2 * Copyright (c) 2008, 2009, 2010, 2011, 2012, 2013, 2014 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.
18 #include "netlink-socket.h"
22 #include <sys/types.h>
26 #include "dynamic-string.h"
30 #include "netlink-protocol.h"
31 #include "odp-netlink.h"
33 #include "ovs-thread.h"
34 #include "poll-loop.h"
36 #include "socket-util.h"
40 VLOG_DEFINE_THIS_MODULE(netlink_socket);
42 COVERAGE_DEFINE(netlink_overflow);
43 COVERAGE_DEFINE(netlink_received);
44 COVERAGE_DEFINE(netlink_recv_jumbo);
45 COVERAGE_DEFINE(netlink_sent);
47 /* Linux header file confusion causes this to be undefined. */
49 #define SOL_NETLINK 270
53 static struct ovs_mutex portid_mutex = OVS_MUTEX_INITIALIZER;
54 static uint32_t g_last_portid = 0;
56 /* Port IDs must be unique! */
59 OVS_GUARDED_BY(portid_mutex)
66 /* A single (bad) Netlink message can in theory dump out many, many log
67 * messages, so the burst size is set quite high here to avoid missing useful
68 * information. Also, at high logging levels we log *all* Netlink messages. */
69 static struct vlog_rate_limit rl = VLOG_RATE_LIMIT_INIT(60, 600);
71 static uint32_t nl_sock_allocate_seq(struct nl_sock *, unsigned int n);
72 static void log_nlmsg(const char *function, int error,
73 const void *message, size_t size, int protocol);
75 static int get_sock_pid_from_kernel(struct nl_sock *sock, uint32_t *pid);
78 /* Netlink sockets. */
89 unsigned int rcvbuf; /* Receive buffer size (SO_RCVBUF). */
92 /* Compile-time limit on iovecs, so that we can allocate a maximum-size array
93 * of iovecs on the stack. */
96 /* Maximum number of iovecs that may be passed to sendmsg, capped at a
97 * minimum of _XOPEN_IOV_MAX (16) and a maximum of MAX_IOVS.
99 * Initialized by nl_sock_create(). */
102 static int nl_pool_alloc(int protocol, struct nl_sock **sockp);
103 static void nl_pool_release(struct nl_sock *);
105 /* Creates a new netlink socket for the given netlink 'protocol'
106 * (NETLINK_ROUTE, NETLINK_GENERIC, ...). Returns 0 and sets '*sockp' to the
107 * new socket if successful, otherwise returns a positive errno value. */
109 nl_sock_create(int protocol, struct nl_sock **sockp)
111 static struct ovsthread_once once = OVSTHREAD_ONCE_INITIALIZER;
112 struct nl_sock *sock;
114 struct sockaddr_nl local, remote;
116 socklen_t local_size;
120 if (ovsthread_once_start(&once)) {
121 int save_errno = errno;
124 max_iovs = sysconf(_SC_UIO_MAXIOV);
125 if (max_iovs < _XOPEN_IOV_MAX) {
126 if (max_iovs == -1 && errno) {
127 VLOG_WARN("sysconf(_SC_UIO_MAXIOV): %s", ovs_strerror(errno));
129 max_iovs = _XOPEN_IOV_MAX;
130 } else if (max_iovs > MAX_IOVS) {
135 ovsthread_once_done(&once);
139 sock = xmalloc(sizeof *sock);
142 sock->handle = CreateFileA("\\\\.\\OpenVSwitchDevice",
143 GENERIC_READ | GENERIC_WRITE,
144 FILE_SHARE_READ | FILE_SHARE_WRITE,
146 FILE_ATTRIBUTE_NORMAL, NULL);
148 int last_error = GetLastError();
150 if (sock->handle == INVALID_HANDLE_VALUE) {
151 VLOG_ERR("fcntl: %s", ovs_strerror(last_error));
155 sock->fd = socket(AF_NETLINK, SOCK_RAW, protocol);
157 VLOG_ERR("fcntl: %s", ovs_strerror(errno));
162 sock->protocol = protocol;
165 rcvbuf = 1024 * 1024;
167 sock->rcvbuf = rcvbuf;
168 retval = get_sock_pid_from_kernel(sock, &sock->pid);
173 if (setsockopt(sock->fd, SOL_SOCKET, SO_RCVBUFFORCE,
174 &rcvbuf, sizeof rcvbuf)) {
175 /* Only root can use SO_RCVBUFFORCE. Everyone else gets EPERM.
