libmultipath/uevent.c

   1 /*
   2  * uevent.c - trigger upon netlink uevents from the kernel
   3  *
   4  *      Only kernels from version 2.6.10* on provide the uevent netlink socket.
   5  *      Until the libc-kernel-headers are updated, you need to compile with:
   6  *
   7  *        gcc -I /lib/modules/`uname -r`/build/include -o uevent_listen uevent_listen.c
   8  *
   9  * Copyright (C) 2004 Kay Sievers <kay.sievers@vrfy.org>
  10  *
  11  *      This program is free software; you can redistribute it and/or modify it
  12  *      under the terms of the GNU General Public License as published by the
  13  *      Free Software Foundation version 2 of the License.
  14  *
  15  *      This program is distributed in the hope that it will be useful, but
  16  *      WITHOUT ANY WARRANTY; without even the implied warranty of
  17  *      MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  18  *      General Public License for more details.
  19  *
  20  *      You should have received a copy of the GNU General Public License along
  21  *      with this program; if not, write to the Free Software Foundation, Inc.,
  22  *      675 Mass Ave, Cambridge, MA 02139, USA.
  23  *
  24  */
  25
  26 #include <unistd.h>
  27 #include <stdio.h>
  28 #include <errno.h>
  29 #include <stdlib.h>
  30 #include <stddef.h>
  31 #include <string.h>
  32 #include <fcntl.h>
  33 #include <time.h>
  34 #include <sys/socket.h>
  35 #include <sys/user.h>
  36 #include <sys/un.h>
  37 #include <linux/types.h>
  38 #include <linux/netlink.h>
  39 #include <pthread.h>
  40 #include <limits.h>
  41 #include <sys/mman.h>
  42 #include <libudev.h>
  43 #include <errno.h>
  44
  45 #include "memory.h"
  46 #include "debug.h"
  47 #include "list.h"
  48 #include "uevent.h"
  49 #include "vector.h"
  50 #include "config.h"
  51
  52 typedef int (uev_trigger)(struct uevent *, void * trigger_data);
  53
  54 pthread_t uevq_thr;
  55 LIST_HEAD(uevq);
  56 pthread_mutex_t uevq_lock, *uevq_lockp = &uevq_lock;
  57 pthread_cond_t  uev_cond,  *uev_condp  = &uev_cond;
  58 uev_trigger *my_uev_trigger;
  59 void * my_trigger_data;
  60 int servicing_uev;
  61
  62 int is_uevent_busy(void)
  63 {
  64         int empty;
  65
  66         pthread_mutex_lock(uevq_lockp);
  67         empty = list_empty(&uevq);
  68         pthread_mutex_unlock(uevq_lockp);
  69         return (!empty || servicing_uev);
  70 }
  71
  72 struct uevent * alloc_uevent (void)
  73 {
  74         struct uevent *uev = MALLOC(sizeof(struct uevent));
  75
  76         if (uev)
  77                 INIT_LIST_HEAD(&uev->node);
  78
  79         return uev;
  80 }
  81
  82 void
  83 setup_thread_attr(pthread_attr_t *attr, size_t stacksize, int detached)
  84 {
  85         if (pthread_attr_init(attr)) {
  86                 fprintf(stderr, "can't initialize thread attr: %s\n",
  87                         strerror(errno));
  88                 exit(1);
  89         }
  90         if (stacksize < PTHREAD_STACK_MIN)
  91                 stacksize = PTHREAD_STACK_MIN;
  92
  93         if (pthread_attr_setstacksize(attr, stacksize)) {
  94                 fprintf(stderr, "can't set thread stack size to %lu: %s\n",
  95                         (unsigned long)stacksize, strerror(errno));
  96                 exit(1);
  97         }
  98         if (detached && pthread_attr_setdetachstate(attr,
  99                                                     PTHREAD_CREATE_DETACHED)) {
 100                 fprintf(stderr, "can't set thread to detached: %s\n",
 101                         strerror(errno));
 102                 exit(1);
 103         }
 104 }
 105
 106 /*
 107  * Called with uevq_lockp held
 108  */
 109 void
 110 service_uevq(struct list_head *tmpq)
 111 {
 112         struct uevent *uev, *tmp;
 113
 114         list_for_each_entry_safe(uev, tmp, tmpq, node) {
 115                 list_del_init(&uev->node);
 116
 117                 if (my_uev_trigger && my_uev_trigger(uev, my_trigger_data))
 118                         condlog(0, "uevent trigger error");
 119
 120                 FREE(uev);
 121         }
 122 }
 123
 124 static void uevq_stop(void *arg)
 125 {
 126         condlog(3, "Stopping uev queue");
 127         pthread_mutex_lock(uevq_lockp);
 128         my_uev_trigger = NULL;
 129         pthread_cond_signal(uev_condp);
 130         pthread_mutex_unlock(uevq_lockp);
 131 }
 132
 133 /*
 134  * Service the uevent queue.
