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"
36 #include "poll-loop.h"
37 #include "rtnetlink.h"
42 #define THIS_MODULE VLM_dpif_linux
44 /* Datapath interface for the openvswitch Linux kernel module. */
49 /* Used by dpif_linux_get_all_names(). */
53 /* Change notification. */
54 int local_ifindex
; /* Ifindex of local port. */
55 struct svec changed_ports
; /* Ports that have changed. */
56 struct rtnetlink_notifier port_notifier
;
60 static struct vlog_rate_limit error_rl
= VLOG_RATE_LIMIT_INIT(9999, 5);
62 static int do_ioctl(const struct dpif
*, int cmd
, const void *arg
);
63 static int lookup_minor(const char *name
, int *minor
);
64 static int finish_open(struct dpif
*, const char *local_ifname
);
65 static int get_openvswitch_major(void);
66 static int create_minor(const char *name
, int minor
, struct dpif
**dpifp
);
67 static int open_minor(int minor
, struct dpif
**dpifp
);
68 static int make_openvswitch_device(int minor
, char **fnp
);
69 static void dpif_linux_port_changed(const struct rtnetlink_change
*,
72 static struct dpif_linux
*
73 dpif_linux_cast(const struct dpif
*dpif
)
75 dpif_assert_class(dpif
, &dpif_linux_class
);
76 return CONTAINER_OF(dpif
, struct dpif_linux
, dpif
);
80 dpif_linux_enumerate(struct svec
*all_dps
)
86 /* Check that the Open vSwitch module is loaded. */
87 major
= get_openvswitch_major();
93 for (i
= 0; i
< ODP_MAX
; i
++) {
98 sprintf(devname
, "dp%d", i
);
99 retval
= dpif_open(devname
, "system", &dpif
);
101 svec_add(all_dps
, devname
);
102 dpif_uninit(dpif
, true);
103 } else if (retval
!= ENODEV
&& !error
) {
111 dpif_linux_open(const char *name
, const char *type OVS_UNUSED
, bool create
,
116 minor
= !strncmp(name
, "dp", 2)
117 && isdigit((unsigned char)name
[2]) ? atoi(name
+ 2) : -1;
120 return create_minor(name
, minor
, dpifp
);
122 /* Scan for unused minor number. */
123 for (minor
= 0; minor
< ODP_MAX
; minor
++) {
124 int error
= create_minor(name
, minor
, dpifp
);
125 if (error
!= EBUSY
) {
130 /* All datapath numbers in use. */
134 struct dpif_linux
*dpif
;
135 struct odp_port port
;
139 error
= lookup_minor(name
, &minor
);
145 error
= open_minor(minor
, dpifp
);
149 dpif
= dpif_linux_cast(*dpifp
);
151 /* We need the local port's ifindex for the poll function. Start by
152 * getting the local port's name. */
153 memset(&port
, 0, sizeof port
);
154 port
.port
= ODPP_LOCAL
;
155 if (ioctl(dpif
->fd
, ODP_PORT_QUERY
, &port
)) {
157 if (error
!= ENODEV
) {
158 VLOG_WARN("%s: probe returned unexpected error: %s",
159 dpif_name(*dpifp
), strerror(error
));
161 dpif_uninit(*dpifp
, true);
165 /* Then use that to finish up opening. */
166 return finish_open(&dpif
->dpif
, port
.devname
);
171 dpif_linux_close(struct dpif
*dpif_
)
173 struct dpif_linux
*dpif
= dpif_linux_cast(dpif_
);
174 rtnetlink_notifier_unregister(&dpif
->port_notifier
);
175 svec_destroy(&dpif
->changed_ports
);
176 free(dpif
->local_ifname
);
182 dpif_linux_get_all_names(const struct dpif
*dpif_
, struct svec
*all_names
)
184 struct dpif_linux
*dpif
= dpif_linux_cast(dpif_
);
186 svec_add_nocopy(all_names
