2 * Copyright (c) 2008, 2009, 2010 Nicira Networks.
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:
8 * http://www.apache.org/licenses/LICENSE-2.0
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.
26 #include <linux/ethtool.h>
27 #include <linux/rtnetlink.h>
28 #include <linux/sockios.h>
30 #include <sys/ioctl.h>
33 #include "dpif-provider.h"
35 #include "poll-loop.h"
36 #include "rtnetlink.h"
41 #define THIS_MODULE VLM_dpif_linux
43 /* Datapath interface for the openvswitch Linux kernel module. */
48 /* Used by dpif_linux_get_all_names(). */
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;
59 static struct vlog_rate_limit error_rl = VLOG_RATE_LIMIT_INIT(9999, 5);
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 *,
71 static struct dpif_linux *
72 dpif_linux_cast(const struct dpif *dpif)
74 dpif_assert_class(dpif, &dpif_linux_class);
75 return CONTAINER_OF(dpif, struct dpif_linux, dpif);
79 dpif_linux_enumerate(struct svec *all_dps)
85 /* Check that the Open vSwitch module is loaded. */
86 major = get_openvswitch_major();
92 for (i = 0; i < ODP_MAX; i++) {
97 sprintf(devname, "dp%d", i);
98 retval = dpif_open(devname, &dpif);
100 svec_add(all_dps, devname);
102 } else if (retval != ENODEV && !error) {
110 dpif_linux_open(const char *name OVS_UNUSED, char *suffix, bool create,
115 minor = !strncmp(name, "dp", 2)
116 && isdigit((unsigned char)name[2]) ? atoi(name + 2) : -1;
119 return create_minor(suffix, minor, dpifp);
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) {
129 /* All datapath numbers in use. */
133 struct dpif_linux *dpif;
134 struct odp_port port;
138 error = lookup_minor(suffix, &minor);
144 error = open_minor(minor, dpifp);
148 dpif = dpif_linux_cast(*dpifp);
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)) {
156 if (error != ENODEV) {
157 VLOG_WARN("%s: probe returned unexpected error: %s",
158 dpif_name(*dpifp), strerror(error));
164 /* Then use that to finish up opening. */
165 return finish_open(&dpif->dpif, port.devname);
170 dpif_linux_close(struct dpif *dpif_)
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);
181 dpif_linux_get_all_names(const struct dpif *dpif_, struct svec *all_names)
183 struct dpif_linux *dpif = dpif_linux_cast(dpif_);
185 svec_add_nocopy(all_names, xasprintf("dp%d", dpif->minor));
186 svec_add(all_names, dpif->local_ifname);
191 dpif_linux_delete(struct dpif *dpif_)
193 return do_ioctl(dpif_, ODP_DP_DESTROY, NULL);
197 dpif_linux_get_stats(const struct dpif *dpif_, struct odp_stats *stats)
199 memset(stats, 0, sizeof *stats);
200 return do_ioctl(dpif_, ODP_DP_STATS, stats);
204 dpif_linux_get_drop_frags(const struct dpif *dpif_, bool *drop_fragsp)
209 error = do_ioctl(dpif_, ODP_GET_DROP_FRAGS, &drop_frags);
211 *drop_fragsp = drop_frags & 1;
217 dpif_linux_set_drop_frags(struct dpif *dpif_, bool drop_frags)
219 int drop_frags_int = drop_frags;
220 return do_ioctl(dpif_, ODP_SET_DROP_FRAGS, &drop_frags_int);
224 dpif_linux_port_add(struct dpif *dpif_, const char *devname, uint16_t flags,
227 struct odp_port port;
230 memset(&port, 0, sizeof port);
231 strncpy(port.devname, devname, sizeof port.devname);
233 error = do_ioctl(dpif_, ODP_PORT_ADD, &port);
235 *port_no = port.port;
241 dpif_linux_port_del(struct dpif *dpif_, uint16_t port_no)
244 return do_ioctl(dpif_, ODP_PORT_DEL, &tmp);
248 dpif_linux_port_query_by_number(const struct dpif *dpif_, uint16_t port_no,