 135  */
 136 int uevent_dispatch(int (*uev_trigger)(struct uevent *, void * trigger_data),
 137                     void * trigger_data)
 138 {
 139         my_uev_trigger = uev_trigger;
 140         my_trigger_data = trigger_data;
 141
 142         mlockall(MCL_CURRENT | MCL_FUTURE);
 143
 144         while (1) {
 145                 LIST_HEAD(uevq_tmp);
 146
 147                 pthread_mutex_lock(uevq_lockp);
 148                 servicing_uev = 0;
 149                 /*
 150                  * Condition signals are unreliable,
 151                  * so make sure we only wait if we have to.
 152                  */
 153                 if (list_empty(&uevq)) {
 154                         pthread_cond_wait(uev_condp, uevq_lockp);
 155                 }
 156                 servicing_uev = 1;
 157                 list_splice_init(&uevq, &uevq_tmp);
 158                 pthread_mutex_unlock(uevq_lockp);
 159                 if (!my_uev_trigger)
 160                         break;
 161                 service_uevq(&uevq_tmp);
 162         }
 163         condlog(3, "Terminating uev service queue");
 164         return 0;
 165 }
 166
 167 int failback_listen(void)
 168 {
 169         int sock;
 170         struct sockaddr_nl snl;
 171         struct sockaddr_un sun;
 172         socklen_t addrlen;
 173         int retval;
 174         int rcvbufsz = 128*1024;
 175         int rcvsz = 0;
 176         int rcvszsz = sizeof(rcvsz);
 177         unsigned int *prcvszsz = (unsigned int *)&rcvszsz;
 178         const int feature_on = 1;
 179         /*
 180          * First check whether we have a udev socket
 181          */
 182         memset(&sun, 0x00, sizeof(struct sockaddr_un));
 183         sun.sun_family = AF_LOCAL;
 184         strcpy(&sun.sun_path[1], "/org/kernel/dm/multipath_event");
 185         addrlen = offsetof(struct sockaddr_un, sun_path) + strlen(sun.sun_path+1) + 1;
 186
 187         sock = socket(AF_LOCAL, SOCK_DGRAM, 0);
 188         if (sock >= 0) {
 189
 190                 condlog(3, "reading events from udev socket.");
 191
 192                 /* the bind takes care of ensuring only one copy running */
 193                 retval = bind(sock, (struct sockaddr *) &sun, addrlen);
 194                 if (retval < 0) {
 195                         condlog(0, "bind failed, exit");
 196                         goto exit;
 197                 }
 198
 199                 /* enable receiving of the sender credentials */
 200                 setsockopt(sock, SOL_SOCKET, SO_PASSCRED,
 201                            &feature_on, sizeof(feature_on));
 202
 203         } else {
 204                 /* Fallback to read kernel netlink events */
 205                 memset(&snl, 0x00, sizeof(struct sockaddr_nl));
 206                 snl.nl_family = AF_NETLINK;
 207                 snl.nl_pid = getpid();
 208                 snl.nl_groups = 0x01;
 209
 210                 sock = socket(PF_NETLINK, SOCK_DGRAM, NETLINK_KOBJECT_UEVENT);
 211                 if (sock == -1) {
 212                         condlog(0, "error getting socket, exit");
 213                         return 1;
 214                 }
 215
 216                 condlog(3, "reading events from kernel.");
 217
 218                 /*
 219                  * try to avoid dropping uevents, even so, this is not a guarantee,
 220                  * but it does help to change the netlink uevent socket's
 221                  * receive buffer threshold from the default value of 106,496 to
 222                  * the maximum value of 262,142.