, xasprintf("dp%d", dpif
->minor
));
187 svec_add(all_names
, dpif
->local_ifname
);
192 dpif_linux_destroy(struct dpif
*dpif_
)
194 struct odp_port
*ports
;
199 err
= dpif_port_list(dpif_
, &ports
, &n_ports
);
204 for (i
= 0; i
< n_ports
; i
++) {
205 if (ports
[i
].port
!= ODPP_LOCAL
) {
206 err
= do_ioctl(dpif_
, ODP_VPORT_DEL
, ports
[i
].devname
);
208 VLOG_WARN_RL(&error_rl
, "%s: error deleting port %s (%s)",
209 dpif_name(dpif_
), ports
[i
].devname
, strerror(err
));
216 return do_ioctl(dpif_
, ODP_DP_DESTROY
, NULL
);
220 dpif_linux_get_stats(const struct dpif
*dpif_
, struct odp_stats
*stats
)
222 memset(stats
, 0, sizeof *stats
);
223 return do_ioctl(dpif_
, ODP_DP_STATS
, stats
);
227 dpif_linux_get_drop_frags(const struct dpif
*dpif_
, bool *drop_fragsp
)
232 error
= do_ioctl(dpif_
, ODP_GET_DROP_FRAGS
, &drop_frags
);
234 *drop_fragsp
= drop_frags
& 1;
240 dpif_linux_set_drop_frags(struct dpif
*dpif_
, bool drop_frags
)
242 int drop_frags_int
= drop_frags
;
243 return do_ioctl(dpif_
, ODP_SET_DROP_FRAGS
, &drop_frags_int
);
247 dpif_linux_port_add(struct dpif
*dpif_
, const char *devname
, uint16_t flags
,
250 struct odp_port port
;
253 memset(&port
, 0, sizeof port
);
254 strncpy(port
.devname
, devname
, sizeof port
.devname
);
256 error
= do_ioctl(dpif_
, ODP_PORT_ATTACH
, &port
);
258 *port_no
= port
.port
;
264 dpif_linux_port_del(struct dpif
*dpif_
, uint16_t port_no
)
268 struct odp_port port
;
270 err
= dpif_port_query_by_number(dpif_
, port_no
, &port
);
275 err
= do_ioctl(dpif_
, ODP_PORT_DETACH
, &tmp
);
280 if (!netdev_is_open(port
.devname
)) {
281 /* Try deleting the port if no one has it open. This shouldn't
282 * actually be necessary unless the config changed while we weren't
283 * running but it won't hurt anything if the port is already gone. */
284 do_ioctl(dpif_
, ODP_VPORT_DEL
, port
.devname
);
291 dpif_linux_port_query_by_number(const struct dpif
*dpif_
, uint16_t port_no
,
292 struct odp_port
*port
)
294 memset(port
, 0, sizeof *port
);
295 port
->port
= port_no
;
296 return do_ioctl(dpif_
, ODP_PORT_QUERY
, port
);
300 dpif_linux_port_query_by_name(const struct dpif
*dpif_
, const char *devname
,
301 struct odp_port
*port
)
303 memset(port
, 0, sizeof *port
);
304 strncpy(port
->devname
, devname
, sizeof port
->devname
);
305 return do_ioctl(dpif_
, ODP_PORT_QUERY
, port
);
309 dpif_linux_flow_flush(struct dpif
*dpif_
)
311 return do_ioctl(dpif_
, ODP_FLOW_FLUSH
, NULL
);
315 dpif_linux_port_list(const struct dpif
*dpif_
, struct odp_port
*ports
, int n
)
317 struct odp_portvec pv
;
322 error
= do_ioctl(dpif_
, ODP_PORT_LIST
, &pv
);
323 return error
? -error
: pv
.n_ports
;
327 dpif_linux_port_poll(const struct dpif
*dpif_
, char **devnamep
)
329 struct dpif_linux
*dpif
= dpif_linux_cast(dpif_
);
331 if (dpif
->change_error
) {
332 dpif
->change_error
= false;
333 svec_clear(&dpif
->changed_ports
);
335 } else if (dpif
->changed_ports
.n
) {
336 *devnamep
= dpif
->changed_ports
.names
[--dpif
->changed_ports
.n
];
344 dpif_linux_port_poll_wait(const struct dpif
*dpif_
)
346 struct dpif_linux
*dpif