249 struct odp_port *port)
251 memset(port, 0, sizeof *port);
252 port->port = port_no;
253 return do_ioctl(dpif_, ODP_PORT_QUERY, port);
257 dpif_linux_port_query_by_name(const struct dpif *dpif_, const char *devname,
258 struct odp_port *port)
260 memset(port, 0, sizeof *port);
261 strncpy(port->devname, devname, sizeof port->devname);
262 return do_ioctl(dpif_, ODP_PORT_QUERY, port);
266 dpif_linux_flow_flush(struct dpif *dpif_)
268 return do_ioctl(dpif_, ODP_FLOW_FLUSH, NULL);
272 dpif_linux_port_list(const struct dpif *dpif_, struct odp_port *ports, int n)
274 struct odp_portvec pv;
279 error = do_ioctl(dpif_, ODP_PORT_LIST, &pv);
280 return error ? -error : pv.n_ports;
284 dpif_linux_port_poll(const struct dpif *dpif_, char **devnamep)
286 struct dpif_linux *dpif = dpif_linux_cast(dpif_);
288 if (dpif->change_error) {
289 dpif->change_error = false;
290 svec_clear(&dpif->changed_ports);
292 } else if (dpif->changed_ports.n) {
293 *devnamep = dpif->changed_ports.names[--dpif->changed_ports.n];
301 dpif_linux_port_poll_wait(const struct dpif *dpif_)
303 struct dpif_linux *dpif = dpif_linux_cast(dpif_);
304 if (dpif->changed_ports.n || dpif->change_error) {
305 poll_immediate_wake();
307 rtnetlink_notifier_wait();
312 dpif_linux_port_group_get(const struct dpif *dpif_, int group,
313 uint16_t ports[], int n)
315 struct odp_port_group pg;
318 assert(n <= UINT16_MAX);
322 error = do_ioctl(dpif_, ODP_PORT_GROUP_GET, &pg);
323 return error ? -error : pg.n_ports;
327 dpif_linux_port_group_set(struct dpif *dpif_, int group,
328 const uint16_t ports[], int n)
330 struct odp_port_group pg;
332 assert(n <= UINT16_MAX);
334 pg.ports = (uint16_t *) ports;
336 return do_ioctl(dpif_, ODP_PORT_GROUP_SET, &pg);
340 dpif_linux_flow_get(const struct dpif *dpif_, struct odp_flow flows[], int n)
342 struct odp_flowvec fv;
345 return do_ioctl(dpif_, ODP_FLOW_GET, &fv);
349 dpif_linux_flow_put(struct dpif *dpif_, struct odp_flow_put *put)
351 return do_ioctl(dpif_, ODP_FLOW_PUT, put);
355 dpif_linux_flow_del(struct dpif *dpif_, struct odp_flow *flow)
357 return do_ioctl(dpif_, ODP_FLOW_DEL, flow);
361 dpif_linux_flow_list(const struct dpif *dpif_, struct odp_flow flows[], int n)
363 struct odp_flowvec fv;
368 error = do_ioctl(dpif_, ODP_FLOW_LIST, &fv);
369 return error ? -error : fv.n_flows;
373 dpif_linux_execute(struct dpif *dpif_, uint16_t in_port,
374 const union odp_action actions[], int n_actions,
375 const struct ofpbuf *buf)
377 struct odp_execute execute;
378 memset(&execute, 0, sizeof execute);
379 execute.in_port = in_port;
380 execute.actions = (union odp_action *) actions;
381 execute.n_actions = n_actions;
382 execute.data = buf->data;
383 execute.length = buf->size;
384 return do_ioctl(dpif_, ODP_EXECUTE, &execute);
388 dpif_linux_recv_get_mask(const struct dpif *dpif_, int *listen_mask)
390 return do_ioctl(dpif_, ODP_GET_LISTEN_MASK, listen_mask);
394 dpif_linux_recv_set_mask(struct dpif *dpif_, int listen_mask)
396 return do_ioctl(dpif_, ODP_SET_LISTEN_MASK, &listen_mask);
400 dpif_linux_get_sflow_probability(const struct dpif *dpif_,
401 uint32_t *probability)
403 return do_ioctl(dpif_, ODP_GET_SFLOW_PROBABILITY, probability);
407 dpif_linux_set_sflow_probability(struct dpif *dpif_, uint32_t probability)
409 return do_ioctl(dpif_, ODP_SET_SFLOW_PROBABILITY, &probability);
