xref: /dpdk/doc/guides/prog_guide/vhost_lib.rst (revision 4aa1f88ac13d9031a399f73d48b8100fe4e86cc9)
1..  SPDX-License-Identifier: BSD-3-Clause
2    Copyright(c) 2010-2016 Intel Corporation.
3
4Vhost Library
5=============
6
7The vhost library implements a user space virtio net server allowing the user
8to manipulate the virtio ring directly. In another words, it allows the user
9to fetch/put packets from/to the VM virtio net device. To achieve this, a
10vhost library should be able to:
11
12* Access the guest memory:
13
14  For QEMU, this is done by using the ``-object memory-backend-file,share=on,...``
15  option. Which means QEMU will create a file to serve as the guest RAM.
16  The ``share=on`` option allows another process to map that file, which
17  means it can access the guest RAM.
18
19* Know all the necessary information about the vring:
20
21  Information such as where the available ring is stored. Vhost defines some
22  messages (passed through a Unix domain socket file) to tell the backend all
23  the information it needs to know how to manipulate the vring.
24
25
26Vhost API Overview
27------------------
28
29The following is an overview of some key Vhost API functions:
30
31* ``rte_vhost_driver_register(path, flags)``
32
33  This function registers a vhost driver into the system. ``path`` specifies
34  the Unix domain socket file path.
35
36  Currently supported flags are:
37
38  - ``RTE_VHOST_USER_CLIENT``
39
40    DPDK vhost-user will act as the client when this flag is given. See below
41    for an explanation.
42
43  - ``RTE_VHOST_USER_NO_RECONNECT``
44
45    When DPDK vhost-user acts as the client it will keep trying to reconnect
46    to the server (QEMU) until it succeeds. This is useful in two cases:
47
48    * When QEMU is not started yet.
49    * When QEMU restarts (for example due to a guest OS reboot).
50
51    This reconnect option is enabled by default. However, it can be turned off
52    by setting this flag.
53
54  - ``RTE_VHOST_USER_IOMMU_SUPPORT``
55
56    IOMMU support will be enabled when this flag is set. It is disabled by
57    default.
58
59    Enabling this flag makes possible to use guest vIOMMU to protect vhost
60    from accessing memory the virtio device isn't allowed to, when the feature
61    is negotiated and an IOMMU device is declared.
62
63  - ``RTE_VHOST_USER_POSTCOPY_SUPPORT``
64
65    Postcopy live-migration support will be enabled when this flag is set.
66    It is disabled by default.
67
68    Enabling this flag should only be done when the calling application does
69    not pre-fault the guest shared memory, otherwise migration would fail.
70
71  - ``RTE_VHOST_USER_LINEARBUF_SUPPORT``
72
73    Enabling this flag forces vhost dequeue function to only provide linear
74    pktmbuf (no multi-segmented pktmbuf).
75
76    The vhost library by default provides a single pktmbuf for given a
77    packet, but if for some reason the data doesn't fit into a single
78    pktmbuf (e.g., TSO is enabled), the library will allocate additional
79    pktmbufs from the same mempool and chain them together to create a
80    multi-segmented pktmbuf.
81
82    However, the vhost application needs to support multi-segmented format.
83    If the vhost application does not support that format and requires large
84    buffers to be dequeue, this flag should be enabled to force only linear
85    buffers (see RTE_VHOST_USER_EXTBUF_SUPPORT) or drop the packet.
86
87    It is disabled by default.
88
89  - ``RTE_VHOST_USER_EXTBUF_SUPPORT``
90
91    Enabling this flag allows vhost dequeue function to allocate and attach
92    an external buffer to a pktmbuf if the pkmbuf doesn't provide enough
93    space to store all data.
94
95    This is useful when the vhost application wants to support large packets
96    but doesn't want to increase the default mempool object size nor to
97    support multi-segmented mbufs (non-linear). In this case, a fresh buffer
98    is allocated using rte_malloc() which gets attached to a pktmbuf using
99    rte_pktmbuf_attach_extbuf().
