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);
78 /* Netlink sockets. */
83 OVERLAPPED overlapped;
91 unsigned int rcvbuf; /* Receive buffer size (SO_RCVBUF). */
94 /* Compile-time limit on iovecs, so that we can allocate a maximum-size array
95 * of iovecs on the stack. */
98 /* Maximum number of iovecs that may be passed to sendmsg, capped at a
99 * minimum of _XOPEN_IOV_MAX (16) and a maximum of MAX_IOVS.
101 * Initialized by nl_sock_create(). */
104 static int nl_pool_alloc(int protocol, struct nl_sock **sockp);
105 static void nl_pool_release(struct nl_sock *);
107 /* Creates a new netlink socket for the given netlink 'protocol'
108 * (NETLINK_ROUTE, NETLINK_GENERIC, ...). Returns 0 and sets '*sockp' to the
109 * new socket if successful, otherwise returns a positive errno value. */
111 nl_sock_create(int protocol, struct nl_sock **sockp)
113 static struct ovsthread_once once = OVSTHREAD_ONCE_INITIALIZER;
114 struct nl_sock *sock;
116 struct sockaddr_nl local, remote;
118 socklen_t local_size;
122 if (ovsthread_once_start(&once)) {
123 int save_errno = errno;
126 max_iovs = sysconf(_SC_UIO_MAXIOV);
127 if (max_iovs < _XOPEN_IOV_MAX) {
128 if (max_iovs == -1 && errno) {
129 VLOG_WARN("sysconf(_SC_UIO_MAXIOV): %s", ovs_strerror(errno));
131 max_iovs = _XOPEN_IOV_MAX;
132 } else if (max_iovs > MAX_IOVS) {
137 ovsthread_once_done(&once);
141 sock = xmalloc(sizeof *sock);
144 sock->handle = CreateFile(OVS_DEVICE_NAME_USER,
145 GENERIC_READ | GENERIC_WRITE,
146 FILE_SHARE_READ | FILE_SHARE_WRITE,
148 FILE_FLAG_OVERLAPPED, NULL);
150 if (sock->handle == INVALID_HANDLE_VALUE) {
151 VLOG_ERR("fcntl: %s", ovs_lasterror_to_string());
155 memset(&sock->overlapped, 0, sizeof sock->overlapped);
156 sock->overlapped.hEvent = CreateEvent(NULL, FALSE, FALSE, NULL);
157 if (sock->overlapped.hEvent == NULL) {
158 VLOG_ERR("fcntl: %s", ovs_lasterror_to_string());
161 /* Initialize the type/ioctl to Generic */
162 sock->read_ioctl = OVS_IOCTL_READ;
164 sock->fd = socket(AF_NETLINK, SOCK_RAW, protocol);
166 VLOG_ERR("fcntl: %s", ovs_strerror(errno));
171 sock->protocol = protocol;
174 rcvbuf = 1024 * 1024;
176 sock->rcvbuf = rcvbuf;
177 retval = get_sock_pid_from_kernel(sock);
182 if (setsockopt(sock->fd, SOL_SOCKET, SO_RCVBUFFORCE,
183 &rcvbuf, sizeof rcvbuf)) {
184 /* Only root can use SO_RCVBUFFORCE. Everyone else gets EPERM.
185 * Warn only if the failure is therefore unexpected. */
186 if (errno != EPERM) {
187 VLOG_WARN_RL(&rl, "setting %d-byte socket receive buffer failed "
188 "(%s)", rcvbuf, ovs_strerror(errno));
192 retval = get_socket_rcvbuf(sock->fd);
197 sock->rcvbuf = retval;
199 /* Connect to kernel (pid 0) as remote address. */
200 memset(&remote, 0, sizeof remote);
201 remote.nl_family = AF_NETLINK;
203 if (connect(sock->fd, (struct sockaddr *) &remote, sizeof remote) < 0) {
204 VLOG_ERR("connect(0): %s", ovs_strerror(errno));
208 /* Obtain pid assigned by kernel. */
209 local_size = sizeof local;
210 if (getsockname(sock->fd, (struct sockaddr *) &local, &local_size) < 0) {
211 VLOG_ERR("getsockname: %s", ovs_strerror(errno));
214 if (local_size < sizeof local || local.nl_family != AF_NETLINK) {
215 VLOG_ERR("getsockname returned bad Netlink name");
219 sock->pid = local.nl_pid;
233 if (sock->overlapped.hEvent) {
234 CloseHandle(sock->overlapped.hEvent);
236 if (sock->handle != INVALID_HANDLE_VALUE) {
237 CloseHandle(sock->handle);
248 /* Creates a new netlink socket for the same protocol as 'src'. Returns 0 and
249 * sets '*sockp' to the new socket if successful, otherwise returns a positive
252 nl_sock_clone(const struct nl_sock *src, struct nl_sock **sockp)
254 return nl_sock_create(src->protocol, sockp);
257 /* Destroys netlink socket 'sock'. */
259 nl_sock_destroy(struct nl_sock *sock)
263 if (sock->overlapped.hEvent) {
264 CloseHandle(sock->overlapped.hEvent);
266 CloseHandle(sock->handle);
275 /* Reads the pid for 'sock' generated in the kernel datapath. The function
276 * follows a transaction semantic. Eventually this function should call into
279 get_sock_pid_from_kernel(struct nl_sock *sock)
281 struct nl_transaction txn;
282 struct ofpbuf request;
283 uint64_t request_stub[128];
285 uint64_t reply_stub[128];
286 struct ovs_header *ovs_header;
287 struct nlmsghdr *nlmsg;
291 int ovs_msg_size = sizeof (struct nlmsghdr) + sizeof (struct genlmsghdr) +
292 sizeof (struct ovs_header);
294 ofpbuf_use_stub(&request, request_stub, sizeof request_stub);
295 txn.request = &request;
296 ofpbuf_use_stub(&reply, reply_stub, sizeof reply_stub);
299 seq = nl_sock_allocate_seq(sock, 1);
300 nl_msg_put_genlmsghdr(&request, 0, OVS_WIN_NL_CTRL_FAMILY_ID, 0,
301 OVS_CTRL_CMD_WIN_GET_PID, OVS_WIN_CONTROL_VERSION);
302 nlmsg = nl_msg_nlmsghdr(txn.request);
303 nlmsg->nlmsg_seq = seq;
305 ovs_header = ofpbuf_put_uninit(&request, sizeof *ovs_header);
306 ovs_header->dp_ifindex = 0;
307 ovs_header = ofpbuf_put_uninit(&reply, ovs_msg_size);
309 if (!DeviceIoControl(sock->handle, OVS_IOCTL_TRANSACT,
310 ofpbuf_data(txn.request), ofpbuf_size(txn.request),
311 ofpbuf_data(txn.reply), ofpbuf_size(txn.reply),
316 if (bytes < ovs_msg_size) {
321 nlmsg = nl_msg_nlmsghdr(txn.reply);
322 if (nlmsg->nlmsg_seq != seq) {
326 sock->pid = nlmsg->nlmsg_pid;
331 ofpbuf_uninit(&request);
332 ofpbuf_uninit(&reply);
339 nl_sock_mcgroup(struct nl_sock *sock, unsigned int multicast_group, bool join)
341 struct ofpbuf request;
342 uint64_t request_stub[128];
343 struct ovs_header *ovs_header;
344 struct nlmsghdr *nlmsg;
347 ofpbuf_use_stub(&request, request_stub, sizeof request_stub);
349 nl_msg_put_genlmsghdr(&request, 0, OVS_WIN_NL_CTRL_FAMILY_ID, 0,
350 OVS_CTRL_CMD_MC_SUBSCRIBE_REQ,
351 OVS_WIN_CONTROL_VERSION);
353 ovs_header = ofpbuf_put_uninit(&request, sizeof *ovs_header);
354 ovs_header->dp_ifindex = 0;
356 nl_msg_put_u32(&request, OVS_NL_ATTR_MCAST_GRP, multicast_group);
357 nl_msg_put_u8(&request, OVS_NL_ATTR_MCAST_JOIN, join ? 1 : 0);
359 error = nl_sock_send(sock, &request, true);
360 ofpbuf_uninit(&request);
364 /* Tries to add 'sock' as a listener for 'multicast_group'. Returns 0 if
365 * successful, otherwise a positive errno value.
