[openvswitch] / lib / socket-util.c

/*
 * Copyright (c) 2008, 2009, 2010, 2011, 2012 Nicira, Inc.
 *
 * Licensed under the Apache License, Version 2.0 (the "License");
 * you may not use this file except in compliance with the License.
 * You may obtain a copy of the License at:
 *
 *     http://www.apache.org/licenses/LICENSE-2.0
 *
 * Unless required by applicable law or agreed to in writing, software
 * distributed under the License is distributed on an "AS IS" BASIS,
 * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 * See the License for the specific language governing permissions and
 * limitations under the License.
 */

#include <config.h>
#include "socket-util.h"
#include <arpa/inet.h>
#include <assert.h>
#include <errno.h>
#include <fcntl.h>
#include <net/if.h>
#include <netdb.h>
#include <poll.h>
#include <stddef.h>
#include <stdio.h>
#include <stdlib.h>
#include <string.h>
#include <sys/resource.h>
#include <sys/socket.h>
#include <sys/stat.h>
#include <sys/un.h>
#include <unistd.h>
#include "dynamic-string.h"
#include "fatal-signal.h"
#include "packets.h"
#include "util.h"
#include "vlog.h"
#if AF_PACKET && __linux__
#include <linux/if_packet.h>
#endif
#ifdef HAVE_NETLINK
#include "netlink-protocol.h"
#include "netlink-socket.h"
#endif

VLOG_DEFINE_THIS_MODULE(socket_util);

/* #ifdefs make it a pain to maintain code: you have to try to build both ways.
 * Thus, this file compiles all of the code regardless of the target, by
 * writing "if (LINUX)" instead of "#ifdef __linux__". */
#ifdef __linux__
#define LINUX 1
#else
#define LINUX 0
#endif

#ifndef O_DIRECTORY
#define O_DIRECTORY 0
#endif

static int getsockopt_int(int fd, int level, int option, const char *optname,
                          int *valuep);

/* Sets 'fd' to non-blocking mode.  Returns 0 if successful, otherwise a
 * positive errno value. */
int
set_nonblocking(int fd)
{
    int flags = fcntl(fd, F_GETFL, 0);
    if (flags != -1) {
        if (fcntl(fd, F_SETFL, flags | O_NONBLOCK) != -1) {
            return 0;
        } else {
            VLOG_ERR("fcntl(F_SETFL) failed: %s", strerror(errno));
            return errno;
        }
    } else {
        VLOG_ERR("fcntl(F_GETFL) failed: %s", strerror(errno));
        return errno;
    }
}

void
xset_nonblocking(int fd)
{
    if (set_nonblocking(fd)) {
        exit(EXIT_FAILURE);
    }
}

static int
set_dscp(int fd, uint8_t dscp)
{
    if (dscp > 63) {
        return EINVAL;
    }

    dscp = dscp << 2;
    if (setsockopt(fd, IPPROTO_IP, IP_TOS, &dscp, sizeof dscp)) {
        return errno;
    }

    return 0;
}

static bool
rlim_is_finite(rlim_t limit)
{
    if (limit == RLIM_INFINITY) {
        return false;
    }

#ifdef RLIM_SAVED_CUR           /* FreeBSD 8.0 lacks RLIM_SAVED_CUR. */
    if (limit == RLIM_SAVED_CUR) {
        return false;
    }
#endif

#ifdef RLIM_SAVED_MAX           /* FreeBSD 8.0 lacks RLIM_SAVED_MAX. */
    if (limit == RLIM_SAVED_MAX) {
        return false;
    }
#endif

    return true;
}

/* Returns the maximum valid FD value, plus 1. */
int
get_max_fds(void)
{
    static int max_fds = -1;
    if (max_fds < 0) {
        struct rlimit r;
        if (!getrlimit(RLIMIT_NOFILE, &r) && rlim_is_finite(r.rlim_cur)) {
            max_fds = r.rlim_cur;
        } else {
            VLOG_WARN("failed to obtain fd limit, defaulting to 1024");
            max_fds = 1024;
        }
    }
    return max_fds;
}

