]> git.proxmox.com Git - mirror_ovs.git/blob - lib/dpif-linux.c
Merge citrix branch into master.
[mirror_ovs.git] / lib / dpif-linux.c
1 /*
2 * Copyright (c) 2008, 2009 Nicira Networks.
3 *
4 * Licensed under the Apache License, Version 2.0 (the "License");
5 * you may not use this file except in compliance with the License.
6 * You may obtain a copy of the License at:
7 *
8 * http://www.apache.org/licenses/LICENSE-2.0
9 *
10 * Unless required by applicable law or agreed to in writing, software
11 * distributed under the License is distributed on an "AS IS" BASIS,
12 * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
13 * See the License for the specific language governing permissions and
14 * limitations under the License.
15 */
16
17 #include <config.h>
18 #include "dpif.h"
19
20 #include <assert.h>
21 #include <ctype.h>
22 #include <errno.h>
23 #include <fcntl.h>
24 #include <inttypes.h>
25 #include <net/if.h>
26 #include <linux/ethtool.h>
27 #include <linux/rtnetlink.h>
28 #include <linux/sockios.h>
29 #include <stdlib.h>
30 #include <sys/ioctl.h>
31 #include <unistd.h>
32
33 #include "dpif-provider.h"
34 #include "ofpbuf.h"
35 #include "poll-loop.h"
36 #include "rtnetlink.h"
37 #include "svec.h"
38 #include "util.h"
39
40 #include "vlog.h"
41 #define THIS_MODULE VLM_dpif_linux
42
43 /* Datapath interface for the openvswitch Linux kernel module. */
44 struct dpif_linux {
45 struct dpif dpif;
46 int fd;
47
48 /* Used by dpif_linux_get_all_names(). */
49 char *local_ifname;
50 int minor;
51
52 /* Change notification. */
53 int local_ifindex; /* Ifindex of local port. */
54 struct svec changed_ports; /* Ports that have changed. */
55 struct rtnetlink_notifier port_notifier;
56 bool change_error;
57 };
58
59 static struct vlog_rate_limit error_rl = VLOG_RATE_LIMIT_INIT(9999, 5);
60
61 static int do_ioctl(const struct dpif *, int cmd, const void *arg);
62 static int lookup_minor(const char *name, int *minor);
63 static int finish_open(struct dpif *, const char *local_ifname);
64 static int get_openvswitch_major(void);
65 static int create_minor(const char *name, int minor, struct dpif **dpifp);
66 static int open_minor(int minor, struct dpif **dpifp);
67 static int make_openvswitch_device(int minor, char **fnp);
68 static void dpif_linux_port_changed(const struct rtnetlink_change *,
69 void *dpif);
70
71 static struct dpif_linux *
72 dpif_linux_cast(const struct dpif *dpif)
73 {
74 dpif_assert_class(dpif, &dpif_linux_class);
75 return CONTAINER_OF(dpif, struct dpif_linux, dpif);
76 }
77
78 static int
79 dpif_linux_enumerate(struct svec *all_dps)
80 {
81 int major;
82 int error;
83 int i;
84
85 /* Check that the Open vSwitch module is loaded. */
86 major = get_openvswitch_major();
87 if (major < 0) {
88 return -major;
89 }
90
91 error = 0;
92 for (i = 0; i < ODP_MAX; i++) {
93 struct dpif *dpif;
94 char devname[16];
95 int retval;
96
97 sprintf(devname, "dp%d", i);
98 retval = dpif_open(devname, &dpif);
99 if (!retval) {
100 svec_add(all_dps, devname);
101 dpif_close(dpif);
102 } else if (retval != ENODEV && !error) {
103 error = retval;
104 }
105 }
106 return error;
107 }
108
109 static int
110 dpif_linux_open(const char *name UNUSED, char *suffix, bool create,
111 struct dpif **dpifp)
112 {
113 int minor;
114
115 minor = !strncmp(name, "dp", 2)