 223                  */
 224                 retval = setsockopt(sock, SOL_SOCKET, SO_RCVBUF, &rcvbufsz,
 225                                     sizeof(rcvbufsz));
 226
 227                 if (retval < 0) {
 228                         condlog(0, "error setting receive buffer size for socket, exit");
 229                         exit(1);
 230                 }
 231                 retval = getsockopt(sock, SOL_SOCKET, SO_RCVBUF, &rcvsz, prcvszsz);
 232                 if (retval < 0) {
 233                         condlog(0, "error setting receive buffer size for socket, exit");
 234                         exit(1);
 235                 }
 236                 condlog(3, "receive buffer size for socket is %u.", rcvsz);
 237
 238                 /* enable receiving of the sender credentials */
 239                 setsockopt(sock, SOL_SOCKET, SO_PASSCRED,
 240                            &feature_on, sizeof(feature_on));
 241
 242                 retval = bind(sock, (struct sockaddr *) &snl,
 243                               sizeof(struct sockaddr_nl));
 244                 if (retval < 0) {
 245                         condlog(0, "bind failed, exit");
 246                         goto exit;
 247                 }
 248         }
 249
 250         while (1) {
 251                 int i;
 252                 char *pos;
 253                 size_t bufpos;
 254                 ssize_t buflen;
 255                 struct uevent *uev;
 256                 char *buffer;
 257                 struct msghdr smsg;
 258                 struct iovec iov;
 259                 char cred_msg[CMSG_SPACE(sizeof(struct ucred))];
 260                 struct cmsghdr *cmsg;
 261                 struct ucred *cred;
 262                 static char buf[HOTPLUG_BUFFER_SIZE + OBJECT_SIZE];
 263
 264                 memset(buf, 0x00, sizeof(buf));
 265                 iov.iov_base = &buf;
 266                 iov.iov_len = sizeof(buf);
 267                 memset (&smsg, 0x00, sizeof(struct msghdr));
 268                 smsg.msg_iov = &iov;
 269                 smsg.msg_iovlen = 1;
 270                 smsg.msg_control = cred_msg;
 271                 smsg.msg_controllen = sizeof(cred_msg);
 272
 273                 buflen = recvmsg(sock, &smsg, 0);
 274                 if (buflen < 0) {
 275                         if (errno != EINTR)
 276                                 condlog(0, "error receiving message, errno %d", errno);
 277                         continue;
 278                 }
 279
 280                 cmsg = CMSG_FIRSTHDR(&smsg);
 281                 if (cmsg == NULL || cmsg->cmsg_type != SCM_CREDENTIALS) {
 282                         condlog(3, "no sender credentials received, message ignored");
 283                         continue;
 284                 }
 285
 286                 cred = (struct ucred *)CMSG_DATA(cmsg);
 287                 if (cred->uid != 0) {
 288                         condlog(3, "sender uid=%d, message ignored", cred->uid);
 289                         continue;
 290                 }
 291
 292                 /* skip header */
 293                 bufpos = strlen(buf) + 1;
 294                 if (bufpos < sizeof("a@/d") || bufpos >= sizeof(buf)) {
 295                         condlog(3, "invalid message length");
 296                         continue;
 297                 }
 298
 299                 /* check message header */
 300                 if (strstr(buf, "@/") == NULL) {
 301                         condlog(3, "unrecognized message header");
 302                         continue;
 303                 }
 304                 if ((size_t)buflen > sizeof(buf)-1) {
 305                         condlog(2, "buffer overflow for received uevent");
 306                         buflen = sizeof(buf)-1;
 307                 }
 308
 309                 uev = alloc_uevent();
 310
 311                 if (!uev) {
 312                         condlog(1, "lost uevent, oom");
 313                         continue;
 314                 }
 315
 316                 if ((size_t)buflen > sizeof(buf)-1)
 317                         buflen = sizeof(buf)-1;
 318
 319                 /*
 320                  * Copy the shared receive buffer contents to buffer private
 321                  * to this uevent so we can immediately reuse the shared buffer.
 322                  */
 323                 memcpy(uev->buffer, buf, HOTPLUG_BUFFER_SIZE + OBJECT_SIZE);
 324                 buffer = uev->buffer;
 325                 buffer[buflen] = '\0';
 326
 327                 /* save start of payload */
 328                 bufpos = strlen(buffer) + 1;
 329
 330                 /* action string */
 331                 uev->action = buffer;
 332                 pos = strchr(buffer, '@');
 333                 if (!pos) {
 334                         condlog(3, "bad action string '%s'", buffer);
 335                         continue;
 336                 }
 337                 pos[0] = '\0';
 338
 339                 /* sysfs path */
 340                 uev->devpath = &pos[1];
 341
 342                 /* hotplug events have the environment attached - reconstruct envp[] */
 343                 for (i = 0; (bufpos < (size_t)buflen) && (i < HOTPLUG_NUM_ENVP-1); i++) {
 344                         int keylen;
 345                         char *key;
 346
 347                         key = &buffer[bufpos];
 348                         keylen = strlen(key);
 349                         uev->envp[i] = key;
 350                         bufpos += keylen + 1;
 351                 }
 352                 uev->envp[i] = NULL;
 353
 354                 condlog(3, "uevent '%s' from '%s'", uev->action, uev->devpath);
 355                 uev->kernel = strrchr(uev->devpath, '/');
 356                 if (uev->kernel)
 357                         uev->kernel++;
 358
 359                 /* print payload environment */
 360                 for (i = 0; uev->envp[i] != NULL; i++)
 361                         condlog(5, "%s", uev->envp[i]);
 362
 363                 /*
 364                  * Queue uevent and poke service pthread.
