Currently, when a new MR is set up, the old MR is deleted. MR deletion
is about 30-40% the time of MR creation. As deleting the old MR is not
important for the process of setting up the new MR, this operation
can be postponed.
This series adds a workqueue that does MR garbage collection at a later
point. If the MR lock is taken, the handler will back off and
reschedule. The exception during shutdown: then the handler must
not postpone the work.
Note that this is only a speculative optimization: if there is some
mapping operation that is triggered while the garbage collector handler
has the lock taken, this operation it will have to wait for the handler
to finish.
Signed-off-by: Dragos Tatulea <dtatulea@nvidia.com>
Reviewed-by: Cosmin Ratiu <cratiu@nvidia.com>
Message-Id: <20240830105838.2666587-9-dtatulea@nvidia.com>
Signed-off-by: Michael S. Tsirkin <mst@redhat.com>
There's currently not a lot of action happening during
the init/destroy of MR resources. But more will be added
in the upcoming patches.
As the mr mutex lock init/destroy has been moved to these
new functions, the lifetime has now shifted away from
mlx5_vdpa_alloc_resources() / mlx5_vdpa_free_resources()
into these new functions. However, the lifetime at the
outer scope remains the same:
mlx5_vdpa_dev_add() / mlx5_vdpa_dev_free()
Signed-off-by: Dragos Tatulea <dtatulea@nvidia.com>
Reviewed-by: Cosmin Ratiu <cratiu@nvidia.com>
Message-Id: <20240830105838.2666587-8-dtatulea@nvidia.com>
Signed-off-by: Michael S. Tsirkin <mst@redhat.com>
change_num_qps() is still suspending/resuming VQs one by one.
This change switches to parallel suspend/resume.
When increasing the number of queues the flow has changed a bit for
simplicity: the setup_vq() function will always be called before
resume_vqs(). If the VQ is initialized, setup_vq() will exit early. If
the VQ is not initialized, setup_vq() will create it and resume_vqs()
will resume it.
Signed-off-by: Dragos Tatulea <dtatulea@nvidia.com>
Reviewed-by: Tariq Toukan <tariqt@nvidia.com>
Message-Id: <20240816090159.1967650-11-dtatulea@nvidia.com>
Signed-off-by: Michael S. Tsirkin <mst@redhat.com>
Acked-by: Eugenio Pérez <eperezma@redhat.com>
Tested-by: Lei Yang <leiyang@redhat.com>
Unregistering notifiers is a costly operation. Instead of removing
the notifiers during device suspend and adding them back at resume,
simply ignore the call when the device is suspended.
At resume time call queue_link_work() to make sure that the device state
is propagated in case there were changes.
For 1 vDPA device x 32 VQs (16 VQPs) attached to a large VM (256 GB RAM,
32 CPUs x 2 threads per core), the device suspend time is reduced from
~13 ms to ~2.5 ms.
Signed-off-by: Dragos Tatulea <dtatulea@nvidia.com>
Reviewed-by: Tariq Toukan <tariqt@nvidia.com>
Acked-by: Eugenio Pérez <eperezma@redhat.com>
Message-Id: <20240816090159.1967650-9-dtatulea@nvidia.com>
Signed-off-by: Michael S. Tsirkin <mst@redhat.com>
Tested-by: Lei Yang <leiyang@redhat.com>
Currently device resume works on vqs serially. Building up on previous
changes that converted vq operations to the async api, this patch
parallelizes the device resume.
For 1 vDPA device x 32 VQs (16 VQPs) attached to a large VM (256 GB RAM,
32 CPUs x 2 threads per core), the device resume time is reduced from
~16 ms to ~4.5 ms.
Signed-off-by: Dragos Tatulea <dtatulea@nvidia.com>
Reviewed-by: Tariq Toukan <tariqt@nvidia.com>
Acked-by: Eugenio Pérez <eperezma@redhat.com>
Message-Id: <20240816090159.1967650-8-dtatulea@nvidia.com>
Signed-off-by: Michael S. Tsirkin <mst@redhat.com>
Tested-by: Lei Yang <leiyang@redhat.com>
Currently device suspend works on vqs serially. Building up on previous
changes that converted vq operations to the async api, this patch
parallelizes the device suspend:
1) Suspend all active vqs parallel.
2) Query suspended vqs in parallel.