116 && isdigit((unsigned char)name[2]) ? atoi(name + 2) : -1;
117 if (create) {
118 if (minor >= 0) {
119 return create_minor(suffix, minor, dpifp);
120 } else {
121 /* Scan for unused minor number. */
122 for (minor = 0; minor < ODP_MAX; minor++) {
123 int error = create_minor(suffix, minor, dpifp);
124 if (error != EBUSY) {
125 return error;
126 }
127 }
128
129 /* All datapath numbers in use. */
130 return ENOBUFS;
131 }
132 } else {
133 struct dpif_linux *dpif;
134 struct odp_port port;
135 int error;
136
137 if (minor < 0) {
138 error = lookup_minor(suffix, &minor);
139 if (error) {
140 return error;
141 }
142 }
143
144 error = open_minor(minor, dpifp);
145 if (error) {
146 return error;
147 }
148 dpif = dpif_linux_cast(*dpifp);
149
150 /* We need the local port's ifindex for the poll function. Start by
151 * getting the local port's name. */
152 memset(&port, 0, sizeof port);
153 port.port = ODPP_LOCAL;
154 if (ioctl(dpif->fd, ODP_PORT_QUERY, &port)) {
155 error = errno;
156 if (error != ENODEV) {
157 VLOG_WARN("%s: probe returned unexpected error: %s",
158 dpif_name(*dpifp), strerror(error));
159 }
160 dpif_close(*dpifp);
161 return error;
162 }
163
164 /* Then use that to finish up opening. */
165 return finish_open(&dpif->dpif, port.devname);
166 }
167 }
168
169 static void
170 dpif_linux_close(struct dpif *dpif_)
171 {
172 struct dpif_linux *dpif = dpif_linux_cast(dpif_);
173 rtnetlink_notifier_unregister(&dpif->port_notifier);
174 svec_destroy(&dpif->changed_ports);
175 free(dpif->local_ifname);
176 close(dpif->fd);
177 free(dpif);
178 }
179
180 static int
181 dpif_linux_get_all_names(const struct dpif *dpif_, struct svec *all_names)
182 {
183 struct dpif_linux *dpif = dpif_linux_cast(dpif_);
184
185 svec_add_nocopy(all_names, xasprintf("dp%d", dpif->minor));
186 svec_add(all_names, dpif->local_ifname);
187 return 0;
188 }
189
190 static int
191 dpif_linux_delete(struct dpif *dpif_)
192 {
193 return do_ioctl(dpif_, ODP_DP_DESTROY, NULL);
194 }
195
196 static int
197 dpif_linux_get_stats(const struct dpif *dpif_, struct odp_stats *stats)
198 {
199 return do_ioctl(dpif_, ODP_DP_STATS, stats);
200 }
201
202 static int
203 dpif_linux_get_drop_frags(const struct dpif *dpif_, bool *drop_fragsp)
204 {
205 int drop_frags;
206 int error;
207
208 error = do_ioctl(dpif_, ODP_GET_DROP_FRAGS, &drop_frags);
209 if (!error) {
210 *drop_fragsp = drop_frags & 1;
211 }
212 return error;
213 }
214
215 static int
216 dpif_linux_set_drop_frags(struct dpif *dpif_, bool drop_frags)
217 {
218 int drop_frags_int = drop_frags;
219 return do_ioctl(dpif_, ODP_SET_DROP_FRAGS, &drop_frags_int);
220 }
221
222 static int
223 dpif_linux_port_add(struct dpif *dpif_, const char *devname, uint16_t flags,
224 uint16_t *port_no)
225 {
226 struct odp_port port;
227 int error;
228
229 memset(&port, 0, sizeof port);
230 strncpy(port.devname, devname, sizeof port.devname);
231 port.flags = flags;
232 error = do_ioctl(dpif_, ODP_PORT_ADD, &port);
233 if (!error) {
234 *port_no = port.port;
235 }
236 return error;
237 }
238
239 static int
240 dpif_linux_port_del(struct dpif *dpif_, uint16_t port_no)
241 {
242 int tmp = port_no;
243 return do_ioctl(dpif_, ODP_PORT_DEL, &tmp);
244 }
245
246 static int
247 dpif_linux_port_query_by_number(const struct dpif *dpif_, uint16_t port_no,
