FD.io VPP  v19.08-27-gf4dcae4
Vector Packet Processing
vhost_user.c
Go to the documentation of this file.
1 /*
2  *------------------------------------------------------------------
3  * vhost.c - vhost-user
4  *
5  * Copyright (c) 2014-2018 Cisco and/or its affiliates.
6  * Licensed under the Apache License, Version 2.0 (the "License");
7  * you may not use this file except in compliance with the License.
8  * You may obtain a copy of the License at:
9  *
10  * http://www.apache.org/licenses/LICENSE-2.0
11  *
12  * Unless required by applicable law or agreed to in writing, software
13  * distributed under the License is distributed on an "AS IS" BASIS,
14  * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
15  * See the License for the specific language governing permissions and
16  * limitations under the License.
17  *------------------------------------------------------------------
18  */
19 
20 #include <fcntl.h> /* for open */
21 #include <sys/ioctl.h>
22 #include <sys/socket.h>
23 #include <sys/un.h>
24 #include <sys/stat.h>
25 #include <sys/types.h>
26 #include <sys/uio.h> /* for iovec */
27 #include <netinet/in.h>
28 #include <sys/vfs.h>
29 
30 #include <linux/if_arp.h>
31 #include <linux/if_tun.h>
32 
33 #include <vlib/vlib.h>
34 #include <vlib/unix/unix.h>
35 
36 #include <vnet/ip/ip.h>
37 
38 #include <vnet/ethernet/ethernet.h>
39 #include <vnet/devices/devices.h>
40 #include <vnet/feature/feature.h>
41 
44 
45 /**
46  * @file
47  * @brief vHost User Device Driver.
48  *
49  * This file contains the source code for vHost User interface.
50  */
51 
52 
54 
55 /* *INDENT-OFF* */
56 vhost_user_main_t vhost_user_main = {
57  .mtu_bytes = 1518,
58 };
59 
60 VNET_HW_INTERFACE_CLASS (vhost_interface_class, static) = {
61  .name = "vhost-user",
62 };
63 /* *INDENT-ON* */
64 
65 static long
67 {
68  struct statfs s;
69  fstatfs (fd, &s);
70  return s.f_bsize;
71 }
72 
73 static void
75 {
76  int i, r, q;
78 
79  for (i = 0; i < vui->nregions; i++)
80  {
81  if (vui->region_mmap_addr[i] != MAP_FAILED)
82  {
83 
84  long page_sz = get_huge_page_size (vui->region_mmap_fd[i]);
85 
86  ssize_t map_sz = (vui->regions[i].memory_size +
87  vui->regions[i].mmap_offset +
88  page_sz - 1) & ~(page_sz - 1);
89 
90  r =
91  munmap (vui->region_mmap_addr[i] - vui->regions[i].mmap_offset,
92  map_sz);
93 
94  vu_log_debug (vui, "unmap memory region %d addr 0x%lx len 0x%lx "
95  "page_sz 0x%x", i, vui->region_mmap_addr[i], map_sz,
96  page_sz);
97 
98  vui->region_mmap_addr[i] = MAP_FAILED;
99 
100  if (r == -1)
101  {
102  vu_log_err (vui, "failed to unmap memory region (errno %d)",
103  errno);
104  }
105  close (vui->region_mmap_fd[i]);
106  }
107  }
108  vui->nregions = 0;
109 
110  for (q = 0; q < VHOST_VRING_MAX_N; q++)
111  {
112  vq = &vui->vrings[q];
113  vq->avail = 0;
114  vq->used = 0;
115  vq->desc = 0;
116  }
117 }
118 
121 {
122  //Let's try to assign one queue to each thread
123  u32 qid;
124  u32 thread_index = 0;
125 
126  vui->use_tx_spinlock = 0;
127  while (1)
128  {
129  for (qid = 0; qid < VHOST_VRING_MAX_N / 2; qid++)
130  {
131  vhost_user_vring_t *rxvq = &vui->vrings[VHOST_VRING_IDX_RX (qid)];
132  if (!rxvq->started || !rxvq->enabled)
133  continue;
134 
135  vui->per_cpu_tx_qid[thread_index] = qid;
136  thread_index++;
137  if (thread_index == vlib_get_thread_main ()->n_vlib_mains)
138  return;
139  }
140  //We need to loop, meaning the spinlock has to be used
141  vui->use_tx_spinlock = 1;
142  if (thread_index == 0)
143  {
144  //Could not find a single valid one
145  for (thread_index = 0;
146  thread_index < vlib_get_thread_main ()->n_vlib_mains;
147  thread_index++)
148  {
149  vui->per_cpu_tx_qid[thread_index] = 0;
150  }
151  return;
152  }
153  }
154 }
155 
156 /**
157  * @brief Unassign existing interface/queue to thread mappings and re-assign
158  * new interface/queue to thread mappings
159  */
162 {
163  vhost_user_vring_t *txvq = &vui->vrings[qid];
164  vnet_main_t *vnm = vnet_get_main ();
165  int rv;
166  u32 q = qid >> 1;
167 
168  ASSERT ((qid & 1) == 1); // should be odd
169  // Assign new queue mappings for the interface
171  vhost_user_input_node.index);
174  /* Set polling as the default */
176  txvq->qid = q;
177  rv = vnet_hw_interface_set_rx_mode (vnm, vui->hw_if_index, q, txvq->mode);
178  if (rv)
179  vu_log_warn (vui, "unable to set rx mode for interface %d, "
180  "queue %d: rc=%d", vui->hw_if_index, q, rv);
181 }
182 
183 /** @brief Returns whether at least one TX and one RX vring are enabled */
186 {
187  int i, found[2] = { }; //RX + TX
188 
189  for (i = 0; i < VHOST_VRING_MAX_N; i++)
190  if (vui->vrings[i].started && vui->vrings[i].enabled)
191  found[i & 1] = 1;
192 
193  return found[0] && found[1];
194 }
195 
198 {
199  /* if we have pointers to descriptor table, go up */
200  int is_ready = vhost_user_intf_ready (vui);
201  if (is_ready != vui->is_ready)
202  {
203  vu_log_debug (vui, "interface %d %s", vui->sw_if_index,
204  is_ready ? "ready" : "down");
205  if (vui->admin_up)
208  : 0);
209  vui->is_ready = is_ready;
210  }
211 }
212 
213 static void
215 {
216  u32 qid;
217  vnet_main_t *vnm = vnet_get_main ();
218 
219  qid = ifq & 0xff;
220  if ((qid & 1) == 0)
221  /* Only care about the odd number, or TX, virtqueue */
222  return;
223 
224  if (vhost_user_intf_ready (vui))
225  // qid >> 1 is to convert virtqueue number to vring queue index
227 }
228 
229 static clib_error_t *
231 {
232  __attribute__ ((unused)) int n;
233  u8 buff[8];
234 
235  n = read (uf->file_descriptor, ((char *) &buff), 8);
236 
237  return 0;
238 }
239 
242 {
243  if (qid & 1) // RX is odd, TX is even
244  {
245  if (vui->vrings[qid].qid == -1)
247  }
248  else
250 }
251 
252 static clib_error_t *
254 {
255  __attribute__ ((unused)) int n;
256  u8 buff[8];
257  vhost_user_intf_t *vui =
258  pool_elt_at_index (vhost_user_main.vhost_user_interfaces,
259  uf->private_data >> 8);
260  u32 qid = uf->private_data & 0xff;
261 
262  n = read (uf->file_descriptor, ((char *) &buff), 8);
263  vu_log_debug (vui, "if %d KICK queue %d", uf->private_data >> 8, qid);
264  if (!vui->vrings[qid].started ||
265  (vhost_user_intf_ready (vui) != vui->is_ready))
266  {
267  if (vui->vrings[qid].started == 0)
268  {
269  vui->vrings[qid].started = 1;
270  vhost_user_thread_placement (vui, qid);
272  }
273  }
274 
276  return 0;
277 }
278 
281 {
282  vhost_user_vring_t *vring = &vui->vrings[qid];
283  clib_memset (vring, 0, sizeof (*vring));
284  vring->kickfd_idx = ~0;
285  vring->callfd_idx = ~0;
286  vring->errfd = -1;
287  vring->qid = -1;
288 
289  /*
290  * We have a bug with some qemu 2.5, and this may be a fix.
291  * Feel like interpretation holy text, but this is from vhost-user.txt.
292  * "
293  * One queue pair is enabled initially. More queues are enabled
294  * dynamically, by sending message VHOST_USER_SET_VRING_ENABLE.
295  * "
296  * Don't know who's right, but this is what DPDK does.
297  */
298  if (qid == 0 || qid == 1)
299  vring->enabled = 1;
300 }
301 
304 {
305  vhost_user_vring_t *vring = &vui->vrings[qid];
306 
307  if (vring->kickfd_idx != ~0)
308  {
310  vring->kickfd_idx);
311  clib_file_del (&file_main, uf);
312  vring->kickfd_idx = ~0;
313  }
314  if (vring->callfd_idx != ~0)
315  {
317  vring->callfd_idx);
318  clib_file_del (&file_main, uf);
319  vring->callfd_idx = ~0;
320  }
321  if (vring->errfd != -1)
322  {
323  close (vring->errfd);
324  vring->errfd = -1;
325  }
326 
327  // save the qid so that we don't need to unassign and assign_rx_thread
328  // when the interface comes back up. They are expensive calls.