176 * Warn only if the failure is therefore unexpected. */
177 if (errno != EPERM) {
178 VLOG_WARN_RL(&rl, "setting %d-byte socket receive buffer failed "
179 "(%s)", rcvbuf, ovs_strerror(errno));
183 retval = get_socket_rcvbuf(sock->fd);
188 sock->rcvbuf = retval;
190 /* Connect to kernel (pid 0) as remote address. */
191 memset(&remote, 0, sizeof remote);
192 remote.nl_family = AF_NETLINK;
194 if (connect(sock->fd, (struct sockaddr *) &remote, sizeof remote) < 0) {
195 VLOG_ERR("connect(0): %s", ovs_strerror(errno));
199 /* Obtain pid assigned by kernel. */
200 local_size = sizeof local;
201 if (getsockname(sock->fd, (struct sockaddr *) &local, &local_size) < 0) {
202 VLOG_ERR("getsockname: %s", ovs_strerror(errno));
205 if (local_size < sizeof local || local.nl_family != AF_NETLINK) {
206 VLOG_ERR("getsockname returned bad Netlink name");
210 sock->pid = local.nl_pid;
224 if (sock->handle != INVALID_HANDLE_VALUE) {
225 CloseHandle(sock->handle);
236 /* Creates a new netlink socket for the same protocol as 'src'. Returns 0 and
237 * sets '*sockp' to the new socket if successful, otherwise returns a positive
240 nl_sock_clone(const struct nl_sock *src, struct nl_sock **sockp)
242 return nl_sock_create(src->protocol, sockp);
245 /* Destroys netlink socket 'sock'. */
247 nl_sock_destroy(struct nl_sock *sock)
251 CloseHandle(sock->handle);
260 /* Reads the pid for 'sock' generated in the kernel datapath. The function
261 * follows a transaction semantic. Eventually this function should call into
264 get_sock_pid_from_kernel(struct nl_sock *sock, uint32_t *pid)
266 struct nl_transaction txn;
267 struct ofpbuf request;
268 uint64_t request_stub[128];
270 uint64_t reply_stub[128];
271 struct ovs_header *ovs_header;
272 struct nlmsghdr *nlmsg;
276 int ovs_msg_size = sizeof (struct nlmsghdr) + sizeof (struct genlmsghdr) +
277 sizeof (struct ovs_header);
279 ofpbuf_use_stub(&request, request_stub, sizeof request_stub);
280 txn.request = &request;
281 ofpbuf_use_stub(&reply, reply_stub, sizeof reply_stub);
284 seq = nl_sock_allocate_seq(sock, 1);
285 nl_msg_put_genlmsghdr(&request, 0, OVS_WIN_NL_CTRL_FAMILY_ID, 0,
286 OVS_CTRL_CMD_WIN_GET_PID, OVS_WIN_CONTROL_VERSION);
287 nlmsg = nl_msg_nlmsghdr(txn.request);
288 nlmsg->nlmsg_seq = seq;
290 ovs_header = ofpbuf_put_uninit(&request, sizeof *ovs_header);
291 ovs_header->dp_ifindex = 0;
292 ovs_header = ofpbuf_put_uninit(&reply, ovs_msg_size);
294 if (!DeviceIoControl(sock->handle, OVS_IOCTL_TRANSACT,
295 ofpbuf_data(txn.request), ofpbuf_size(txn.request),
296 ofpbuf_data(txn.reply), ofpbuf_size(txn.reply),
301 if (bytes < ovs_msg_size) {
306 nlmsg = nl_msg_nlmsghdr(txn.request);
307 if (nlmsg->nlmsg_seq != seq) {
311 *pid = nlmsg->nlmsg_pid;
316 ofpbuf_uninit(&request);
317 ofpbuf_uninit(&reply);
322 /* Tries to add 'sock' as a listener for 'multicast_group'. Returns 0 if
323 * successful, otherwise a positive errno value.
325 * A socket that is subscribed to a multicast group that receives asynchronous
326 * notifications must not be used for Netlink transactions or dumps, because
327 * transactions and dumps can cause notifications to be lost.
329 * Multicast group numbers are always positive.
331 * It is not an error to attempt to join a multicast group to which a socket
332 * already belongs. */
334 nl_sock_join_mcgroup(struct nl_sock *sock, unsigned int multicast_group)
337 #define OVS_VPORT_MCGROUP_FALLBACK_ID 33
338 struct ofpbuf msg_buf;
339 struct message_multicast
342 /* if true, join; if else, leave */
344 unsigned int groupId;
347 struct message_multicast msg = { 0 };
349 msg.nlmsg_len = sizeof(struct message_multicast);
350 msg.nlmsg_type = OVS_VPORT_MCGROUP_FALLBACK_ID;
353 msg.nlmsg_pid = sock->pid;
356 msg.groupId = multicast_group;
357 msg_buf.base_ = &msg;
358 msg_buf.data_ = &msg;
359 msg_buf.size_ = msg.nlmsg_len;
361 nl_sock_send__(sock, &msg_buf, msg.nlmsg_seq, 0);
363 if (setsockopt(sock->fd, SOL_NETLINK, NETLINK_ADD_MEMBERSHIP,
364 &multicast_group, sizeof multicast_group) < 0) {
365 VLOG_WARN("could not join multicast group %u (%s)",
366 multicast_group, ovs_strerror(errno));
373 /* Tries to make 'sock' stop listening to 'multicast_group'. Returns 0 if
374 * successful, otherwise a positive errno value.
376 * Multicast group numbers are always positive.
378 * It is not an error to attempt to leave a multicast group to which a socket
381 * On success, reading from 'sock' will still return any messages that were
382 * received on 'multicast_group' before the group was left. */
384 nl_sock_leave_mcgroup(struct nl_sock *sock, unsigned int multicast_group)
387 struct ofpbuf msg_buf;
388 struct message_multicast
391 /* if true, join; if else, leave*/
395 struct message_multicast msg = { 0 };
396 nl_msg_put_nlmsghdr(&msg, sizeof(struct message_multicast),
400 msg_buf.base_ = &msg;
401 msg_buf.data_ = &msg;
402 msg_buf.size_ = msg.nlmsg_len;
404 nl_sock_send__(sock, &msg_buf, msg.nlmsg_seq, 0);
406 if (setsockopt(sock->fd, SOL_NETLINK, NETLINK_DROP_MEMBERSHIP,
407 &multicast_group, sizeof multicast_group) < 0) {
408 VLOG_WARN("could not leave multicast group %u (%s)",
409 multicast_group, ovs_strerror(errno));
417 nl_sock_send__(struct nl_sock *sock, const struct ofpbuf *msg,
418 uint32_t nlmsg_seq, bool wait)
420 struct nlmsghdr *nlmsg = nl_msg_nlmsghdr(msg);
423 nlmsg->nlmsg_len = ofpbuf_size(msg);
424 nlmsg->nlmsg_seq = nlmsg_seq;
425 nlmsg->nlmsg_pid = sock->pid;
430 DWORD last_error = 0;
431 result = WriteFile(sock->handle, ofpbuf_data(msg), ofpbuf_size(msg),
433 last_error = GetLastError();
434 if (last_error != ERROR_SUCCESS && !result) {
439 retval = send(sock->fd, ofpbuf_data(msg), ofpbuf_size(msg), wait ? 0 : MSG_DONTWAIT);
441 error = retval < 0 ? errno : 0;
442 } while (error == EINTR);
443 log_nlmsg(__func__, error, ofpbuf_data(msg), ofpbuf_size(msg), sock->protocol);
445 COVERAGE_INC(netlink_sent);
450 /* Tries to send 'msg', which must contain a Netlink message, to the kernel on
451 * 'sock'. nlmsg_len in 'msg' will be finalized to match ofpbuf_size(msg), nlmsg_pid
452 * will be set to 'sock''s pid, and nlmsg_seq will be initialized to a fresh
453 * sequence number, before the message is sent.
