X-Git-Url: https://pintos-os.org/cgi-bin/gitweb.cgi?a=blobdiff_plain;f=lib%2Fnetlink-socket.c;h=bc462353434ddc215cc48027eafcfd0aae130dbb;hb=b5d29991cc4722aec39c346c3f82291581e92aa0;hp=c9402fd9fa42943c64c24030e889161b83b1a188;hpb=2fe27d5ad27f3c7879ea696209bcf9702d9b7109;p=openvswitch

diff --git a/lib/netlink-socket.c b/lib/netlink-socket.c
index c9402fd9..bc462353 100644
--- a/lib/netlink-socket.c
+++ b/lib/netlink-socket.c
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2008, 2009, 2010 Nicira Networks.
+ * Copyright (c) 2008, 2009, 2010, 2011 Nicira Networks.
  *
  * Licensed under the Apache License, Version 2.0 (the "License");
  * you may not use this file except in compliance with the License.
@@ -21,21 +21,26 @@
 #include <inttypes.h>
 #include <stdlib.h>
 #include <sys/types.h>
+#include <sys/uio.h>
 #include <unistd.h>
 #include "coverage.h"
 #include "dynamic-string.h"
+#include "hash.h"
+#include "hmap.h"
 #include "netlink.h"
 #include "netlink-protocol.h"
 #include "ofpbuf.h"
 #include "poll-loop.h"
+#include "socket-util.h"
 #include "stress.h"
+#include "util.h"
 #include "vlog.h"
 
 VLOG_DEFINE_THIS_MODULE(netlink_socket);
 
 COVERAGE_DEFINE(netlink_overflow);
 COVERAGE_DEFINE(netlink_received);
-COVERAGE_DEFINE(netlink_recv_retry);
+COVERAGE_DEFINE(netlink_recv_jumbo);
 COVERAGE_DEFINE(netlink_send);
 COVERAGE_DEFINE(netlink_sent);
 
@@ -50,7 +55,7 @@ COVERAGE_DEFINE(netlink_sent);
 static struct vlog_rate_limit rl = VLOG_RATE_LIMIT_INIT(60, 600);
 
 static void log_nlmsg(const char *function, int error,
-                      const void *message, size_t size);
+                      const void *message, size_t size, int protocol);
 
 /* Netlink sockets. */
 
@@ -58,30 +63,51 @@ struct nl_sock
 {
     int fd;
     uint32_t pid;
+    int protocol;
+    struct nl_dump *dump;
+    unsigned int rcvbuf;        /* Receive buffer size (SO_RCVBUF). */
 };
 
-static int alloc_pid(uint32_t *);
-static void free_pid(uint32_t);
+/* Compile-time limit on iovecs, so that we can allocate a maximum-size array
+ * of iovecs on the stack. */
+#define MAX_IOVS 128
+
+/* Maximum number of iovecs that may be passed to sendmsg, capped at a
+ * minimum of _XOPEN_IOV_MAX (16) and a maximum of MAX_IOVS.
+ *
+ * Initialized by nl_sock_create(). */
+static int max_iovs;
+
+static int nl_sock_cow__(struct nl_sock *);
 
 /* Creates a new netlink socket for the given netlink 'protocol'
  * (NETLINK_ROUTE, NETLINK_GENERIC, ...).  Returns 0 and sets '*sockp' to the
- * new socket if successful, otherwise returns a positive errno value.
- *
- * If 'multicast_group' is nonzero, the new socket subscribes to the specified
- * netlink multicast group.  (A netlink socket may listen to an arbitrary
- * number of multicast groups, but so far we only need one at a time.)
- *
- * Nonzero 'so_sndbuf' or 'so_rcvbuf' override the kernel default send or
- * receive buffer size, respectively.
- */
+ * new socket if successful, otherwise returns a positive errno value.  */
 int
-nl_sock_create(int protocol, int multicast_group,
-               size_t so_sndbuf, size_t so_rcvbuf, struct nl_sock **sockp)
+nl_sock_create(int protocol, struct nl_sock **sockp)
 {
     struct nl_sock *sock;
     struct sockaddr_nl local, remote;
+    socklen_t local_size;
     int retval = 0;
 
+    if (!max_iovs) {
+        int save_errno = errno;
+        errno = 0;
+
+        max_iovs = sysconf(_SC_UIO_MAXIOV);
+        if (max_iovs < _XOPEN_IOV_MAX) {
+            if (max_iovs == -1 && errno) {
+                VLOG_WARN("sysconf(_SC_UIO_MAXIOV): %s", strerror(errno));
+            }
+            max_iovs = _XOPEN_IOV_MAX;
+        } else if (max_iovs > MAX_IOVS) {
+            max_iovs = MAX_IOVS;
+        }
+
+        errno = save_errno;
+    }
+
     *sockp = NULL;
     sock = malloc(sizeof *sock);
     if (sock == NULL) {
@@ -93,71 +119,41 @@ nl_sock_create(int protocol, int multicast_group,
         VLOG_ERR("fcntl: %s", strerror(errno));
         goto error;
     }
+    sock->protocol = protocol;
+    sock->dump = NULL;
 
-    retval = alloc_pid(&sock->pid);
-    if (retval) {
+    retval = get_socket_rcvbuf(sock->fd);
+    if (retval < 0) {
+        retval = -retval;
         goto error;
     }
+    sock->rcvbuf = retval;
 
-    if (so_sndbuf != 0
-        && setsockopt(sock->fd, SOL_SOCKET, SO_SNDBUF,
-                      &so_sndbuf, sizeof so_sndbuf) < 0) {
-        VLOG_ERR("setsockopt(SO_SNDBUF,%zu): %s", so_sndbuf, strerror(errno));
-        goto error_free_pid;
-    }
-
-    if (so_rcvbuf != 0
-        && setsockopt(sock->fd, SOL_SOCKET, SO_RCVBUF,
-                      &so_rcvbuf, sizeof so_rcvbuf) < 0) {
-        VLOG_ERR("setsockopt(SO_RCVBUF,%zu): %s", so_rcvbuf, strerror(errno));
-        goto error_free_pid;
-    }
-
-    /* Bind local address as our selected pid. */
-    memset(&local, 0, sizeof local);
-    local.nl_family = AF_NETLINK;
-    local.nl_pid = sock->pid;
-    if (multicast_group > 0 && multicast_group <= 32) {
-        /* This method of joining multicast groups is supported by old kernels,
-         * but it only allows 32 multicast groups per protocol. */
-        local.nl_groups |= 1ul << (multicast_group - 1);
-    }
-    if (bind(sock->fd, (struct sockaddr *) &local, sizeof local) < 0) {
-        VLOG_ERR("bind(%"PRIu32"): %s", sock->pid, strerror(errno));
-        goto error_free_pid;
-    }
-
-    /* Bind remote address as the kernel (pid 0). */
+    /* Connect to kernel (pid 0) as remote address. */
     memset(&remote, 0, sizeof remote);
     remote.nl_family = AF_NETLINK;
     remote.nl_pid = 0;
     if (connect(sock->fd, (struct sockaddr *) &remote, sizeof remote) < 0) {
         VLOG_ERR("connect(0): %s", strerror(errno));
-        goto error_free_pid;
+        goto error;
     }
 