329  u16 q = vui->vrings[qid].qid;
330  vhost_user_vring_init (vui, qid);
331  vui->vrings[qid].qid = q;
332 }
333 
336 {
337  vnet_main_t *vnm = vnet_get_main ();
338  int q;
339 
341 
342  if (vui->clib_file_index != ~0)
343  {
345  vui->clib_file_index = ~0;
346  }
347 
348  vui->is_ready = 0;
349 
350  for (q = 0; q < VHOST_VRING_MAX_N; q++)
351  vhost_user_vring_close (vui, q);
352 
353  unmap_all_mem_regions (vui);
354  vu_log_debug (vui, "interface ifindex %d disconnected", vui->sw_if_index);
355 }
356 
357 static clib_error_t *
359 {
360  int n, i, j;
361  int fd, number_of_fds = 0;
362  int fds[VHOST_MEMORY_MAX_NREGIONS];
363  vhost_user_msg_t msg;
364  struct msghdr mh;
365  struct iovec iov[1];
367  vhost_user_intf_t *vui;
368  struct cmsghdr *cmsg;
369  u8 q;
370  clib_file_t template = { 0 };
371  vnet_main_t *vnm = vnet_get_main ();
373 
374  vui = pool_elt_at_index (vum->vhost_user_interfaces, uf->private_data);
375 
376  char control[CMSG_SPACE (VHOST_MEMORY_MAX_NREGIONS * sizeof (int))];
377 
378  clib_memset (&mh, 0, sizeof (mh));
379  clib_memset (control, 0, sizeof (control));
380 
381  for (i = 0; i < VHOST_MEMORY_MAX_NREGIONS; i++)
382  fds[i] = -1;
383 
384  /* set the payload */
385  iov[0].iov_base = (void *) &msg;
386  iov[0].iov_len = VHOST_USER_MSG_HDR_SZ;
387 
388  mh.msg_iov = iov;
389  mh.msg_iovlen = 1;
390  mh.msg_control = control;
391  mh.msg_controllen = sizeof (control);
392 
393  n = recvmsg (uf->file_descriptor, &mh, 0);
394 
395  if (n != VHOST_USER_MSG_HDR_SZ)
396  {
397  if (n == -1)
398  {
399  vu_log_debug (vui, "recvmsg returned error %d %s", errno,
400  strerror (errno));
401  }
402  else
403  {
404  vu_log_debug (vui, "n (%d) != VHOST_USER_MSG_HDR_SZ (%d)",
406  }
407  goto close_socket;
408  }
409 
410  if (mh.msg_flags & MSG_CTRUNC)
411  {
412  vu_log_debug (vui, "MSG_CTRUNC is set");
413  goto close_socket;
414  }
415 
416  cmsg = CMSG_FIRSTHDR (&mh);
417 
418  if (cmsg && (cmsg->cmsg_len > 0) && (cmsg->cmsg_level == SOL_SOCKET) &&
419  (cmsg->cmsg_type == SCM_RIGHTS) &&
420  (cmsg->cmsg_len - CMSG_LEN (0) <=
421  VHOST_MEMORY_MAX_NREGIONS * sizeof (int)))
422  {
423  number_of_fds = (cmsg->cmsg_len - CMSG_LEN (0)) / sizeof (int);
424  clib_memcpy_fast (fds, CMSG_DATA (cmsg), number_of_fds * sizeof (int));
425  }
426 
427  /* version 1, no reply bit set */
428  if ((msg.flags & 7) != 1)
429  {
430  vu_log_debug (vui, "malformed message received. closing socket");
431  goto close_socket;
432  }
433 
434  {
435  int rv;
436  rv =
437  read (uf->file_descriptor, ((char *) &msg) + VHOST_USER_MSG_HDR_SZ,
438  msg.size);
439  if (rv < 0)
440  {
441  vu_log_debug (vui, "read failed %s", strerror (errno));
442  goto close_socket;
443  }
444  else if (rv != msg.size)
445  {
446  vu_log_debug (vui, "message too short (read %dB should be %dB)", rv,
447  msg.size);
448  goto close_socket;
449  }
450  }
451 
452  switch (msg.request)
453  {
455  msg.flags |= 4;
456  msg.u64 = (1ULL << FEAT_VIRTIO_NET_F_MRG_RXBUF) |
457  (1ULL << FEAT_VIRTIO_NET_F_CTRL_VQ) |
458  (1ULL << FEAT_VIRTIO_F_ANY_LAYOUT) |
459  (1ULL << FEAT_VIRTIO_F_INDIRECT_DESC) |
460  (1ULL << FEAT_VHOST_F_LOG_ALL) |
461  (1ULL << FEAT_VIRTIO_NET_F_GUEST_ANNOUNCE) |
462  (1ULL << FEAT_VIRTIO_NET_F_MQ) |
463  (1ULL << FEAT_VHOST_USER_F_PROTOCOL_FEATURES) |
464  (1ULL << FEAT_VIRTIO_F_VERSION_1);
465  msg.u64 &= vui->feature_mask;
466 
467  if (vui->enable_gso)
469 
470  msg.size = sizeof (msg.u64);
471  vu_log_debug (vui, "if %d msg VHOST_USER_GET_FEATURES - reply "
472  "0x%016llx", vui->hw_if_index, msg.u64);
473  n =
474  send (uf->file_descriptor, &msg, VHOST_USER_MSG_HDR_SZ + msg.size, 0);
475  if (n != (msg.size + VHOST_USER_MSG_HDR_SZ))
476  {
477  vu_log_debug (vui, "could not send message response");
478  goto close_socket;
479  }
480  break;
481 
483  vu_log_debug (vui, "if %d msg VHOST_USER_SET_FEATURES features "
484  "0x%016llx", vui->hw_if_index, msg.u64);
485 
486  vui->features = msg.u64;
487 
488  if (vui->features &
489  ((1 << FEAT_VIRTIO_NET_F_MRG_RXBUF) |
490  (1ULL << FEAT_VIRTIO_F_VERSION_1)))
491  vui->virtio_net_hdr_sz = 12;
492  else
493  vui->virtio_net_hdr_sz = 10;
494 
495  vui->is_any_layout =
496  (vui->features & (1 << FEAT_VIRTIO_F_ANY_LAYOUT)) ? 1 : 0;
497 
500  if (vui->enable_gso &&
503  hw->flags |=
506  else
510  vui->is_ready = 0;
512  break;
513 
515  vu_log_debug (vui, "if %d msg VHOST_USER_SET_MEM_TABLE nregions %d",
516  vui->hw_if_index, msg.memory.nregions);
517 
518  if ((msg.memory.nregions < 1) ||
519  (msg.memory.nregions > VHOST_MEMORY_MAX_NREGIONS))
520  {
521  vu_log_debug (vui, "number of mem regions must be between 1 and %i",
522  VHOST_MEMORY_MAX_NREGIONS);
523  goto close_socket;
524  }
525 
526  if (msg.memory.nregions != number_of_fds)
527  {
528  vu_log_debug (vui, "each memory region must have FD");
529  goto close_socket;
530  }
531 
532  /* Do the mmap without barrier sync */
533  void *region_mmap_addr[VHOST_MEMORY_MAX_NREGIONS];
534  for (i = 0; i < msg.memory.nregions; i++)
535  {
536  long page_sz = get_huge_page_size (fds[i]);
537 
538  /* align size to page */
539  ssize_t map_sz = (msg.memory.regions[i].memory_size +
540  msg.memory.regions[i].mmap_offset +
541  page_sz - 1) & ~(page_sz - 1);
542 
543  region_mmap_addr[i] = mmap (0, map_sz, PROT_READ | PROT_WRITE,
544  MAP_SHARED, fds[i], 0);
545  if (region_mmap_addr[i] == MAP_FAILED)
546  {
547  vu_log_err (vui, "failed to map memory. errno is %d", errno);
548  for (j = 0; j < i; j++)
549  munmap (region_mmap_addr[j], map_sz);
550  goto close_socket;
551  }
552  vu_log_debug (vui, "map memory region %d addr 0 len 0x%lx fd %d "
553  "mapped 0x%lx page_sz 0x%x", i, map_sz, fds[i],
554  region_mmap_addr[i], page_sz);
555  }
556 
558  unmap_all_mem_regions (vui);
559  for (i = 0; i < msg.memory.nregions; i++)
560  {
561  clib_memcpy_fast (&(vui->regions[i]), &msg.memory.regions[i],
562  sizeof (vhost_user_memory_region_t));
563 
564  vui->region_mmap_addr[i] = region_mmap_addr[i];
565  vui->region_guest_addr_lo[i] = vui->regions[i].guest_phys_addr;
566  vui->region_guest_addr_hi[i] = vui->regions[i].guest_phys_addr +
567  vui->regions[i].memory_size;
568 
569  vui->region_mmap_addr[i] += vui->regions[i].mmap_offset;
570  vui->region_mmap_fd[i] = fds[i];
571 
572  vui->nregions++;
573  }
575  break;
576 
578  vu_log_debug (vui, "if %d msg VHOST_USER_SET_VRING_NUM idx %d num %d",
579  vui->hw_if_index, msg.state.index, msg.state.num);
580 
581  if ((msg.state.num > 32768) || /* maximum ring size is 32768 */
582  (msg.state.num == 0) || /* it cannot be zero */
583  ((msg.state.num - 1) & msg.state.num)) /* must be power of 2 */
584  goto close_socket;
585  vui->vrings[msg.state.index].qsz_mask = msg.state.num - 1;
586  break;
587 
589  vu_log_debug (vui, "if %d msg VHOST_USER_SET_VRING_ADDR idx %d",
590  vui->hw_if_index, msg.state.index);
591 
592  if (msg.state.index >= VHOST_VRING_MAX_N)
593  {
594  vu_log_debug (vui, "invalid vring index VHOST_USER_SET_VRING_ADDR:"
595  " %d >= %d", msg.state.index, VHOST_VRING_MAX_N);
596  goto close_socket;
597  }
598 
599  if (msg.size < sizeof (msg.addr))
600  {
601  vu_log_debug (vui, "vhost message is too short (%d < %d)",
602  msg.size, sizeof (msg.addr));
603  goto close_socket;
604  }
605 
606  vring_desc_t *desc = map_user_mem (vui, msg.addr.desc_user_addr);
607  vring_used_t *used = map_user_mem (vui, msg.addr.used_user_addr);
608  vring_avail_t *avail = map_user_mem (vui, msg.addr.avail_user_addr);
609 
610  if ((desc == NULL) || (used == NULL) || (avail == NULL))
611  {
612  vu_log_debug (vui, "failed to map user memory for hw_if_index %d",
613  vui->hw_if_index);
614  goto close_socket;
615  }
616 
618  vui->vrings[msg.state.index].desc = desc;
619  vui->vrings[msg.state.index].used = used;
620  vui->vrings[msg.state.index].avail = avail;
621 
622  vui->vrings[msg.state.index].log_guest_addr = msg.addr.log_guest_addr;
623  vui->vrings[msg.state.index].log_used =
624  (msg.addr.flags & (1 << VHOST_VRING_F_LOG)) ? 1 : 0;
625 
626  /* Spec says: If VHOST_USER_F_PROTOCOL_FEATURES has not been negotiated,
627  the ring is initialized in an enabled state. */
628  if (!(vui->features & (1 << FEAT_VHOST_USER_F_PROTOCOL_FEATURES)))
629  vui->vrings[msg.state.index].enabled = 1;
630 
631  vui->vrings[msg.state.index].last_used_idx =
632  vui->vrings[msg.state.index].last_avail_idx =
633  vui->vrings[msg.state.index].used->idx;
634 
635  /* tell driver that we don't want interrupts */
636  vui->vrings[msg.state.index].used->flags = VRING_USED_F_NO_NOTIFY;
639  break;
640 
642  vu_log_debug (vui, "if %d msg VHOST_USER_SET_OWNER", vui->hw_if_index);
643  break;
644 
646  vu_log_debug (vui, "if %d msg VHOST_USER_RESET_OWNER",
647  vui->hw_if_index);
648  break;
649 
651  vu_log_debug (vui, "if %d msg VHOST_USER_SET_VRING_CALL %d",
652  vui->hw_if_index, msg.u64);
653 
654  q = (u8) (msg.u64 & 0xFF);
655 
656  /* if there is old fd, delete and close it */
657  if (vui->vrings[q].callfd_idx != ~0)
658  {
660  vui->vrings[q].callfd_idx);
661  clib_file_del (&file_main, uf);
662  vui->vrings[q].callfd_idx = ~0;
663  }
664 
665  if (!(msg.u64 & VHOST_USER_VRING_NOFD_MASK))
666  {
667  if (number_of_fds != 1)
668  {
669  vu_log_debug (vui, "More than one fd received !");
670  goto close_socket;
671  }
672 
673  template.read_function = vhost_user_callfd_read_ready;
674  template.file_descriptor = fds[0];
675  template.private_data =
676  ((vui - vhost_user_main.vhost_user_interfaces) << 8) + q;
677  vui->vrings[q].callfd_idx = clib_file_add (&file_main, &template);
678  }
679  else
680  vui->vrings[q].callfd_idx = ~0;
681  break;
682 
684  vu_log_debug (vui, "if %d msg VHOST_USER_SET_VRING_KICK %d",
685  vui->hw_if_index, msg.u64);
686 
687  q = (u8) (msg.u64 & 0xFF);
688 
689  if (vui->vrings[q].kickfd_idx != ~0)
690  {
692  vui->vrings[q].kickfd_idx);
693  clib_file_del (&file_main, uf);
694  vui->vrings[q].kickfd_idx = ~0;
695  }
696 
697  if (!(msg.u64 & VHOST_USER_VRING_NOFD_MASK))
698  {
699  if (number_of_fds != 1)
700  {
701  vu_log_debug (vui, "More than one fd received !");
702  goto close_socket;
703  }
704 
705  template.read_function = vhost_user_kickfd_read_ready;
706  template.file_descriptor = fds[0];
707  template.private_data =
708  (((uword) (vui - vhost_user_main.vhost_user_interfaces)) << 8) +
709  q;
710  vui->vrings[q].kickfd_idx = clib_file_add (&file_main, &template);
711  }
712  else
713  {
714  //When no kickfd is set, the queue is initialized as started
715  vui->vrings[q].kickfd_idx = ~0;
716  vui->vrings[q].started = 1;
718  }
720  break;
721 
723  vu_log_debug (vui, "if %d msg VHOST_USER_SET_VRING_ERR %d",
724  vui->hw_if_index, msg.u64);
725 
726  q = (u8) (msg.u64 & 0xFF);
727 
728  if (vui->vrings[q].errfd != -1)
729  close (vui->vrings[q].errfd);
730 
731  if (!(msg.u64 & VHOST_USER_VRING_NOFD_MASK))
732  {
733  if (number_of_fds != 1)
734  goto close_socket;
735 
736  vui->vrings[q].errfd = fds[0];
737  }
738  else
739  vui->vrings[q].errfd = -1;
740  break;
741 
743  vu_log_debug (vui, "if %d msg VHOST_USER_SET_VRING_BASE idx %d num %d",
744  vui->hw_if_index, msg.state.index, msg.state.num);
746  vui->vrings[msg.state.index].last_avail_idx = msg.state.num;
748  break;
749 
751  if (msg.state.index >= VHOST_VRING_MAX_N)
752  {