/* Translates 'host_name', which must be a string representation of an IP
 * address, into a numeric IP address in '*addr'.  Returns 0 if successful,
 * otherwise a positive errno value. */
int
lookup_ip(const char *host_name, struct in_addr *addr)
{
    if (!inet_aton(host_name, addr)) {
        static struct vlog_rate_limit rl = VLOG_RATE_LIMIT_INIT(1, 5);
        VLOG_ERR_RL(&rl, "\"%s\" is not a valid IP address", host_name);
        return ENOENT;
    }
    return 0;
}

/* Translates 'host_name', which must be a string representation of an IPv6
 * address, into a numeric IPv6 address in '*addr'.  Returns 0 if successful,
 * otherwise a positive errno value. */
int
lookup_ipv6(const char *host_name, struct in6_addr *addr)
{
    if (inet_pton(AF_INET6, host_name, addr) != 1) {
        static struct vlog_rate_limit rl = VLOG_RATE_LIMIT_INIT(1, 5);
        VLOG_ERR_RL(&rl, "\"%s\" is not a valid IPv6 address", host_name);
        return ENOENT;
    }
    return 0;
}

/* Translates 'host_name', which must be a host name or a string representation
 * of an IP address, into a numeric IP address in '*addr'.  Returns 0 if
 * successful, otherwise a positive errno value.
 *
 * Most Open vSwitch code should not use this because it causes deadlocks:
 * gethostbyname() sends out a DNS request but that starts a new flow for which
 * OVS must set up a flow, but it can't because it's waiting for a DNS reply.
 * The synchronous lookup also delays other activty.  (Of course we can solve
 * this but it doesn't seem worthwhile quite yet.)  */
int
lookup_hostname(const char *host_name, struct in_addr *addr)
{
    struct hostent *h;

    if (inet_aton(host_name, addr)) {
        return 0;
    }

    h = gethostbyname(host_name);
    if (h) {
        *addr = *(struct in_addr *) h->h_addr;
        return 0;
    }

    return (h_errno == HOST_NOT_FOUND ? ENOENT
            : h_errno == TRY_AGAIN ? EAGAIN
            : h_errno == NO_RECOVERY ? EIO
            : h_errno == NO_ADDRESS ? ENXIO
            : EINVAL);
}

/* Returns the error condition associated with socket 'fd' and resets the
 * socket's error status. */
int
get_socket_error(int fd)
{
    int error;

    if (getsockopt_int(fd, SOL_SOCKET, SO_ERROR, "SO_ERROR", &error)) {
        error = errno;
    }
    return error;
}

int
check_connection_completion(int fd)
{
    struct pollfd pfd;
    int retval;

    pfd.fd = fd;
    pfd.events = POLLOUT;
    do {
        retval = poll(&pfd, 1, 0);
    } while (retval < 0 && errno == EINTR);
    if (retval == 1) {
        return get_socket_error(fd);
    } else if (retval < 0) {
        static struct vlog_rate_limit rl = VLOG_RATE_LIMIT_INIT(5, 10);
        VLOG_ERR_RL(&rl, "poll: %s", strerror(errno));
        return errno;
    } else {
        return EAGAIN;
    }
}

/* Drain all the data currently in the receive queue of a datagram socket (and
 * possibly additional data).  There is no way to know how many packets are in
 * the receive queue, but we do know that the total number of bytes queued does
 * not exceed the receive buffer size, so we pull packets until none are left
 * or we've read that many bytes. */
int
drain_rcvbuf(int fd)
{
    int rcvbuf;

    rcvbuf = get_socket_rcvbuf(fd);
    if (rcvbuf < 0) {
        return -rcvbuf;
    }

    while (rcvbuf > 0) {
        /* In Linux, specifying MSG_TRUNC in the flags argument causes the
         * datagram length to be returned, even if that is longer than the
         * buffer provided.  Thus, we can use a 1-byte buffer to discard the
         * incoming datagram and still be able to account how many bytes were
         * removed from the receive buffer.
         *
         * On other Unix-like OSes, MSG_TRUNC has no effect in the flags
         * argument. */
        char buffer[LINUX ? 1 : 2048];
        ssize_t n_bytes = recv(fd, buffer, sizeof buffer,
                               MSG_TRUNC | MSG_DONTWAIT);
        if (n_bytes <= 0 || n_bytes >= rcvbuf) {
            break;
        }
        rcvbuf -= n_bytes;
    }
    return 0;
}

