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, "system", &dpif);
100 svec_add(all_dps, devname);
102 } else if (retval != ENODEV && !error) {
110 dpif_linux_open(const char *name, const char *type UNUSED, bool create,
115 minor = !strncmp(name, "dp", 2)
116 && isdigit((unsigned char)name[2]) ? atoi(name + 2) : -1;
119 return create_minor(name, minor, dpifp);
121 /* Scan for unused minor number. */
122 for (minor = 0; minor < ODP_MAX; minor++) {
123 int error = create_minor(name, 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(name, &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 return do_ioctl(dpif_, ODP_DP_STATS, stats);
203 dpif_linux_get_drop_frags(const struct dpif *dpif_, bool *drop_fragsp)
208 error = do_ioctl(dpif_, ODP_GET_DROP_FRAGS, &drop_frags);
210 *drop_fragsp = drop_frags & 1;
216 dpif_linux_set_drop_frags(struct dpif *dpif_, bool drop_frags)
218 int drop_frags_int = drop_frags;
219 return do_ioctl(dpif_, ODP_SET_DROP_FRAGS, &drop_frags_int);
223 dpif_linux_port_add(struct dpif *dpif_, const char *devname, uint16_t flags,
226 struct odp_port port;
229 memset(&port, 0, sizeof port);
230 strncpy(port.devname, devname, sizeof port.devname);
232 error = do_ioctl(dpif_, ODP_PORT_ADD, &port);
234 *port_no = port.port;
240 dpif_linux_port_del(struct dpif *dpif_, uint16_t port_no)
243 return do_ioctl(dpif_, ODP_PORT_DEL, &tmp);
247 dpif_linux_port_query_by_number(const struct dpif *dpif_, uint16_t port_no,
248 struct odp_port *port)
250 memset(port, 0, sizeof *port);
251 port->port = port_no;
252 return do_ioctl(dpif_, ODP_PORT_QUERY, port);
256 dpif_linux_port_query_by_name(const struct dpif *dpif_, const char *devname,
257 struct odp_port *port)
259 memset(port, 0, sizeof *port);
260 strncpy(port->devname, devname, sizeof port->devname);
261 return do_ioctl(dpif_, ODP_PORT_QUERY, port);
265 dpif_linux_flow_flush(struct dpif *dpif_)
267 return do_ioctl(dpif_, ODP_FLOW_FLUSH, NULL);
271 dpif_linux_port_list(const struct dpif *dpif_, struct odp_port *ports, int n)
273 struct odp_portvec pv;
278 error = do_ioctl(dpif_, ODP_PORT_LIST, &pv);
279 return error ? -error : pv.n_ports;
283 dpif_linux_port_poll(const struct dpif *dpif_, char **devnamep)
285 struct dpif_linux *dpif = dpif_linux_cast(dpif_);
287 if (dpif->change_error) {
288 dpif->change_error = false;
289 svec_clear(&dpif->changed_ports);
291 } else if (dpif->changed_ports.n) {
292 *devnamep = dpif->changed_ports.names[--dpif->changed_ports.n];
300 dpif_linux_port_poll_wait(const struct dpif *dpif_)
302 struct dpif_linux *dpif = dpif_linux_cast(dpif_);
303 if (dpif->changed_ports.n || dpif->change_error) {
304 poll_immediate_wake();
306 rtnetlink_notifier_wait();
311 dpif_linux_port_group_get(const struct dpif *dpif_, int group,
312 uint16_t ports[], int n)
314 struct odp_port_group pg;
317 assert(n <= UINT16_MAX);
321 error = do_ioctl(dpif_, ODP_PORT_GROUP_GET, &pg);
322 return error ? -error : pg.n_ports;
326 dpif_linux_port_group_set(struct dpif *dpif_, int group,
327 const uint16_t ports[], int n)
329 struct odp_port_group pg;
331 assert(n <= UINT16_MAX);
333 pg.ports = (uint16_t *) ports;
335 return do_ioctl(dpif_, ODP_PORT_GROUP_SET, &pg);
339 dpif_linux_flow_get(const struct dpif *dpif_, struct odp_flow flows[], int n)
341 struct odp_flowvec fv;
344 return do_ioctl(dpif_, ODP_FLOW_GET, &fv);
348 dpif_linux_flow_put(struct dpif *dpif_, struct odp_flow_put *put)
350 return do_ioctl(dpif_, ODP_FLOW_PUT, put);
354 dpif_linux_flow_del(struct dpif *dpif_, struct odp_flow *flow)
356 return do_ioctl(dpif_, ODP_FLOW_DEL, flow);