367 * A socket that is subscribed to a multicast group that receives asynchronous
368 * notifications must not be used for Netlink transactions or dumps, because
369 * transactions and dumps can cause notifications to be lost.
371 * Multicast group numbers are always positive.
373 * It is not an error to attempt to join a multicast group to which a socket
374 * already belongs. */
376 nl_sock_join_mcgroup(struct nl_sock *sock, unsigned int multicast_group)
379 /* Set the socket type as a "multicast" socket */
380 sock->read_ioctl = OVS_IOCTL_READ_EVENT;
381 int error = nl_sock_mcgroup(sock, multicast_group, true);
383 sock->read_ioctl = OVS_IOCTL_READ;
384 VLOG_WARN("could not join multicast group %u (%s)",
385 multicast_group, ovs_strerror(error));
389 if (setsockopt(sock->fd, SOL_NETLINK, NETLINK_ADD_MEMBERSHIP,
390 &multicast_group, sizeof multicast_group) < 0) {
391 VLOG_WARN("could not join multicast group %u (%s)",
392 multicast_group, ovs_strerror(errno));
401 nl_sock_subscribe_packets(struct nl_sock *sock)
405 if (sock->read_ioctl != OVS_IOCTL_READ) {
409 error = nl_sock_subscribe_packet__(sock, true);
411 VLOG_WARN("could not unsubscribe packets (%s)",
412 ovs_strerror(errno));
415 sock->read_ioctl = OVS_IOCTL_READ_PACKET;
421 nl_sock_unsubscribe_packets(struct nl_sock *sock)
423 ovs_assert(sock->read_ioctl == OVS_IOCTL_READ_PACKET);
425 int error = nl_sock_subscribe_packet__(sock, false);
427 VLOG_WARN("could not subscribe to packets (%s)",
428 ovs_strerror(errno));
432 sock->read_ioctl = OVS_IOCTL_READ;
437 nl_sock_subscribe_packet__(struct nl_sock *sock, bool subscribe)
439 struct ofpbuf request;
440 uint64_t request_stub[128];
441 struct ovs_header *ovs_header;
442 struct nlmsghdr *nlmsg;
445 ofpbuf_use_stub(&request, request_stub, sizeof request_stub);
446 nl_msg_put_genlmsghdr(&request, 0, OVS_WIN_NL_CTRL_FAMILY_ID, 0,
447 OVS_CTRL_CMD_PACKET_SUBSCRIBE_REQ,
448 OVS_WIN_CONTROL_VERSION);
450 ovs_header = ofpbuf_put_uninit(&request, sizeof *ovs_header);
451 ovs_header->dp_ifindex = 0;
452 nl_msg_put_u8(&request, OVS_NL_ATTR_PACKET_SUBSCRIBE, subscribe ? 1 : 0);
453 nl_msg_put_u32(&request, OVS_NL_ATTR_PACKET_PID, sock->pid);
455 error = nl_sock_send(sock, &request, true);
456 ofpbuf_uninit(&request);
461 /* Tries to make 'sock' stop listening to 'multicast_group'. Returns 0 if
462 * successful, otherwise a positive errno value.
464 * Multicast group numbers are always positive.
466 * It is not an error to attempt to leave a multicast group to which a socket
469 * On success, reading from 'sock' will still return any messages that were
470 * received on 'multicast_group' before the group was left. */
472 nl_sock_leave_mcgroup(struct nl_sock *sock, unsigned int multicast_group)
475 int error = nl_sock_mcgroup(sock, multicast_group, false);
477 VLOG_WARN("could not leave multicast group %u (%s)",
478 multicast_group, ovs_strerror(error));
481 sock->read_ioctl = OVS_IOCTL_READ;
483 if (setsockopt(sock->fd, SOL_NETLINK, NETLINK_DROP_MEMBERSHIP,
484 &multicast_group, sizeof multicast_group) < 0) {
485 VLOG_WARN("could not leave multicast group %u (%s)",
486 multicast_group, ovs_strerror(errno));
494 nl_sock_send__(struct nl_sock *sock, const struct ofpbuf *msg,
495 uint32_t nlmsg_seq, bool wait)
497 struct nlmsghdr *nlmsg = nl_msg_nlmsghdr(msg);
500 nlmsg->nlmsg_len = ofpbuf_size(msg);
501 nlmsg->nlmsg_seq = nlmsg_seq;
502 nlmsg->nlmsg_pid = sock->pid;
508 if (!DeviceIoControl(sock->handle, OVS_IOCTL_WRITE,
509 ofpbuf_data(msg), ofpbuf_size(msg), NULL, 0,
512 /* XXX: Map to a more appropriate error based on GetLastError(). */
515 retval = ofpbuf_size(msg);
518 retval = send(sock->fd, ofpbuf_data(msg), ofpbuf_size(msg),
519 wait ? 0 : MSG_DONTWAIT);
521 error = retval < 0 ? errno : 0;
522 } while (error == EINTR);
523 log_nlmsg(__func__, error, ofpbuf_data(msg), ofpbuf_size(msg), sock->protocol);
525 COVERAGE_INC(netlink_sent);
530 /* Tries to send 'msg', which must contain a Netlink message, to the kernel on
531 * 'sock'. nlmsg_len in 'msg' will be finalized to match ofpbuf_size(msg), nlmsg_pid
532 * will be set to 'sock''s pid, and nlmsg_seq will be initialized to a fresh
533 * sequence number, before the message is sent.
535 * Returns 0 if successful, otherwise a positive errno value. If
536 * 'wait' is true, then the send will wait until buffer space is ready;
537 * otherwise, returns EAGAIN if the 'sock' send buffer is full. */
539 nl_sock_send(struct nl_sock *sock, const struct ofpbuf *msg, bool wait)
541 return nl_sock_send_seq(sock, msg, nl_sock_allocate_seq(sock, 1), wait);
544 /* Tries to send 'msg', which must contain a Netlink message, to the kernel on
545 * 'sock'. nlmsg_len in 'msg' will be finalized to match ofpbuf_size(msg), nlmsg_pid
546 * will be set to 'sock''s pid, and nlmsg_seq will be initialized to
547 * 'nlmsg_seq', before the message is sent.