/* Returns the size of socket 'sock''s receive buffer (SO_RCVBUF), or a
 * negative errno value if an error occurs. */
int
get_socket_rcvbuf(int sock)
{
    int rcvbuf;
    int error;

    error = getsockopt_int(sock, SOL_SOCKET, SO_RCVBUF, "SO_RCVBUF", &rcvbuf);
    return error ? -error : rcvbuf;
}

/* Reads and discards up to 'n' datagrams from 'fd', stopping as soon as no
 * more data can be immediately read.  ('fd' should therefore be in
 * non-blocking mode.)*/
void
drain_fd(int fd, size_t n_packets)
{
    for (; n_packets > 0; n_packets--) {
        /* 'buffer' only needs to be 1 byte long in most circumstances.  This
         * size is defensive against the possibility that we someday want to
         * use a Linux tap device without TUN_NO_PI, in which case a buffer
         * smaller than sizeof(struct tun_pi) will give EINVAL on read. */
        char buffer[128];
        if (read(fd, buffer, sizeof buffer) <= 0) {
            break;
        }
    }
}

/* Stores in '*un' a sockaddr_un that refers to file 'name'.  Stores in
 * '*un_len' the size of the sockaddr_un. */
static void
make_sockaddr_un__(const char *name, struct sockaddr_un *un, socklen_t *un_len)
{
    un->sun_family = AF_UNIX;
    ovs_strzcpy(un->sun_path, name, sizeof un->sun_path);
    *un_len = (offsetof(struct sockaddr_un, sun_path)
                + strlen (un->sun_path) + 1);
}

/* Stores in '*un' a sockaddr_un that refers to file 'name'.  Stores in
 * '*un_len' the size of the sockaddr_un.
 *
 * Returns 0 on success, otherwise a positive errno value.  On success,
 * '*dirfdp' is either -1 or a nonnegative file descriptor that the caller
 * should close after using '*un' to bind or connect.  On failure, '*dirfdp' is
 * -1. */
static int
make_sockaddr_un(const char *name, struct sockaddr_un *un, socklen_t *un_len,
                 int *dirfdp)
{
    enum { MAX_UN_LEN = sizeof un->sun_path - 1 };

    *dirfdp = -1;
    if (strlen(name) > MAX_UN_LEN) {
        static struct vlog_rate_limit rl = VLOG_RATE_LIMIT_INIT(1, 1);

        if (LINUX) {
            /* 'name' is too long to fit in a sockaddr_un, but we have a
             * workaround for that on Linux: shorten it by opening a file
             * descriptor for the directory part of the name and indirecting
             * through /proc/self/fd/<dirfd>/<basename>. */
            char *dir, *base;
            char *short_name;
            int dirfd;

            dir = dir_name(name);
            base = base_name(name);

            dirfd = open(dir, O_DIRECTORY | O_RDONLY);
            if (dirfd < 0) {
                free(base);
                free(dir);
                return errno;
            }

            short_name = xasprintf("/proc/self/fd/%d/%s", dirfd, base);
            free(dir);
            free(base);

            if (strlen(short_name) <= MAX_UN_LEN) {
                make_sockaddr_un__(short_name, un, un_len);
                free(short_name);
                *dirfdp = dirfd;
                return 0;
            }
            free(short_name);
            close(dirfd);

            VLOG_WARN_RL(&rl, "Unix socket name %s is longer than maximum "
                         "%d bytes (even shortened)", name, MAX_UN_LEN);
        } else {
            /* 'name' is too long and we have no workaround. */
            VLOG_WARN_RL(&rl, "Unix socket name %s is longer than maximum "
                         "%d bytes", name, MAX_UN_LEN);
        }

        return ENAMETOOLONG;
    } else {
        make_sockaddr_un__(name, un, un_len);
        return 0;
    }
}

