On Sun, Feb 27, 2022 at 9:42 PM Eugenio P?rez <eperezma at redhat.com>
wrote:>
> This series enable shadow virtqueue (SVQ) for vhost-vdpa devices. This
> is intended as a new method of tracking the memory the devices touch
> during a migration process: Instead of relay on vhost device's dirty
> logging capability, SVQ intercepts the VQ dataplane forwarding the
> descriptors between VM and device. This way qemu is the effective
> writer of guests memory, like in qemu's virtio device operation.
>
> When SVQ is enabled qemu offers a new virtual address space to the
> device to read and write into, and it maps new vrings and the guest
> memory in it. SVQ also intercepts kicks and calls between the device
> and the guest. Used buffers relay would cause dirty memory being
> tracked.
>
> This effectively means that vDPA device passthrough is intercepted by
> qemu. While SVQ should only be enabled at migration time, the switching
> from regular mode to SVQ mode is left for a future series.
>
> It is based on the ideas of DPDK SW assisted LM, in the series of
> DPDK's https://patchwork.dpdk.org/cover/48370/ . However, these does
> not map the shadow vq in guest's VA, but in qemu's.
>
> For qemu to use shadow virtqueues the guest virtio driver must not use
> features like event_idx, indirect descriptors, packed and in_order.
> These features are easy to implement on top of this base, but is left
> for a future series for simplicity.
>
> SVQ needs to be enabled at qemu start time with vdpa cmdline parameter:
>
> -netdev type=vhost-vdpa,vhostdev=vhost-vdpa-0,id=vhost-vdpa0,x-svq=off
>
> The first three patches enables notifications forwarding with
> assistance of qemu. It's easy to enable only this if the relevant
> cmdline part of the last patch is applied on top of these.
>
> Next four patches implement the actual buffer forwarding. However,
> address are not translated from HVA so they will need a host device with
> an iommu allowing them to access all of the HVA range.
>
> The last part of the series uses properly the host iommu, so qemu
> creates a new iova address space in the device's range and translates
> the buffers in it. Finally, it adds the cmdline parameter.
>
> Some simple performance tests with netperf were done. They used a nested
> guest with vp_vdpa, vhost-kernel at L0 host. Starting with no svq and a
> baseline average of ~9980.13Mbps:
> Recv Send Send
> Socket Socket Message Elapsed
> Size Size Size Time Throughput
> bytes bytes bytes secs. 10^6bits/sec
>
> 131072 16384 16384 30.01 9910.61
> 131072 16384 16384 30.00 10030.94
> 131072 16384 16384 30.01 9998.84
>
> To enable the notifications interception reduced performance to an
> average of ~9577.73Mbit/s:
> Recv Send Send
> Socket Socket Message Elapsed
> Size Size Size Time Throughput
> bytes bytes bytes secs. 10^6bits/sec
>
> 131072 16384 16384 30.00 9563.03
> 131072 16384 16384 30.01 9626.65
> 131072 16384 16384 30.01 9543.51
>
> Finally, to enable buffers forwarding reduced the throughput again to
> ~8902.92Mbit/s:
> Recv Send Send
> Socket Socket Message Elapsed
> Size Size Size Time Throughput
> bytes bytes bytes secs. 10^6bits/sec
>
> 131072 16384 16384 30.01 8643.19
> 131072 16384 16384 30.01 9033.56
> 131072 16384 16384 30.01 9032.02
>
> However, many performance improvements were left out of this series for
> simplicity, so difference if performance should shrink in the future.
I think the performance should be acceptable as a start.
>
> Comments are welcome.
>
> TODO in future series:
> * Event, indirect, packed, and others features of virtio.
> * To support different set of features between the device<->SVQ and
the
> SVQ<->guest communication.
> * Support of device host notifier memory regions.
> * To sepparate buffers forwarding in its own AIO context, so we can
> throw more threads to that task and we don't need to stop the main
> event loop.
> * Support multiqueue virtio-net vdpa.
> * Proper documentation.
>
> Changes from v1:
> * Feature set at device->SVQ is now the same as SVQ->guest.
> * Size of SVQ is not max available device size anymore, but guest's
> negotiated.
> * Add VHOST_FILE_UNBIND kick and call fd treatment.
> * Make SVQ a public struct
> * Come back to previous approach to iova-tree
> * Some assertions are now fail paths. Some errors are now log_guest.
> * Only mask _F_LOG feature at vdpa_set_features svq enable path.
> * Refactor some errors and messages. Add missing error unwindings.
> * Add memory barrier at _F_NO_NOTIFY set.
> * Stop checking for features flags out of transport range.
> v1 link:
> https://lore.kernel.org/virtualization/7d86c715-6d71-8a27-91f5-8d47b71e3201
at redhat.com/
>
> Changes from v4 RFC:
> * Support of allocating / freeing iova ranges in IOVA tree. Extending
> already present iova-tree for that.
> * Proper validation of guest features. Now SVQ can negotiate a
> different set of features with the device when enabled.