For 1 vDPA device x 32 VQs (16 VQPs) attached to a large VM (256 GB RAM,
32 CPUs x 2 threads per core), the device suspend time is reduced from
~37 ms to ~13 ms.
A later patch will remove the link unregister operation which will make
it even faster.
Signed-off-by: Dragos Tatulea <dtatulea@nvidia.com>
Reviewed-by: Tariq Toukan <tariqt@nvidia.com>
Acked-by: Eugenio Pérez <eperezma@redhat.com>
Message-Id: <20240816090159.1967650-7-dtatulea@nvidia.com>
Signed-off-by: Michael S. Tsirkin <mst@redhat.com>
Tested-by: Lei Yang <leiyang@redhat.com>
Switch firmware vq modify command to be issued via the async API to
allow future parallelization. The new refactored function applies the
modify on a range of vqs and waits for their execution to complete.
For now the command is still used in a serial fashion. A later patch
will switch to modifying multiple vqs in parallel.
Signed-off-by: Dragos Tatulea <dtatulea@nvidia.com>
Reviewed-by: Tariq Toukan <tariqt@nvidia.com>
Message-Id: <20240816090159.1967650-6-dtatulea@nvidia.com>
Signed-off-by: Michael S. Tsirkin <mst@redhat.com>
Acked-by: Eugenio Pérez <eperezma@redhat.com>
Tested-by: Lei Yang <leiyang@redhat.com>
Switch firmware vq query command to be issued via the async API to
allow future parallelization.
For now the command is still serial but the infrastructure is there
to issue commands in parallel, including ratelimiting the number
of issued async commands to firmware.
A later patch will switch to issuing more commands at a time.
Signed-off-by: Dragos Tatulea <dtatulea@nvidia.com>
Reviewed-by: Tariq Toukan <tariqt@nvidia.com>
Message-Id: <20240816090159.1967650-5-dtatulea@nvidia.com>
Signed-off-by: Michael S. Tsirkin <mst@redhat.com>
Tested-by: Lei Yang <leiyang@redhat.com>
Add the function to support setting the MAC address.
For vdpa/mlx5, the function will use mlx5_mpfs_add_mac
to set the mac address
Tested in ConnectX-6 Dx device
Signed-off-by: Cindy Lu <lulu@redhat.com>
Message-Id: <20240731031653.1047692-4-lulu@redhat.com>
Signed-off-by: Michael S. Tsirkin <mst@redhat.com>
Acked-by: Jason Wang <jasowang@redhat.com>
The vdpa device can be reset many times in sequence without any
significant state changes in between. Previously this was not a problem:
VQs were torn down only on first reset. But after VQ pre-creation was
introduced, each reset will delete and re-create the hardware VQs and
their associated resources.
To solve this problem, avoid resetting hardware VQs if the VQs are still
in a blank state.
Reviewed-by: Cosmin Ratiu <cratiu@nvidia.com>
Acked-by: Eugenio Pérez <eperezma@redhat.com>
Signed-off-by: Dragos Tatulea <dtatulea@nvidia.com>
Message-Id: <20240626-stage-vdpa-vq-precreate-v2-23-560c491078df@nvidia.com>
Signed-off-by: Michael S. Tsirkin <mst@redhat.com>
There are a few conditions under which the hardware VQs need a full
teardown and setup:
- VQ size changed to something else than default value. Hardware VQ size
modification is not supported.
- User turns off certain device features: mergeable buffers, checksum
virtio 1.0 compliance. In these cases, the TIR and RQT need to be
re-created.
Add a needs_teardown configuration variable and set it when detecting
the above scenarios. On next DRIVER_OK, the resources will be torn down
first.
Reviewed-by: Cosmin Ratiu <cratiu@nvidia.com>
Acked-by: Eugenio Pérez <eperezma@redhat.com>
Signed-off-by: Dragos Tatulea <dtatulea@nvidia.com>
Message-Id: <20240626-stage-vdpa-vq-precreate-v2-22-560c491078df@nvidia.com>
Signed-off-by: Michael S. Tsirkin <mst@redhat.com>
Currently, hardware VQs are created right when the vdpa device gets into
DRIVER_OK state. That is easier because most of the VQ state is known by
then.