360 dpif_linux_flow_list(const struct dpif *dpif_, struct odp_flow flows[], int n)
362 struct odp_flowvec fv;
367 error = do_ioctl(dpif_, ODP_FLOW_LIST, &fv);
368 return error ? -error : fv.n_flows;
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)
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);
387 dpif_linux_recv_get_mask(const struct dpif *dpif_, int *listen_mask)
389 return do_ioctl(dpif_, ODP_GET_LISTEN_MASK, listen_mask);
393 dpif_linux_recv_set_mask(struct dpif *dpif_, int listen_mask)
395 return do_ioctl(dpif_, ODP_SET_LISTEN_MASK, &listen_mask);
399 dpif_linux_recv(struct dpif *dpif_, struct ofpbuf **bufp)
401 struct dpif_linux *dpif = dpif_linux_cast(dpif_);
406 buf = ofpbuf_new(65536);
407 retval = read(dpif->fd, ofpbuf_tail(buf), ofpbuf_tailroom(buf));
410 if (error != EAGAIN) {
411 VLOG_WARN_RL(&error_rl, "%s: read failed: %s",
412 dpif_name(dpif_), strerror(error));
414 } else if (retval >= sizeof(struct odp_msg)) {
415 struct odp_msg *msg = buf->data;
416 if (msg->length <= retval) {
421 VLOG_WARN_RL(&error_rl, "%s: discarding message truncated "
422 "from %"PRIu32" bytes to %d",
423 dpif_name(dpif_), msg->length, retval);
426 } else if (!retval) {
427 VLOG_WARN_RL(&error_rl, "%s: unexpected end of file", dpif_name(dpif_));
430 VLOG_WARN_RL(&error_rl,
431 "%s: discarding too-short message (%d bytes)",
432 dpif_name(dpif_), retval);
442 dpif_linux_recv_wait(struct dpif *dpif_)
444 struct dpif_linux *dpif = dpif_linux_cast(dpif_);
445 poll_fd_wait(dpif->fd, POLLIN);
448 const struct dpif_class dpif_linux_class = {
452 dpif_linux_enumerate,
455 dpif_linux_get_all_names,
457 dpif_linux_get_stats,
458 dpif_linux_get_drop_frags,
459 dpif_linux_set_drop_frags,
462 dpif_linux_port_query_by_number,
463 dpif_linux_port_query_by_name,
464 dpif_linux_port_list,
465 dpif_linux_port_poll,
466 dpif_linux_port_poll_wait,
467 dpif_linux_port_group_get,
468 dpif_linux_port_group_set,
472 dpif_linux_flow_flush,
473 dpif_linux_flow_list,
475 dpif_linux_recv_get_mask,
476 dpif_linux_recv_set_mask,
478 dpif_linux_recv_wait,
481 static int get_openvswitch_major(void);
482 static int get_major(const char *target);
485 do_ioctl(const struct dpif *dpif_, int cmd, const void *arg)
487 struct dpif_linux *dpif = dpif_linux_cast(dpif_);
488 return ioctl(dpif->fd, cmd, arg) ? errno : 0;
492 lookup_minor(const char *name, int *minorp)
494 struct ethtool_drvinfo drvinfo;
500 sock = socket(AF_INET, SOCK_DGRAM, 0);
502 VLOG_WARN("socket(AF_INET) failed: %s", strerror(errno));
507 memset(&ifr, 0, sizeof ifr);
508 strncpy(ifr.ifr_name, name, sizeof ifr.ifr_name);
509 ifr.ifr_data = (caddr_t) &drvinfo;
511 memset(&drvinfo, 0, sizeof drvinfo);
512 drvinfo.cmd = ETHTOOL_GDRVINFO;
513 if (ioctl(sock, SIOCETHTOOL, &ifr)) {
514 VLOG_WARN("ioctl(SIOCETHTOOL) failed: %s", strerror(errno));
516 goto error_close_sock;
519 if (strcmp(drvinfo.driver, "openvswitch")) {
520 VLOG_WARN("%s is not an openvswitch device", name);
522 goto error_close_sock;
525 if (sscanf(drvinfo.bus_info, "%d.%d", &minor, &port_no) != 2) {
526 VLOG_WARN("%s ethtool bus_info has unexpected format", name);
528 goto error_close_sock;
529 } else if (port_no != ODPP_LOCAL) {
530 /* This is an Open vSwitch device but not the local port. We
531 * intentionally support only using the name of the local port as the
532 * name of a datapath; otherwise, it would be too difficult to
533 * enumerate all the names of a datapath. */
535 goto error_close_sock;
549 make_openvswitch_device(int minor, char **fnp)
551 const char dirname[] = "/dev/net";
557 major = get_openvswitch_major();
561 dev = makedev(major, minor);