-    /* Older kernel headers failed to define this macro.  We want our programs
-     * to support the newer kernel features even if compiled with older
-     * headers, so define it ourselves in such a case. */
-#ifndef NETLINK_ADD_MEMBERSHIP
-#define NETLINK_ADD_MEMBERSHIP 1
-#endif
-
-    /* This method of joining multicast groups is only supported by newish
-     * kernels, but it allows for an arbitrary number of multicast groups. */
-    if (multicast_group > 32
-        && setsockopt(sock->fd, SOL_NETLINK, NETLINK_ADD_MEMBERSHIP,
-                      &multicast_group, sizeof multicast_group) < 0) {
-        VLOG_ERR("setsockopt(NETLINK_ADD_MEMBERSHIP,%d): %s",
-                 multicast_group, strerror(errno));
-        goto error_free_pid;
+    /* Obtain pid assigned by kernel. */
+    local_size = sizeof local;
+    if (getsockname(sock->fd, (struct sockaddr *) &local, &local_size) < 0) {
+        VLOG_ERR("getsockname: %s", strerror(errno));
+        goto error;
+    }
+    if (local_size < sizeof local || local.nl_family != AF_NETLINK) {
+        VLOG_ERR("getsockname returned bad Netlink name");
+        retval = EINVAL;
+        goto error;
     }
+    sock->pid = local.nl_pid;
 
     *sockp = sock;
     return 0;
 
-error_free_pid:
-    free_pid(sock->pid);
 error:
     if (retval == 0) {
         retval = errno;
@@ -172,26 +168,81 @@ error:
     return retval;
 }
 
+/* Creates a new netlink socket for the same protocol as 'src'.  Returns 0 and
+ * sets '*sockp' to the new socket if successful, otherwise returns a positive
+ * errno value.  */
+int
+nl_sock_clone(const struct nl_sock *src, struct nl_sock **sockp)
+{
+    return nl_sock_create(src->protocol, sockp);
+}
+
 /* Destroys netlink socket 'sock'. */
 void
 nl_sock_destroy(struct nl_sock *sock)
 {
     if (sock) {
-        close(sock->fd);
-        free_pid(sock->pid);
-        free(sock);
+        if (sock->dump) {
+            sock->dump = NULL;
+        } else {
+            close(sock->fd);
+            free(sock);
+        }
     }
 }
 
-/* Tries to send 'msg', which must contain a Netlink message, to the kernel on
- * 'sock'.  nlmsg_len in 'msg' will be finalized to match msg->size, and
- * nlmsg_pid will be set to 'sock''s pid, before the message is sent.
+/* Tries to add 'sock' as a listener for 'multicast_group'.  Returns 0 if
+ * successful, otherwise a positive errno value.
  *
- * Returns 0 if successful, otherwise a positive errno value.  If
- * 'wait' is true, then the send will wait until buffer space is ready;
- * otherwise, returns EAGAIN if the 'sock' send buffer is full. */
+ * A socket that is subscribed to a multicast group that receives asynchronous
+ * notifications must not be used for Netlink transactions or dumps, because
+ * transactions and dumps can cause notifications to be lost.
+ *
+ * Multicast group numbers are always positive.
+ *
+ * It is not an error to attempt to join a multicast group to which a socket
+ * already belongs. */
 int
-nl_sock_send(struct nl_sock *sock, const struct ofpbuf *msg, bool wait)
+nl_sock_join_mcgroup(struct nl_sock *sock, unsigned int multicast_group)
+{
+    int error = nl_sock_cow__(sock);
+    if (error) {
+        return error;
+    }
+    if (setsockopt(sock->fd, SOL_NETLINK, NETLINK_ADD_MEMBERSHIP,
+                   &multicast_group, sizeof multicast_group) < 0) {
+        VLOG_WARN("could not join multicast group %u (%s)",
+                  multicast_group, strerror(errno));
+        return errno;
+    }
+    return 0;
+}
+
+/* Tries to make 'sock' stop listening to 'multicast_group'.  Returns 0 if
+ * successful, otherwise a positive errno value.
+ *
+ * Multicast group numbers are always positive.
+ *
+ * It is not an error to attempt to leave a multicast group to which a socket
+ * does not belong.
+ *
+ * On success, reading from 'sock' will still return any messages that were
+ * received on 'multicast_group' before the group was left. */
+int
+nl_sock_leave_mcgroup(struct nl_sock *sock, unsigned int multicast_group)
+{
+    assert(!sock->dump);
+    if (setsockopt(sock->fd, SOL_NETLINK, NETLINK_DROP_MEMBERSHIP,
+                   &multicast_group, sizeof multicast_group) < 0) {
+        VLOG_WARN("could not leave multicast group %u (%s)",
+                  multicast_group, strerror(errno));
+        return errno;
+    }
+    return 0;
+}
+
+static int
+nl_sock_send__(struct nl_sock *sock, const struct ofpbuf *msg, bool wait)
 {
     struct nlmsghdr *nlmsg = nl_msg_nlmsghdr(msg);
     int error;
@@ -203,43 +254,28 @@ nl_sock_send(struct nl_sock *sock, const struct ofpbuf *msg, bool wait)
         retval = send(sock->fd, msg->data, msg->size, wait ? 0 : MSG_DONTWAIT);
         error = retval < 0 ? errno : 0;
     } while (error == EINTR);
-    log_nlmsg(__func__, error, msg->data, msg->size);
+    log_nlmsg(__func__, error, msg->data, msg->size, sock->protocol);
     if (!error) {
         COVERAGE_INC(netlink_sent);
     }
     return error;
 }
 