248 struct odp_port *port)
249 {
250 memset(port, 0, sizeof *port);
251 port->port = port_no;
252 return do_ioctl(dpif_, ODP_PORT_QUERY, port);
253 }
254
255 static int
256 dpif_linux_port_query_by_name(const struct dpif *dpif_, const char *devname,
257 struct odp_port *port)
258 {
259 memset(port, 0, sizeof *port);
260 strncpy(port->devname, devname, sizeof port->devname);
261 return do_ioctl(dpif_, ODP_PORT_QUERY, port);
262 }
263
264 static int
265 dpif_linux_flow_flush(struct dpif *dpif_)
266 {
267 return do_ioctl(dpif_, ODP_FLOW_FLUSH, NULL);
268 }
269
270 static int
271 dpif_linux_port_list(const struct dpif *dpif_, struct odp_port *ports, int n)
272 {
273 struct odp_portvec pv;
274 int error;
275
276 pv.ports = ports;
277 pv.n_ports = n;
278 error = do_ioctl(dpif_, ODP_PORT_LIST, &pv);
279 return error ? -error : pv.n_ports;
280 }
281
282 static int
283 dpif_linux_port_poll(const struct dpif *dpif_, char **devnamep)
284 {
285 struct dpif_linux *dpif = dpif_linux_cast(dpif_);
286
287 if (dpif->change_error) {
288 dpif->change_error = false;
289 svec_clear(&dpif->changed_ports);
290 return ENOBUFS;
291 } else if (dpif->changed_ports.n) {
292 *devnamep = dpif->changed_ports.names[--dpif->changed_ports.n];
293 return 0;
294 } else {
295 return EAGAIN;
296 }
297 }
298
299 static void
300 dpif_linux_port_poll_wait(const struct dpif *dpif_)
301 {
302 struct dpif_linux *dpif = dpif_linux_cast(dpif_);
303 if (dpif->changed_ports.n || dpif->change_error) {
304 poll_immediate_wake();
305 } else {
306 rtnetlink_notifier_wait();
307 }
308 }
309
310 static int
311 dpif_linux_port_group_get(const struct dpif *dpif_, int group,
312 uint16_t ports[], int n)
313 {
314 struct odp_port_group pg;
315 int error;
316
317 assert(n <= UINT16_MAX);
318 pg.group = group;
319 pg.ports = ports;
320 pg.n_ports = n;
321 error = do_ioctl(dpif_, ODP_PORT_GROUP_GET, &pg);
322 return error ? -error : pg.n_ports;
323 }
324
325 static int
326 dpif_linux_port_group_set(struct dpif *dpif_, int group,
327 const uint16_t ports[], int n)
328 {
329 struct odp_port_group pg;
330
331 assert(n <= UINT16_MAX);
332 pg.group = group;
333 pg.ports = (uint16_t *) ports;
334 pg.n_ports = n;
335 return do_ioctl(dpif_, ODP_PORT_GROUP_SET, &pg);
336 }
337
338 static int
339 dpif_linux_flow_get(const struct dpif *dpif_, struct odp_flow flows[], int n)
340 {
341 struct odp_flowvec fv;
342 fv.flows = flows;
343 fv.n_flows = n;
344 return do_ioctl(dpif_, ODP_FLOW_GET, &fv);
345 }
346
347 static int
348 dpif_linux_flow_put(struct dpif *dpif_, struct odp_flow_put *put)
349 {
350 return do_ioctl(dpif_, ODP_FLOW_PUT, put);
351 }
352
353 static int
354 dpif_linux_flow_del(struct dpif *dpif_, struct odp_flow *flow)
355 {
356 return do_ioctl(dpif_, ODP_FLOW_DEL, flow);
357 }
358
359 static int
360 dpif_linux_flow_list(const struct dpif *dpif_, struct odp_flow flows[], int n)
361 {
362 struct odp_flowvec fv;
363 int error;
364
365 fv.flows = flows;
366 fv.n_flows = n;
367 error = do_ioctl(dpif_, ODP_FLOW_LIST, &fv);
368 return error ? -error : fv.n_flows;
369 }
370
371 static int
372 dpif_linux_execute(struct dpif *dpif_, uint16_t in_port,
373 const union odp_action actions[], int n_actions,
374 const struct ofpbuf *buf)
375 {
376 struct odp_execute execute;
377 memset(&execute, 0, sizeof execute);