455 * Returns 0 if successful, otherwise a positive errno value. If
456 * 'wait' is true, then the send will wait until buffer space is ready;
457 * otherwise, returns EAGAIN if the 'sock' send buffer is full. */
459 nl_sock_send(struct nl_sock *sock, const struct ofpbuf *msg, bool wait)
461 return nl_sock_send_seq(sock, msg, nl_sock_allocate_seq(sock, 1), wait);
464 /* Tries to send 'msg', which must contain a Netlink message, to the kernel on
465 * 'sock'. nlmsg_len in 'msg' will be finalized to match ofpbuf_size(msg), nlmsg_pid
466 * will be set to 'sock''s pid, and nlmsg_seq will be initialized to
467 * 'nlmsg_seq', before the message is sent.
469 * Returns 0 if successful, otherwise a positive errno value. If
470 * 'wait' is true, then the send will wait until buffer space is ready;
471 * otherwise, returns EAGAIN if the 'sock' send buffer is full.
473 * This function is suitable for sending a reply to a request that was received
474 * with sequence number 'nlmsg_seq'. Otherwise, use nl_sock_send() instead. */
476 nl_sock_send_seq(struct nl_sock *sock, const struct ofpbuf *msg,
477 uint32_t nlmsg_seq, bool wait)
479 return nl_sock_send__(sock, msg, nlmsg_seq, wait);
483 nl_sock_recv__(struct nl_sock *sock, struct ofpbuf *buf, bool wait)
485 /* We can't accurately predict the size of the data to be received. The
486 * caller is supposed to have allocated enough space in 'buf' to handle the
487 * "typical" case. To handle exceptions, we make available enough space in
488 * 'tail' to allow Netlink messages to be up to 64 kB long (a reasonable
489 * figure since that's the maximum length of a Netlink attribute). */
490 struct nlmsghdr *nlmsghdr;
492 #define MAX_STACK_LENGTH 81920
493 uint8_t tail[MAX_STACK_LENGTH];
502 ovs_assert(buf->allocated >= sizeof *nlmsghdr);
505 iov[0].iov_base = ofpbuf_base(buf);
506 iov[0].iov_len = buf->allocated;
507 iov[1].iov_base = tail;
508 iov[1].iov_len = sizeof tail;
510 memset(&msg, 0, sizeof msg);
514 /* Receive a Netlink message from the kernel.
516 * This works around a kernel bug in which the kernel returns an error code
517 * as if it were the number of bytes read. It doesn't actually modify
518 * anything in the receive buffer in that case, so we can initialize the
519 * Netlink header with an impossible message length and then, upon success,
520 * check whether it changed. */
521 nlmsghdr = ofpbuf_base(buf);
523 nlmsghdr->nlmsg_len = UINT32_MAX;
525 boolean result = false;
526 DWORD last_error = 0;
527 result = ReadFile(sock->handle, tail, MAX_STACK_LENGTH, &retval, NULL);
528 last_error = GetLastError();
529 if (last_error != ERROR_SUCCESS && !result) {
533 ofpbuf_put(buf, tail, retval);
536 retval = recvmsg(sock->fd, &msg, wait ? 0 : MSG_DONTWAIT);
538 error = (retval < 0 ? errno
539 : retval == 0 ? ECONNRESET /* not possible? */
540 : nlmsghdr->nlmsg_len != UINT32_MAX ? 0
542 } while (error == EINTR);
544 if (error == ENOBUFS) {
545 /* Socket receive buffer overflow dropped one or more messages that
546 * the kernel tried to send to us. */
547 COVERAGE_INC(netlink_overflow);
552 if (msg.msg_flags & MSG_TRUNC) {
553 VLOG_ERR_RL(&rl, "truncated message (longer than %"PRIuSIZE" bytes)",
558 if (retval < sizeof *nlmsghdr
559 || nlmsghdr->nlmsg_len < sizeof *nlmsghdr
560 || nlmsghdr->nlmsg_len > retval) {
561 VLOG_ERR_RL(&rl, "received invalid nlmsg (%"PRIuSIZE" bytes < %"PRIuSIZE")",
562 retval, sizeof *nlmsghdr);
566 ofpbuf_set_size(buf, MIN(retval, buf->allocated));
567 if (retval > buf->allocated) {
568 COVERAGE_INC(netlink_recv_jumbo);
569 ofpbuf_put(buf, tail, retval - buf->allocated);
573 log_nlmsg(__func__, 0, ofpbuf_data(buf), ofpbuf_size(buf), sock->protocol);
574 COVERAGE_INC(netlink_received);
579 /* Tries to receive a Netlink message from the kernel on 'sock' into 'buf'. If
580 * 'wait' is true, waits for a message to be ready. Otherwise, fails with
581 * EAGAIN if the 'sock' receive buffer is empty.
583 * The caller must have initialized 'buf' with an allocation of at least
584 * NLMSG_HDRLEN bytes. For best performance, the caller should allocate enough
585 * space for a "typical" message.
587 * On success, returns 0 and replaces 'buf''s previous content by the received
588 * message. This function expands 'buf''s allocated memory, as necessary, to
589 * hold the actual size of the received message.