/* Binds Unix domain socket 'fd' to a file with permissions 0700. */
static int
bind_unix_socket(int fd, struct sockaddr *sun, socklen_t sun_len)
{
    /* According to _Unix Network Programming_, umask should affect bind(). */
    mode_t old_umask = umask(0077);
    int error = bind(fd, sun, sun_len) ? errno : 0;
    umask(old_umask);
    return error;
}

/* Creates a Unix domain socket in the given 'style' (either SOCK_DGRAM or
 * SOCK_STREAM) that is bound to '*bind_path' (if 'bind_path' is non-null) and
 * connected to '*connect_path' (if 'connect_path' is non-null).  If 'nonblock'
 * is true, the socket is made non-blocking.
 *
 * Returns the socket's fd if successful, otherwise a negative errno value. */
int
make_unix_socket(int style, bool nonblock,
                 const char *bind_path, const char *connect_path)
{
    int error;
    int fd;

    fd = socket(PF_UNIX, style, 0);
    if (fd < 0) {
        return -errno;
    }

    /* Set nonblocking mode right away, if we want it.  This prevents blocking
     * in connect(), if connect_path != NULL.  (In turn, that's a corner case:
     * it will only happen if style is SOCK_STREAM or SOCK_SEQPACKET, and only
     * if a backlog of un-accepted connections has built up in the kernel.)  */
    if (nonblock) {
        int flags = fcntl(fd, F_GETFL, 0);
        if (flags == -1) {
            error = errno;
            goto error;
        }
        if (fcntl(fd, F_SETFL, flags | O_NONBLOCK) == -1) {
            error = errno;
            goto error;
        }
    }

    if (bind_path) {
        struct sockaddr_un un;
        socklen_t un_len;
        int dirfd;

        if (unlink(bind_path) && errno != ENOENT) {
            VLOG_WARN("unlinking \"%s\": %s\n", bind_path, strerror(errno));
        }
        fatal_signal_add_file_to_unlink(bind_path);

        error = make_sockaddr_un(bind_path, &un, &un_len, &dirfd);
        if (!error) {
            error = bind_unix_socket(fd, (struct sockaddr *) &un, un_len);
        }
        if (dirfd >= 0) {
            close(dirfd);
        }
        if (error) {
            goto error;
        }
    }

    if (connect_path) {
        struct sockaddr_un un;
        socklen_t un_len;
        int dirfd;

        error = make_sockaddr_un(connect_path, &un, &un_len, &dirfd);
        if (!error
            && connect(fd, (struct sockaddr*) &un, un_len)
            && errno != EINPROGRESS) {
            error = errno;
        }
        if (dirfd >= 0) {
            close(dirfd);
        }
        if (error) {
            goto error;
        }
    }

    return fd;

error:
    if (error == EAGAIN) {
        error = EPROTO;
    }
    if (bind_path) {
        fatal_signal_unlink_file_now(bind_path);
    }
    close(fd);
    return -error;
}

int
get_unix_name_len(socklen_t sun_len)
{
    return (sun_len >= offsetof(struct sockaddr_un, sun_path)
            ? sun_len - offsetof(struct sockaddr_un, sun_path)
            : 0);
}

ovs_be32
guess_netmask(ovs_be32 ip_)
{
    uint32_t ip = ntohl(ip_);
    return ((ip >> 31) == 0 ? htonl(0xff000000)   /* Class A */
            : (ip >> 30) == 2 ? htonl(0xffff0000) /* Class B */
            : (ip >> 29) == 6 ? htonl(0xffffff00) /* Class C */
            : htonl(0));                          /* ??? */
}

/* Parses 'target', which should be a string in the format "<host>[:<port>]".
 * <host> is required.  If 'default_port' is nonzero then <port> is optional
 * and defaults to 'default_port'.
 *
 * On success, returns true and stores the parsed remote address into '*sinp'.
 * On failure, logs an error, stores zeros into '*sinp', and returns false. */
bool
inet_parse_active(const char *target_, uint16_t default_port,
                  struct sockaddr_in *sinp)
{
    char *target = xstrdup(target_);
    char *save_ptr = NULL;
    const char *host_name;
    const char *port_string;
    bool ok = false;

    /* Defaults. */
    sinp->sin_family = AF_INET;
    sinp->sin_port = htons(default_port);