100
101    See RTE_VHOST_USER_LINEARBUF_SUPPORT as well to disable multi-segmented
102    mbufs for application that doesn't support chained mbufs.
103
104    It is disabled by default.
105
106  - ``RTE_VHOST_USER_ASYNC_COPY``
107
108    Asynchronous data path will be enabled when this flag is set. Async
109    data path allows applications to enable DMA acceleration for vhost
110    queues. Vhost leverages the registered DMA channels to free CPU from
111    memory copy operations in data path. A set of async data path APIs are
112    defined for DPDK applications to make use of the async capability. Only
113    packets enqueued/dequeued by async APIs are processed through the async
114    data path.
115
116    Currently this feature is only implemented on split ring enqueue data
117    path.
118
119    It is disabled by default.
120
121  - ``RTE_VHOST_USER_NET_COMPLIANT_OL_FLAGS``
122
123    Since v16.04, the vhost library forwards checksum and gso requests for
124    packets received from a virtio driver by filling Tx offload metadata in
125    the mbuf. This behavior is inconsistent with other drivers but it is left
126    untouched for existing applications that might rely on it.
127
128    This flag disables the legacy behavior and instead ask vhost to simply
129    populate Rx offload metadata in the mbuf.
130
131    It is disabled by default.
132
133  - ``RTE_VHOST_USER_NET_STATS_ENABLE``
134
135  Per-virtqueue statistics collection will be enabled when this flag is set.
136  When enabled, the application may use rte_vhost_stats_get_names() and
137  rte_vhost_stats_get() to collect statistics, and rte_vhost_stats_reset() to
138  reset them.
139
140  It is disabled by default
141
142* ``rte_vhost_driver_set_features(path, features)``
143
144  This function sets the feature bits the vhost-user driver supports. The
145  vhost-user driver could be vhost-user net, yet it could be something else,
146  say, vhost-user SCSI.
147
148* ``rte_vhost_driver_callback_register(path, vhost_device_ops)``
149
150  This function registers a set of callbacks, to let DPDK applications take
151  the appropriate action when some events happen. The following events are
152  currently supported:
153
154  * ``new_device(int vid)``
155
156    This callback is invoked when a virtio device becomes ready. ``vid``
157    is the vhost device ID.
158
159  * ``destroy_device(int vid)``
160
161    This callback is invoked when a virtio device is paused or shut down.
162
163  * ``vring_state_changed(int vid, uint16_t queue_id, int enable)``
164
165    This callback is invoked when a specific queue's state is changed, for
166    example to enabled or disabled.
167
168  * ``features_changed(int vid, uint64_t features)``
169
170    This callback is invoked when the features is changed. For example,
171    ``VHOST_F_LOG_ALL`` will be set/cleared at the start/end of live
172    migration, respectively.
173
174  * ``new_connection(int vid)``
175
176    This callback is invoked on new vhost-user socket connection. If DPDK
177    acts as the server the device should not be deleted before
178    ``destroy_connection`` callback is received.
179
180  * ``destroy_connection(int vid)``
181
182    This callback is invoked when vhost-user socket connection is closed.
183    It indicates that device with id ``vid`` is no longer in use and can be
184    safely deleted.
185
186* ``rte_vhost_driver_disable/enable_features(path, features))``
187
188  This function disables/enables some features. For example, it can be used to
189  disable mergeable buffers and TSO features, which both are enabled by
190  default.
191
192* ``rte_vhost_driver_start(path)``
193
194  This function triggers the vhost-user negotiation. It should be invoked at
195  the end of initializing a vhost-user driver.
196
197* ``rte_vhost_enqueue_burst(vid, queue_id, pkts, count)``
198
199  Transmits (enqueues) ``count`` packets from host to guest.
200
201* ``rte_vhost_dequeue_burst(vid, queue_id, mbuf_pool, pkts, count)``
202
203  Receives (dequeues) ``count`` packets from guest, and stored them at ``pkts``.
204
205* ``rte_vhost_crypto_create(vid, cryptodev_id, sess_mempool, socket_id)``
206
207  As an extension of new_device(), this function adds virtio-crypto workload
208  acceleration capability to the device. All crypto workload is processed by
209  DPDK cryptodev with the device ID of ``cryptodev_id``.
210
211* ``rte_vhost_crypto_free(vid)``
212
213  Frees the memory and vhost-user message handlers created in
214  rte_vhost_crypto_create().