591 * On failure, returns a positive errno value and clears 'buf' to zero length.
592 * 'buf' retains its previous memory allocation.
594 * Regardless of success or failure, this function resets 'buf''s headroom to
597 nl_sock_recv(struct nl_sock *sock, struct ofpbuf *buf, bool wait)
599 return nl_sock_recv__(sock, buf, wait);
603 nl_sock_record_errors__(struct nl_transaction **transactions, size_t n,
608 for (i = 0; i < n; i++) {
609 struct nl_transaction *txn = transactions[i];
613 ofpbuf_clear(txn->reply);
619 nl_sock_transact_multiple__(struct nl_sock *sock,
620 struct nl_transaction **transactions, size_t n,
623 uint64_t tmp_reply_stub[1024 / 8];
624 struct nl_transaction tmp_txn;
625 struct ofpbuf tmp_reply;
628 struct iovec iovs[MAX_IOVS];
633 base_seq = nl_sock_allocate_seq(sock, n);
635 for (i = 0; i < n; i++) {
636 struct nl_transaction *txn = transactions[i];
637 struct nlmsghdr *nlmsg = nl_msg_nlmsghdr(txn->request);
639 nlmsg->nlmsg_len = ofpbuf_size(txn->request);
640 nlmsg->nlmsg_seq = base_seq + i;
641 nlmsg->nlmsg_pid = sock->pid;
643 iovs[i].iov_base = ofpbuf_data(txn->request);
644 iovs[i].iov_len = ofpbuf_size(txn->request);
647 memset(&msg, 0, sizeof msg);
652 DWORD last_error = 0;
654 for (i = 0; i < n; i++) {
655 result = WriteFile((HANDLE)sock->handle, iovs[i].iov_base, iovs[i].iov_len,
657 last_error = GetLastError();
658 if (last_error != ERROR_SUCCESS && !result) {
666 error = sendmsg(sock->fd, &msg, 0) < 0 ? errno : 0;
668 } while (error == EINTR);
670 for (i = 0; i < n; i++) {
671 struct nl_transaction *txn = transactions[i];
673 log_nlmsg(__func__, error, ofpbuf_data(txn->request), ofpbuf_size(txn->request),
677 COVERAGE_ADD(netlink_sent, n);
684 ofpbuf_use_stub(&tmp_reply, tmp_reply_stub, sizeof tmp_reply_stub);
685 tmp_txn.request = NULL;
686 tmp_txn.reply = &tmp_reply;
689 struct nl_transaction *buf_txn, *txn;
692 /* Find a transaction whose buffer we can use for receiving a reply.
693 * If no such transaction is left, use tmp_txn. */
695 for (i = 0; i < n; i++) {
696 if (transactions[i]->reply) {
697 buf_txn = transactions[i];
702 /* Receive a reply. */
703 error = nl_sock_recv__(sock, buf_txn->reply, false);
705 if (error == EAGAIN) {
706 nl_sock_record_errors__(transactions, n, 0);
713 /* Match the reply up with a transaction. */
714 seq = nl_msg_nlmsghdr(buf_txn->reply)->nlmsg_seq;
715 if (seq < base_seq || seq >= base_seq + n) {
716 VLOG_DBG_RL(&rl, "ignoring unexpected seq %#"PRIx32, seq);
720 txn = transactions[i];
722 /* Fill in the results for 'txn'. */
723 if (nl_msg_nlmsgerr(buf_txn->reply, &txn->error)) {
725 ofpbuf_clear(txn->reply);
728 VLOG_DBG_RL(&rl, "received NAK error=%d (%s)",
729 error, ovs_strerror(txn->error));
733 if (txn->reply && txn != buf_txn) {
735 struct ofpbuf *reply = buf_txn->reply;
736 buf_txn->reply = txn->reply;
741 /* Fill in the results for transactions before 'txn'. (We have to do
742 * this after the results for 'txn' itself because of the buffer swap
744 nl_sock_record_errors__(transactions, i, 0);
748 transactions += i + 1;
752 ofpbuf_uninit(&tmp_reply);
758 nl_sock_transact_multiple(struct nl_sock *sock,
759 struct nl_transaction **transactions, size_t n)
768 /* In theory, every request could have a 64 kB reply. But the default and
769 * maximum socket rcvbuf size with typical Dom0 memory sizes both tend to
770 * be a bit below 128 kB, so that would only allow a single message in a
771 * "batch". So we assume that replies average (at most) 4 kB, which allows
772 * a good deal of batching.
774 * In practice, most of the requests that we batch either have no reply at
775 * all or a brief reply. */
776 max_batch_count = MAX(sock->rcvbuf / 4096, 1);
777 max_batch_count = MIN(max_batch_count, max_iovs);
783 /* Batch up to 'max_batch_count' transactions. But cap it at about a
784 * page of requests total because big skbuffs are expensive to
785 * allocate in the kernel. */
786 #if defined(PAGESIZE)
787 enum { MAX_BATCH_BYTES = MAX(1, PAGESIZE - 512) };
789 enum { MAX_BATCH_BYTES = 4096 - 512 };
791 bytes = ofpbuf_size(transactions[0]->request);
792 for (count = 1; count < n && count < max_batch_count; count++) {
793 if (bytes + ofpbuf_size(transactions[count]->request) > MAX_BATCH_BYTES) {
796 bytes += ofpbuf_size(transactions[count]->request);
799 error = nl_sock_transact_multiple__(sock, transactions, count, &done);
800 transactions += done;
803 if (error == ENOBUFS) {
804 VLOG_DBG_RL(&rl, "receive buffer overflow, resending request");
806 VLOG_ERR_RL(&rl, "transaction error (%s)", ovs_strerror(error));
807 nl_sock_record_errors__(transactions, n, error);
813 nl_sock_transact(struct nl_sock *sock, const struct ofpbuf *request,
814 struct ofpbuf **replyp)
816 struct nl_transaction *transactionp;
817 struct nl_transaction transaction;
819 transaction.request = CONST_CAST(struct ofpbuf *, request);
820 transaction.reply = replyp ? ofpbuf_new(1024) : NULL;
821 transactionp = &transaction;
823 nl_sock_transact_multiple(sock, &transactionp, 1);
826 if (transaction.error) {
827 ofpbuf_delete(transaction.reply);
830 *replyp = transaction.reply;
834 return transaction.error;
837 /* Drain all the messages currently in 'sock''s receive queue. */
839 nl_sock_drain(struct nl_sock *sock)
844 return drain_rcvbuf(sock->fd);
848 /* Starts a Netlink "dump" operation, by sending 'request' to the kernel on a
849 * Netlink socket created with the given 'protocol', and initializes 'dump' to
850 * reflect the state of the operation.
