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, see <http://www.gnu.org/licenses/>.
  22  *
  23  */
  24
  25 #include <unistd.h>
  26 #include <stdio.h>
  27 #include <stdbool.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 <poll.h>
  38 #include <linux/types.h>
  39 #include <linux/netlink.h>
  40 #include <pthread.h>
  41 #include <sys/mman.h>
  42 #include <sys/time.h>
  43 #include <libudev.h>
  44 #include <errno.h>
  45
  46 #include "debug.h"
  47 #include "list.h"
  48 #include "uevent.h"
  49 #include "vector.h"
  50 #include "structs.h"
  51 #include "util.h"
  52 #include "config.h"
  53 #include "blacklist.h"
  54 #include "devmapper.h"
  55
  56 #define MAX_ACCUMULATION_COUNT 2048
  57 #define MAX_ACCUMULATION_TIME 30*1000
  58 #define MIN_BURST_SPEED 10
  59
  60 typedef int (uev_trigger)(struct uevent *, void * trigger_data);
  61
  62 static LIST_HEAD(uevq);
  63 static pthread_mutex_t uevq_lock = PTHREAD_MUTEX_INITIALIZER;
  64 static pthread_mutex_t *uevq_lockp = &uevq_lock;
  65 static pthread_cond_t uev_cond = PTHREAD_COND_INITIALIZER;
  66 static pthread_cond_t *uev_condp = &uev_cond;
  67 static uev_trigger *my_uev_trigger;
  68 static void *my_trigger_data;
  69 static int servicing_uev;
  70
  71 int is_uevent_busy(void)
  72 {
  73         int empty;
  74
  75         pthread_mutex_lock(uevq_lockp);
  76         empty = list_empty(&uevq);
  77         pthread_mutex_unlock(uevq_lockp);
  78         return (!empty || servicing_uev);
  79 }
  80
  81 struct uevent * alloc_uevent (void)
  82 {
  83         struct uevent *uev = calloc(1, sizeof(struct uevent));
  84
  85         if (uev) {
  86                 INIT_LIST_HEAD(&uev->node);
  87                 INIT_LIST_HEAD(&uev->merge_node);
  88         }
  89
  90         return uev;
  91 }
  92
  93 static void uevq_cleanup(struct list_head *tmpq);
  94
  95 static void cleanup_uev(void *arg)
  96 {
  97         struct uevent *uev = arg;
  98
  99         uevq_cleanup(&uev->merge_node);
 100         if (uev->udev)
 101                 udev_device_unref(uev->udev);
 102         free(uev);
 103 }
 104
 105 static void uevq_cleanup(struct list_head *tmpq)
 106 {
 107         struct uevent *uev, *tmp;
 108
 109         list_for_each_entry_safe(uev, tmp, tmpq, node) {
 110                 list_del_init(&uev->node);
 111                 cleanup_uev(uev);
 112         }
 113 }
 114
 115 static const char* uevent_get_env_var(const struct uevent *uev,
 116                                       const char *attr)
 117 {
 118         int i;
 119         size_t len;
 120         const char *p = NULL;
 121
 122         if (attr == NULL)
 123                 goto invalid;
 124
 125         len = strlen(attr);
 126         if (len == 0)
 127                 goto invalid;
 128
 129         for (i = 0; uev->envp[i] != NULL; i++) {
 130                 const char *var = uev->envp[i];
 131
 132                 if (strlen(var) > len &&
 133                     !memcmp(var, attr, len) && var[len] == '=') {
 134                         p = var + len + 1;
 135                         break;
 136                 }
 137         }
 138
 139         condlog(4, "%s: %s -> '%s'", __func__, attr, p ?: "(null)");
 140         return p;
 141
 142 invalid:
 143         condlog(2, "%s: empty variable name", __func__);
 144         return NULL;
 145 }
 146
 147 int uevent_get_env_positive_int(const struct uevent *uev,
 148                                        const char *attr)
 149 {
 150         const char *p = uevent_get_env_var(uev, attr);
 151         char *q;
 152         int ret;
 153
 154         if (p == NULL || *p == '\0')
 155                 return -1;
 156
 157         ret = strtoul(p, &q, 10);
 158         if (*q != '\0' || ret < 0) {
 159                 condlog(2, "%s: invalid %s: '%s'", __func__, attr, p);
 160                 return -1;
 161         }
 162         return ret;
 163 }
 164
 165 void
 166 uevent_get_wwid(struct uevent *uev)
 167 {
