lib/socket-util-unix.c

   1 /*
   2  * Copyright (c) 2014 Nicira, Inc.
   3  *
   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:
   7  *
   8  *     http://www.apache.org/licenses/LICENSE-2.0
   9  *
  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.
  15  */
  16
  17 #include <config.h>
  18 #include "socket-util.h"
  19 #include <errno.h>
  20 #include <fcntl.h>
  21 #include <net/if.h>
  22 #include <sys/ioctl.h>
  23 #include <sys/types.h>
  24 #include <sys/stat.h>
  25 #include <sys/un.h>
  26 #include <sys/wait.h>
  27 #include <unistd.h>
  28 #include "fatal-signal.h"
  29 #include "random.h"
  30 #include "util.h"
  31 #include "openvswitch/vlog.h"
  32
  33 VLOG_DEFINE_THIS_MODULE(socket_util_unix);
  34
  35 /* #ifdefs make it a pain to maintain code: you have to try to build both ways.
  36  * Thus, this file compiles all of the code regardless of the target, by
  37  * writing "if (LINUX)" instead of "#ifdef __linux__". */
  38 #ifdef __linux__
  39 #define LINUX 1
  40 #else
  41 #define LINUX 0
  42 #endif
  43
  44 #ifndef O_DIRECTORY
  45 #define O_DIRECTORY 0
  46 #endif
  47
  48 /* Maximum length of the sun_path member in a struct sockaddr_un, excluding
  49  * space for a null terminator. */
  50 #define MAX_UN_LEN (sizeof(((struct sockaddr_un *) 0)->sun_path) - 1)
  51
  52 void
  53 xpipe(int fds[2])
  54 {
  55     if (pipe(fds)) {
  56         VLOG_FATAL("failed to create pipe (%s)", ovs_strerror(errno));
  57     }
  58 }
  59
  60 void
  61 xpipe_nonblocking(int fds[2])
  62 {
  63     xpipe(fds);
  64     xset_nonblocking(fds[0]);
  65     xset_nonblocking(fds[1]);
  66 }
  67
  68 /* Drain all the data currently in the receive queue of a datagram socket (and
  69  * possibly additional data).  There is no way to know how many packets are in
  70  * the receive queue, but we do know that the total number of bytes queued does
  71  * not exceed the receive buffer size, so we pull packets until none are left
  72  * or we've read that many bytes. */
  73 int
  74 drain_rcvbuf(int fd)
  75 {
  76     int rcvbuf;
  77
  78     rcvbuf = get_socket_rcvbuf(fd);
  79     if (rcvbuf < 0) {
  80         return -rcvbuf;
  81     }
  82
  83     while (rcvbuf > 0) {
  84         /* In Linux, specifying MSG_TRUNC in the flags argument causes the
  85          * datagram length to be returned, even if that is longer than the
  86          * buffer provided.  Thus, we can use a 1-byte buffer to discard the
  87          * incoming datagram and still be able to account how many bytes were
  88          * removed from the receive buffer.
  89          *
  90          * On other Unix-like OSes, MSG_TRUNC has no effect in the flags
  91          * argument. */
  92         char buffer[LINUX ? 1 : 2048];
  93         ssize_t n_bytes = recv(fd, buffer, sizeof buffer,
  94                                MSG_TRUNC | MSG_DONTWAIT);
  95         if (n_bytes <= 0 || n_bytes >= rcvbuf) {
  96             break;
  97         }
  98         rcvbuf -= n_bytes;
  99     }
 100     return 0;
 101 }
 102
 103 /* Attempts to shorten 'name' by opening a file descriptor for the directory
 104  * part of the name and indirecting through /proc/self/fd/<dirfd>/<basename>.
 105  * On systems with Linux-like /proc, this works as long as <basename> isn't too
 106  * long.