549 * Returns 0 if successful, otherwise a positive errno value. If
550 * 'wait' is true, then the send will wait until buffer space is ready;
551 * otherwise, returns EAGAIN if the 'sock' send buffer is full.
553 * This function is suitable for sending a reply to a request that was received
554 * with sequence number 'nlmsg_seq'. Otherwise, use nl_sock_send() instead. */
556 nl_sock_send_seq(struct nl_sock *sock, const struct ofpbuf *msg,
557 uint32_t nlmsg_seq, bool wait)
559 return nl_sock_send__(sock, msg, nlmsg_seq, wait);
563 nl_sock_recv__(struct nl_sock *sock, struct ofpbuf *buf, bool wait)
565 /* We can't accurately predict the size of the data to be received. The
566 * caller is supposed to have allocated enough space in 'buf' to handle the
567 * "typical" case. To handle exceptions, we make available enough space in
568 * 'tail' to allow Netlink messages to be up to 64 kB long (a reasonable
569 * figure since that's the maximum length of a Netlink attribute). */
570 struct nlmsghdr *nlmsghdr;
577 ovs_assert(buf->allocated >= sizeof *nlmsghdr);
580 iov[0].iov_base = ofpbuf_base(buf);
581 iov[0].iov_len = buf->allocated;
582 iov[1].iov_base = tail;
583 iov[1].iov_len = sizeof tail;
585 memset(&msg, 0, sizeof msg);
589 /* Receive a Netlink message from the kernel.
591 * This works around a kernel bug in which the kernel returns an error code
592 * as if it were the number of bytes read. It doesn't actually modify
593 * anything in the receive buffer in that case, so we can initialize the
594 * Netlink header with an impossible message length and then, upon success,
595 * check whether it changed. */
596 nlmsghdr = ofpbuf_base(buf);
598 nlmsghdr->nlmsg_len = UINT32_MAX;
601 if (!DeviceIoControl(sock->handle, sock->read_ioctl,
602 NULL, 0, tail, sizeof tail, &bytes, NULL)) {
611 if (retval >= buf->allocated) {
612 ofpbuf_reinit(buf, retval);
613 nlmsghdr = ofpbuf_base(buf);
614 nlmsghdr->nlmsg_len = UINT32_MAX;
616 memcpy(ofpbuf_data(buf), tail, retval);
617 ofpbuf_set_size(buf, retval);
621 retval = recvmsg(sock->fd, &msg, wait ? 0 : MSG_DONTWAIT);
623 error = (retval < 0 ? errno
624 : retval == 0 ? ECONNRESET /* not possible? */
625 : nlmsghdr->nlmsg_len != UINT32_MAX ? 0
627 } while (error == EINTR);
629 if (error == ENOBUFS) {
630 /* Socket receive buffer overflow dropped one or more messages that
631 * the kernel tried to send to us. */
632 COVERAGE_INC(netlink_overflow);
637 if (msg.msg_flags & MSG_TRUNC) {
638 VLOG_ERR_RL(&rl, "truncated message (longer than %"PRIuSIZE" bytes)",
643 if (retval < sizeof *nlmsghdr
644 || nlmsghdr->nlmsg_len < sizeof *nlmsghdr
645 || nlmsghdr->nlmsg_len > retval) {
646 VLOG_ERR_RL(&rl, "received invalid nlmsg (%"PRIuSIZE" bytes < %"PRIuSIZE")",
647 retval, sizeof *nlmsghdr);
651 ofpbuf_set_size(buf, MIN(retval, buf->allocated));
652 if (retval > buf->allocated) {
653 COVERAGE_INC(netlink_recv_jumbo);
654 ofpbuf_put(buf, tail, retval - buf->allocated);
658 log_nlmsg(__func__, 0, ofpbuf_data(buf), ofpbuf_size(buf), sock->protocol);
659 COVERAGE_INC(netlink_received);
664 /* Tries to receive a Netlink message from the kernel on 'sock' into 'buf'. If
665 * 'wait' is true, waits for a message to be ready. Otherwise, fails with
666 * EAGAIN if the 'sock' receive buffer is empty.
668 * The caller must have initialized 'buf' with an allocation of at least
669 * NLMSG_HDRLEN bytes. For best performance, the caller should allocate enough
670 * space for a "typical" message.
672 * On success, returns 0 and replaces 'buf''s previous content by the received
673 * message. This function expands 'buf''s allocated memory, as necessary, to
674 * hold the actual size of the received message.
676 * On failure, returns a positive errno value and clears 'buf' to zero length.
677 * 'buf' retains its previous memory allocation.
679 * Regardless of success or failure, this function resets 'buf''s headroom to
682 nl_sock_recv(struct nl_sock *sock, struct ofpbuf *buf, bool wait)
684 return nl_sock_recv__(sock, buf, wait);
688 nl_sock_record_errors__(struct nl_transaction **transactions, size_t n,
693 for (i = 0; i < n; i++) {
694 struct nl_transaction *txn = transactions[i];
698 ofpbuf_clear(txn->reply);
704 nl_sock_transact_multiple__(struct nl_sock *sock,
705 struct nl_transaction **transactions, size_t n,
708 uint64_t tmp_reply_stub[1024 / 8];
709 struct nl_transaction tmp_txn;
710 struct ofpbuf tmp_reply;
713 struct iovec iovs[MAX_IOVS];
718 base_seq = nl_sock_allocate_seq(sock, n);
720 for (i = 0; i < n; i++) {
721 struct nl_transaction *txn = transactions[i];
722 struct nlmsghdr *nlmsg = nl_msg_nlmsghdr(txn->request);
724 nlmsg->nlmsg_len = ofpbuf_size(txn->request);
725 nlmsg->nlmsg_seq = base_seq + i;
726 nlmsg->nlmsg_pid = sock->pid;
728 iovs[i].iov_base = ofpbuf_data(txn->request);
729 iovs[i].iov_len = ofpbuf_size(txn->request);
733 memset(&msg, 0, sizeof msg);
737 error = sendmsg(sock->fd, &msg, 0) < 0 ? errno : 0;
738 } while (error == EINTR);
740 for (i = 0; i < n; i++) {
741 struct nl_transaction *txn = transactions[i];
743 log_nlmsg(__func__, error, ofpbuf_data(txn->request),
744 ofpbuf_size(txn->request), sock->protocol);
747 COVERAGE_ADD(netlink_sent, n);
754 ofpbuf_use_stub(&tmp_reply, tmp_reply_stub, sizeof tmp_reply_stub);
755 tmp_txn.request = NULL;
756 tmp_txn.reply = &tmp_reply;
759 struct nl_transaction *buf_txn, *txn;
762 /* Find a transaction whose buffer we can use for receiving a reply.