> * Support of host notifiers memory regions
> * Handling of SVQ full queue in case guest's descriptors span to
> different memory regions (qemu's VA chunks).
> * Flush pending used buffers at end of SVQ operation.
> * QMP command now looks by NetClientState name. Other devices will need
> to implement it's way to enable vdpa.
> * Rename QMP command to set, so it looks more like a way of working
> * Better use of qemu error system
> * Make a few assertions proper error-handling paths.
> * Add more documentation
> * Less coupling of virtio / vhost, that could cause friction on changes
> * Addressed many other small comments and small fixes.
>
> Changes from v3 RFC:
> * Move everything to vhost-vdpa backend. A big change, this allowed
> some cleanup but more code has been added in other places.
> * More use of glib utilities, especially to manage memory.
> v3 link:
> https://lists.nongnu.org/archive/html/qemu-devel/2021-05/msg06032.html
>
> Changes from v2 RFC:
> * Adding vhost-vdpa devices support
> * Fixed some memory leaks pointed by different comments
> v2 link:
> https://lists.nongnu.org/archive/html/qemu-devel/2021-03/msg05600.html
>
> Changes from v1 RFC:
> * Use QMP instead of migration to start SVQ mode.
> * Only accepting IOMMU devices, closer behavior with target devices
> (vDPA)
> * Fix invalid masking/unmasking of vhost call fd.
> * Use of proper methods for synchronization.
> * No need to modify VirtIO device code, all of the changes are
> contained in vhost code.
> * Delete superfluous code.
> * An intermediate RFC was sent with only the notifications forwarding
> changes. It can be seen in
> https://patchew.org/QEMU/20210129205415.876290-1-eperezma at
redhat.com/
> v1 link:
> https://lists.gnu.org/archive/html/qemu-devel/2020-11/msg05372.html
>
> Eugenio P?rez (20):
> virtio: Add VIRTIO_F_QUEUE_STATE
> virtio-net: Honor VIRTIO_CONFIG_S_DEVICE_STOPPED
> virtio: Add virtio_queue_is_host_notifier_enabled
> vhost: Make vhost_virtqueue_{start,stop} public
> vhost: Add x-vhost-enable-shadow-vq qmp
> vhost: Add VhostShadowVirtqueue
> vdpa: Register vdpa devices in a list
> vhost: Route guest->host notification through shadow virtqueue
> Add vhost_svq_get_svq_call_notifier
> Add vhost_svq_set_guest_call_notifier
> vdpa: Save call_fd in vhost-vdpa
> vhost-vdpa: Take into account SVQ in vhost_vdpa_set_vring_call
> vhost: Route host->guest notification through shadow virtqueue
> virtio: Add vhost_shadow_vq_get_vring_addr
> vdpa: Save host and guest features
> vhost: Add vhost_svq_valid_device_features to shadow vq
> vhost: Shadow virtqueue buffers forwarding
> vhost: Add VhostIOVATree
> vhost: Use a tree to store memory mappings
> vdpa: Add custom IOTLB translations to SVQ
This list seems wrong btw :)
Thanks
>
> Eugenio P?rez (14):
> vhost: Add VhostShadowVirtqueue
> vhost: Add Shadow VirtQueue kick forwarding capabilities
> vhost: Add Shadow VirtQueue call forwarding capabilities
> vhost: Add vhost_svq_valid_features to shadow vq
> virtio: Add vhost_shadow_vq_get_vring_addr
> vdpa: adapt vhost_ops callbacks to svq
> vhost: Shadow virtqueue buffers forwarding
> util: Add iova_tree_alloc
> vhost: Add VhostIOVATree
> vdpa: Add custom IOTLB translations to SVQ
> vdpa: Adapt vhost_vdpa_get_vring_base to SVQ
> vdpa: Never set log_base addr if SVQ is enabled
> vdpa: Expose VHOST_F_LOG_ALL on SVQ
> vdpa: Add x-svq to NetdevVhostVDPAOptions
>
> qapi/net.json | 5 +-
> hw/virtio/vhost-iova-tree.h | 27 ++
> hw/virtio/vhost-shadow-virtqueue.h | 90 ++++
> include/hw/virtio/vhost-vdpa.h | 8 +
> include/qemu/iova-tree.h | 18 +
> hw/virtio/vhost-iova-tree.c | 155 +++++++
> hw/virtio/vhost-shadow-virtqueue.c | 632 +++++++++++++++++++++++++++++
> hw/virtio/vhost-vdpa.c | 551 ++++++++++++++++++++++++-
> net/vhost-vdpa.c | 48 ++-
> util/iova-tree.c | 133 ++++++
> hw/virtio/meson.build | 2 +-
> 11 files changed, 1644 insertions(+), 25 deletions(-)
> create mode 100644 hw/virtio/vhost-iova-tree.h
> create mode 100644 hw/virtio/vhost-shadow-virtqueue.h
> create mode 100644 hw/virtio/vhost-iova-tree.c
> create mode 100644 hw/virtio/vhost-shadow-virtqueue.c
>
> --
> 2.27.0
>
>