    /* Tokenize. */
    host_name = strtok_r(target, ":", &save_ptr);
    port_string = strtok_r(NULL, ":", &save_ptr);
    if (!host_name) {
        VLOG_ERR("%s: bad peer name format", target_);
        goto exit;
    }

    /* Look up IP, port. */
    if (lookup_ip(host_name, &sinp->sin_addr)) {
        goto exit;
    }
    if (port_string && atoi(port_string)) {
        sinp->sin_port = htons(atoi(port_string));
    } else if (!default_port) {
        VLOG_ERR("%s: port number must be specified", target_);
        goto exit;
    }

    ok = true;

exit:
    if (!ok) {
        memset(sinp, 0, sizeof *sinp);
    }
    free(target);
    return ok;
}

/* Opens a non-blocking IPv4 socket of the specified 'style' and connects to
 * 'target', which should be a string in the format "<host>[:<port>]".  <host>
 * is required.  If 'default_port' is nonzero then <port> is optional and
 * defaults to 'default_port'.
 *
 * 'style' should be SOCK_STREAM (for TCP) or SOCK_DGRAM (for UDP).
 *
 * On success, returns 0 (indicating connection complete) or EAGAIN (indicating
 * connection in progress), in which case the new file descriptor is stored
 * into '*fdp'.  On failure, returns a positive errno value other than EAGAIN
 * and stores -1 into '*fdp'.
 *
 * If 'sinp' is non-null, then on success the target address is stored into
 * '*sinp'.
 *
 * 'dscp' becomes the DSCP bits in the IP headers for the new connection.  It
 * should be in the range [0, 63] and will automatically be shifted to the
 * appropriately place in the IP tos field. */
int
inet_open_active(int style, const char *target, uint16_t default_port,
                 struct sockaddr_in *sinp, int *fdp, uint8_t dscp)
{
    struct sockaddr_in sin;
    int fd = -1;
    int error;

    /* Parse. */
    if (!inet_parse_active(target, default_port, &sin)) {
        error = EAFNOSUPPORT;
        goto exit;
    }

    /* Create non-blocking socket. */
    fd = socket(AF_INET, style, 0);
    if (fd < 0) {
        VLOG_ERR("%s: socket: %s", target, strerror(errno));
        error = errno;
        goto exit;
    }
    error = set_nonblocking(fd);
    if (error) {
        goto exit;
    }

    /* The dscp bits must be configured before connect() to ensure that the TOS
     * field is set during the connection establishment.  If set after
     * connect(), the handshake SYN frames will be sent with a TOS of 0. */
    error = set_dscp(fd, dscp);
    if (error) {
        VLOG_ERR("%s: socket: %s", target, strerror(error));
        goto exit;
    }

    /* Connect. */
    error = connect(fd, (struct sockaddr *) &sin, sizeof sin) == 0 ? 0 : errno;
    if (error == EINPROGRESS) {
        error = EAGAIN;
    }

exit:
    if (!error || error == EAGAIN) {
        if (sinp) {
            *sinp = sin;
        }
    } else if (fd >= 0) {
        close(fd);
    }
    *fdp = fd;
    return error;
}

/* Parses 'target', which should be a string in the format "[<port>][:<ip>]":
 *
 *      - If 'default_port' is -1, then <port> is required.  Otherwise, if
 *        <port> is omitted, then 'default_port' is used instead.
 *
 *      - If <port> (or 'default_port', if used) is 0, then no port is bound
 *        and the TCP/IP stack will select a port.
 *
 *      - If <ip> is omitted then the IP address is wildcarded.
 *
 * If successful, stores the address into '*sinp' and returns true; otherwise
 * zeros '*sinp' and returns false. */
bool
inet_parse_passive(const char *target_, int default_port,
                   struct sockaddr_in *sinp)
{
    char *target = xstrdup(target_);
    char *string_ptr = target;
    const char *host_name;
    const char *port_string;
    bool ok = false;
    int port;

    /* Address defaults. */
    memset(sinp, 0, sizeof *sinp);
    sinp->sin_family = AF_INET;
    sinp->sin_addr.s_addr = htonl(INADDR_ANY);
    sinp->sin_port = htons(default_port);