 107  *
 108  * On success, returns 0 and stores the short name in 'short_name' and a
 109  * directory file descriptor to eventually be closed in '*dirfpd'. */
 110 static int
 111 shorten_name_via_proc(const char *name, char short_name[MAX_UN_LEN + 1],
 112                       int *dirfdp)
 113 {
 114     char *dir, *base;
 115     int dirfd;
 116     int len;
 117
 118     if (!LINUX) {
 119         return ENAMETOOLONG;
 120     }
 121
 122     dir = dir_name(name);
 123     dirfd = open(dir, O_DIRECTORY | O_RDONLY);
 124     if (dirfd < 0) {
 125         static struct vlog_rate_limit rl = VLOG_RATE_LIMIT_INIT(1, 1);
 126         int error = errno;
 127
 128         VLOG_WARN_RL(&rl, "%s: open failed (%s)", dir, ovs_strerror(error));
 129         free(dir);
 130
 131         return error;
 132     }
 133     free(dir);
 134
 135     base = base_name(name);
 136     len = snprintf(short_name, MAX_UN_LEN + 1,
 137                    "/proc/self/fd/%d/%s", dirfd, base);
 138     free(base);
 139
 140     if (len >= 0 && len <= MAX_UN_LEN) {
 141         *dirfdp = dirfd;
 142         return 0;
 143     } else {
 144         close(dirfd);
 145         return ENAMETOOLONG;
 146     }
 147 }
 148
 149 /* Attempts to shorten 'name' by creating a symlink for the directory part of
 150  * the name and indirecting through <symlink>/<basename>.  This works on
 151  * systems that support symlinks, as long as <basename> isn't too long.
 152  *
 153  * On success, returns 0 and stores the short name in 'short_name' and the
 154  * symbolic link to eventually delete in 'linkname'. */
 155 static int
 156 shorten_name_via_symlink(const char *name, char short_name[MAX_UN_LEN + 1],
 157                          char linkname[MAX_UN_LEN + 1])
 158 {
 159     char *abs, *dir, *base;
 160     const char *tmpdir;
 161     int error;
 162     int i;
 163
 164     abs = abs_file_name(NULL, name);
 165     dir = dir_name(abs);
 166     base = base_name(abs);
 167     free(abs);
 168
 169     tmpdir = getenv("TMPDIR");
 170     if (tmpdir == NULL) {
 171         tmpdir = "/tmp";
 172     }
 173
 174     for (i = 0; i < 1000; i++) {
 175         int len;
 176
 177         len = snprintf(linkname, MAX_UN_LEN + 1,
 178                        "%s/ovs-un-c-%"PRIu32, tmpdir, random_uint32());
 179         error = (len < 0 || len > MAX_UN_LEN ? ENAMETOOLONG
 180                  : symlink(dir, linkname) ? errno
 181                  : 0);
 182         if (error != EEXIST) {
 183             break;
 184         }
 185     }
 186
 187     if (!error) {
 188         int len;
 189
 190         fatal_signal_add_file_to_unlink(linkname);
 191
 192         len = snprintf(short_name, MAX_UN_LEN + 1, "%s/%s", linkname, base);
 193         if (len < 0 || len > MAX_UN_LEN) {
 194             fatal_signal_unlink_file_now(linkname);
 195             error = ENAMETOOLONG;
 196         }
 197     }
 198
 199     if (error) {
 200         linkname[0] = '\0';
 201     }
 202     free(dir);
 203     free(base);
 204
 205     return error;
 206 }
 207
 208 /* Stores in '*un' a sockaddr_un that refers to file 'name'.  Stores in
 209  * '*un_len' the size of the sockaddr_un.
 210  *
 211  * Returns 0 on success, otherwise a positive errno value.