= dpif_linux_cast(dpif_
);
347 if (dpif
->changed_ports
.n
|| dpif
->change_error
) {
348 poll_immediate_wake();
350 rtnetlink_notifier_wait();
355 dpif_linux_port_group_get(const struct dpif
*dpif_
, int group
,
356 uint16_t ports
[], int n
)
358 struct odp_port_group pg
;
361 assert(n
<= UINT16_MAX
);
365 error
= do_ioctl(dpif_
, ODP_PORT_GROUP_GET
, &pg
);
366 return error
? -error
: pg
.n_ports
;
370 dpif_linux_port_group_set(struct dpif
*dpif_
, int group
,
371 const uint16_t ports
[], int n
)
373 struct odp_port_group pg
;
375 assert(n
<= UINT16_MAX
);
377 pg
.ports
= (uint16_t *) ports
;
379 return do_ioctl(dpif_
, ODP_PORT_GROUP_SET
, &pg
);
383 dpif_linux_flow_get(const struct dpif
*dpif_
, struct odp_flow flows
[], int n
)
385 struct odp_flowvec fv
;
388 return do_ioctl(dpif_
, ODP_FLOW_GET
, &fv
);
392 dpif_linux_flow_put(struct dpif
*dpif_
, struct odp_flow_put
*put
)
394 return do_ioctl(dpif_
, ODP_FLOW_PUT
, put
);
398 dpif_linux_flow_del(struct dpif
*dpif_
, struct odp_flow
*flow
)
400 return do_ioctl(dpif_
, ODP_FLOW_DEL
, flow
);
404 dpif_linux_flow_list(const struct dpif
*dpif_
, struct odp_flow flows
[], int n
)
406 struct odp_flowvec fv
;
411 error
= do_ioctl(dpif_
, ODP_FLOW_LIST
, &fv
);
412 return error
? -error
: fv
.n_flows
;
416 dpif_linux_execute(struct dpif
*dpif_
, uint16_t in_port
,
417 const union odp_action actions
[], int n_actions
,
418 const struct ofpbuf
*buf
)
420 struct odp_execute execute
;
421 memset(&execute
, 0, sizeof execute
);
422 execute
.in_port
= in_port
;
423 execute
.actions
= (union odp_action
*) actions
;
424 execute
.n_actions
= n_actions
;
425 execute
.data
= buf
->data
;
426 execute
.length
= buf
->size
;
427 return do_ioctl(dpif_
, ODP_EXECUTE
, &execute
);
431 dpif_linux_recv_get_mask(const struct dpif
*dpif_
, int *listen_mask
)
433 return do_ioctl(dpif_
, ODP_GET_LISTEN_MASK
, listen_mask
);
437 dpif_linux_recv_set_mask(struct dpif
*dpif_
, int listen_mask
)
439 return do_ioctl(dpif_
, ODP_SET_LISTEN_MASK
, &listen_mask
);
443 dpif_linux_get_sflow_probability(const struct dpif
*dpif_
,
444 uint32_t *probability
)
446 return do_ioctl(dpif_
, ODP_GET_SFLOW_PROBABILITY
, probability
);
450 dpif_linux_set_sflow_probability(struct dpif
*dpif_
, uint32_t probability
)
452 return do_ioctl(dpif_
, ODP_SET_SFLOW_PROBABILITY
, &probability
);
456 dpif_linux_recv(struct dpif
*dpif_
, struct ofpbuf
**bufp
)
458 struct dpif_linux
*dpif
= dpif_linux_cast(dpif_
);
463 buf
= ofpbuf_new(65536);
464 retval
= read(dpif
->fd
, ofpbuf_tail(buf
), ofpbuf_tailroom(buf
));
467 if (error
!= EAGAIN
) {
468 VLOG_WARN_RL(&error_rl
, "%s: read failed: %s",
469 dpif_name(dpif_
), strerror(error
));
471 } else if (retval
>= sizeof(struct odp_msg
)) {
472 struct odp_msg
*msg
= buf
->data
;
473 if (msg
->length
<= retval
) {
478 VLOG_WARN_RL(&error_rl
, "%s: discarding message truncated "
479 "from %"PRIu32
" bytes to %d",
480 dpif_name(dpif_
), msg
->length
, retval
);
483 } else if (!retval
) {
484 VLOG_WARN_RL(&error_rl
, "%s: unexpected end of file", dpif_name(dpif_
));
487 VLOG_WARN_RL(&error_rl
,
488 "%s: discarding too-short message (%d bytes)",