    /* Parse optional port number. */
    port_string = strsep(&string_ptr, ":");
    if (port_string && str_to_int(port_string, 10, &port)) {
        sinp->sin_port = htons(port);
    } else if (default_port < 0) {
        VLOG_ERR("%s: port number must be specified", target_);
        goto exit;
    }

    /* Parse optional bind IP. */
    host_name = strsep(&string_ptr, ":");
    if (host_name && host_name[0] && lookup_ip(host_name, &sinp->sin_addr)) {
        goto exit;
    }

    ok = true;

exit:
    if (!ok) {
        memset(sinp, 0, sizeof *sinp);
    }
    free(target);
    return ok;
}


/* Opens a non-blocking IPv4 socket of the specified 'style', binds to
 * 'target', and listens for incoming connections.  Parses 'target' in the same
 * way was inet_parse_passive().
 *
 * 'style' should be SOCK_STREAM (for TCP) or SOCK_DGRAM (for UDP).
 *
 * For TCP, the socket will have SO_REUSEADDR turned on.
 *
 * On success, returns a non-negative file descriptor.  On failure, returns a
 * negative errno value.
 *
 * If 'sinp' is non-null, then on success the bound address is stored into
 * '*sinp'.
 *
 * 'dscp' becomes the DSCP bits in the IP headers for the new connection.  It
 * should be in the range [0, 63] and will automatically be shifted to the
 * appropriately place in the IP tos field. */
int
inet_open_passive(int style, const char *target, int default_port,
                  struct sockaddr_in *sinp, uint8_t dscp)
{
    struct sockaddr_in sin;
    int fd = 0, error;
    unsigned int yes = 1;

    if (!inet_parse_passive(target, default_port, &sin)) {
        return -EAFNOSUPPORT;
    }

    /* Create non-blocking socket, set SO_REUSEADDR. */
    fd = socket(AF_INET, style, 0);
    if (fd < 0) {
        error = errno;
        VLOG_ERR("%s: socket: %s", target, strerror(error));
        return -error;
    }
    error = set_nonblocking(fd);
    if (error) {
        goto error;
    }
    if (style == SOCK_STREAM
        && setsockopt(fd, SOL_SOCKET, SO_REUSEADDR, &yes, sizeof yes) < 0) {
        error = errno;
        VLOG_ERR("%s: setsockopt(SO_REUSEADDR): %s", target, strerror(error));
        goto error;
    }

    /* Bind. */
    if (bind(fd, (struct sockaddr *) &sin, sizeof sin) < 0) {
        error = errno;
        VLOG_ERR("%s: bind: %s", target, strerror(error));
        goto error;
    }

    /* The dscp bits must be configured before connect() to ensure that the TOS
     * field is set during the connection establishment.  If set after
     * connect(), the handshake SYN frames will be sent with a TOS of 0. */
    error = set_dscp(fd, dscp);
    if (error) {
        VLOG_ERR("%s: socket: %s", target, strerror(error));
        goto error;
    }

    /* Listen. */
    if (style == SOCK_STREAM && listen(fd, 10) < 0) {
        error = errno;
        VLOG_ERR("%s: listen: %s", target, strerror(error));
        goto error;
    }

    if (sinp) {
        socklen_t sin_len = sizeof sin;
        if (getsockname(fd, (struct sockaddr *) &sin, &sin_len) < 0){
            error = errno;
            VLOG_ERR("%s: getsockname: %s", target, strerror(error));
            goto error;
        }
        if (sin.sin_family != AF_INET || sin_len != sizeof sin) {
            error = EAFNOSUPPORT;
            VLOG_ERR("%s: getsockname: invalid socket name", target);
            goto error;
        }
        *sinp = sin;
    }

    return fd;

error:
    close(fd);
    return -error;
}

/* Returns a readable and writable fd for /dev/null, if successful, otherwise
 * a negative errno value.  The caller must not close the returned fd (because
 * the same fd will be handed out to subsequent callers). */
int
get_null_fd(void)
{
    static int null_fd = -1;
    if (null_fd < 0) {
        null_fd = open("/dev/null", O_RDWR);
        if (null_fd < 0) {
            int error = errno;
            VLOG_ERR("could not open /dev/null: %s", strerror(error));
            return -error;
        }
    }
    return null_fd;
}

int
read_fully(int fd, void *p_, size_t size, size_t *bytes_read)
{
    uint8_t *p = p_;