753  vu_log_debug (vui, "invalid vring index VHOST_USER_GET_VRING_BASE:"
754  " %d >= %d", msg.state.index, VHOST_VRING_MAX_N);
755  goto close_socket;
756  }
757 
758  /* protection is needed to prevent rx/tx from changing last_avail_idx */
760  /*
761  * Copy last_avail_idx from the vring before closing it because
762  * closing the vring also initializes the vring last_avail_idx
763  */
764  msg.state.num = vui->vrings[msg.state.index].last_avail_idx;
765  msg.flags |= 4;
766  msg.size = sizeof (msg.state);
767 
768  /*
769  * Spec says: Client must [...] stop ring upon receiving
770  * VHOST_USER_GET_VRING_BASE
771  */
772  vhost_user_vring_close (vui, msg.state.index);
774  vu_log_debug (vui, "if %d msg VHOST_USER_GET_VRING_BASE idx %d num %d",
775  vui->hw_if_index, msg.state.index, msg.state.num);
776  n =
777  send (uf->file_descriptor, &msg, VHOST_USER_MSG_HDR_SZ + msg.size, 0);
778  if (n != (msg.size + VHOST_USER_MSG_HDR_SZ))
779  {
780  vu_log_debug (vui, "could not send message response");
781  goto close_socket;
782  }
784  break;
785 
786  case VHOST_USER_NONE:
787  vu_log_debug (vui, "if %d msg VHOST_USER_NONE", vui->hw_if_index);
788  break;
789 
791  vu_log_debug (vui, "if %d msg VHOST_USER_SET_LOG_BASE",
792  vui->hw_if_index);
793 
794  if (msg.size != sizeof (msg.log))
795  {
796  vu_log_debug (vui, "invalid msg size for VHOST_USER_SET_LOG_BASE:"
797  " %d instead of %d", msg.size, sizeof (msg.log));
798  goto close_socket;
799  }
800 
802  {
803  vu_log_debug (vui, "VHOST_USER_PROTOCOL_F_LOG_SHMFD not set but "
804  "VHOST_USER_SET_LOG_BASE received");
805  goto close_socket;
806  }
807 
808  fd = fds[0];
809  /* align size to page */
810  long page_sz = get_huge_page_size (fd);
811  ssize_t map_sz =
812  (msg.log.size + msg.log.offset + page_sz - 1) & ~(page_sz - 1);
813 
814  void *log_base_addr = mmap (0, map_sz, PROT_READ | PROT_WRITE,
815  MAP_SHARED, fd, 0);
816 
817  vu_log_debug (vui, "map log region addr 0 len 0x%lx off 0x%lx fd %d "
818  "mapped 0x%lx", map_sz, msg.log.offset, fd,
819  log_base_addr);
820 
821  if (log_base_addr == MAP_FAILED)
822  {
823  vu_log_err (vui, "failed to map memory. errno is %d", errno);
824  goto close_socket;
825  }
826 
828  vui->log_base_addr = log_base_addr;
829  vui->log_base_addr += msg.log.offset;
830  vui->log_size = msg.log.size;
832 
833  msg.flags |= 4;
834  msg.size = sizeof (msg.u64);
835  n =
836  send (uf->file_descriptor, &msg, VHOST_USER_MSG_HDR_SZ + msg.size, 0);
837  if (n != (msg.size + VHOST_USER_MSG_HDR_SZ))
838  {
839  vu_log_debug (vui, "could not send message response");
840  goto close_socket;
841  }
842  break;
843 
845  vu_log_debug (vui, "if %d msg VHOST_USER_SET_LOG_FD", vui->hw_if_index);
846  break;
847 
849  msg.flags |= 4;
850  msg.u64 = (1 << VHOST_USER_PROTOCOL_F_LOG_SHMFD) |
852  msg.size = sizeof (msg.u64);
853  vu_log_debug (vui, "if %d msg VHOST_USER_GET_PROTOCOL_FEATURES - "
854  "reply 0x%016llx", vui->hw_if_index, msg.u64);
855  n =
856  send (uf->file_descriptor, &msg, VHOST_USER_MSG_HDR_SZ + msg.size, 0);
857  if (n != (msg.size + VHOST_USER_MSG_HDR_SZ))
858  {
859  vu_log_debug (vui, "could not send message response");
860  goto close_socket;
861  }
862  break;
863 
865  vu_log_debug (vui, "if %d msg VHOST_USER_SET_PROTOCOL_FEATURES "
866  "features 0x%016llx", vui->hw_if_index, msg.u64);
867  vui->protocol_features = msg.u64;
868  break;
869 
871  msg.flags |= 4;
872  msg.u64 = VHOST_VRING_MAX_N;
873  msg.size = sizeof (msg.u64);
874  vu_log_debug (vui, "if %d msg VHOST_USER_GET_QUEUE_NUM - reply %d",
875  vui->hw_if_index, msg.u64);
876  n =
877  send (uf->file_descriptor, &msg, VHOST_USER_MSG_HDR_SZ + msg.size, 0);
878  if (n != (msg.size + VHOST_USER_MSG_HDR_SZ))
879  {
880  vu_log_debug (vui, "could not send message response");
881  goto close_socket;
882  }
883  break;
884 
886  vu_log_debug (vui, "if %d VHOST_USER_SET_VRING_ENABLE: %s queue %d",
887  vui->hw_if_index, msg.state.num ? "enable" : "disable",
888  msg.state.index);
889  if (msg.state.index >= VHOST_VRING_MAX_N)
890  {
891  vu_log_debug (vui, "invalid vring idx VHOST_USER_SET_VRING_ENABLE:"
892  " %d >= %d", msg.state.index, VHOST_VRING_MAX_N);
893  goto close_socket;
894  }
895 
896  vui->vrings[msg.state.index].enabled = msg.state.num;
897  vhost_user_thread_placement (vui, msg.state.index);
899  break;
900 
901  default:
902  vu_log_debug (vui, "unknown vhost-user message %d received. "
903  "closing socket", msg.request);
904  goto close_socket;
905  }
906 
907  return 0;
908 
909 close_socket:
914  return 0;
915 }
916 
917 static clib_error_t *
919 {
922  vhost_user_intf_t *vui =
924 
925  vu_log_debug (vui, "socket error on if %d", vui->sw_if_index);
929  return 0;
930 }
931 
932 static clib_error_t *
934 {
935  int client_fd, client_len;
936  struct sockaddr_un client;
937  clib_file_t template = { 0 };
939  vhost_user_intf_t *vui;
940 
942 
943  client_len = sizeof (client);
944  client_fd = accept (uf->file_descriptor,
945  (struct sockaddr *) &client,
946  (socklen_t *) & client_len);
947 
948  if (client_fd < 0)
949  return clib_error_return_unix (0, "accept");
950 
951  if (vui->clib_file_index != ~0)
952  {
953  vu_log_debug (vui, "Close client socket for vhost interface %d, fd %d",
956  }
957 
958  vu_log_debug (vui, "New client socket for vhost interface %d, fd %d",
959  vui->sw_if_index, client_fd);
960  template.read_function = vhost_user_socket_read;
961  template.error_function = vhost_user_socket_error;
962  template.file_descriptor = client_fd;
963  template.private_data = vui - vhost_user_main.vhost_user_interfaces;
964  vui->clib_file_index = clib_file_add (&file_main, &template);
965  return 0;
966 }
967 
968 static clib_error_t *
970 {
973 
974  vum->log_default = vlib_log_register_class ("vhost-user", 0);
975 
976  vum->coalesce_frames = 32;
977  vum->coalesce_time = 1e-3;
978 
979  vec_validate (vum->cpus, tm->n_vlib_mains - 1);
980 
981  vhost_cpu_t *cpu;
982  vec_foreach (cpu, vum->cpus)
983  {
984  /* This is actually not necessary as validate already zeroes it
985  * Just keeping the loop here for later because I am lazy. */
986  cpu->rx_buffers_len = 0;
987  }
988 
989  vum->random = random_default_seed ();
990 
992 
993  return 0;
994 }
995 
996 /* *INDENT-OFF* */
998 {
999  .runs_after = VLIB_INITS("ip4_init"),
1000 };
1001 /* *INDENT-ON* */
1002 
1003 static uword
1006 {
1007  vhost_user_intf_t *vui;
1008  f64 timeout = 3153600000.0 /* 100 years */ ;
1009  uword event_type, *event_data = 0;
1011  u16 qid;
1012  f64 now, poll_time_remaining;
1013  f64 next_timeout;
1014  u8 stop_timer = 0;
1015 
1016  while (1)
1017  {
1018  poll_time_remaining =
1020  event_type = vlib_process_get_events (vm, &event_data);
1021  vec_reset_length (event_data);
1022 
1023  /*
1024  * Use the remaining timeout if it is less than coalesce time to avoid
1025  * resetting the existing timer in the middle of expiration
1026  */
1027  timeout = poll_time_remaining;
1028  if (vlib_process_suspend_time_is_zero (timeout) ||
1029  (timeout > vum->coalesce_time))
1030  timeout = vum->coalesce_time;
1031 
1032  now = vlib_time_now (vm);
1033  switch (event_type)
1034  {
1036  stop_timer = 1;
1037  break;
1038 
1040  stop_timer = 0;
1041  if (!vlib_process_suspend_time_is_zero (poll_time_remaining))
1042  break;
1043  /* fall through */
1044 
1045  case ~0:
1046  /* *INDENT-OFF* */
1047  pool_foreach (vui, vum->vhost_user_interfaces, {
1048  next_timeout = timeout;
1049  for (qid = 0; qid < VHOST_VRING_MAX_N / 2; qid += 2)
1050  {
1051  vhost_user_vring_t *rxvq = &vui->vrings[qid];
1052  vhost_user_vring_t *txvq = &vui->vrings[qid + 1];
1053 
1054  if (txvq->qid == -1)
1055  continue;
1056  if (txvq->n_since_last_int)
1057  {
1058  if (now >= txvq->int_deadline)
1059  vhost_user_send_call (vm, txvq);
1060  else
1061  next_timeout = txvq->int_deadline - now;
1062  }
1063 
1064  if (rxvq->n_since_last_int)
1065  {
1066  if (now >= rxvq->int_deadline)
1067  vhost_user_send_call (vm, rxvq);
1068  else
1069  next_timeout = rxvq->int_deadline - now;
1070  }
1071 
1072  if ((next_timeout < timeout) && (next_timeout > 0.0))
1073  timeout = next_timeout;
1074  }
1075  });
1076  /* *INDENT-ON* */
1077  break;
1078 
1079  default:
1080  clib_warning ("BUG: unhandled event type %d", event_type);
1081  break;
1082  }
1083  /* No less than 1 millisecond */
1084  if (timeout < 1e-3)
1085  timeout = 1e-3;
1086  if (stop_timer)
1087  timeout = 3153600000.0;
1088  }
1089  return 0;
1090 }
1091 
1092 /* *INDENT-OFF* */
1095  .type = VLIB_NODE_TYPE_PROCESS,
1096  .name = "vhost-user-send-interrupt-process",
1097 };
1098 /* *INDENT-ON* */
1099 
1100 static uword
1103 {
1105  vhost_user_intf_t *vui;
1106  struct sockaddr_un sun;
1107  int sockfd;
1108  clib_file_t template = { 0 };
1109  f64 timeout = 3153600000.0 /* 100 years */ ;
1110  uword *event_data = 0;
1111 
1112  sockfd = -1;
1113  sun.sun_family = AF_UNIX;
1115  template.error_function = vhost_user_socket_error;
1116 
1117  while (1)
1118  {
1120  vlib_process_get_events (vm, &event_data);
1121  vec_reset_length (event_data);
1122 
1123  timeout = 3.0;
1124 
1125  /* *INDENT-OFF* */
1126  pool_foreach (vui, vum->vhost_user_interfaces, {
1127 
1128  if (vui->unix_server_index == ~0) { //Nothing to do for server sockets
1129  if (vui->clib_file_index == ~0)
1130  {
1131  if ((sockfd < 0) &&
1132  ((sockfd = socket (AF_UNIX, SOCK_STREAM, 0)) < 0))
1133  {
1134  /*
1135  * 1st time error or new error for this interface,
1136  * spit out the message and record the error
1137  */
1138  if (!vui->sock_errno || (vui->sock_errno != errno))
1139  {
1140  clib_unix_warning
1141  ("Error: Could not open unix socket for %s",
1142  vui->sock_filename);
1143  vui->sock_errno = errno;
1144  }
1145  continue;
1146  }
1147 
1148  /* try to connect */
1149  strncpy (sun.sun_path, (char *) vui->sock_filename,
1150  sizeof (sun.sun_path) - 1);
1151 
1152  /* Avoid hanging VPP if the other end does not accept */
1153  if (fcntl(sockfd, F_SETFL, O_NONBLOCK) < 0)
1154  clib_unix_warning ("fcntl");
1155 
1156  if (connect (sockfd, (struct sockaddr *) &sun,
1157  sizeof (struct sockaddr_un)) == 0)
1158  {
1159  /* Set the socket to blocking as it was before */
1160  if (fcntl(sockfd, F_SETFL, 0) < 0)
1161  clib_unix_warning ("fcntl2");
1162 
1163  vui->sock_errno = 0;
1164  template.file_descriptor = sockfd;
1165  template.private_data =
1166  vui - vhost_user_main.vhost_user_interfaces;
1167  vui->clib_file_index = clib_file_add (&file_main, &template);
1168 
1169  /* This sockfd is considered consumed */
1170  sockfd = -1;
1171  }
1172  else
1173  {
1174  vui->sock_errno = errno;
1175  }
1176  }
1177  else
1178  {
1179  /* check if socket is alive */
1180  int error = 0;
1181  socklen_t len = sizeof (error);
1182  int fd = UNIX_GET_FD(vui->clib_file_index);
1183  int retval =
1184  getsockopt (fd, SOL_SOCKET, SO_ERROR, &error, &len);
1185 
1186  if (retval)
1187  {
1188  vu_log_debug (vui, "getsockopt returned %d", retval);
1189  vhost_user_if_disconnect (vui);
1190  }
1191  }
1192  }
1193  });
1194  /* *INDENT-ON* */
1195  }
1196  return 0;
1197 }
1198 
1199 /* *INDENT-OFF* */
1201  .function = vhost_user_process,
1202  .type = VLIB_NODE_TYPE_PROCESS,
1203  .name = "vhost-user-process",
1204 };
1205 /* *INDENT-ON* */
1206 
1207 /**
1208  * Disables and reset interface structure.
