FD.io VPP  v18.01.2-1-g9b554f3
Vector Packet Processing
pci.c
Go to the documentation of this file.
1 /*
2  * Copyright (c) 2016 Cisco and/or its affiliates.
3  * Licensed under the Apache License, Version 2.0 (the "License");
4  * you may not use this file except in compliance with the License.
5  * You may obtain a copy of the License at:
6  *
7  * http://www.apache.org/licenses/LICENSE-2.0
8  *
9  * Unless required by applicable law or agreed to in writing, software
10  * distributed under the License is distributed on an "AS IS" BASIS,
11  * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
12  * See the License for the specific language governing permissions and
13  * limitations under the License.
14  */
15 /*
16  * pci.c: Linux user space PCI bus management.
17  *
18  * Copyright (c) 2008 Eliot Dresselhaus
19  *
20  * Permission is hereby granted, free of charge, to any person obtaining
21  * a copy of this software and associated documentation files (the
22  * "Software"), to deal in the Software without restriction, including
23  * without limitation the rights to use, copy, modify, merge, publish,
24  * distribute, sublicense, and/or sell copies of the Software, and to
25  * permit persons to whom the Software is furnished to do so, subject to
26  * the following conditions:
27  *
28  * The above copyright notice and this permission notice shall be
29  * included in all copies or substantial portions of the Software.
30  *
31  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
32  * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
33  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
34  * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS BE
35  * LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION
36  * OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION
37  * WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
38  */
39 
40 #include <vppinfra/linux/sysfs.h>
41 
42 #include <vlib/vlib.h>
43 #include <vlib/pci/pci.h>
44 #include <vlib/unix/unix.h>
45 
46 #include <sys/types.h>
47 #include <sys/stat.h>
48 #include <fcntl.h>
49 #include <dirent.h>
50 #include <sys/ioctl.h>
51 #include <net/if.h>
52 #include <linux/ethtool.h>
53 #include <linux/sockios.h>
54 #include <linux/vfio.h>
55 #include <sys/eventfd.h>
56 
57 static const char *sysfs_pci_dev_path = "/sys/bus/pci/devices";
58 static const char *sysfs_pci_drv_path = "/sys/bus/pci/drivers";
59 static char *sysfs_mod_vfio_noiommu =
60  "/sys/module/vfio/parameters/enable_unsafe_noiommu_mode";
61 
62 typedef struct
63 {
65  vlib_pci_addr_t addr;
66 
67  /* Resource file descriptors. */
69 
70  /* File descriptor for config space read/write. */
71  int config_fd;
72 
73  /* File descriptor for /dev/uio%d */
74  int uio_fd;
75 
76  /* Minor device for uio device. */
78 
79  /* vfio */
81 
82  /* Index given by clib_file_add. */
84 
85  /* Interrupt handler */
86  void (*interrupt_handler) (vlib_pci_dev_handle_t h);
87 
88  /* private data */
90 
92 
93 typedef struct
94 {
95  int group;
96  int fd;
97  int refcnt;
99 
100 /* Pool of PCI devices. */
101 typedef struct
102 {
105 
106  /* VFIO */
109 
110  /* pool of IOMMU groups */
112 
113  /* iommu group pool index by group id hash */
115 
117 
119 
122 {
124  return pool_elt_at_index (lpm->linux_pci_devices, h);
125 }
126 
127 uword
129 {
131  return d->private_data;
132 }
133 
134 void
136 {
138  d->private_data = private_data;
139 }
140 
141 vlib_pci_addr_t *
143 {
145  return &lpm->addr;
146 }
147 
148 /* Call to allocate/initialize the pci subsystem.
149  This is not an init function so that users can explicitly enable
150  pci only when it's needed. */
152 
154 
156 vlib_pci_get_device_info (vlib_pci_addr_t * addr, clib_error_t ** error)
157 {
159  clib_error_t *err;
161  u8 *f = 0;
162  u32 tmp;
163  int fd;
164 
165  di = clib_mem_alloc (sizeof (vlib_pci_device_info_t));
166  memset (di, 0, sizeof (vlib_pci_device_info_t));
167  di->addr.as_u32 = addr->as_u32;
168 
169  u8 *dev_dir_name = format (0, "%s/%U", sysfs_pci_dev_path,
170  format_vlib_pci_addr, addr);
171 
172  f = format (0, "%v/config%c", dev_dir_name, 0);
173  fd = open ((char *) f, O_RDWR);
174 
175  /* Try read-only access if write fails. */
176  if (fd < 0)
177  fd = open ((char *) f, O_RDONLY);
178 
179  if (fd < 0)
180  {
181  err = clib_error_return_unix (0, "open `%s'", f);
