FD.io VPP  v19.04-6-g6f05f72
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  msg.size = sizeof (msg.u64);
467  vu_log_debug (vui, "if %d msg VHOST_USER_GET_FEATURES - reply "
468  "0x%016llx", vui->hw_if_index, msg.u64);
469  n =
470  send (uf->file_descriptor, &msg, VHOST_USER_MSG_HDR_SZ + msg.size, 0);
471  if (n != (msg.size + VHOST_USER_MSG_HDR_SZ))
472  {
473  vu_log_debug (vui, "could not send message response");
474  goto close_socket;
475  }
476  break;
477 
479  vu_log_debug (vui, "if %d msg VHOST_USER_SET_FEATURES features "
480  "0x%016llx", vui->hw_if_index, msg.u64);
481 
482  vui->features = msg.u64;
483 
484  if (vui->features &
485  ((1 << FEAT_VIRTIO_NET_F_MRG_RXBUF) |
486  (1ULL << FEAT_VIRTIO_F_VERSION_1)))
487  vui->virtio_net_hdr_sz = 12;
488  else
489  vui->virtio_net_hdr_sz = 10;
490 
491  vui->is_any_layout =
492  (vui->features & (1 << FEAT_VIRTIO_F_ANY_LAYOUT)) ? 1 : 0;
493 
496  vui->is_ready = 0;
497  break;
498 
500  vu_log_debug (vui, "if %d msg VHOST_USER_SET_MEM_TABLE nregions %d",
501  vui->hw_if_index, msg.memory.nregions);
502 
503  if ((msg.memory.nregions < 1) ||
504  (msg.memory.nregions > VHOST_MEMORY_MAX_NREGIONS))
505  {
506  vu_log_debug (vui, "number of mem regions must be between 1 and %i",
507  VHOST_MEMORY_MAX_NREGIONS);
508  goto close_socket;
509  }
510 
511  if (msg.memory.nregions != number_of_fds)
512  {
513  vu_log_debug (vui, "each memory region must have FD");
514  goto close_socket;
515  }
516 
517  /* Do the mmap without barrier sync */
518  void *region_mmap_addr[VHOST_MEMORY_MAX_NREGIONS];
519  for (i = 0; i < msg.memory.nregions; i++)
520  {
521  long page_sz = get_huge_page_size (fds[i]);
522 
523  /* align size to page */
524  ssize_t map_sz = (msg.memory.regions[i].memory_size +
525  msg.memory.regions[i].mmap_offset +
526  page_sz - 1) & ~(page_sz - 1);
527 
528  region_mmap_addr[i] = mmap (0, map_sz, PROT_READ | PROT_WRITE,
529  MAP_SHARED, fds[i], 0);
530  if (region_mmap_addr[i] == MAP_FAILED)
531  {
532  vu_log_err (vui, "failed to map memory. errno is %d", errno);
533  for (j = 0; j < i; j++)
534  munmap (region_mmap_addr[j], map_sz);
535  goto close_socket;
536  }
537  vu_log_debug (vui, "map memory region %d addr 0 len 0x%lx fd %d "
538  "mapped 0x%lx page_sz 0x%x", i, map_sz, fds[i],
539  region_mmap_addr[i], page_sz);
540  }
541 
543  unmap_all_mem_regions (vui);
544  for (i = 0; i < msg.memory.nregions; i++)
545  {
546  clib_memcpy_fast (&(vui->regions[i]), &msg.memory.regions[i],
547  sizeof (vhost_user_memory_region_t));
548 
549  vui->region_mmap_addr[i] = region_mmap_addr[i];
550  vui->region_guest_addr_lo[i] = vui->regions[i].guest_phys_addr;
551  vui->region_guest_addr_hi[i] = vui->regions[i].guest_phys_addr +
552  vui->regions[i].memory_size;
553 
554  vui->region_mmap_addr[i] += vui->regions[i].mmap_offset;
555  vui->region_mmap_fd[i] = fds[i];
556 
557  vui->nregions++;
558  }
560  break;
561 
563  vu_log_debug (vui, "if %d msg VHOST_USER_SET_VRING_NUM idx %d num %d",
564  vui->hw_if_index, msg.state.index, msg.state.num);
565 
566  if ((msg.state.num > 32768) || /* maximum ring size is 32768 */
567  (msg.state.num == 0) || /* it cannot be zero */
568  ((msg.state.num - 1) & msg.state.num)) /* must be power of 2 */
569  goto close_socket;
570  vui->vrings[msg.state.index].qsz_mask = msg.state.num - 1;
571  break;
572 
574  vu_log_debug (vui, "if %d msg VHOST_USER_SET_VRING_ADDR idx %d",
575  vui->hw_if_index, msg.state.index);
576 
577  if (msg.state.index >= VHOST_VRING_MAX_N)
578  {
579  vu_log_debug (vui, "invalid vring index VHOST_USER_SET_VRING_ADDR:"
580  " %d >= %d", msg.state.index, VHOST_VRING_MAX_N);
581  goto close_socket;
582  }
583 
584  if (msg.size < sizeof (msg.addr))
585  {
586  vu_log_debug (vui, "vhost message is too short (%d < %d)",
587  msg.size, sizeof (msg.addr));
588  goto close_socket;
589  }
590 
591  vring_desc_t *desc = map_user_mem (vui, msg.addr.desc_user_addr);
592  vring_used_t *used = map_user_mem (vui, msg.addr.used_user_addr);
593  vring_avail_t *avail = map_user_mem (vui, msg.addr.avail_user_addr);
594 
595  if ((desc == NULL) || (used == NULL) || (avail == NULL))
596  {
597  vu_log_debug (vui, "failed to map user memory for hw_if_index %d",
598  vui->hw_if_index);
599  goto close_socket;
600  }
601 
603  vui->vrings[msg.state.index].desc = desc;
604  vui->vrings[msg.state.index].used = used;
605  vui->vrings[msg.state.index].avail = avail;
606 
607  vui->vrings[msg.state.index].log_guest_addr = msg.addr.log_guest_addr;
608  vui->vrings[msg.state.index].log_used =
609  (msg.addr.flags & (1 << VHOST_VRING_F_LOG)) ? 1 : 0;
610 
611  /* Spec says: If VHOST_USER_F_PROTOCOL_FEATURES has not been negotiated,
612  the ring is initialized in an enabled state. */
613  if (!(vui->features & (1 << FEAT_VHOST_USER_F_PROTOCOL_FEATURES)))
614  vui->vrings[msg.state.index].enabled = 1;
615 
616  vui->vrings[msg.state.index].last_used_idx =
617  vui->vrings[msg.state.index].last_avail_idx =
618  vui->vrings[msg.state.index].used->idx;
619 
620  /* tell driver that we don't want interrupts */
621  vui->vrings[msg.state.index].used->flags = VRING_USED_F_NO_NOTIFY;
624  break;
625 
627  vu_log_debug (vui, "if %d msg VHOST_USER_SET_OWNER", vui->hw_if_index);
628  break;
629 
631  vu_log_debug (vui, "if %d msg VHOST_USER_RESET_OWNER",
632  vui->hw_if_index);
633  break;
634 
636  vu_log_debug (vui, "if %d msg VHOST_USER_SET_VRING_CALL %d",
637  vui->hw_if_index, msg.u64);
638 
639  q = (u8) (msg.u64 & 0xFF);
640 
641  /* if there is old fd, delete and close it */
642  if (vui->vrings[q].callfd_idx != ~0)
643  {
645  vui->vrings[q].callfd_idx);
646  clib_file_del (&file_main, uf);
647  vui->vrings[q].callfd_idx = ~0;
648  }
649 
650  if (!(msg.u64 & VHOST_USER_VRING_NOFD_MASK))
651  {
652  if (number_of_fds != 1)
653  {
654  vu_log_debug (vui, "More than one fd received !");
655  goto close_socket;
656  }
657 
658  template.read_function = vhost_user_callfd_read_ready;
659  template.file_descriptor = fds[0];
660  template.private_data =
661  ((vui - vhost_user_main.vhost_user_interfaces) << 8) + q;
662  vui->vrings[q].callfd_idx = clib_file_add (&file_main, &template);
663  }
664  else
665  vui->vrings[q].callfd_idx = ~0;
666  break;
667 
669  vu_log_debug (vui, "if %d msg VHOST_USER_SET_VRING_KICK %d",
670  vui->hw_if_index, msg.u64);
671 
672  q = (u8) (msg.u64 & 0xFF);
673 
674  if (vui->vrings[q].kickfd_idx != ~0)
675  {