 168         char *uid_attribute;
 169         const char *val;
 170         struct config * conf;
 171
 172         conf = get_multipath_config();
 173         pthread_cleanup_push(put_multipath_config, conf);
 174         uid_attribute = get_uid_attribute_by_attrs(conf, uev->kernel);
 175         pthread_cleanup_pop(1);
 176
 177         val = uevent_get_env_var(uev, uid_attribute);
 178         if (val)
 179                 uev->wwid = val;
 180 }
 181
 182 static bool uevent_need_merge(void)
 183 {
 184         struct config * conf;
 185         bool need_merge = false;
 186
 187         conf = get_multipath_config();
 188         if (VECTOR_SIZE(&conf->uid_attrs) > 0)
 189                 need_merge = true;
 190         put_multipath_config(conf);
 191
 192         return need_merge;
 193 }
 194
 195 static bool uevent_can_discard(struct uevent *uev)
 196 {
 197         int invalid = 0;
 198         struct config * conf;
 199
 200         /*
 201          * do not filter dm devices by devnode
 202          */
 203         if (!strncmp(uev->kernel, "dm-", 3))
 204                 return false;
 205         /*
 206          * filter paths devices by devnode
 207          */
 208         conf = get_multipath_config();
 209         pthread_cleanup_push(put_multipath_config, conf);
 210         if (filter_devnode(conf->blist_devnode, conf->elist_devnode,
 211                            uev->kernel) > 0)
 212                 invalid = 1;
 213         pthread_cleanup_pop(1);
 214
 215         if (invalid)
 216                 return true;
 217         return false;
 218 }
 219
 220 static bool
 221 uevent_can_filter(struct uevent *earlier, struct uevent *later)
 222 {
 223
 224         /*
 225          * filter earlier uvents if path has removed later. Eg:
 226          * "add path1 |chang path1 |add path2 |remove path1"
 227          * can filter as:
 228          * "add path2 |remove path1"
 229          * uevents "add path1" and "chang path1" are filtered out
 230          */
 231         if (!strcmp(earlier->kernel, later->kernel) &&
 232                 !strcmp(later->action, "remove") &&
 233                 strncmp(later->kernel, "dm-", 3)) {
 234                 return true;
 235         }
 236
 237         /*
 238          * filter change uvents if add uevents exist. Eg:
 239          * "change path1| add path1 |add path2"
 240          * can filter as:
 241          * "add path1 |add path2"
 242          * uevent "chang path1" is filtered out
 243          */
 244         if (!strcmp(earlier->kernel, later->kernel) &&
 245                 !strcmp(earlier->action, "change") &&
 246                 !strcmp(later->action, "add") &&
 247                 strncmp(later->kernel, "dm-", 3)) {
 248                 return true;
 249         }
 250
 251         return false;
 252 }
 253
 254 static bool
 255 merge_need_stop(struct uevent *earlier, struct uevent *later)
 256 {
 257         /*
 258          * dm uevent do not try to merge with left uevents
 259          */
 260         if (!strncmp(later->kernel, "dm-", 3))
 261                 return true;
 262
 263         /*
 264          * we can not make a jugement without wwid,
 265          * so it is sensible to stop merging
 266          */
 267         if (!earlier->wwid || !later->wwid)
 268                 return true;
 269         /*
 270          * uevents merging stopped
 271          * when we meet an opposite action uevent from the same LUN to AVOID
 272          * "add path1 |remove path1 |add path2 |remove path2 |add path3"
 273          * to merge as "remove path1, path2" and "add path1, path2, path3"
 274          * OR
 275          * "remove path1 |add path1 |remove path2 |add path2 |remove path3"
 276          * to merge as "add path1, path2" and "remove path1, path2, path3"
 277          * SO
 278          * when we meet a non-change uevent from the same LUN
 279          * with the same wwid and different action
 280          * it would be better to stop merging.