763 * If no such transaction is left, use tmp_txn. */
765 for (i = 0; i < n; i++) {
766 if (transactions[i]->reply) {
767 buf_txn = transactions[i];
772 /* Receive a reply. */
773 error = nl_sock_recv__(sock, buf_txn->reply, false);
775 if (error == EAGAIN) {
776 nl_sock_record_errors__(transactions, n, 0);
783 /* Match the reply up with a transaction. */
784 seq = nl_msg_nlmsghdr(buf_txn->reply)->nlmsg_seq;
785 if (seq < base_seq || seq >= base_seq + n) {
786 VLOG_DBG_RL(&rl, "ignoring unexpected seq %#"PRIx32, seq);
790 txn = transactions[i];
792 /* Fill in the results for 'txn'. */
793 if (nl_msg_nlmsgerr(buf_txn->reply, &txn->error)) {
795 ofpbuf_clear(txn->reply);
798 VLOG_DBG_RL(&rl, "received NAK error=%d (%s)",
799 error, ovs_strerror(txn->error));
803 if (txn->reply && txn != buf_txn) {
805 struct ofpbuf *reply = buf_txn->reply;
806 buf_txn->reply = txn->reply;
811 /* Fill in the results for transactions before 'txn'. (We have to do
812 * this after the results for 'txn' itself because of the buffer swap
814 nl_sock_record_errors__(transactions, i, 0);
818 transactions += i + 1;
822 ofpbuf_uninit(&tmp_reply);
825 uint8_t reply_buf[65536];
826 for (i = 0; i < n; i++) {
828 struct nl_transaction *txn = transactions[i];
829 struct nlmsghdr *request_nlmsg, *reply_nlmsg;
831 if (!DeviceIoControl(sock->handle, OVS_IOCTL_TRANSACT,
832 ofpbuf_data(txn->request),
833 ofpbuf_size(txn->request),
834 reply_buf, sizeof reply_buf,
836 /* XXX: Map to a more appropriate error. */
841 if (reply_len < sizeof *reply_nlmsg) {
842 nl_sock_record_errors__(transactions, n, 0);
843 VLOG_DBG_RL(&rl, "insufficient length of reply %#"PRIu32
844 " for seq: %#"PRIx32, reply_len, request_nlmsg->nlmsg_seq);
848 /* Validate the sequence number in the reply. */
849 request_nlmsg = nl_msg_nlmsghdr(txn->request);
850 reply_nlmsg = (struct nlmsghdr *)reply_buf;
852 if (request_nlmsg->nlmsg_seq != reply_nlmsg->nlmsg_seq) {
853 ovs_assert(request_nlmsg->nlmsg_seq == reply_nlmsg->nlmsg_seq);
854 VLOG_DBG_RL(&rl, "mismatched seq request %#"PRIx32
855 ", reply %#"PRIx32, request_nlmsg->nlmsg_seq,
856 reply_nlmsg->nlmsg_seq);
860 /* Handle errors embedded within the netlink message. */
861 ofpbuf_use_stub(&tmp_reply, reply_buf, sizeof reply_buf);
862 ofpbuf_set_size(&tmp_reply, sizeof reply_buf);
863 if (nl_msg_nlmsgerr(&tmp_reply, &txn->error)) {
865 ofpbuf_clear(txn->reply);
868 VLOG_DBG_RL(&rl, "received NAK error=%d (%s)",
869 error, ovs_strerror(txn->error));
874 /* Copy the reply to the buffer specified by the caller. */
875 if (reply_len > txn->reply->allocated) {
876 ofpbuf_reinit(txn->reply, reply_len);
878 memcpy(ofpbuf_data(txn->reply), reply_buf, reply_len);
879 ofpbuf_set_size(txn->reply, reply_len);
882 ofpbuf_uninit(&tmp_reply);
884 /* Count the number of successful transactions. */
890 COVERAGE_ADD(netlink_sent, n);
898 nl_sock_transact_multiple(struct nl_sock *sock,
899 struct nl_transaction **transactions, size_t n)
908 /* In theory, every request could have a 64 kB reply. But the default and
909 * maximum socket rcvbuf size with typical Dom0 memory sizes both tend to
910 * be a bit below 128 kB, so that would only allow a single message in a
911 * "batch". So we assume that replies average (at most) 4 kB, which allows
912 * a good deal of batching.
914 * In practice, most of the requests that we batch either have no reply at
915 * all or a brief reply. */
916 max_batch_count = MAX(sock->rcvbuf / 4096, 1);
917 max_batch_count = MIN(max_batch_count, max_iovs);
923 /* Batch up to 'max_batch_count' transactions. But cap it at about a
924 * page of requests total because big skbuffs are expensive to
925 * allocate in the kernel. */
926 #if defined(PAGESIZE)
927 enum { MAX_BATCH_BYTES = MAX(1, PAGESIZE - 512) };
929 enum { MAX_BATCH_BYTES = 4096 - 512 };
931 bytes = ofpbuf_size(transactions[0]->request);
932 for (count = 1; count < n && count < max_batch_count; count++) {
933 if (bytes + ofpbuf_size(transactions[count]->request) > MAX_BATCH_BYTES) {
936 bytes += ofpbuf_size(transactions[count]->request);
939 error = nl_sock_transact_multiple__(sock, transactions, count, &done);
940 transactions += done;
943 if (error == ENOBUFS) {
944 VLOG_DBG_RL(&rl, "receive buffer overflow, resending request");
946 VLOG_ERR_RL(&rl, "transaction error (%s)", ovs_strerror(error));
947 nl_sock_record_errors__(transactions, n, error);
953 nl_sock_transact(struct nl_sock *sock, const struct ofpbuf *request,
954 struct ofpbuf **replyp)
956 struct nl_transaction *transactionp;
957 struct nl_transaction transaction;
959 transaction.request = CONST_CAST(struct ofpbuf *, request);
960 transaction.reply = replyp ? ofpbuf_new(1024) : NULL;
961 transactionp = &transaction;
963 nl_sock_transact_multiple(sock, &transactionp, 1);
966 if (transaction.error) {
967 ofpbuf_delete(transaction.reply);
970 *replyp = transaction.reply;
974 return transaction.error;
977 /* Drain all the messages currently in 'sock''s receive queue. */
979 nl_sock_drain(struct nl_sock *sock)
984 return drain_rcvbuf(sock->fd);
988 /* Starts a Netlink "dump" operation, by sending 'request' to the kernel on a
989 * Netlink socket created with the given 'protocol', and initializes 'dump' to
990 * reflect the state of the operation.
992 * 'request' must contain a Netlink message. Before sending the message,
993 * nlmsg_len will be finalized to match request->size, and nlmsg_pid will be
994 * set to the Netlink socket's pid. NLM_F_DUMP and NLM_F_ACK will be set in
997 * The design of this Netlink socket library ensures that the dump is reliable.
999 * This function provides no status indication. nl_dump_done() provides an
1000 * error status for the entire dump operation.
1002 * The caller must eventually destroy 'request'.
1005 nl_dump_start(struct nl_dump *dump, int protocol, const struct ofpbuf *request)
1007 nl_msg_nlmsghdr(request)->nlmsg_flags |= NLM_F_DUMP | NLM_F_ACK;
1009 ovs_mutex_init(&dump->mutex);
1010 ovs_mutex_lock(&dump->mutex);
1011 dump->status = nl_pool_alloc(protocol, &dump->sock);
1012 if (!dump->status) {
1013 dump->status = nl_sock_send__(dump->sock, request,
1014 nl_sock_allocate_seq(dump->sock, 1),
1017 dump->nl_seq = nl_msg_nlmsghdr(request)->nlmsg_seq;
1018 ovs_mutex_unlock(&dump->mutex);
1022 nl_dump_refill(struct nl_dump *dump, struct ofpbuf *buffer)
1023 OVS_REQUIRES(dump->mutex)
1025 struct nlmsghdr *nlmsghdr;
1028 while (!ofpbuf_size(buffer)) {
1029 error = nl_sock_recv__(dump->sock, buffer, false);
1031 /* The kernel never blocks providing the results of a dump, so
1032 * error == EAGAIN means that we've read the whole thing, and
1033 * therefore transform it into EOF. (The kernel always provides
1034 * NLMSG_DONE as a sentinel. Some other thread must have received
1035 * that already but not yet signaled it in 'status'.)
