_ERR(ndev))
 		return PTR_ERR(ndev);
 
-- 
2.35.1

[PATCH RFC v1 3/3] vdpa/mlx5: Disable VLAN support to support live migration

[Eli Cohen]

Current qemu code does not support live migration for devices supporting
VLAN. Disable it.

Note: this patch is provided just to enable testing with current qemu.

Signed-off-by: Eli Cohen <elic@nvidia.com>
---
 drivers/vdpa/mlx5/net/mlx5_vnet.c | 1 -
 1 file changed, 1 deletion(-)

diff --git a/drivers/vdpa/mlx5/net/mlx5_vnet.c b/drivers/vdpa/mlx5/net/mlx5_vnet.c
index cb637dee2ea8..a468e02fd869 100644
--- a/drivers/vdpa/mlx5/net/mlx5_vnet.c
+++ b/drivers/vdpa/mlx5/net/mlx5_vnet.c
@@ -2172,7 +2172,6 @@ static u64 get_supported_features(struct mlx5_core_dev *mdev)
 	mlx_vdpa_features |= BIT_ULL(VIRTIO_NET_F_MQ);
 	mlx_vdpa_features |= BIT_ULL(VIRTIO_NET_F_STATUS);
 	mlx_vdpa_features |= BIT_ULL(VIRTIO_NET_F_MTU);
-	mlx_vdpa_features |= BIT_ULL(VIRTIO_NET_F_CTRL_VLAN);
 
 	return mlx_vdpa_features;
 }
-- 
2.35.1