1209  * It can then be either init again, or removed from used interfaces.
1210  */
1211 static void
1213 {
1214  int q;
1216 
1217  // disconnect interface sockets
1219  vhost_user_update_gso_interface_count (vui, 0 /* delete */ );
1221 
1222  for (q = 0; q < VHOST_VRING_MAX_N; q++)
1223  {
1224  // Remove existing queue mapping for the interface
1225  if (q & 1)
1226  {
1227  int rv;
1228  vnet_main_t *vnm = vnet_get_main ();
1229  vhost_user_vring_t *txvq = &vui->vrings[q];
1230 
1231  if (txvq->qid != -1)
1232  {
1234  vui->hw_if_index,
1235  q >> 1);
1236  if (rv)
1237  vu_log_warn (vui, "unable to unassign interface %d, "
1238  "queue %d: rc=%d", vui->hw_if_index, q >> 1, rv);
1239  }
1240  }
1241 
1242  clib_mem_free ((void *) vui->vring_locks[q]);
1243  }
1244 
1245  if (vui->unix_server_index != ~0)
1246  {
1247  //Close server socket
1249  vui->unix_server_index);
1250  clib_file_del (&file_main, uf);
1251  vui->unix_server_index = ~0;
1252  unlink (vui->sock_filename);
1253  }
1254 
1256  &vui->if_index);
1257 }
1258 
1259 int
1261 {
1263  vhost_user_intf_t *vui;
1264  int rv = 0;
1265  vnet_hw_interface_t *hwif;
1266  u16 qid;
1267 
1268  if (!
1269  (hwif =
1271  || hwif->dev_class_index != vhost_user_device_class.index)
1272  return VNET_API_ERROR_INVALID_SW_IF_INDEX;
1273 
1275 
1276  vu_log_debug (vui, "Deleting vhost-user interface %s (instance %d)",
1277  hwif->name, hwif->dev_instance);
1278 
1279  for (qid = 1; qid < VHOST_VRING_MAX_N / 2; qid += 2)
1280  {
1281  vhost_user_vring_t *txvq = &vui->vrings[qid];
1282 
1283  if (txvq->qid == -1)
1284  continue;
1285  if ((vum->ifq_count > 0) &&
1288  {
1289  vum->ifq_count--;
1290  // Stop the timer if there is no more interrupt interface/queue
1291  if ((vum->ifq_count == 0) &&
1292  (vum->coalesce_time > 0.0) && (vum->coalesce_frames > 0))
1293  {
1297  break;
1298  }
1299  }
1300  }
1301 
1302  // Disable and reset interface
1303  vhost_user_term_if (vui);
1304 
1305  // Reset renumbered iface
1306  if (hwif->dev_instance <
1309 
1310  // Delete ethernet interface
1312 
1313  // Back to pool
1314  pool_put (vum->vhost_user_interfaces, vui);
1315 
1316  return rv;
1317 }
1318 
1319 static clib_error_t *
1321 {
1322  vnet_main_t *vnm = vnet_get_main ();
1324  vhost_user_intf_t *vui;
1325 
1327  /* *INDENT-OFF* */
1328  pool_foreach (vui, vum->vhost_user_interfaces, {
1329  vhost_user_delete_if (vnm, vm, vui->sw_if_index);
1330  });
1331  /* *INDENT-ON* */
1333  return 0;
1334 }
1335 
1337 
1338 /**
1339  * Open server unix socket on specified sock_filename.
1340  */
1341 static int
1342 vhost_user_init_server_sock (const char *sock_filename, int *sock_fd)
1343 {
1344  int rv = 0;
1345  struct sockaddr_un un = { };
1346  int fd;
1347  /* create listening socket */
1348  if ((fd = socket (AF_UNIX, SOCK_STREAM, 0)) < 0)
1349  return VNET_API_ERROR_SYSCALL_ERROR_1;
1350 
1351  un.sun_family = AF_UNIX;
1352  strncpy ((char *) un.sun_path, (char *) sock_filename,
1353  sizeof (un.sun_path) - 1);
1354 
1355  /* remove if exists */
1356  unlink ((char *) sock_filename);
1357 
1358  if (bind (fd, (struct sockaddr *) &un, sizeof (un)) == -1)
1359  {
1360  rv = VNET_API_ERROR_SYSCALL_ERROR_2;
1361  goto error;
1362  }
1363 
1364  if (listen (fd, 1) == -1)
1365  {
1366  rv = VNET_API_ERROR_SYSCALL_ERROR_3;
1367  goto error;
1368  }
1369 
1370  *sock_fd = fd;
1371  return 0;
1372 
1373 error:
1374  close (fd);
1375  return rv;
1376 }
1377 
1378 /**
1379  * Create ethernet interface for vhost user interface.
1380  */
1381 static void
1383  vhost_user_intf_t * vui, u8 * hwaddress)
1384 {
1386  u8 hwaddr[6];
1387  clib_error_t *error;
1388 
1389  /* create hw and sw interface */
1390  if (hwaddress)
1391  {
1392  clib_memcpy (hwaddr, hwaddress, 6);
1393  }
1394  else
1395  {
1396  random_u32 (&vum->random);
1397  clib_memcpy (hwaddr + 2, &vum->random, sizeof (vum->random));
1398  hwaddr[0] = 2;
1399  hwaddr[1] = 0xfe;
1400  }
1401 
1403  (vnm,
1405  vui - vum->vhost_user_interfaces /* device instance */ ,
1406  hwaddr /* ethernet address */ ,
1407  &vui->hw_if_index, 0 /* flag change */ );
1408 
1409  if (error)
1410  clib_error_report (error);
1411 }
1412 
1413 /*
1414  * Initialize vui with specified attributes
1415  */
1416 static void
1418  vhost_user_intf_t * vui,
1419  int server_sock_fd,
1420  const char *sock_filename,
1421  u64 feature_mask, u32 * sw_if_index, u8 enable_gso)
1422 {
1423  vnet_sw_interface_t *sw;
1424  int q;
1426  vnet_hw_interface_t *hw;
1427 
1428  hw = vnet_get_hw_interface (vnm, vui->hw_if_index);
1429  sw = vnet_get_hw_sw_interface (vnm, vui->hw_if_index);
1430  if (server_sock_fd != -1)
1431  {
1432  clib_file_t template = { 0 };
1434  template.file_descriptor = server_sock_fd;
1435  template.private_data = vui - vum->vhost_user_interfaces; //hw index
1436  vui->unix_server_index = clib_file_add (&file_main, &template);
1437  }
1438  else
1439  {
1440  vui->unix_server_index = ~0;
1441  }
1442 
1443  vui->sw_if_index = sw->sw_if_index;
1444  strncpy (vui->sock_filename, sock_filename,
1445  ARRAY_LEN (vui->sock_filename) - 1);
1446  vui->sock_errno = 0;
1447  vui->is_ready = 0;
1448  vui->feature_mask = feature_mask;
1449  vui->clib_file_index = ~0;
1450  vui->log_base_addr = 0;
1451  vui->if_index = vui - vum->vhost_user_interfaces;
1452  vui->enable_gso = enable_gso;
1453  /*
1454  * enable_gso takes precedence over configurable feature mask if there
1455  * is a clash.
1456  * if feature mask disables gso, but enable_gso is configured,
1457  * then gso is enable
1458  * if feature mask enables gso, but enable_gso is not configured,
1459  * then gso is enable
1460  *
1461  * if gso is enable via feature mask, it must enable both host and guest
1462  * gso feature mask, we don't support one sided GSO or partial GSO.
1463  */
1464  if ((vui->enable_gso == 0) &&
1466  (FEATURE_VIRTIO_NET_F_HOST_GUEST_TSO_FEATURE_BITS)))
1467  vui->enable_gso = 1;
1468  vhost_user_update_gso_interface_count (vui, 1 /* add */ );
1470  &vui->if_index, 0);
1471 
1472  for (q = 0; q < VHOST_VRING_MAX_N; q++)
1473  vhost_user_vring_init (vui, q);
1474 
1477 
1478  if (sw_if_index)
1479  *sw_if_index = vui->sw_if_index;
1480 
1481  for (q = 0; q < VHOST_VRING_MAX_N; q++)
1482  {
1485  clib_memset ((void *) vui->vring_locks[q], 0, CLIB_CACHE_LINE_BYTES);
1486  }
1487 
1489  vlib_get_thread_main ()->n_vlib_mains - 1);
1491 }
1492 
1493 int
1495  const char *sock_filename,
1496  u8 is_server,
1497  u32 * sw_if_index,
1498  u64 feature_mask,
1499  u8 renumber, u32 custom_dev_instance, u8 * hwaddr,
1500  u8 enable_gso)
1501 {
1502  vhost_user_intf_t *vui = NULL;
1503  u32 sw_if_idx = ~0;
1504  int rv = 0;
1505  int server_sock_fd = -1;
1507  uword *if_index;
1508 
1509  if (sock_filename == NULL || !(strlen (sock_filename) > 0))
1510  {
1511  return VNET_API_ERROR_INVALID_ARGUMENT;
1512  }
1513 
1514  if_index = mhash_get (&vum->if_index_by_sock_name, (void *) sock_filename);
1515  if (if_index)
1516  {
1517  if (sw_if_index)
1518  {
1519  vui = &vum->vhost_user_interfaces[*if_index];
1520  *sw_if_index = vui->sw_if_index;
1521  }
1522  return VNET_API_ERROR_IF_ALREADY_EXISTS;
1523  }
1524 
1525  if (is_server)
1526  {
1527  if ((rv =
1528  vhost_user_init_server_sock (sock_filename, &server_sock_fd)) != 0)
1529  {
1530  return rv;
1531  }
1532  }
1533 
1534  /* Protect the uninitialized vui from being dispatched by rx/tx */
1536  pool_get (vhost_user_main.vhost_user_interfaces, vui);
1537  vhost_user_create_ethernet (vnm, vm, vui, hwaddr);
1539 
1540  vhost_user_vui_init (vnm, vui, server_sock_fd, sock_filename,
1541  feature_mask, &sw_if_idx, enable_gso);
1542  vnet_sw_interface_set_mtu (vnm, vui->sw_if_index, 9000);
1544 
1545  if (renumber)
1546  vnet_interface_name_renumber (sw_if_idx, custom_dev_instance);
1547 
1548  if (sw_if_index)
1549  *sw_if_index = sw_if_idx;
1550 
1551  // Process node must connect
1553 
1554  return rv;
1555 }
1556 
1557 int
1559  const char *sock_filename,
1560  u8 is_server,
1561  u32 sw_if_index,
1562  u64 feature_mask, u8 renumber, u32 custom_dev_instance,
1563  u8 enable_gso)
1564 {
1566  vhost_user_intf_t *vui = NULL;
1567  u32 sw_if_idx = ~0;
1568  int server_sock_fd = -1;
1569  int rv = 0;
1570  vnet_hw_interface_t *hwif;
1571  uword *if_index;
1572 
1573  if (!