677  vui->vrings[q].kickfd_idx);
678  clib_file_del (&file_main, uf);
679  vui->vrings[q].kickfd_idx = ~0;
680  }
681 
682  if (!(msg.u64 & VHOST_USER_VRING_NOFD_MASK))
683  {
684  if (number_of_fds != 1)
685  {
686  vu_log_debug (vui, "More than one fd received !");
687  goto close_socket;
688  }
689 
690  template.read_function = vhost_user_kickfd_read_ready;
691  template.file_descriptor = fds[0];
692  template.private_data =
693  (((uword) (vui - vhost_user_main.vhost_user_interfaces)) << 8) +
694  q;
695  vui->vrings[q].kickfd_idx = clib_file_add (&file_main, &template);
696  }
697  else
698  {
699  //When no kickfd is set, the queue is initialized as started
700  vui->vrings[q].kickfd_idx = ~0;
701  vui->vrings[q].started = 1;
703  }
705  break;
706 
708  vu_log_debug (vui, "if %d msg VHOST_USER_SET_VRING_ERR %d",
709  vui->hw_if_index, msg.u64);
710 
711  q = (u8) (msg.u64 & 0xFF);
712 
713  if (vui->vrings[q].errfd != -1)
714  close (vui->vrings[q].errfd);
715 
716  if (!(msg.u64 & VHOST_USER_VRING_NOFD_MASK))
717  {
718  if (number_of_fds != 1)
719  goto close_socket;
720 
721  vui->vrings[q].errfd = fds[0];
722  }
723  else
724  vui->vrings[q].errfd = -1;
725  break;
726 
728  vu_log_debug (vui, "if %d msg VHOST_USER_SET_VRING_BASE idx %d num %d",
729  vui->hw_if_index, msg.state.index, msg.state.num);
731  vui->vrings[msg.state.index].last_avail_idx = msg.state.num;
733  break;
734 
736  if (msg.state.index >= VHOST_VRING_MAX_N)
737  {
738  vu_log_debug (vui, "invalid vring index VHOST_USER_GET_VRING_BASE:"
739  " %d >= %d", msg.state.index, VHOST_VRING_MAX_N);
740  goto close_socket;
741  }
742 
743  /* protection is needed to prevent rx/tx from changing last_avail_idx */
745  /*
746  * Copy last_avail_idx from the vring before closing it because
747  * closing the vring also initializes the vring last_avail_idx
748  */
749  msg.state.num = vui->vrings[msg.state.index].last_avail_idx;
750  msg.flags |= 4;
751  msg.size = sizeof (msg.state);
752 
753  /*
754  * Spec says: Client must [...] stop ring upon receiving
755  * VHOST_USER_GET_VRING_BASE
756  */
757  vhost_user_vring_close (vui, msg.state.index);
759  vu_log_debug (vui, "if %d msg VHOST_USER_GET_VRING_BASE idx %d num %d",
760  vui->hw_if_index, msg.state.index, msg.state.num);
761  n =
762  send (uf->file_descriptor, &msg, VHOST_USER_MSG_HDR_SZ + msg.size, 0);
763  if (n != (msg.size + VHOST_USER_MSG_HDR_SZ))
764  {
765  vu_log_debug (vui, "could not send message response");
766  goto close_socket;
767  }
769  break;
770 
771  case VHOST_USER_NONE:
772  vu_log_debug (vui, "if %d msg VHOST_USER_NONE", vui->hw_if_index);
773  break;
774 
776  vu_log_debug (vui, "if %d msg VHOST_USER_SET_LOG_BASE",
777  vui->hw_if_index);
778 
779  if (msg.size != sizeof (msg.log))
780  {
781  vu_log_debug (vui, "invalid msg size for VHOST_USER_SET_LOG_BASE:"
782  " %d instead of %d", msg.size, sizeof (msg.log));
783  goto close_socket;
784  }
785 
787  {
788  vu_log_debug (vui, "VHOST_USER_PROTOCOL_F_LOG_SHMFD not set but "
789  "VHOST_USER_SET_LOG_BASE received");
790  goto close_socket;
791  }
792 
793  fd = fds[0];
794  /* align size to page */
795  long page_sz = get_huge_page_size (fd);
796  ssize_t map_sz =
797  (msg.log.size + msg.log.offset + page_sz - 1) & ~(page_sz - 1);
798 
799  void *log_base_addr = mmap (0, map_sz, PROT_READ | PROT_WRITE,
800  MAP_SHARED, fd, 0);
801 
802  vu_log_debug (vui, "map log region addr 0 len 0x%lx off 0x%lx fd %d "
803  "mapped 0x%lx", map_sz, msg.log.offset, fd,
804  log_base_addr);
805 
806  if (log_base_addr == MAP_FAILED)
807  {
808  vu_log_err (vui, "failed to map memory. errno is %d", errno);
809  goto close_socket;
810  }
811 
813  vui->log_base_addr = log_base_addr;
814  vui->log_base_addr += msg.log.offset;
815  vui->log_size = msg.log.size;
817 
818  msg.flags |= 4;
819  msg.size = sizeof (msg.u64);
820  n =
821  send (uf->file_descriptor, &msg, VHOST_USER_MSG_HDR_SZ + msg.size, 0);
822  if (n != (msg.size + VHOST_USER_MSG_HDR_SZ))
823  {
824  vu_log_debug (vui, "could not send message response");
825  goto close_socket;
826  }
827  break;
828 
830  vu_log_debug (vui, "if %d msg VHOST_USER_SET_LOG_FD", vui->hw_if_index);
831  break;
832 
834  msg.flags |= 4;
835  msg.u64 = (1 << VHOST_USER_PROTOCOL_F_LOG_SHMFD) |
837  msg.size = sizeof (msg.u64);
838  vu_log_debug (vui, "if %d msg VHOST_USER_GET_PROTOCOL_FEATURES - "
839  "reply 0x%016llx", vui->hw_if_index, msg.u64);
840  n =
841  send (uf->file_descriptor, &msg, VHOST_USER_MSG_HDR_SZ + msg.size, 0);
842  if (n != (msg.size + VHOST_USER_MSG_HDR_SZ))
843  {
844  vu_log_debug (vui, "could not send message response");
845  goto close_socket;
846  }
847  break;
848 
850  vu_log_debug (vui, "if %d msg VHOST_USER_SET_PROTOCOL_FEATURES "
851  "features 0x%016llx", vui->hw_if_index, msg.u64);
852  vui->protocol_features = msg.u64;
853  break;
854 
856  msg.flags |= 4;
857  msg.u64 = VHOST_VRING_MAX_N;
858  msg.size = sizeof (msg.u64);
859  vu_log_debug (vui, "if %d msg VHOST_USER_GET_QUEUE_NUM - reply %d",
860  vui->hw_if_index, msg.u64);
861  n =
862  send (uf->file_descriptor, &msg, VHOST_USER_MSG_HDR_SZ + msg.size, 0);
863  if (n != (msg.size + VHOST_USER_MSG_HDR_SZ))
864  {
865  vu_log_debug (vui, "could not send message response");
866  goto close_socket;
867  }
868  break;
869 
871  vu_log_debug (vui, "if %d VHOST_USER_SET_VRING_ENABLE: %s queue %d",
872  vui->hw_if_index, msg.state.num ? "enable" : "disable",
873  msg.state.index);
874  if (msg.state.index >= VHOST_VRING_MAX_N)
875  {
876  vu_log_debug (vui, "invalid vring idx VHOST_USER_SET_VRING_ENABLE:"
877  " %d >= %d", msg.state.index, VHOST_VRING_MAX_N);
878  goto close_socket;
879  }
880 
881  vui->vrings[msg.state.index].enabled = msg.state.num;
882  vhost_user_thread_placement (vui, msg.state.index);
884  break;
885 
886  default:
887  vu_log_debug (vui, "unknown vhost-user message %d received. "
888  "closing socket", msg.request);
889  goto close_socket;
890  }
891 
892  return 0;
893 
894 close_socket:
899  return 0;
900 }
901 
902 static clib_error_t *
904 {
907  vhost_user_intf_t *vui =
909 
910  vu_log_debug (vui, "socket error on if %d", vui->sw_if_index);
914  return 0;
915 }
916 
917 static clib_error_t *
919 {
920  int client_fd, client_len;
921  struct sockaddr_un client;
922  clib_file_t template = { 0 };
924  vhost_user_intf_t *vui;
925 
927 
928  client_len = sizeof (client);
929  client_fd = accept (uf->file_descriptor,
