2 * Copyright (c) 2008, 2009, 2010, 2011, 2012, 2013, 2014, 2015 Nicira, Inc.
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.
19 #include "dpif-linux.h"
26 #include <linux/types.h>
27 #include <linux/pkt_sched.h>
31 #include <sys/epoll.h>
36 #include "dpif-provider.h"
37 #include "dynamic-string.h"
39 #include "fat-rwlock.h"
41 #include "netdev-linux.h"
42 #include "netdev-vport.h"
43 #include "netlink-notifier.h"
44 #include "netlink-socket.h"
49 #include "poll-loop.h"
54 #include "unaligned.h"
58 VLOG_DEFINE_THIS_MODULE(dpif_linux);
59 enum { MAX_PORTS = USHRT_MAX };
61 /* This ethtool flag was introduced in Linux 2.6.24, so it might be
62 * missing if we have old headers. */
63 #define ETH_FLAG_LRO (1 << 15) /* LRO is enabled */
65 struct dpif_linux_dp {
66 /* Generic Netlink header. */
69 /* struct ovs_header. */
73 const char *name; /* OVS_DP_ATTR_NAME. */
74 const uint32_t *upcall_pid; /* OVS_DP_ATTR_UPCALL_PID. */
75 uint32_t user_features; /* OVS_DP_ATTR_USER_FEATURES */
76 struct ovs_dp_stats stats; /* OVS_DP_ATTR_STATS. */
77 struct ovs_dp_megaflow_stats megaflow_stats;
78 /* OVS_DP_ATTR_MEGAFLOW_STATS.*/
81 static void dpif_linux_dp_init(struct dpif_linux_dp *);
82 static int dpif_linux_dp_from_ofpbuf(struct dpif_linux_dp *,
83 const struct ofpbuf *);
84 static void dpif_linux_dp_dump_start(struct nl_dump *);
85 static int dpif_linux_dp_transact(const struct dpif_linux_dp *request,
86 struct dpif_linux_dp *reply,
87 struct ofpbuf **bufp);
88 static int dpif_linux_dp_get(const struct dpif *, struct dpif_linux_dp *reply,
89 struct ofpbuf **bufp);
91 struct dpif_linux_flow {
92 /* Generic Netlink header. */
95 /* struct ovs_header. */
96 unsigned int nlmsg_flags;
101 * The 'stats' member points to 64-bit data that might only be aligned on
102 * 32-bit boundaries, so get_unaligned_u64() should be used to access its
105 * If 'actions' is nonnull then OVS_FLOW_ATTR_ACTIONS will be included in
106 * the Netlink version of the command, even if actions_len is zero. */
107 const struct nlattr *key; /* OVS_FLOW_ATTR_KEY. */
109 const struct nlattr *mask; /* OVS_FLOW_ATTR_MASK. */
111 const struct nlattr *actions; /* OVS_FLOW_ATTR_ACTIONS. */
113 const struct ovs_flow_stats *stats; /* OVS_FLOW_ATTR_STATS. */
114 const uint8_t *tcp_flags; /* OVS_FLOW_ATTR_TCP_FLAGS. */
115 const ovs_32aligned_u64 *used; /* OVS_FLOW_ATTR_USED. */
116 bool clear; /* OVS_FLOW_ATTR_CLEAR. */
119 static void dpif_linux_flow_init(struct dpif_linux_flow *);
120 static int dpif_linux_flow_from_ofpbuf(struct dpif_linux_flow *,
121 const struct ofpbuf *);
122 static void dpif_linux_flow_to_ofpbuf(const struct dpif_linux_flow *,
124 static int dpif_linux_flow_transact(struct dpif_linux_flow *request,
125 struct dpif_linux_flow *reply,
126 struct ofpbuf **bufp);
127 static void dpif_linux_flow_get_stats(const struct dpif_linux_flow *,
128 struct dpif_flow_stats *);
130 /* One of the dpif channels between the kernel and userspace. */
131 struct dpif_channel {
132 struct nl_sock *sock; /* Netlink socket. */
133 long long int last_poll; /* Last time this channel was polled. */
136 struct dpif_handler {
137 struct dpif_channel *channels;/* Array of channels for each handler. */
138 struct epoll_event *epoll_events;
139 int epoll_fd; /* epoll fd that includes channel socks. */
140 int n_events; /* Num events returned by epoll_wait(). */
141 int event_offset; /* Offset into 'epoll_events'. */
144 /* Datapath interface for the openvswitch Linux kernel module. */
149 /* Upcall messages. */
150 struct fat_rwlock upcall_lock;
151 struct dpif_handler *handlers;
152 uint32_t n_handlers; /* Num of upcall handlers. */
153 int uc_array_size; /* Size of 'handler->channels' and */
154 /* 'handler->epoll_events'. */
156 /* Change notification. */
157 struct nl_sock *port_notifier; /* vport multicast group subscriber. */
158 bool refresh_channels;
161 static void report_loss(struct dpif_linux *, struct dpif_channel *,
162 uint32_t ch_idx, uint32_t handler_id);
164 static struct vlog_rate_limit error_rl = VLOG_RATE_LIMIT_INIT(9999, 5);
166 /* Generic Netlink family numbers for OVS.
168 * Initialized by dpif_linux_init(). */
169 static int ovs_datapath_family;
170 static int ovs_vport_family;
171 static int ovs_flow_family;
172 static int ovs_packet_family;
174 /* Generic Netlink multicast groups for OVS.
176 * Initialized by dpif_linux_init(). */
177 static unsigned int ovs_vport_mcgroup;
179 static int dpif_linux_init(void);
180 static int open_dpif(const struct dpif_linux_dp *, struct dpif **);
181 static uint32_t dpif_linux_port_get_pid(const struct dpif *,
182 odp_port_t port_no, uint32_t hash);
183 static int dpif_linux_refresh_channels(struct dpif_linux *,
184 uint32_t n_handlers);
185 static void dpif_linux_vport_to_ofpbuf(const struct dpif_linux_vport *,
187 static int dpif_linux_vport_from_ofpbuf(struct dpif_linux_vport *,
188 const struct ofpbuf *);
190 static struct dpif_linux *
191 dpif_linux_cast(const struct dpif *dpif)
193 dpif_assert_class(dpif, &dpif_linux_class);
194 return CONTAINER_OF(dpif, struct dpif_linux, dpif);
198 dpif_linux_enumerate(struct sset *all_dps)
201 uint64_t reply_stub[NL_DUMP_BUFSIZE / 8];
202 struct ofpbuf msg, buf;
205 error = dpif_linux_init();
210 ofpbuf_use_stub(&buf, reply_stub, sizeof reply_stub);
211 dpif_linux_dp_dump_start(&dump);
212 while (nl_dump_next(&dump, &msg, &buf)) {
213 struct dpif_linux_dp dp;
215 if (!dpif_linux_dp_from_ofpbuf(&dp, &msg)) {
216 sset_add(all_dps, dp.name);
220 return nl_dump_done(&dump);
224 dpif_linux_open(const struct dpif_class *class OVS_UNUSED, const char *name,
225 bool create, struct dpif **dpifp)
227 struct dpif_linux_dp dp_request, dp;
232 error = dpif_linux_init();
237 /* Create or look up datapath. */
238 dpif_linux_dp_init(&dp_request);
240 dp_request.cmd = OVS_DP_CMD_NEW;
242 dp_request.upcall_pid = &upcall_pid;
244 /* Use OVS_DP_CMD_SET to report user features */
245 dp_request.cmd = OVS_DP_CMD_SET;
247 dp_request.name = name;
248 dp_request.user_features |= OVS_DP_F_UNALIGNED;
249 dp_request.user_features |= OVS_DP_F_VPORT_PIDS;
250 error = dpif_linux_dp_transact(&dp_request, &dp, &buf);
255 error = open_dpif(&dp, dpifp);
261 open_dpif(const struct dpif_linux_dp *dp, struct dpif **dpifp)
263 struct dpif_linux *dpif;
265 dpif = xzalloc(sizeof *dpif);
266 dpif->port_notifier = NULL;
267 fat_rwlock_init(&dpif->upcall_lock);
269 dpif_init(&dpif->dpif, &dpif_linux_class, dp->name,
270 dp->dp_ifindex, dp->dp_ifindex);
272 dpif->dp_ifindex = dp->dp_ifindex;
273 *dpifp = &dpif->dpif;
278 /* Destroys the netlink sockets pointed by the elements in 'socksp'
279 * and frees the 'socksp'. */
281 vport_del_socksp(struct nl_sock **socksp, uint32_t n_socks)
285 for (i = 0; i < n_socks; i++) {
286 nl_sock_destroy(socksp[i]);
292 /* Creates an array of netlink sockets. Returns an array of the
293 * corresponding pointers. Records the error in 'error'. */
294 static struct nl_sock **
295 vport_create_socksp(uint32_t n_socks, int *error)
297 struct nl_sock **socksp = xzalloc(n_socks * sizeof *socksp);
300 for (i = 0; i < n_socks; i++) {
301 *error = nl_sock_create(NETLINK_GENERIC, &socksp[i]);
310 vport_del_socksp(socksp, n_socks);
315 /* Given the array of pointers to netlink sockets 'socksp', returns
316 * the array of corresponding pids. If the 'socksp' is NULL, returns
317 * a single-element array of value 0. */
319 vport_socksp_to_pids(struct nl_sock **socksp, uint32_t n_socks)
324 pids = xzalloc(sizeof *pids);
328 pids = xzalloc(n_socks * sizeof *pids);
329 for (i = 0; i < n_socks; i++) {
330 pids[i] = nl_sock_pid(socksp[i]);
337 /* Given the port number 'port_idx', extracts the pids of netlink sockets
338 * associated to the port and assigns it to 'upcall_pids'. */
340 vport_get_pids(struct dpif_linux *dpif, uint32_t port_idx,
341 uint32_t **upcall_pids)
346 /* Since the nl_sock can only be assigned in either all
347 * or none "dpif->handlers" channels, the following check
