Enable vdpa net migration with features depending on CVQ
cover Enable vdpa net migration with features depending on CVQ
At this moment the migration of net features that depends on CVQ is not possible, as there is no reliable way to restore the device state like mac address, number of enabled queues, etc to the destination. This is mainly caused because the device must only read CVQ, and process all the commands before resuming the dataplane.
This series lift that requirement, sending the VHOST_VDPA_SET_VRING_ENABLE ioctl for dataplane vqs only after the device has processed all commands.
---
From FRC:
* Enable vqs early in case CVQ cannot be shadowed.
Eugenio Pérez (7):
vdpa: export vhost_vdpa_set_vring_ready
vdpa: add should_enable op
vdpa: use virtio_ops->should_enable at vhost_vdpa_set_vrings_ready
vdpa: add stub vhost_vdpa_should_enable
vdpa: delay enable of data vqs
vdpa: enable cvq svq if data vq are shadowed
vdpa: remove net cvq migration blocker
include/hw/virtio/vhost-vdpa.h | 9 +++++ hw/virtio/vhost-vdpa.c | 33 ++++++++++++---- net/vhost-vdpa.c | 69 ++++++++++++++++++++++++++-------- hw/virtio/trace-events | 2 +- 4 files changed, 89 insertions(+), 24 deletions(-)
Signature · 1 line
2.39.3
PATCH 1/7 vdpa: export vhost_vdpa_set_vring_ready
The vhost-vdpa net backend needs to enable vrings in a different order than default, so export it.
No functional change intended except for tracing, that now includes the (virtio) index being enabled and the return value of the ioctl.
Still ignoring return value of this function if called from vhost_vdpa_dev_start, as reorganize calling code around it is out of the scope of this series.
--- include/hw/virtio/vhost-vdpa.h | 1 + hw/virtio/vhost-vdpa.c | 28 ++++++++++++++++++++-------- hw/virtio/trace-events | 2 +- 3 files changed, 22 insertions(+), 9 deletions(-)
include/hw/virtio/vhost-vdpa.h +1 -0
@@ -57,6 +57,7 @@ typedef struct vhost_vdpa { } VhostVDPA;int vhost_vdpa_get_iova_range (int fd,struct vhost_vdpa_iova_range *iova_range); +int vhost_vdpa_set_vring_ready (struct vhost_vdpa *v,unsigned idx);int vhost_vdpa_dma_map (struct vhost_vdpa *v,uint32_t asid,hwaddr iova,hwaddr size,void *vaddr,bool readonly);
hw/virtio/vhost-vdpa.c +20 -8
@@ -532,6 +532,19 @@ int vhost_vdpa_get_iova_range(int fd, struct vhost_vdpa_iova_range *iova_range)return ret <0 ? -errno :0 ; } +int vhost_vdpa_set_vring_ready (struct vhost_vdpa *v,unsigned idx) +{ +struct vhost_dev *dev = v->dev; +struct vhost_vring_state state = { + .index = idx, + .num =1 , + }; +int r =vhost_vdpa_call (dev,VHOST_VDPA_SET_VRING_ENABLE , &state); + +trace_vhost_vdpa_set_vring_ready (dev, idx, r); +return r; +} +/* * The use of this function is for requests that only need to be * applied once. Typically such request occurs at the beginning
@@ -876,16 +889,15 @@ static int vhost_vdpa_get_vq_index(struct vhost_dev *dev, int idx)return idx; } -static int vhost_vdpa_set_vring_ready (struct vhost_dev *dev) +static int vhost_vdpa_set_vrings_ready (struct vhost_dev *dev) { +struct vhost_vdpa *v = dev->opaque;int i; -trace_vhost_vdpa_set_vring_ready (dev); + +assert (dev->vhost_ops->backend_type ==VHOST_BACKEND_TYPE_VDPA ); +for (i =0 ; i < dev->nvqs; ++i) { -struct vhost_vring_state state = { - .index = dev->vq_index + i, - .num =1 , - }; -vhost_vdpa_call (dev,VHOST_VDPA_SET_VRING_ENABLE , &state); +vhost_vdpa_set_vring_ready (v, dev->vq_index + i); }return 0 ; }
@@ -1298,7 +1310,7 @@ static int vhost_vdpa_dev_start(struct vhost_dev *dev, bool started)if (unlikely (!ok)) {return -1 ; } -vhost_vdpa_set_vring_ready (dev); +vhost_vdpa_set_vrings_ready (dev); }else {vhost_vdpa_suspend (dev);vhost_vdpa_svqs_stop (dev);