This patch switches to creating all VQs and their associated resources
at device creation time. The motivation is to reduce the vdpa device
live migration downtime by moving the expensive operation of creating
all the hardware VQs and their associated resources out of downtime on
the destination VM.
The VQs are now created in a blank state. The VQ configuration will
happen later, on DRIVER_OK. Then the configuration will be applied when
the VQs are moved to the Ready state.
When .set_vq_ready() is called on a VQ before DRIVER_OK, special care is
needed: now that the VQ is already created a resume_vq() will be
triggered too early when no mr has been configured yet. Skip calling
resume_vq() in this case, let it be handled during DRIVER_OK.
For virtio-vdpa, the device configuration is done earlier during
.vdpa_dev_add() by vdpa_register_device(). Avoid calling
setup_vq_resources() a second time in that case.
On a 64 CPU, 256 GB VM with 1 vDPA device of 16 VQps, the full VQ
resource creation + resume time was ~370ms. Now it's down to 60 ms
(only VQ config and resume). The measurements were done on a ConnectX6DX
based vDPA device.
Signed-off-by: Dragos Tatulea <dtatulea@nvidia.com>
Reviewed-by: Cosmin Ratiu <cratiu@nvidia.com>
Message-Id: <20240626-stage-vdpa-vq-precreate-v2-21-560c491078df@nvidia.com>
Signed-off-by: Michael S. Tsirkin <mst@redhat.com>
Resume a VQ if it is already created when the number of VQ pairs
increases. This is done in preparation for VQ pre-creation which is
coming in a later patch. It is necessary because calling setup_vq() on
an already created VQ will return early and will not enable the queue.
For symmetry, suspend a VQ instead of tearing it down when the number of
VQ pairs decreases. But only if the resume operation is supported.
Reviewed-by: Cosmin Ratiu <cratiu@nvidia.com>
Acked-by: Eugenio Pérez <eperezma@redhat.com>
Signed-off-by: Dragos Tatulea <dtatulea@nvidia.com>
Message-Id: <20240626-stage-vdpa-vq-precreate-v2-20-560c491078df@nvidia.com>
Signed-off-by: Michael S. Tsirkin <mst@redhat.com>
There are a few more places modifying the VQ to Ready directly. Let's
consolidate them into resume_vq().
The redundant warnings for resume_vq() errors can also be dropped.
There is one special case that needs to be handled for virtio-vdpa:
the initialized flag must be set to true earlier in setup_vq() so that
resume_vq() doesn't return early.
Reviewed-by: Cosmin Ratiu <cratiu@nvidia.com>
Acked-by: Eugenio Pérez <eperezma@redhat.com>
Signed-off-by: Dragos Tatulea <dtatulea@nvidia.com>
Message-Id: <20240626-stage-vdpa-vq-precreate-v2-18-560c491078df@nvidia.com>
Signed-off-by: Michael S. Tsirkin <mst@redhat.com>
Currently rqt_size is initialized during device flag configuration.
That's because it is the earliest moment when device knows if MQ
(multi queue) is on or off.
Shift this configuration earlier to device creation time. This implies
that non-MQ devices will have a larger RQT size. But the configuration
will still be correct.
This is done in preparation for the pre-creation of hardware virtqueues
at device add time. When that change will be added, RQT will be created
at device creation time so it needs to be initialized to its max size.
Reviewed-by: Cosmin Ratiu <cratiu@nvidia.com>
Acked-by: Eugenio Pérez <eperezma@redhat.com>
Signed-off-by: Dragos Tatulea <dtatulea@nvidia.com>
Message-Id: <20240626-stage-vdpa-vq-precreate-v2-13-560c491078df@nvidia.com>
Signed-off-by: Michael S. Tsirkin <mst@redhat.com>
The virtqueue size is a pre-requisite for setting up any virtqueue
resources. For the upcoming optimization of creating virtqueues at
device add, the virtqueue size has to be configured.
The queue size check in setup_vq() will always be false. So remove it.
Signed-off-by: Dragos Tatulea <dtatulea@nvidia.com>
Reviewed-by: Cosmin Ratiu <cratiu@nvidia.com>
Message-Id: <20240626-stage-vdpa-vq-precreate-v2-12-560c491078df@nvidia.com>
Signed-off-by: Michael S. Tsirkin <mst@redhat.com>
The hardware VQ configuration is mirrored by data in struct
mlx5_vdpa_virtqueue . Instead of clearing just a few fields at reset,
fully clear the struct and initialize with the appropriate default
values.