349 if (!dpif->handlers[0].channels[port_idx].sock) {
353 pids = xzalloc(dpif->n_handlers * sizeof *pids);
355 for (i = 0; i < dpif->n_handlers; i++) {
356 pids[i] = nl_sock_pid(dpif->handlers[i].channels[port_idx].sock);
365 vport_add_channels(struct dpif_linux *dpif, odp_port_t port_no,
366 struct nl_sock **socksp)
368 struct epoll_event event;
369 uint32_t port_idx = odp_to_u32(port_no);
373 if (dpif->handlers == NULL) {
377 /* We assume that the datapath densely chooses port numbers, which can
378 * therefore be used as an index into 'channels' and 'epoll_events' of
379 * 'dpif->handler'. */
380 if (port_idx >= dpif->uc_array_size) {
381 uint32_t new_size = port_idx + 1;
383 if (new_size > MAX_PORTS) {
384 VLOG_WARN_RL(&error_rl, "%s: datapath port %"PRIu32" too big",
385 dpif_name(&dpif->dpif), port_no);
389 for (i = 0; i < dpif->n_handlers; i++) {
390 struct dpif_handler *handler = &dpif->handlers[i];
392 handler->channels = xrealloc(handler->channels,
393 new_size * sizeof *handler->channels);
395 for (j = dpif->uc_array_size; j < new_size; j++) {
396 handler->channels[j].sock = NULL;
399 handler->epoll_events = xrealloc(handler->epoll_events,
400 new_size * sizeof *handler->epoll_events);
403 dpif->uc_array_size = new_size;
406 memset(&event, 0, sizeof event);
407 event.events = EPOLLIN;
408 event.data.u32 = port_idx;
410 for (i = 0; i < dpif->n_handlers; i++) {
411 struct dpif_handler *handler = &dpif->handlers[i];
413 if (epoll_ctl(handler->epoll_fd, EPOLL_CTL_ADD, nl_sock_fd(socksp[i]),
418 dpif->handlers[i].channels[port_idx].sock = socksp[i];
419 dpif->handlers[i].channels[port_idx].last_poll = LLONG_MIN;
425 for (j = 0; j < i; j++) {
426 epoll_ctl(dpif->handlers[j].epoll_fd, EPOLL_CTL_DEL,
427 nl_sock_fd(socksp[j]), NULL);
428 dpif->handlers[j].channels[port_idx].sock = NULL;
435 vport_del_channels(struct dpif_linux *dpif, odp_port_t port_no)
437 uint32_t port_idx = odp_to_u32(port_no);
440 if (!dpif->handlers || port_idx >= dpif->uc_array_size) {
444 /* Since the sock can only be assigned in either all or none
445 * of "dpif->handlers" channels, the following check would
447 if (!dpif->handlers[0].channels[port_idx].sock) {
451 for (i = 0; i < dpif->n_handlers; i++) {
452 struct dpif_handler *handler = &dpif->handlers[i];
454 epoll_ctl(handler->epoll_fd, EPOLL_CTL_DEL,
455 nl_sock_fd(handler->channels[port_idx].sock), NULL);
456 nl_sock_destroy(handler->channels[port_idx].sock);
457 handler->channels[port_idx].sock = NULL;
458 handler->event_offset = handler->n_events = 0;
463 destroy_all_channels(struct dpif_linux *dpif) OVS_REQ_WRLOCK(dpif->upcall_lock)
467 if (!dpif->handlers) {
471 for (i = 0; i < dpif->uc_array_size; i++ ) {
472 struct dpif_linux_vport vport_request;
473 uint32_t upcall_pids = 0;
475 /* Since the sock can only be assigned in either all or none
476 * of "dpif->handlers" channels, the following check would
478 if (!dpif->handlers[0].channels[i].sock) {
482 /* Turn off upcalls. */
483 dpif_linux_vport_init(&vport_request);
484 vport_request.cmd = OVS_VPORT_CMD_SET;
485 vport_request.dp_ifindex = dpif->dp_ifindex;
486 vport_request.port_no = u32_to_odp(i);
487 vport_request.n_upcall_pids = 1;
488 vport_request.upcall_pids = &upcall_pids;
489 dpif_linux_vport_transact(&vport_request, NULL, NULL);
491 vport_del_channels(dpif, u32_to_odp(i));
494 for (i = 0; i < dpif->n_handlers; i++) {
495 struct dpif_handler *handler = &dpif->handlers[i];
497 close(handler->epoll_fd);
498 free(handler->epoll_events);
499 free(handler->channels);
502 free(dpif->handlers);
503 dpif->handlers = NULL;
504 dpif->n_handlers = 0;
505 dpif->uc_array_size = 0;
509 dpif_linux_close(struct dpif *dpif_)
511 struct dpif_linux *dpif = dpif_linux_cast(dpif_);
513 nl_sock_destroy(dpif->port_notifier);
515 fat_rwlock_wrlock(&dpif->upcall_lock);
516 destroy_all_channels(dpif);
517 fat_rwlock_unlock(&dpif->upcall_lock);
519 fat_rwlock_destroy(&dpif->upcall_lock);
524 dpif_linux_destroy(struct dpif *dpif_)
526 struct dpif_linux *dpif = dpif_linux_cast(dpif_);
527 struct dpif_linux_dp dp;
529 dpif_linux_dp_init(&dp);
530 dp.cmd = OVS_DP_CMD_DEL;
531 dp.dp_ifindex = dpif->dp_ifindex;
532 return dpif_linux_dp_transact(&dp, NULL, NULL);
536 dpif_linux_run(struct dpif *dpif_)
538 struct dpif_linux *dpif = dpif_linux_cast(dpif_);
540 if (dpif->refresh_channels) {
541 dpif->refresh_channels = false;
542 fat_rwlock_wrlock(&dpif->upcall_lock);
543 dpif_linux_refresh_channels(dpif, dpif->n_handlers);
544 fat_rwlock_unlock(&dpif->upcall_lock);
549 dpif_linux_get_stats(const struct dpif *dpif_, struct dpif_dp_stats *stats)
551 struct dpif_linux_dp dp;
555 error = dpif_linux_dp_get(dpif_, &dp, &buf);
557 stats->n_hit = dp.stats.n_hit;
558 stats->n_missed = dp.stats.n_missed;
559 stats->n_lost = dp.stats.n_lost;
560 stats->n_flows = dp.stats.n_flows;
561 stats->n_masks = dp.megaflow_stats.n_masks;
562 stats->n_mask_hit = dp.megaflow_stats.n_mask_hit;
569 get_vport_type(const struct dpif_linux_vport *vport)
571 static struct vlog_rate_limit rl = VLOG_RATE_LIMIT_INIT(5, 20);
573 switch (vport->type) {
574 case OVS_VPORT_TYPE_NETDEV: {
575 const char *type = netdev_get_type_from_name(vport->name);
577 return type ? type : "system";
580 case OVS_VPORT_TYPE_INTERNAL:
583 case OVS_VPORT_TYPE_GRE:
586 case OVS_VPORT_TYPE_GRE64:
589 case OVS_VPORT_TYPE_VXLAN:
592 case OVS_VPORT_TYPE_LISP:
595 case OVS_VPORT_TYPE_UNSPEC:
596 case __OVS_VPORT_TYPE_MAX:
600 VLOG_WARN_RL(&rl, "dp%d: port `%s' has unsupported type %u",
601 vport->dp_ifindex, vport->name, (unsigned int) vport->type);
605 static enum ovs_vport_type
606 netdev_to_ovs_vport_type(const struct netdev *netdev)
608 const char *type = netdev_get_type(netdev);
610 if (!strcmp(type, "tap") || !strcmp(type, "system")) {
611 return OVS_VPORT_TYPE_NETDEV;
612 } else if (!strcmp(type, "internal")) {
613 return OVS_VPORT_TYPE_INTERNAL;
614 } else if (strstr(type, "gre64")) {
615 return OVS_VPORT_TYPE_GRE64;
616 } else if (strstr(type, "gre")) {
617 return OVS_VPORT_TYPE_GRE;
618 } else if (!strcmp(type, "vxlan")) {
619 return OVS_VPORT_TYPE_VXLAN;
620 } else if (!strcmp(type, "lisp")) {
621 return OVS_VPORT_TYPE_LISP;
623 return OVS_VPORT_TYPE_UNSPEC;
628 dpif_linux_port_add__(struct dpif_linux *dpif, struct netdev *netdev,
629 odp_port_t *port_nop)
630 OVS_REQ_WRLOCK(dpif->upcall_lock)
632 const struct netdev_tunnel_config *tnl_cfg;
633 char namebuf[NETDEV_VPORT_NAME_BUFSIZE];
634 const char *name = netdev_vport_get_dpif_port(netdev,
635 namebuf, sizeof namebuf);
636 const char *type = netdev_get_type(netdev);
637 struct dpif_linux_vport request, reply;
639 uint64_t options_stub[64 / 8];
640 struct ofpbuf options;
641 struct nl_sock **socksp = NULL;
642 uint32_t *upcall_pids;
645 if (dpif->handlers) {
646 socksp = vport_create_socksp(dpif->n_handlers, &error);
652 dpif_linux_vport_init(&request);
653 request.cmd = OVS_VPORT_CMD_NEW;
654 request.dp_ifindex = dpif->dp_ifindex;
655 request.type = netdev_to_ovs_vport_type(netdev);
656 if (request.type == OVS_VPORT_TYPE_UNSPEC) {
657 VLOG_WARN_RL(&error_rl, "%s: cannot create port `%s' because it has "
658 "unsupported type `%s'",
659 dpif_name(&dpif->dpif), name, type);
660 vport_del_socksp(socksp, dpif->n_handlers);
665 if (request.type == OVS_VPORT_TYPE_NETDEV) {
666 netdev_linux_ethtool_set_flag(netdev, ETH_FLAG_LRO, "LRO", false);
669 tnl_cfg = netdev_get_tunnel_config(netdev);
670 if (tnl_cfg && tnl_cfg->dst_port != 0) {
671 ofpbuf_use_stack(&options, options_stub, sizeof options_stub);
672 nl_msg_put_u16(&options, OVS_TUNNEL_ATTR_DST_PORT,
673 ntohs(tnl_cfg->dst_port));
674 request.options = ofpbuf_data(&options);
675 request.options_len = ofpbuf_size(&options);
678 request.port_no = *port_nop;
679 upcall_pids = vport_socksp_to_pids(socksp, dpif->n_handlers);
680 request.n_upcall_pids = socksp ? dpif->n_handlers : 1;
681 request.upcall_pids = upcall_pids;
683 error = dpif_linux_vport_transact(&request, &reply, &buf);