930  (struct sockaddr *) &client,
931  (socklen_t *) & client_len);
932 
933  if (client_fd < 0)
934  return clib_error_return_unix (0, "accept");
935 
936  if (vui->clib_file_index != ~0)
937  {
938  vu_log_debug (vui, "Close client socket for vhost interface %d, fd %d",
941  }
942 
943  vu_log_debug (vui, "New client socket for vhost interface %d, fd %d",
944  vui->sw_if_index, client_fd);
945  template.read_function = vhost_user_socket_read;
946  template.error_function = vhost_user_socket_error;
947  template.file_descriptor = client_fd;
948  template.private_data = vui - vhost_user_main.vhost_user_interfaces;
949  vui->clib_file_index = clib_file_add (&file_main, &template);
950  return 0;
951 }
952 
953 static clib_error_t *
955 {
956  clib_error_t *error;
959 
960  error = vlib_call_init_function (vm, ip4_init);
961  if (error)
962  return error;
963 
964  vum->log_default = vlib_log_register_class ("vhost-user", 0);
965 
966  vum->coalesce_frames = 32;
967  vum->coalesce_time = 1e-3;
968 
969  vec_validate (vum->cpus, tm->n_vlib_mains - 1);
970 
971  vhost_cpu_t *cpu;
972  vec_foreach (cpu, vum->cpus)
973  {
974  /* This is actually not necessary as validate already zeroes it
975  * Just keeping the loop here for later because I am lazy. */
976  cpu->rx_buffers_len = 0;
977  }
978 
979  vum->random = random_default_seed ();
980 
982 
983  return 0;
984 }
985 
987 
988 static uword
991 {
992  vhost_user_intf_t *vui;
993  f64 timeout = 3153600000.0 /* 100 years */ ;
994  uword event_type, *event_data = 0;
996  u16 qid;
997  f64 now, poll_time_remaining;
998  f64 next_timeout;
999  u8 stop_timer = 0;
1000 
1001  while (1)
1002  {
1003  poll_time_remaining =
1005  event_type = vlib_process_get_events (vm, &event_data);
1006  vec_reset_length (event_data);
1007 
1008  /*
1009  * Use the remaining timeout if it is less than coalesce time to avoid
1010  * resetting the existing timer in the middle of expiration
1011  */
1012  timeout = poll_time_remaining;
1013  if (vlib_process_suspend_time_is_zero (timeout) ||
1014  (timeout > vum->coalesce_time))
1015  timeout = vum->coalesce_time;
1016 
1017  now = vlib_time_now (vm);
1018  switch (event_type)
1019  {
1021  stop_timer = 1;
1022  break;
1023 
1025  stop_timer = 0;
1026  if (!vlib_process_suspend_time_is_zero (poll_time_remaining))
1027  break;
1028  /* fall through */
1029 
1030  case ~0:
1031  /* *INDENT-OFF* */
1032  pool_foreach (vui, vum->vhost_user_interfaces, {
1033  next_timeout = timeout;
1034  for (qid = 0; qid < VHOST_VRING_MAX_N / 2; qid += 2)
1035  {
1036  vhost_user_vring_t *rxvq = &vui->vrings[qid];
1037  vhost_user_vring_t *txvq = &vui->vrings[qid + 1];
1038 
1039  if (txvq->qid == -1)
1040  continue;
1041  if (txvq->n_since_last_int)
1042  {
1043  if (now >= txvq->int_deadline)
1044  vhost_user_send_call (vm, txvq);
1045  else
1046  next_timeout = txvq->int_deadline - now;
1047  }
1048 
1049  if (rxvq->n_since_last_int)
1050  {
1051  if (now >= rxvq->int_deadline)
1052  vhost_user_send_call (vm, rxvq);
1053  else
1054  next_timeout = rxvq->int_deadline - now;
1055  }
1056 
1057  if ((next_timeout < timeout) && (next_timeout > 0.0))
1058  timeout = next_timeout;
1059  }
1060  });
1061  /* *INDENT-ON* */
1062  break;
1063 
1064  default:
1065  clib_warning ("BUG: unhandled event type %d", event_type);
1066  break;
1067  }
1068  /* No less than 1 millisecond */
1069  if (timeout < 1e-3)
1070  timeout = 1e-3;
1071  if (stop_timer)
1072  timeout = 3153600000.0;
1073  }
1074  return 0;
1075 }
1076 
1077 /* *INDENT-OFF* */
1080  .type = VLIB_NODE_TYPE_PROCESS,
1081  .name = "vhost-user-send-interrupt-process",
1082 };
1083 /* *INDENT-ON* */
1084 
1085 static uword
1088 {
1090  vhost_user_intf_t *vui;
1091  struct sockaddr_un sun;
1092  int sockfd;
1093  clib_file_t template = { 0 };
1094  f64 timeout = 3153600000.0 /* 100 years */ ;
1095  uword *event_data = 0;
1096 
1097  sockfd = -1;
1098  sun.sun_family = AF_UNIX;
1100  template.error_function = vhost_user_socket_error;
1101 
1102  while (1)
1103  {
1105  vlib_process_get_events (vm, &event_data);
1106  vec_reset_length (event_data);
1107 
1108  timeout = 3.0;
1109 
1110  /* *INDENT-OFF* */
1111  pool_foreach (vui, vum->vhost_user_interfaces, {
1112 
1113  if (vui->unix_server_index == ~0) { //Nothing to do for server sockets
1114  if (vui->clib_file_index == ~0)
1115  {
1116  if ((sockfd < 0) &&
1117  ((sockfd = socket (AF_UNIX, SOCK_STREAM, 0)) < 0))
1118  {
1119  /*
1120  * 1st time error or new error for this interface,
1121  * spit out the message and record the error
1122  */
1123  if (!vui->sock_errno || (vui->sock_errno != errno))
1124  {
1125  clib_unix_warning
1126  ("Error: Could not open unix socket for %s",
1127  vui->sock_filename);
1128  vui->sock_errno = errno;
1129  }
1130  continue;
1131  }
1132 
1133  /* try to connect */
1134  strncpy (sun.sun_path, (char *) vui->sock_filename,
1135  sizeof (sun.sun_path) - 1);
1136 
1137  /* Avoid hanging VPP if the other end does not accept */
1138  if (fcntl(sockfd, F_SETFL, O_NONBLOCK) < 0)
1139  clib_unix_warning ("fcntl");
1140 
1141  if (connect (sockfd, (struct sockaddr *) &sun,
1142  sizeof (struct sockaddr_un)) == 0)
1143  {
1144  /* Set the socket to blocking as it was before */
1145  if (fcntl(sockfd, F_SETFL, 0) < 0)
1146  clib_unix_warning ("fcntl2");
1147 
1148  vui->sock_errno = 0;
1149  template.file_descriptor = sockfd;
1150  template.private_data =
1151  vui - vhost_user_main.vhost_user_interfaces;
1152  vui->clib_file_index = clib_file_add (&file_main, &template);
1153 
1154  /* This sockfd is considered consumed */
1155  sockfd = -1;
1156  }
1157  else
1158  {
1159  vui->sock_errno = errno;
1160  }
1161  }
1162  else
1163  {
1164  /* check if socket is alive */
1165  int error = 0;
1166  socklen_t len = sizeof (error);
1167  int fd = UNIX_GET_FD(vui->clib_file_index);
1168  int retval =
1169  getsockopt (fd, SOL_SOCKET, SO_ERROR, &error, &len);
1170 
1171  if (retval)
1172  {
1173  vu_log_debug (vui, "getsockopt returned %d", retval);
1174  vhost_user_if_disconnect (vui);
1175  }
1176  }
1177  }
1178  });
1179  /* *INDENT-ON* */
1180  }
1181  return 0;
1182 }
1183 
1184 /* *INDENT-OFF* */
1186  .function = vhost_user_process,
1187  .type = VLIB_NODE_TYPE_PROCESS,
1188  .name = "vhost-user-process",
1189 };
1190 /* *INDENT-ON* */
1191 
1192 /**
1193  * Disables and reset interface structure.
1194  * It can then be either init again, or removed from used interfaces.