413 dpif_linux_recv(struct dpif *dpif_, struct ofpbuf **bufp)
415 struct dpif_linux *dpif = dpif_linux_cast(dpif_);
420 buf = ofpbuf_new(65536);
421 retval = read(dpif->fd, ofpbuf_tail(buf), ofpbuf_tailroom(buf));
424 if (error != EAGAIN) {
425 VLOG_WARN_RL(&error_rl, "%s: read failed: %s",
426 dpif_name(dpif_), strerror(error));
428 } else if (retval >= sizeof(struct odp_msg)) {
429 struct odp_msg *msg = buf->data;
430 if (msg->length <= retval) {
435 VLOG_WARN_RL(&error_rl, "%s: discarding message truncated "
436 "from %"PRIu32" bytes to %d",
437 dpif_name(dpif_), msg->length, retval);
440 } else if (!retval) {
441 VLOG_WARN_RL(&error_rl, "%s: unexpected end of file", dpif_name(dpif_));
444 VLOG_WARN_RL(&error_rl,
445 "%s: discarding too-short message (%d bytes)",
446 dpif_name(dpif_), retval);
456 dpif_linux_recv_wait(struct dpif *dpif_)
458 struct dpif_linux *dpif = dpif_linux_cast(dpif_);
459 poll_fd_wait(dpif->fd, POLLIN);
462 const struct dpif_class dpif_linux_class = {
463 "", /* This is the default class. */
467 dpif_linux_enumerate,
470 dpif_linux_get_all_names,
472 dpif_linux_get_stats,
473 dpif_linux_get_drop_frags,
474 dpif_linux_set_drop_frags,
477 dpif_linux_port_query_by_number,
478 dpif_linux_port_query_by_name,
479 dpif_linux_port_list,
480 dpif_linux_port_poll,
481 dpif_linux_port_poll_wait,
482 dpif_linux_port_group_get,
483 dpif_linux_port_group_set,
487 dpif_linux_flow_flush,
488 dpif_linux_flow_list,
490 dpif_linux_recv_get_mask,
491 dpif_linux_recv_set_mask,
492 dpif_linux_get_sflow_probability,
493 dpif_linux_set_sflow_probability,
495 dpif_linux_recv_wait,
498 static int get_openvswitch_major(void);
499 static int get_major(const char *target);
502 do_ioctl(const struct dpif *dpif_, int cmd, const void *arg)
504 struct dpif_linux *dpif = dpif_linux_cast(dpif_);
505 return ioctl(dpif->fd, cmd, arg) ? errno : 0;
509 lookup_minor(const char *name, int *minorp)
511 struct ethtool_drvinfo drvinfo;
517 sock = socket(AF_INET, SOCK_DGRAM, 0);
519 VLOG_WARN("socket(AF_INET) failed: %s", strerror(errno));
524 memset(&ifr, 0, sizeof ifr);
525 strncpy(ifr.ifr_name, name, sizeof ifr.ifr_name);
526 ifr.ifr_data = (caddr_t) &drvinfo;
528 memset(&drvinfo, 0, sizeof drvinfo);
529 drvinfo.cmd = ETHTOOL_GDRVINFO;
530 if (ioctl(sock, SIOCETHTOOL, &ifr)) {
531 VLOG_WARN("ioctl(SIOCETHTOOL) failed: %s", strerror(errno));
533 goto error_close_sock;
536 if (strcmp(drvinfo.driver, "openvswitch")) {
537 VLOG_WARN("%s is not an openvswitch device", name);
539 goto error_close_sock;
542 if (sscanf(drvinfo.bus_info, "%d.%d", &minor, &port_no) != 2) {
543 VLOG_WARN("%s ethtool bus_info has unexpected format", name);
545 goto error_close_sock;
546 } else if (port_no != ODPP_LOCAL) {
547 /* This is an Open vSwitch device but not the local port. We
548 * intentionally support only using the name of the local port as the
549 * name of a datapath; otherwise, it would be too difficult to
550 * enumerate all the names of a datapath. */
552 goto error_close_sock;
566 make_openvswitch_device(int minor, char **fnp)
568 const char dirname[] = "/dev/net";
576 major = get_openvswitch_major();
580 dev = makedev(major, minor);
582 sprintf(fn, "%s/dp%d", dirname, minor);
584 if (!S_ISCHR(s.st_mode)) {
585 VLOG_WARN_RL(&error_rl, "%s is not a character device, fixing",
587 } else if (s.st_rdev != dev) {
588 VLOG_WARN_RL(&error_rl,