852 * 'request' must contain a Netlink message. Before sending the message,
853 * nlmsg_len will be finalized to match request->size, and nlmsg_pid will be
854 * set to the Netlink socket's pid. NLM_F_DUMP and NLM_F_ACK will be set in
857 * The design of this Netlink socket library ensures that the dump is reliable.
859 * This function provides no status indication. nl_dump_done() provides an
860 * error status for the entire dump operation.
862 * The caller must eventually destroy 'request'.
865 nl_dump_start(struct nl_dump *dump, int protocol, const struct ofpbuf *request)
867 nl_msg_nlmsghdr(request)->nlmsg_flags |= NLM_F_DUMP | NLM_F_ACK;
869 ovs_mutex_init(&dump->mutex);
870 ovs_mutex_lock(&dump->mutex);
871 dump->status = nl_pool_alloc(protocol, &dump->sock);
873 dump->status = nl_sock_send__(dump->sock, request,
874 nl_sock_allocate_seq(dump->sock, 1),
877 dump->nl_seq = nl_msg_nlmsghdr(request)->nlmsg_seq;
878 ovs_mutex_unlock(&dump->mutex);
882 nl_dump_refill(struct nl_dump *dump, struct ofpbuf *buffer)
883 OVS_REQUIRES(dump->mutex)
885 struct nlmsghdr *nlmsghdr;
888 while (!ofpbuf_size(buffer)) {
889 error = nl_sock_recv__(dump->sock, buffer, false);
891 /* The kernel never blocks providing the results of a dump, so
892 * error == EAGAIN means that we've read the whole thing, and
893 * therefore transform it into EOF. (The kernel always provides
894 * NLMSG_DONE as a sentinel. Some other thread must have received
895 * that already but not yet signaled it in 'status'.)
897 * Any other error is just an error. */
898 return error == EAGAIN ? EOF : error;
901 nlmsghdr = nl_msg_nlmsghdr(buffer);
902 if (dump->nl_seq != nlmsghdr->nlmsg_seq) {
903 VLOG_DBG_RL(&rl, "ignoring seq %#"PRIx32" != expected %#"PRIx32,
904 nlmsghdr->nlmsg_seq, dump->nl_seq);
905 ofpbuf_clear(buffer);
909 if (nl_msg_nlmsgerr(buffer, &error) && error) {
910 VLOG_INFO_RL(&rl, "netlink dump request error (%s)",
911 ovs_strerror(error));
912 ofpbuf_clear(buffer);
920 nl_dump_next__(struct ofpbuf *reply, struct ofpbuf *buffer)
922 struct nlmsghdr *nlmsghdr = nl_msg_next(buffer, reply);
924 VLOG_WARN_RL(&rl, "netlink dump contains message fragment");
926 } else if (nlmsghdr->nlmsg_type == NLMSG_DONE) {
933 /* Attempts to retrieve another reply from 'dump' into 'buffer'. 'dump' must
934 * have been initialized with nl_dump_start(), and 'buffer' must have been
935 * initialized. 'buffer' should be at least NL_DUMP_BUFSIZE bytes long.
937 * If successful, returns true and points 'reply->data' and
938 * 'ofpbuf_size(reply)' to the message that was retrieved. The caller must not
939 * modify 'reply' (because it points within 'buffer', which will be used by
940 * future calls to this function).
942 * On failure, returns false and sets 'reply->data' to NULL and
943 * 'ofpbuf_size(reply)' to 0. Failure might indicate an actual error or merely
944 * the end of replies. An error status for the entire dump operation is
945 * provided when it is completed by calling nl_dump_done().
947 * Multiple threads may call this function, passing the same nl_dump, however
948 * each must provide independent buffers. This function may cache multiple
949 * replies in the buffer, and these will be processed before more replies are
950 * fetched. When this function returns false, other threads may continue to
951 * process replies in their buffers, but they will not fetch more replies.
954 nl_dump_next(struct nl_dump *dump, struct ofpbuf *reply, struct ofpbuf *buffer)
958 /* If the buffer is empty, refill it.
960 * If the buffer is not empty, we don't check the dump's status.
961 * Otherwise, we could end up skipping some of the dump results if thread A
962 * hits EOF while thread B is in the midst of processing a batch. */
963 if (!ofpbuf_size(buffer)) {
964 ovs_mutex_lock(&dump->mutex);
966 /* Take the mutex here to avoid an in-kernel race. If two threads
967 * try to read from a Netlink dump socket at once, then the socket
968 * error can be set to EINVAL, which will be encountered on the
969 * next recv on that socket, which could be anywhere due to the way
970 * that we pool Netlink sockets. Serializing the recv calls avoids
972 dump->status = nl_dump_refill(dump, buffer);
974 retval = dump->status;
975 ovs_mutex_unlock(&dump->mutex);
978 /* Fetch the next message from the buffer. */
980 retval = nl_dump_next__(reply, buffer);
982 /* Record 'retval' as the dump status, but don't overwrite an error
984 ovs_mutex_lock(&dump->mutex);
985 if (dump->status <= 0) {
986 dump->status = retval;
988 ovs_mutex_unlock(&dump->mutex);
993 ofpbuf_set_data(reply, NULL);
994 ofpbuf_set_size(reply, 0);
999 /* Completes Netlink dump operation 'dump', which must have been initialized
1000 * with nl_dump_start(). Returns 0 if the dump operation was error-free,
1001 * otherwise a positive errno value describing the problem. */
1003 nl_dump_done(struct nl_dump *dump)
1007 ovs_mutex_lock(&dump->mutex);
1008 status = dump->status;
1009 ovs_mutex_unlock(&dump->mutex);
1011 /* Drain any remaining messages that the client didn't read. Otherwise the
1012 * kernel will continue to queue them up and waste buffer space.