564 sprintf(fn, "%s/dp%d", dirname, minor);
566 if (!S_ISCHR(s.st_mode)) {
567 VLOG_WARN_RL(&error_rl, "%s is not a character device, fixing",
569 } else if (s.st_rdev != dev) {
570 VLOG_WARN_RL(&error_rl,
571 "%s is device %u:%u but should be %u:%u, fixing",
572 fn, major(s.st_rdev), minor(s.st_rdev),
573 major(dev), minor(dev));
578 VLOG_WARN_RL(&error_rl, "%s: unlink failed (%s)",
579 fn, strerror(errno));
582 } else if (errno == ENOENT) {
583 if (stat(dirname, &s)) {
584 if (errno == ENOENT) {
585 if (mkdir(dirname, 0755)) {
586 VLOG_WARN_RL(&error_rl, "%s: mkdir failed (%s)",
587 dirname, strerror(errno));
591 VLOG_WARN_RL(&error_rl, "%s: stat failed (%s)",
592 dirname, strerror(errno));
597 VLOG_WARN_RL(&error_rl, "%s: stat failed (%s)", fn, strerror(errno));
601 /* The device needs to be created. */
602 if (mknod(fn, S_IFCHR | 0700, dev)) {
603 VLOG_WARN_RL(&error_rl,
604 "%s: creating character device %u:%u failed (%s)",
605 fn, major(dev), minor(dev), strerror(errno));
614 /* Return the major device number of the Open vSwitch device. If it
615 * cannot be determined, a negative errno is returned. */
617 get_openvswitch_major(void)
619 static int openvswitch_major = -1;
620 if (openvswitch_major < 0) {
621 openvswitch_major = get_major("openvswitch");
623 return openvswitch_major;
627 get_major(const char *target)
629 const char fn[] = "/proc/devices";
634 file = fopen(fn, "r");
636 VLOG_ERR("opening %s failed (%s)", fn, strerror(errno));
640 for (ln = 1; fgets(line, sizeof line, file); ln++) {
644 if (!strncmp(line, "Character", 9) || line[0] == '\0') {
646 } else if (!strncmp(line, "Block", 5)) {
647 /* We only want character devices, so skip the rest of the file. */
649 } else if (sscanf(line, "%d %63s", &major, name)) {
650 if (!strcmp(name, target)) {
657 VLOG_WARN("%s:%d: syntax error", fn, ln);
663 VLOG_ERR("%s: %s major not found (is the module loaded?)", fn, target);
668 finish_open(struct dpif *dpif_, const char *local_ifname)
670 struct dpif_linux *dpif = dpif_linux_cast(dpif_);
671 dpif->local_ifname = strdup(local_ifname);
672 dpif->local_ifindex = if_nametoindex(local_ifname);
673 if (!dpif->local_ifindex) {
676 VLOG_WARN("could not get ifindex of %s device: %s",
677 local_ifname, strerror(errno));
684 create_minor(const char *name, int minor, struct dpif **dpifp)
686 int error = open_minor(minor, dpifp);
688 error = do_ioctl(*dpifp, ODP_DP_CREATE, name);
690 error = finish_open(*dpifp, name);
699 open_minor(int minor, struct dpif **dpifp)
705 error = make_openvswitch_device(minor, &fn);
710 fd = open(fn, O_RDONLY | O_NONBLOCK);
712 struct dpif_linux *dpif = xmalloc(sizeof *dpif);
713 error = rtnetlink_notifier_register(&dpif->port_notifier,
714 dpif_linux_port_changed, dpif);
718 name = xasprintf("dp%d", minor);
719 dpif_init(&dpif->dpif, &dpif_linux_class, name, minor, minor);
723 dpif->local_ifname = NULL;
725 dpif->local_ifindex = 0;
726 svec_init(&dpif->changed_ports);
727 dpif->change_error = false;
728 *dpifp = &dpif->dpif;
734 VLOG_WARN("%s: open failed (%s)", fn, strerror(error));
742 dpif_linux_port_changed(const struct rtnetlink_change *change, void *dpif_)
744 struct dpif_linux *dpif = dpif_;
747 if (change->master_ifindex == dpif->local_ifindex
748 && (change->nlmsg_type == RTM_NEWLINK
749 || change->nlmsg_type == RTM_DELLINK))
751 /* Our datapath changed, either adding a new port or deleting an
753 if (!svec_contains(&dpif->changed_ports, change->ifname)) {
754 svec_add(&dpif->changed_ports, change->ifname);
755 svec_sort(&dpif->changed_ports);
759 dpif->change_error = true;