512).
* zero copy is really good for VM2VM case. For iperf between two VMs, the
- boost could be above 70% (when TSO is enableld).
+ boost could be above 70% (when TSO is enabled).
- * for VM2NIC case, the ``nb_tx_desc`` has to be small enough: <= 64 if virtio
- indirect feature is not enabled and <= 128 if it is enabled.
+ * For zero copy in VM2NIC case, guest Tx used vring may be starved if the
+ PMD driver consume the mbuf but not release them timely.
- This is because when dequeue zero copy is enabled, guest Tx used vring will
- be updated only when corresponding mbuf is freed. Thus, the nb_tx_desc
- has to be small enough so that the PMD driver will run out of available
- Tx descriptors and free mbufs timely. Otherwise, guest Tx vring would be
- starved.
+ For example, i40e driver has an optimization to maximum NIC pipeline which
+ postpones returning transmitted mbuf until only tx_free_threshold free
+ descs left. The virtio TX used ring will be starved if the formula
+ (num_i40e_tx_desc - num_virtio_tx_desc > tx_free_threshold) is true, since
+ i40e will not return back mbuf.
+
+ A performance tip for tuning zero copy in VM2NIC case is to adjust the
+ frequency of mbuf free (i.e. adjust tx_free_threshold of i40e driver) to
+ balance consumer and producer.
* Guest memory should be backended with huge pages to achieve better
performance. Using 1G page size is the best.
of those segments, thus the fewer the segments, the quicker we will get
the mapping. NOTE: we may speed it by using tree searching in future.
+ * zero copy can not work when using vfio-pci with iommu mode currently, this
+ is because we don't setup iommu dma mapping for guest memory. If you have
+ to use vfio-pci driver, please insert vfio-pci kernel module in noiommu
+ mode.
+
+ * The consumer of zero copy mbufs should consume these mbufs as soon as
+ possible, otherwise it may block the operations in vhost.
+
- ``RTE_VHOST_USER_IOMMU_SUPPORT``
IOMMU support will be enabled when this flag is set. It is disabled by
from accessing memory the virtio device isn't allowed to, when the feature
is negotiated and an IOMMU device is declared.
- However, this feature enables vhost-user's reply-ack protocol feature,
- which implementation is buggy in Qemu v2.7.0-v2.9.0 when doing multiqueue.
- Enabling this flag with these Qemu version results in Qemu being blocked
- when multiple queue pairs are declared.
+ - ``RTE_VHOST_USER_POSTCOPY_SUPPORT``
+
+ Postcopy live-migration support will be enabled when this flag is set.
+ It is disabled by default.
+
+ Enabling this flag should only be done when the calling application does
+ not pre-fault the guest shared memory, otherwise migration would fail.
+
+ - ``RTE_VHOST_USER_LINEARBUF_SUPPORT``
+
+ Enabling this flag forces vhost dequeue function to only provide linear
+ pktmbuf (no multi-segmented pktmbuf).
+
+ The vhost library by default provides a single pktmbuf for given a
+ packet, but if for some reason the data doesn't fit into a single
+ pktmbuf (e.g., TSO is enabled), the library will allocate additional
+ pktmbufs from the same mempool and chain them together to create a
+ multi-segmented pktmbuf.
+
+ However, the vhost application needs to support multi-segmented format.
+ If the vhost application does not support that format and requires large
+ buffers to be dequeue, this flag should be enabled to force only linear
+ buffers (see RTE_VHOST_USER_EXTBUF_SUPPORT) or drop the packet.
+
+ It is disabled by default.
+
+ - ``RTE_VHOST_USER_EXTBUF_SUPPORT``
+
+ Enabling this flag allows vhost dequeue function to allocate and attach
+ an external buffer to a pktmbuf if the pkmbuf doesn't provide enough
+ space to store all data.
+
+ This is useful when the vhost application wants to support large packets
+ but doesn't want to increase the default mempool object size nor to
+ support multi-segmented mbufs (non-linear). In this case, a fresh buffer
+ is allocated using rte_malloc() which gets attached to a pktmbuf using
+ rte_pktmbuf_attach_extbuf().
+
+ See RTE_VHOST_USER_LINEARBUF_SUPPORT as well to disable multi-segmented
+ mbufs for application that doesn't support chained mbufs.
+
+ It is disabled by default.
+
+ - ``RTE_VHOST_USER_ASYNC_COPY``
+
+ Asynchronous data path will be enabled when this flag is set. Async data
+ path allows applications to register async copy devices (typically
+ hardware DMA channels) to the vhost queues. Vhost leverages the copy
+ device registered to free CPU from memory copy operations. A set of
+ async data path APIs are defined for DPDK applications to make use of
+ the async capability. Only packets enqueued/dequeued by async APIs are
+ processed through the async data path.
+
+ Currently this feature is only implemented on split ring enqueue data
+ path.
+
+ It is disabled by default.
* ``rte_vhost_driver_set_features(path, features)``
Receives (dequeues) ``count`` packets from guest, and stored them at ``pkts``.
+* ``rte_vhost_crypto_create(vid, cryptodev_id, sess_mempool, socket_id)``
+
+ As an extension of new_device(), this function adds virtio-crypto workload
+ acceleration capability to the device. All crypto workload is processed by
+ DPDK cryptodev with the device ID of ``cryptodev_id``.
+
+* ``rte_vhost_crypto_free(vid)``
+
+ Frees the memory and vhost-user message handlers created in
+ rte_vhost_crypto_create().