 281          */
 282         if (!strcmp(earlier->wwid, later->wwid) &&
 283             strcmp(earlier->action, later->action) &&
 284             strcmp(earlier->action, "change") &&
 285             strcmp(later->action, "change"))
 286                 return true;
 287
 288         return false;
 289 }
 290
 291 static bool
 292 uevent_can_merge(struct uevent *earlier, struct uevent *later)
 293 {
 294         /* merge paths uevents
 295          * whose wwids exsit and are same
 296          * and actions are same,
 297          * and actions are addition or deletion
 298          */
 299         if (earlier->wwid && later->wwid &&
 300             !strcmp(earlier->wwid, later->wwid) &&
 301             !strcmp(earlier->action, later->action) &&
 302             strncmp(earlier->action, "change", 6) &&
 303             strncmp(earlier->kernel, "dm-", 3)) {
 304                 return true;
 305         }
 306
 307         return false;
 308 }
 309
 310 static void
 311 uevent_prepare(struct list_head *tmpq)
 312 {
 313         struct uevent *uev, *tmp;
 314
 315         list_for_each_entry_reverse_safe(uev, tmp, tmpq, node) {
 316                 if (uevent_can_discard(uev)) {
 317                         list_del_init(&uev->node);
 318                         if (uev->udev)
 319                                 udev_device_unref(uev->udev);
 320                         free(uev);
 321                         continue;
 322                 }
 323
 324                 if (strncmp(uev->kernel, "dm-", 3) &&
 325                     uevent_need_merge())
 326                         uevent_get_wwid(uev);
 327         }
 328 }
 329
 330 static void
 331 uevent_filter(struct uevent *later, struct list_head *tmpq)
 332 {
 333         struct uevent *earlier, *tmp;
 334
 335         list_for_some_entry_reverse_safe(earlier, tmp, &later->node, tmpq, node) {
 336                 /*
 337                  * filter unnessary earlier uevents
 338                  * by the later uevent
 339                  */
 340                 if (uevent_can_filter(earlier, later)) {
 341                         condlog(3, "uevent: %s-%s has filtered by uevent: %s-%s",
 342                                 earlier->kernel, earlier->action,
 343                                 later->kernel, later->action);
 344
 345                         list_del_init(&earlier->node);
 346                         if (earlier->udev)
 347                                 udev_device_unref(earlier->udev);
 348                         free(earlier);
 349                 }
 350         }
 351 }
 352
 353 static void
 354 uevent_merge(struct uevent *later, struct list_head *tmpq)
 355 {
 356         struct uevent *earlier, *tmp;
 357
 358         list_for_some_entry_reverse_safe(earlier, tmp, &later->node, tmpq, node) {
 359                 if (merge_need_stop(earlier, later))
 360                         break;
 361                 /*
 362                  * merge earlier uevents to the later uevent
 363                  */
 364                 if (uevent_can_merge(earlier, later)) {
 365                         condlog(3, "merged uevent: %s-%s-%s with uevent: %s-%s-%s",
 366                                 earlier->action, earlier->kernel, earlier->wwid,
 367                                 later->action, later->kernel, later->wwid);
 368
 369                         list_move(&earlier->node, &later->merge_node);
 370                 }
 371         }
 372 }
 373
 374 static void
 375 merge_uevq(struct list_head *tmpq)
 376 {
 377         struct uevent *later;
 378
 379         uevent_prepare(tmpq);
 380         list_for_each_entry_reverse(later, tmpq, node) {
 381                 uevent_filter(later, tmpq);
 382                 if(uevent_need_merge())
 383                         uevent_merge(later, tmpq);
 384         }
 385 }
 386
 387 static void
 388 service_uevq(struct list_head *tmpq)
 389 {
 390         struct uevent *uev, *tmp;
 391
 392         list_for_each_entry_safe(uev, tmp, tmpq, node) {
 393                 list_del_init(&uev->node);
 394
 395                 pthread_cleanup_push(cleanup_uev, uev);
 396                 if (my_uev_trigger && my_uev_trigger(uev, my_trigger_data))
 397                         condlog(0, "uevent trigger error");
 398                 pthread_cleanup_pop(1);
 399         }
 400 }
 401
 402 static void uevent_cleanup(void *arg)
 403 {
 404         struct udev *udev = arg;
 405
 406         condlog(3, "Releasing uevent_listen() resources");
 407         udev_unref(udev);
 408 }
 409
 410 static void monitor_cleanup(void *arg)
 411 {
 412         struct udev_monitor *monitor = arg;
 413
 414         condlog(3, "Releasing uevent_monitor() resources");
 415         udev_monitor_unref(monitor);
 416 }
 417
 418 static void cleanup_uevq(void *arg)
 419 {
 420         uevq_cleanup(arg);
 421 }
 422
 423 static void cleanup_global_uevq(void *arg __attribute__((unused)))
 424 {
 425         pthread_mutex_lock(uevq_lockp);
 426         uevq_cleanup(&uevq);
 427         pthread_mutex_unlock(uevq_lockp);
 428 }
 429
 430 /*
 431  * Service the uevent queue.
 432  */
 433 int uevent_dispatch(int (*uev_trigger)(struct uevent *, void * trigger_data),
 434                     void * trigger_data)
 435 {
 436         my_uev_trigger = uev_trigger;
 437         my_trigger_data = trigger_data;
 438
 439         mlockall(MCL_CURRENT | MCL_FUTURE);
 440
 441         while (1) {
 442                 LIST_HEAD(uevq_tmp);
 443
 444                 pthread_cleanup_push(cleanup_mutex, uevq_lockp);
 445                 pthread_mutex_lock(uevq_lockp);
 446                 servicing_uev = 0;
 447                 /*
 448                  * Condition signals are unreliable,
 449                  * so make sure we only wait if we have to.