1195  */
1196 static void
1198 {
1199  int q;
1201 
1202  // disconnect interface sockets
1205 
1206  for (q = 0; q < VHOST_VRING_MAX_N; q++)
1207  {
1208  // Remove existing queue mapping for the interface
1209  if (q & 1)
1210  {
1211  int rv;
1212  vnet_main_t *vnm = vnet_get_main ();
1213  vhost_user_vring_t *txvq = &vui->vrings[q];
1214 
1215  if (txvq->qid != -1)
1216  {
1218  vui->hw_if_index,
1219  q >> 1);
1220  if (rv)
1221  vu_log_warn (vui, "unable to unassign interface %d, "
1222  "queue %d: rc=%d", vui->hw_if_index, q >> 1, rv);
1223  }
1224  }
1225 
1226  clib_mem_free ((void *) vui->vring_locks[q]);
1227  }
1228 
1229  if (vui->unix_server_index != ~0)
1230  {
1231  //Close server socket
1233  vui->unix_server_index);
1234  clib_file_del (&file_main, uf);
1235  vui->unix_server_index = ~0;
1236  unlink (vui->sock_filename);
1237  }
1238 
1240  &vui->if_index);
1241 }
1242 
1243 int
1245 {
1247  vhost_user_intf_t *vui;
1248  int rv = 0;
1249  vnet_hw_interface_t *hwif;
1250  u16 qid;
1251 
1252  if (!(hwif = vnet_get_sup_hw_interface (vnm, sw_if_index)) ||
1254  return VNET_API_ERROR_INVALID_SW_IF_INDEX;
1255 
1257 
1258  vu_log_debug (vui, "Deleting vhost-user interface %s (instance %d)",
1259  hwif->name, hwif->dev_instance);
1260 
1261  for (qid = 1; qid < VHOST_VRING_MAX_N / 2; qid += 2)
1262  {
1263  vhost_user_vring_t *txvq = &vui->vrings[qid];
1264 
1265  if (txvq->qid == -1)
1266  continue;
1267  if ((vum->ifq_count > 0) &&
1270  {
1271  vum->ifq_count--;
1272  // Stop the timer if there is no more interrupt interface/queue
1273  if ((vum->ifq_count == 0) &&
1274  (vum->coalesce_time > 0.0) && (vum->coalesce_frames > 0))
1275  {
1279  break;
1280  }
1281  }
1282  }
1283 
1284  // Disable and reset interface
1285  vhost_user_term_if (vui);
1286 
1287  // Reset renumbered iface
1288  if (hwif->dev_instance <
1291 
1292  // Delete ethernet interface
1294 
1295  // Back to pool
1296  pool_put (vum->vhost_user_interfaces, vui);
1297 
1298  return rv;
1299 }
1300 
1301 static clib_error_t *
1303 {
1304  vnet_main_t *vnm = vnet_get_main ();
1306  vhost_user_intf_t *vui;
1307 
1309  /* *INDENT-OFF* */
1310  pool_foreach (vui, vum->vhost_user_interfaces, {
1311  vhost_user_delete_if (vnm, vm, vui->sw_if_index);
1312  });
1313  /* *INDENT-ON* */
1315  return 0;
1316 }
1317 
1319 
1320 /**
1321  * Open server unix socket on specified sock_filename.
1322  */
1323 static int
1324 vhost_user_init_server_sock (const char *sock_filename, int *sock_fd)
1325 {
1326  int rv = 0;
1327  struct sockaddr_un un = { };
1328  int fd;
1329  /* create listening socket */
1330  if ((fd = socket (AF_UNIX, SOCK_STREAM, 0)) < 0)
1331  return VNET_API_ERROR_SYSCALL_ERROR_1;
1332 
1333  un.sun_family = AF_UNIX;
1334  strncpy ((char *) un.sun_path, (char *) sock_filename,
1335  sizeof (un.sun_path) - 1);
1336 
1337  /* remove if exists */
1338  unlink ((char *) sock_filename);
1339 
1340  if (bind (fd, (struct sockaddr *) &un, sizeof (un)) == -1)
1341  {
1342  rv = VNET_API_ERROR_SYSCALL_ERROR_2;
1343  goto error;
1344  }
1345 
1346  if (listen (fd, 1) == -1)
1347  {
1348  rv = VNET_API_ERROR_SYSCALL_ERROR_3;
1349  goto error;
1350  }
1351 
1352  *sock_fd = fd;
1353  return 0;
1354 
1355 error:
1356  close (fd);
1357  return rv;
1358 }
1359 
1360 /**
1361  * Create ethernet interface for vhost user interface.
1362  */
1363 static void
1365  vhost_user_intf_t * vui, u8 * hwaddress)
1366 {
1368  u8 hwaddr[6];
1369  clib_error_t *error;
1370 
1371  /* create hw and sw interface */
1372  if (hwaddress)
1373  {
1374  clib_memcpy (hwaddr, hwaddress, 6);
1375  }
1376  else
1377  {
1378  random_u32 (&vum->random);
1379  clib_memcpy (hwaddr + 2, &vum->random, sizeof (vum->random));
1380  hwaddr[0] = 2;
1381  hwaddr[1] = 0xfe;
1382  }
1383 
1385  (vnm,
1387  vui - vum->vhost_user_interfaces /* device instance */ ,
1388  hwaddr /* ethernet address */ ,
1389  &vui->hw_if_index, 0 /* flag change */ );
1390 
1391  if (error)
1392  clib_error_report (error);
1393 }
1394 
1395 /*
1396  * Initialize vui with specified attributes
1397  */
1398 static void
1400  vhost_user_intf_t * vui,
1401  int server_sock_fd,
1402  const char *sock_filename,
1403  u64 feature_mask, u32 * sw_if_index)
1404 {
1405  vnet_sw_interface_t *sw;
1406  int q;
1408  vnet_hw_interface_t *hw;
1409 
1410  hw = vnet_get_hw_interface (vnm, vui->hw_if_index);
1411  sw = vnet_get_hw_sw_interface (vnm, vui->hw_if_index);
1412  if (server_sock_fd != -1)
1413  {
1414  clib_file_t template = { 0 };
1416  template.file_descriptor = server_sock_fd;
1417  template.private_data = vui - vum->vhost_user_interfaces; //hw index
1418  vui->unix_server_index = clib_file_add (&file_main, &template);
1419  }
1420  else
1421  {
1422  vui->unix_server_index = ~0;
1423  }
1424 
1425  vui->sw_if_index = sw->sw_if_index;
1426  strncpy (vui->sock_filename, sock_filename,
1427  ARRAY_LEN (vui->sock_filename) - 1);
1428  vui->sock_errno = 0;
1429  vui->is_ready = 0;
1430  vui->feature_mask = feature_mask;
1431  vui->clib_file_index = ~0;
1432  vui->log_base_addr = 0;
1433  vui->if_index = vui - vum->vhost_user_interfaces;
1435  &vui->if_index, 0);
1436 
1437  for (q = 0; q < VHOST_VRING_MAX_N; q++)
1438  vhost_user_vring_init (vui, q);
1439 
1442 
1443  if (sw_if_index)
1444  *sw_if_index = vui->sw_if_index;
1445 
1446  for (q = 0; q < VHOST_VRING_MAX_N; q++)
1447  {
1450  clib_memset ((void *) vui->vring_locks[q], 0, CLIB_CACHE_LINE_BYTES);
1451  }
1452 
1454  vlib_get_thread_main ()->n_vlib_mains - 1);
1456 }
1457 
1458 int
1460  const char *sock_filename,
1461  u8 is_server,
1462  u32 * sw_if_index,
1463  u64 feature_mask,
1464  u8 renumber, u32 custom_dev_instance, u8 * hwaddr)
1465 {
1466  vhost_user_intf_t *vui = NULL;
1467  u32 sw_if_idx = ~0;
1468  int rv = 0;
1469  int server_sock_fd = -1;
1471  uword *if_index;
1472 
1473  if (sock_filename == NULL || !(strlen (sock_filename) > 0))
1474  {
1475  return VNET_API_ERROR_INVALID_ARGUMENT;
1476  }
1477 
1478  if_index = mhash_get (&vum->if_index_by_sock_name, (void *) sock_filename);
1479  if (if_index)
1480  {
1481  if (sw_if_index)
1482  {
1483  vui = &vum->vhost_user_interfaces[*if_index];
1484  *sw_if_index = vui->sw_if_index;
1485  }
1486  return VNET_API_ERROR_IF_ALREADY_EXISTS;
1487  }
1488 
1489  if (is_server)
1490  {
1491  if ((rv =
1492  vhost_user_init_server_sock (sock_filename, &server_sock_fd)) != 0)
1493  {
1494  return rv;
1495  }
1496  }
1497 
1498  /* Protect the uninitialized vui from being dispatched by rx/tx */
1500  pool_get (vhost_user_main.vhost_user_interfaces, vui);
1501  vhost_user_create_ethernet (vnm, vm, vui, hwaddr);
1503 
1504  vhost_user_vui_init (vnm, vui, server_sock_fd, sock_filename,
1505  feature_mask, &sw_if_idx);
1506  vnet_sw_interface_set_mtu (vnm, vui->sw_if_index, 9000);
1508 
1509  if (renumber)
1510  vnet_interface_name_renumber (sw_if_idx, custom_dev_instance);