1574  (hwif =
1576  || hwif->dev_class_index != vhost_user_device_class.index)
1577  return VNET_API_ERROR_INVALID_SW_IF_INDEX;
1578 
1579  if (sock_filename == NULL || !(strlen (sock_filename) > 0))
1580  return VNET_API_ERROR_INVALID_ARGUMENT;
1581 
1583 
1584  /*
1585  * Disallow changing the interface to have the same path name
1586  * as other interface
1587  */
1588  if_index = mhash_get (&vum->if_index_by_sock_name, (void *) sock_filename);
1589  if (if_index && (*if_index != vui->if_index))
1590  return VNET_API_ERROR_IF_ALREADY_EXISTS;
1591 
1592  // First try to open server socket
1593  if (is_server)
1594  if ((rv = vhost_user_init_server_sock (sock_filename,
1595  &server_sock_fd)) != 0)
1596  return rv;
1597 
1598  vhost_user_term_if (vui);
1599  vhost_user_vui_init (vnm, vui, server_sock_fd,
1600  sock_filename, feature_mask, &sw_if_idx, enable_gso);
1601 
1602  if (renumber)
1603  vnet_interface_name_renumber (sw_if_idx, custom_dev_instance);
1604 
1605  // Process node must connect
1607 
1608  return rv;
1609 }
1610 
1611 clib_error_t *
1613  unformat_input_t * input,
1614  vlib_cli_command_t * cmd)
1615 {
1616  unformat_input_t _line_input, *line_input = &_line_input;
1617  u8 *sock_filename = NULL;
1618  u32 sw_if_index;
1619  u8 is_server = 0;
1620  u64 feature_mask = (u64) ~ (0ULL);
1621  u8 renumber = 0;
1622  u32 custom_dev_instance = ~0;
1623  u8 hwaddr[6];
1624  u8 *hw = NULL;
1625  clib_error_t *error = NULL;
1626  u8 enable_gso = 0;
1627 
1628  /* Get a line of input. */
1629  if (!unformat_user (input, unformat_line_input, line_input))
1630  return 0;
1631 
1632  /* GSO feature is disable by default */
1634  while (unformat_check_input (line_input) != UNFORMAT_END_OF_INPUT)
1635  {
1636  if (unformat (line_input, "socket %s", &sock_filename))
1637  ;
1638  else if (unformat (line_input, "server"))
1639  is_server = 1;
1640  else if (unformat (line_input, "gso"))
1641  enable_gso = 1;
1642  else if (unformat (line_input, "feature-mask 0x%llx", &feature_mask))
1643  ;
1644  else
1645  if (unformat
1646  (line_input, "hwaddr %U", unformat_ethernet_address, hwaddr))
1647  hw = hwaddr;
1648  else if (unformat (line_input, "renumber %d", &custom_dev_instance))
1649  {
1650  renumber = 1;
1651  }
1652  else
1653  {
1654  error = clib_error_return (0, "unknown input `%U'",
1655  format_unformat_error, line_input);
1656  goto done;
1657  }
1658  }
1659 
1660  vnet_main_t *vnm = vnet_get_main ();
1661 
1662  int rv;
1663  if ((rv = vhost_user_create_if (vnm, vm, (char *) sock_filename,
1664  is_server, &sw_if_index, feature_mask,
1665  renumber, custom_dev_instance, hw,
1666  enable_gso)))
1667  {
1668  error = clib_error_return (0, "vhost_user_create_if returned %d", rv);
1669  goto done;
1670  }
1671 
1673  sw_if_index);
1674 
1675 done:
1676  vec_free (sock_filename);
1677  unformat_free (line_input);
1678 
1679  return error;
1680 }
1681 
1682 clib_error_t *
1684  unformat_input_t * input,
1685  vlib_cli_command_t * cmd)
1686 {
1687  unformat_input_t _line_input, *line_input = &_line_input;
1688  u32 sw_if_index = ~0;
1689  vnet_main_t *vnm = vnet_get_main ();
1690  clib_error_t *error = NULL;
1691 
1692  /* Get a line of input. */
1693  if (!unformat_user (input, unformat_line_input, line_input))
1694  return 0;
1695 
1696  while (unformat_check_input (line_input) != UNFORMAT_END_OF_INPUT)
1697  {
1698  if (unformat (line_input, "sw_if_index %d", &sw_if_index))
1699  ;
1700  else if (unformat
1701  (line_input, "%U", unformat_vnet_sw_interface, vnm,
1702  &sw_if_index))
1703  {
1704  vnet_hw_interface_t *hwif =
1706  if (hwif == NULL ||
1708  {
1709  error = clib_error_return (0, "Not a vhost interface");
1710  goto done;
1711  }
1712  }
1713  else
1714  {
1715  error = clib_error_return (0, "unknown input `%U'",
1716  format_unformat_error, line_input);
1717  goto done;
1718  }
1719  }
1720 
1721  vhost_user_delete_if (vnm, vm, sw_if_index);
1722 
1723 done:
1724  unformat_free (line_input);
1725 
1726  return error;
1727 }
1728 
1729 int
1731  vhost_user_intf_details_t ** out_vuids)
1732 {
1733  int rv = 0;
1735  vhost_user_intf_t *vui;
1736  vhost_user_intf_details_t *r_vuids = NULL;
1738  u32 *hw_if_indices = 0;
1740  u8 *s = NULL;
1741  int i;
1742 
1743  if (!out_vuids)
1744  return -1;
1745 
1747  vec_add1 (hw_if_indices, vui->hw_if_index);
1748  );
1749 
1750  for (i = 0; i < vec_len (hw_if_indices); i++)
1751  {
1752  hi = vnet_get_hw_interface (vnm, hw_if_indices[i]);
1754 
1755  vec_add2 (r_vuids, vuid, 1);
1756  vuid->sw_if_index = vui->sw_if_index;
1757  vuid->virtio_net_hdr_sz = vui->virtio_net_hdr_sz;
1758  vuid->features = vui->features;
1759  vuid->num_regions = vui->nregions;
1760  vuid->is_server = vui->unix_server_index != ~0;
1761  vuid->sock_errno = vui->sock_errno;
1762  strncpy ((char *) vuid->sock_filename, (char *) vui->sock_filename,
1763  sizeof (vuid->sock_filename));
1764  vuid->sock_filename[ARRAY_LEN (vuid->sock_filename) - 1] = '\0';
1765  s = format (s, "%v%c", hi->name, 0);
1766 
1767  strncpy ((char *) vuid->if_name, (char *) s,
1768  ARRAY_LEN (vuid->if_name) - 1);
1769  _vec_len (s) = 0;
1770  }
1771 
1772  vec_free (s);
1773  vec_free (hw_if_indices);
1774 
1775  *out_vuids = r_vuids;
1776 
1777  return rv;
1778 }
1779 
1780 clib_error_t *
1782  unformat_input_t * input,
1783  vlib_cli_command_t * cmd)
1784 {
1785  clib_error_t *error = 0;
1786  vnet_main_t *vnm = vnet_get_main ();
1788  vhost_user_intf_t *vui;
1789  u32 hw_if_index, *hw_if_indices = 0;
1791  u16 qid;
1792  u32 ci;
1793  int i, j, q;
1794  int show_descr = 0;
1795  struct feat_struct
1796  {
1797  u8 bit;
1798  char *str;
1799  };
1800  struct feat_struct *feat_entry;
1801 
1802  static struct feat_struct feat_array[] = {
1803 #define _(s,b) { .str = #s, .bit = b, },
1805 #undef _
1806  {.str = NULL}
1807  };
1808 
1809 #define foreach_protocol_feature \
1810  _(VHOST_USER_PROTOCOL_F_MQ) \
1811  _(VHOST_USER_PROTOCOL_F_LOG_SHMFD)
1812 
1813  static struct feat_struct proto_feat_array[] = {
1814 #define _(s) { .str = #s, .bit = s},
1816 #undef _
1817  {.str = NULL}
1818  };
1819 
1820  while (unformat_check_input (input) != UNFORMAT_END_OF_INPUT)
1821  {
1822  if (unformat
1823  (input, "%U", unformat_vnet_hw_interface, vnm, &hw_if_index))
1824  {
1825  hi = vnet_get_hw_interface (vnm, hw_if_index);
1826  if (vhost_user_device_class.index != hi->dev_class_index)
1827  {
1828  error = clib_error_return (0, "unknown input `%U'",
1829  format_unformat_error, input);
1830  goto done;
1831  }
1832  vec_add1 (hw_if_indices, hw_if_index);
1833  }
1834  else if (unformat (input, "descriptors") || unformat (input, "desc"))
1835  show_descr = 1;
1836  else
1837  {
1838  error = clib_error_return (0, "unknown input `%U'",
1839  format_unformat_error, input);
1840  goto done;
1841  }
1842  }
1843  if (vec_len (hw_if_indices) == 0)
1844  {
1846  vec_add1 (hw_if_indices, vui->hw_if_index);
1847  );
1848  }
1849  vlib_cli_output (vm, "Virtio vhost-user interfaces");
1850  vlib_cli_output (vm, "Global:\n coalesce frames %d time %e",
1851  vum->coalesce_frames, vum->coalesce_time);
1852  vlib_cli_output (vm, " Number of rx virtqueues in interrupt mode: %d",
1853  vum->ifq_count);
1854  vlib_cli_output (vm, " Number of GSO interfaces: %d", vum->gso_count);
1855 
1856  for (i = 0; i < vec_len (hw_if_indices); i++)
1857  {
1858  hi = vnet_get_hw_interface (vnm, hw_if_indices[i]);
1860  vlib_cli_output (vm, "Interface: %U (ifindex %d)",
1861  format_vnet_hw_if_index_name, vnm, hw_if_indices[i],
1862  hw_if_indices[i]);
1863  if (vui->enable_gso)
1864  vlib_cli_output (vm, " GSO enable");
1865 
1866  vlib_cli_output (vm, "virtio_net_hdr_sz %d\n"
1867  " features mask (0x%llx): \n"
1868  " features (0x%llx): \n",
1869  vui->virtio_net_hdr_sz, vui->feature_mask,
1870  vui->features);
1871 
1872  feat_entry = (struct feat_struct *) &feat_array;
1873  while (feat_entry->str)
1874  {
1875  if (vui->features & (1ULL << feat_entry->bit))
1876  vlib_cli_output (vm, " %s (%d)", feat_entry->str,
1877  feat_entry->bit);
1878  feat_entry++;
1879  }
1880 
1881  vlib_cli_output (vm, " protocol features (0x%llx)",
1882  vui->protocol_features);
1883  feat_entry = (struct feat_struct *) &proto_feat_array;
1884  while (feat_entry->str)
1885  {
1886  if (vui->protocol_features & (1ULL << feat_entry->bit))
1887  vlib_cli_output (vm, " %s (%d)", feat_entry->str,
1888  feat_entry->bit);
1889  feat_entry++;
1890  }
1891 
1892  vlib_cli_output (vm, "\n");
1893 
1894  vlib_cli_output (vm, " socket filename %s type %s errno \"%s\"\n\n",
1895  vui->sock_filename,
1896  (vui->unix_server_index != ~0) ? "server" : "client",
1897  strerror (vui->sock_errno));
1898 
1899  vlib_cli_output (vm, " rx placement: ");
1900 
1901  for (qid = 1; qid < VHOST_VRING_MAX_N / 2; qid += 2)
1902  {
1903  vnet_main_t *vnm = vnet_get_main ();
1904  uword thread_index;
1906  vhost_user_vring_t *txvq = &vui->vrings[qid];
1907 
1908  if (txvq->qid == -1)
1909  continue;
1910  thread_index =
1912  qid >> 1);
1913  vnet_hw_interface_get_rx_mode (vnm, vui->hw_if_index, qid >> 1,
1914  &mode);
1915  vlib_cli_output (vm, " thread %d on vring %d, %U\n",
1916  thread_index, qid,
1918  }
1919 
1920  vlib_cli_output (vm, " tx placement: %s\n",
1921  vui->use_tx_spinlock ? "spin-lock" : "lock-free");
1922 
1924  {
1925  vlib_cli_output (vm, " thread %d on vring %d\n", ci,
1926  VHOST_VRING_IDX_RX (vui->per_cpu_tx_qid[ci]));
1927  }
1928 
1929  vlib_cli_output (vm, "\n");
1930 
1931  vlib_cli_output (vm, " Memory regions (total %d)\n", vui->nregions);
1932 
1933  if (vui->nregions)
1934  {
1935  vlib_cli_output (vm,
1936  " region fd guest_phys_addr memory_size userspace_addr mmap_offset mmap_addr\n");
1937  vlib_cli_output (vm,
1938  " ====== ===== ================== ================== ================== ================== ==================\n");