-/* Tries to send the 'n_iov' chunks of data in 'iov' to the kernel on 'sock' as
- * a single Netlink message.  (The message must be fully formed and not require
- * finalization of its nlmsg_len or nlmsg_pid fields.)
+/* Tries to send 'msg', which must contain a Netlink message, to the kernel on
+ * 'sock'.  nlmsg_len in 'msg' will be finalized to match msg->size, and
+ * nlmsg_pid will be set to 'sock''s pid, before the message is sent.
  *
- * Returns 0 if successful, otherwise a positive errno value.  If 'wait' is
- * true, then the send will wait until buffer space is ready; otherwise,
- * returns EAGAIN if the 'sock' send buffer is full. */
+ * Returns 0 if successful, otherwise a positive errno value.  If
+ * 'wait' is true, then the send will wait until buffer space is ready;
+ * otherwise, returns EAGAIN if the 'sock' send buffer is full. */
 int
-nl_sock_sendv(struct nl_sock *sock, const struct iovec iov[], size_t n_iov,
-              bool wait)
+nl_sock_send(struct nl_sock *sock, const struct ofpbuf *msg, bool wait)
 {
-    struct msghdr msg;
-    int error;
-
-    COVERAGE_INC(netlink_send);
-    memset(&msg, 0, sizeof msg);
-    msg.msg_iov = (struct iovec *) iov;
-    msg.msg_iovlen = n_iov;
-    do {
-        int retval;
-        retval = sendmsg(sock->fd, &msg, wait ? 0 : MSG_DONTWAIT);
-        error = retval < 0 ? errno : 0;
-    } while (error == EINTR);
-    if (error != EAGAIN) {
-        log_nlmsg(__func__, error, iov[0].iov_base, iov[0].iov_len);
-        if (!error) {
-            COVERAGE_INC(netlink_sent);
-        }
+    int error = nl_sock_cow__(sock);
+    if (error) {
+        return error;
     }
-    return error;
+    return nl_sock_send__(sock, msg, wait);
 }
 
 /* This stress option is useful for testing that OVS properly tolerates
@@ -251,84 +287,71 @@ STRESS_OPTION(
     netlink_overflow, "simulate netlink socket receive buffer overflow",
     5, 1, -1, 100);
 
-/* Tries to receive a netlink message from the kernel on 'sock'.  If
- * successful, stores the received message into '*bufp' and returns 0.  The
- * caller is responsible for destroying the message with ofpbuf_delete().  On
- * failure, returns a positive errno value and stores a null pointer into
- * '*bufp'.
- *
- * If 'wait' is true, nl_sock_recv waits for a message to be ready; otherwise,
- * returns EAGAIN if the 'sock' receive buffer is empty. */
-int
-nl_sock_recv(struct nl_sock *sock, struct ofpbuf **bufp, bool wait)
+static int
+nl_sock_recv__(struct nl_sock *sock, struct ofpbuf **bufp, bool wait)
 {
-    uint8_t tmp;
-    ssize_t bufsize = 2048;
-    ssize_t nbytes, nbytes2;
-    struct ofpbuf *buf;
+    /* We can't accurately predict the size of the data to be received.  Most
+     * received data will fit in a 2 kB buffer, so we allocate that much space.
+     * In case the data is actually bigger than that, we make available enough
+     * additional space to allow Netlink messages to be up to 64 kB long (a
+     * reasonable figure since that's the maximum length of a Netlink
+     * attribute). */
+    enum { MAX_SIZE = 65536 };
+    enum { HEAD_SIZE = 2048 };
+    enum { TAIL_SIZE = MAX_SIZE - HEAD_SIZE };
+
     struct nlmsghdr *nlmsghdr;
-    struct iovec iov;
-    struct msghdr msg = {
-        .msg_name = NULL,
-        .msg_namelen = 0,
-        .msg_iov = &iov,
-        .msg_iovlen = 1,
-        .msg_control = NULL,
-        .msg_controllen = 0,
-        .msg_flags = 0
-    };
+    uint8_t tail[TAIL_SIZE];
+    struct iovec iov[2];
+    struct ofpbuf *buf;
+    struct msghdr msg;
+    ssize_t retval;
 
-    buf = ofpbuf_new(bufsize);
     *bufp = NULL;
 
-try_again:
-    /* Attempt to read the message.  We don't know the size of the data
-     * yet, so we take a guess at 2048.  If we're wrong, we keep trying
-     * and doubling the buffer size each time.
-     */
-    nlmsghdr = ofpbuf_put_uninit(buf, bufsize);
-    iov.iov_base = nlmsghdr;
-    iov.iov_len = bufsize;
+    buf = ofpbuf_new(HEAD_SIZE);
+    iov[0].iov_base = buf->data;
+    iov[0].iov_len = HEAD_SIZE;
+    iov[1].iov_base = tail;
+    iov[1].iov_len = TAIL_SIZE;
+
+    memset(&msg, 0, sizeof msg);
+    msg.msg_iov = iov;
+    msg.msg_iovlen = 2;
+
     do {
-        nbytes = recvmsg(sock->fd, &msg, (wait ? 0 : MSG_DONTWAIT) | MSG_PEEK);
-    } while (nbytes < 0 && errno == EINTR);
-    if (nbytes < 0) {
+        retval = recvmsg(sock->fd, &msg, wait ? 0 : MSG_DONTWAIT);
+    } while (retval < 0 && errno == EINTR);
+
+    if (retval < 0) {
+        int error = errno;
+        if (error == ENOBUFS) {
+            /* Socket receive buffer overflow dropped one or more messages that
+             * the kernel tried to send to us. */
+            COVERAGE_INC(netlink_overflow);
+        }
         ofpbuf_delete(buf);
-        return errno;
+        return error;
     }
+
     if (msg.msg_flags & MSG_TRUNC) {
-        COVERAGE_INC(netlink_recv_retry);
-        bufsize *= 2;
-        ofpbuf_reinit(buf, bufsize);
-        goto try_again;
+        VLOG_ERR_RL(&rl, "truncated message (longer than %d bytes)", MAX_SIZE);
+        ofpbuf_delete(buf);
+        return E2BIG;
     }
-    buf->size = nbytes;
 