1511 
1512  if (sw_if_index)
1513  *sw_if_index = sw_if_idx;
1514 
1515  // Process node must connect
1517 
1518  return rv;
1519 }
1520 
1521 int
1523  const char *sock_filename,
1524  u8 is_server,
1525  u32 sw_if_index,
1526  u64 feature_mask, u8 renumber, u32 custom_dev_instance)
1527 {
1529  vhost_user_intf_t *vui = NULL;
1530  u32 sw_if_idx = ~0;
1531  int server_sock_fd = -1;
1532  int rv = 0;
1533  vnet_hw_interface_t *hwif;
1534  uword *if_index;
1535 
1536  if (!(hwif = vnet_get_sup_hw_interface (vnm, sw_if_index)) ||
1538  return VNET_API_ERROR_INVALID_SW_IF_INDEX;
1539 
1540  if (sock_filename == NULL || !(strlen (sock_filename) > 0))
1541  return VNET_API_ERROR_INVALID_ARGUMENT;
1542 
1544 
1545  /*
1546  * Disallow changing the interface to have the same path name
1547  * as other interface
1548  */
1549  if_index = mhash_get (&vum->if_index_by_sock_name, (void *) sock_filename);
1550  if (if_index && (*if_index != vui->if_index))
1551  return VNET_API_ERROR_IF_ALREADY_EXISTS;
1552 
1553  // First try to open server socket
1554  if (is_server)
1555  if ((rv = vhost_user_init_server_sock (sock_filename,
1556  &server_sock_fd)) != 0)
1557  return rv;
1558 
1559  vhost_user_term_if (vui);
1560  vhost_user_vui_init (vnm, vui, server_sock_fd,
1561  sock_filename, feature_mask, &sw_if_idx);
1562 
1563  if (renumber)
1564  vnet_interface_name_renumber (sw_if_idx, custom_dev_instance);
1565 
1566  // Process node must connect
1568 
1569  return rv;
1570 }
1571 
1572 clib_error_t *
1574  unformat_input_t * input,
1575  vlib_cli_command_t * cmd)
1576 {
1577  unformat_input_t _line_input, *line_input = &_line_input;
1578  u8 *sock_filename = NULL;
1579  u32 sw_if_index;
1580  u8 is_server = 0;
1581  u64 feature_mask = (u64) ~ (0ULL);
1582  u8 renumber = 0;
1583  u32 custom_dev_instance = ~0;
1584  u8 hwaddr[6];
1585  u8 *hw = NULL;
1586  clib_error_t *error = NULL;
1587 
1588  /* Get a line of input. */
1589  if (!unformat_user (input, unformat_line_input, line_input))
1590  return 0;
1591 
1592  while (unformat_check_input (line_input) != UNFORMAT_END_OF_INPUT)
1593  {
1594  if (unformat (line_input, "socket %s", &sock_filename))
1595  ;
1596  else if (unformat (line_input, "server"))
1597  is_server = 1;
1598  else if (unformat (line_input, "feature-mask 0x%llx", &feature_mask))
1599  ;
1600  else
1601  if (unformat
1602  (line_input, "hwaddr %U", unformat_ethernet_address, hwaddr))
1603  hw = hwaddr;
1604  else if (unformat (line_input, "renumber %d", &custom_dev_instance))
1605  {
1606  renumber = 1;
1607  }
1608  else
1609  {
1610  error = clib_error_return (0, "unknown input `%U'",
1611  format_unformat_error, line_input);
1612  goto done;
1613  }
1614  }
1615 
1616  vnet_main_t *vnm = vnet_get_main ();
1617 
1618  int rv;
1619  if ((rv = vhost_user_create_if (vnm, vm, (char *) sock_filename,
1620  is_server, &sw_if_index, feature_mask,
1621  renumber, custom_dev_instance, hw)))
1622  {
1623  error = clib_error_return (0, "vhost_user_create_if returned %d", rv);
1624  goto done;
1625  }
1626 
1628  sw_if_index);
1629 
1630 done:
1631  vec_free (sock_filename);
1632  unformat_free (line_input);
1633 
1634  return error;
1635 }
1636 
1637 clib_error_t *
1639  unformat_input_t * input,
1640  vlib_cli_command_t * cmd)
1641 {
1642  unformat_input_t _line_input, *line_input = &_line_input;
1643  u32 sw_if_index = ~0;
1644  vnet_main_t *vnm = vnet_get_main ();
1645  clib_error_t *error = NULL;
1646 
1647  /* Get a line of input. */
1648  if (!unformat_user (input, unformat_line_input, line_input))
1649  return 0;
1650 
1651  while (unformat_check_input (line_input) != UNFORMAT_END_OF_INPUT)
1652  {
1653  if (unformat (line_input, "sw_if_index %d", &sw_if_index))
1654  ;
1655  else if (unformat
1656  (line_input, "%U", unformat_vnet_sw_interface, vnm,
1657  &sw_if_index))
1658  {
1659  vnet_hw_interface_t *hwif =
1660  vnet_get_sup_hw_interface (vnm, sw_if_index);
1661  if (hwif == NULL ||
1663  {
1664  error = clib_error_return (0, "Not a vhost interface");
1665  goto done;
1666  }
1667  }
1668  else
1669  {
1670  error = clib_error_return (0, "unknown input `%U'",
1671  format_unformat_error, line_input);
1672  goto done;
1673  }
1674  }
1675 
1676  vhost_user_delete_if (vnm, vm, sw_if_index);
1677 
1678 done:
1679  unformat_free (line_input);
1680 
1681  return error;
1682 }
1683 
1684 int
1686  vhost_user_intf_details_t ** out_vuids)
1687 {
1688  int rv = 0;
1690  vhost_user_intf_t *vui;
1691  vhost_user_intf_details_t *r_vuids = NULL;
1693  u32 *hw_if_indices = 0;
1695  u8 *s = NULL;
1696  int i;
1697 
1698  if (!out_vuids)
1699  return -1;
1700 
1702  vec_add1 (hw_if_indices, vui->hw_if_index);
1703  );
1704 
1705  for (i = 0; i < vec_len (hw_if_indices); i++)
1706  {
1707  hi = vnet_get_hw_interface (vnm, hw_if_indices[i]);
1709 
1710  vec_add2 (r_vuids, vuid, 1);
1711  vuid->sw_if_index = vui->sw_if_index;
1712  vuid->virtio_net_hdr_sz = vui->virtio_net_hdr_sz;
1713  vuid->features = vui->features;
1714  vuid->num_regions = vui->nregions;
1715  vuid->is_server = vui->unix_server_index != ~0;
1716  vuid->sock_errno = vui->sock_errno;
1717  strncpy ((char *) vuid->sock_filename, (char *) vui->sock_filename,
1718  sizeof (vuid->sock_filename));
1719  vuid->sock_filename[ARRAY_LEN (vuid->sock_filename) - 1] = '\0';
1720  s = format (s, "%v%c", hi->name, 0);
1721 
1722  strncpy ((char *) vuid->if_name, (char *) s,
1723  ARRAY_LEN (vuid->if_name) - 1);
1724  _vec_len (s) = 0;
1725  }
1726 
1727  vec_free (s);
1728  vec_free (hw_if_indices);
1729 
1730  *out_vuids = r_vuids;
1731 
1732  return rv;
1733 }
1734 
1735 clib_error_t *
1737  unformat_input_t * input,
1738  vlib_cli_command_t * cmd)
1739 {
1740  clib_error_t *error = 0;
1741  vnet_main_t *vnm = vnet_get_main ();
1743  vhost_user_intf_t *vui;
1744  u32 hw_if_index, *hw_if_indices = 0;
1746  u16 qid;
1747  u32 ci;
1748  int i, j, q;
1749  int show_descr = 0;
1750  struct feat_struct
1751  {
1752  u8 bit;
1753  char *str;
1754  };
1755  struct feat_struct *feat_entry;
1756 
1757  static struct feat_struct feat_array[] = {
1758 #define _(s,b) { .str = #s, .bit = b, },
1760 #undef _
1761  {.str = NULL}
1762  };
1763 
1764 #define foreach_protocol_feature \
1765  _(VHOST_USER_PROTOCOL_F_MQ) \
1766  _(VHOST_USER_PROTOCOL_F_LOG_SHMFD)
1767 
1768  static struct feat_struct proto_feat_array[] = {
1769 #define _(s) { .str = #s, .bit = s},
1771 #undef _
1772  {.str = NULL}
1773  };
1774 
1775  while (unformat_check_input (input) != UNFORMAT_END_OF_INPUT)
1776  {
1777  if (unformat
1778  (input, "%U", unformat_vnet_hw_interface, vnm, &hw_if_index))
1779  {
1780  hi = vnet_get_hw_interface (vnm, hw_if_index);
1781  if (vhost_user_device_class.index != hi->dev_class_index)
1782  {
1783  error = clib_error_return (0, "unknown input `%U'",
1784  format_unformat_error, input);