182  goto error;
183  }
184 
185  /* You can only read more that 64 bytes of config space as root; so we try to
186  read the full space but fall back to just the first 64 bytes. */
187  if (read (fd, &di->config_data, sizeof (di->config_data)) !=
188  sizeof (di->config_data)
189  && read (fd, &di->config0,
190  sizeof (di->config0)) != sizeof (di->config0))
191  {
192  err = clib_error_return_unix (0, "read `%s'", f);
193  close (fd);
194  goto error;
195  }
196 
197  {
198  static pci_config_header_t all_ones;
199  if (all_ones.vendor_id == 0)
200  memset (&all_ones, ~0, sizeof (all_ones));
201 
202  if (!memcmp (&di->config0.header, &all_ones, sizeof (all_ones)))
203  {
204  err = clib_error_return (0, "invalid PCI config for `%s'", f);
205  close (fd);
206  goto error;
207  }
208  }
209 
210  if (di->config0.header.header_type == 0)
212  else
214 
215  di->numa_node = -1;
216  vec_reset_length (f);
217  f = format (f, "%v/numa_node%c", dev_dir_name, 0);
218  err = clib_sysfs_read ((char *) f, "%u", &di->numa_node);
219  if (err)
220  {
221  di->numa_node = -1;
222  clib_error_free (err);
223  }
224 
225  vec_reset_length (f);
226  f = format (f, "%v/class%c", dev_dir_name, 0);
227  err = clib_sysfs_read ((char *) f, "0x%x", &tmp);
228  if (err)
229  goto error;
230  di->device_class = tmp >> 8;
231 
232  vec_reset_length (f);
233  f = format (f, "%v/vendor%c", dev_dir_name, 0);
234  err = clib_sysfs_read ((char *) f, "0x%x", &tmp);
235  if (err)
236  goto error;
237  di->vendor_id = tmp;
238 
239  vec_reset_length (f);
240  f = format (f, "%v/device%c", dev_dir_name, 0);
241  err = clib_sysfs_read ((char *) f, "0x%x", &tmp);
242  if (err)
243  goto error;
244  di->device_id = tmp;
245 
246  vec_reset_length (f);
247  f = format (f, "%v/driver%c", dev_dir_name, 0);
248  di->driver_name = clib_sysfs_link_to_name ((char *) f);
249 
250  di->iommu_group = -1;
251  if (lpm->vfio_container_fd != -1)
252  {
253  u8 *tmpstr;
254  vec_reset_length (f);
255  f = format (f, "%v/iommu_group%c", dev_dir_name, 0);
256  tmpstr = clib_sysfs_link_to_name ((char *) f);
257  if (tmpstr)
258  {
259  di->iommu_group = atoi ((char *) tmpstr);
260  vec_free (tmpstr);
261  }
262  }
263 
264  close (fd);
265 
266  vec_reset_length (f);
267  f = format (f, "%v/vpd%c", dev_dir_name, 0);
268  fd = open ((char *) f, O_RDONLY);
269  if (fd >= 0)
270  {
271  while (1)
272  {
273  u8 tag[3];
274  u8 *data = 0;
275  uword len;
276 
277  if (read (fd, &tag, 3) != 3)
278  break;
279 
280  if (tag[0] != 0x82 && tag[0] != 0x90 && tag[0] != 0x91)
281  break;
282 
283  len = (tag[2] << 8) | tag[1];
284  vec_validate (data, len);
285 
286  if (read (fd, data, len) != len)
287  {
288  vec_free (data);
289  break;
290  }
291  if (tag[0] == 0x82)
292  di->product_name = data;
293  else if (tag[0] == 0x90)
294  di->vpd_r = data;
295  else if (tag[0] == 0x91)
296  di->vpd_w = data;
297 
298  data = 0;
299  }
300  close (fd);
301  }
302 
303  goto done;
304 
305 error:
307  di = 0;
308 
309 done:
310  vec_free (f);
311  vec_free (dev_dir_name);
312  if (error)
313  *error = err;
314  else
315  clib_error_free (err);
316  return di;
317 }
318 
319 static int
320 directory_exists (char *path)
321 {
322  struct stat s = { 0 };
323  if (stat (path, &s) == -1)
324  return 0;
325 
326  return S_ISDIR (s.st_mode);
327 }
328 
329 clib_error_t *
330 vlib_pci_bind_to_uio (vlib_pci_addr_t * addr, char *uio_drv_name)
331 {
332  clib_error_t *error = 0;
333  u8 *s = 0, *driver_name = 0;
334  DIR *dir = 0;
335  struct dirent *e;
337  int fd, clear_driver_override = 0;
338  u8 *dev_dir_name = format (0, "%s/%U", sysfs_pci_dev_path,
339  format_vlib_pci_addr, addr);
340 
341  di = vlib_pci_get_device_info (addr, &error);
342 
343  if (error)
344  return error;
345 
346  if (strncmp ("auto", uio_drv_name, 5) == 0)
347  {
348  int vfio_pci_loaded = 0;
349 
350  if (directory_exists ("/sys/module/vfio_pci"))
351  vfio_pci_loaded = 1;
352 
353  if (di->iommu_group != -1)
354  {
355  /* device is bound to IOMMU group */
356  if (!vfio_pci_loaded)
357  {
358  error = clib_error_return (0, "Skipping PCI device %U: device "
359  "is bound to IOMMU group and "
360  "vfio-pci driver is not loaded",
361  format_vlib_pci_addr, addr);
362  goto done;
363  }
364  else
365  uio_drv_name = "vfio-pci";
366  }
367  else