 212  *
 213  * Uses '*dirfdp' and 'linkname' to store references to data when the caller no
 214  * longer needs to use 'un'.  On success, freeing these references with
 215  * free_sockaddr_un() is mandatory to avoid a leak; on failure, freeing them is
 216  * unnecessary but harmless. */
 217 static int
 218 make_sockaddr_un(const char *name, struct sockaddr_un *un, socklen_t *un_len,
 219                  int *dirfdp, char linkname[MAX_UN_LEN + 1])
 220 {
 221     char short_name[MAX_UN_LEN + 1];
 222
 223     *dirfdp = -1;
 224     linkname[0] = '\0';
 225     if (strlen(name) > MAX_UN_LEN) {
 226         /* 'name' is too long to fit in a sockaddr_un.  Try a workaround. */
 227         int error = shorten_name_via_proc(name, short_name, dirfdp);
 228         if (error == ENAMETOOLONG) {
 229             error = shorten_name_via_symlink(name, short_name, linkname);
 230         }
 231         if (error) {
 232             static struct vlog_rate_limit rl = VLOG_RATE_LIMIT_INIT(1, 1);
 233
 234             VLOG_WARN_RL(&rl, "Unix socket name %s is longer than maximum "
 235                          "%"PRIuSIZE" bytes", name, MAX_UN_LEN);
 236             return error;
 237         }
 238
 239         name = short_name;
 240     }
 241
 242     un->sun_family = AF_UNIX;
 243     ovs_strzcpy(un->sun_path, name, sizeof un->sun_path);
 244     *un_len = (offsetof(struct sockaddr_un, sun_path)
 245                 + strlen (un->sun_path) + 1);
 246     return 0;
 247 }
 248
 249 /* Clean up after make_sockaddr_un(). */
 250 static void
 251 free_sockaddr_un(int dirfd, const char *linkname)
 252 {
 253     if (dirfd >= 0) {
 254         close(dirfd);
 255     }
 256     if (linkname[0]) {
 257         fatal_signal_unlink_file_now(linkname);
 258     }
 259 }
 260
 261 /* Binds Unix domain socket 'fd' to a file with permissions 0700. */
 262 static int bind_unix_socket(int fd, struct sockaddr *sun, socklen_t sun_len)
 263 {
 264     const mode_t mode = 0770;    /* Allow both user and group access. */
 265
 266     if (LINUX) {
 267         /* On Linux, the fd's permissions become the file's permissions.
 268          * fchmod() does not affect other files, like umask() does. */
 269         if (fchmod(fd, mode)) {
 270             return errno;
 271         }
 272
 273         /* Must be after fchmod(). */
 274         if (bind(fd, sun, sun_len)) {
 275             return errno;
 276         }
 277         return 0;
 278     } else {
 279         /* On FreeBSD and NetBSD, only the umask affects permissions.  The
 280          * umask is process-wide rather than thread-specific, so we have to use
 281          * a subprocess for safety. */
 282         pid_t pid = fork();
 283
 284         if (!pid) {
 285             umask(mode ^ 0777);
 286             _exit(bind(fd, sun, sun_len) ? errno : 0);
 287         } else if (pid > 0) {
 288             int status;
 289             int error;
 290
 291             do {
 292                 error = waitpid(pid, &status, 0) < 0 ? errno : 0;
 293             } while (error == EINTR);
 294
 295             return (error ? error
 296                     : WIFEXITED(status) ? WEXITSTATUS(status)
 297                     : WIFSIGNALED(status) ? EINTR
 298                     : ECHILD /* WTF? */);
 299         } else {
 300             return errno;
 301         }
 302     }
 303 }
 304
 305 /* Creates a Unix domain socket in the given 'style' (either SOCK_DGRAM or
 306  * SOCK_STREAM) that is bound to '*bind_path' (if 'bind_path' is non-null) and
 307  * connected to '*connect_path' (if 'connect_path' is non-null).  If 'nonblock'
 308  * is true, the socket is made non-blocking.