As clear_vqs_ready() is used only during reset, get rid of it.
Reviewed-by: Cosmin Ratiu <cratiu@nvidia.com>
Acked-by: Eugenio Pérez <eperezma@redhat.com>
Signed-off-by: Dragos Tatulea <dtatulea@nvidia.com>
Message-Id: <20240626-stage-vdpa-vq-precreate-v2-8-560c491078df@nvidia.com>
Signed-off-by: Michael S. Tsirkin <mst@redhat.com>
The virtio spec says that a vdpa device should start off with one queue
pair. The driver is already compliant.
This patch moves the initialization to device add and reset times. This
is done in preparation for the pre-creation of hardware virtqueues at
device add time.
Signed-off-by: Dragos Tatulea <dtatulea@nvidia.com>
Reviewed-by: Cosmin Ratiu <cratiu@nvidia.com>
Message-Id: <20240626-stage-vdpa-vq-precreate-v2-7-560c491078df@nvidia.com>
Signed-off-by: Michael S. Tsirkin <mst@redhat.com>
Acked-by: Eugenio Pérez <eperezma@redhat.com>
The MLX driver was not updating its control virtqueue size at set_vq_num
and instead always initialized to MLX5_CVQ_MAX_ENT (16) at
setup_cvq_vring.
Qemu would try to set the size to 64 by default, however, because the
CVQ size always was initialized to 16, an error would be thrown when
sending >16 control messages (as used-ring entry 17 is initialized to 0).
For example, starting a guest with x-svq=on and then executing the
following command would produce the error below:
# for i in {1..20}; do ifconfig eth0 hw ether XX:xx:XX:xx:XX:XX; done
qemu-system-x86_64: Insufficient written data (0)
[ 435.331223] virtio_net virtio0: Failed to set mac address by vq command.
SIOCSIFHWADDR: Invalid argument
Acked-by: Dragos Tatulea <dtatulea@nvidia.com>
Acked-by: Eugenio Pérez <eperezma@redhat.com>
Signed-off-by: Jonah Palmer <jonah.palmer@oracle.com>
Message-Id: <20240216142502.78095-1-jonah.palmer@oracle.com>
Signed-off-by: Michael S. Tsirkin <mst@redhat.com>
Tested-by: Lei Yang <leiyang@redhat.com>
Fixes: 5262912ef3 ("vdpa/mlx5: Add support for control VQ and MAC setting")
Deleting the old mr during mr update (.set_map) and then modifying the
vqs with the new mr is not a good flow for firmware. The firmware
expects that mkeys are deleted after there are no more vqs referencing
them.
Introduce reference counting for mrs to fix this. It is the only way to
make sure that mkeys are not in use by vqs.
An mr reference is taken when the mr is associated to the mr asid table
and when the mr is linked to the vq on create/modify. The reference is
released when the mkey is unlinked from the vq (trough modify/destroy)
and from the mr asid table.
To make things consistent, get rid of mlx5_vdpa_destroy_mr and use
get/put semantics everywhere.
Reviewed-by: Gal Pressman <gal@nvidia.com>
Acked-by: Eugenio Pérez <eperezma@redhat.com>
Acked-by: Jason Wang <jasowang@redhat.com>
Signed-off-by: Dragos Tatulea <dtatulea@nvidia.com>
Message-Id: <20231225151203.152687-8-dtatulea@nvidia.com>
Signed-off-by: Michael S. Tsirkin <mst@redhat.com>
Instead of tearing down and setting up vq resources, use vq
suspend/resume during .set_map to speed things up a bit.
The vq mr is updated with the new mapping while the vqs are suspended.
If the device doesn't support resumable vqs, do the old teardown and
setup dance.
Reviewed-by: Gal Pressman <gal@nvidia.com>
Acked-by: Eugenio Pérez <eperezma@redhat.com>
Acked-by: Jason Wang <jasowang@redhat.com>
Signed-off-by: Dragos Tatulea <dtatulea@nvidia.com>
Message-Id: <20231225151203.152687-7-dtatulea@nvidia.com>
Signed-off-by: Michael S. Tsirkin <mst@redhat.com>
.set_vq_state will set the indices and mark the fields to be modified in
the hw vq.