489 dpif_name(dpif_
), retval
);
499 dpif_linux_recv_wait(struct dpif
*dpif_
)
501 struct dpif_linux
*dpif
= dpif_linux_cast(dpif_
);
502 poll_fd_wait(dpif
->fd
, POLLIN
);
505 const struct dpif_class dpif_linux_class
= {
509 dpif_linux_enumerate
,
512 dpif_linux_get_all_names
,
514 dpif_linux_get_stats
,
515 dpif_linux_get_drop_frags
,
516 dpif_linux_set_drop_frags
,
519 dpif_linux_port_query_by_number
,
520 dpif_linux_port_query_by_name
,
521 dpif_linux_port_list
,
522 dpif_linux_port_poll
,
523 dpif_linux_port_poll_wait
,
524 dpif_linux_port_group_get
,
525 dpif_linux_port_group_set
,
529 dpif_linux_flow_flush
,
530 dpif_linux_flow_list
,
532 dpif_linux_recv_get_mask
,
533 dpif_linux_recv_set_mask
,
534 dpif_linux_get_sflow_probability
,
535 dpif_linux_set_sflow_probability
,
537 dpif_linux_recv_wait
,
540 static int get_openvswitch_major(void);
541 static int get_major(const char *target
);
544 do_ioctl(const struct dpif
*dpif_
, int cmd
, const void *arg
)
546 struct dpif_linux
*dpif
= dpif_linux_cast(dpif_
);
547 return ioctl(dpif
->fd
, cmd
, arg
) ? errno
: 0;
551 lookup_minor(const char *name
, int *minorp
)
553 struct ethtool_drvinfo drvinfo
;
559 sock
= socket(AF_INET
, SOCK_DGRAM
, 0);
561 VLOG_WARN("socket(AF_INET) failed: %s", strerror(errno
));
566 memset(&ifr
, 0, sizeof ifr
);
567 strncpy(ifr
.ifr_name
, name
, sizeof ifr
.ifr_name
);
568 ifr
.ifr_data
= (caddr_t
) &drvinfo
;
570 memset(&drvinfo
, 0, sizeof drvinfo
);
571 drvinfo
.cmd
= ETHTOOL_GDRVINFO
;
572 if (ioctl(sock
, SIOCETHTOOL
, &ifr
)) {
573 VLOG_WARN("ioctl(SIOCETHTOOL) failed: %s", strerror(errno
));
575 goto error_close_sock
;
578 if (strcmp(drvinfo
.driver
, "openvswitch")) {
579 VLOG_WARN("%s is not an openvswitch device", name
);
581 goto error_close_sock
;
584 if (sscanf(drvinfo
.bus_info
, "%d.%d", &minor
, &port_no
) != 2) {
585 VLOG_WARN("%s ethtool bus_info has unexpected format", name
);
587 goto error_close_sock
;
588 } else if (port_no
!= ODPP_LOCAL
) {
589 /* This is an Open vSwitch device but not the local port. We
590 * intentionally support only using the name of the local port as the
591 * name of a datapath; otherwise, it would be too difficult to
592 * enumerate all the names of a datapath. */
594 goto error_close_sock
;
608 make_openvswitch_device(int minor
, char **fnp
)
610 const char dirname
[] = "/dev/net";
618 major
= get_openvswitch_major();
622 dev
= makedev(major
, minor
);
624 sprintf(fn
, "%s/dp%d", dirname
, minor
);
626 if (!S_ISCHR(s
.st_mode
)) {
627 VLOG_WARN_RL(&error_rl
, "%s is not a character device, fixing",
629 } else if (s
.st_rdev
!= dev
) {
630 VLOG_WARN_RL(&error_rl
,
631 "%s is device %u:%u but should be %u:%u, fixing",
632 fn
, major(s
.st_rdev
), minor(s
.st_rdev
),
633 major(dev
), minor(dev
));
638 VLOG_WARN_RL(&error_rl
, "%s: unlink failed (%s)",
639 fn
, strerror(errno
));
642 } else if (errno
== ENOENT
) {
643 if (stat(dirname
, &s
)) {
644 if (errno
== ENOENT
) {
645 if (mkdir(dirname
, 0755)) {
646 VLOG_WARN_RL(&error_rl
, "%s: mkdir failed (%s)",
647 dirname
, strerror(errno
));
651 VLOG_WARN_RL(&error_rl