    *bytes_read = 0;
    while (size > 0) {
        ssize_t retval = read(fd, p, size);
        if (retval > 0) {
            *bytes_read += retval;
            size -= retval;
            p += retval;
        } else if (retval == 0) {
            return EOF;
        } else if (errno != EINTR) {
            return errno;
        }
    }
    return 0;
}

int
write_fully(int fd, const void *p_, size_t size, size_t *bytes_written)
{
    const uint8_t *p = p_;

    *bytes_written = 0;
    while (size > 0) {
        ssize_t retval = write(fd, p, size);
        if (retval > 0) {
            *bytes_written += retval;
            size -= retval;
            p += retval;
        } else if (retval == 0) {
            VLOG_WARN("write returned 0");
            return EPROTO;
        } else if (errno != EINTR) {
            return errno;
        }
    }
    return 0;
}

/* Given file name 'file_name', fsyncs the directory in which it is contained.
 * Returns 0 if successful, otherwise a positive errno value. */
int
fsync_parent_dir(const char *file_name)
{
    int error = 0;
    char *dir;
    int fd;

    dir = dir_name(file_name);
    fd = open(dir, O_RDONLY);
    if (fd >= 0) {
        if (fsync(fd)) {
            if (errno == EINVAL || errno == EROFS) {
                /* This directory does not support synchronization.  Not
                 * really an error. */
            } else {
                error = errno;
                VLOG_ERR("%s: fsync failed (%s)", dir, strerror(error));
            }
        }
        close(fd);
    } else {
        error = errno;
        VLOG_ERR("%s: open failed (%s)", dir, strerror(error));
    }
    free(dir);

    return error;
}

/* Obtains the modification time of the file named 'file_name' to the greatest
 * supported precision.  If successful, stores the mtime in '*mtime' and
 * returns 0.  On error, returns a positive errno value and stores zeros in
 * '*mtime'. */
int
get_mtime(const char *file_name, struct timespec *mtime)
{
    struct stat s;

    if (!stat(file_name, &s)) {
        mtime->tv_sec = s.st_mtime;

#if HAVE_STRUCT_STAT_ST_MTIM_TV_NSEC
        mtime->tv_nsec = s.st_mtim.tv_nsec;
#elif HAVE_STRUCT_STAT_ST_MTIMENSEC
        mtime->tv_nsec = s.st_mtimensec;
#else
        mtime->tv_nsec = 0;
#endif

        return 0;
    } else {
        mtime->tv_sec = mtime->tv_nsec = 0;
        return errno;
    }
}

void
xpipe(int fds[2])
{
    if (pipe(fds)) {
        VLOG_FATAL("failed to create pipe (%s)", strerror(errno));
    }
}

static int
getsockopt_int(int fd, int level, int option, const char *optname, int *valuep)
{
    static struct vlog_rate_limit rl = VLOG_RATE_LIMIT_INIT(5, 10);
    socklen_t len;
    int value;
    int error;

    len = sizeof value;
    if (getsockopt(fd, level, option, &value, &len)) {
        error = errno;
        VLOG_ERR_RL(&rl, "getsockopt(%s): %s", optname, strerror(error));
    } else if (len != sizeof value) {
        error = EINVAL;
        VLOG_ERR_RL(&rl, "getsockopt(%s): value is %u bytes (expected %zu)",
                    optname, (unsigned int) len, sizeof value);
    } else {
        error = 0;
    }