685 *port_nop = reply.port_no;
687 if (error == EBUSY && *port_nop != ODPP_NONE) {
688 VLOG_INFO("%s: requested port %"PRIu32" is in use",
689 dpif_name(&dpif->dpif), *port_nop);
692 vport_del_socksp(socksp, dpif->n_handlers);
697 error = vport_add_channels(dpif, *port_nop, socksp);
699 VLOG_INFO("%s: could not add channel for port %s",
700 dpif_name(&dpif->dpif), name);
702 /* Delete the port. */
703 dpif_linux_vport_init(&request);
704 request.cmd = OVS_VPORT_CMD_DEL;
705 request.dp_ifindex = dpif->dp_ifindex;
706 request.port_no = *port_nop;
707 dpif_linux_vport_transact(&request, NULL, NULL);
708 vport_del_socksp(socksp, dpif->n_handlers);
722 dpif_linux_port_add(struct dpif *dpif_, struct netdev *netdev,
723 odp_port_t *port_nop)
725 struct dpif_linux *dpif = dpif_linux_cast(dpif_);
728 fat_rwlock_wrlock(&dpif->upcall_lock);
729 error = dpif_linux_port_add__(dpif, netdev, port_nop);
730 fat_rwlock_unlock(&dpif->upcall_lock);
736 dpif_linux_port_del__(struct dpif_linux *dpif, odp_port_t port_no)
737 OVS_REQ_WRLOCK(dpif->upcall_lock)
739 struct dpif_linux_vport vport;
742 dpif_linux_vport_init(&vport);
743 vport.cmd = OVS_VPORT_CMD_DEL;
744 vport.dp_ifindex = dpif->dp_ifindex;
745 vport.port_no = port_no;
746 error = dpif_linux_vport_transact(&vport, NULL, NULL);
748 vport_del_channels(dpif, port_no);
754 dpif_linux_port_del(struct dpif *dpif_, odp_port_t port_no)
756 struct dpif_linux *dpif = dpif_linux_cast(dpif_);
759 fat_rwlock_wrlock(&dpif->upcall_lock);
760 error = dpif_linux_port_del__(dpif, port_no);
761 fat_rwlock_unlock(&dpif->upcall_lock);
767 dpif_linux_port_query__(const struct dpif_linux *dpif, odp_port_t port_no,
768 const char *port_name, struct dpif_port *dpif_port)
770 struct dpif_linux_vport request;
771 struct dpif_linux_vport reply;
775 dpif_linux_vport_init(&request);
776 request.cmd = OVS_VPORT_CMD_GET;
777 request.dp_ifindex = dpif->dp_ifindex;
778 request.port_no = port_no;
779 request.name = port_name;
781 error = dpif_linux_vport_transact(&request, &reply, &buf);
783 if (reply.dp_ifindex != request.dp_ifindex) {
784 /* A query by name reported that 'port_name' is in some datapath
785 * other than 'dpif', but the caller wants to know about 'dpif'. */
787 } else if (dpif_port) {
788 dpif_port->name = xstrdup(reply.name);
789 dpif_port->type = xstrdup(get_vport_type(&reply));
790 dpif_port->port_no = reply.port_no;
798 dpif_linux_port_query_by_number(const struct dpif *dpif_, odp_port_t port_no,
799 struct dpif_port *dpif_port)
801 struct dpif_linux *dpif = dpif_linux_cast(dpif_);
803 return dpif_linux_port_query__(dpif, port_no, NULL, dpif_port);
807 dpif_linux_port_query_by_name(const struct dpif *dpif_, const char *devname,
808 struct dpif_port *dpif_port)
810 struct dpif_linux *dpif = dpif_linux_cast(dpif_);
812 return dpif_linux_port_query__(dpif, 0, devname, dpif_port);
816 dpif_linux_port_get_pid__(const struct dpif_linux *dpif, odp_port_t port_no,
818 OVS_REQ_RDLOCK(dpif->upcall_lock)
820 uint32_t port_idx = odp_to_u32(port_no);
823 if (dpif->handlers && dpif->uc_array_size > 0) {
824 /* The ODPP_NONE "reserved" port number uses the "ovs-system"'s
825 * channel, since it is not heavily loaded. */
826 uint32_t idx = port_idx >= dpif->uc_array_size ? 0 : port_idx;
827 struct dpif_handler *h = &dpif->handlers[hash % dpif->n_handlers];
829 /* Needs to check in case the socket pointer is changed in between
830 * the holding of upcall_lock. A known case happens when the main
831 * thread deletes the vport while the handler thread is handling
832 * the upcall from that port. */
833 if (h->channels[idx].sock) {
834 pid = nl_sock_pid(h->channels[idx].sock);
842 dpif_linux_port_get_pid(const struct dpif *dpif_, odp_port_t port_no,
845 const struct dpif_linux *dpif = dpif_linux_cast(dpif_);
848 fat_rwlock_rdlock(&dpif->upcall_lock);
849 ret = dpif_linux_port_get_pid__(dpif, port_no, hash);
850 fat_rwlock_unlock(&dpif->upcall_lock);
856 dpif_linux_flow_flush(struct dpif *dpif_)
858 const struct dpif_linux *dpif = dpif_linux_cast(dpif_);
859 struct dpif_linux_flow flow;
861 dpif_linux_flow_init(&flow);
862 flow.cmd = OVS_FLOW_CMD_DEL;
863 flow.dp_ifindex = dpif->dp_ifindex;
864 return dpif_linux_flow_transact(&flow, NULL, NULL);
867 struct dpif_linux_port_state {
873 dpif_linux_port_dump_start__(const struct dpif_linux *dpif,
874 struct nl_dump *dump)
876 struct dpif_linux_vport request;
879 dpif_linux_vport_init(&request);
880 request.cmd = OVS_VPORT_CMD_GET;
881 request.dp_ifindex = dpif->dp_ifindex;
883 buf = ofpbuf_new(1024);
884 dpif_linux_vport_to_ofpbuf(&request, buf);
885 nl_dump_start(dump, NETLINK_GENERIC, buf);
890 dpif_linux_port_dump_start(const struct dpif *dpif_, void **statep)
892 struct dpif_linux *dpif = dpif_linux_cast(dpif_);
893 struct dpif_linux_port_state *state;
895 *statep = state = xmalloc(sizeof *state);
896 dpif_linux_port_dump_start__(dpif, &state->dump);
898 ofpbuf_init(&state->buf, NL_DUMP_BUFSIZE);
903 dpif_linux_port_dump_next__(const struct dpif_linux *dpif, struct nl_dump *dump,
904 struct dpif_linux_vport *vport,
905 struct ofpbuf *buffer)
910 if (!nl_dump_next(dump, &buf, buffer)) {
914 error = dpif_linux_vport_from_ofpbuf(vport, &buf);
916 VLOG_WARN_RL(&error_rl, "%s: failed to parse vport record (%s)",
917 dpif_name(&dpif->dpif), ovs_strerror(error));
923 dpif_linux_port_dump_next(const struct dpif *dpif_, void *state_,
924 struct dpif_port *dpif_port)
926 struct dpif_linux *dpif = dpif_linux_cast(dpif_);
927 struct dpif_linux_port_state *state = state_;
928 struct dpif_linux_vport vport;
931 error = dpif_linux_port_dump_next__(dpif, &state->dump, &vport,
936 dpif_port->name = CONST_CAST(char *, vport.name);
937 dpif_port->type = CONST_CAST(char *, get_vport_type(&vport));
938 dpif_port->port_no = vport.port_no;
943 dpif_linux_port_dump_done(const struct dpif *dpif_ OVS_UNUSED, void *state_)
945 struct dpif_linux_port_state *state = state_;
946 int error = nl_dump_done(&state->dump);
948 ofpbuf_uninit(&state->buf);
954 dpif_linux_port_poll(const struct dpif *dpif_, char **devnamep)
956 struct dpif_linux *dpif = dpif_linux_cast(dpif_);
958 /* Lazily create the Netlink socket to listen for notifications. */
959 if (!dpif->port_notifier) {
960 struct nl_sock *sock;
963 error = nl_sock_create(NETLINK_GENERIC, &sock);
968 error = nl_sock_join_mcgroup(sock, ovs_vport_mcgroup);
970 nl_sock_destroy(sock);
973 dpif->port_notifier = sock;
975 /* We have no idea of the current state so report that everything
981 static struct vlog_rate_limit rl = VLOG_RATE_LIMIT_INIT(1, 5);
982 uint64_t buf_stub[4096 / 8];
986 ofpbuf_use_stub(&buf, buf_stub, sizeof buf_stub);
987 error = nl_sock_recv(dpif->port_notifier, &buf, false);
989 struct dpif_linux_vport vport;
991 error = dpif_linux_vport_from_ofpbuf(&vport, &buf);
993 if (vport.dp_ifindex == dpif->dp_ifindex
994 && (vport.cmd == OVS_VPORT_CMD_NEW
995 || vport.cmd == OVS_VPORT_CMD_DEL
996 || vport.cmd == OVS_VPORT_CMD_SET)) {
997 VLOG_DBG("port_changed: dpif:%s vport:%s cmd:%"PRIu8,
998 dpif->dpif.full_name, vport.name, vport.cmd);
999 if (vport.cmd == OVS_VPORT_CMD_DEL && dpif->handlers) {
1000 dpif->refresh_channels = true;
1002 *devnamep = xstrdup(vport.name);
1003 ofpbuf_uninit(&buf);
1007 } else if (error != EAGAIN) {
1008 VLOG_WARN_RL(&rl, "error reading or parsing netlink (%s)",
1009 ovs_strerror(error));
1010 nl_sock_drain(dpif->port_notifier);
1014 ofpbuf_uninit(&buf);
1022 dpif_linux_port_poll_wait(const struct dpif *dpif_)
1024 const struct dpif_linux *dpif = dpif_linux_cast(dpif_);
1026 if (dpif->port_notifier) {
1027 nl_sock_wait(dpif->port_notifier, POLLIN);
1029 poll_immediate_wake();
1034 dpif_linux_flow_get__(const struct dpif_linux *dpif,
1035 const struct nlattr *key, size_t key_len,
1036 struct dpif_linux_flow *reply, struct ofpbuf **bufp)
1038 struct dpif_linux_flow request;
1040 dpif_linux_flow_init(&request);
1041 request.cmd = OVS_FLOW_CMD_GET;