368  {
369  /* device is not bound to IOMMU group so we have multiple options */
370  if (vfio_pci_loaded &&
371  (error = clib_sysfs_write (sysfs_mod_vfio_noiommu, "Y")) == 0)
372  uio_drv_name = "vfio-pci";
373  else if (directory_exists ("/sys/module/uio_pci_generic"))
374  uio_drv_name = "uio_pci_generic";
375  else if (directory_exists ("/sys/module/igb_uio"))
376  uio_drv_name = "igb_uio";
377  else
378  {
379  clib_error_free (error);
380  error = clib_error_return (0, "Skipping PCI device %U: missing "
381  "kernel VFIO or UIO driver",
382  format_vlib_pci_addr, addr);
383  goto done;
384  }
385  clib_error_free (error);
386  }
387  }
388 
389  s = format (s, "%v/driver%c", dev_dir_name, 0);
390  driver_name = clib_sysfs_link_to_name ((char *) s);
391  vec_reset_length (s);
392 
393  if (driver_name &&
394  ((strcmp ("vfio-pci", (char *) driver_name) == 0) ||
395  (strcmp ("uio_pci_generic", (char *) driver_name) == 0) ||
396  (strcmp ("igb_uio", (char *) driver_name) == 0)))
397  goto done;
398 
399  /* walk trough all linux interfaces and if interface belonging to
400  this device is founf check if interface is admin up */
401  dir = opendir ("/sys/class/net");
402  s = format (s, "%U%c", format_vlib_pci_addr, addr, 0);
403 
404  if (!dir)
405  {
406  error = clib_error_return (0, "Skipping PCI device %U: failed to "
407  "read /sys/class/net",
408  format_vlib_pci_addr, addr);
409  goto done;
410  }
411 
412  fd = socket (PF_INET, SOCK_DGRAM, 0);
413  if (fd < 0)
414  {
415  error = clib_error_return_unix (0, "socket");
416  goto done;
417  }
418 
419  while ((e = readdir (dir)))
420  {
421  struct ifreq ifr;
422  struct ethtool_drvinfo drvinfo;
423 
424  if (e->d_name[0] == '.') /* skip . and .. */
425  continue;
426 
427  memset (&ifr, 0, sizeof ifr);
428  memset (&drvinfo, 0, sizeof drvinfo);
429  ifr.ifr_data = (char *) &drvinfo;
430  strncpy (ifr.ifr_name, e->d_name, IFNAMSIZ - 1);
431  drvinfo.cmd = ETHTOOL_GDRVINFO;
432  if (ioctl (fd, SIOCETHTOOL, &ifr) < 0)
433  {
434  /* Some interfaces (eg "lo") don't support this ioctl */
435  if ((errno != ENOTSUP) && (errno != ENODEV))
436  clib_unix_warning ("ioctl fetch intf %s bus info error",
437  e->d_name);
438  continue;
439  }
440 
441  if (strcmp ((char *) s, drvinfo.bus_info))
442  continue;
443 
444  memset (&ifr, 0, sizeof (ifr));
445  strncpy (ifr.ifr_name, e->d_name, IFNAMSIZ - 1);
446  if (ioctl (fd, SIOCGIFFLAGS, &ifr) < 0)
447  {
448  error = clib_error_return_unix (0, "ioctl fetch intf %s flags",
449  e->d_name);
450  close (fd);
451  goto done;
452  }
453 
454  if (ifr.ifr_flags & IFF_UP)
455  {
456  error = clib_error_return (0, "Skipping PCI device %U as host "
457  "interface %s is up",
458  format_vlib_pci_addr, addr, e->d_name);
459  close (fd);
460  goto done;
461  }
462  }
463 
464  close (fd);
465  vec_reset_length (s);
466 
467  s = format (s, "%v/driver/unbind%c", dev_dir_name, 0);
468  clib_sysfs_write ((char *) s, "%U", format_vlib_pci_addr, addr);
469  vec_reset_length (s);
470 
471  s = format (s, "%v/driver_override%c", dev_dir_name, 0);
472  if (access ((char *) s, F_OK) == 0)
473  {
474  clib_sysfs_write ((char *) s, "%s", uio_drv_name);
475  clear_driver_override = 1;
476  }
477  else
478  {
479  vec_reset_length (s);
480  s = format (s, "%s/%s/new_id%c", sysfs_pci_drv_path, uio_drv_name, 0);
481  clib_sysfs_write ((char *) s, "0x%04x 0x%04x", di->vendor_id,
482  di->device_id);
483  }
484  vec_reset_length (s);
485 
486  s = format (s, "%s/%s/bind%c", sysfs_pci_drv_path, uio_drv_name, 0);
487  clib_sysfs_write ((char *) s, "%U", format_vlib_pci_addr, addr);
488  vec_reset_length (s);
489 
490  if (clear_driver_override)
491  {
492  s = format (s, "%v/driver_override%c", dev_dir_name, 0);
493  clib_sysfs_write ((char *) s, "%c", 0);
494  vec_reset_length (s);
495  }
496 
497 done:
498  closedir (dir);
499  vec_free (s);
500  vec_free (dev_dir_name);
501  vec_free (driver_name);
502  return error;
503 }
504 
505 
506 static clib_error_t *
507 scan_uio_dir (void *arg, u8 * path_name, u8 * file_name)
508 {
509  linux_pci_device_t *l = arg;
510  unformat_input_t input;
511 
512  unformat_init_string (&input, (char *) file_name, vec_len (file_name));
513 
514  if (!unformat (&input, "uio%d", &l->uio_minor))
515  abort ();
516 
517  unformat_free (&input);
518  return 0;