hw/virtio/trace-events +1 -1
@@ -48,7 +48,7 @@ vhost_vdpa_set_features(void *dev, uint64_t features) "dev: %p features: 0x%"PRI vhost_vdpa_get_device_id(void *dev, uint32_t device_id) "dev: %p device_id %"PRIu32 vhost_vdpa_reset_device(void *dev) "dev: %p" vhost_vdpa_get_vq_index(void *dev, int idx, int vq_idx) "dev: %p idx: %d vq idx: %d" -vhost_vdpa_set_vring_ready(void *dev) "dev: %p" +vhost_vdpa_set_vring_ready(void *dev, unsigned i, int r) "dev: %p, idx: %u, r: %d" vhost_vdpa_dump_config(void *dev, const char *line) "dev: %p %s" vhost_vdpa_set_config(void *dev, uint32_t offset, uint32_t size, uint32_t flags) "dev: %p offset: %"PRIu32" size: %"PRIu32" flags: 0x%"PRIx32 vhost_vdpa_get_config(void *dev, void *config, uint32_t config_len) "dev: %p config: %p config_len: %"PRIu32
Signature · 1 line
2.39.3
PATCH 2/7 vdpa: add should_enable op
In reply to Eugenio PérezTo restore the device at the destination of a live migration we send the commands through control virtqueue. For a device to read CVQ it must have received the DRIVER_OK status bit.
However this opens a window where the device could start receiving packets in rx queue 0 before it receives the RSS configuration. To avoid that, we will not send vring_enable until all configuration is used by the device.
As a first step, enable a new vitio ops per vhost_vdpa device to know if we should enable a virtqueue or not. This srtuct can be reused in the future to add more actions to vhost_vdpa that depend on the virtIO kind of device.
--- include/hw/virtio/vhost-vdpa.h | 8 ++++++++ 1 file changed, 8 insertions(+)
include/hw/virtio/vhost-vdpa.h +8 -0
@@ -30,6 +30,13 @@ typedef struct VhostVDPAHostNotifier {void *addr; } VhostVDPAHostNotifier; +struct vhost_vdpa ; +typedef bool (*vhost_vdpa_virtio_should_enable_op )(const struct vhost_vdpa *v); + +typedef struct VhostVDPAVirtIOOps { +vhost_vdpa_virtio_should_enable_op should_enable; +}VhostVDPAVirtIOOps ; + typedef struct vhost_vdpa { int device_fd; int index;
@@ -48,6 +55,7 @@ typedef struct vhost_vdpa {VhostIOVATree *iova_tree;GPtrArray *shadow_vqs;const VhostShadowVirtqueueOps *shadow_vq_ops; +const VhostVDPAVirtIOOps *virtio_ops;void *shadow_vq_ops_opaque;struct vhost_dev *dev;Error *migration_blocker;
Signature · 1 line
2.39.3
PATCH 3/7 vdpa: use virtio_ops->should_enable at vhost_vdpa_set_vrings_ready
In reply to Eugenio PérezThis allow to skip some rings that qemu does not want to enable.
--- hw/virtio/vhost-vdpa.c | 5 +++++ 1 file changed, 5 insertions(+)
hw/virtio/vhost-vdpa.c +5 -0
@@ -896,6 +896,11 @@ static int vhost_vdpa_set_vrings_ready(struct vhost_dev *dev)assert (dev->vhost_ops->backend_type ==VHOST_BACKEND_TYPE_VDPA ); +if (v->virtio_ops && v->virtio_ops->should_enable + && !(v->virtio_ops->should_enable (v))) { +return 0 ; + } +for (i =0 ; i < dev->nvqs; ++i) {vhost_vdpa_set_vring_ready (v, dev->vq_index + i); }
Signature · 1 line
2.39.3
PATCH 4/7 vdpa: add stub vhost_vdpa_should_enable
In reply to Eugenio PérezTo restore the device at the destination of a live migration we send the commands through control virtqueue. For a device to read CVQ it must have received the DRIVER_OK status bit.
However this opens a window where the device could start receiving packets in rx queue 0 before it receives the RSS configuration. To avoid that, we will not send vring_enable until all configuration is used by the device.
Add vhost-vdpa net vhost_vdpa_should_enable. Do not change the behavior in this commit, only introduce the op.