1014 * XXX We could just destroy and discard the socket in this case. */
1016 uint64_t tmp_reply_stub[NL_DUMP_BUFSIZE / 8];
1017 struct ofpbuf reply, buf;
1019 ofpbuf_use_stub(&buf, tmp_reply_stub, sizeof tmp_reply_stub);
1020 while (nl_dump_next(dump, &reply, &buf)) {
1021 /* Nothing to do. */
1023 ofpbuf_uninit(&buf);
1025 ovs_mutex_lock(&dump->mutex);
1026 status = dump->status;
1027 ovs_mutex_unlock(&dump->mutex);
1031 nl_pool_release(dump->sock);
1032 ovs_mutex_destroy(&dump->mutex);
1034 return status == EOF ? 0 : status;
1037 /* Causes poll_block() to wake up when any of the specified 'events' (which is
1038 * a OR'd combination of POLLIN, POLLOUT, etc.) occur on 'sock'. */
1040 nl_sock_wait(const struct nl_sock *sock, short int events)
1043 poll_fd_wait(sock->handle, events);
1045 poll_fd_wait(sock->fd, events);
1049 /* Returns the underlying fd for 'sock', for use in "poll()"-like operations
1050 * that can't use nl_sock_wait().
1052 * It's a little tricky to use the returned fd correctly, because nl_sock does
1053 * "copy on write" to allow a single nl_sock to be used for notifications,
1054 * transactions, and dumps. If 'sock' is used only for notifications and
1055 * transactions (and never for dump) then the usage is safe. */
1057 nl_sock_fd(const struct nl_sock *sock)
1060 return sock->handle;
1066 /* Returns the PID associated with this socket. */
1068 nl_sock_pid(const struct nl_sock *sock)
1073 /* Miscellaneous. */
1075 struct genl_family {
1076 struct hmap_node hmap_node;
1081 static struct hmap genl_families = HMAP_INITIALIZER(&genl_families);
1083 static const struct nl_policy family_policy[CTRL_ATTR_MAX + 1] = {
1084 [CTRL_ATTR_FAMILY_ID] = {.type = NL_A_U16},
1085 [CTRL_ATTR_MCAST_GROUPS] = {.type = NL_A_NESTED, .optional = true},
1088 static struct genl_family *
1089 find_genl_family_by_id(uint16_t id)
1091 struct genl_family *family;
1093 HMAP_FOR_EACH_IN_BUCKET (family, hmap_node, hash_int(id, 0),
1095 if (family->id == id) {
1103 define_genl_family(uint16_t id, const char *name)
1105 struct genl_family *family = find_genl_family_by_id(id);
1108 if (!strcmp(family->name, name)) {
1113 family = xmalloc(sizeof *family);
1115 hmap_insert(&genl_families, &family->hmap_node, hash_int(id, 0));
1117 family->name = xstrdup(name);
1121 genl_family_to_name(uint16_t id)
1123 if (id == GENL_ID_CTRL) {
1126 struct genl_family *family = find_genl_family_by_id(id);
1127 return family ? family->name : "unknown";
1132 do_lookup_genl_family(const char *name, struct nlattr **attrs,
1133 struct ofpbuf **replyp)
1135 struct nl_sock *sock;
1136 struct ofpbuf request, *reply;
1140 error = nl_sock_create(NETLINK_GENERIC, &sock);
1145 ofpbuf_init(&request, 0);
1146 nl_msg_put_genlmsghdr(&request, 0, GENL_ID_CTRL, NLM_F_REQUEST,
1147 CTRL_CMD_GETFAMILY, 1);
1148 nl_msg_put_string(&request, CTRL_ATTR_FAMILY_NAME, name);
1149 error = nl_sock_transact(sock, &request, &reply);
1150 ofpbuf_uninit(&request);
1152 nl_sock_destroy(sock);
1156 if (!nl_policy_parse(reply, NLMSG_HDRLEN + GENL_HDRLEN,
1157 family_policy, attrs, ARRAY_SIZE(family_policy))
1158 || nl_attr_get_u16(attrs[CTRL_ATTR_FAMILY_ID]) == 0) {
1159 nl_sock_destroy(sock);
1160 ofpbuf_delete(reply);
1164 nl_sock_destroy(sock);
1169 /* Finds the multicast group called 'group_name' in genl family 'family_name'.
1170 * When successful, writes its result to 'multicast_group' and returns 0.
1171 * Otherwise, clears 'multicast_group' and returns a positive error code.