-    /* We successfully read the message, so recv again to clear the queue */
-    iov.iov_base = &tmp;
-    iov.iov_len = 1;
-    do {
-        nbytes2 = recvmsg(sock->fd, &msg, MSG_DONTWAIT);
-    } while (nbytes2 < 0 && errno == EINTR);
-    if (nbytes2 < 0) {
-        if (errno == ENOBUFS) {
-            /* The kernel is notifying us that a message it tried to send to us
-             * was dropped.  We have to pass this along to the caller in case
-             * it wants to retry a request.  So kill the buffer, which we can
-             * re-read next time. */
-            COVERAGE_INC(netlink_overflow);
-            ofpbuf_delete(buf);
-            return ENOBUFS;
-        } else {
-            VLOG_ERR_RL(&rl, "failed to remove nlmsg from socket: %s\n",
-                        strerror(errno));
-        }
+    ofpbuf_put_uninit(buf, MIN(retval, HEAD_SIZE));
+    if (retval > HEAD_SIZE) {
+        COVERAGE_INC(netlink_recv_jumbo);
+        ofpbuf_put(buf, tail, retval - HEAD_SIZE);
     }
-    if (nbytes < sizeof *nlmsghdr
+
+    nlmsghdr = buf->data;
+    if (retval < sizeof *nlmsghdr
         || nlmsghdr->nlmsg_len < sizeof *nlmsghdr
-        || nlmsghdr->nlmsg_len > nbytes) {
+        || nlmsghdr->nlmsg_len > retval) {
         VLOG_ERR_RL(&rl, "received invalid nlmsg (%zd bytes < %d)",
-                    bufsize, NLMSG_HDRLEN);
+                    retval, NLMSG_HDRLEN);
         ofpbuf_delete(buf);
         return EPROTO;
     }
@@ -339,12 +362,226 @@ try_again:
     }
 
     *bufp = buf;
-    log_nlmsg(__func__, 0, buf->data, buf->size);
+    log_nlmsg(__func__, 0, buf->data, buf->size, sock->protocol);
     COVERAGE_INC(netlink_received);
 
     return 0;
 }
 
+/* Tries to receive a netlink message from the kernel on 'sock'.  If
+ * successful, stores the received message into '*bufp' and returns 0.  The
+ * caller is responsible for destroying the message with ofpbuf_delete().  On
+ * failure, returns a positive errno value and stores a null pointer into
+ * '*bufp'.
+ *
+ * If 'wait' is true, nl_sock_recv waits for a message to be ready; otherwise,
+ * returns EAGAIN if the 'sock' receive buffer is empty. */
+int
+nl_sock_recv(struct nl_sock *sock, struct ofpbuf **bufp, bool wait)
+{
+    int error = nl_sock_cow__(sock);
+    if (error) {
+        return error;
+    }
+    return nl_sock_recv__(sock, bufp, wait);
+}
+
+static int
+find_nl_transaction_by_seq(struct nl_transaction **transactions, size_t n,
+                           uint32_t seq)
+{
+    int i;
+
+    for (i = 0; i < n; i++) {
+        struct nl_transaction *t = transactions[i];
+
+        if (seq == nl_msg_nlmsghdr(t->request)->nlmsg_seq) {
+            return i;
+        }
+    }
+
+    return -1;
+}
+
+static void
+nl_sock_record_errors__(struct nl_transaction **transactions, size_t n,
+                        int error)
+{
+    size_t i;
+
+    for (i = 0; i < n; i++) {
+        transactions[i]->error = error;
+        transactions[i]->reply = NULL;
+    }
+}
+
+static int
+nl_sock_transact_multiple__(struct nl_sock *sock,
+                            struct nl_transaction **transactions, size_t n,
+                            size_t *done)
+{
+    struct iovec iovs[MAX_IOVS];
+    struct msghdr msg;
+    int error;
+    int i;
+
+    *done = 0;
+    for (i = 0; i < n; i++) {
+        struct ofpbuf *request = transactions[i]->request;
+        struct nlmsghdr *nlmsg = nl_msg_nlmsghdr(request);
+
+        nlmsg->nlmsg_len = request->size;
+        nlmsg->nlmsg_pid = sock->pid;
+        if (i == n - 1) {
+            /* Ensure that we get a reply even if the final request doesn't
+             * ordinarily call for one. */
+            nlmsg->nlmsg_flags |= NLM_F_ACK;
+        }
+
+        iovs[i].iov_base = request->data;
+        iovs[i].iov_len = request->size;
+    }
+
+    memset(&msg, 0, sizeof msg);
+    msg.msg_iov = iovs;
+    msg.msg_iovlen = n;
+    do {
+        error = sendmsg(sock->fd, &msg, 0) < 0 ? errno : 0;
+    } while (error == EINTR);
+
+    for (i = 0; i < n; i++) {
+        struct ofpbuf *request = transactions[i]->request;
+
+        log_nlmsg(__func__, error, request->data, request->size,
+                  sock->protocol);
+    }
+    if (!error) {
+        COVERAGE_ADD(netlink_sent, n);
+    }
+
+    if (error) {
+        return error;
+    }
+
+    while (n > 0) {
+        struct ofpbuf *reply;
+
+        error = nl_sock_recv__(sock, &reply, true);
+        if (error) {
+            return error;
+        }
+
+        i = find_nl_transaction_by_seq(transactions, n,
+                                       nl_msg_nlmsghdr(reply)->nlmsg_seq);
+        if (i < 0) {
+            VLOG_DBG_RL(&rl, "ignoring unexpected seq %#"PRIx32,
+                        nl_msg_nlmsghdr(reply)->nlmsg_seq);
+            ofpbuf_delete(reply);
+            continue;
+        }
+
+        nl_sock_record_errors__(transactions, i, 0);
+        if (nl_msg_nlmsgerr(reply, &error)) {
+            transactions[i]->reply = NULL;
+            transactions[i]->error = error;
+            if (error) {
+                VLOG_DBG_RL(&rl, "received NAK error=%d (%s)",
+                            error, strerror(error));
+            }
+            ofpbuf_delete(reply);
+        } else {
+            transactions[i]->reply = reply;
+            transactions[i]->error = 0;
+        }
+
+        *done += i + 1;
+        transactions += i + 1;
+        n -= i + 1;
+    }
+
+    return 0;
+}
+
+/* Sends the 'request' member of the 'n' transactions in 'transactions' to the
+ * kernel, in order, and waits for responses to all of them.  Fills in the
+ * 'error' member of each transaction with 0 if it was successful, otherwise
+ * with a positive errno value.  'reply' will be NULL on error or if the
+ * transaction was successful but had no reply beyond an indication of success.
+ * For a successful transaction that did have a more detailed reply, 'reply'
+ * will be set to the reply message.
+ *
+ * The caller is responsible for destroying each request and reply, and the
+ * transactions array itself.
+ *
+ * Before sending each message, this function will finalize nlmsg_len in each
+ * 'request' to match the ofpbuf's size, and set nlmsg_pid to 'sock''s pid.
+ * NLM_F_ACK will be added to some requests' nlmsg_flags.
+ *
+ * Bare Netlink is an unreliable transport protocol.  This function layers
+ * reliable delivery and reply semantics on top of bare Netlink.  See
+ * nl_sock_transact() for some caveats.
+ */
+void
+nl_sock_transact_multiple(struct nl_sock *sock,
+                          struct nl_transaction **transactions, size_t n)
+{
+    int max_batch_count;
+    int error;
+
+    if (!n) {
+        return;
+    }
+
+    error = nl_sock_cow__(sock);
+    if (error) {
+        nl_sock_record_errors__(transactions, n, error);
+        return;
+    }
+
+    /* In theory, every request could have a 64 kB reply.  But the default and
+     * maximum socket rcvbuf size with typical Dom0 memory sizes both tend to
+     * be a bit below 128 kB, so that would only allow a single message in a
+     * "batch".  So we assume that replies average (at most) 4 kB, which allows
+     * a good deal of batching.
+     *
+     * In practice, most of the requests that we batch either have no reply at
+     * all or a brief reply. */
+    max_batch_count = MAX(sock->rcvbuf / 4096, 1);
+    max_batch_count = MIN(max_batch_count, max_iovs);
+
+    while (n > 0) {
+        size_t count, bytes;
+        size_t done;
+
+        /* Batch up to 'max_batch_count' transactions.  But cap it at about a
+         * page of requests total because big skbuffs are expensive to
+         * allocate in the kernel.  */
+#if defined(PAGESIZE)
+        enum { MAX_BATCH_BYTES = MAX(1, PAGESIZE - 512) };
+#else
+        enum { MAX_BATCH_BYTES = 4096 - 512 };
+#endif
+        bytes = transactions[0]->request->size;
+        for (count = 1; count < n && count < max_batch_count; count++) {
+            if (bytes + transactions[count]->request->size > MAX_BATCH_BYTES) {
+                break;
+            }
+            bytes += transactions[count]->request->size;
+        }
+
+        error = nl_sock_transact_multiple__(sock, transactions, count, &done);
+        transactions += done;
+        n -= done;
+
+        if (error == ENOBUFS) {
+            VLOG_DBG_RL(&rl, "receive buffer overflow, resending request");
+        } else if (error) {
+            VLOG_ERR_RL(&rl, "transaction error (%s)", strerror(error));
+            nl_sock_record_errors__(transactions, n, error);
+        }
+    }
+}
+
 /* Sends 'request' to the kernel via 'sock' and waits for a response.  If
  * successful, returns 0.  On failure, returns a positive errno value.
  *
@@ -386,67 +623,65 @@ try_again:
  *         needs to be idempotent.
  */
 int
-nl_sock_transact(struct nl_sock *sock,
-                 const struct ofpbuf *request, struct ofpbuf **replyp)
+nl_sock_transact(struct nl_sock *sock, const struct ofpbuf *request,
+                 struct ofpbuf **replyp)
 {
-    uint32_t seq = nl_msg_nlmsghdr(request)->nlmsg_seq;
-    struct nlmsghdr *nlmsghdr;
-    struct ofpbuf *reply;
-    int retval;
+    struct nl_transaction *transactionp;
+    struct nl_transaction transaction;
 