1785  goto done;
1786  }
1787  vec_add1 (hw_if_indices, hw_if_index);
1788  }
1789  else if (unformat (input, "descriptors") || unformat (input, "desc"))
1790  show_descr = 1;
1791  else
1792  {
1793  error = clib_error_return (0, "unknown input `%U'",
1794  format_unformat_error, input);
1795  goto done;
1796  }
1797  }
1798  if (vec_len (hw_if_indices) == 0)
1799  {
1801  vec_add1 (hw_if_indices, vui->hw_if_index);
1802  );
1803  }
1804  vlib_cli_output (vm, "Virtio vhost-user interfaces");
1805  vlib_cli_output (vm, "Global:\n coalesce frames %d time %e",
1806  vum->coalesce_frames, vum->coalesce_time);
1807  vlib_cli_output (vm, " number of rx virtqueues in interrupt mode: %d",
1808  vum->ifq_count);
1809 
1810  for (i = 0; i < vec_len (hw_if_indices); i++)
1811  {
1812  hi = vnet_get_hw_interface (vnm, hw_if_indices[i]);
1814  vlib_cli_output (vm, "Interface: %U (ifindex %d)",
1815  format_vnet_hw_if_index_name, vnm, hw_if_indices[i],
1816  hw_if_indices[i]);
1817 
1818  vlib_cli_output (vm, "virtio_net_hdr_sz %d\n"
1819  " features mask (0x%llx): \n"
1820  " features (0x%llx): \n",
1821  vui->virtio_net_hdr_sz, vui->feature_mask,
1822  vui->features);
1823 
1824  feat_entry = (struct feat_struct *) &feat_array;
1825  while (feat_entry->str)
1826  {
1827  if (vui->features & (1ULL << feat_entry->bit))
1828  vlib_cli_output (vm, " %s (%d)", feat_entry->str,
1829  feat_entry->bit);
1830  feat_entry++;
1831  }
1832 
1833  vlib_cli_output (vm, " protocol features (0x%llx)",
1834  vui->protocol_features);
1835  feat_entry = (struct feat_struct *) &proto_feat_array;
1836  while (feat_entry->str)
1837  {
1838  if (vui->protocol_features & (1ULL << feat_entry->bit))
1839  vlib_cli_output (vm, " %s (%d)", feat_entry->str,
1840  feat_entry->bit);
1841  feat_entry++;
1842  }
1843 
1844  vlib_cli_output (vm, "\n");
1845 
1846  vlib_cli_output (vm, " socket filename %s type %s errno \"%s\"\n\n",
1847  vui->sock_filename,
1848  (vui->unix_server_index != ~0) ? "server" : "client",
1849  strerror (vui->sock_errno));
1850 
1851  vlib_cli_output (vm, " rx placement: ");
1852 
1853  for (qid = 1; qid < VHOST_VRING_MAX_N / 2; qid += 2)
1854  {
1855  vnet_main_t *vnm = vnet_get_main ();
1856  uword thread_index;
1858  vhost_user_vring_t *txvq = &vui->vrings[qid];
1859 
1860  if (txvq->qid == -1)
1861  continue;
1862  thread_index =
1864  qid >> 1);
1865  vnet_hw_interface_get_rx_mode (vnm, vui->hw_if_index, qid >> 1,
1866  &mode);
1867  vlib_cli_output (vm, " thread %d on vring %d, %U\n",
1868  thread_index, qid,
1870  }
1871 
1872  vlib_cli_output (vm, " tx placement: %s\n",
1873  vui->use_tx_spinlock ? "spin-lock" : "lock-free");
1874 
1876  {
1877  vlib_cli_output (vm, " thread %d on vring %d\n", ci,
1878  VHOST_VRING_IDX_RX (vui->per_cpu_tx_qid[ci]));
1879  }
1880 
1881  vlib_cli_output (vm, "\n");
1882 
1883  vlib_cli_output (vm, " Memory regions (total %d)\n", vui->nregions);
1884 
1885  if (vui->nregions)
1886  {
1887  vlib_cli_output (vm,
1888  " region fd guest_phys_addr memory_size userspace_addr mmap_offset mmap_addr\n");
1889  vlib_cli_output (vm,
1890  " ====== ===== ================== ================== ================== ================== ==================\n");
1891  }
1892  for (j = 0; j < vui->nregions; j++)
1893  {
1894  vlib_cli_output (vm,
1895  " %d %-5d 0x%016lx 0x%016lx 0x%016lx 0x%016lx 0x%016lx\n",
1896  j, vui->region_mmap_fd[j],
1897  vui->regions[j].guest_phys_addr,
1898  vui->regions[j].memory_size,
1899  vui->regions[j].userspace_addr,
1900  vui->regions[j].mmap_offset,
1902  }
1903  for (q = 0; q < VHOST_VRING_MAX_N; q++)
1904  {
1905  if (!vui->vrings[q].started)
1906  continue;
1907 
1908  vlib_cli_output (vm, "\n Virtqueue %d (%s%s)\n", q,
1909  (q & 1) ? "RX" : "TX",
1910  vui->vrings[q].enabled ? "" : " disabled");
1911 
1912  vlib_cli_output (vm,
1913  " qsz %d last_avail_idx %d last_used_idx %d\n",
1914  vui->vrings[q].qsz_mask + 1,
1915  vui->vrings[q].last_avail_idx,
1916  vui->vrings[q].last_used_idx);
1917 
1918  if (vui->vrings[q].avail && vui->vrings[q].used)
1919  vlib_cli_output (vm,
1920  " avail.flags %x avail.idx %d used.flags %x used.idx %d\n",
1921  vui->vrings[q].avail->flags,
1922  vui->vrings[q].avail->idx,
1923  vui->vrings[q].used->flags,
1924  vui->vrings[q].used->idx);
1925 
1926  int kickfd = UNIX_GET_FD (vui->vrings[q].kickfd_idx);
1927  int callfd = UNIX_GET_FD (vui->vrings[q].callfd_idx);
1928  vlib_cli_output (vm, " kickfd %d callfd %d errfd %d\n",
1929  kickfd, callfd, vui->vrings[q].errfd);
1930 
1931  if (show_descr)
1932  {
1933  vlib_cli_output (vm, "\n descriptor table:\n");
1934  vlib_cli_output (vm,
1935  " id addr len flags next user_addr\n");
1936  vlib_cli_output (vm,
1937  " ===== ================== ===== ====== ===== ==================\n");
1938  for (j = 0; j < vui->vrings[q].qsz_mask + 1; j++)
1939  {
1940  u32 mem_hint = 0;
1941  vlib_cli_output (vm,
1942  " %-5d 0x%016lx %-5d 0x%04x %-5d 0x%016lx\n",
1943  j, vui->vrings[q].desc[j].addr,
1944  vui->vrings[q].desc[j].len,
1945  vui->vrings[q].desc[j].flags,
1946  vui->vrings[q].desc[j].next,
1948  (vui,
1949  vui->vrings[q].desc[j].
1950  addr, &mem_hint)));
1951  }
1952  }
1953  }
1954  vlib_cli_output (vm, "\n");
1955  }
1956 done:
1957  vec_free (hw_if_indices);
1958  return error;
1959 }
1960 
1961 /*
1962  * CLI functions
1963  */
1964 
1965 /*?
1966  * Create a vHost User interface. Once created, a new virtual interface
1967  * will exist with the name '<em>VirtualEthernet0/0/x</em>', where '<em>x</em>'
1968  * is the next free index.
1969  *
1970  * There are several parameters associated with a vHost interface:
1971  *
1972  * - <b>socket <socket-filename></b> - Name of the linux socket used by hypervisor
1973  * and VPP to manage the vHost interface. If in '<em>server</em>' mode, VPP will
1974  * create the socket if it does not already exist. If in '<em>client</em>' mode,
1975  * hypervisor will create the socket if it does not already exist. The VPP code
1976  * is indifferent to the file location. However, if SELinux is enabled, then the
1977  * socket needs to be created in '<em>/var/run/vpp/</em>'.
1978  *
1979  * - <b>server</b> - Optional flag to indicate that VPP should be the server for
1980  * the linux socket. If not provided, VPP will be the client. In '<em>server</em>'
1981  * mode, the VM can be reset without tearing down the vHost Interface. In
1982  * '<em>client</em>' mode, VPP can be reset without bringing down the VM and
1983  * tearing down the vHost Interface.
1984  *
1985  * - <b>feature-mask <hex></b> - Optional virtio/vhost feature set negotiated at
1986  * startup. <b>This is intended for degugging only.</b> It is recommended that this
1987  * parameter not be used except by experienced users. By default, all supported
1988  * features will be advertised. Otherwise, provide the set of features desired.