--- net/vhost-vdpa.c | 10 ++++++++++ 1 file changed, 10 insertions(+)
net/vhost-vdpa.c +10 -0
@@ -1255,6 +1255,15 @@ static const VhostShadowVirtqueueOps vhost_vdpa_net_svq_ops = { .avail_handler = vhost_vdpa_net_handle_ctrl_avail, }; +static bool vhost_vdpa_should_enable (const struct vhost_vdpa *v) +{ +return true; +} + +static const VhostVDPAVirtIOOps vhost_vdpa_virtio_net_ops = { + .should_enable = vhost_vdpa_should_enable, +}; +/** * Probe if CVQ is isolated *
@@ -1378,6 +1387,7 @@ static NetClientState *net_vhost_vdpa_init(NetClientState *peer, s->vhost_vdpa.shadow_vqs_enabled = svq; s->vhost_vdpa.iova_range = iova_range; s->vhost_vdpa.shadow_data = svq; + s->vhost_vdpa.virtio_ops = &vhost_vdpa_virtio_net_ops;if (queue_pair_index ==0 ) {vhost_vdpa_net_valid_svq_features (features, &s->vhost_vdpa.migration_blocker);
Signature · 1 line
2.39.3
PATCH 5/7 vdpa: delay enable of data vqs
In reply to Eugenio PérezTo restore the device at the destination of a live migration we send the commands through control virtqueue. For a device to read CVQ it must have received the DRIVER_OK status bit.
However this opens a window where the device could start receiving packets in rx queue 0 before it receives the RSS configuration. To avoid that, we do not send vring_enable until all configuration is used by the device.
---
v2: Enable the dataplane vqs if cannot shadow CVQ because of device features or ASID.
--- net/vhost-vdpa.c | 44 +++++++++++++++++++++++++++++++++++++++++++- 1 file changed, 43 insertions(+), 1 deletion(-)
net/vhost-vdpa.c +43 -1
@@ -283,6 +283,15 @@ static VhostVDPAState *vhost_vdpa_net_first_nc_vdpa(VhostVDPAState *s)return DO_UPCAST (VhostVDPAState, nc, nc0); } +/** From any vdpa net client, get the netclient of the last queue pair */ +static VhostVDPAState *vhost_vdpa_net_last_nc_vdpa (VhostVDPAState *s) +{ +VirtIONet *n =qemu_get_nic_opaque (s->nc.peer); +NetClientState *nc =qemu_get_peer (n->nic->ncs, n->max_ncs -1 ); + +return DO_UPCAST (VhostVDPAState, nc, nc); +} +static void vhost_vdpa_net_log_global_enable (VhostVDPAState *s,bool enable) {struct vhost_vdpa *v = &s->vhost_vdpa;
@@ -996,6 +1005,13 @@ static int vhost_vdpa_net_load(NetClientState *nc)return r; } +for (int i =0 ; i < v->dev->vq_index; ++i) { + r =vhost_vdpa_set_vring_ready (v, i); +if (unlikely (r)) { +return r; + } + } +return 0 ; }
@@ -1255,9 +1271,35 @@ static const VhostShadowVirtqueueOps vhost_vdpa_net_svq_ops = { .avail_handler = vhost_vdpa_net_handle_ctrl_avail, }; +/** +* Check if a vhost_vdpa device should enable before DRIVER_OK +* +* CVQ must always start first if we want to restore the state safely. Do not +* start data vqs if the device has CVQ. +*/ static bool vhost_vdpa_should_enable (const struct vhost_vdpa *v) { -return true; +struct vhost_dev *dev = v->dev; +VhostVDPAState *s =container_of (v, VhostVDPAState, vhost_vdpa); +VhostVDPAState *cvq_s =vhost_vdpa_net_last_nc_vdpa (s); + +if (!(dev->vq_index_end %2 )) { +/* vDPA device does not have CVQ */ +return true; + } + +if (dev->vq_index +1 == dev->vq_index_end) { +/* We're evaluating CVQ, that must always enable first */ +return true; + } + +if (!vhost_vdpa_net_valid_svq_features (v->dev->features,NULL ) || + !cvq_s->cvq_isolated) { +/* CVQ is not isolated, so let's enable as usual */ +return true; + } + +return false; }static const VhostVDPAVirtIOOps vhost_vdpa_virtio_net_ops = {
Signature · 1 line
2.39.3
PATCH 6/7 vdpa: enable cvq svq if data vq are shadowed
In reply to Eugenio PérezPrevious to this commit, it was assumed that data can only be shadowed with x-cvq, or if a migration was in place. So CVQ did not need to check for migration.