+    transaction.request = (struct ofpbuf *) request;
+    transactionp = &transaction;
+    nl_sock_transact_multiple(sock, &transactionp, 1);
     if (replyp) {
-        *replyp = NULL;
+        *replyp = transaction.reply;
+    } else {
+        ofpbuf_delete(transaction.reply);
     }
+    return transaction.error;
+}
 
-    /* Ensure that we get a reply even if this message doesn't ordinarily call
-     * for one. */
-    nl_msg_nlmsghdr(request)->nlmsg_flags |= NLM_F_ACK;
-
-send:
-    retval = nl_sock_send(sock, request, true);
-    if (retval) {
-        return retval;
+/* Drain all the messages currently in 'sock''s receive queue. */
+int
+nl_sock_drain(struct nl_sock *sock)
+{
+    int error = nl_sock_cow__(sock);
+    if (error) {
+        return error;
     }
+    return drain_rcvbuf(sock->fd);
+}
 
-recv:
-    retval = nl_sock_recv(sock, &reply, true);
-    if (retval) {
-        if (retval == ENOBUFS) {
-            COVERAGE_INC(netlink_overflow);
-            VLOG_DBG_RL(&rl, "receive buffer overflow, resending request");
-            goto send;
-        } else {
-            return retval;
-        }
-    }
-    nlmsghdr = nl_msg_nlmsghdr(reply);
-    if (seq != nlmsghdr->nlmsg_seq) {
-        VLOG_DBG_RL(&rl, "ignoring seq %"PRIu32" != expected %"PRIu32,
-                    nl_msg_nlmsghdr(reply)->nlmsg_seq, seq);
-        ofpbuf_delete(reply);
-        goto recv;
-    }
+/* The client is attempting some operation on 'sock'.  If 'sock' has an ongoing
+ * dump operation, then replace 'sock''s fd with a new socket and hand 'sock''s
+ * old fd over to the dump. */
+static int
+nl_sock_cow__(struct nl_sock *sock)
+{
+    struct nl_sock *copy;
+    uint32_t tmp_pid;
+    int tmp_fd;
+    int error;
 