1037 * Any other error is just an error. */
1038 return error == EAGAIN ? EOF : error;
1041 nlmsghdr = nl_msg_nlmsghdr(buffer);
1042 if (dump->nl_seq != nlmsghdr->nlmsg_seq) {
1043 VLOG_DBG_RL(&rl, "ignoring seq %#"PRIx32" != expected %#"PRIx32,
1044 nlmsghdr->nlmsg_seq, dump->nl_seq);
1045 ofpbuf_clear(buffer);
1049 if (nl_msg_nlmsgerr(buffer, &error) && error) {
1050 VLOG_INFO_RL(&rl, "netlink dump request error (%s)",
1051 ovs_strerror(error));
1052 ofpbuf_clear(buffer);
1060 nl_dump_next__(struct ofpbuf *reply, struct ofpbuf *buffer)
1062 struct nlmsghdr *nlmsghdr = nl_msg_next(buffer, reply);
1064 VLOG_WARN_RL(&rl, "netlink dump contains message fragment");
1066 } else if (nlmsghdr->nlmsg_type == NLMSG_DONE) {
1073 /* Attempts to retrieve another reply from 'dump' into 'buffer'. 'dump' must
1074 * have been initialized with nl_dump_start(), and 'buffer' must have been
1075 * initialized. 'buffer' should be at least NL_DUMP_BUFSIZE bytes long.
1077 * If successful, returns true and points 'reply->data' and
1078 * 'ofpbuf_size(reply)' to the message that was retrieved. The caller must not
1079 * modify 'reply' (because it points within 'buffer', which will be used by
1080 * future calls to this function).
1082 * On failure, returns false and sets 'reply->data' to NULL and
1083 * 'ofpbuf_size(reply)' to 0. Failure might indicate an actual error or merely
1084 * the end of replies. An error status for the entire dump operation is
1085 * provided when it is completed by calling nl_dump_done().
1087 * Multiple threads may call this function, passing the same nl_dump, however
1088 * each must provide independent buffers. This function may cache multiple
1089 * replies in the buffer, and these will be processed before more replies are
1090 * fetched. When this function returns false, other threads may continue to
1091 * process replies in their buffers, but they will not fetch more replies.
1094 nl_dump_next(struct nl_dump *dump, struct ofpbuf *reply, struct ofpbuf *buffer)
1098 /* If the buffer is empty, refill it.
1100 * If the buffer is not empty, we don't check the dump's status.
1101 * Otherwise, we could end up skipping some of the dump results if thread A
1102 * hits EOF while thread B is in the midst of processing a batch. */
1103 if (!ofpbuf_size(buffer)) {
1104 ovs_mutex_lock(&dump->mutex);
1105 if (!dump->status) {
1106 /* Take the mutex here to avoid an in-kernel race. If two threads
1107 * try to read from a Netlink dump socket at once, then the socket
1108 * error can be set to EINVAL, which will be encountered on the
1109 * next recv on that socket, which could be anywhere due to the way
1110 * that we pool Netlink sockets. Serializing the recv calls avoids
1112 dump->status = nl_dump_refill(dump, buffer);
1114 retval = dump->status;
1115 ovs_mutex_unlock(&dump->mutex);
1118 /* Fetch the next message from the buffer. */
1120 retval = nl_dump_next__(reply, buffer);
1122 /* Record 'retval' as the dump status, but don't overwrite an error
1124 ovs_mutex_lock(&dump->mutex);
1125 if (dump->status <= 0) {
1126 dump->status = retval;
1128 ovs_mutex_unlock(&dump->mutex);
1133 ofpbuf_set_data(reply, NULL);
1134 ofpbuf_set_size(reply, 0);
1139 /* Completes Netlink dump operation 'dump', which must have been initialized
1140 * with nl_dump_start(). Returns 0 if the dump operation was error-free,
1141 * otherwise a positive errno value describing the problem. */
1143 nl_dump_done(struct nl_dump *dump)
1147 ovs_mutex_lock(&dump->mutex);
1148 status = dump->status;
1149 ovs_mutex_unlock(&dump->mutex);
1151 /* Drain any remaining messages that the client didn't read. Otherwise the
1152 * kernel will continue to queue them up and waste buffer space.
1154 * XXX We could just destroy and discard the socket in this case. */
1156 uint64_t tmp_reply_stub[NL_DUMP_BUFSIZE / 8];
1157 struct ofpbuf reply, buf;
1159 ofpbuf_use_stub(&buf, tmp_reply_stub, sizeof tmp_reply_stub);
1160 while (nl_dump_next(dump, &reply, &buf)) {
1161 /* Nothing to do. */
1163 ofpbuf_uninit(&buf);
1165 ovs_mutex_lock(&dump->mutex);
1166 status = dump->status;
1167 ovs_mutex_unlock(&dump->mutex);
1171 nl_pool_release(dump->sock);
1172 ovs_mutex_destroy(&dump->mutex);
1174 return status == EOF ? 0 : status;
1178 /* Pend an I/O request in the driver. The driver completes the I/O whenever
1179 * an event or a packet is ready to be read. Once the I/O is completed
1180 * the overlapped structure event associated with the pending I/O will be set
1183 pend_io_request(struct nl_sock *sock)
1185 struct ofpbuf request;
1186 uint64_t request_stub[128];
1187 struct ovs_header *ovs_header;
1188 struct nlmsghdr *nlmsg;
1193 OVERLAPPED *overlapped = CONST_CAST(OVERLAPPED *, &sock->overlapped);
1195 int ovs_msg_size = sizeof (struct nlmsghdr) + sizeof (struct genlmsghdr) +
1196 sizeof (struct ovs_header);
1198 ofpbuf_use_stub(&request, request_stub, sizeof request_stub);
1200 seq = nl_sock_allocate_seq(sock, 1);
1201 nl_msg_put_genlmsghdr(&request, 0, OVS_WIN_NL_CTRL_FAMILY_ID, 0,
1202 OVS_CTRL_CMD_WIN_PEND_REQ, OVS_WIN_CONTROL_VERSION);
1203 nlmsg = nl_msg_nlmsghdr(&request);
1204 nlmsg->nlmsg_seq = seq;
1206 ovs_header = ofpbuf_put_uninit(&request, sizeof *ovs_header);
1207 ovs_header->dp_ifindex = 0;
1209 if (!DeviceIoControl(sock->handle, OVS_IOCTL_WRITE,
1210 ofpbuf_data(&request), ofpbuf_size(&request),
1211 NULL, 0, &bytes, overlapped)) {
1212 error = GetLastError();
1213 /* Check if the I/O got pended */
1214 if (error != ERROR_IO_INCOMPLETE && error != ERROR_IO_PENDING) {
1215 VLOG_ERR("nl_sock_wait failed - %s\n", ovs_format_message(error));
1220 /* The I/O was completed synchronously */
1221 poll_immediate_wake();
1226 ofpbuf_uninit(&request);
1231 /* Causes poll_block() to wake up when any of the specified 'events' (which is
1232 * a OR'd combination of POLLIN, POLLOUT, etc.) occur on 'sock'.