378 execute.in_port = in_port;
379 execute.actions = (union odp_action *) actions;
380 execute.n_actions = n_actions;
381 execute.data = buf->data;
382 execute.length = buf->size;
383 return do_ioctl(dpif_, ODP_EXECUTE, &execute);
384 }
385
386 static int
387 dpif_linux_recv_get_mask(const struct dpif *dpif_, int *listen_mask)
388 {
389 return do_ioctl(dpif_, ODP_GET_LISTEN_MASK, listen_mask);
390 }
391
392 static int
393 dpif_linux_recv_set_mask(struct dpif *dpif_, int listen_mask)
394 {
395 return do_ioctl(dpif_, ODP_SET_LISTEN_MASK, &listen_mask);
396 }
397
398 static int
399 dpif_linux_recv(struct dpif *dpif_, struct ofpbuf **bufp)
400 {
401 struct dpif_linux *dpif = dpif_linux_cast(dpif_);
402 struct ofpbuf *buf;
403 int retval;
404 int error;
405
406 buf = ofpbuf_new(65536);
407 retval = read(dpif->fd, ofpbuf_tail(buf), ofpbuf_tailroom(buf));
408 if (retval < 0) {
409 error = errno;
410 if (error != EAGAIN) {
411 VLOG_WARN_RL(&error_rl, "%s: read failed: %s",
412 dpif_name(dpif_), strerror(error));
413 }
414 } else if (retval >= sizeof(struct odp_msg)) {
415 struct odp_msg *msg = buf->data;
416 if (msg->length <= retval) {
417 buf->size += retval;
418 *bufp = buf;
419 return 0;
420 } else {
421 VLOG_WARN_RL(&error_rl, "%s: discarding message truncated "
422 "from %"PRIu32" bytes to %d",
423 dpif_name(dpif_), msg->length, retval);
424 error = ERANGE;
425 }
426 } else if (!retval) {
427 VLOG_WARN_RL(&error_rl, "%s: unexpected end of file", dpif_name(dpif_));
428 error = EPROTO;
429 } else {
430 VLOG_WARN_RL(&error_rl,
431 "%s: discarding too-short message (%d bytes)",
432 dpif_name(dpif_), retval);
433 error = ERANGE;
434 }
435
436 *bufp = NULL;
437 ofpbuf_delete(buf);
438 return error;
439 }
440
441 static void
442 dpif_linux_recv_wait(struct dpif *dpif_)
443 {
444 struct dpif_linux *dpif = dpif_linux_cast(dpif_);
445 poll_fd_wait(dpif->fd, POLLIN);
446 }
447
448 const struct dpif_class dpif_linux_class = {
449 "", /* This is the default class. */
450 "linux",
451 NULL,
452 NULL,
453 dpif_linux_enumerate,
454 dpif_linux_open,
455 dpif_linux_close,
456 dpif_linux_get_all_names,
457 dpif_linux_delete,
458 dpif_linux_get_stats,
459 dpif_linux_get_drop_frags,
460 dpif_linux_set_drop_frags,
461 dpif_linux_port_add,
462 dpif_linux_port_del,
463 dpif_linux_port_query_by_number,
464 dpif_linux_port_query_by_name,
465 dpif_linux_port_list,
466 dpif_linux_port_poll,
467 dpif_linux_port_poll_wait,
468 dpif_linux_port_group_get,
469 dpif_linux_port_group_set,
470 dpif_linux_flow_get,
471 dpif_linux_flow_put,
472 dpif_linux_flow_del,
473 dpif_linux_flow_flush,
474 dpif_linux_flow_list,
475 dpif_linux_execute,
476 dpif_linux_recv_get_mask,
477 dpif_linux_recv_set_mask,
478 dpif_linux_recv,
479 dpif_linux_recv_wait,
480 };
481 \f
482 static int get_openvswitch_major(void);
483 static int get_major(const char *target);
484
485 static int
486 do_ioctl(const struct dpif *dpif_, int cmd, const void *arg)
487 {
488 struct dpif_linux *dpif = dpif_linux_cast(dpif_);
489 return ioctl(dpif->fd, cmd, arg) ? errno : 0;
490 }
491
492 static int
493 lookup_minor(const char *name, int *minorp)
494 {
495 struct ethtool_drvinfo drvinfo;
496 int minor, port_no;
497 struct ifreq ifr;
498 int error;
499 int sock;
500
501 sock = socket(AF_INET, SOCK_DGRAM, 0);
502 if (sock < 0) {