1042 request.dp_ifindex = dpif->dp_ifindex;
1044 request.key_len = key_len;
1045 return dpif_linux_flow_transact(&request, reply, bufp);
1049 dpif_linux_flow_get(const struct dpif *dpif_,
1050 const struct nlattr *key, size_t key_len,
1051 struct ofpbuf **bufp,
1052 struct nlattr **maskp, size_t *mask_len,
1053 struct nlattr **actionsp, size_t *actions_len,
1054 struct dpif_flow_stats *stats)
1056 const struct dpif_linux *dpif = dpif_linux_cast(dpif_);
1057 struct dpif_linux_flow reply;
1060 error = dpif_linux_flow_get__(dpif, key, key_len, &reply, bufp);
1063 *maskp = CONST_CAST(struct nlattr *, reply.mask);
1064 *mask_len = reply.mask_len;
1067 *actionsp = CONST_CAST(struct nlattr *, reply.actions);
1068 *actions_len = reply.actions_len;
1071 dpif_linux_flow_get_stats(&reply, stats);
1078 dpif_linux_init_flow_put(struct dpif_linux *dpif, const struct dpif_flow_put *put,
1079 struct dpif_linux_flow *request)
1081 static const struct nlattr dummy_action;
1083 dpif_linux_flow_init(request);
1084 request->cmd = (put->flags & DPIF_FP_CREATE
1085 ? OVS_FLOW_CMD_NEW : OVS_FLOW_CMD_SET);
1086 request->dp_ifindex = dpif->dp_ifindex;
1087 request->key = put->key;
1088 request->key_len = put->key_len;
1089 request->mask = put->mask;
1090 request->mask_len = put->mask_len;
1091 /* Ensure that OVS_FLOW_ATTR_ACTIONS will always be included. */
1092 request->actions = (put->actions
1094 : CONST_CAST(struct nlattr *, &dummy_action));
1095 request->actions_len = put->actions_len;
1096 if (put->flags & DPIF_FP_ZERO_STATS) {
1097 request->clear = true;
1099 request->nlmsg_flags = put->flags & DPIF_FP_MODIFY ? 0 : NLM_F_CREATE;
1103 dpif_linux_flow_put(struct dpif *dpif_, const struct dpif_flow_put *put)
1105 struct dpif_linux *dpif = dpif_linux_cast(dpif_);
1106 struct dpif_linux_flow request, reply;
1110 dpif_linux_init_flow_put(dpif, put, &request);
1111 error = dpif_linux_flow_transact(&request,
1112 put->stats ? &reply : NULL,
1113 put->stats ? &buf : NULL);
1114 if (!error && put->stats) {
1115 dpif_linux_flow_get_stats(&reply, put->stats);
1122 dpif_linux_init_flow_del(struct dpif_linux *dpif, const struct dpif_flow_del *del,
1123 struct dpif_linux_flow *request)
1125 dpif_linux_flow_init(request);
1126 request->cmd = OVS_FLOW_CMD_DEL;
1127 request->dp_ifindex = dpif->dp_ifindex;
1128 request->key = del->key;
1129 request->key_len = del->key_len;
1133 dpif_linux_flow_del(struct dpif *dpif_, const struct dpif_flow_del *del)
1135 struct dpif_linux *dpif = dpif_linux_cast(dpif_);
1136 struct dpif_linux_flow request, reply;
1140 dpif_linux_init_flow_del(dpif, del, &request);
1141 error = dpif_linux_flow_transact(&request,
1142 del->stats ? &reply : NULL,
1143 del->stats ? &buf : NULL);
1144 if (!error && del->stats) {
1145 dpif_linux_flow_get_stats(&reply, del->stats);
1151 struct dpif_linux_flow_state {
1152 struct dpif_linux_flow flow;
1153 struct dpif_flow_stats stats;
1154 struct ofpbuf buffer; /* Always used to store flows. */
1155 struct ofpbuf *tmp; /* Used if kernel does not supply actions. */
1158 struct dpif_linux_flow_iter {
1159 struct nl_dump dump;
1164 dpif_linux_flow_dump_state_init(void **statep)
1166 struct dpif_linux_flow_state *state;
1168 *statep = state = xmalloc(sizeof *state);
1169 ofpbuf_init(&state->buffer, NL_DUMP_BUFSIZE);
1174 dpif_linux_flow_dump_state_uninit(void *state_)
1176 struct dpif_linux_flow_state *state = state_;
1178 ofpbuf_uninit(&state->buffer);
1179 ofpbuf_delete(state->tmp);
1184 dpif_linux_flow_dump_start(const struct dpif *dpif_, void **iterp)
1186 const struct dpif_linux *dpif = dpif_linux_cast(dpif_);
1187 struct dpif_linux_flow_iter *iter;
1188 struct dpif_linux_flow request;
1191 *iterp = iter = xmalloc(sizeof *iter);
1193 dpif_linux_flow_init(&request);
1194 request.cmd = OVS_FLOW_CMD_GET;
1195 request.dp_ifindex = dpif->dp_ifindex;
1197 buf = ofpbuf_new(1024);
1198 dpif_linux_flow_to_ofpbuf(&request, buf);
1199 nl_dump_start(&iter->dump, NETLINK_GENERIC, buf);
1201 atomic_init(&iter->status, 0);
1207 dpif_linux_flow_dump_next(const struct dpif *dpif_, void *iter_, void *state_,
1208 const struct nlattr **key, size_t *key_len,
1209 const struct nlattr **mask, size_t *mask_len,
1210 const struct nlattr **actions, size_t *actions_len,
1211 const struct dpif_flow_stats **stats)
1213 const struct dpif_linux *dpif = dpif_linux_cast(dpif_);
1214 struct dpif_linux_flow_iter *iter = iter_;
1215 struct dpif_linux_flow_state *state = state_;
1220 ofpbuf_delete(state->tmp);
1223 if (!nl_dump_next(&iter->dump, &buf, &state->buffer)) {
1227 error = dpif_linux_flow_from_ofpbuf(&state->flow, &buf);
1229 atomic_store(&iter->status, error);
1233 if (actions && !state->flow.actions) {
1234 struct dpif_linux_flow reply;
1236 /* Keys are required to be allocated from 'state->buffer' so
1237 * they're preserved across calls. Therefore we need a separate
1238 * reply to prevent them from being overwritten. Actions, however,
1239 * don't have this requirement, so it's that fine they're destroyed
1240 * on the next call. */
1241 error = dpif_linux_flow_get__(dpif, state->flow.key,
1242 state->flow.key_len,
1243 &reply, &state->tmp);
1244 state->flow.actions = reply.actions;
1245 state->flow.actions_len = reply.actions_len;
1247 if (error == ENOENT) {
1248 VLOG_DBG("dumped flow disappeared on get");
1250 VLOG_WARN("error fetching dumped flow: %s",
1251 ovs_strerror(error));
1257 *actions = state->flow.actions;
1258 *actions_len = state->flow.actions_len;
1261 *key = state->flow.key;
1262 *key_len = state->flow.key_len;
1265 *mask = state->flow.mask;
1266 *mask_len = state->flow.mask ? state->flow.mask_len : 0;
1269 dpif_linux_flow_get_stats(&state->flow, &state->stats);
1270 *stats = &state->stats;
1276 dpif_linux_flow_dump_next_may_destroy_keys(void *state_)
1278 struct dpif_linux_flow_state *state = state_;
1279 struct dpif_linux_flow flow;
1280 struct ofpbuf nlmsg;
1282 /* Check whether there's a flow remaining in the buffer that includes
1283 * actions. (If it does not include actions, then we could end up
1284 * destroying keys previously returned trying to retrieve its actions
1286 return (!nl_dump_peek(&nlmsg, &state->buffer)
1287 || dpif_linux_flow_from_ofpbuf(&flow, &nlmsg)
1292 dpif_linux_flow_dump_done(const struct dpif *dpif OVS_UNUSED, void *iter_)
1294 struct dpif_linux_flow_iter *iter = iter_;
1296 unsigned int nl_status = nl_dump_done(&iter->dump);
1298 atomic_read(&iter->status, &dump_status);
1300 return dump_status ? dump_status : nl_status;
1304 dpif_linux_encode_execute(int dp_ifindex, const struct dpif_execute *d_exec,
1307 struct ovs_header *k_exec;
1310 ofpbuf_prealloc_tailroom(buf, (64
1311 + ofpbuf_size(d_exec->packet)
1312 + ODP_KEY_METADATA_SIZE
1313 + d_exec->actions_len));
1315 nl_msg_put_genlmsghdr(buf, 0, ovs_packet_family, NLM_F_REQUEST,
1316 OVS_PACKET_CMD_EXECUTE, OVS_PACKET_VERSION);
1318 k_exec = ofpbuf_put_uninit(buf, sizeof *k_exec);
1319 k_exec->dp_ifindex = dp_ifindex;
1321 nl_msg_put_unspec(buf, OVS_PACKET_ATTR_PACKET,
1322 ofpbuf_data(d_exec->packet),
1323 ofpbuf_size(d_exec->packet));
1325 key_ofs = nl_msg_start_nested(buf, OVS_PACKET_ATTR_KEY);
1326 odp_key_from_pkt_metadata(buf, &d_exec->md);
1327 nl_msg_end_nested(buf, key_ofs);
1329 nl_msg_put_unspec(buf, OVS_PACKET_ATTR_ACTIONS,
1330 d_exec->actions, d_exec->actions_len);
1334 dpif_linux_execute__(int dp_ifindex, const struct dpif_execute *execute)
1336 uint64_t request_stub[1024 / 8];
1337 struct ofpbuf request;
1340 ofpbuf_use_stub(&request, request_stub, sizeof request_stub);
1341 dpif_linux_encode_execute(dp_ifindex, execute, &request);
1342 error = nl_transact(NETLINK_GENERIC, &request, NULL);
1343 ofpbuf_uninit(&request);
1349 dpif_linux_execute(struct dpif *dpif_, struct dpif_execute *execute)
1351 const struct dpif_linux *dpif = dpif_linux_cast(dpif_);
1353 return dpif_linux_execute__(dpif->dp_ifindex, execute);
1356 /* Executes, against 'dpif', up to the first 'n_ops' operations in 'ops'.