215
216* ``rte_vhost_crypto_fetch_requests(vid, queue_id, ops, nb_ops)``
217
218  Receives (dequeues) ``nb_ops`` virtio-crypto requests from guest, parses
219  them to DPDK Crypto Operations, and fills the ``ops`` with parsing results.
220
221* ``rte_vhost_crypto_finalize_requests(queue_id, ops, nb_ops)``
222
223  After the ``ops`` are dequeued from Cryptodev, finalizes the jobs and
224  notifies the guest(s).
225
226* ``rte_vhost_crypto_set_zero_copy(vid, option)``
227
228  Enable or disable zero copy feature of the vhost crypto backend.
229
230* ``rte_vhost_async_dma_configure(dma_id, vchan_id)``
231
232  Tell vhost which DMA vChannel is going to use. This function needs to
233  be called before register async data-path for vring.
234
235* ``rte_vhost_async_channel_register(vid, queue_id)``
236
237  Register async DMA acceleration for a vhost queue after vring is enabled.
238
239* ``rte_vhost_async_channel_register_thread_unsafe(vid, queue_id)``
240
241  Register async DMA acceleration for a vhost queue without performing
242  any locking.
243
244  This function is only safe to call in vhost callback functions
245  (i.e., struct rte_vhost_device_ops).
246
247* ``rte_vhost_async_channel_unregister(vid, queue_id)``
248
249  Unregister the async DMA acceleration from a vhost queue.
250  Unregistration will fail, if the vhost queue has in-flight
251  packets that are not completed.
252
253  Unregister async DMA acceleration in vring_state_changed() may
254  fail, as this API tries to acquire the spinlock of vhost
255  queue. The recommended way is to unregister async copy
256  devices for all vhost queues in destroy_device(), when a
257  virtio device is paused or shut down.
258
259* ``rte_vhost_async_channel_unregister_thread_unsafe(vid, queue_id)``
260
261  Unregister async DMA acceleration for a vhost queue without performing
262  any locking.
263
264  This function is only safe to call in vhost callback functions
265  (i.e., struct rte_vhost_device_ops).
266
267* ``rte_vhost_submit_enqueue_burst(vid, queue_id, pkts, count, dma_id, vchan_id)``
268
269  Submit an enqueue request to transmit ``count`` packets from host to guest
270  by async data path. Applications must not free the packets submitted for
271  enqueue until the packets are completed.
272
273* ``rte_vhost_poll_enqueue_completed(vid, queue_id, pkts, count, dma_id, vchan_id)``
274
275  Poll enqueue completion status from async data path. Completed packets
276  are returned to applications through ``pkts``.
277
278* ``rte_vhost_async_get_inflight(vid, queue_id)``
279
280  This function returns the amount of in-flight packets for the vhost
281  queue using async acceleration.
282
283 * ``rte_vhost_async_get_inflight_thread_unsafe(vid, queue_id)``
284
285  Get the number of inflight packets for a vhost queue without performing
286  any locking. It should only be used within the vhost ops, which already
287  holds the lock.
288
289* ``rte_vhost_clear_queue_thread_unsafe(vid, queue_id, **pkts, count, dma_id, vchan_id)``
290
291  Clear in-flight packets which are submitted to async channel in vhost
292  async data path without performing locking on virtqueue. Completed
293  packets are returned to applications through ``pkts``.
294
295* ``rte_vhost_clear_queue(vid, queue_id, **pkts, count, dma_id, vchan_id)``
296
297  Clear in-flight packets which are submitted to async channel in vhost async data
298  path. Completed packets are returned to applications through ``pkts``.
299
300* ``rte_vhost_vring_call_nonblock(int vid, uint16_t vring_idx)``
301
302  Notify the guest that used descriptors have been added to the vring. This function
303  will return -EAGAIN when vq's access lock is held by other thread, user should try
304  again later.