Advertise that the device supports changing the vq state when the device
is in DRIVER_OK state and suspended.
Reviewed-by: Gal Pressman <gal@nvidia.com>
Signed-off-by: Dragos Tatulea <dtatulea@nvidia.com>
Acked-by: Jason Wang <jasowang@redhat.com>
Message-Id: <20231225151203.152687-6-dtatulea@nvidia.com>
Signed-off-by: Michael S. Tsirkin <mst@redhat.com>
mlx5_vdpa does not preserve userland's view of vring base for the control
queue in the following sequence:
ioctl VHOST_SET_VRING_BASE
ioctl VHOST_VDPA_SET_STATUS VIRTIO_CONFIG_S_DRIVER_OK
mlx5_vdpa_set_status()
setup_cvq_vring()
vringh_init_iotlb()
vringh_init_kern()
vrh->last_avail_idx = 0;
ioctl VHOST_GET_VRING_BASE
To fix, restore the value of cvq->vring.last_avail_idx after calling
vringh_init_iotlb.
Fixes: 5262912ef3 ("vdpa/mlx5: Add support for control VQ and MAC setting")
Signed-off-by: Steve Sistare <steven.sistare@oracle.com>
Acked-by: Eugenio Pérez <eperezma@redhat.com>
Acked-by: Jason Wang <jasowang@redhat.com>
Message-Id: <1699014387-194368-1-git-send-email-steven.sistare@oracle.com>
Signed-off-by: Michael S. Tsirkin <mst@redhat.com>
Since commit 6f5312f801 ("vdpa/mlx5: Add support for running with
virtio_vdpa"), mlx5_vdpa starts with preallocate 1:1 DMA MR at device
creation time. This 1:1 DMA MR will be implicitly destroyed while the
first .set_map call is invoked, in which case callers like vhost-vdpa
will start to set up custom mappings. When the .reset callback is
invoked, the custom mappings will be cleared and the 1:1 DMA MR will be
re-created.
In order to reduce excessive memory mapping cost in live migration, it
is desirable to decouple the vhost-vdpa IOTLB abstraction from the
virtio device life cycle, i.e. mappings can be kept around intact across
virtio device reset. Leverage the .reset_map callback, which is meant to
destroy the regular MR (including cvq mapping) on the given ASID and
recreate the initial DMA mapping. That way, the device .reset op runs
free from having to maintain and clean up memory mappings by itself.
Additionally, implement .compat_reset to cater for older userspace,
which may wish to see mapping to be cleared during reset.
Co-developed-by: Dragos Tatulea <dtatulea@nvidia.com>
Signed-off-by: Dragos Tatulea <dtatulea@nvidia.com>
Signed-off-by: Si-Wei Liu <si-wei.liu@oracle.com>
Message-Id: <1697880319-4937-7-git-send-email-si-wei.liu@oracle.com>
Signed-off-by: Michael S. Tsirkin <mst@redhat.com>
Tested-by: Lei Yang <leiyang@redhat.com>
Offer this backend feature as mlx5 is compatible with it. It allows it
to do live migration with CVQ, dynamically switching between passthrough
and shadow virtqueue.
Signed-off-by: Eugenio Pérez <eperezma@redhat.com>
Message-Id: <20230703142514.363256-1-eperezma@redhat.com>
Signed-off-by: Michael S. Tsirkin <mst@redhat.com>
For the following sequence:
- cvq group is in ASID 0
- .set_map(1, cvq_iotlb)
- .set_group_asid(cvq_group, 1)
... the cvq mapping from ASID 0 will be used. This is not always correct
behaviour.
This patch adds support for the above mentioned flow by saving the iotlb
on each .set_map and updating the cvq iotlb with it on a cvq group change.
Acked-by: Jason Wang <jasowang@redhat.com>
Acked-by: Eugenio Pérez <eperezma@redhat.com>
Signed-off-by: Dragos Tatulea <dtatulea@nvidia.com>
Message-Id: <20231018171456.1624030-18-dtatulea@nvidia.com>
Signed-off-by: Michael S. Tsirkin <mst@redhat.com>
Reviewed-by: Si-Wei Liu <si-wei.liu@oracle.com>
Tested-by: Si-Wei Liu <si-wei.liu@oracle.com>
Tested-by: Lei Yang <leiyang@redhat.com>