 309  *
 310  * Returns the socket's fd if successful, otherwise a negative errno value. */
 311 int
 312 make_unix_socket(int style, bool nonblock,
 313                  const char *bind_path, const char *connect_path)
 314 {
 315     int error;
 316     int fd;
 317
 318     fd = socket(PF_UNIX, style, 0);
 319     if (fd < 0) {
 320         return -errno;
 321     }
 322
 323     /* Set nonblocking mode right away, if we want it.  This prevents blocking
 324      * in connect(), if connect_path != NULL.  (In turn, that's a corner case:
 325      * it will only happen if style is SOCK_STREAM or SOCK_SEQPACKET, and only
 326      * if a backlog of un-accepted connections has built up in the kernel.)  */
 327     if (nonblock) {
 328         error = set_nonblocking(fd);
 329         if (error) {
 330             goto error;
 331         }
 332     }
 333
 334     if (bind_path) {
 335         char linkname[MAX_UN_LEN + 1];
 336         struct sockaddr_un un;
 337         socklen_t un_len;
 338         int dirfd;
 339
 340         if (unlink(bind_path) && errno != ENOENT) {
 341             VLOG_WARN("unlinking \"%s\": %s\n",
 342                       bind_path, ovs_strerror(errno));
 343         }
 344         fatal_signal_add_file_to_unlink(bind_path);
 345
 346         error = make_sockaddr_un(bind_path, &un, &un_len, &dirfd, linkname);
 347         if (!error) {
 348             error = bind_unix_socket(fd, (struct sockaddr *) &un, un_len);
 349         }
 350         free_sockaddr_un(dirfd, linkname);
 351
 352         if (error) {
 353             goto error;
 354         }
 355     }
 356
 357     if (connect_path) {
 358         char linkname[MAX_UN_LEN + 1];
 359         struct sockaddr_un un;
 360         socklen_t un_len;
 361         int dirfd;
 362
 363         error = make_sockaddr_un(connect_path, &un, &un_len, &dirfd, linkname);
 364         if (!error
 365             && connect(fd, (struct sockaddr*) &un, un_len)
 366             && errno != EINPROGRESS) {
 367             error = errno;
 368         }
 369         free_sockaddr_un(dirfd, linkname);
 370
 371         if (error) {
 372             goto error;
 373         }
 374     }
 375
 376     return fd;
 377
 378 error:
 379     if (error == EAGAIN) {
 380         error = EPROTO;
 381     }
 382     if (bind_path) {
 383         fatal_signal_unlink_file_now(bind_path);
 384     }
 385     close(fd);
 386     return -error;
 387 }
 388
 389 int
 390 get_unix_name_len(socklen_t sun_len)
 391 {
 392     return (sun_len >= offsetof(struct sockaddr_un, sun_path)
 393             ? sun_len - offsetof(struct sockaddr_un, sun_path)
 394             : 0);
 395 }
 396
 397 /* Calls ioctl() on an AF_INET sock, passing the specified 'command' and
 398  * 'arg'.  Returns 0 if successful, otherwise a positive errno value. */
 399 int
 400 af_inet_ioctl(unsigned long int command, const void *arg)
 401 {
 402     static struct ovsthread_once once = OVSTHREAD_ONCE_INITIALIZER;
 403     static int sock;
 404
 405     if (ovsthread_once_start(&once)) {
 406         sock = socket(AF_INET, SOCK_DGRAM, 0);
 407         if (sock < 0) {
 408             int error = sock_errno();
 409             VLOG_ERR("failed to create inet socket: %s", sock_strerror(error));
 410             sock = -error;
 411         }
 412         ovsthread_once_done(&once);
 413     }
 414
 415     return (sock < 0 ? -sock
 416             : ioctl(sock, command, arg) == -1 ? errno
 417             : 0);
 418 }
 419
 420 int
 421 af_inet_ifreq_ioctl(const char *name, struct ifreq *ifr, unsigned long int cmd,
 422                     const char *cmd_name)
 423 {
 424     int error;
 425
 426     ovs_strzcpy(ifr->ifr_name, name, sizeof ifr->ifr_name);
 427     error = af_inet_ioctl(cmd, ifr);
 428     if (error) {
 429         static struct vlog_rate_limit rl = VLOG_RATE_LIMIT_INIT(5, 20);
 430         VLOG_DBG_RL(&rl, "%s: ioctl(%s) failed: %s", name, cmd_name,
 431                     ovs_strerror(error));
 432     }
 433     return error;
 434 }