503 VLOG_WARN("socket(AF_INET) failed: %s", strerror(errno));
504 error = errno;
505 goto error;
506 }
507
508 memset(&ifr, 0, sizeof ifr);
509 strncpy(ifr.ifr_name, name, sizeof ifr.ifr_name);
510 ifr.ifr_data = (caddr_t) &drvinfo;
511
512 memset(&drvinfo, 0, sizeof drvinfo);
513 drvinfo.cmd = ETHTOOL_GDRVINFO;
514 if (ioctl(sock, SIOCETHTOOL, &ifr)) {
515 VLOG_WARN("ioctl(SIOCETHTOOL) failed: %s", strerror(errno));
516 error = errno;
517 goto error_close_sock;
518 }
519
520 if (strcmp(drvinfo.driver, "openvswitch")) {
521 VLOG_WARN("%s is not an openvswitch device", name);
522 error = EOPNOTSUPP;
523 goto error_close_sock;
524 }
525
526 if (sscanf(drvinfo.bus_info, "%d.%d", &minor, &port_no) != 2) {
527 VLOG_WARN("%s ethtool bus_info has unexpected format", name);
528 error = EPROTOTYPE;
529 goto error_close_sock;
530 } else if (port_no != ODPP_LOCAL) {
531 /* This is an Open vSwitch device but not the local port. We
532 * intentionally support only using the name of the local port as the
533 * name of a datapath; otherwise, it would be too difficult to
534 * enumerate all the names of a datapath. */
535 error = EOPNOTSUPP;
536 goto error_close_sock;
537 }
538
539 *minorp = minor;
540 close(sock);
541 return 0;
542
543 error_close_sock:
544 close(sock);
545 error:
546 return error;
547 }
548
549 static int
550 make_openvswitch_device(int minor, char **fnp)
551 {
552 const char dirname[] = "/dev/net";
553 int major;
554 dev_t dev;
555 struct stat s;
556 char fn[128];
557
558 major = get_openvswitch_major();
559 if (major < 0) {
560 return -major;
561 }
562 dev = makedev(major, minor);
563
564 *fnp = NULL;
565 sprintf(fn, "%s/dp%d", dirname, minor);
566 if (!stat(fn, &s)) {
567 if (!S_ISCHR(s.st_mode)) {
568 VLOG_WARN_RL(&error_rl, "%s is not a character device, fixing",
569 fn);
570 } else if (s.st_rdev != dev) {
571 VLOG_WARN_RL(&error_rl,
572 "%s is device %u:%u but should be %u:%u, fixing",
573 fn, major(s.st_rdev), minor(s.st_rdev),
574 major(dev), minor(dev));
575 } else {
576 goto success;
577 }
578 if (unlink(fn)) {
579 VLOG_WARN_RL(&error_rl, "%s: unlink failed (%s)",
580 fn, strerror(errno));
581 return errno;
582 }
583 } else if (errno == ENOENT) {
584 if (stat(dirname, &s)) {
585 if (errno == ENOENT) {
586 if (mkdir(dirname, 0755)) {
587 VLOG_WARN_RL(&error_rl, "%s: mkdir failed (%s)",
588 dirname, strerror(errno));
589 return errno;
590 }
591 } else {
592 VLOG_WARN_RL(&error_rl, "%s: stat failed (%s)",
593 dirname, strerror(errno));
594 return errno;
595 }
596 }
597 } else {
598 VLOG_WARN_RL(&error_rl, "%s: stat failed (%s)", fn, strerror(errno));
599 return errno;
600 }
601
602 /* The device needs to be created. */
603 if (mknod(fn, S_IFCHR | 0700, dev)) {
604 VLOG_WARN_RL(&error_rl,
605 "%s: creating character device %u:%u failed (%s)",
606 fn, major(dev), minor(dev), strerror(errno));
607 return errno;
608 }
609
610 success:
611 *fnp = xstrdup(fn);
612 return 0;
613 }
614
615 /* Return the major device number of the Open vSwitch device. If it
616 * cannot be determined, a negative errno is returned. */
617 static int
618 get_openvswitch_major(void)
619 {
620 static int openvswitch_major = -1;
621 if (openvswitch_major < 0) {
622 openvswitch_major = get_major("openvswitch");
623 }
624 return openvswitch_major;
625 }
626
627 static int