-    /* If the reply is an error, discard the reply and return the error code.
-     *
-     * Except: if the reply is just an acknowledgement (error code of 0), and
-     * the caller is interested in the reply (replyp != NULL), pass the reply
-     * up to the caller.  Otherwise the caller will get a return value of 0
-     * and null '*replyp', which makes unwary callers likely to segfault. */
-    if (nl_msg_nlmsgerr(reply, &retval) && (retval || !replyp)) {
-        ofpbuf_delete(reply);
-        if (retval) {
-            VLOG_DBG_RL(&rl, "received NAK error=%d (%s)",
-                        retval, strerror(retval));
-        }
-        return retval != EAGAIN ? retval : EPROTO;
+    if (!sock->dump) {
+        return 0;
     }
 
-    if (replyp) {
-        *replyp = reply;
-    } else {
-        ofpbuf_delete(reply);
+    error = nl_sock_clone(sock, &copy);
+    if (error) {
+        return error;
     }
+
+    tmp_fd = sock->fd;
+    sock->fd = copy->fd;
+    copy->fd = tmp_fd;
+
+    tmp_pid = sock->pid;
+    sock->pid = copy->pid;
+    copy->pid = tmp_pid;
+
+    sock->dump->sock = copy;
+    sock->dump = NULL;
+
     return 0;
 }
 
@@ -457,22 +692,23 @@ recv:
  * be set to 'sock''s pid, before the message is sent.  NLM_F_DUMP and
  * NLM_F_ACK will be set in nlmsg_flags.
  *
- * The properties of Netlink make dump operations reliable as long as all of
- * the following are true:
- *
- *   - At most a single dump is in progress at a time on a given nl_sock.
+ * This Netlink socket library is designed to ensure that the dump is reliable
+ * and that it will not interfere with other operations on 'sock', including
+ * destroying or sending and receiving messages on 'sock'.  One corner case is
+ * not handled:
  *
- *   - The nl_sock is not subscribed to any multicast groups.
- *
- *   - The nl_sock is not used to send any other messages before the dump
- *     operation is complete.
+ *   - If 'sock' has been used to send a request (e.g. with nl_sock_send())
+ *     whose response has not yet been received (e.g. with nl_sock_recv()).
+ *     This is unusual: usually nl_sock_transact() is used to send a message
+ *     and receive its reply all in one go.
  *
  * This function provides no status indication.  An error status for the entire
  * dump operation is provided when it is completed by calling nl_dump_done().
  *
- * The caller is responsible for destroying 'request'.  The caller must not
- * close 'sock' before it completes the dump operation (by calling
- * nl_dump_done()).
+ * The caller is responsible for destroying 'request'.
+ *
+ * The new 'dump' is independent of 'sock'.  'sock' and 'dump' may be destroyed
+ * in either order.
  */
 void
 nl_dump_start(struct nl_dump *dump,
@@ -481,9 +717,20 @@ nl_dump_start(struct nl_dump *dump,
     struct nlmsghdr *nlmsghdr = nl_msg_nlmsghdr(request);
     nlmsghdr->nlmsg_flags |= NLM_F_DUMP | NLM_F_ACK;
     dump->seq = nlmsghdr->nlmsg_seq;
-    dump->sock = sock;
-    dump->status = nl_sock_send(sock, request, true);
     dump->buffer = NULL;
+    if (sock->dump) {
+        /* 'sock' already has an ongoing dump.  Clone the socket because
+         * Netlink only allows one dump at a time. */
+        dump->status = nl_sock_clone(sock, &dump->sock);
+        if (dump->status) {
+            return;
+        }
+    } else {
+        sock->dump = dump;
+        dump->sock = sock;
+        dump->status = 0;
+    }
+    dump->status = nl_sock_send__(sock, request, true);
 }
 
 /* Helper function for nl_dump_next(). */
@@ -494,7 +741,7 @@ nl_dump_recv(struct nl_dump *dump, struct ofpbuf **bufferp)
     struct ofpbuf *buffer;
     int retval;
 
-    retval = nl_sock_recv(dump->sock, bufferp, true);
+    retval = nl_sock_recv__(dump->sock, bufferp, true);
     if (retval) {
         return retval == EINTR ? EAGAIN : retval;
     }
@@ -502,7 +749,7 @@ nl_dump_recv(struct nl_dump *dump, struct ofpbuf **bufferp)
 
     nlmsghdr = nl_msg_nlmsghdr(buffer);
     if (dump->seq != nlmsghdr->nlmsg_seq) {
-        VLOG_DBG_RL(&rl, "ignoring seq %"PRIu32" != expected %"PRIu32,
+        VLOG_DBG_RL(&rl, "ignoring seq %#"PRIx32" != expected %#"PRIx32,
                     nlmsghdr->nlmsg_seq, dump->seq);
         return EAGAIN;
     }
@@ -583,6 +830,13 @@ nl_dump_done(struct nl_dump *dump)
         }
     }
 
+    if (dump->sock) {
+        if (dump->sock->dump) {
+            dump->sock->dump = NULL;
+        } else {
+            nl_sock_destroy(dump->sock);
+        }
+    }
     ofpbuf_delete(dump->buffer);
     return dump->status == EOF ? 0 : dump->status;
 }
@@ -594,116 +848,210 @@ nl_sock_wait(const struct nl_sock *sock, short int events)
 {
     poll_fd_wait(sock->fd, events);
 }
+
+/* Returns the underlying fd for 'sock', for use in "poll()"-like operations
+ * that can't use nl_sock_wait().
+ *
+ * It's a little tricky to use the returned fd correctly, because nl_sock does
+ * "copy on write" to allow a single nl_sock to be used for notifications,
+ * transactions, and dumps.  If 'sock' is used only for notifications and
+ * transactions (and never for dump) then the usage is safe. */
+int
+nl_sock_fd(const struct nl_sock *sock)
+{
+    return sock->fd;
+}
+
+/* Returns the PID associated with this socket. */
+uint32_t
+nl_sock_pid(const struct nl_sock *sock)
+{
+    return sock->pid;
+}
 
 /* Miscellaneous.  */
 
+struct genl_family {
+    struct hmap_node hmap_node;
+    uint16_t id;
+    char *name;
+};
+
+static struct hmap genl_families = HMAP_INITIALIZER(&genl_families);
+
 static const struct nl_policy family_policy[CTRL_ATTR_MAX + 1] = {
     [CTRL_ATTR_FAMILY_ID] = {.type = NL_A_U16},
+    [CTRL_ATTR_MCAST_GROUPS] = {.type = NL_A_NESTED, .optional = true},
 };
 