1357 * Returns the number actually executed (at least 1, if 'n_ops' is
1360 dpif_linux_operate__(struct dpif_linux *dpif, struct dpif_op **ops, size_t n_ops)
1362 enum { MAX_OPS = 50 };
1365 struct nl_transaction txn;
1367 struct ofpbuf request;
1368 uint64_t request_stub[1024 / 8];
1370 struct ofpbuf reply;
1371 uint64_t reply_stub[1024 / 8];
1374 struct nl_transaction *txnsp[MAX_OPS];
1377 n_ops = MIN(n_ops, MAX_OPS);
1378 for (i = 0; i < n_ops; i++) {
1379 struct op_auxdata *aux = &auxes[i];
1380 struct dpif_op *op = ops[i];
1381 struct dpif_flow_put *put;
1382 struct dpif_flow_del *del;
1383 struct dpif_linux_flow flow;
1385 ofpbuf_use_stub(&aux->request,
1386 aux->request_stub, sizeof aux->request_stub);
1387 aux->txn.request = &aux->request;
1389 ofpbuf_use_stub(&aux->reply, aux->reply_stub, sizeof aux->reply_stub);
1390 aux->txn.reply = NULL;
1393 case DPIF_OP_FLOW_PUT:
1394 put = &op->u.flow_put;
1395 dpif_linux_init_flow_put(dpif, put, &flow);
1397 flow.nlmsg_flags |= NLM_F_ECHO;
1398 aux->txn.reply = &aux->reply;
1400 dpif_linux_flow_to_ofpbuf(&flow, &aux->request);
1403 case DPIF_OP_FLOW_DEL:
1404 del = &op->u.flow_del;
1405 dpif_linux_init_flow_del(dpif, del, &flow);
1407 flow.nlmsg_flags |= NLM_F_ECHO;
1408 aux->txn.reply = &aux->reply;
1410 dpif_linux_flow_to_ofpbuf(&flow, &aux->request);
1413 case DPIF_OP_EXECUTE:
1414 /* Can't execute a packet that won't fit in a Netlink attribute. */
1415 if (OVS_UNLIKELY(nl_attr_oversized(
1416 ofpbuf_size(op->u.execute.packet)))) {
1417 /* Report an error immediately if this is the first operation.
1418 * Otherwise the easiest thing to do is to postpone to the next
1419 * call (when this will be the first operation). */
1421 VLOG_ERR_RL(&error_rl,
1422 "dropping oversized %"PRIu32"-byte packet",
1423 ofpbuf_size(op->u.execute.packet));
1424 op->error = ENOBUFS;
1429 dpif_linux_encode_execute(dpif->dp_ifindex, &op->u.execute,
1440 for (i = 0; i < n_ops; i++) {
1441 txnsp[i] = &auxes[i].txn;
1443 nl_transact_multiple(NETLINK_GENERIC, txnsp, n_ops);
1445 for (i = 0; i < n_ops; i++) {
1446 struct op_auxdata *aux = &auxes[i];
1447 struct nl_transaction *txn = &auxes[i].txn;
1448 struct dpif_op *op = ops[i];
1449 struct dpif_flow_put *put;
1450 struct dpif_flow_del *del;
1452 op->error = txn->error;
1455 case DPIF_OP_FLOW_PUT:
1456 put = &op->u.flow_put;
1459 struct dpif_linux_flow reply;
1461 op->error = dpif_linux_flow_from_ofpbuf(&reply,
1464 dpif_linux_flow_get_stats(&reply, put->stats);
1469 memset(put->stats, 0, sizeof *put->stats);
1474 case DPIF_OP_FLOW_DEL:
1475 del = &op->u.flow_del;
1478 struct dpif_linux_flow reply;
1480 op->error = dpif_linux_flow_from_ofpbuf(&reply,
1483 dpif_linux_flow_get_stats(&reply, del->stats);
1488 memset(del->stats, 0, sizeof *del->stats);
1493 case DPIF_OP_EXECUTE:
1500 ofpbuf_uninit(&aux->request);
1501 ofpbuf_uninit(&aux->reply);
1508 dpif_linux_operate(struct dpif *dpif_, struct dpif_op **ops, size_t n_ops)
1510 struct dpif_linux *dpif = dpif_linux_cast(dpif_);
1513 size_t chunk = dpif_linux_operate__(dpif, ops, n_ops);
1519 /* Synchronizes 'channels' in 'dpif->handlers' with the set of vports
1520 * currently in 'dpif' in the kernel, by adding a new set of channels for
1521 * any kernel vport that lacks one and deleting any channels that have no
1522 * backing kernel vports. */
1524 dpif_linux_refresh_channels(struct dpif_linux *dpif, uint32_t n_handlers)
1525 OVS_REQ_WRLOCK(dpif->upcall_lock)
1527 unsigned long int *keep_channels;
1528 struct dpif_linux_vport vport;
1529 size_t keep_channels_nbits;
1530 struct nl_dump dump;
1531 uint64_t reply_stub[NL_DUMP_BUFSIZE / 8];
1536 if (dpif->n_handlers != n_handlers) {
1537 destroy_all_channels(dpif);
1538 dpif->handlers = xzalloc(n_handlers * sizeof *dpif->handlers);
1539 for (i = 0; i < n_handlers; i++) {
1540 struct dpif_handler *handler = &dpif->handlers[i];
1542 handler->epoll_fd = epoll_create(10);
1543 if (handler->epoll_fd < 0) {
1546 for (j = 0; j < i; j++) {
1547 close(dpif->handlers[j].epoll_fd);
1549 free(dpif->handlers);
1550 dpif->handlers = NULL;
1555 dpif->n_handlers = n_handlers;
1558 for (i = 0; i < n_handlers; i++) {
1559 struct dpif_handler *handler = &dpif->handlers[i];
1561 handler->event_offset = handler->n_events = 0;
1564 keep_channels_nbits = dpif->uc_array_size;
1565 keep_channels = bitmap_allocate(keep_channels_nbits);
1567 ofpbuf_use_stub(&buf, reply_stub, sizeof reply_stub);
1568 dpif_linux_port_dump_start__(dpif, &dump);
1569 while (!dpif_linux_port_dump_next__(dpif, &dump, &vport, &buf)) {
1570 uint32_t port_no = odp_to_u32(vport.port_no);
1571 uint32_t *upcall_pids = NULL;
1574 if (port_no >= dpif->uc_array_size
1575 || !vport_get_pids(dpif, port_no, &upcall_pids)) {
1576 struct nl_sock **socksp = vport_create_socksp(dpif->n_handlers,
1583 error = vport_add_channels(dpif, vport.port_no, socksp);
1585 VLOG_INFO("%s: could not add channels for port %s",
1586 dpif_name(&dpif->dpif), vport.name);
1587 vport_del_socksp(socksp, dpif->n_handlers);
1591 upcall_pids = vport_socksp_to_pids(socksp, dpif->n_handlers);
1595 /* Configure the vport to deliver misses to 'sock'. */
1596 if (vport.upcall_pids[0] == 0
1597 || vport.n_upcall_pids != dpif->n_handlers
1598 || memcmp(upcall_pids, vport.upcall_pids, n_handlers * sizeof
1600 struct dpif_linux_vport vport_request;
1602 dpif_linux_vport_init(&vport_request);
1603 vport_request.cmd = OVS_VPORT_CMD_SET;
1604 vport_request.dp_ifindex = dpif->dp_ifindex;
1605 vport_request.port_no = vport.port_no;
1606 vport_request.n_upcall_pids = dpif->n_handlers;
1607 vport_request.upcall_pids = upcall_pids;
1608 error = dpif_linux_vport_transact(&vport_request, NULL, NULL);
1610 VLOG_WARN_RL(&error_rl,
1611 "%s: failed to set upcall pid on port: %s",
1612 dpif_name(&dpif->dpif), ovs_strerror(error));
1614 if (error != ENODEV && error != ENOENT) {
1617 /* The vport isn't really there, even though the dump says
1618 * it is. Probably we just hit a race after a port
1625 if (port_no < keep_channels_nbits) {
1626 bitmap_set1(keep_channels, port_no);
1633 vport_del_channels(dpif, vport.port_no);
1635 nl_dump_done(&dump);
1636 ofpbuf_uninit(&buf);
1638 /* Discard any saved channels that we didn't reuse. */
1639 for (i = 0; i < keep_channels_nbits; i++) {
1640 if (!bitmap_is_set(keep_channels, i)) {
1641 vport_del_channels(dpif, u32_to_odp(i));
1644 free(keep_channels);
1650 dpif_linux_recv_set__(struct dpif_linux *dpif, bool enable)
1651 OVS_REQ_WRLOCK(dpif->upcall_lock)
1653 if ((dpif->handlers != NULL) == enable) {
1655 } else if (!enable) {
1656 destroy_all_channels(dpif);
1659 return dpif_linux_refresh_channels(dpif, 1);
1664 dpif_linux_recv_set(struct dpif *dpif_, bool enable)
1666 struct dpif_linux *dpif = dpif_linux_cast(dpif_);
1669 fat_rwlock_wrlock(&dpif->upcall_lock);
1670 error = dpif_linux_recv_set__(dpif, enable);
1671 fat_rwlock_unlock(&dpif->upcall_lock);
1677 dpif_linux_handlers_set(struct dpif *dpif_, uint32_t n_handlers)
1679 struct dpif_linux *dpif = dpif_linux_cast(dpif_);
1682 fat_rwlock_wrlock(&dpif->upcall_lock);
1683 if (dpif->handlers) {
1684 error = dpif_linux_refresh_channels(dpif, n_handlers);
1686 fat_rwlock_unlock(&dpif->upcall_lock);