519 }
520 
521 static clib_error_t *
523 {
524  int __attribute__ ((unused)) rv;
527 
528  u32 icount;
529  rv = read (uf->file_descriptor, &icount, 4);
530 
531  if (p->interrupt_handler)
532  p->interrupt_handler (h);
533 
535 
536  return /* no error */ 0;
537 }
538 
539 static clib_error_t *
541 {
542  clib_error_t *err = 0;
543  struct vfio_irq_set i = {
544  .argsz = sizeof (struct vfio_irq_set),
545  .start = 0,
546  .count = 1,
547  .index = VFIO_PCI_INTX_IRQ_INDEX,
548  .flags = VFIO_IRQ_SET_DATA_NONE | VFIO_IRQ_SET_ACTION_UNMASK,
549  };
550 
551  if (ioctl (d->vfio_device_fd, VFIO_DEVICE_SET_IRQS, &i) < 0)
552  {
553  err = clib_error_return_unix (0, "ioctl(VFIO_DEVICE_SET_IRQS) '%U'",
555  }
556  return err;
557 }
558 
559 static clib_error_t *
561 {
562  u32 error_index = (u32) uf->private_data;
563 
564  return clib_error_return (0, "pci device %d: error", error_index);
565 }
566 
567 static clib_error_t *
569 {
570  int __attribute__ ((unused)) rv;
573 
574  u64 icount;
575  rv = read (uf->file_descriptor, &icount, sizeof (icount));
576 
577  if (p->interrupt_handler)
578  p->interrupt_handler (h);
579 
581 
582  return /* no error */ 0;
583 }
584 
585 static clib_error_t *
587 {
588  u32 error_index = (u32) uf->private_data;
589 
590  return clib_error_return (0, "pci device %d: error", error_index);
591 }
592 
593 static clib_error_t *
596 {
597  clib_error_t *err = 0;
598  clib_file_t t = { 0 };
599  u8 *s = 0;
600 
601  p->addr.as_u32 = di->addr.as_u32;
602  p->uio_fd = -1;
603 
604  s = format (s, "%s/%U/config%c", sysfs_pci_dev_path,
605  format_vlib_pci_addr, &di->addr, 0);
606 
607  p->config_fd = open ((char *) s, O_RDWR);
608  vec_reset_length (s);
609 
610  if (p->config_fd == -1)
611  {
612  err = clib_error_return_unix (0, "open '%s'", s);
613  goto error;
614  }
615 
616  s = format (0, "%s/%U/uio", sysfs_pci_dev_path,
617  format_vlib_pci_addr, &di->addr);
618  foreach_directory_file ((char *) s, scan_uio_dir, p, /* scan_dirs */
619  1);
620  vec_reset_length (s);
621 
622  s = format (s, "/dev/uio%d%c", p->uio_minor, 0);
623  p->uio_fd = open ((char *) s, O_RDWR);
624  if (p->uio_fd < 0)
625  {
626  err = clib_error_return_unix (0, "open '%s'", s);
627  goto error;
628  }
629 
631  t.file_descriptor = p->uio_fd;
633  t.private_data = p->handle;
634 
636  p->interrupt_handler = r->interrupt_handler;
637  err = r->init_function (vm, p->handle);
638 
639 error:
640  free (s);
641  if (err)
642  {
643  if (p->config_fd != -1)
644  close (p->config_fd);
645  if (p->uio_fd != -1)
646  close (p->uio_fd);
647  }
648  return err;
649 }
650 
653 {
655  uword *p;
656 
657  p = hash_get (lpm->iommu_pool_index_by_group, group);
658 
659  return p ? pool_elt_at_index (lpm->iommu_groups, p[0]) : 0;
660 }
661 
662 static clib_error_t *
664 {
667  clib_error_t *err = 0;
668  struct vfio_group_status group_status;
669  u8 *s = 0;
670  int fd;
671 
672  g = get_vfio_iommu_group (group);
673  if (g)
674  {
675  g->refcnt++;
676  return 0;
677  }
678  s = format (s, "/dev/vfio/%u%c", group, 0);
679  fd = open ((char *) s, O_RDWR);
680  if (fd < 0)
681  return clib_error_return_unix (0, "open '%s'", s);
682 
683  group_status.argsz = sizeof (group_status);
684  if (ioctl (fd, VFIO_GROUP_GET_STATUS, &group_status) < 0)
685  {
686  err = clib_error_return_unix (0, "ioctl(VFIO_GROUP_GET_STATUS) '%s'",
687  s);
688  goto error;
689  }
690 
691  if (!(group_status.flags & VFIO_GROUP_FLAGS_VIABLE))
692  {
693  err = clib_error_return (0, "iommu group %d is not viable (not all "
694  "devices in this group bound to vfio-pci)",
695  group);
696  goto error;
697  }
698 
699  if (ioctl (fd, VFIO_GROUP_SET_CONTAINER, &lpm->vfio_container_fd) < 0)
700  {
701  err = clib_error_return_unix (0, "ioctl(VFIO_GROUP_SET_CONTAINER) '%s'",
702  s);
703  goto error;
704  }
705 
706  if (lpm->vfio_iommu_mode == 0)
707  {
708  if (ioctl (lpm->vfio_container_fd, VFIO_SET_IOMMU, VFIO_TYPE1_IOMMU) <
709  0)
710  {
711  err = clib_error_return_unix (0, "ioctl(VFIO_SET_IOMMU) "
712  "'/dev/vfio/vfio'");
713  goto error;
714  }
715  lpm->vfio_iommu_mode = VFIO_TYPE1_IOMMU;
716  }
717 
718 
719  pool_get (lpm->iommu_groups, g);
720  g->fd = fd;
721  g->refcnt = 1;
722  hash_set (lpm->iommu_pool_index_by_group, group, g - lpm->iommu_groups);
723  vec_free (s);
724  return 0;
725 error:
726  close (fd);
727  return err;
728 }
729 
730 static clib_error_t *
733 {
735  struct vfio_device_info device_info;
736  struct vfio_irq_info irq_info = { 0 };
737  clib_error_t *err = 0;
738  u8 *s = 0;
739  int dfd;
740 
741  p->addr.as_u32 = di->addr.as_u32;
742 
743  if (di->driver_name == 0 ||
744  (strcmp ("vfio-pci", (char *) di->driver_name) != 0))
745  return clib_error_return (0, "Device '%U' (iommu group %d) not bound to "
746  "vfio-pci", format_vlib_pci_addr, &di->addr,
747  di->iommu_group);
748 
749  if ((err = open_vfio_iommu_group (di->iommu_group)))
750  return err;
751 
753 
754  s = format (s, "%U%c", format_vlib_pci_addr, &di->addr, 0);
755  if ((dfd = ioctl (g->fd, VFIO_GROUP_GET_DEVICE_FD, (char *) s)) < 0)
756  {
757  err = clib_error_return_unix (0, "ioctl(VFIO_GROUP_GET_DEVICE_FD) '%U'",
758  format_vlib_pci_addr, &di->addr);
759  goto error;
760  }
761  vec_reset_length (s);
762  p->vfio_device_fd = dfd;
763 
764  device_info.argsz = sizeof (device_info);
765  if (ioctl (dfd, VFIO_DEVICE_GET_INFO, &device_info) < 0)
766  {
767  err = clib_error_return_unix (0, "ioctl(VFIO_DEVICE_GET_INFO) '%U'",
768  format_vlib_pci_addr, &di->addr);
769  goto error;
770  }
771 
772  irq_info.argsz = sizeof (struct vfio_irq_info);
773  irq_info.index = VFIO_PCI_INTX_IRQ_INDEX;
774  if (ioctl (dfd, VFIO_DEVICE_GET_IRQ_INFO, &irq_info) < 0)
775  {
776  err = clib_error_return_unix (0, "ioctl(VFIO_DEVICE_GET_IRQ_INFO) "
777  "'%U'", format_vlib_pci_addr, &di->addr);
778  goto error;
779  }
780 
781  /* reset if device supports it */
782  if (device_info.flags & VFIO_DEVICE_FLAGS_RESET)
783  if (ioctl (dfd, VFIO_DEVICE_RESET) < 0)
784  {
785  err = clib_error_return_unix (0, "ioctl(VFIO_DEVICE_RESET) '%U'",
786  format_vlib_pci_addr, &di->addr);
787  goto error;
788  }
789 
790  if ((irq_info.flags & VFIO_IRQ_INFO_EVENTFD) && irq_info.count == 1)
791  {
792  u8 buf[sizeof (struct vfio_irq_set) + sizeof (int)] = { 0 };
793  struct vfio_irq_set *irq_set = (struct vfio_irq_set *) buf;
794  clib_file_t t = { 0 };
795  int efd = eventfd (0, EFD_NONBLOCK);
796 
797  irq_set->argsz = sizeof (buf);
798  irq_set->count = 1;
799  irq_set->index = VFIO_PCI_INTX_IRQ_INDEX;
800  irq_set->flags =
801  VFIO_IRQ_SET_DATA_EVENTFD | VFIO_IRQ_SET_ACTION_TRIGGER;
802  clib_memcpy (&irq_set->data, &efd, sizeof (int));
803 
804  if (ioctl (dfd, VFIO_DEVICE_SET_IRQS, irq_set) < 0)
805  {
806  err = clib_error_return_unix (0, "ioctl(VFIO_DEVICE_SET_IRQS) '%U'",
807  format_vlib_pci_addr, &di->addr);
808  goto error;
809  }
810 
812  t.file_descriptor = efd;
814  t.private_data = p->handle;
816 
817  /* unmask the interrupt */
819  }
820 
821  p->interrupt_handler = r->interrupt_handler;
822 
823  s = format (s, "%s/%U/config%c", sysfs_pci_dev_path,
824  format_vlib_pci_addr, &di->addr, 0);
825 
826  p->config_fd = open ((char *) s, O_RDWR);
827  vec_reset_length (s);
828 
829  if (p->config_fd == -1)
830  {
831  err = clib_error_return_unix (0, "open '%s'", s);
832  goto error;
833  }
834 
835  err = r->init_function (vm, p->handle);
836 
837 error:
838  vec_free (s);
839  if (err)
840  {
841  if (dfd != -1)
842  close (dfd);
843  if (p->config_fd != -1)
844  close (p->config_fd);
845  }
846  return err;
847 }
848 
849 /* Configuration space read/write. */
850 clib_error_t *
852  vlib_read_or_write_t read_or_write,
853  uword address, void *data, u32 n_bytes)
854 {
856  int n;
857 
858  if (read_or_write == VLIB_READ)
859  n = pread (p->config_fd, data, n_bytes, address);
860  else
861  n = pwrite (p->config_fd, data, n_bytes, address);
862 
863  if (n != n_bytes)
864  return clib_error_return_unix (0, "%s",
865  read_or_write == VLIB_READ
866  ? "read" : "write");
867 
868  return 0;
869 }
870 
871 static clib_error_t *
873  u32 resource, u8 * addr, void **result)
874 {
876  struct stat stat_buf;
877  u8 *file_name;
878  int fd;
879  clib_error_t *error;
880  int flags = MAP_SHARED;
881 
882  error = 0;
883 
884  file_name = format (0, "%s/%U/resource%d%c", sysfs_pci_dev_path,
885  format_vlib_pci_addr, &p->addr, resource, 0);
886 
887  fd = open ((char *) file_name, O_RDWR);
888  if (fd < 0)
889  {
890  error = clib_error_return_unix (0, "open `%s'", file_name);
891  goto done;
892  }
893 
894  if (fstat (fd, &stat_buf) < 0)
895  {
896  error = clib_error_return_unix (0, "fstat `%s'", file_name);