1233 * On Windows, 'sock' is not treated as const, and may be modified. */
1235 nl_sock_wait(const struct nl_sock *sock, short int events)
1238 if (sock->overlapped.Internal != STATUS_PENDING) {
1239 pend_io_request(CONST_CAST(struct nl_sock *, sock));
1240 /* XXX: poll_wevent_wait(sock->overlapped.hEvent); */
1242 poll_immediate_wake(); /* XXX: temporary. */
1244 poll_fd_wait(sock->fd, events);
1248 /* Returns the underlying fd for 'sock', for use in "poll()"-like operations
1249 * that can't use nl_sock_wait().
1251 * It's a little tricky to use the returned fd correctly, because nl_sock does
1252 * "copy on write" to allow a single nl_sock to be used for notifications,
1253 * transactions, and dumps. If 'sock' is used only for notifications and
1254 * transactions (and never for dump) then the usage is safe. */
1256 nl_sock_fd(const struct nl_sock *sock)
1259 BUILD_ASSERT_DECL(sizeof sock->handle == sizeof(int));
1260 return (int)sock->handle;
1266 /* Returns the PID associated with this socket. */
1268 nl_sock_pid(const struct nl_sock *sock)
1273 /* Miscellaneous. */
1275 struct genl_family {
1276 struct hmap_node hmap_node;
1281 static struct hmap genl_families = HMAP_INITIALIZER(&genl_families);
1283 static const struct nl_policy family_policy[CTRL_ATTR_MAX + 1] = {
1284 [CTRL_ATTR_FAMILY_ID] = {.type = NL_A_U16},
1285 [CTRL_ATTR_MCAST_GROUPS] = {.type = NL_A_NESTED, .optional = true},
1288 static struct genl_family *
1289 find_genl_family_by_id(uint16_t id)
1291 struct genl_family *family;
1293 HMAP_FOR_EACH_IN_BUCKET (family, hmap_node, hash_int(id, 0),
1295 if (family->id == id) {
1303 define_genl_family(uint16_t id, const char *name)
1305 struct genl_family *family = find_genl_family_by_id(id);
1308 if (!strcmp(family->name, name)) {
1313 family = xmalloc(sizeof *family);
1315 hmap_insert(&genl_families, &family->hmap_node, hash_int(id, 0));
1317 family->name = xstrdup(name);
1321 genl_family_to_name(uint16_t id)
1323 if (id == GENL_ID_CTRL) {
1326 struct genl_family *family = find_genl_family_by_id(id);
1327 return family ? family->name : "unknown";
1333 do_lookup_genl_family(const char *name, struct nlattr **attrs,
1334 struct ofpbuf **replyp)
1336 struct nl_sock *sock;
1337 struct ofpbuf request, *reply;
1341 error = nl_sock_create(NETLINK_GENERIC, &sock);
1346 ofpbuf_init(&request, 0);
1347 nl_msg_put_genlmsghdr(&request, 0, GENL_ID_CTRL, NLM_F_REQUEST,
1348 CTRL_CMD_GETFAMILY, 1);
1349 nl_msg_put_string(&request, CTRL_ATTR_FAMILY_NAME, name);
1350 error = nl_sock_transact(sock, &request, &reply);
1351 ofpbuf_uninit(&request);
1353 nl_sock_destroy(sock);
1357 if (!nl_policy_parse(reply, NLMSG_HDRLEN + GENL_HDRLEN,
1358 family_policy, attrs, ARRAY_SIZE(family_policy))
1359 || nl_attr_get_u16(attrs[CTRL_ATTR_FAMILY_ID]) == 0) {
1360 nl_sock_destroy(sock);
1361 ofpbuf_delete(reply);
1365 nl_sock_destroy(sock);
1371 do_lookup_genl_family(const char *name, struct nlattr **attrs,
1372 struct ofpbuf **replyp)
1374 struct nlmsghdr *nlmsg;
1375 struct ofpbuf *reply;
1378 const char *family_name;
1379 uint32_t family_version;
1380 uint32_t family_attrmax;
1381 uint32_t mcgrp_id = OVS_WIN_NL_INVALID_MCGRP_ID;
1382 const char *mcgrp_name = NULL;
1385 reply = ofpbuf_new(1024);
1387 /* CTRL_ATTR_MCAST_GROUPS is supported only for VPORT family. */
1388 if (!strcmp(name, OVS_WIN_CONTROL_FAMILY)) {
1389 family_id = OVS_WIN_NL_CTRL_FAMILY_ID;
1390 family_name = OVS_WIN_CONTROL_FAMILY;
1391 family_version = OVS_WIN_CONTROL_VERSION;
1392 family_attrmax = OVS_WIN_CONTROL_ATTR_MAX;
1393 } else if (!strcmp(name, OVS_DATAPATH_FAMILY)) {
1394 family_id = OVS_WIN_NL_DATAPATH_FAMILY_ID;
1395 family_name = OVS_DATAPATH_FAMILY;
1396 family_version = OVS_DATAPATH_VERSION;
1397 family_attrmax = OVS_DP_ATTR_MAX;
1398 } else if (!strcmp(name, OVS_PACKET_FAMILY)) {
1399 family_id = OVS_WIN_NL_PACKET_FAMILY_ID;
1400 family_name = OVS_PACKET_FAMILY;
1401 family_version = OVS_PACKET_VERSION;
1402 family_attrmax = OVS_PACKET_ATTR_MAX;
1403 } else if (!strcmp(name, OVS_VPORT_FAMILY)) {
1404 family_id = OVS_WIN_NL_VPORT_FAMILY_ID;
1405 family_name = OVS_VPORT_FAMILY;
1406 family_version = OVS_VPORT_VERSION;
1407 family_attrmax = OVS_VPORT_ATTR_MAX;
1408 mcgrp_id = OVS_WIN_NL_VPORT_MCGRP_ID;
1409 mcgrp_name = OVS_VPORT_MCGROUP;
1410 } else if (!strcmp(name, OVS_FLOW_FAMILY)) {
1411 family_id = OVS_WIN_NL_FLOW_FAMILY_ID;
1412 family_name = OVS_FLOW_FAMILY;
1413 family_version = OVS_FLOW_VERSION;
1414 family_attrmax = OVS_FLOW_ATTR_MAX;
1415 } else if (!strcmp(name, OVS_WIN_NETDEV_FAMILY)) {
1416 family_id = OVS_WIN_NL_NETDEV_FAMILY_ID;
1417 family_name = OVS_WIN_NETDEV_FAMILY;
1418 family_version = OVS_WIN_NETDEV_VERSION;
1419 family_attrmax = OVS_WIN_NETDEV_ATTR_MAX;
1421 ofpbuf_delete(reply);
1425 nl_msg_put_genlmsghdr(reply, 0, GENL_ID_CTRL, 0,
1426 CTRL_CMD_NEWFAMILY, family_version);
1427 /* CTRL_ATTR_HDRSIZE and CTRL_ATTR_OPS are not populated, but the
1428 * callers do not seem to need them. */
1429 nl_msg_put_u16(reply, CTRL_ATTR_FAMILY_ID, family_id);
1430 nl_msg_put_string(reply, CTRL_ATTR_FAMILY_NAME, family_name);
1431 nl_msg_put_u32(reply, CTRL_ATTR_VERSION, family_version);
1432 nl_msg_put_u32(reply, CTRL_ATTR_MAXATTR, family_attrmax);
1434 if (mcgrp_id != OVS_WIN_NL_INVALID_MCGRP_ID) {
1435 size_t mcgrp_ofs1 = nl_msg_start_nested(reply, CTRL_ATTR_MCAST_GROUPS);
1436 size_t mcgrp_ofs2= nl_msg_start_nested(reply,
1437 OVS_WIN_NL_VPORT_MCGRP_ID - OVS_WIN_NL_MCGRP_START_ID);
1438 nl_msg_put_u32(reply, CTRL_ATTR_MCAST_GRP_ID, mcgrp_id);
1439 ovs_assert(mcgrp_name != NULL);
1440 nl_msg_put_string(reply, CTRL_ATTR_MCAST_GRP_NAME, mcgrp_name);
1441 nl_msg_end_nested(reply, mcgrp_ofs2);
1442 nl_msg_end_nested(reply, mcgrp_ofs1);
1445 /* Set the total length of the netlink message. */
1446 nlmsg = nl_msg_nlmsghdr(reply);
1447 nlmsg->nlmsg_len = ofpbuf_size(reply);
1449 if (!nl_policy_parse(reply, NLMSG_HDRLEN + GENL_HDRLEN,
1450 family_policy, attrs, ARRAY_SIZE(family_policy))
1451 || nl_attr_get_u16(attrs[CTRL_ATTR_FAMILY_ID]) == 0) {
1452 ofpbuf_delete(reply);
1461 /* Finds the multicast group called 'group_name' in genl family 'family_name'.