    *valuep = error ? 0 : value;
    return error;
}

static void
describe_sockaddr(struct ds *string, int fd,
                  int (*getaddr)(int, struct sockaddr *, socklen_t *))
{
    struct sockaddr_storage ss;
    socklen_t len = sizeof ss;

    if (!getaddr(fd, (struct sockaddr *) &ss, &len)) {
        if (ss.ss_family == AF_INET) {
            struct sockaddr_in sin;

            memcpy(&sin, &ss, sizeof sin);
            ds_put_format(string, IP_FMT":%"PRIu16,
                          IP_ARGS(&sin.sin_addr.s_addr), ntohs(sin.sin_port));
        } else if (ss.ss_family == AF_UNIX) {
            struct sockaddr_un sun;
            const char *null;
            size_t maxlen;

            memcpy(&sun, &ss, sizeof sun);
            maxlen = len - offsetof(struct sockaddr_un, sun_path);
            null = memchr(sun.sun_path, '\0', maxlen);
            ds_put_buffer(string, sun.sun_path,
                          null ? null - sun.sun_path : maxlen);
        }
#ifdef HAVE_NETLINK
        else if (ss.ss_family == AF_NETLINK) {
            int protocol;

/* SO_PROTOCOL was introduced in 2.6.32.  Support it regardless of the version
 * of the Linux kernel headers in use at build time. */
#ifndef SO_PROTOCOL
#define SO_PROTOCOL 38
#endif

            if (!getsockopt_int(fd, SOL_SOCKET, SO_PROTOCOL, "SO_PROTOCOL",
                                &protocol)) {
                switch (protocol) {
                case NETLINK_ROUTE:
                    ds_put_cstr(string, "NETLINK_ROUTE");
                    break;

                case NETLINK_GENERIC:
                    ds_put_cstr(string, "NETLINK_GENERIC");
                    break;

                default:
                    ds_put_format(string, "AF_NETLINK family %d", protocol);
                    break;
                }
            } else {
                ds_put_cstr(string, "AF_NETLINK");
            }
        }
#endif
#if AF_PACKET && __linux__
        else if (ss.ss_family == AF_PACKET) {
            struct sockaddr_ll sll;

            memcpy(&sll, &ss, sizeof sll);
            ds_put_cstr(string, "AF_PACKET");
            if (sll.sll_ifindex) {
                char name[IFNAMSIZ];

                if (if_indextoname(sll.sll_ifindex, name)) {
                    ds_put_format(string, "(%s)", name);
                } else {
                    ds_put_format(string, "(ifindex=%d)", sll.sll_ifindex);
                }
            }
            if (sll.sll_protocol) {
                ds_put_format(string, "(protocol=0x%"PRIu16")",
                              ntohs(sll.sll_protocol));
            }
        }
#endif
        else if (ss.ss_family == AF_UNSPEC) {
            ds_put_cstr(string, "AF_UNSPEC");
        } else {
            ds_put_format(string, "AF_%d", (int) ss.ss_family);
        }
    }
}


#ifdef __linux__
static void
put_fd_filename(struct ds *string, int fd)
{
    char buf[1024];
    char *linkname;
    int n;

    linkname = xasprintf("/proc/self/fd/%d", fd);
    n = readlink(linkname, buf, sizeof buf);
    if (n > 0) {
        ds_put_char(string, ' ');
        ds_put_buffer(string, buf, n);
        if (n > sizeof buf) {
            ds_put_cstr(string, "...");
        }
    }
    free(linkname);
}
#endif

/* Returns a malloc()'d string describing 'fd', for use in logging. */
char *
describe_fd(int fd)
{
    struct ds string;
    struct stat s;

    ds_init(&string);
    if (fstat(fd, &s)) {
        ds_put_format(&string, "fstat failed (%s)", strerror(errno));
    } else if (S_ISSOCK(s.st_mode)) {
        describe_sockaddr(&string, fd, getsockname);
        ds_put_cstr(&string, "<->");
        describe_sockaddr(&string, fd, getpeername);
    } else {
        ds_put_cstr(&string, (isatty(fd) ? "tty"
                              : S_ISDIR(s.st_mode) ? "directory"
                              : S_ISCHR(s.st_mode) ? "character device"
                              : S_ISBLK(s.st_mode) ? "block device"
                              : S_ISREG(s.st_mode) ? "file"
                              : S_ISFIFO(s.st_mode) ? "FIFO"
                              : S_ISLNK(s.st_mode) ? "symbolic link"
                              : "unknown"));
#ifdef __linux__
        put_fd_filename(&string, fd);
#endif
    }
    return ds_steal_cstr(&string);
}