1692 dpif_linux_queue_to_priority(const struct dpif *dpif OVS_UNUSED,
1693 uint32_t queue_id, uint32_t *priority)
1695 if (queue_id < 0xf000) {
1696 *priority = TC_H_MAKE(1 << 16, queue_id + 1);
1704 parse_odp_packet(struct ofpbuf *buf, struct dpif_upcall *upcall,
1707 static const struct nl_policy ovs_packet_policy[] = {
1708 /* Always present. */
1709 [OVS_PACKET_ATTR_PACKET] = { .type = NL_A_UNSPEC,
1710 .min_len = ETH_HEADER_LEN },
1711 [OVS_PACKET_ATTR_KEY] = { .type = NL_A_NESTED },
1713 /* OVS_PACKET_CMD_ACTION only. */
1714 [OVS_PACKET_ATTR_USERDATA] = { .type = NL_A_UNSPEC, .optional = true },
1717 struct ovs_header *ovs_header;
1718 struct nlattr *a[ARRAY_SIZE(ovs_packet_policy)];
1719 struct nlmsghdr *nlmsg;
1720 struct genlmsghdr *genl;
1724 ofpbuf_use_const(&b, ofpbuf_data(buf), ofpbuf_size(buf));
1726 nlmsg = ofpbuf_try_pull(&b, sizeof *nlmsg);
1727 genl = ofpbuf_try_pull(&b, sizeof *genl);
1728 ovs_header = ofpbuf_try_pull(&b, sizeof *ovs_header);
1729 if (!nlmsg || !genl || !ovs_header
1730 || nlmsg->nlmsg_type != ovs_packet_family
1731 || !nl_policy_parse(&b, 0, ovs_packet_policy, a,
1732 ARRAY_SIZE(ovs_packet_policy))) {
1736 type = (genl->cmd == OVS_PACKET_CMD_MISS ? DPIF_UC_MISS
1737 : genl->cmd == OVS_PACKET_CMD_ACTION ? DPIF_UC_ACTION
1743 /* (Re)set ALL fields of '*upcall' on successful return. */
1744 upcall->type = type;
1745 upcall->key = CONST_CAST(struct nlattr *,
1746 nl_attr_get(a[OVS_PACKET_ATTR_KEY]));
1747 upcall->key_len = nl_attr_get_size(a[OVS_PACKET_ATTR_KEY]);
1748 upcall->userdata = a[OVS_PACKET_ATTR_USERDATA];
1750 /* Allow overwriting the netlink attribute header without reallocating. */
1751 ofpbuf_use_stub(&upcall->packet,
1752 CONST_CAST(struct nlattr *,
1753 nl_attr_get(a[OVS_PACKET_ATTR_PACKET])) - 1,
1754 nl_attr_get_size(a[OVS_PACKET_ATTR_PACKET]) +
1755 sizeof(struct nlattr));
1756 ofpbuf_set_data(&upcall->packet,
1757 (char *)ofpbuf_data(&upcall->packet) + sizeof(struct nlattr));
1758 ofpbuf_set_size(&upcall->packet, nl_attr_get_size(a[OVS_PACKET_ATTR_PACKET]));
1760 *dp_ifindex = ovs_header->dp_ifindex;
1766 dpif_linux_recv__(struct dpif_linux *dpif, uint32_t handler_id,
1767 struct dpif_upcall *upcall, struct ofpbuf *buf)
1768 OVS_REQ_RDLOCK(dpif->upcall_lock)
1770 struct dpif_handler *handler;
1773 if (!dpif->handlers || handler_id >= dpif->n_handlers) {
1777 handler = &dpif->handlers[handler_id];
1778 if (handler->event_offset >= handler->n_events) {
1781 handler->event_offset = handler->n_events = 0;
1784 retval = epoll_wait(handler->epoll_fd, handler->epoll_events,
1785 dpif->uc_array_size, 0);
1786 } while (retval < 0 && errno == EINTR);
1788 static struct vlog_rate_limit rl = VLOG_RATE_LIMIT_INIT(1, 1);
1789 VLOG_WARN_RL(&rl, "epoll_wait failed (%s)", ovs_strerror(errno));
1790 } else if (retval > 0) {
1791 handler->n_events = retval;
1795 while (handler->event_offset < handler->n_events) {
1796 int idx = handler->epoll_events[handler->event_offset].data.u32;
1797 struct dpif_channel *ch = &dpif->handlers[handler_id].channels[idx];
1799 handler->event_offset++;
1805 if (++read_tries > 50) {
1809 error = nl_sock_recv(ch->sock, buf, false);
1810 if (error == ENOBUFS) {
1811 /* ENOBUFS typically means that we've received so many
1812 * packets that the buffer overflowed. Try again
1813 * immediately because there's almost certainly a packet
1814 * waiting for us. */
1815 report_loss(dpif, ch, idx, handler_id);
1819 ch->last_poll = time_msec();
1821 if (error == EAGAIN) {
1827 error = parse_odp_packet(buf, upcall, &dp_ifindex);
1828 if (!error && dp_ifindex == dpif->dp_ifindex) {
1840 dpif_linux_recv(struct dpif *dpif_, uint32_t handler_id,
1841 struct dpif_upcall *upcall, struct ofpbuf *buf)
1843 struct dpif_linux *dpif = dpif_linux_cast(dpif_);
1846 fat_rwlock_rdlock(&dpif->upcall_lock);
1847 error = dpif_linux_recv__(dpif, handler_id, upcall, buf);
1848 fat_rwlock_unlock(&dpif->upcall_lock);
1854 dpif_linux_recv_wait__(struct dpif_linux *dpif, uint32_t handler_id)
1855 OVS_REQ_RDLOCK(dpif->upcall_lock)
1857 if (dpif->handlers && handler_id < dpif->n_handlers) {
1858 struct dpif_handler *handler = &dpif->handlers[handler_id];
1860 poll_fd_wait(handler->epoll_fd, POLLIN);
1865 dpif_linux_recv_wait(struct dpif *dpif_, uint32_t handler_id)
1867 struct dpif_linux *dpif = dpif_linux_cast(dpif_);
1869 fat_rwlock_rdlock(&dpif->upcall_lock);
1870 dpif_linux_recv_wait__(dpif, handler_id);
1871 fat_rwlock_unlock(&dpif->upcall_lock);
1875 dpif_linux_recv_purge__(struct dpif_linux *dpif)
1876 OVS_REQ_WRLOCK(dpif->upcall_lock)
1878 if (dpif->handlers) {
1881 for (i = 0; i < dpif->uc_array_size; i++ ) {
1882 if (!dpif->handlers[0].channels[i].sock) {
1886 for (j = 0; j < dpif->n_handlers; j++) {
1887 nl_sock_drain(dpif->handlers[j].channels[i].sock);
1894 dpif_linux_recv_purge(struct dpif *dpif_)
1896 struct dpif_linux *dpif = dpif_linux_cast(dpif_);
1898 fat_rwlock_wrlock(&dpif->upcall_lock);
1899 dpif_linux_recv_purge__(dpif);
1900 fat_rwlock_unlock(&dpif->upcall_lock);
1903 const struct dpif_class dpif_linux_class = {
1905 dpif_linux_enumerate,
1912 dpif_linux_get_stats,
1913 dpif_linux_port_add,
1914 dpif_linux_port_del,
1915 dpif_linux_port_query_by_number,
1916 dpif_linux_port_query_by_name,
1917 dpif_linux_port_get_pid,
1918 dpif_linux_port_dump_start,
1919 dpif_linux_port_dump_next,
1920 dpif_linux_port_dump_done,
1921 dpif_linux_port_poll,
1922 dpif_linux_port_poll_wait,
1923 dpif_linux_flow_get,
1924 dpif_linux_flow_put,
1925 dpif_linux_flow_del,
1926 dpif_linux_flow_flush,
1927 dpif_linux_flow_dump_state_init,
1928 dpif_linux_flow_dump_start,
1929 dpif_linux_flow_dump_next,
1930 dpif_linux_flow_dump_next_may_destroy_keys,
1931 dpif_linux_flow_dump_done,
1932 dpif_linux_flow_dump_state_uninit,
1935 dpif_linux_recv_set,
1936 dpif_linux_handlers_set,
1937 dpif_linux_queue_to_priority,
1939 dpif_linux_recv_wait,
1940 dpif_linux_recv_purge,
1944 dpif_linux_init(void)
1946 static struct ovsthread_once once = OVSTHREAD_ONCE_INITIALIZER;
1949 if (ovsthread_once_start(&once)) {
1950 error = nl_lookup_genl_family(OVS_DATAPATH_FAMILY,
1951 &ovs_datapath_family);
1953 VLOG_ERR("Generic Netlink family '%s' does not exist. "
1954 "The Open vSwitch kernel module is probably not loaded.",
1955 OVS_DATAPATH_FAMILY);
1958 error = nl_lookup_genl_family(OVS_VPORT_FAMILY, &ovs_vport_family);
1961 error = nl_lookup_genl_family(OVS_FLOW_FAMILY, &ovs_flow_family);
1964 error = nl_lookup_genl_family(OVS_PACKET_FAMILY,
1965 &ovs_packet_family);
1968 error = nl_lookup_genl_mcgroup(OVS_VPORT_FAMILY, OVS_VPORT_MCGROUP,
1969 &ovs_vport_mcgroup);
1972 ovsthread_once_done(&once);
1979 dpif_linux_is_internal_device(const char *name)
1981 struct dpif_linux_vport reply;
1985 error = dpif_linux_vport_get(name, &reply, &buf);
1988 } else if (error != ENODEV && error != ENOENT) {
1989 VLOG_WARN_RL(&error_rl, "%s: vport query failed (%s)",
1990 name, ovs_strerror(error));
1993 return reply.type == OVS_VPORT_TYPE_INTERNAL;
1996 /* Parses the contents of 'buf', which contains a "struct ovs_header" followed
1997 * by Netlink attributes, into 'vport'. Returns 0 if successful, otherwise a
1998 * positive errno value.