1462 * When successful, writes its result to 'multicast_group' and returns 0.
1463 * Otherwise, clears 'multicast_group' and returns a positive error code.
1466 nl_lookup_genl_mcgroup(const char *family_name, const char *group_name,
1467 unsigned int *multicast_group)
1469 struct nlattr *family_attrs[ARRAY_SIZE(family_policy)];
1470 const struct nlattr *mc;
1471 struct ofpbuf *reply;
1475 *multicast_group = 0;
1476 error = do_lookup_genl_family(family_name, family_attrs, &reply);
1481 if (!family_attrs[CTRL_ATTR_MCAST_GROUPS]) {
1486 NL_NESTED_FOR_EACH (mc, left, family_attrs[CTRL_ATTR_MCAST_GROUPS]) {
1487 static const struct nl_policy mc_policy[] = {
1488 [CTRL_ATTR_MCAST_GRP_ID] = {.type = NL_A_U32},
1489 [CTRL_ATTR_MCAST_GRP_NAME] = {.type = NL_A_STRING},
1492 struct nlattr *mc_attrs[ARRAY_SIZE(mc_policy)];
1493 const char *mc_name;
1495 if (!nl_parse_nested(mc, mc_policy, mc_attrs, ARRAY_SIZE(mc_policy))) {
1500 mc_name = nl_attr_get_string(mc_attrs[CTRL_ATTR_MCAST_GRP_NAME]);
1501 if (!strcmp(group_name, mc_name)) {
1503 nl_attr_get_u32(mc_attrs[CTRL_ATTR_MCAST_GRP_ID]);
1511 ofpbuf_delete(reply);
1515 /* If '*number' is 0, translates the given Generic Netlink family 'name' to a
1516 * number and stores it in '*number'. If successful, returns 0 and the caller
1517 * may use '*number' as the family number. On failure, returns a positive
1518 * errno value and '*number' caches the errno value. */
1520 nl_lookup_genl_family(const char *name, int *number)
1523 struct nlattr *attrs[ARRAY_SIZE(family_policy)];
1524 struct ofpbuf *reply;
1527 error = do_lookup_genl_family(name, attrs, &reply);
1529 *number = nl_attr_get_u16(attrs[CTRL_ATTR_FAMILY_ID]);
1530 define_genl_family(*number, name);
1534 ofpbuf_delete(reply);
1536 ovs_assert(*number != 0);
1538 return *number > 0 ? 0 : -*number;
1542 struct nl_sock *socks[16];
1546 static struct ovs_mutex pool_mutex = OVS_MUTEX_INITIALIZER;
1547 static struct nl_pool pools[MAX_LINKS] OVS_GUARDED_BY(pool_mutex);
1550 nl_pool_alloc(int protocol, struct nl_sock **sockp)
1552 struct nl_sock *sock = NULL;
1553 struct nl_pool *pool;
1555 ovs_assert(protocol >= 0 && protocol < ARRAY_SIZE(pools));
1557 ovs_mutex_lock(&pool_mutex);
1558 pool = &pools[protocol];
1560 sock = pool->socks[--pool->n];
1562 ovs_mutex_unlock(&pool_mutex);
1568 return nl_sock_create(protocol, sockp);
1573 nl_pool_release(struct nl_sock *sock)
1576 struct nl_pool *pool = &pools[sock->protocol];
1578 ovs_mutex_lock(&pool_mutex);
1579 if (pool->n < ARRAY_SIZE(pool->socks)) {
1580 pool->socks[pool->n++] = sock;
1583 ovs_mutex_unlock(&pool_mutex);
1585 nl_sock_destroy(sock);
1589 /* Sends 'request' to the kernel on a Netlink socket for the given 'protocol'
1590 * (e.g. NETLINK_ROUTE or NETLINK_GENERIC) and waits for a response. If
1591 * successful, returns 0. On failure, returns a positive errno value.
1593 * If 'replyp' is nonnull, then on success '*replyp' is set to the kernel's
1594 * reply, which the caller is responsible for freeing with ofpbuf_delete(), and
1595 * on failure '*replyp' is set to NULL. If 'replyp' is null, then the kernel's
1596 * reply, if any, is discarded.
1598 * Before the message is sent, nlmsg_len in 'request' will be finalized to
1599 * match ofpbuf_size(msg), nlmsg_pid will be set to the pid of the socket used
1600 * for sending the request, and nlmsg_seq will be initialized.
1602 * The caller is responsible for destroying 'request'.
1604 * Bare Netlink is an unreliable transport protocol. This function layers
1605 * reliable delivery and reply semantics on top of bare Netlink.
1607 * In Netlink, sending a request to the kernel is reliable enough, because the
1608 * kernel will tell us if the message cannot be queued (and we will in that
1609 * case put it on the transmit queue and wait until it can be delivered).
1611 * Receiving the reply is the real problem: if the socket buffer is full when
1612 * the kernel tries to send the reply, the reply will be dropped. However, the
1613 * kernel sets a flag that a reply has been dropped. The next call to recv
1614 * then returns ENOBUFS. We can then re-send the request.
1618 * 1. Netlink depends on sequence numbers to match up requests and
1619 * replies. The sender of a request supplies a sequence number, and
1620 * the reply echos back that sequence number.