305
306* ``rte_vhost_vring_stats_get_names(int vid, uint16_t queue_id, struct rte_vhost_stat_name *names, unsigned int size)``
307
308  This function returns the names of the queue statistics. It requires
309  statistics collection to be enabled at registration time.
310
311* ``rte_vhost_vring_stats_get(int vid, uint16_t queue_id, struct rte_vhost_stat *stats, unsigned int n)``
312
313  This function returns the queue statistics. It requires statistics
314  collection to be enabled at registration time.
315
316* ``rte_vhost_vring_stats_reset(int vid, uint16_t queue_id)``
317
318  This function resets the queue statistics. It requires statistics
319  collection to be enabled at registration time.
320
321* ``rte_vhost_async_try_dequeue_burst(vid, queue_id, mbuf_pool, pkts, count,
322  nr_inflight, dma_id, vchan_id)``
323
324  Receive ``count`` packets from guest to host in async data path,
325  and store them at ``pkts``.
326
327* ``rte_vhost_driver_get_vdpa_dev_type(path, type)``
328
329  Get device type of vDPA device, such as VDPA_DEVICE_TYPE_NET,
330  VDPA_DEVICE_TYPE_BLK.
331
332* ``rte_vhost_async_dma_unconfigure(dma_id, vchan_id)``
333
334  Clean DMA vChannel finished to use. After this function is called,
335  the specified DMA vChannel should no longer be used by the Vhost library.
336
337* ``rte_vhost_notify_guest(int vid, uint16_t queue_id)``
338
339  Inject the offloaded interrupt received by the 'guest_notify' callback,
340  into the vhost device's queue.
341
342* ``rte_vhost_driver_set_max_queue_num(const char *path, uint32_t max_queue_pairs)``
343
344  Set the maximum number of queue pairs supported by the device.
345
346Vhost-user Implementations
347--------------------------
348
349Vhost-user uses Unix domain sockets for passing messages. This means the DPDK
350vhost-user implementation has two options:
351
352* DPDK vhost-user acts as the server.
353
354  DPDK will create a Unix domain socket server file and listen for
355  connections from the frontend.
356
357  Note, this is the default mode, and the only mode before DPDK v16.07.
358
359
360* DPDK vhost-user acts as the client.
361
362  Unlike the server mode, this mode doesn't create the socket file;
363  it just tries to connect to the server (which responses to create the
364  file instead).
365
366  When the DPDK vhost-user application restarts, DPDK vhost-user will try to
367  connect to the server again. This is how the "reconnect" feature works.
368
369  .. Note::
370     * The "reconnect" feature requires **QEMU v2.7** (or above).
371
372     * The vhost supported features must be exactly the same before and
373       after the restart. For example, if TSO is disabled and then enabled,
374       nothing will work and undefined issues might happen.
375
376No matter which mode is used, once a connection is established, DPDK
377vhost-user will start receiving and processing vhost messages from QEMU.
378
379For messages with a file descriptor, the file descriptor can be used directly
380in the vhost process as it is already installed by the Unix domain socket.
381
382The supported vhost messages are:
383
384* ``VHOST_SET_MEM_TABLE``
385* ``VHOST_SET_VRING_KICK``
386* ``VHOST_SET_VRING_CALL``
387* ``VHOST_SET_LOG_FD``
388* ``VHOST_SET_VRING_ERR``
389
390For ``VHOST_SET_MEM_TABLE`` message, QEMU will send information for each
391memory region and its file descriptor in the ancillary data of the message.
392The file descriptor is used to map that region.
393
394``VHOST_SET_VRING_KICK`` is used as the signal to put the vhost device into
395the data plane, and ``VHOST_GET_VRING_BASE`` is used as the signal to remove
396the vhost device from the data plane.
397
398When the socket connection is closed, vhost will destroy the device.