-static int do_lookup_genl_family(const char *name)
+static struct genl_family *
+find_genl_family_by_id(uint16_t id)
+{
+    struct genl_family *family;
+
+    HMAP_FOR_EACH_IN_BUCKET (family, hmap_node, hash_int(id, 0),
+                             &genl_families) {
+        if (family->id == id) {
+            return family;
+        }
+    }
+    return NULL;
+}
+
+static void
+define_genl_family(uint16_t id, const char *name)
+{
+    struct genl_family *family = find_genl_family_by_id(id);
+
+    if (family) {
+        if (!strcmp(family->name, name)) {
+            return;
+        }
+        free(family->name);
+    } else {
+        family = xmalloc(sizeof *family);
+        family->id = id;
+        hmap_insert(&genl_families, &family->hmap_node, hash_int(id, 0));
+    }
+    family->name = xstrdup(name);
+}
+
+static const char *
+genl_family_to_name(uint16_t id)
+{
+    if (id == GENL_ID_CTRL) {
+        return "control";
+    } else {
+        struct genl_family *family = find_genl_family_by_id(id);
+        return family ? family->name : "unknown";
+    }
+}
+
+static int
+do_lookup_genl_family(const char *name, struct nlattr **attrs,
+                      struct ofpbuf **replyp)
 {
     struct nl_sock *sock;
     struct ofpbuf request, *reply;
-    struct nlattr *attrs[ARRAY_SIZE(family_policy)];
-    int retval;
+    int error;
 
-    retval = nl_sock_create(NETLINK_GENERIC, 0, 0, 0, &sock);
-    if (retval) {
-        return -retval;
+    *replyp = NULL;
+    error = nl_sock_create(NETLINK_GENERIC, &sock);
+    if (error) {
+        return error;
     }
 
     ofpbuf_init(&request, 0);
     nl_msg_put_genlmsghdr(&request, 0, GENL_ID_CTRL, NLM_F_REQUEST,
                           CTRL_CMD_GETFAMILY, 1);
     nl_msg_put_string(&request, CTRL_ATTR_FAMILY_NAME, name);
-    retval = nl_sock_transact(sock, &request, &reply);
+    error = nl_sock_transact(sock, &request, &reply);
     ofpbuf_uninit(&request);
-    if (retval) {
+    if (error) {
         nl_sock_destroy(sock);
-        return -retval;
+        return error;
     }
 
     if (!nl_policy_parse(reply, NLMSG_HDRLEN + GENL_HDRLEN,
-                         family_policy, attrs, ARRAY_SIZE(family_policy))) {
+                         family_policy, attrs, ARRAY_SIZE(family_policy))
+        || nl_attr_get_u16(attrs[CTRL_ATTR_FAMILY_ID]) == 0) {
         nl_sock_destroy(sock);
         ofpbuf_delete(reply);
-        return -EPROTO;
+        return EPROTO;
     }
 
-    retval = nl_attr_get_u16(attrs[CTRL_ATTR_FAMILY_ID]);
-    if (retval == 0) {
-        retval = -EPROTO;
-    }
     nl_sock_destroy(sock);
-    ofpbuf_delete(reply);
-    return retval;
+    *replyp = reply;
+    return 0;
 }
 
-/* If '*number' is 0, translates the given Generic Netlink family 'name' to a
- * number and stores it in '*number'.  If successful, returns 0 and the caller
- * may use '*number' as the family number.  On failure, returns a positive
- * errno value and '*number' caches the errno value. */
+/* Finds the multicast group called 'group_name' in genl family 'family_name'.
+ * When successful, writes its result to 'multicast_group' and returns 0.
+ * Otherwise, clears 'multicast_group' and returns a positive error code.
+ *
+ * Some kernels do not support looking up a multicast group with this function.
+ * In this case, 'multicast_group' will be populated with 'fallback'. */
 int
-nl_lookup_genl_family(const char *name, int *number)
+nl_lookup_genl_mcgroup(const char *family_name, const char *group_name,
+                       unsigned int *multicast_group, unsigned int fallback)
 {
-    if (*number == 0) {
-        *number = do_lookup_genl_family(name);
-        assert(*number != 0);
+    struct nlattr *family_attrs[ARRAY_SIZE(family_policy)];
+    const struct nlattr *mc;
+    struct ofpbuf *reply;
+    unsigned int left;
+    int error;
+
+    *multicast_group = 0;
+    error = do_lookup_genl_family(family_name, family_attrs, &reply);
+    if (error) {
+        return error;
     }
-    return *number > 0 ? 0 : -*number;
-}
-
-/* Netlink PID.
- *
- * Every Netlink socket must be bound to a unique 32-bit PID.  By convention,
- * programs that have a single Netlink socket use their Unix process ID as PID,
- * and programs with multiple Netlink sockets add a unique per-socket
- * identifier in the bits above the Unix process ID.
- *
- * The kernel has Netlink PID 0.
- */
 
-/* Parameters for how many bits in the PID should come from the Unix process ID
- * and how many unique per-socket. */
-#define SOCKET_BITS 10
-#define MAX_SOCKETS (1u << SOCKET_BITS)
+    if (!family_attrs[CTRL_ATTR_MCAST_GROUPS]) {
+        *multicast_group = fallback;
+        VLOG_WARN("%s-%s: has no multicast group, using fallback %d",
+                  family_name, group_name, *multicast_group);
+        error = 0;
+        goto exit;
+    }
 
-#define PROCESS_BITS (32 - SOCKET_BITS)
-#define MAX_PROCESSES (1u << PROCESS_BITS)
-#define PROCESS_MASK ((uint32_t) (MAX_PROCESSES - 1))
+    NL_NESTED_FOR_EACH (mc, left, family_attrs[CTRL_ATTR_MCAST_GROUPS]) {
+        static const struct nl_policy mc_policy[] = {
+            [CTRL_ATTR_MCAST_GRP_ID] = {.type = NL_A_U32},
+            [CTRL_ATTR_MCAST_GRP_NAME] = {.type = NL_A_STRING},
+        };
 
-/* Bit vector of unused socket identifiers. */
-static uint32_t avail_sockets[ROUND_UP(MAX_SOCKETS, 32)];
+        struct nlattr *mc_attrs[ARRAY_SIZE(mc_policy)];
+        const char *mc_name;
 