 365                  */
 366                 pthread_mutex_lock(uevq_lockp);
 367                 list_add_tail(&uev->node, &uevq);
 368                 pthread_cond_signal(uev_condp);
 369                 pthread_mutex_unlock(uevq_lockp);
 370         }
 371
 372 exit:
 373         close(sock);
 374         return 1;
 375 }
 376
 377 int uevent_listen(void)
 378 {
 379         int err;
 380         struct udev_monitor *monitor = NULL;
 381         int fd, socket_flags;
 382         int need_failback = 1;
 383         /*
 384          * Queue uevents for service by dedicated thread so that the uevent
 385          * listening thread does not block on multipathd locks (vecs->lock)
 386          * thereby not getting to empty the socket's receive buffer queue
 387          * often enough.
 388          */
 389         INIT_LIST_HEAD(&uevq);
 390
 391         pthread_mutex_init(uevq_lockp, NULL);
 392         pthread_cond_init(uev_condp, NULL);
 393         pthread_cleanup_push(uevq_stop, NULL);
 394
 395         monitor = udev_monitor_new_from_netlink(conf->udev, "udev");
 396         if (!monitor) {
 397                 condlog(2, "failed to create udev monitor");
 398                 goto out;
 399         }
 400         if (udev_monitor_set_receive_buffer_size(monitor, 128 * 1024 * 1024))
 401                 condlog(2, "failed to increase buffer size");
 402         fd = udev_monitor_get_fd(monitor);
 403         if (fd < 0) {
 404                 condlog(2, "failed to get monitor fd");
 405                 goto out;
 406         }
 407         socket_flags = fcntl(fd, F_GETFL);
 408         if (socket_flags < 0) {
 409                 condlog(2, "failed to get monitor socket flags : %s",
 410                         strerror(errno));
 411                 goto out;
 412         }
 413         if (fcntl(fd, F_SETFL, socket_flags & ~O_NONBLOCK) < 0) {
 414                 condlog(2, "failed to set monitor socket flags : %s",
 415                         strerror(errno));
 416                 goto out;
 417         }
 418         err = udev_monitor_filter_add_match_subsystem_devtype(monitor, "block",
 419                                                               NULL);
 420         if (err)
 421                 condlog(2, "failed to create filter : %s\n", strerror(-err));
 422         err = udev_monitor_enable_receiving(monitor);
 423         if (err) {
 424                 condlog(2, "failed to enable receiving : %s\n", strerror(-err));
 425                 goto out;
 426         }
 427         while (1) {
 428                 int i = 0;
 429                 char *pos, *end;
 430                 struct uevent *uev;
 431                 struct udev_device *dev;
 432                 struct udev_list_entry *list_entry;
 433
 434                 dev = udev_monitor_receive_device(monitor);
 435                 if (!dev) {
 436                         condlog(0, "failed getting udev device");
 437                         continue;
 438                 }
 439
 440                 uev = alloc_uevent();
 441                 if (!uev) {
 442                         condlog(1, "lost uevent, oom");
 443                         continue;
 444                 }
 445                 pos = uev->buffer;
 446                 end = pos + HOTPLUG_BUFFER_SIZE + OBJECT_SIZE - 1;
 447                 udev_list_entry_foreach(list_entry, udev_device_get_properties_list_entry(dev)) {
 448                         const char *name, *value;
 449                         int bytes;
 450
 451                         name = udev_list_entry_get_name(list_entry);
 452                         if (!name)
 453                                 name = "(null)";
 454                         value = udev_list_entry_get_value(list_entry);
 455                         if (!value)
 456                                 value = "(null)";
 457                         bytes = snprintf(pos, end - pos, "%s=%s", name,
 458                                         value);
 459                         if (pos + bytes >= end) {
 460                                 condlog(2, "buffer overflow for uevent");
 461                                 break;
 462                         }
 463                         uev->envp[i] = pos;
 464                         pos += bytes;
 465                         *pos = '\0';
 466                         pos++;
 467                         if (strcmp(name, "DEVPATH") == 0)
 468                                 uev->devpath = uev->envp[i] + 8;
 469                         if (strcmp(name, "ACTION") == 0)
 470                                 uev->action = uev->envp[i] + 7;
 471                         i++;
 472                         if (i == HOTPLUG_NUM_ENVP - 1)
 473                                 break;
 474                 }
 475                 udev_device_unref(dev);
 476                 uev->envp[i] = NULL;
 477
 478                 condlog(3, "uevent '%s' from '%s'", uev->action, uev->devpath);
 479                 uev->kernel = strrchr(uev->devpath, '/');
 480                 if (uev->kernel)
 481                         uev->kernel++;
 482
 483                 /* print payload environment */
 484                 for (i = 0; uev->envp[i] != NULL; i++)
 485                         condlog(5, "%s", uev->envp[i]);
 486
 487                 /*
 488                  * Queue uevent and poke service pthread.