1939  }
1940  for (j = 0; j < vui->nregions; j++)
1941  {
1942  vlib_cli_output (vm,
1943  " %d %-5d 0x%016lx 0x%016lx 0x%016lx 0x%016lx 0x%016lx\n",
1944  j, vui->region_mmap_fd[j],
1945  vui->regions[j].guest_phys_addr,
1946  vui->regions[j].memory_size,
1947  vui->regions[j].userspace_addr,
1948  vui->regions[j].mmap_offset,
1950  }
1951  for (q = 0; q < VHOST_VRING_MAX_N; q++)
1952  {
1953  if (!vui->vrings[q].started)
1954  continue;
1955 
1956  vlib_cli_output (vm, "\n Virtqueue %d (%s%s)\n", q,
1957  (q & 1) ? "RX" : "TX",
1958  vui->vrings[q].enabled ? "" : " disabled");
1959 
1960  vlib_cli_output (vm,
1961  " qsz %d last_avail_idx %d last_used_idx %d\n",
1962  vui->vrings[q].qsz_mask + 1,
1963  vui->vrings[q].last_avail_idx,
1964  vui->vrings[q].last_used_idx);
1965 
1966  if (vui->vrings[q].avail && vui->vrings[q].used)
1967  vlib_cli_output (vm,
1968  " avail.flags %x avail.idx %d used.flags %x used.idx %d\n",
1969  vui->vrings[q].avail->flags,
1970  vui->vrings[q].avail->idx,
1971  vui->vrings[q].used->flags,
1972  vui->vrings[q].used->idx);
1973 
1974  int kickfd = UNIX_GET_FD (vui->vrings[q].kickfd_idx);
1975  int callfd = UNIX_GET_FD (vui->vrings[q].callfd_idx);
1976  vlib_cli_output (vm, " kickfd %d callfd %d errfd %d\n",
1977  kickfd, callfd, vui->vrings[q].errfd);
1978 
1979  if (show_descr)
1980  {
1981  vlib_cli_output (vm, "\n descriptor table:\n");
1982  vlib_cli_output (vm,
1983  " id addr len flags next user_addr\n");
1984  vlib_cli_output (vm,
1985  " ===== ================== ===== ====== ===== ==================\n");
1986  for (j = 0; j < vui->vrings[q].qsz_mask + 1; j++)
1987  {
1988  u32 mem_hint = 0;
1989  vlib_cli_output (vm,
1990  " %-5d 0x%016lx %-5d 0x%04x %-5d 0x%016lx\n",
1991  j, vui->vrings[q].desc[j].addr,
1992  vui->vrings[q].desc[j].len,
1993  vui->vrings[q].desc[j].flags,
1994  vui->vrings[q].desc[j].next,
1996  (vui,
1997  vui->vrings[q].desc[j].
1998  addr, &mem_hint)));
1999  }
2000  }
2001  }
2002  vlib_cli_output (vm, "\n");
2003  }
2004 done:
2005  vec_free (hw_if_indices);
2006  return error;
2007 }
2008 
2009 /*
2010  * CLI functions
2011  */
2012 
2013 /*?
2014  * Create a vHost User interface. Once created, a new virtual interface
2015  * will exist with the name '<em>VirtualEthernet0/0/x</em>', where '<em>x</em>'
2016  * is the next free index.
2017  *
2018  * There are several parameters associated with a vHost interface:
2019  *
2020  * - <b>socket <socket-filename></b> - Name of the linux socket used by hypervisor
2021  * and VPP to manage the vHost interface. If in '<em>server</em>' mode, VPP will
2022  * create the socket if it does not already exist. If in '<em>client</em>' mode,
2023  * hypervisor will create the socket if it does not already exist. The VPP code
2024  * is indifferent to the file location. However, if SELinux is enabled, then the
2025  * socket needs to be created in '<em>/var/run/vpp/</em>'.
2026  *
2027  * - <b>server</b> - Optional flag to indicate that VPP should be the server for
2028  * the linux socket. If not provided, VPP will be the client. In '<em>server</em>'
2029  * mode, the VM can be reset without tearing down the vHost Interface. In
2030  * '<em>client</em>' mode, VPP can be reset without bringing down the VM and
2031  * tearing down the vHost Interface.
2032  *
2033  * - <b>feature-mask <hex></b> - Optional virtio/vhost feature set negotiated at
2034  * startup. <b>This is intended for degugging only.</b> It is recommended that this
2035  * parameter not be used except by experienced users. By default, all supported
2036  * features will be advertised. Otherwise, provide the set of features desired.
2037  * - 0x000008000 (15) - VIRTIO_NET_F_MRG_RXBUF
2038  * - 0x000020000 (17) - VIRTIO_NET_F_CTRL_VQ
2039  * - 0x000200000 (21) - VIRTIO_NET_F_GUEST_ANNOUNCE
2040  * - 0x000400000 (22) - VIRTIO_NET_F_MQ
2041  * - 0x004000000 (26) - VHOST_F_LOG_ALL
2042  * - 0x008000000 (27) - VIRTIO_F_ANY_LAYOUT
2043  * - 0x010000000 (28) - VIRTIO_F_INDIRECT_DESC
2044  * - 0x040000000 (30) - VHOST_USER_F_PROTOCOL_FEATURES
2045  * - 0x100000000 (32) - VIRTIO_F_VERSION_1
2046  *
2047  * - <b>hwaddr <mac-addr></b> - Optional ethernet address, can be in either
2048  * X:X:X:X:X:X unix or X.X.X cisco format.
2049  *
2050  * - <b>renumber <dev_instance></b> - Optional parameter which allows the instance
2051  * in the name to be specified. If instance already exists, name will be used
2052  * anyway and multiple instances will have the same name. Use with caution.
2053  *
2054  * @cliexpar
2055  * Example of how to create a vhost interface with VPP as the client and all features enabled:
2056  * @cliexstart{create vhost-user socket /var/run/vpp/vhost1.sock}
2057  * VirtualEthernet0/0/0
2058  * @cliexend
2059  * Example of how to create a vhost interface with VPP as the server and with just
2060  * multiple queues enabled:
2061  * @cliexstart{create vhost-user socket /var/run/vpp/vhost2.sock server feature-mask 0x40400000}
2062  * VirtualEthernet0/0/1
2063  * @cliexend
2064  * Once the vHost interface is created, enable the interface using:
2065  * @cliexcmd{set interface state VirtualEthernet0/0/0 up}
2066 ?*/
2067 /* *INDENT-OFF* */
2068 VLIB_CLI_COMMAND (vhost_user_connect_command, static) = {
2069  .path = "create vhost-user",
2070  .short_help = "create vhost-user socket <socket-filename> [server] "
2071  "[feature-mask <hex>] [hwaddr <mac-addr>] [renumber <dev_instance>] [gso]",
2072  .function = vhost_user_connect_command_fn,
2073  .is_mp_safe = 1,
2074 };
2075 /* *INDENT-ON* */
2076 
2077 /*?
2078  * Delete a vHost User interface using the interface name or the
2079  * software interface index. Use the '<em>show interface</em>'
2080  * command to determine the software interface index. On deletion,
2081  * the linux socket will not be deleted.
2082  *
2083  * @cliexpar
2084  * Example of how to delete a vhost interface by name:
2085  * @cliexcmd{delete vhost-user VirtualEthernet0/0/1}
2086  * Example of how to delete a vhost interface by software interface index:
2087  * @cliexcmd{delete vhost-user sw_if_index 1}
2088 ?*/
2089 /* *INDENT-OFF* */
2090 VLIB_CLI_COMMAND (vhost_user_delete_command, static) = {
2091  .path = "delete vhost-user",
2092  .short_help = "delete vhost-user {<interface> | sw_if_index <sw_idx>}",
2093  .function = vhost_user_delete_command_fn,
2094 };
2095 
2096 /*?
2097  * Display the attributes of a single vHost User interface (provide interface
2098  * name), multiple vHost User interfaces (provide a list of interface names seperated
2099  * by spaces) or all Vhost User interfaces (omit an interface name to display all
2100  * vHost interfaces).
2101  *
2102  * @cliexpar
2103  * @parblock
2104  * Example of how to display a vhost interface:
2105  * @cliexstart{show vhost-user VirtualEthernet0/0/0}
2106  * Virtio vhost-user interfaces
2107  * Global:
2108  * coalesce frames 32 time 1e-3
2109  * Interface: VirtualEthernet0/0/0 (ifindex 1)
2110  * virtio_net_hdr_sz 12
2111  * features mask (0xffffffffffffffff):
2112  * features (0x50408000):
2113  * VIRTIO_NET_F_MRG_RXBUF (15)
2114  * VIRTIO_NET_F_MQ (22)
2115  * VIRTIO_F_INDIRECT_DESC (28)
2116  * VHOST_USER_F_PROTOCOL_FEATURES (30)
2117  * protocol features (0x3)
2118  * VHOST_USER_PROTOCOL_F_MQ (0)
2119  * VHOST_USER_PROTOCOL_F_LOG_SHMFD (1)
2120  *
2121  * socket filename /var/run/vpp/vhost1.sock type client errno "Success"
2122  *
2123  * rx placement:
2124  * thread 1 on vring 1
2125  * thread 1 on vring 5
2126  * thread 2 on vring 3
2127  * thread 2 on vring 7
2128  * tx placement: spin-lock
2129  * thread 0 on vring 0
2130  * thread 1 on vring 2
2131  * thread 2 on vring 0
2132  *
2133  * Memory regions (total 2)
2134  * region fd guest_phys_addr memory_size userspace_addr mmap_offset mmap_addr
2135  * ====== ===== ================== ================== ================== ================== ==================
2136  * 0 60 0x0000000000000000 0x00000000000a0000 0x00002aaaaac00000 0x0000000000000000 0x00002aab2b400000
2137  * 1 61 0x00000000000c0000 0x000000003ff40000 0x00002aaaaacc0000 0x00000000000c0000 0x00002aababcc0000
2138  *
2139  * Virtqueue 0 (TX)
2140  * qsz 256 last_avail_idx 0 last_used_idx 0
2141  * avail.flags 1 avail.idx 128 used.flags 1 used.idx 0
2142  * kickfd 62 callfd 64 errfd -1
2143  *
2144  * Virtqueue 1 (RX)
2145  * qsz 256 last_avail_idx 0 last_used_idx 0
2146  * avail.flags 1 avail.idx 0 used.flags 1 used.idx 0
2147  * kickfd 65 callfd 66 errfd -1
2148  *
2149  * Virtqueue 2 (TX)
2150  * qsz 256 last_avail_idx 0 last_used_idx 0
2151  * avail.flags 1 avail.idx 128 used.flags 1 used.idx 0
2152  * kickfd 63 callfd 70 errfd -1
2153  *
2154  * Virtqueue 3 (RX)
2155  * qsz 256 last_avail_idx 0 last_used_idx 0
2156  * avail.flags 1 avail.idx 0 used.flags 1 used.idx 0
2157  * kickfd 72 callfd 74 errfd -1
2158  *
2159  * Virtqueue 4 (TX disabled)
2160  * qsz 256 last_avail_idx 0 last_used_idx 0
2161  * avail.flags 1 avail.idx 0 used.flags 1 used.idx 0
2162  * kickfd 76 callfd 78 errfd -1
2163  *
2164  * Virtqueue 5 (RX disabled)
2165  * qsz 256 last_avail_idx 0 last_used_idx 0
2166  * avail.flags 1 avail.idx 0 used.flags 1 used.idx 0
2167  * kickfd 80 callfd 82 errfd -1
2168  *
2169  * Virtqueue 6 (TX disabled)
2170  * qsz 256 last_avail_idx 0 last_used_idx 0
2171  * avail.flags 1 avail.idx 0 used.flags 1 used.idx 0
2172  * kickfd 84 callfd 86 errfd -1
2173  *
2174  * Virtqueue 7 (RX disabled)
2175  * qsz 256 last_avail_idx 0 last_used_idx 0
2176  * avail.flags 1 avail.idx 0 used.flags 1 used.idx 0
2177  * kickfd 88 callfd 90 errfd -1
2178  *
2179  * @cliexend
2180  *
2181  * The optional '<em>descriptors</em>' parameter will display the same output as
2182  * the previous example but will include the descriptor table for each queue.