-/* Allocates and returns a new Netlink PID. */
-static int
-alloc_pid(uint32_t *pid)
-{
-    int i;
+        if (!nl_parse_nested(mc, mc_policy, mc_attrs, ARRAY_SIZE(mc_policy))) {
+            error = EPROTO;
+            goto exit;
+        }
 
-    for (i = 0; i < MAX_SOCKETS; i++) {
-        if ((avail_sockets[i / 32] & (1u << (i % 32))) == 0) {
-            avail_sockets[i / 32] |= 1u << (i % 32);
-            *pid = (getpid() & PROCESS_MASK) | (i << PROCESS_BITS);
-            return 0;
+        mc_name = nl_attr_get_string(mc_attrs[CTRL_ATTR_MCAST_GRP_NAME]);
+        if (!strcmp(group_name, mc_name)) {
+            *multicast_group =
+                nl_attr_get_u32(mc_attrs[CTRL_ATTR_MCAST_GRP_ID]);
+            error = 0;
+            goto exit;
         }
     }
-    VLOG_ERR("netlink pid space exhausted");
-    return ENOBUFS;
+    error = EPROTO;
+
+exit:
+    ofpbuf_delete(reply);
+    return error;
 }
 
-/* Makes the specified 'pid' available for reuse. */
-static void
-free_pid(uint32_t pid)
+/* If '*number' is 0, translates the given Generic Netlink family 'name' to a
+ * number and stores it in '*number'.  If successful, returns 0 and the caller
+ * may use '*number' as the family number.  On failure, returns a positive
+ * errno value and '*number' caches the errno value. */
+int
+nl_lookup_genl_family(const char *name, int *number)
 {
-    int sock = pid >> PROCESS_BITS;
-    assert(avail_sockets[sock / 32] & (1u << (sock % 32)));
-    avail_sockets[sock / 32] &= ~(1u << (sock % 32));
+    if (*number == 0) {
+        struct nlattr *attrs[ARRAY_SIZE(family_policy)];
+        struct ofpbuf *reply;
+        int error;
+
+        error = do_lookup_genl_family(name, attrs, &reply);
+        if (!error) {
+            *number = nl_attr_get_u16(attrs[CTRL_ATTR_FAMILY_ID]);
+            define_genl_family(*number, name);
+        } else {
+            *number = -error;
+        }
+        ofpbuf_delete(reply);
+
+        assert(*number != 0);
+    }
+    return *number > 0 ? 0 : -*number;
 }
 
 static void
-nlmsghdr_to_string(const struct nlmsghdr *h, struct ds *ds)
+nlmsghdr_to_string(const struct nlmsghdr *h, int protocol, struct ds *ds)
 {
     struct nlmsg_flag {
         unsigned int bits;
@@ -734,6 +1082,8 @@ nlmsghdr_to_string(const struct nlmsghdr *h, struct ds *ds)
         ds_put_cstr(ds, "(overrun)");
     } else if (h->nlmsg_type < NLMSG_MIN_TYPE) {
         ds_put_cstr(ds, "(reserved)");
+    } else if (protocol == NETLINK_GENERIC) {
+        ds_put_format(ds, "(%s)", genl_family_to_name(h->nlmsg_type));
     } else {
         ds_put_cstr(ds, "(family-defined)");
     }
@@ -748,19 +1098,17 @@ nlmsghdr_to_string(const struct nlmsghdr *h, struct ds *ds)
     if (flags_left) {
         ds_put_format(ds, "[OTHER:%"PRIx16"]", flags_left);
     }
-    ds_put_format(ds, ", seq=%"PRIx32", pid=%"PRIu32"(%d:%d))",
-                  h->nlmsg_seq, h->nlmsg_pid,
-                  (int) (h->nlmsg_pid & PROCESS_MASK),
-                  (int) (h->nlmsg_pid >> PROCESS_BITS));
+    ds_put_format(ds, ", seq=%"PRIx32", pid=%"PRIu32,
+                  h->nlmsg_seq, h->nlmsg_pid);
 }
 
 static char *
-nlmsg_to_string(const struct ofpbuf *buffer)
+nlmsg_to_string(const struct ofpbuf *buffer, int protocol)
 {
     struct ds ds = DS_EMPTY_INITIALIZER;
     const struct nlmsghdr *h = ofpbuf_at(buffer, 0, NLMSG_HDRLEN);
     if (h) {
-        nlmsghdr_to_string(h, &ds);
+        nlmsghdr_to_string(h, protocol, &ds);
         if (h->nlmsg_type == NLMSG_ERROR) {
             const struct nlmsgerr *e;
             e = ofpbuf_at(buffer, NLMSG_HDRLEN,
@@ -771,7 +1119,7 @@ nlmsg_to_string(const struct ofpbuf *buffer)
                     ds_put_format(&ds, "(%s)", strerror(-e->error));
                 }
                 ds_put_cstr(&ds, ", in-reply-to(");
-                nlmsghdr_to_string(&e->msg, &ds);
+                nlmsghdr_to_string(&e->msg, protocol, &ds);
                 ds_put_cstr(&ds, "))");
             } else {
                 ds_put_cstr(&ds, " error(truncated)");
@@ -787,6 +1135,12 @@ nlmsg_to_string(const struct ofpbuf *buffer)
             } else {
                 ds_put_cstr(&ds, " done(truncated)");
             }
+        } else if (protocol == NETLINK_GENERIC) {
+            struct genlmsghdr *genl = nl_msg_genlmsghdr(buffer);
+            if (genl) {
+                ds_put_format(&ds, ",genl(cmd=%"PRIu8",version=%"PRIu8")",
+                              genl->cmd, genl->version);
+            }
         }
     } else {
         ds_put_cstr(&ds, "nl(truncated)");
@@ -796,7 +1150,7 @@ nlmsg_to_string(const struct ofpbuf *buffer)
 
 static void
 log_nlmsg(const char *function, int error,
-          const void *message, size_t size)
+          const void *message, size_t size, int protocol)
 {
     struct ofpbuf buffer;
     char *nlmsg;
@@ -806,7 +1160,7 @@ log_nlmsg(const char *function, int error,
     }
 
     ofpbuf_use_const(&buffer, message, size);
-    nlmsg = nlmsg_to_string(&buffer);
+    nlmsg = nlmsg_to_string(&buffer, protocol);
     VLOG_DBG_RL(&rl, "%s (%s): %s", function, strerror(error), nlmsg);
     free(nlmsg);
 }