897  goto done;
898  }
899 
900  vec_validate (p->resource_fds, resource);
901  p->resource_fds[resource] = fd;
902  if (addr != 0)
903  flags |= MAP_FIXED;
904 
905  *result = mmap (addr,
906  /* size */ stat_buf.st_size,
907  PROT_READ | PROT_WRITE, flags,
908  /* file */ fd,
909  /* offset */ 0);
910  if (*result == (void *) -1)
911  {
912  error = clib_error_return_unix (0, "mmap `%s'", file_name);
913  goto done;
914  }
915 
916 done:
917  if (error)
918  {
919  if (fd >= 0)
920  close (fd);
921  }
922  vec_free (file_name);
923  return error;
924 }
925 
926 clib_error_t *
928 {
929  return (vlib_pci_map_resource_int (h, resource, 0 /* addr */ , result));
930 }
931 
932 clib_error_t *
934  u32 resource, u8 * addr, void **result)
935 {
936  return (vlib_pci_map_resource_int (h, resource, addr, result));
937 }
938 
939 void
941 {
942  vlib_pci_main_t *pm = &pci_main;
946  clib_error_t *err = 0;
948 
949  pool_get (lpm->linux_pci_devices, p);
950  p->handle = p - lpm->linux_pci_devices;
951 
952  r = pm->pci_device_registrations;
953 
954  while (r)
955  {
956  for (i = r->supported_devices; i->vendor_id != 0; i++)
957  if (i->vendor_id == di->config0.header.vendor_id &&
958  i->device_id == di->config0.header.device_id)
959  {
960  if (di->iommu_group != -1)
961  err = add_device_vfio (vm, p, di, r);
962  else
963  err = add_device_uio (vm, p, di, r);
964 
965  if (err)
966  clib_error_report (err);
967  else
968  return;
969  }
970  r = r->next_registration;
971  }
972 
973  /* No driver, close the PCI config-space FD */
974  memset (p, 0, sizeof (linux_pci_device_t));
975  pool_put (lpm->linux_pci_devices, p);
976 }
977 
978 static clib_error_t *
979 scan_pci_addr (void *arg, u8 * dev_dir_name, u8 * ignored)
980 {
981  vlib_pci_addr_t addr, **addrv = arg;
982  unformat_input_t input;
983  clib_error_t *err = 0;
984 
985  unformat_init_string (&input, (char *) dev_dir_name,
986  vec_len (dev_dir_name));
987 
988  if (!unformat (&input, "/sys/bus/pci/devices/%U",
989  unformat_vlib_pci_addr, &addr))
990  err = clib_error_return (0, "unformat error `%v`", dev_dir_name);
991 
992  unformat_free (&input);
993 
994  if (err)
995  return err;
996 
997  vec_add1 (*addrv, addr);
998  return 0;
999 }
1000 
1001 static int
1002 pci_addr_cmp (void *v1, void *v2)
1003 {
1004  vlib_pci_addr_t *a1 = v1;
1005  vlib_pci_addr_t *a2 = v2;
1006 
1007  if (a1->domain > a2->domain)
1008  return 1;
1009  if (a1->domain < a2->domain)
1010  return -1;
1011  if (a1->bus > a2->bus)
1012  return 1;
1013  if (a1->bus < a2->bus)
1014  return -1;
1015  if (a1->slot > a2->slot)
1016  return 1;
1017  if (a1->slot < a2->slot)
1018  return -1;
1019  if (a1->function > a2->function)
1020  return 1;
1021  if (a1->function < a2->function)
1022  return -1;
1023  return 0;
1024 }
1025 
1026 vlib_pci_addr_t *
1028 {
1029  vlib_pci_addr_t *addrs = 0;
1030  clib_error_t *err;
1032  &addrs, /* scan_dirs */ 0);
1033  if (err)
1034  {
1035  vec_free (addrs);
1036  return 0;
1037  }
1038 
1040 
1041  return addrs;
1042 }
1043 
1044 clib_error_t *
1046 {
1047  vlib_pci_main_t *pm = &pci_main;
1049  vlib_pci_addr_t *addr = 0, *addrs;
1050  clib_error_t *error;
1051  int fd;
1052 
1053  pm->vlib_main = vm;
1054 
1055  if ((error = vlib_call_init_function (vm, unix_input_init)))
1056  return error;
1057 
1058  ASSERT (sizeof (vlib_pci_addr_t) == sizeof (u32));
1059 
1060  fd = open ("/dev/vfio/vfio", O_RDWR);
1061 
1062  if ((fd != -1) && (ioctl (fd, VFIO_GET_API_VERSION) != VFIO_API_VERSION))
1063  {
1064  close (fd);
1065  fd = -1;
1066  }
1067 
1068  if ((fd != -1) && (ioctl (fd, VFIO_CHECK_EXTENSION, VFIO_TYPE1_IOMMU) == 0))
1069  {
1070  close (fd);
1071  fd = -1;
1072  }
1073 
1074  lpm->vfio_container_fd = fd;
1075  lpm->iommu_pool_index_by_group = hash_create (0, sizeof (uword));
1076 
1077  addrs = vlib_pci_get_all_dev_addrs ();
1078  /* *INDENT-OFF* */
1079  vec_foreach (addr, addrs)
1080  {
1082  if ((d = vlib_pci_get_device_info (addr, 0)))
1083  {
1086  }
1087  }
1088  /* *INDENT-ON* */
1089 
1090  return error;
1091 }
1092 
1094 
1095 /*
1096  * fd.io coding-style-patch-verification: ON
1097  *
1098  * Local Variables:
1099  * eval: (c-set-style "gnu")
1100  * End:
1101  */
#define vec_validate(V, I)
Make sure vector is long enough for given index (no header, unspecified alignment) ...