+
+* ``rte_vhost_crypto_fetch_requests(vid, queue_id, ops, nb_ops)``
+
+ Receives (dequeues) ``nb_ops`` virtio-crypto requests from guest, parses
+ them to DPDK Crypto Operations, and fills the ``ops`` with parsing results.
+
+* ``rte_vhost_crypto_finalize_requests(queue_id, ops, nb_ops)``
+
+ After the ``ops`` are dequeued from Cryptodev, finalizes the jobs and
+ notifies the guest(s).
+
+* ``rte_vhost_crypto_set_zero_copy(vid, option)``
+
+ Enable or disable zero copy feature of the vhost crypto backend.
+
+* ``rte_vhost_async_channel_register(vid, queue_id, features, ops)``
+
+ Register a vhost queue with async copy device channel.
+ Following device ``features`` must be specified together with the
+ registration:
+
+ * ``async_inorder``
+
+ Async copy device can guarantee the ordering of copy completion
+ sequence. Copies are completed in the same order with that at
+ the submission time.
+
+ Currently, only ``async_inorder`` capable device is supported by vhost.
+
+ * ``async_threshold``
+
+ The copy length (in bytes) below which CPU copy will be used even if
+ applications call async vhost APIs to enqueue/dequeue data.
+
+ Typical value is 512~1024 depending on the async device capability.
+
+ Applications must provide following ``ops`` callbacks for vhost lib to
+ work with the async copy devices:
+
+ * ``transfer_data(vid, queue_id, descs, opaque_data, count)``
+
+ vhost invokes this function to submit copy data to the async devices.
+ For non-async_inorder capable devices, ``opaque_data`` could be used
+ for identifying the completed packets.
+
+ * ``check_completed_copies(vid, queue_id, opaque_data, max_packets)``
+
+ vhost invokes this function to get the copy data completed by async
+ devices.
+
+* ``rte_vhost_async_channel_unregister(vid, queue_id)``
+
+ Unregister the async copy device channel from a vhost queue.
+
+* ``rte_vhost_submit_enqueue_burst(vid, queue_id, pkts, count)``
+
+ Submit an enqueue request to transmit ``count`` packets from host to guest
+ by async data path. Enqueue is not guaranteed to finish upon the return of
+ this API call.
+
+ Applications must not free the packets submitted for enqueue until the
+ packets are completed.
+
+* ``rte_vhost_poll_enqueue_completed(vid, queue_id, pkts, count)``
+
+ Poll enqueue completion status from async data path. Completed packets
+ are returned to applications through ``pkts``.
+
Vhost-user Implementations
--------------------------
* Memory pre-allocation
- For non-zerocopy, guest memory pre-allocation is not a must. This can help
- save of memory. If users really want the guest memory to be pre-allocated
- (e.g., for performance reason), we can add option ``-mem-prealloc`` when
- starting QEMU. Or, we can lock all memory at vhost side which will force
- memory to be allocated when mmap at vhost side; option --mlockall in
- ovs-dpdk is an example in hand.
+ For non-zerocopy non-async data path, guest memory pre-allocation is not a
+ must. This can help save of memory. If users really want the guest memory
+ to be pre-allocated (e.g., for performance reason), we can add option
+ ``-mem-prealloc`` when starting QEMU. Or, we can lock all memory at vhost
+ side which will force memory to be allocated when mmap at vhost side;
+ option --mlockall in ovs-dpdk is an example in hand.
- For zerocopy, we force the VM memory to be pre-allocated at vhost lib when
- mapping the guest memory; and also we need to lock the memory to prevent
- pages being swapped out to disk.
+ For async and zerocopy data path, we force the VM memory to be
+ pre-allocated at vhost lib when mapping the guest memory; and also we need
+ to lock the memory to prevent pages being swapped out to disk.
* Memory sharing
For more vhost details and how to support vhost in vSwitch, please refer to
the vhost example in the DPDK Sample Applications Guide.
+
+Vhost data path acceleration (vDPA)
+-----------------------------------
+
+vDPA supports selective datapath in vhost-user lib by enabling virtio ring
+compatible devices to serve virtio driver directly for datapath acceleration.
+
+``rte_vhost_driver_attach_vdpa_device`` is used to configure the vhost device
+with accelerated backend.
+
+Also vhost device capabilities are made configurable to adopt various devices.
+Such capabilities include supported features, protocol features, queue number.
+
+Finally, a set of device ops is defined for device specific operations:
+
+* ``get_queue_num``
+
+ Called to get supported queue number of the device.
+
+* ``get_features``
+
+ Called to get supported features of the device.
+
+* ``get_protocol_features``
+
+ Called to get supported protocol features of the device.
+
+* ``dev_conf``
+
+ Called to configure the actual device when the virtio device becomes ready.
+
+* ``dev_close``
+
+ Called to close the actual device when the virtio device is stopped.
+
+* ``set_vring_state``
+
+ Called to change the state of the vring in the actual device when vring state
+ changes.
+
+* ``set_features``
+
+ Called to set the negotiated features to device.
+
+* ``migration_done``
+
+ Called to allow the device to response to RARP sending.
+
+* ``get_vfio_group_fd``
+
+ Called to get the VFIO group fd of the device.
+
+* ``get_vfio_device_fd``
+
+ Called to get the VFIO device fd of the device.
+
+* ``get_notify_area``
+
+ Called to get the notify area info of the queue.
git.droids-corp.org / dpdk.git / blobdiff