, "%s: stat failed (%s)",
652 dirname
, strerror(errno
));
657 VLOG_WARN_RL(&error_rl
, "%s: stat failed (%s)", fn
, strerror(errno
));
661 /* The device needs to be created. */
662 if (mknod(fn
, S_IFCHR
| 0700, dev
)) {
663 VLOG_WARN_RL(&error_rl
,
664 "%s: creating character device %u:%u failed (%s)",
665 fn
, major(dev
), minor(dev
), strerror(errno
));
674 /* Return the major device number of the Open vSwitch device. If it
675 * cannot be determined, a negative errno is returned. */
677 get_openvswitch_major(void)
679 static int openvswitch_major
= -1;
680 if (openvswitch_major
< 0) {
681 openvswitch_major
= get_major("openvswitch");
683 return openvswitch_major
;
687 get_major(const char *target
)
689 const char fn
[] = "/proc/devices";
694 file
= fopen(fn
, "r");
696 VLOG_ERR("opening %s failed (%s)", fn
, strerror(errno
));
700 for (ln
= 1; fgets(line
, sizeof line
, file
); ln
++) {
704 if (!strncmp(line
, "Character", 9) || line
[0] == '\0') {
706 } else if (!strncmp(line
, "Block", 5)) {
707 /* We only want character devices, so skip the rest of the file. */
709 } else if (sscanf(line
, "%d %63s", &major
, name
)) {
710 if (!strcmp(name
, target
)) {
717 VLOG_WARN("%s:%d: syntax error", fn
, ln
);
725 VLOG_ERR("%s: %s major not found (is the module loaded?)", fn
, target
);
730 finish_open(struct dpif
*dpif_
, const char *local_ifname
)
732 struct dpif_linux
*dpif
= dpif_linux_cast(dpif_
);
733 dpif
->local_ifname
= xstrdup(local_ifname
);
734 dpif
->local_ifindex
= if_nametoindex(local_ifname
);
735 if (!dpif
->local_ifindex
) {
737 dpif_uninit(dpif_
, true);
738 VLOG_WARN("could not get ifindex of %s device: %s",
739 local_ifname
, strerror(errno
));
746 create_minor(const char *name
, int minor
, struct dpif
**dpifp
)
748 int error
= open_minor(minor
, dpifp
);
750 error
= do_ioctl(*dpifp
, ODP_DP_CREATE
, name
);
752 error
= finish_open(*dpifp
, name
);
754 dpif_uninit(*dpifp
, true);
761 open_minor(int minor
, struct dpif
**dpifp
)
767 error
= make_openvswitch_device(minor
, &fn
);
772 fd
= open(fn
, O_RDONLY
| O_NONBLOCK
);
774 struct dpif_linux
*dpif
= xmalloc(sizeof *dpif
);
775 error
= rtnetlink_notifier_register(&dpif
->port_notifier
,
776 dpif_linux_port_changed
, dpif
);
780 name
= xasprintf("dp%d", minor
);
781 dpif_init(&dpif
->dpif
, &dpif_linux_class
, name
, minor
, minor
);
785 dpif
->local_ifname
= NULL
;
787 dpif
->local_ifindex
= 0;
788 svec_init(&dpif
->changed_ports
);
789 dpif
->change_error
= false;
790 *dpifp
= &dpif
->dpif
;
796 VLOG_WARN("%s: open failed (%s)", fn
, strerror(error
));
804 dpif_linux_port_changed(const struct rtnetlink_change
*change
, void *dpif_
)
806 struct dpif_linux
*dpif
= dpif_
;
809 if (change
->master_ifindex
== dpif
->local_ifindex
810 && (change
->nlmsg_type
== RTM_NEWLINK
811 || change
->nlmsg_type
== RTM_DELLINK
))
813 /* Our datapath changed, either adding a new port or deleting an
815 if (!svec_contains(&dpif
->changed_ports
, change
->ifname
)) {
816 svec_add(&dpif
->changed_ports
, change
->ifname
);
817 svec_sort(&dpif
->changed_ports
);
821 dpif
->change_error
= true;