1989  * - 0x000008000 (15) - VIRTIO_NET_F_MRG_RXBUF
1990  * - 0x000020000 (17) - VIRTIO_NET_F_CTRL_VQ
1991  * - 0x000200000 (21) - VIRTIO_NET_F_GUEST_ANNOUNCE
1992  * - 0x000400000 (22) - VIRTIO_NET_F_MQ
1993  * - 0x004000000 (26) - VHOST_F_LOG_ALL
1994  * - 0x008000000 (27) - VIRTIO_F_ANY_LAYOUT
1995  * - 0x010000000 (28) - VIRTIO_F_INDIRECT_DESC
1996  * - 0x040000000 (30) - VHOST_USER_F_PROTOCOL_FEATURES
1997  * - 0x100000000 (32) - VIRTIO_F_VERSION_1
1998  *
1999  * - <b>hwaddr <mac-addr></b> - Optional ethernet address, can be in either
2000  * X:X:X:X:X:X unix or X.X.X cisco format.
2001  *
2002  * - <b>renumber <dev_instance></b> - Optional parameter which allows the instance
2003  * in the name to be specified. If instance already exists, name will be used
2004  * anyway and multiple instances will have the same name. Use with caution.
2005  *
2006  * @cliexpar
2007  * Example of how to create a vhost interface with VPP as the client and all features enabled:
2008  * @cliexstart{create vhost-user socket /var/run/vpp/vhost1.sock}
2009  * VirtualEthernet0/0/0
2010  * @cliexend
2011  * Example of how to create a vhost interface with VPP as the server and with just
2012  * multiple queues enabled:
2013  * @cliexstart{create vhost-user socket /var/run/vpp/vhost2.sock server feature-mask 0x40400000}
2014  * VirtualEthernet0/0/1
2015  * @cliexend
2016  * Once the vHost interface is created, enable the interface using:
2017  * @cliexcmd{set interface state VirtualEthernet0/0/0 up}
2018 ?*/
2019 /* *INDENT-OFF* */
2020 VLIB_CLI_COMMAND (vhost_user_connect_command, static) = {
2021  .path = "create vhost-user",
2022  .short_help = "create vhost-user socket <socket-filename> [server] "
2023  "[feature-mask <hex>] [hwaddr <mac-addr>] [renumber <dev_instance>] ",
2024  .function = vhost_user_connect_command_fn,
2025  .is_mp_safe = 1,
2026 };
2027 /* *INDENT-ON* */
2028 
2029 /*?
2030  * Delete a vHost User interface using the interface name or the
2031  * software interface index. Use the '<em>show interface</em>'
2032  * command to determine the software interface index. On deletion,
2033  * the linux socket will not be deleted.
2034  *
2035  * @cliexpar
2036  * Example of how to delete a vhost interface by name:
2037  * @cliexcmd{delete vhost-user VirtualEthernet0/0/1}
2038  * Example of how to delete a vhost interface by software interface index:
2039  * @cliexcmd{delete vhost-user sw_if_index 1}
2040 ?*/
2041 /* *INDENT-OFF* */
2042 VLIB_CLI_COMMAND (vhost_user_delete_command, static) = {
2043  .path = "delete vhost-user",
2044  .short_help = "delete vhost-user {<interface> | sw_if_index <sw_idx>}",
2045  .function = vhost_user_delete_command_fn,
2046 };
2047 
2048 /*?
2049  * Display the attributes of a single vHost User interface (provide interface
2050  * name), multiple vHost User interfaces (provide a list of interface names seperated
2051  * by spaces) or all Vhost User interfaces (omit an interface name to display all
2052  * vHost interfaces).
2053  *
2054  * @cliexpar
2055  * @parblock
2056  * Example of how to display a vhost interface:
2057  * @cliexstart{show vhost-user VirtualEthernet0/0/0}
2058  * Virtio vhost-user interfaces
2059  * Global:
2060  * coalesce frames 32 time 1e-3
2061  * Interface: VirtualEthernet0/0/0 (ifindex 1)
2062  * virtio_net_hdr_sz 12
2063  * features mask (0xffffffffffffffff):
2064  * features (0x50408000):
2065  * VIRTIO_NET_F_MRG_RXBUF (15)
2066  * VIRTIO_NET_F_MQ (22)
2067  * VIRTIO_F_INDIRECT_DESC (28)
2068  * VHOST_USER_F_PROTOCOL_FEATURES (30)
2069  * protocol features (0x3)
2070  * VHOST_USER_PROTOCOL_F_MQ (0)
2071  * VHOST_USER_PROTOCOL_F_LOG_SHMFD (1)
2072  *
2073  * socket filename /var/run/vpp/vhost1.sock type client errno "Success"
2074  *
2075  * rx placement:
2076  * thread 1 on vring 1
2077  * thread 1 on vring 5
2078  * thread 2 on vring 3
2079  * thread 2 on vring 7
2080  * tx placement: spin-lock
2081  * thread 0 on vring 0
2082  * thread 1 on vring 2
2083  * thread 2 on vring 0
2084  *
2085  * Memory regions (total 2)
2086  * region fd guest_phys_addr memory_size userspace_addr mmap_offset mmap_addr
2087  * ====== ===== ================== ================== ================== ================== ==================
2088  * 0 60 0x0000000000000000 0x00000000000a0000 0x00002aaaaac00000 0x0000000000000000 0x00002aab2b400000
2089  * 1 61 0x00000000000c0000 0x000000003ff40000 0x00002aaaaacc0000 0x00000000000c0000 0x00002aababcc0000
2090  *
2091  * Virtqueue 0 (TX)
2092  * qsz 256 last_avail_idx 0 last_used_idx 0
2093  * avail.flags 1 avail.idx 128 used.flags 1 used.idx 0
2094  * kickfd 62 callfd 64 errfd -1
2095  *
2096  * Virtqueue 1 (RX)
2097  * qsz 256 last_avail_idx 0 last_used_idx 0
2098  * avail.flags 1 avail.idx 0 used.flags 1 used.idx 0
2099  * kickfd 65 callfd 66 errfd -1
2100  *
2101  * Virtqueue 2 (TX)
2102  * qsz 256 last_avail_idx 0 last_used_idx 0
2103  * avail.flags 1 avail.idx 128 used.flags 1 used.idx 0
2104  * kickfd 63 callfd 70 errfd -1
2105  *
2106  * Virtqueue 3 (RX)
2107  * qsz 256 last_avail_idx 0 last_used_idx 0
2108  * avail.flags 1 avail.idx 0 used.flags 1 used.idx 0
2109  * kickfd 72 callfd 74 errfd -1
2110  *
2111  * Virtqueue 4 (TX disabled)
2112  * qsz 256 last_avail_idx 0 last_used_idx 0
2113  * avail.flags 1 avail.idx 0 used.flags 1 used.idx 0
2114  * kickfd 76 callfd 78 errfd -1
2115  *
2116  * Virtqueue 5 (RX disabled)
2117  * qsz 256 last_avail_idx 0 last_used_idx 0
2118  * avail.flags 1 avail.idx 0 used.flags 1 used.idx 0
2119  * kickfd 80 callfd 82 errfd -1
2120  *
2121  * Virtqueue 6 (TX disabled)
2122  * qsz 256 last_avail_idx 0 last_used_idx 0
2123  * avail.flags 1 avail.idx 0 used.flags 1 used.idx 0
2124  * kickfd 84 callfd 86 errfd -1
2125  *
2126  * Virtqueue 7 (RX disabled)
2127  * qsz 256 last_avail_idx 0 last_used_idx 0
2128  * avail.flags 1 avail.idx 0 used.flags 1 used.idx 0
2129  * kickfd 88 callfd 90 errfd -1
2130  *
2131  * @cliexend
2132  *
2133  * The optional '<em>descriptors</em>' parameter will display the same output as
2134  * the previous example but will include the descriptor table for each queue.