1174 nl_lookup_genl_mcgroup(const char *family_name, const char *group_name,
1175 unsigned int *multicast_group)
1177 struct nlattr *family_attrs[ARRAY_SIZE(family_policy)];
1178 const struct nlattr *mc;
1179 struct ofpbuf *reply;
1183 *multicast_group = 0;
1184 error = do_lookup_genl_family(family_name, family_attrs, &reply);
1189 if (!family_attrs[CTRL_ATTR_MCAST_GROUPS]) {
1194 NL_NESTED_FOR_EACH (mc, left, family_attrs[CTRL_ATTR_MCAST_GROUPS]) {
1195 static const struct nl_policy mc_policy[] = {
1196 [CTRL_ATTR_MCAST_GRP_ID] = {.type = NL_A_U32},
1197 [CTRL_ATTR_MCAST_GRP_NAME] = {.type = NL_A_STRING},
1200 struct nlattr *mc_attrs[ARRAY_SIZE(mc_policy)];
1201 const char *mc_name;
1203 if (!nl_parse_nested(mc, mc_policy, mc_attrs, ARRAY_SIZE(mc_policy))) {
1208 mc_name = nl_attr_get_string(mc_attrs[CTRL_ATTR_MCAST_GRP_NAME]);
1209 if (!strcmp(group_name, mc_name)) {
1211 nl_attr_get_u32(mc_attrs[CTRL_ATTR_MCAST_GRP_ID]);
1219 ofpbuf_delete(reply);
1223 /* If '*number' is 0, translates the given Generic Netlink family 'name' to a
1224 * number and stores it in '*number'. If successful, returns 0 and the caller
1225 * may use '*number' as the family number. On failure, returns a positive
1226 * errno value and '*number' caches the errno value. */
1228 nl_lookup_genl_family(const char *name, int *number)
1231 struct nlattr *attrs[ARRAY_SIZE(family_policy)];
1232 struct ofpbuf *reply;
1235 error = do_lookup_genl_family(name, attrs, &reply);
1237 *number = nl_attr_get_u16(attrs[CTRL_ATTR_FAMILY_ID]);
1238 define_genl_family(*number, name);
1242 ofpbuf_delete(reply);
1244 ovs_assert(*number != 0);
1246 return *number > 0 ? 0 : -*number;
1250 struct nl_sock *socks[16];
1254 static struct ovs_mutex pool_mutex = OVS_MUTEX_INITIALIZER;
1255 static struct nl_pool pools[MAX_LINKS] OVS_GUARDED_BY(pool_mutex);
1258 nl_pool_alloc(int protocol, struct nl_sock **sockp)
1260 struct nl_sock *sock = NULL;
1261 struct nl_pool *pool;
1263 ovs_assert(protocol >= 0 && protocol < ARRAY_SIZE(pools));
1265 ovs_mutex_lock(&pool_mutex);
1266 pool = &pools[protocol];
1268 sock = pool->socks[--pool->n];
1270 ovs_mutex_unlock(&pool_mutex);
1276 return nl_sock_create(protocol, sockp);
1281 nl_pool_release(struct nl_sock *sock)
1284 struct nl_pool *pool = &pools[sock->protocol];
1286 ovs_mutex_lock(&pool_mutex);
1287 if (pool->n < ARRAY_SIZE(pool->socks)) {
1288 pool->socks[pool->n++] = sock;
1291 ovs_mutex_unlock(&pool_mutex);
1293 nl_sock_destroy(sock);
1297 /* Sends 'request' to the kernel on a Netlink socket for the given 'protocol'
1298 * (e.g. NETLINK_ROUTE or NETLINK_GENERIC) and waits for a response. If
1299 * successful, returns 0. On failure, returns a positive errno value.
1301 * If 'replyp' is nonnull, then on success '*replyp' is set to the kernel's
1302 * reply, which the caller is responsible for freeing with ofpbuf_delete(), and
1303 * on failure '*replyp' is set to NULL. If 'replyp' is null, then the kernel's
1304 * reply, if any, is discarded.
1306 * Before the message is sent, nlmsg_len in 'request' will be finalized to
1307 * match ofpbuf_size(msg), nlmsg_pid will be set to the pid of the socket used
1308 * for sending the request, and nlmsg_seq will be initialized.
1310 * The caller is responsible for destroying 'request'.
1312 * Bare Netlink is an unreliable transport protocol. This function layers
1313 * reliable delivery and reply semantics on top of bare Netlink.
1315 * In Netlink, sending a request to the kernel is reliable enough, because the
1316 * kernel will tell us if the message cannot be queued (and we will in that
1317 * case put it on the transmit queue and wait until it can be delivered).
1319 * Receiving the reply is the real problem: if the socket buffer is full when
1320 * the kernel tries to send the reply, the reply will be dropped. However, the
1321 * kernel sets a flag that a reply has been dropped. The next call to recv
1322 * then returns ENOBUFS. We can then re-send the request.
1326 * 1. Netlink depends on sequence numbers to match up requests and
1327 * replies. The sender of a request supplies a sequence number, and
1328 * the reply echos back that sequence number.
1330 * This is fine, but (1) some kernel netlink implementations are
1331 * broken, in that they fail to echo sequence numbers and (2) this
1332 * function will drop packets with non-matching sequence numbers, so
1333 * that only a single request can be usefully transacted at a time.
1335 * 2. Resending the request causes it to be re-executed, so the request
1336 * needs to be idempotent.
1339 nl_transact(int protocol, const struct ofpbuf *request,
1340 struct ofpbuf **replyp)
1342 struct nl_sock *sock;
1345 error = nl_pool_alloc(protocol, &sock);
1351 error = nl_sock_transact(sock, request, replyp);
1353 nl_pool_release(sock);
1357 /* Sends the 'request' member of the 'n' transactions in 'transactions' on a
1358 * Netlink socket for the given 'protocol' (e.g. NETLINK_ROUTE or
1359 * NETLINK_GENERIC), in order, and receives responses to all of them. Fills in
1360 * the 'error' member of each transaction with 0 if it was successful,
1361 * otherwise with a positive errno value. If 'reply' is nonnull, then it will
1362 * be filled with the reply if the message receives a detailed reply. In other
1363 * cases, i.e. where the request failed or had no reply beyond an indication of
1364 * success, 'reply' will be cleared if it is nonnull.
1366 * The caller is responsible for destroying each request and reply, and the
1367 * transactions array itself.