2183  * The output is truncated below:
2184  * @cliexstart{show vhost-user VirtualEthernet0/0/0 descriptors}
2185  * Virtio vhost-user interfaces
2186  * Global:
2187  * coalesce frames 32 time 1e-3
2188  * Interface: VirtualEthernet0/0/0 (ifindex 1)
2189  * virtio_net_hdr_sz 12
2190  * features mask (0xffffffffffffffff):
2191  * features (0x50408000):
2192  * VIRTIO_NET_F_MRG_RXBUF (15)
2193  * VIRTIO_NET_F_MQ (22)
2194  * :
2195  * Virtqueue 0 (TX)
2196  * qsz 256 last_avail_idx 0 last_used_idx 0
2197  * avail.flags 1 avail.idx 128 used.flags 1 used.idx 0
2198  * kickfd 62 callfd 64 errfd -1
2199  *
2200  * descriptor table:
2201  * id addr len flags next user_addr
2202  * ===== ================== ===== ====== ===== ==================
2203  * 0 0x0000000010b6e974 2060 0x0002 1 0x00002aabbc76e974
2204  * 1 0x0000000010b6e034 2060 0x0002 2 0x00002aabbc76e034
2205  * 2 0x0000000010b6d6f4 2060 0x0002 3 0x00002aabbc76d6f4
2206  * 3 0x0000000010b6cdb4 2060 0x0002 4 0x00002aabbc76cdb4
2207  * 4 0x0000000010b6c474 2060 0x0002 5 0x00002aabbc76c474
2208  * 5 0x0000000010b6bb34 2060 0x0002 6 0x00002aabbc76bb34
2209  * 6 0x0000000010b6b1f4 2060 0x0002 7 0x00002aabbc76b1f4
2210  * 7 0x0000000010b6a8b4 2060 0x0002 8 0x00002aabbc76a8b4
2211  * 8 0x0000000010b69f74 2060 0x0002 9 0x00002aabbc769f74
2212  * 9 0x0000000010b69634 2060 0x0002 10 0x00002aabbc769634
2213  * 10 0x0000000010b68cf4 2060 0x0002 11 0x00002aabbc768cf4
2214  * :
2215  * 249 0x0000000000000000 0 0x0000 250 0x00002aab2b400000
2216  * 250 0x0000000000000000 0 0x0000 251 0x00002aab2b400000
2217  * 251 0x0000000000000000 0 0x0000 252 0x00002aab2b400000
2218  * 252 0x0000000000000000 0 0x0000 253 0x00002aab2b400000
2219  * 253 0x0000000000000000 0 0x0000 254 0x00002aab2b400000
2220  * 254 0x0000000000000000 0 0x0000 255 0x00002aab2b400000
2221  * 255 0x0000000000000000 0 0x0000 32768 0x00002aab2b400000
2222  *
2223  * Virtqueue 1 (RX)
2224  * qsz 256 last_avail_idx 0 last_used_idx 0
2225  * :
2226  * @cliexend
2227  * @endparblock
2228 ?*/
2229 /* *INDENT-OFF* */
2230 VLIB_CLI_COMMAND (show_vhost_user_command, static) = {
2231  .path = "show vhost-user",
2232  .short_help = "show vhost-user [<interface> [<interface> [..]]] [descriptors]",
2233  .function = show_vhost_user_command_fn,
2234 };
2235 /* *INDENT-ON* */
2236 
2237 
2238 static clib_error_t *
2240 {
2242 
2243  while (unformat_check_input (input) != UNFORMAT_END_OF_INPUT)
2244  {
2245  if (unformat (input, "coalesce-frames %d", &vum->coalesce_frames))
2246  ;
2247  else if (unformat (input, "coalesce-time %f", &vum->coalesce_time))
2248  ;
2249  else if (unformat (input, "dont-dump-memory"))
2250  vum->dont_dump_vhost_user_memory = 1;
2251  else
2252  return clib_error_return (0, "unknown input `%U'",
2253  format_unformat_error, input);
2254  }
2255 
2256  return 0;
2257 }
2258 
2259 /* vhost-user { ... } configuration. */
2260 VLIB_CONFIG_FUNCTION (vhost_user_config, "vhost-user");
2261 
2262 void
2264 {
2266  vhost_user_intf_t *vui;
2267 
2268  if (vum->dont_dump_vhost_user_memory)
2269  {
2271  unmap_all_mem_regions (vui);
2272  );
2273  }
2274 }
2275 
2276 /*
2277  * fd.io coding-style-patch-verification: ON
2278  *
2279  * Local Variables:
2280  * eval: (c-set-style "gnu")
2281  * End:
2282  */
unformat_function_t unformat_vnet_hw_interface
vlib_log_class_t vlib_log_register_class(char *class, char *subclass)
Definition: log.c:176
#define vec_validate(V, I)
Make sure vector is long enough for given index (no header, unspecified alignment) ...
Definition: vec.h:439
u32 len
Definition: pci.h:212
vmrglw vmrglh hi
#define vec_foreach_index(var, v)
Iterate over vector indices.
static void clib_file_del(clib_file_main_t *um, clib_file_t *f)
Definition: file.h:109
format_function_t format_vnet_hw_if_index_name
#define vu_log_warn(dev, f,...)
Definition: vhost_user.h:55
vring_desc_t * desc
Definition: vhost_user.h:263
static uword random_default_seed(void)
Default random seed (unix/linux user-mode)
Definition: random.h:91
#define VHOST_USER_PROTOCOL_F_LOG_SHMFD
Definition: vhost_user.h:32
static f64 vlib_process_wait_for_event_or_clock(vlib_main_t *vm, f64 dt)
Suspend a cooperative multi-tasking thread Waits for an event, or for the indicated number of seconds...
Definition: node_funcs.h:673
VNET_HW_INTERFACE_CLASS(vhost_interface_class, static)
vhost_cpu_t * cpus
Per-CPU data for vhost-user.
Definition: vhost_user.h:378
void ethernet_delete_interface(vnet_main_t *vnm, u32 hw_if_index)
Definition: interface.c:324
vnet_main_t * vnet_get_main(void)
Definition: misc.c:46
u64 region_guest_addr_hi[VHOST_MEMORY_MAX_NREGIONS]
Definition: vhost_user.h:314
static clib_error_t * vhost_user_socksvr_accept_ready(clib_file_t *uf)
Definition: vhost_user.c:933
unsigned long u64
Definition: types.h:89
#define VHOST_VRING_MAX_N
Definition: vhost_user.h:22
#define clib_memcpy_fast(a, b, c)
Definition: string.h:81
#define NULL
Definition: clib.h:58
uword mhash_unset(mhash_t *h, void *key, uword *old_value)
Definition: mhash.c:346
static f64 vlib_time_now(vlib_main_t *vm)
Definition: main.h:258
static_always_inline void vhost_user_update_iface_state(vhost_user_intf_t *vui)
Definition: vhost_user.c:197
u16 next
Definition: pci.h:214
#define VLIB_BUFFER_PRE_DATA_SIZE
Definition: buffer.h:51
vring_avail_t * avail
Definition: vhost_user.h:264
static vnet_hw_interface_t * vnet_get_hw_interface(vnet_main_t *vnm, u32 hw_if_index)
u32 file_descriptor
Definition: file.h:54
vnet_device_class_t vhost_user_device_class
#define vec_add1(V, E)
Add 1 element to end of vector (unspecified alignment).
Definition: vec.h:522
int vnet_interface_name_renumber(u32 sw_if_index, u32 new_show_dev_instance)
Definition: interface.c:1374
static clib_error_t * vhost_user_callfd_read_ready(clib_file_t *uf)
Definition: vhost_user.c:230
#define vec_add2(V, P, N)
Add N elements to end of vector V, return pointer to new elements in P.
Definition: vec.h:560
int i
#define VHOST_USER_EVENT_START_TIMER
Definition: vhost_user.h:289
uword unformat_user(unformat_input_t *input, unformat_function_t *func,...)
Definition: unformat.c:989
clib_memset(h->entries, 0, sizeof(h->entries[0])*entries)
static_always_inline void vhost_user_vring_init(vhost_user_intf_t *vui, u32 qid)
Definition: vhost_user.c:280
u8 * format(u8 *s, const char *fmt,...)
Definition: format.c:424
unformat_function_t unformat_vnet_sw_interface
u16 idx
Definition: pci.h:220
u16 flags
Definition: pci.h:233
int vhost_user_create_if(vnet_main_t *vnm, vlib_main_t *vm, const char *sock_filename, u8 is_server, u32 *sw_if_index, u64 feature_mask, u8 renumber, u32 custom_dev_instance, u8 *hwaddr, u8 enable_gso)
Definition: vhost_user.c:1494
#define pool_get(P, E)
Allocate an object E from a pool P (unspecified alignment).
Definition: pool.h:236
vring_used_t * used
Definition: vhost_user.h:265
vhost_vring_addr_t addr
Definition: vhost_user.h:147
format_function_t format_vnet_sw_if_index_name
#define VHOST_USER_VRING_NOFD_MASK
Definition: vhost_user.h:26
unsigned char u8
Definition: types.h:56
static_always_inline void vhost_user_rx_thread_placement(vhost_user_intf_t *vui, u32 qid)
Unassign existing interface/queue to thread mappings and re-assign new interface/queue to thread mapp...
Definition: vhost_user.c:161
static uword vlib_process_suspend_time_is_zero(f64 dt)
Returns TRUE if a process suspend time is less than 10us.
Definition: node_funcs.h:410
#define vec_reset_length(v)
Reset vector length to zero NULL-pointer tolerant.
clib_file_function_t * read_function
Definition: file.h:67
double f64
Definition: types.h:142
#define vlib_worker_thread_barrier_sync(X)
Definition: threads.h:204
static vnet_sw_interface_t * vnet_get_hw_sw_interface(vnet_main_t *vnm, u32 hw_if_index)
#define clib_memcpy(d, s, n)
Definition: string.h:180
static void vhost_user_vui_init(vnet_main_t *vnm, vhost_user_intf_t *vui, int server_sock_fd, const char *sock_filename, u64 feature_mask, u32 *sw_if_index, u8 enable_gso)
Definition: vhost_user.c:1417
#define vu_log_debug(dev, f,...)
Definition: vhost_user.h:48
clib_file_t * file_pool
Definition: file.h:88
u64 addr
Definition: pci.h:211
vnet_hw_interface_rx_mode
Definition: interface.h:53
#define static_always_inline
Definition: clib.h:99
#define pool_foreach(VAR, POOL, BODY)
Iterate through pool.
Definition: pool.h:493
vl_api_interface_index_t sw_if_index
Definition: gre.api:50
#define VLIB_INIT_FUNCTION(x)
Definition: init.h:173
static uword vlib_process_get_events(vlib_main_t *vm, uword **data_vector)
Return the first event type which has occurred and a vector of per-event data of that type...
Definition: node_funcs.h:516
static clib_error_t * vhost_user_socket_read(clib_file_t *uf)
Definition: vhost_user.c:358
#define UNIX_GET_FD(unixfd_idx)
Definition: vhost_user.h:68
void * log_base_addr
Definition: vhost_user.h:324
static_always_inline void vhost_user_thread_placement(vhost_user_intf_t *vui, u32 qid)
Definition: vhost_user.c:241
static_always_inline void * map_guest_mem(vhost_user_intf_t *vui, uword addr, u32 *hint)
static_always_inline void vnet_device_input_set_interrupt_pending(vnet_main_t *vnm, u32 hw_if_index, u16 queue_id)
Definition: devices.h:136
vnet_hw_interface_flags_t flags
Definition: interface.h:505
#define vec_elt_at_index(v, i)
Get vector value at index i checking that i is in bounds.