399
400Guest memory requirement
401------------------------
402
403* Memory pre-allocation
404
405  For non-async data path guest memory pre-allocation is not a
406  must but can help save memory. To do this we can add option
407  ``-mem-prealloc`` when starting QEMU, or we can lock all memory at vhost
408  side which will force memory to be allocated when it calls mmap
409  (option --mlockall in ovs-dpdk is an example in hand).
410
411
412  For async data path, we force the VM memory to be pre-allocated at vhost
413  lib when mapping the guest memory; and also we need to lock the memory to
414  prevent pages being swapped out to disk.
415
416* Memory sharing
417
418  Make sure ``share=on`` QEMU option is given. The vhost-user will not work with
419  a QEMU instance without shared memory mapping.
420
421Vhost supported vSwitch reference
422---------------------------------
423
424For more vhost details and how to support vhost in vSwitch, please refer to
425the vhost example in the DPDK Sample Applications Guide.
426
427Vhost data path acceleration (vDPA)
428-----------------------------------
429
430vDPA supports selective datapath in vhost-user lib by enabling virtio ring
431compatible devices to serve virtio driver directly for datapath acceleration.
432
433``rte_vhost_driver_attach_vdpa_device`` is used to configure the vhost device
434with accelerated backend.
435
436Also vhost device capabilities are made configurable to adopt various devices.
437Such capabilities include supported features, protocol features, queue number.
438
439Finally, a set of device ops is defined for device specific operations:
440
441* ``get_queue_num``
442
443  Called to get supported queue number of the device.
444
445* ``get_features``
446
447  Called to get supported features of the device.
448
449* ``get_protocol_features``
450
451  Called to get supported protocol features of the device.
452
453* ``dev_conf``
454
455  Called to configure the actual device when the virtio device becomes ready.
456
457* ``dev_close``
458
459  Called to close the actual device when the virtio device is stopped.
460
461* ``set_vring_state``
462
463  Called to change the state of the vring in the actual device when vring state
464  changes.
465
466* ``set_features``
467
468  Called to set the negotiated features to device.
469
470* ``migration_done``
471
472  Called to allow the device to response to RARP sending.
473
474* ``get_vfio_group_fd``
475
476   Called to get the VFIO group fd of the device.
477
478* ``get_vfio_device_fd``
479
480  Called to get the VFIO device fd of the device.
481
482* ``get_notify_area``
483
484  Called to get the notify area info of the queue.
485
486Vhost asynchronous data path
487----------------------------
488
489Vhost asynchronous data path leverages DMA devices to offload memory
490copies from the CPU and it is implemented in an asynchronous way. It
491enables applications, like OVS, to save CPU cycles and hide memory copy
492overhead, thus achieving higher throughput.
493
494Vhost doesn't manage DMA devices and applications, like OVS, need to
495manage and configure DMA devices. Applications need to tell vhost what
496DMA devices to use in every data path function call. This design enables
497the flexibility for applications to dynamically use DMA channels in
498different function modules, not limited in vhost.
499
500In addition, vhost supports M:N mapping between vrings and DMA virtual
501channels. Specifically, one vring can use multiple different DMA channels
502and one DMA channel can be shared by multiple vrings at the same time.
503The reason of enabling one vring to use multiple DMA channels is that
504it's possible that more than one dataplane threads enqueue packets to
505the same vring with their own DMA virtual channels. Besides, the number
506of DMA devices is limited. For the purpose of scaling, it's necessary to
507support sharing DMA channels among vrings.
508
509* Async enqueue API usage
510
511  In async enqueue path, rte_vhost_poll_enqueue_completed() needs to be
512  called in time to notify the guest of DMA copy completed packets.
513  Moreover, calling rte_vhost_submit_enqueue_burst() all the time but
514  not poll completed will cause the DMA ring to be full, which will
515  result in packet loss eventually.
516
517* Recommended IOVA mode in async datapath
518
519  When DMA devices are bound to VFIO driver, VA mode is recommended.
520  For PA mode, page by page mapping may exceed IOMMU's max capability,
521  better to use 1G guest hugepage.
522
523  For UIO driver or kernel driver, any VFIO related error messages
524  can be ignored.
525