1369 * Before sending each message, this function will finalize nlmsg_len in each
1370 * 'request' to match the ofpbuf's size, set nlmsg_pid to the pid of the socket
1371 * used for the transaction, and initialize nlmsg_seq.
1373 * Bare Netlink is an unreliable transport protocol. This function layers
1374 * reliable delivery and reply semantics on top of bare Netlink. See
1375 * nl_transact() for some caveats.
1378 nl_transact_multiple(int protocol,
1379 struct nl_transaction **transactions, size_t n)
1381 struct nl_sock *sock;
1384 error = nl_pool_alloc(protocol, &sock);
1386 nl_sock_transact_multiple(sock, transactions, n);
1387 nl_pool_release(sock);
1389 nl_sock_record_errors__(transactions, n, error);
1395 nl_sock_allocate_seq(struct nl_sock *sock, unsigned int n)
1397 uint32_t seq = sock->next_seq;
1399 sock->next_seq += n;
1401 /* Make it impossible for the next request for sequence numbers to wrap
1402 * around to 0. Start over with 1 to avoid ever using a sequence number of
1403 * 0, because the kernel uses sequence number 0 for notifications. */
1404 if (sock->next_seq >= UINT32_MAX / 2) {
1412 nlmsghdr_to_string(const struct nlmsghdr *h, int protocol, struct ds *ds)
1418 static const struct nlmsg_flag flags[] = {
1419 { NLM_F_REQUEST, "REQUEST" },
1420 { NLM_F_MULTI, "MULTI" },
1421 { NLM_F_ACK, "ACK" },
1422 { NLM_F_ECHO, "ECHO" },
1423 { NLM_F_DUMP, "DUMP" },
1424 { NLM_F_ROOT, "ROOT" },
1425 { NLM_F_MATCH, "MATCH" },
1426 { NLM_F_ATOMIC, "ATOMIC" },
1428 const struct nlmsg_flag *flag;
1429 uint16_t flags_left;
1431 ds_put_format(ds, "nl(len:%"PRIu32", type=%"PRIu16,
1432 h->nlmsg_len, h->nlmsg_type);
1433 if (h->nlmsg_type == NLMSG_NOOP) {
1434 ds_put_cstr(ds, "(no-op)");
1435 } else if (h->nlmsg_type == NLMSG_ERROR) {
1436 ds_put_cstr(ds, "(error)");
1437 } else if (h->nlmsg_type == NLMSG_DONE) {
1438 ds_put_cstr(ds, "(done)");
1439 } else if (h->nlmsg_type == NLMSG_OVERRUN) {
1440 ds_put_cstr(ds, "(overrun)");
1441 } else if (h->nlmsg_type < NLMSG_MIN_TYPE) {
1442 ds_put_cstr(ds, "(reserved)");
1443 } else if (protocol == NETLINK_GENERIC) {
1444 ds_put_format(ds, "(%s)", genl_family_to_name(h->nlmsg_type));
1446 ds_put_cstr(ds, "(family-defined)");
1448 ds_put_format(ds, ", flags=%"PRIx16, h->nlmsg_flags);
1449 flags_left = h->nlmsg_flags;
1450 for (flag = flags; flag < &flags[ARRAY_SIZE(flags)]; flag++) {
1451 if ((flags_left & flag->bits) == flag->bits) {
1452 ds_put_format(ds, "[%s]", flag->name);
1453 flags_left &= ~flag->bits;
1457 ds_put_format(ds, "[OTHER:%"PRIx16"]", flags_left);
1459 ds_put_format(ds, ", seq=%"PRIx32", pid=%"PRIu32,
1460 h->nlmsg_seq, h->nlmsg_pid);
1464 nlmsg_to_string(const struct ofpbuf *buffer, int protocol)
1466 struct ds ds = DS_EMPTY_INITIALIZER;
1467 const struct nlmsghdr *h = ofpbuf_at(buffer, 0, NLMSG_HDRLEN);
1469 nlmsghdr_to_string(h, protocol, &ds);
1470 if (h->nlmsg_type == NLMSG_ERROR) {
1471 const struct nlmsgerr *e;
1472 e = ofpbuf_at(buffer, NLMSG_HDRLEN,
1473 NLMSG_ALIGN(sizeof(struct nlmsgerr)));
1475 ds_put_format(&ds, " error(%d", e->error);
1477 ds_put_format(&ds, "(%s)", ovs_strerror(-e->error));
1479 ds_put_cstr(&ds, ", in-reply-to(");
1480 nlmsghdr_to_string(&e->msg, protocol, &ds);
1481 ds_put_cstr(&ds, "))");
1483 ds_put_cstr(&ds, " error(truncated)");
1485 } else if (h->nlmsg_type == NLMSG_DONE) {
1486 int *error = ofpbuf_at(buffer, NLMSG_HDRLEN, sizeof *error);
1488 ds_put_format(&ds, " done(%d", *error);
1490 ds_put_format(&ds, "(%s)", ovs_strerror(-*error));
1492 ds_put_cstr(&ds, ")");
1494 ds_put_cstr(&ds, " done(truncated)");
1496 } else if (protocol == NETLINK_GENERIC) {
1497 struct genlmsghdr *genl = nl_msg_genlmsghdr(buffer);
1499 ds_put_format(&ds, ",genl(cmd=%"PRIu8",version=%"PRIu8")",
1500 genl->cmd, genl->version);
1504 ds_put_cstr(&ds, "nl(truncated)");
1510 log_nlmsg(const char *function, int error,
1511 const void *message, size_t size, int protocol)
1513 struct ofpbuf buffer;
1516 if (!VLOG_IS_DBG_ENABLED()) {
1520 ofpbuf_use_const(&buffer, message, size);
1521 nlmsg = nlmsg_to_string(&buffer, protocol);
1522 VLOG_DBG_RL(&rl, "%s (%s): %s", function, ovs_strerror(error), nlmsg);