 489                  */
 490                 pthread_mutex_lock(uevq_lockp);
 491                 list_add_tail(&uev->node, &uevq);
 492                 pthread_cond_signal(uev_condp);
 493                 pthread_mutex_unlock(uevq_lockp);
 494         }
 495         need_failback = 0;
 496 out:
 497         if (monitor)
 498                 udev_monitor_unref(monitor);
 499         if (need_failback)
 500                 err = failback_listen();
 501         pthread_cleanup_pop(1);
 502         pthread_mutex_destroy(uevq_lockp);
 503         pthread_cond_destroy(uev_condp);
 504         return err;
 505 }
 506
 507 extern int
 508 uevent_get_major(struct uevent *uev)
 509 {
 510         char *p, *q;
 511         int i, major = -1;
 512
 513         for (i = 0; uev->envp[i] != NULL; i++) {
 514                 if (!strncmp(uev->envp[i], "MAJOR", 5) && strlen(uev->envp[i]) > 6) {
 515                         p = uev->envp[i] + 6;
 516                         major = strtoul(p, &q, 10);
 517                         if (p == q) {
 518                                 condlog(2, "invalid major '%s'", p);
 519                                 major = -1;
 520                         }
 521                         break;
 522                 }
 523         }
 524         return major;
 525 }
 526
 527 extern int
 528 uevent_get_minor(struct uevent *uev)
 529 {
 530         char *p, *q;
 531         int i, minor = -1;
 532
 533         for (i = 0; uev->envp[i] != NULL; i++) {
 534                 if (!strncmp(uev->envp[i], "MINOR", 5) && strlen(uev->envp[i]) > 6) {
 535                         p = uev->envp[i] + 6;
 536                         minor = strtoul(p, &q, 10);
 537                         if (p == q) {
 538                                 condlog(2, "invalid minor '%s'", p);
 539                                 minor = -1;
 540                         }
 541                         break;
 542                 }
 543         }
 544         return minor;
 545 }
 546
 547 extern int
 548 uevent_get_disk_ro(struct uevent *uev)
 549 {
 550         char *p, *q;
 551         int i, ro = -1;
 552
 553         for (i = 0; uev->envp[i] != NULL; i++) {
 554                 if (!strncmp(uev->envp[i], "DISK_RO", 6) && strlen(uev->envp[i]) > 7) {
 555                         p = uev->envp[i] + 8;
 556                         ro = strtoul(p, &q, 10);
 557                         if (p == q) {
 558                                 condlog(2, "invalid read_only setting '%s'", p);
 559                                 ro = -1;
 560                         }
 561                         break;
 562                 }
 563         }
 564         return ro;
 565 }
 566
 567 extern char *
 568 uevent_get_dm_name(struct uevent *uev)
 569 {
 570         char *p = NULL;
 571         int i;
 572
 573         for (i = 0; uev->envp[i] != NULL; i++) {
 574                 if (!strncmp(uev->envp[i], "DM_NAME", 6) &&
 575                     strlen(uev->envp[i]) > 7) {
 576                         p = MALLOC(strlen(uev->envp[i] + 8) + 1);
 577                         strcpy(p, uev->envp[i] + 8);
 578                         break;
 579                 }
 580         }
 581         return p;
 582 }