2000 * 'vport' will contain pointers into 'buf', so the caller should not free
2001 * 'buf' while 'vport' is still in use. */
2003 dpif_linux_vport_from_ofpbuf(struct dpif_linux_vport *vport,
2004 const struct ofpbuf *buf)
2006 static const struct nl_policy ovs_vport_policy[] = {
2007 [OVS_VPORT_ATTR_PORT_NO] = { .type = NL_A_U32 },
2008 [OVS_VPORT_ATTR_TYPE] = { .type = NL_A_U32 },
2009 [OVS_VPORT_ATTR_NAME] = { .type = NL_A_STRING, .max_len = IFNAMSIZ },
2010 [OVS_VPORT_ATTR_UPCALL_PID] = { .type = NL_A_UNSPEC },
2011 [OVS_VPORT_ATTR_STATS] = { NL_POLICY_FOR(struct ovs_vport_stats),
2013 [OVS_VPORT_ATTR_OPTIONS] = { .type = NL_A_NESTED, .optional = true },
2016 struct nlattr *a[ARRAY_SIZE(ovs_vport_policy)];
2017 struct ovs_header *ovs_header;
2018 struct nlmsghdr *nlmsg;
2019 struct genlmsghdr *genl;
2022 dpif_linux_vport_init(vport);
2024 ofpbuf_use_const(&b, ofpbuf_data(buf), ofpbuf_size(buf));
2025 nlmsg = ofpbuf_try_pull(&b, sizeof *nlmsg);
2026 genl = ofpbuf_try_pull(&b, sizeof *genl);
2027 ovs_header = ofpbuf_try_pull(&b, sizeof *ovs_header);
2028 if (!nlmsg || !genl || !ovs_header
2029 || nlmsg->nlmsg_type != ovs_vport_family
2030 || !nl_policy_parse(&b, 0, ovs_vport_policy, a,
2031 ARRAY_SIZE(ovs_vport_policy))) {
2035 vport->cmd = genl->cmd;
2036 vport->dp_ifindex = ovs_header->dp_ifindex;
2037 vport->port_no = nl_attr_get_odp_port(a[OVS_VPORT_ATTR_PORT_NO]);
2038 vport->type = nl_attr_get_u32(a[OVS_VPORT_ATTR_TYPE]);
2039 vport->name = nl_attr_get_string(a[OVS_VPORT_ATTR_NAME]);
2040 if (a[OVS_VPORT_ATTR_UPCALL_PID]) {
2041 vport->n_upcall_pids = nl_attr_get_size(a[OVS_VPORT_ATTR_UPCALL_PID])
2042 / (sizeof *vport->upcall_pids);
2043 vport->upcall_pids = nl_attr_get(a[OVS_VPORT_ATTR_UPCALL_PID]);
2046 if (a[OVS_VPORT_ATTR_STATS]) {
2047 vport->stats = nl_attr_get(a[OVS_VPORT_ATTR_STATS]);
2049 if (a[OVS_VPORT_ATTR_OPTIONS]) {
2050 vport->options = nl_attr_get(a[OVS_VPORT_ATTR_OPTIONS]);
2051 vport->options_len = nl_attr_get_size(a[OVS_VPORT_ATTR_OPTIONS]);
2056 /* Appends to 'buf' (which must initially be empty) a "struct ovs_header"
2057 * followed by Netlink attributes corresponding to 'vport'. */
2059 dpif_linux_vport_to_ofpbuf(const struct dpif_linux_vport *vport,
2062 struct ovs_header *ovs_header;
2064 nl_msg_put_genlmsghdr(buf, 0, ovs_vport_family, NLM_F_REQUEST | NLM_F_ECHO,
2065 vport->cmd, OVS_VPORT_VERSION);
2067 ovs_header = ofpbuf_put_uninit(buf, sizeof *ovs_header);
2068 ovs_header->dp_ifindex = vport->dp_ifindex;
2070 if (vport->port_no != ODPP_NONE) {
2071 nl_msg_put_odp_port(buf, OVS_VPORT_ATTR_PORT_NO, vport->port_no);
2074 if (vport->type != OVS_VPORT_TYPE_UNSPEC) {
2075 nl_msg_put_u32(buf, OVS_VPORT_ATTR_TYPE, vport->type);
2079 nl_msg_put_string(buf, OVS_VPORT_ATTR_NAME, vport->name);
2082 if (vport->upcall_pids) {
2083 nl_msg_put_unspec(buf, OVS_VPORT_ATTR_UPCALL_PID,
2085 vport->n_upcall_pids * sizeof *vport->upcall_pids);
2089 nl_msg_put_unspec(buf, OVS_VPORT_ATTR_STATS,
2090 vport->stats, sizeof *vport->stats);
2093 if (vport->options) {
2094 nl_msg_put_nested(buf, OVS_VPORT_ATTR_OPTIONS,
2095 vport->options, vport->options_len);
2099 /* Clears 'vport' to "empty" values. */
2101 dpif_linux_vport_init(struct dpif_linux_vport *vport)
2103 memset(vport, 0, sizeof *vport);
2104 vport->port_no = ODPP_NONE;
2107 /* Executes 'request' in the kernel datapath. If the command fails, returns a
2108 * positive errno value. Otherwise, if 'reply' and 'bufp' are null, returns 0
2109 * without doing anything else. If 'reply' and 'bufp' are nonnull, then the
2110 * result of the command is expected to be an ovs_vport also, which is decoded
2111 * and stored in '*reply' and '*bufp'. The caller must free '*bufp' when the
2112 * reply is no longer needed ('reply' will contain pointers into '*bufp'). */
2114 dpif_linux_vport_transact(const struct dpif_linux_vport *request,
2115 struct dpif_linux_vport *reply,
2116 struct ofpbuf **bufp)
2118 struct ofpbuf *request_buf;
2121 ovs_assert((reply != NULL) == (bufp != NULL));
2123 error = dpif_linux_init();
2127 dpif_linux_vport_init(reply);
2132 request_buf = ofpbuf_new(1024);
2133 dpif_linux_vport_to_ofpbuf(request, request_buf);
2134 error = nl_transact(NETLINK_GENERIC, request_buf, bufp);
2135 ofpbuf_delete(request_buf);
2139 error = dpif_linux_vport_from_ofpbuf(reply, *bufp);
2142 dpif_linux_vport_init(reply);
2143 ofpbuf_delete(*bufp);
2150 /* Obtains information about the kernel vport named 'name' and stores it into
2151 * '*reply' and '*bufp'. The caller must free '*bufp' when the reply is no
2152 * longer needed ('reply' will contain pointers into '*bufp'). */
2154 dpif_linux_vport_get(const char *name, struct dpif_linux_vport *reply,
2155 struct ofpbuf **bufp)
2157 struct dpif_linux_vport request;
2159 dpif_linux_vport_init(&request);
2160 request.cmd = OVS_VPORT_CMD_GET;
2161 request.name = name;
2163 return dpif_linux_vport_transact(&request, reply, bufp);
2166 /* Parses the contents of 'buf', which contains a "struct ovs_header" followed
2167 * by Netlink attributes, into 'dp'. Returns 0 if successful, otherwise a
2168 * positive errno value.
2170 * 'dp' will contain pointers into 'buf', so the caller should not free 'buf'
2171 * while 'dp' is still in use. */
2173 dpif_linux_dp_from_ofpbuf(struct dpif_linux_dp *dp, const struct ofpbuf *buf)
2175 static const struct nl_policy ovs_datapath_policy[] = {
2176 [OVS_DP_ATTR_NAME] = { .type = NL_A_STRING, .max_len = IFNAMSIZ },
2177 [OVS_DP_ATTR_STATS] = { NL_POLICY_FOR(struct ovs_dp_stats),
2179 [OVS_DP_ATTR_MEGAFLOW_STATS] = {
2180 NL_POLICY_FOR(struct ovs_dp_megaflow_stats),
2184 struct nlattr *a[ARRAY_SIZE(ovs_datapath_policy)];
2185 struct ovs_header *ovs_header;
2186 struct nlmsghdr *nlmsg;
2187 struct genlmsghdr *genl;
2190 dpif_linux_dp_init(dp);
2192 ofpbuf_use_const(&b, ofpbuf_data(buf), ofpbuf_size(buf));
2193 nlmsg = ofpbuf_try_pull(&b, sizeof *nlmsg);
2194 genl = ofpbuf_try_pull(&b, sizeof *genl);
2195 ovs_header = ofpbuf_try_pull(&b, sizeof *ovs_header);
2196 if (!nlmsg || !genl || !ovs_header
2197 || nlmsg->nlmsg_type != ovs_datapath_family
2198 || !nl_policy_parse(&b, 0, ovs_datapath_policy, a,
2199 ARRAY_SIZE(ovs_datapath_policy))) {
2203 dp->cmd = genl->cmd;
2204 dp->dp_ifindex = ovs_header->dp_ifindex;
2205 dp->name = nl_attr_get_string(a[OVS_DP_ATTR_NAME]);
2206 if (a[OVS_DP_ATTR_STATS]) {
2207 /* Can't use structure assignment because Netlink doesn't ensure
2208 * sufficient alignment for 64-bit members. */
2209 memcpy(&dp->stats, nl_attr_get(a[OVS_DP_ATTR_STATS]),
2213 if (a[OVS_DP_ATTR_MEGAFLOW_STATS]) {
2214 /* Can't use structure assignment because Netlink doesn't ensure
2215 * sufficient alignment for 64-bit members. */
2216 memcpy(&dp->megaflow_stats, nl_attr_get(a[OVS_DP_ATTR_MEGAFLOW_STATS]),
2217 sizeof dp->megaflow_stats);
2223 /* Appends to 'buf' the Generic Netlink message described by 'dp'. */
2225 dpif_linux_dp_to_ofpbuf(const struct dpif_linux_dp *dp, struct ofpbuf *buf)
2227 struct ovs_header *ovs_header;
2229 nl_msg_put_genlmsghdr(buf, 0, ovs_datapath_family,
2230 NLM_F_REQUEST | NLM_F_ECHO, dp->cmd,
2231 OVS_DATAPATH_VERSION);
2233 ovs_header = ofpbuf_put_uninit(buf, sizeof *ovs_header);
2234 ovs_header->dp_ifindex = dp->dp_ifindex;
2237 nl_msg_put_string(buf, OVS_DP_ATTR_NAME, dp->name);
2240 if (dp->upcall_pid) {
2241 nl_msg_put_u32(buf, OVS_DP_ATTR_UPCALL_PID, *dp->upcall_pid);
2244 if (dp->user_features) {
2245 nl_msg_put_u32(buf, OVS_DP_ATTR_USER_FEATURES, dp->user_features);
2248 /* Skip OVS_DP_ATTR_STATS since we never have a reason to serialize it. */
2251 /* Clears 'dp' to "empty" values. */
2253 dpif_linux_dp_init(struct dpif_linux_dp *dp)
2255 memset(dp, 0, sizeof *dp);
2256 dp->megaflow_stats.n_masks = UINT32_MAX;
2257 dp->megaflow_stats.n_mask_hit = UINT64_MAX;
2261 dpif_linux_dp_dump_start(struct nl_dump *dump)
2263 struct dpif_linux_dp request;
2266 dpif_linux_dp_init(&request);
2267 request.cmd = OVS_DP_CMD_GET;
2269 buf = ofpbuf_new(1024);
2270 dpif_linux_dp_to_ofpbuf(&request, buf);
2271 nl_dump_start(dump, NETLINK_GENERIC, buf);
2275 /* Executes 'request' in the kernel datapath. If the command fails, returns a
2276 * positive errno value. Otherwise, if 'reply' and 'bufp' are null, returns 0
2277 * without doing anything else. If 'reply' and 'bufp' are nonnull, then the
2278 * result of the command is expected to be of the same form, which is decoded
2279 * and stored in '*reply' and '*bufp'. The caller must free '*bufp' when the
2280 * reply is no longer needed ('reply' will contain pointers into '*bufp'). */
2282 dpif_linux_dp_transact(const struct dpif_linux_dp *request,
2283 struct dpif_linux_dp *reply, struct ofpbuf **bufp)
2285 struct ofpbuf *request_buf;
2288 ovs_assert((reply != NULL) == (bufp != NULL));
2290 request_buf = ofpbuf_new(1024);
2291 dpif_linux_dp_to_ofpbuf(request, request_buf);
2292 error = nl_transact(NETLINK_GENERIC, request_buf, bufp);
2293 ofpbuf_delete(request_buf);
2296 dpif_linux_dp_init(reply);
2298 error = dpif_linux_dp_from_ofpbuf(reply, *bufp);
2301 ofpbuf_delete(*bufp);
2308 /* Obtains information about 'dpif_' and stores it into '*reply' and '*bufp'.