--- net/vhost-vdpa.c | 5 ++--- 1 file changed, 2 insertions(+), 3 deletions(-)
net/vhost-vdpa.c +2 -3
@@ -514,11 +514,10 @@ static int vhost_vdpa_net_cvq_start(NetClientState *nc) s0 =vhost_vdpa_net_first_nc_vdpa (s); v->shadow_data = s0->vhost_vdpa.shadow_vqs_enabled; - v->shadow_vqs_enabled = s->always_svq; + v->shadow_vqs_enabled = v->shadow_data; s->vhost_vdpa.address_space_id =VHOST_VDPA_GUEST_PA_ASID ; -if (s->vhost_vdpa.shadow_data) { -/* SVQ is already configured for all virtqueues */ +if (s->always_svq) { goto out; }
Signature · 1 line
2.39.3
PATCH 7/7 vdpa: remove net cvq migration blocker
In reply to Eugenio PérezNow that we have add migration blockers if the device does not support all the needed features, remove the general blocker applied to all net devices with CVQ.
--- net/vhost-vdpa.c | 12 ------------ 1 file changed, 12 deletions(-)
net/vhost-vdpa.c +0 -12
@@ -1443,18 +1443,6 @@ static NetClientState *net_vhost_vdpa_init(NetClientState *peer, s->vhost_vdpa.shadow_vq_ops = &vhost_vdpa_net_svq_ops; s->vhost_vdpa.shadow_vq_ops_opaque = s; s->cvq_isolated = cvq_isolated; - -/* -* TODO: We cannot migrate devices with CVQ and no x-svq enabled as -* there is no way to set the device state (MAC, MQ, etc) before -* starting the datapath. -* -* Migration blocker ownership now belongs to s->vhost_vdpa. -*/ -if (!svq) { -error_setg (&s->vhost_vdpa.migration_blocker, -"net vdpa cannot migrate with CVQ feature" ); - } } ret =vhost_vdpa_add (nc, (void *)&s->vhost_vdpa, queue_pair_index, nvqs);if (ret) {
Signature · 1 line
2.39.3
Quoted 9 lines · Eugenio Pérez
At this moment the migration of net features that depends on CVQ is not possible, as there is no reliable way to restore the device state like mac address, number of enabled queues, etc to the destination. This is mainly caused because the device must only read CVQ, and process all the commands before resuming the dataplane.
This series lift that requirement, sending the
VHOST_VDPA_SET_VRING_ENABLEioctl for dataplane vqs only after the device has processed all commands.
I think it's better to explain (that is what I don't understand) why we can not simply reorder vhost_net_start_one() in vhost_net_start()?
for (i = 0; i < nvhosts; i++) {
if (i < data_queue_pairs) {
peer = qemu_get_peer(ncs, i);
} else {
peer = qemu_get_peer(ncs, n->max_queue_pairs);
}
if (peer->vring_enable) {
/* restore vring enable state */
r = vhost_set_vring_enable(peer, peer->vring_enable);
if (r < 0) {
goto err_start;
}
}=> r = vhost_net_start_one(get_vhost_net(peer), dev);
if (r < 0) {
goto err_start;
}
}Can we simply start cvq first here?
Thanks
Quoted 23 lines
---
From FRC:
* Enable vqs early in case CVQ cannot be shadowed.Eugenio Pérez (7):
vdpa: exportvhost_vdpa_set_vring_ready
vdpa: addshould_enableop
vdpa: usevirtio_ops->should_enableatvhost_vdpa_set_vrings_ready
vdpa: add stubvhost_vdpa_should_enable
vdpa: delay enable of data vqs
vdpa: enable cvq svq if data vq are shadowed
vdpa: remove net cvq migration blockerinclude/hw/virtio/vhost-vdpa.h | 9 +++++ hw/virtio/vhost-vdpa.c | 33 ++++++++++++---- net/vhost-vdpa.c | 69 ++++++++++++++++++++++++++-------- hw/virtio/trace-events | 2 +- 4 files changed, 89 insertions(+), 24 deletions(-)--
2.39.3
Quoted 31 lines · Jason Wang
Quoted 9 lines · Eugenio Pérez
At this moment the migration of net features that depends on CVQ is not possible, as there is no reliable way to restore the device state like mac address, number of enabled queues, etc to the destination. This is mainly caused because the device must only read CVQ, and process all the commands before resuming the dataplane.
This series lift that requirement, sending the
VHOST_VDPA_SET_VRING_ENABLEioctl for dataplane vqs only after the device has processed all commands.I think it's better to explain (that is what I don't understand) why we can not simply reorder
vhost_net_start_one()invhost_net_start()?for (i = 0; i < nvhosts; i++) { if (i < data_queue_pairs) { peer = qemu_get_peer(ncs, i); } else { peer = qemu_get_peer(ncs, n->max_queue_pairs); } if (peer->vring_enable) { /* restore vring enable state */ r = vhost_set_vring_enable(peer, peer->vring_enable); if (r < 0) { goto err_start; } }
Quoted 8 lines
=> r = vhost_net_start_one(get_vhost_net(peer), dev); if (r < 0) { goto err_start; } }Can we simply start cvq first here?