#define VRING_USED_F_NO_NOTIFY
Definition: vhost_user.h:45
#define clib_error_return(e, args...)
Definition: error.h:99
clib_file_main_t file_main
Definition: main.c:63
unsigned int u32
Definition: types.h:88
static vlib_node_registration_t vhost_user_process_node
(constructor) VLIB_REGISTER_NODE (vhost_user_process_node)
Definition: vhost_user.c:1200
static clib_error_t * vhost_user_kickfd_read_ready(clib_file_t *uf)
Definition: vhost_user.c:253
static clib_error_t * vhost_user_socket_error(clib_file_t *uf)
Definition: vhost_user.c:918
void vhost_user_unmap_all(void)
Definition: vhost_user.c:2263
unformat_function_t unformat_line_input
Definition: format.h:283
static_always_inline void vhost_user_tx_thread_placement(vhost_user_intf_t *vui)
Definition: vhost_user.c:120
#define VHOST_USER_PROTOCOL_F_MQ
Definition: vhost_user.h:31
#define pool_elt_at_index(p, i)
Returns pointer to element at given index.
Definition: pool.h:514
static_always_inline uword vnet_get_device_input_thread_index(vnet_main_t *vnm, u32 hw_if_index, u16 queue_id)
Definition: devices.h:127
static void vhost_user_term_if(vhost_user_intf_t *vui)
Disables and reset interface structure.
Definition: vhost_user.c:1212
static void vlib_process_signal_event(vlib_main_t *vm, uword node_index, uword type_opaque, uword data)
Definition: node_funcs.h:934
static_always_inline int vhost_user_intf_ready(vhost_user_intf_t *vui)
Returns whether at least one TX and one RX vring are enabled.
Definition: vhost_user.c:185
u32 random
Pseudo random iterator.
Definition: vhost_user.h:381
uword mhash_set_mem(mhash_t *h, void *key, uword *new_value, uword *old_value)
Definition: mhash.c:264
vlib_node_registration_t vhost_user_input_node
(constructor) VLIB_REGISTER_NODE (vhost_user_input_node)
struct _unformat_input_t unformat_input_t
unsigned short u16
Definition: types.h:57
#define clib_error_return_unix(e, args...)
Definition: error.h:102
int vnet_hw_interface_get_rx_mode(vnet_main_t *vnm, u32 hw_if_index, u16 queue_id, vnet_hw_interface_rx_mode *mode)
Definition: devices.c:313
#define pool_put(P, E)
Free an object E in pool P.
Definition: pool.h:286
int vhost_user_delete_if(vnet_main_t *vnm, vlib_main_t *vm, u32 sw_if_index)
Definition: vhost_user.c:1260
format_function_t format_vnet_hw_interface_rx_mode
#define VLIB_CONFIG_FUNCTION(x, n,...)
Definition: init.h:182
char sock_filename[256]
Definition: vhost_user.h:299
vhost_user_main_t vhost_user_main
Definition: vhost_user.c:56
#define VHOST_USER_MSG_HDR_SZ
Definition: vhost_user.h:20
u32 region_mmap_fd[VHOST_MEMORY_MAX_NREGIONS]
Definition: vhost_user.h:315
int vhost_user_dump_ifs(vnet_main_t *vnm, vlib_main_t *vm, vhost_user_intf_details_t **out_vuids)
Definition: vhost_user.c:1730
vhost_user_memory_region_t regions[VHOST_MEMORY_MAX_NREGIONS]
Definition: vhost_user.h:311
vlib_log_class_t log_default
Definition: vhost_user.h:387
static_always_inline void vhost_user_update_gso_interface_count(vhost_user_intf_t *vui, u8 add)
#define VHOST_VRING_IDX_RX(qid)
Definition: vhost_user.h:23
u32 * show_dev_instance_by_real_dev_instance
Definition: vhost_user.h:372
static_always_inline void vhost_user_vring_close(vhost_user_intf_t *vui, u32 qid)
Definition: vhost_user.c:303
vhost_user_intf_t * vhost_user_interfaces
Definition: vhost_user.h:371
#define VLIB_REGISTER_NODE(x,...)
Definition: node.h:169
static void mhash_init_c_string(mhash_t *h, uword n_value_bytes)
Definition: mhash.h:78
#define UNFORMAT_END_OF_INPUT
Definition: format.h:145
vlib_main_t * vm
Definition: buffer.c:312
#define vec_free(V)
Free vector&#39;s memory (no header).
Definition: vec.h:341
static void unmap_all_mem_regions(vhost_user_intf_t *vui)
Definition: vhost_user.c:74
static clib_error_t * vhost_user_config(vlib_main_t *vm, unformat_input_t *input)
Definition: vhost_user.c:2239
#define VLIB_MAIN_LOOP_EXIT_FUNCTION(x)
Definition: init.h:178
#define VHOST_USER_EVENT_STOP_TIMER
Definition: vhost_user.h:290
static vnet_hw_interface_t * vnet_get_sup_hw_interface_api_visible_or_null(vnet_main_t *vnm, u32 sw_if_index)
#define clib_warning(format, args...)
Definition: error.h:59
static uword vhost_user_process(vlib_main_t *vm, vlib_node_runtime_t *rt, vlib_frame_t *f)
Definition: vhost_user.c:1101
#define ARRAY_LEN(x)
Definition: clib.h:62
clib_error_t * vhost_user_delete_command_fn(vlib_main_t *vm, unformat_input_t *input, vlib_cli_command_t *cmd)
Definition: vhost_user.c:1683
vl_api_vxlan_gbp_api_tunnel_mode_t mode
Definition: vxlan_gbp.api:44
#define foreach_protocol_feature
u16 flags
Definition: pci.h:219
static int vhost_user_init_server_sock(const char *sock_filename, int *sock_fd)
Open server unix socket on specified sock_filename.
Definition: vhost_user.c:1342
#define VLIB_CLI_COMMAND(x,...)
Definition: cli.h:155
uword unformat_ethernet_address(unformat_input_t *input, va_list *args)
Definition: format.c:233
clib_error_t * show_vhost_user_command_fn(vlib_main_t *vm, unformat_input_t *input, vlib_cli_command_t *cmd)
Definition: vhost_user.c:1781
#define ASSERT(truth)
#define VHOST_VRING_F_LOG
Definition: vhost_user.h:33
void vnet_hw_interface_assign_rx_thread(vnet_main_t *vnm, u32 hw_if_index, u16 queue_id, uword thread_index)
Definition: devices.c:139
static uword * mhash_get(mhash_t *h, const void *key)
Definition: mhash.h:110
#define vu_log_err(dev, f,...)
Definition: vhost_user.h:61
static uword clib_file_add(clib_file_main_t *um, clib_file_t *template)
Definition: file.h:96
static_always_inline void vhost_user_if_disconnect(vhost_user_intf_t *vui)
Definition: vhost_user.c:335
static void clib_mem_free(void *p)
Definition: mem.h:226
volatile u32 * vring_locks[VHOST_VRING_MAX_N]
Definition: vhost_user.h:319
mhash_t if_index_by_sock_name
Definition: vhost_user.h:369
static clib_error_t * vhost_user_exit(vlib_main_t *vm)
Definition: vhost_user.c:1320
#define clib_error_report(e)
Definition: error.h:113
vlib_node_registration_t vhost_user_send_interrupt_node
(constructor) VLIB_REGISTER_NODE (vhost_user_send_interrupt_node)
Definition: vhost_user.c:53
static uword pointer_to_uword(const void *p)
Definition: types.h:131
static void vhost_user_create_ethernet(vnet_main_t *vnm, vlib_main_t *vm, vhost_user_intf_t *vui, u8 *hwaddress)
Create ethernet interface for vhost user interface.
Definition: vhost_user.c:1382
static vlib_main_t * vlib_get_main(void)
Definition: global_funcs.h:23
struct _vlib_node_registration vlib_node_registration_t
void * region_mmap_addr[VHOST_MEMORY_MAX_NREGIONS]
Definition: vhost_user.h:312
int vhost_user_modify_if(vnet_main_t *vnm, vlib_main_t *vm, const char *sock_filename, u8 is_server, u32 sw_if_index, u64 feature_mask, u8 renumber, u32 custom_dev_instance, u8 enable_gso)
Definition: vhost_user.c:1558
static long get_huge_page_size(int fd)
Definition: vhost_user.c:66
#define vec_len(v)
Number of elements in vector (rvalue-only, NULL tolerant)
clib_error_t * ethernet_register_interface(vnet_main_t *vnm, u32 dev_class_index, u32 dev_instance, const u8 *address, u32 *hw_if_index_return, ethernet_flag_change_function_t flag_change)
Definition: interface.c:278
vhost_user_vring_t vrings[VHOST_VRING_MAX_N]
Definition: vhost_user.h:318
clib_error_t * vnet_hw_interface_set_flags(vnet_main_t *vnm, u32 hw_if_index, vnet_hw_interface_flags_t flags)
Definition: interface.c:501
static void vhost_user_set_interrupt_pending(vhost_user_intf_t *vui, u32 ifq)
Definition: vhost_user.c:214
u64 uword
Definition: types.h:112
static void unformat_free(unformat_input_t *i)
Definition: format.h:163
#define VHOST_MEMORY_MAX_NREGIONS
Definition: vhost_user.h:19
clib_error_t * vhost_user_connect_command_fn(vlib_main_t *vm, unformat_input_t *input, vlib_cli_command_t *cmd)
Definition: vhost_user.c:1612
void vnet_sw_interface_set_mtu(vnet_main_t *vnm, u32 sw_if_index, u32 mtu)
Definition: interface.c:663
u32 rx_buffers_len
Definition: vhost_user.h:356
int vnet_hw_interface_unassign_rx_thread(vnet_main_t *vnm, u32 hw_if_index, u16 queue_id)
Definition: devices.c:188
u16 idx
Definition: pci.h:234
static void * clib_mem_alloc_aligned(uword size, uword align)
Definition: mem.h:161
static u32 random_u32(u32 *seed)
32-bit random number generator
Definition: random.h:69
u8 * format_unformat_error(u8 *s, va_list *va)
Definition: unformat.c:91
void vlib_worker_thread_barrier_release(vlib_main_t *vm)
Definition: threads.c:1487
static vlib_thread_main_t * vlib_get_thread_main()
Definition: global_funcs.h:32
u64 region_guest_addr_lo[VHOST_MEMORY_MAX_NREGIONS]
Definition: vhost_user.h:313
u16 flags
Definition: pci.h:213
#define vec_foreach(var, vec)
Vector iterator.
uword private_data
Definition: file.h:64
Definition: file.h:51
int vnet_hw_interface_set_rx_mode(vnet_main_t *vnm, u32 hw_if_index, u16 queue_id, vnet_hw_interface_rx_mode mode)
Definition: devices.c:253
static clib_error_t * vhost_user_init(vlib_main_t *vm)
Definition: vhost_user.c:969
static_always_inline void * map_user_mem(vhost_user_intf_t *vui, uword addr)
#define CLIB_CACHE_LINE_BYTES
Definition: cache.h:59
static uword vhost_user_send_interrupt_process(vlib_main_t *vm, vlib_node_runtime_t *rt, vlib_frame_t *f)
Definition: vhost_user.c:1004
#define FEATURE_VIRTIO_NET_F_HOST_GUEST_TSO_FEATURE_BITS
Definition: vhost_user.h:125
void vlib_cli_output(vlib_main_t *vm, char *fmt,...)
Definition: cli.c:768
vl_api_address_union_t un
Definition: ip_types.api:85
#define VLIB_INITS(...)
Definition: init.h:344
static void vnet_hw_interface_set_input_node(vnet_main_t *vnm, u32 hw_if_index, u32 node_index)
Definition: devices.h:79
uword unformat(unformat_input_t *i, const char *fmt,...)
Definition: unformat.c:978
static uword unformat_check_input(unformat_input_t *i)
Definition: format.h:171
int dont_dump_vhost_user_memory
Definition: vhost_user.h:375