2309 * The caller must free '*bufp' when the reply is no longer needed ('reply'
2310 * will contain pointers into '*bufp'). */
2312 dpif_linux_dp_get(const struct dpif *dpif_, struct dpif_linux_dp *reply,
2313 struct ofpbuf **bufp)
2315 struct dpif_linux *dpif = dpif_linux_cast(dpif_);
2316 struct dpif_linux_dp request;
2318 dpif_linux_dp_init(&request);
2319 request.cmd = OVS_DP_CMD_GET;
2320 request.dp_ifindex = dpif->dp_ifindex;
2322 return dpif_linux_dp_transact(&request, reply, bufp);
2325 /* Parses the contents of 'buf', which contains a "struct ovs_header" followed
2326 * by Netlink attributes, into 'flow'. Returns 0 if successful, otherwise a
2327 * positive errno value.
2329 * 'flow' will contain pointers into 'buf', so the caller should not free 'buf'
2330 * while 'flow' is still in use. */
2332 dpif_linux_flow_from_ofpbuf(struct dpif_linux_flow *flow,
2333 const struct ofpbuf *buf)
2335 static const struct nl_policy ovs_flow_policy[] = {
2336 [OVS_FLOW_ATTR_KEY] = { .type = NL_A_NESTED },
2337 [OVS_FLOW_ATTR_MASK] = { .type = NL_A_NESTED, .optional = true },
2338 [OVS_FLOW_ATTR_ACTIONS] = { .type = NL_A_NESTED, .optional = true },
2339 [OVS_FLOW_ATTR_STATS] = { NL_POLICY_FOR(struct ovs_flow_stats),
2341 [OVS_FLOW_ATTR_TCP_FLAGS] = { .type = NL_A_U8, .optional = true },
2342 [OVS_FLOW_ATTR_USED] = { .type = NL_A_U64, .optional = true },
2343 /* The kernel never uses OVS_FLOW_ATTR_CLEAR. */
2346 struct nlattr *a[ARRAY_SIZE(ovs_flow_policy)];
2347 struct ovs_header *ovs_header;
2348 struct nlmsghdr *nlmsg;
2349 struct genlmsghdr *genl;
2352 dpif_linux_flow_init(flow);
2354 ofpbuf_use_const(&b, ofpbuf_data(buf), ofpbuf_size(buf));
2355 nlmsg = ofpbuf_try_pull(&b, sizeof *nlmsg);
2356 genl = ofpbuf_try_pull(&b, sizeof *genl);
2357 ovs_header = ofpbuf_try_pull(&b, sizeof *ovs_header);
2358 if (!nlmsg || !genl || !ovs_header
2359 || nlmsg->nlmsg_type != ovs_flow_family
2360 || !nl_policy_parse(&b, 0, ovs_flow_policy, a,
2361 ARRAY_SIZE(ovs_flow_policy))) {
2365 flow->nlmsg_flags = nlmsg->nlmsg_flags;
2366 flow->dp_ifindex = ovs_header->dp_ifindex;
2367 flow->key = nl_attr_get(a[OVS_FLOW_ATTR_KEY]);
2368 flow->key_len = nl_attr_get_size(a[OVS_FLOW_ATTR_KEY]);
2370 if (a[OVS_FLOW_ATTR_MASK]) {
2371 flow->mask = nl_attr_get(a[OVS_FLOW_ATTR_MASK]);
2372 flow->mask_len = nl_attr_get_size(a[OVS_FLOW_ATTR_MASK]);
2374 if (a[OVS_FLOW_ATTR_ACTIONS]) {
2375 flow->actions = nl_attr_get(a[OVS_FLOW_ATTR_ACTIONS]);
2376 flow->actions_len = nl_attr_get_size(a[OVS_FLOW_ATTR_ACTIONS]);
2378 if (a[OVS_FLOW_ATTR_STATS]) {
2379 flow->stats = nl_attr_get(a[OVS_FLOW_ATTR_STATS]);
2381 if (a[OVS_FLOW_ATTR_TCP_FLAGS]) {
2382 flow->tcp_flags = nl_attr_get(a[OVS_FLOW_ATTR_TCP_FLAGS]);
2384 if (a[OVS_FLOW_ATTR_USED]) {
2385 flow->used = nl_attr_get(a[OVS_FLOW_ATTR_USED]);
2390 /* Appends to 'buf' (which must initially be empty) a "struct ovs_header"
2391 * followed by Netlink attributes corresponding to 'flow'. */
2393 dpif_linux_flow_to_ofpbuf(const struct dpif_linux_flow *flow,
2396 struct ovs_header *ovs_header;
2398 nl_msg_put_genlmsghdr(buf, 0, ovs_flow_family,
2399 NLM_F_REQUEST | flow->nlmsg_flags,
2400 flow->cmd, OVS_FLOW_VERSION);
2402 ovs_header = ofpbuf_put_uninit(buf, sizeof *ovs_header);
2403 ovs_header->dp_ifindex = flow->dp_ifindex;
2405 if (flow->key_len) {
2406 nl_msg_put_unspec(buf, OVS_FLOW_ATTR_KEY, flow->key, flow->key_len);
2409 if (flow->mask_len) {
2410 nl_msg_put_unspec(buf, OVS_FLOW_ATTR_MASK, flow->mask, flow->mask_len);
2413 if (flow->actions || flow->actions_len) {
2414 nl_msg_put_unspec(buf, OVS_FLOW_ATTR_ACTIONS,
2415 flow->actions, flow->actions_len);
2418 /* We never need to send these to the kernel. */
2419 ovs_assert(!flow->stats);
2420 ovs_assert(!flow->tcp_flags);
2421 ovs_assert(!flow->used);
2424 nl_msg_put_flag(buf, OVS_FLOW_ATTR_CLEAR);
2428 /* Clears 'flow' to "empty" values. */
2430 dpif_linux_flow_init(struct dpif_linux_flow *flow)
2432 memset(flow, 0, sizeof *flow);
2435 /* Executes 'request' in the kernel datapath. If the command fails, returns a
2436 * positive errno value. Otherwise, if 'reply' and 'bufp' are null, returns 0
2437 * without doing anything else. If 'reply' and 'bufp' are nonnull, then the
2438 * result of the command is expected to be a flow also, which is decoded and
2439 * stored in '*reply' and '*bufp'. The caller must free '*bufp' when the reply
2440 * is no longer needed ('reply' will contain pointers into '*bufp'). */
2442 dpif_linux_flow_transact(struct dpif_linux_flow *request,
2443 struct dpif_linux_flow *reply, struct ofpbuf **bufp)
2445 struct ofpbuf *request_buf;
2448 ovs_assert((reply != NULL) == (bufp != NULL));
2451 request->nlmsg_flags |= NLM_F_ECHO;
2454 request_buf = ofpbuf_new(1024);
2455 dpif_linux_flow_to_ofpbuf(request, request_buf);
2456 error = nl_transact(NETLINK_GENERIC, request_buf, bufp);
2457 ofpbuf_delete(request_buf);
2461 error = dpif_linux_flow_from_ofpbuf(reply, *bufp);
2464 dpif_linux_flow_init(reply);
2465 ofpbuf_delete(*bufp);
2473 dpif_linux_flow_get_stats(const struct dpif_linux_flow *flow,
2474 struct dpif_flow_stats *stats)
2477 stats->n_packets = get_unaligned_u64(&flow->stats->n_packets);
2478 stats->n_bytes = get_unaligned_u64(&flow->stats->n_bytes);
2480 stats->n_packets = 0;
2483 stats->used = flow->used ? get_32aligned_u64(flow->used) : 0;
2484 stats->tcp_flags = flow->tcp_flags ? *flow->tcp_flags : 0;
2487 /* Logs information about a packet that was recently lost in 'ch' (in
2490 report_loss(struct dpif_linux *dpif, struct dpif_channel *ch, uint32_t ch_idx,
2491 uint32_t handler_id)
2493 static struct vlog_rate_limit rl = VLOG_RATE_LIMIT_INIT(5, 5);
2496 if (VLOG_DROP_WARN(&rl)) {
2501 if (ch->last_poll != LLONG_MIN) {
2502 ds_put_format(&s, " (last polled %lld ms ago)",
2503 time_msec() - ch->last_poll);
2506 VLOG_WARN("%s: lost packet on port channel %u of handler %u",
2507 dpif_name(&dpif->dpif), ch_idx, handler_id);