Well the current order is:
* set dev features (conditioned by
* Configure all vq addresses
* Configure all vq size
...
* Enable cvq
* DRIVER_OK
* Enable all the rest of the queues.
If we just start CVQ first, we need to modify vhost_vdpa_set_features as minimum. A lot of code that depends on vdev->vq_index{,_end} may be affected.
Also, I'm not sure if all the devices will support configure address, vq size, etc after DRIVER_OK.
Quoted 26 lines
Thanks
Quoted 23 lines
---
From FRC:
* Enable vqs early in case CVQ cannot be shadowed.Eugenio Pérez (7):
vdpa: exportvhost_vdpa_set_vring_ready
vdpa: addshould_enableop
vdpa: usevirtio_ops->should_enableatvhost_vdpa_set_vrings_ready
vdpa: add stubvhost_vdpa_should_enable
vdpa: delay enable of data vqs
vdpa: enable cvq svq if data vq are shadowed
vdpa: remove net cvq migration blockerinclude/hw/virtio/vhost-vdpa.h | 9 +++++ hw/virtio/vhost-vdpa.c | 33 ++++++++++++---- net/vhost-vdpa.c | 69 ++++++++++++++++++++++++++-------- hw/virtio/trace-events | 2 +- 4 files changed, 89 insertions(+), 24 deletions(-)--
2.39.3
Quoted 51 lines · Eugenio Perez Martin
Quoted 31 lines · Jason Wang
Quoted 9 lines · Eugenio Pérez
At this moment the migration of net features that depends on CVQ is not possible, as there is no reliable way to restore the device state like mac address, number of enabled queues, etc to the destination. This is mainly caused because the device must only read CVQ, and process all the commands before resuming the dataplane.
This series lift that requirement, sending the
VHOST_VDPA_SET_VRING_ENABLEioctl for dataplane vqs only after the device has processed all commands.I think it's better to explain (that is what I don't understand) why we can not simply reorder
vhost_net_start_one()invhost_net_start()?for (i = 0; i < nvhosts; i++) { if (i < data_queue_pairs) { peer = qemu_get_peer(ncs, i); } else { peer = qemu_get_peer(ncs, n->max_queue_pairs); } if (peer->vring_enable) { /* restore vring enable state */ r = vhost_set_vring_enable(peer, peer->vring_enable); if (r < 0) { goto err_start; } }Quoted 8 lines
=> r = vhost_net_start_one(get_vhost_net(peer), dev); if (r < 0) { goto err_start; } }Can we simply start cvq first here?
Well the current order is:
* set dev features (conditioned by
* Configure all vq addresses
* Configure all vq size
...
* Enable cvq
*DRIVER_OK
* Enable all the rest of the queues.
Quoted 6 lines
If we just start CVQ first, we need to modify
vhost_vdpa_set_featuresas minimum. A lot of code that depends on vdev->vq_index{,_end} may be affected.Also, I'm not sure if all the devices will support configure address, vq size, etc after
DRIVER_OK.
Ok, so basically what I meant is to seek a way to refactor vhost_net_start() instead of introducing new ops (e.g introducing virtio ops in vhost seems a layer violation anyhow).
Can we simply factor VRING_ENABLE out and then we can enable vring in any order as we want in vhost_net_start()?
Thanks
Quoted 28 lines
Quoted 27 lines
Thanks
Quoted 23 lines
---
From FRC:
* Enable vqs early in case CVQ cannot be shadowed.Eugenio Pérez (7):
vdpa: exportvhost_vdpa_set_vring_ready
vdpa: addshould_enableop
vdpa: usevirtio_ops->should_enableatvhost_vdpa_set_vrings_ready
vdpa: add stubvhost_vdpa_should_enable
vdpa: delay enable of data vqs
vdpa: enable cvq svq if data vq are shadowed
vdpa: remove net cvq migration blockerinclude/hw/virtio/vhost-vdpa.h | 9 +++++ hw/virtio/vhost-vdpa.c | 33 ++++++++++++---- net/vhost-vdpa.c | 69 ++++++++++++++++++++++++++-------- hw/virtio/trace-events | 2 +- 4 files changed, 89 insertions(+), 24 deletions(-)--
2.39.3