1622 * This is fine, but (1) some kernel netlink implementations are
1623 * broken, in that they fail to echo sequence numbers and (2) this
1624 * function will drop packets with non-matching sequence numbers, so
1625 * that only a single request can be usefully transacted at a time.
1627 * 2. Resending the request causes it to be re-executed, so the request
1628 * needs to be idempotent.
1631 nl_transact(int protocol, const struct ofpbuf *request,
1632 struct ofpbuf **replyp)
1634 struct nl_sock *sock;
1637 error = nl_pool_alloc(protocol, &sock);
1643 error = nl_sock_transact(sock, request, replyp);
1645 nl_pool_release(sock);
1649 /* Sends the 'request' member of the 'n' transactions in 'transactions' on a
1650 * Netlink socket for the given 'protocol' (e.g. NETLINK_ROUTE or
1651 * NETLINK_GENERIC), in order, and receives responses to all of them. Fills in
1652 * the 'error' member of each transaction with 0 if it was successful,
1653 * otherwise with a positive errno value. If 'reply' is nonnull, then it will
1654 * be filled with the reply if the message receives a detailed reply. In other
1655 * cases, i.e. where the request failed or had no reply beyond an indication of
1656 * success, 'reply' will be cleared if it is nonnull.
1658 * The caller is responsible for destroying each request and reply, and the
1659 * transactions array itself.
1661 * Before sending each message, this function will finalize nlmsg_len in each
1662 * 'request' to match the ofpbuf's size, set nlmsg_pid to the pid of the socket
1663 * used for the transaction, and initialize nlmsg_seq.
1665 * Bare Netlink is an unreliable transport protocol. This function layers
1666 * reliable delivery and reply semantics on top of bare Netlink. See
1667 * nl_transact() for some caveats.
1670 nl_transact_multiple(int protocol,
1671 struct nl_transaction **transactions, size_t n)
1673 struct nl_sock *sock;
1676 error = nl_pool_alloc(protocol, &sock);
1678 nl_sock_transact_multiple(sock, transactions, n);
1679 nl_pool_release(sock);
1681 nl_sock_record_errors__(transactions, n, error);
1687 nl_sock_allocate_seq(struct nl_sock *sock, unsigned int n)
1689 uint32_t seq = sock->next_seq;
1691 sock->next_seq += n;
1693 /* Make it impossible for the next request for sequence numbers to wrap
1694 * around to 0. Start over with 1 to avoid ever using a sequence number of
1695 * 0, because the kernel uses sequence number 0 for notifications. */
1696 if (sock->next_seq >= UINT32_MAX / 2) {
1704 nlmsghdr_to_string(const struct nlmsghdr *h, int protocol, struct ds *ds)
1710 static const struct nlmsg_flag flags[] = {
1711 { NLM_F_REQUEST, "REQUEST" },
1712 { NLM_F_MULTI, "MULTI" },
1713 { NLM_F_ACK, "ACK" },
1714 { NLM_F_ECHO, "ECHO" },
1715 { NLM_F_DUMP, "DUMP" },
1716 { NLM_F_ROOT, "ROOT" },
1717 { NLM_F_MATCH, "MATCH" },
1718 { NLM_F_ATOMIC, "ATOMIC" },
1720 const struct nlmsg_flag *flag;
1721 uint16_t flags_left;
1723 ds_put_format(ds, "nl(len:%"PRIu32", type=%"PRIu16,
1724 h->nlmsg_len, h->nlmsg_type);
1725 if (h->nlmsg_type == NLMSG_NOOP) {
1726 ds_put_cstr(ds, "(no-op)");
1727 } else if (h->nlmsg_type == NLMSG_ERROR) {
1728 ds_put_cstr(ds, "(error)");
1729 } else if (h->nlmsg_type == NLMSG_DONE) {
1730 ds_put_cstr(ds, "(done)");
1731 } else if (h->nlmsg_type == NLMSG_OVERRUN) {
1732 ds_put_cstr(ds, "(overrun)");
1733 } else if (h->nlmsg_type < NLMSG_MIN_TYPE) {
1734 ds_put_cstr(ds, "(reserved)");
1735 } else if (protocol == NETLINK_GENERIC) {
1736 ds_put_format(ds, "(%s)", genl_family_to_name(h->nlmsg_type));
1738 ds_put_cstr(ds, "(family-defined)");
1740 ds_put_format(ds, ", flags=%"PRIx16, h->nlmsg_flags);
1741 flags_left = h->nlmsg_flags;
1742 for (flag = flags; flag < &flags[ARRAY_SIZE(flags)]; flag++) {
1743 if ((flags_left & flag->bits) == flag->bits) {
1744 ds_put_format(ds, "[%s]", flag->name);
1745 flags_left &= ~flag->bits;
1749 ds_put_format(ds, "[OTHER:%"PRIx16"]", flags_left);
1751 ds_put_format(ds, ", seq=%"PRIx32", pid=%"PRIu32,
1752 h->nlmsg_seq, h->nlmsg_pid);
1756 nlmsg_to_string(const struct ofpbuf *buffer, int protocol)
1758 struct ds ds = DS_EMPTY_INITIALIZER;
1759 const struct nlmsghdr *h = ofpbuf_at(buffer, 0, NLMSG_HDRLEN);
1761 nlmsghdr_to_string(h, protocol, &ds);
1762 if (h->nlmsg_type == NLMSG_ERROR) {
1763 const struct nlmsgerr *e;
1764 e = ofpbuf_at(buffer, NLMSG_HDRLEN,
1765 NLMSG_ALIGN(sizeof(struct nlmsgerr)));
1767 ds_put_format(&ds, " error(%d", e->error);
1769 ds_put_format(&ds, "(%s)", ovs_strerror(-e->error));
1771 ds_put_cstr(&ds, ", in-reply-to(");
1772 nlmsghdr_to_string(&e->msg, protocol, &ds);
1773 ds_put_cstr(&ds, "))");
1775 ds_put_cstr(&ds, " error(truncated)");
1777 } else if (h->nlmsg_type == NLMSG_DONE) {
1778 int *error = ofpbuf_at(buffer, NLMSG_HDRLEN, sizeof *error);
1780 ds_put_format(&ds, " done(%d", *error);
1782 ds_put_format(&ds, "(%s)", ovs_strerror(-*error));
1784 ds_put_cstr(&ds, ")");
1786 ds_put_cstr(&ds, " done(truncated)");
1788 } else if (protocol == NETLINK_GENERIC) {
1789 struct genlmsghdr *genl = nl_msg_genlmsghdr(buffer);
1791 ds_put_format(&ds, ",genl(cmd=%"PRIu8",version=%"PRIu8")",
1792 genl->cmd, genl->version);
1796 ds_put_cstr(&ds, "nl(truncated)");
1802 log_nlmsg(const char *function, int error,
1803 const void *message, size_t size, int protocol)
1805 struct ofpbuf buffer;
1808 if (!VLOG_IS_DBG_ENABLED()) {
1812 ofpbuf_use_const(&buffer, message, size);
1813 nlmsg = nlmsg_to_string(&buffer, protocol);
1814 VLOG_DBG_RL(&rl, "%s (%s): %s", function, ovs_strerror(error), nlmsg);