Definition: vec.h:432
u8 * format_vlib_pci_addr(u8 *s, va_list *va)
Definition: pci.c:140
linux_pci_vfio_iommu_group_t * iommu_groups
Definition: pci.c:111
#define hash_set(h, key, value)
Definition: hash.h:254
sll srl srl sll sra u16x4 i
Definition: vector_sse2.h:337
static clib_error_t * unix_input_init(vlib_main_t *vm)
Definition: input.c:259
uword unformat_vlib_pci_addr(unformat_input_t *input, va_list *args)
Definition: pci.c:123
pci_device_registration_t * pci_device_registrations
Definition: pci.h:140
vlib_read_or_write_t
Definition: defs.h:54
vlib_pci_device_info_t * vlib_pci_get_device_info(vlib_pci_addr_t *addr, clib_error_t **error)
Definition: pci.c:156
vlib_pci_addr_t * vlib_pci_get_addr(vlib_pci_dev_handle_t h)
Definition: pci.c:142
u32 file_descriptor
Definition: file.h:53
#define vec_add1(V, E)
Add 1 element to end of vector (unspecified alignment).
Definition: vec.h:518
u8 * format(u8 *s, const char *fmt,...)
Definition: format.c:419
static clib_error_t * linux_pci_uio_error_ready(clib_file_t *uf)
Definition: pci.c:560
u32 clib_file_index
Definition: pci.c:83
vlib_pci_addr_t addr
Definition: pci.c:65
vlib_pci_addr_t * vlib_pci_get_all_dev_addrs()
Definition: pci.c:1027
#define pool_get(P, E)
Allocate an object E from a pool P (unspecified alignment).
Definition: pool.h:225
clib_error_t * clib_sysfs_write(char *file_name, char *fmt,...)
Definition: sysfs.c:26
static char * sysfs_mod_vfio_noiommu
Definition: pci.c:59
vlib_pci_main_t pci_main
Definition: pci.c:53
#define vec_reset_length(v)
Reset vector length to zero NULL-pointer tolerant.
clib_file_function_t * read_function
Definition: file.h:63
int * resource_fds
Definition: pci.c:68
linux_pci_main_t linux_pci_main
Definition: pci.c:153
uword flags
Definition: clib_error.h:29
static void pci_config_type1_little_to_host(pci_config_type1_regs_t *r)
Definition: pci_config.h:353
#define VLIB_INIT_FUNCTION(x)
Definition: init.h:111
void di(unformat_input_t *i)
Definition: unformat.c:163
clib_error_t * vlib_pci_bind_to_uio(vlib_pci_addr_t *addr, char *uio_drv_name)
Definition: pci.c:330
#define clib_error_return(e, args...)
Definition: error.h:99
clib_file_main_t file_main
Definition: main.c:63
struct _pci_device_registration pci_device_registration_t
unsigned long u64
Definition: types.h:89
vlib_main_t * vlib_main
Definition: pci.c:103
#define vlib_call_init_function(vm, x)
Definition: init.h:162
pci_config_type1_regs_t config1
Definition: pci.h:85
vlib_pci_dev_handle_t handle
Definition: pci.c:64
void unformat_init_string(unformat_input_t *input, char *string, int string_len)
Definition: unformat.c:1023
uword private_data
Definition: pci.c:89
u8 config_data[256]
Definition: pci.h:86
#define hash_get(h, key)
Definition: hash.h:248
#define pool_elt_at_index(p, i)
Returns pointer to element at given index.
Definition: pool.h:459
int config_fd
Definition: pci.c:71
clib_error_t * linux_pci_init(vlib_main_t *vm)
Definition: pci.c:1045
static clib_error_t * scan_pci_addr(void *arg, u8 *dev_dir_name, u8 *ignored)
Definition: pci.c:979
struct _unformat_input_t unformat_input_t
u32 vlib_pci_dev_handle_t
Definition: pci.h:94
#define clib_error_return_unix(e, args...)
Definition: error.h:102
#define pool_put(P, E)
Free an object E in pool P.
Definition: pool.h:271
static const char * sysfs_pci_drv_path
Definition: pci.c:58
static clib_error_t * linux_pci_vfio_error_ready(clib_file_t *uf)
Definition: pci.c:586
clib_error_t * clib_sysfs_read(char *file_name, char *fmt,...)