589 "%s is device %u:%u but should be %u:%u, fixing",
590 fn, major(s.st_rdev), minor(s.st_rdev),
591 major(dev), minor(dev));
596 VLOG_WARN_RL(&error_rl, "%s: unlink failed (%s)",
597 fn, strerror(errno));
600 } else if (errno == ENOENT) {
601 if (stat(dirname, &s)) {
602 if (errno == ENOENT) {
603 if (mkdir(dirname, 0755)) {
604 VLOG_WARN_RL(&error_rl, "%s: mkdir failed (%s)",
605 dirname, strerror(errno));
609 VLOG_WARN_RL(&error_rl, "%s: stat failed (%s)",
610 dirname, strerror(errno));
615 VLOG_WARN_RL(&error_rl, "%s: stat failed (%s)", fn, strerror(errno));
619 /* The device needs to be created. */
620 if (mknod(fn, S_IFCHR | 0700, dev)) {
621 VLOG_WARN_RL(&error_rl,
622 "%s: creating character device %u:%u failed (%s)",
623 fn, major(dev), minor(dev), strerror(errno));
632 /* Return the major device number of the Open vSwitch device. If it
633 * cannot be determined, a negative errno is returned. */
635 get_openvswitch_major(void)
637 static int openvswitch_major = -1;
638 if (openvswitch_major < 0) {
639 openvswitch_major = get_major("openvswitch");
641 return openvswitch_major;
645 get_major(const char *target)
647 const char fn[] = "/proc/devices";
652 file = fopen(fn, "r");
654 VLOG_ERR("opening %s failed (%s)", fn, strerror(errno));
658 for (ln = 1; fgets(line, sizeof line, file); ln++) {
662 if (!strncmp(line, "Character", 9) || line[0] == '\0') {
664 } else if (!strncmp(line, "Block", 5)) {
665 /* We only want character devices, so skip the rest of the file. */
667 } else if (sscanf(line, "%d %63s", &major, name)) {
668 if (!strcmp(name, target)) {
675 VLOG_WARN("%s:%d: syntax error", fn, ln);
681 VLOG_ERR("%s: %s major not found (is the module loaded?)", fn, target);
686 finish_open(struct dpif *dpif_, const char *local_ifname)
688 struct dpif_linux *dpif = dpif_linux_cast(dpif_);
689 dpif->local_ifname = strdup(local_ifname);
690 dpif->local_ifindex = if_nametoindex(local_ifname);
691 if (!dpif->local_ifindex) {
694 VLOG_WARN("could not get ifindex of %s device: %s",
695 local_ifname, strerror(errno));
702 create_minor(const char *name, int minor, struct dpif **dpifp)
704 int error = open_minor(minor, dpifp);
706 error = do_ioctl(*dpifp, ODP_DP_CREATE, name);
708 error = finish_open(*dpifp, name);
717 open_minor(int minor, struct dpif **dpifp)
723 error = make_openvswitch_device(minor, &fn);
728 fd = open(fn, O_RDONLY | O_NONBLOCK);
730 struct dpif_linux *dpif = xmalloc(sizeof *dpif);
731 error = rtnetlink_notifier_register(&dpif->port_notifier,
732 dpif_linux_port_changed, dpif);
736 name = xasprintf("dp%d", minor);
737 dpif_init(&dpif->dpif, &dpif_linux_class, name, minor, minor);
741 dpif->local_ifname = NULL;
743 dpif->local_ifindex = 0;
744 svec_init(&dpif->changed_ports);
745 dpif->change_error = false;
746 *dpifp = &dpif->dpif;
752 VLOG_WARN("%s: open failed (%s)", fn, strerror(error));
760 dpif_linux_port_changed(const struct rtnetlink_change *change, void *dpif_)
762 struct dpif_linux *dpif = dpif_;
765 if (change->master_ifindex == dpif->local_ifindex
766 && (change->nlmsg_type == RTM_NEWLINK
767 || change->nlmsg_type == RTM_DELLINK))
769 /* Our datapath changed, either adding a new port or deleting an
771 if (!svec_contains(&dpif->changed_ports, change->ifname)) {
772 svec_add(&dpif->changed_ports, change->ifname);
773 svec_sort(&dpif->changed_ports);
777 dpif->change_error = true;