628 get_major(const char *target)
629 {
630 const char fn[] = "/proc/devices";
631 char line[128];
632 FILE *file;
633 int ln;
634
635 file = fopen(fn, "r");
636 if (!file) {
637 VLOG_ERR("opening %s failed (%s)", fn, strerror(errno));
638 return -errno;
639 }
640
641 for (ln = 1; fgets(line, sizeof line, file); ln++) {
642 char name[64];
643 int major;
644
645 if (!strncmp(line, "Character", 9) || line[0] == '\0') {
646 /* Nothing to do. */
647 } else if (!strncmp(line, "Block", 5)) {
648 /* We only want character devices, so skip the rest of the file. */
649 break;
650 } else if (sscanf(line, "%d %63s", &major, name)) {
651 if (!strcmp(name, target)) {
652 fclose(file);
653 return major;
654 }
655 } else {
656 static bool warned;
657 if (!warned) {
658 VLOG_WARN("%s:%d: syntax error", fn, ln);
659 }
660 warned = true;
661 }
662 }
663
664 VLOG_ERR("%s: %s major not found (is the module loaded?)", fn, target);
665 return -ENODEV;
666 }
667
668 static int
669 finish_open(struct dpif *dpif_, const char *local_ifname)
670 {
671 struct dpif_linux *dpif = dpif_linux_cast(dpif_);
672 dpif->local_ifname = strdup(local_ifname);
673 dpif->local_ifindex = if_nametoindex(local_ifname);
674 if (!dpif->local_ifindex) {
675 int error = errno;
676 dpif_close(dpif_);
677 VLOG_WARN("could not get ifindex of %s device: %s",
678 local_ifname, strerror(errno));
679 return error;
680 }
681 return 0;
682 }
683
684 static int
685 create_minor(const char *name, int minor, struct dpif **dpifp)
686 {
687 int error = open_minor(minor, dpifp);
688 if (!error) {
689 error = do_ioctl(*dpifp, ODP_DP_CREATE, name);
690 if (!error) {
691 error = finish_open(*dpifp, name);
692 } else {
693 dpif_close(*dpifp);
694 }
695 }
696 return error;
697 }
698
699 static int
700 open_minor(int minor, struct dpif **dpifp)
701 {
702 int error;
703 char *fn;
704 int fd;
705
706 error = make_openvswitch_device(minor, &fn);
707 if (error) {
708 return error;
709 }
710
711 fd = open(fn, O_RDONLY | O_NONBLOCK);
712 if (fd >= 0) {
713 struct dpif_linux *dpif = xmalloc(sizeof *dpif);
714 error = rtnetlink_notifier_register(&dpif->port_notifier,
715 dpif_linux_port_changed, dpif);
716 if (!error) {
717 char *name;
718
719 name = xasprintf("dp%d", minor);
720 dpif_init(&dpif->dpif, &dpif_linux_class, name, minor, minor);
721 free(name);
722
723 dpif->fd = fd;
724 dpif->local_ifname = NULL;
725 dpif->minor = minor;
726 dpif->local_ifindex = 0;
727 svec_init(&dpif->changed_ports);
728 dpif->change_error = false;
729 *dpifp = &dpif->dpif;
730 } else {
731 free(dpif);
732 }
733 } else {
734 error = errno;
735 VLOG_WARN("%s: open failed (%s)", fn, strerror(error));
736 }
737 free(fn);
738
739 return error;
740 }
741
742 static void
743 dpif_linux_port_changed(const struct rtnetlink_change *change, void *dpif_)
744 {
745 struct dpif_linux *dpif = dpif_;
746
747 if (change) {
748 if (change->master_ifindex == dpif->local_ifindex
749 && (change->nlmsg_type == RTM_NEWLINK
750 || change->nlmsg_type == RTM_DELLINK))
751 {
752 /* Our datapath changed, either adding a new port or deleting an
753 * existing one. */
754 if (!svec_contains(&dpif->changed_ports, change->ifname)) {
755 svec_add(&dpif->changed_ports, change->ifname);
756 svec_sort(&dpif->changed_ports);
757 }
758 }
759 } else {
760 dpif->change_error = true;
761 }
762 }