Definition: sysfs.c:50
void(* interrupt_handler)(vlib_pci_dev_handle_t h)
Definition: pci.c:86
static clib_error_t * linux_pci_uio_read_ready(clib_file_t *uf)
Definition: pci.c:522
static void pci_config_type0_little_to_host(pci_config_type0_regs_t *r)
Definition: pci_config.h:279
static clib_error_t * add_device_vfio(vlib_main_t *vm, linux_pci_device_t *p, vlib_pci_device_info_t *di, pci_device_registration_t *r)
Definition: pci.c:731
static clib_error_t * open_vfio_iommu_group(int group)
Definition: pci.c:663
vlib_main_t * vm
Definition: buffer.c:283
vec_header_t h
Definition: buffer.c:282
#define vec_free(V)
Free vector&#39;s memory (no header).
Definition: vec.h:336
u8 * driver_name
Definition: pci.h:79
clib_error_t * pci_bus_init(vlib_main_t *vm)
Definition: pci.c:251
#define clib_memcpy(a, b, c)
Definition: string.h:75
#define hash_create(elts, value_bytes)
Definition: hash.h:681
linux_pci_device_t * linux_pci_get_device(vlib_pci_dev_handle_t h)
Definition: pci.c:121
#define ASSERT(truth)
unsigned int u32
Definition: types.h:88
static uword clib_file_add(clib_file_main_t *um, clib_file_t *template)
Definition: file.h:84
pci_config_type0_regs_t config0
Definition: pci.h:84
#define clib_error_report(e)
Definition: error.h:113
vlib_pci_addr_t addr
Definition: pci.h:63
static void * clib_mem_alloc(uword size)
Definition: mem.h:112
u64 uword
Definition: types.h:112
Definition: defs.h:56
static int pci_addr_cmp(void *v1, void *v2)
Definition: pci.c:1002
clib_error_t * foreach_directory_file(char *dir_name, clib_error_t *(*f)(void *arg, u8 *path_name, u8 *file_name), void *arg, int scan_dirs)
Definition: util.c:49
#define vec_len(v)
Number of elements in vector (rvalue-only, NULL tolerant)
unsigned char u8
Definition: types.h:56
static int directory_exists(char *path)
Definition: pci.c:320
uword vlib_pci_get_private_data(vlib_pci_dev_handle_t h)
Definition: pci.c:128
#define vec_sort_with_function(vec, f)
Sort a vector using the supplied element comparison function.
Definition: vec.h:956
static void unformat_free(unformat_input_t *i)
Definition: format.h:161
#define clib_unix_warning(format, args...)
Definition: error.h:68
void vlib_pci_set_private_data(vlib_pci_dev_handle_t h, uword private_data)
Definition: pci.c:135
static clib_error_t * linux_pci_vfio_unmask_intx(linux_pci_device_t *d)
Definition: pci.c:540
#define clib_error_free(e)
Definition: error.h:86
clib_file_function_t * error_function
Definition: file.h:63
u16 device_id
Definition: pci.h:117
int vfio_device_fd
Definition: pci.c:80
static linux_pci_vfio_iommu_group_t * get_vfio_iommu_group(int group)
Definition: pci.c:652
static clib_error_t * vlib_pci_intr_enable(vlib_pci_dev_handle_t h)
Definition: pci.h:197
static clib_error_t * add_device_uio(vlib_main_t *vm, linux_pci_device_t *p, vlib_pci_device_info_t *di, pci_device_registration_t *r)
Definition: pci.c:594
#define vec_foreach(var, vec)
Vector iterator.
uword private_data
Definition: file.h:60
Definition: file.h:50
int vfio_container_fd
Definition: pci.c:107
vhost_vring_addr_t addr
Definition: vhost-user.h:83
static const char * sysfs_pci_dev_path
Definition: pci.c:57
static clib_error_t * linux_pci_vfio_read_ready(clib_file_t *uf)
Definition: pci.c:568
static void vlib_pci_free_device_info(vlib_pci_device_info_t *di)
Definition: pci.h:104
void init_device_from_registered(vlib_main_t *vm, vlib_pci_device_info_t *di)
Definition: pci.c:940
u32 flags
Definition: vhost-user.h:77
int vfio_iommu_mode
Definition: pci.c:108
clib_error_t * vlib_pci_map_resource_fixed(vlib_pci_dev_handle_t h, u32 resource, u8 *addr, void **result)
Definition: pci.c:933
u16 device_class
Definition: pci.h:69
linux_pci_device_t * linux_pci_devices
Definition: pci.c:104
vlib_main_t * vlib_main
Definition: pci.h:139
pci_config_header_t header
Definition: pci_config.h:247
static clib_error_t * scan_uio_dir(void *arg, u8 *path_name, u8 *file_name)
Definition: pci.c:507
u8 * product_name
Definition: pci.h:74
u16 vendor_id
Definition: pci.h:117
uword unformat(unformat_input_t *i, const char *fmt,...)
Definition: unformat.c:972
u32 uio_minor
Definition: pci.c:77
static clib_error_t * vlib_pci_map_resource_int(vlib_pci_dev_handle_t h, u32 resource, u8 *addr, void **result)
Definition: pci.c:872
u8 * clib_sysfs_link_to_name(char *link)
Definition: sysfs.c:90
uword * iommu_pool_index_by_group
Definition: pci.c:114
clib_error_t * vlib_pci_map_resource(vlib_pci_dev_handle_t h, u32 resource, void **result)
Definition: pci.c:927
clib_error_t * vlib_pci_read_write_config(vlib_pci_dev_handle_t h, vlib_read_or_write_t read_or_write, uword address, void *data, u32 n_bytes)
Definition: pci.c:851