2135  * The output is truncated below:
2136  * @cliexstart{show vhost-user VirtualEthernet0/0/0 descriptors}
2137  * Virtio vhost-user interfaces
2138  * Global:
2139  * coalesce frames 32 time 1e-3
2140  * Interface: VirtualEthernet0/0/0 (ifindex 1)
2141  * virtio_net_hdr_sz 12
2142  * features mask (0xffffffffffffffff):
2143  * features (0x50408000):
2144  * VIRTIO_NET_F_MRG_RXBUF (15)
2145  * VIRTIO_NET_F_MQ (22)
2146  * :
2147  * Virtqueue 0 (TX)
2148  * qsz 256 last_avail_idx 0 last_used_idx 0
2149  * avail.flags 1 avail.idx 128 used.flags 1 used.idx 0
2150  * kickfd 62 callfd 64 errfd -1
2151  *
2152  * descriptor table:
2153  * id addr len flags next user_addr
2154  * ===== ================== ===== ====== ===== ==================
2155  * 0 0x0000000010b6e974 2060 0x0002 1 0x00002aabbc76e974
2156  * 1 0x0000000010b6e034 2060 0x0002 2 0x00002aabbc76e034
2157  * 2 0x0000000010b6d6f4 2060 0x0002 3 0x00002aabbc76d6f4
2158  * 3 0x0000000010b6cdb4 2060 0x0002 4 0x00002aabbc76cdb4
2159  * 4 0x0000000010b6c474 2060 0x0002 5 0x00002aabbc76c474
2160  * 5 0x0000000010b6bb34 2060 0x0002 6 0x00002aabbc76bb34
2161  * 6 0x0000000010b6b1f4 2060 0x0002 7 0x00002aabbc76b1f4
2162  * 7 0x0000000010b6a8b4 2060 0x0002 8 0x00002aabbc76a8b4
2163  * 8 0x0000000010b69f74 2060 0x0002 9 0x00002aabbc769f74
2164  * 9 0x0000000010b69634 2060 0x0002 10 0x00002aabbc769634
2165  * 10 0x0000000010b68cf4 2060 0x0002 11 0x00002aabbc768cf4
2166  * :
2167  * 249 0x0000000000000000 0 0x0000 250 0x00002aab2b400000
2168  * 250 0x0000000000000000 0 0x0000 251 0x00002aab2b400000
2169  * 251 0x0000000000000000 0 0x0000 252 0x00002aab2b400000
2170  * 252 0x0000000000000000 0 0x0000 253 0x00002aab2b400000
2171  * 253 0x0000000000000000 0 0x0000 254 0x00002aab2b400000
2172  * 254 0x0000000000000000 0 0x0000 255 0x00002aab2b400000
2173  * 255 0x0000000000000000 0 0x0000 32768 0x00002aab2b400000
2174  *
2175  * Virtqueue 1 (RX)
2176  * qsz 256 last_avail_idx 0 last_used_idx 0
2177  * :
2178  * @cliexend
2179  * @endparblock
2180 ?*/
2181 /* *INDENT-OFF* */
2182 VLIB_CLI_COMMAND (show_vhost_user_command, static) = {
2183  .path = "show vhost-user",
2184  .short_help = "show vhost-user [<interface> [<interface> [..]]] [descriptors]",
2185  .function = show_vhost_user_command_fn,
2186 };
2187 /* *INDENT-ON* */
2188 
2189 
2190 static clib_error_t *
2192 {
2194 
2195  while (unformat_check_input (input) != UNFORMAT_END_OF_INPUT)
2196  {
2197  if (unformat (input, "coalesce-frames %d", &vum->coalesce_frames))
2198  ;
2199  else if (unformat (input, "coalesce-time %f", &vum->coalesce_time))
2200  ;
2201  else if (unformat (input, "dont-dump-memory"))
2202  vum->dont_dump_vhost_user_memory = 1;
2203  else
2204  return clib_error_return (0, "unknown input `%U'",
2205  format_unformat_error, input);
2206  }
2207 
2208  return 0;
2209 }
2210 
2211 /* vhost-user { ... } configuration. */
2212 VLIB_CONFIG_FUNCTION (vhost_user_config, "vhost-user");
2213 
2214 void
2216 {
2218  vhost_user_intf_t *vui;
2219 
2220  if (vum->dont_dump_vhost_user_memory)
2221  {
2223  unmap_all_mem_regions (vui);
2224  );
2225  }
2226 }
2227 
2228 /*
2229  * fd.io coding-style-patch-verification: ON
2230  *
2231  * Local Variables:
2232  * eval: (c-set-style "gnu")
2233  * End:
2234  */
unformat_function_t unformat_vnet_hw_interface
vlib_log_class_t vlib_log_register_class(char *class, char *subclass)
Definition: log.c:227
#define vec_validate(V, I)
Make sure vector is long enough for given index (no header, unspecified alignment) ...
Definition: vec.h:439
u32 sw_if_index
Definition: ipsec_gre.api:37
u32 len
Definition: pci.h:200
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:237
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:703
VNET_HW_INTERFACE_CLASS(vhost_interface_class, static)
vhost_cpu_t * cpus
Per-CPU data for vhost-user.
Definition: vhost_user.h:350
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:47
static vnet_hw_interface_t * vnet_get_sup_hw_interface(vnet_main_t *vnm, u32 sw_if_index)
u64 region_guest_addr_hi[VHOST_MEMORY_MAX_NREGIONS]
Definition: vhost_user.h:288
static clib_error_t * vhost_user_socksvr_accept_ready(clib_file_t *uf)
Definition: vhost_user.c:918
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:353
static f64 vlib_time_now(vlib_main_t *vm)
Definition: main.h:255
static_always_inline void vhost_user_update_iface_state(vhost_user_intf_t *vui)
Definition: vhost_user.c:197
u16 next
Definition: pci.h:202
#define VLIB_BUFFER_PRE_DATA_SIZE
Definition: buffer.h:51
vring_avail_t * avail
Definition: vhost_user.h:238
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:1354
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)
Definition: vhost_user.c:1399
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:263
uword unformat_user(unformat_input_t *input, unformat_function_t *func,...)
Definition: unformat.c:983
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:208
u16 flags
Definition: pci.h:221
#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:239
vhost_vring_addr_t addr
Definition: vhost_user.h:121
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)
Definition: vhost_user.c:1522
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:440
#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
#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:199
vnet_hw_interface_rx_mode
Definition: interface.h:52
#define static_always_inline
Definition: clib.h:99
#define pool_foreach(VAR, POOL, BODY)
Iterate through pool.
Definition: pool.h:493
#define VLIB_INIT_FUNCTION(x)
Definition: init.h:163
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:546
static clib_error_t * ip4_init(vlib_main_t *vm)
Definition: ip4_input.c:351
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:298
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:494
#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:1185
#define vlib_call_init_function(vm, x)
Definition: init.h:260
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:903
void vhost_user_unmap_all(void)
Definition: vhost_user.c:2215
unformat_function_t unformat_line_input
Definition: format.h:282
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:1197
static void vlib_process_signal_event(vlib_main_t *vm, uword node_index, uword type_opaque, uword data)
Definition: node_funcs.h:964
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:353
uword mhash_set_mem(mhash_t *h, void *key, uword *new_value, uword *old_value)
Definition: mhash.c:271
vlib_node_registration_t vhost_user_input_node
(constructor) VLIB_REGISTER_NODE (vhost_user_input_node)
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)
Definition: vhost_user.c:1459
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:1244
format_function_t format_vnet_hw_interface_rx_mode
#define VLIB_CONFIG_FUNCTION(x, n,...)
Definition: init.h:172
char sock_filename[256]
Definition: vhost_user.h:273
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:289
int vhost_user_dump_ifs(vnet_main_t *vnm, vlib_main_t *vm, vhost_user_intf_details_t **out_vuids)
Definition: vhost_user.c:1685
vhost_user_memory_region_t regions[VHOST_MEMORY_MAX_NREGIONS]
Definition: vhost_user.h:285
vlib_log_class_t log_default
Definition: vhost_user.h:359
#define VHOST_VRING_IDX_RX(qid)
Definition: vhost_user.h:23
u32 * show_dev_instance_by_real_dev_instance
Definition: vhost_user.h:344
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:343
#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:144
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:2191
#define VLIB_MAIN_LOOP_EXIT_FUNCTION(x)
Definition: init.h:168
#define VHOST_USER_EVENT_STOP_TIMER
Definition: vhost_user.h:264
#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:1086
#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:1638
vl_api_vxlan_gbp_api_tunnel_mode_t mode
Definition: vxlan_gbp.api:44
#define foreach_protocol_feature
u16 flags
Definition: pci.h:207
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:1324
#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:1736
#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:205
volatile u32 * vring_locks[VHOST_VRING_MAX_N]
Definition: vhost_user.h:293
mhash_t if_index_by_sock_name
Definition: vhost_user.h:341
static clib_error_t * vhost_user_exit(vlib_main_t *vm)
Definition: vhost_user.c:1302
#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:1364
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:286
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:292
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:504
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:162
#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:1573
void vnet_sw_interface_set_mtu(vnet_main_t *vnm, u32 sw_if_index, u32 mtu)
Definition: interface.c:642
u32 rx_buffers_len
Definition: vhost_user.h:328
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:222
static void * clib_mem_alloc_aligned(uword size, uword align)
Definition: mem.h:140
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:287
u16 flags
Definition: pci.h:201
#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:954
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:989
void vlib_cli_output(vlib_main_t *vm, char *fmt,...)
Definition: cli.c:725
vl_api_address_union_t un
Definition: ip_types.api:32
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:972
static uword unformat_check_input(unformat_input_t *i)
Definition: format.h:170
int dont_dump_vhost_user_memory
Definition: vhost_user.h:347