 450                  */
 451                 if (list_empty(&uevq)) {
 452                         pthread_cond_wait(uev_condp, uevq_lockp);
 453                 }
 454                 servicing_uev = 1;
 455                 list_splice_init(&uevq, &uevq_tmp);
 456                 pthread_cleanup_pop(1);
 457
 458                 if (!my_uev_trigger)
 459                         break;
 460
 461                 pthread_cleanup_push(cleanup_uevq, &uevq_tmp);
 462                 merge_uevq(&uevq_tmp);
 463                 service_uevq(&uevq_tmp);
 464                 pthread_cleanup_pop(1);
 465         }
 466         condlog(3, "Terminating uev service queue");
 467         return 0;
 468 }
 469
 470 static struct uevent *uevent_from_udev_device(struct udev_device *dev)
 471 {
 472         struct uevent *uev;
 473         int i = 0;
 474         char *pos, *end;
 475         struct udev_list_entry *list_entry;
 476
 477         uev = alloc_uevent();
 478         if (!uev) {
 479                 udev_device_unref(dev);
 480                 condlog(1, "lost uevent, oom");
 481                 return NULL;
 482         }
 483         pos = uev->buffer;
 484         end = pos + HOTPLUG_BUFFER_SIZE + OBJECT_SIZE - 1;
 485         udev_list_entry_foreach(list_entry, udev_device_get_properties_list_entry(dev)) {
 486                 const char *name, *value;
 487                 int bytes;
 488
 489                 name = udev_list_entry_get_name(list_entry);
 490                 if (!name)
 491                         name = "(null)";
 492                 value = udev_list_entry_get_value(list_entry);
 493                 if (!value)
 494                         value = "(null)";
 495                 bytes = snprintf(pos, end - pos, "%s=%s", name, value);
 496                 if (pos + bytes >= end) {
 497                         condlog(2, "buffer overflow for uevent");
 498                         break;
 499                 }
 500                 uev->envp[i] = pos;
 501                 pos += bytes;
 502                 *pos = '\0';
 503                 pos++;
 504                 if (strcmp(name, "DEVPATH") == 0)
 505                         uev->devpath = uev->envp[i] + 8;
 506                 if (strcmp(name, "ACTION") == 0)
 507                         uev->action = uev->envp[i] + 7;
 508                 i++;
 509                 if (i == HOTPLUG_NUM_ENVP - 1)
 510                         break;
 511         }
 512         if (!uev->devpath || ! uev->action) {
 513                 udev_device_unref(dev);
 514                 condlog(1, "uevent missing necessary fields");
 515                 free(uev);
 516                 return NULL;
 517         }
 518         uev->udev = dev;
 519         uev->envp[i] = NULL;
 520
 521         condlog(3, "uevent '%s' from '%s'", uev->action, uev->devpath);
 522         uev->kernel = strrchr(uev->devpath, '/');
 523         if (uev->kernel)
 524                 uev->kernel++;
 525
 526         /* print payload environment */
 527         for (i = 0; uev->envp[i] != NULL; i++)
 528                 condlog(5, "%s", uev->envp[i]);
 529         return uev;
 530 }
 531
 532 static bool uevent_burst(struct timeval *start_time, int events)
 533 {
 534         struct timeval diff_time, end_time;
 535         unsigned long speed;
 536         unsigned long eclipse_ms;
 537
 538         if(events > MAX_ACCUMULATION_COUNT) {
 539                 condlog(2, "burst got %u uevents, too much uevents, stopped", events);
 540                 return false;
 541         }
 542
 543         gettimeofday(&end_time, NULL);
 544         timersub(&end_time, start_time, &diff_time);
 545
 546         eclipse_ms = diff_time.tv_sec * 1000 + diff_time.tv_usec / 1000;
 547
 548         if (eclipse_ms == 0)
 549                 return true;
 550
 551         if (eclipse_ms > MAX_ACCUMULATION_TIME) {
 552                 condlog(2, "burst continued %lu ms, too long time, stopped", eclipse_ms);
 553                 return false;
 554         }
 555
 556         speed = (events * 1000) / eclipse_ms;
 557         if (speed > MIN_BURST_SPEED)
 558                 return true;
 559
 560         return false;
 561 }
 562
 563 int uevent_listen(struct udev *udev)
 564 {
 565         int err = 2;
 566         struct udev_monitor *monitor = NULL;
 567         int fd, socket_flags, events;
 568         struct timeval start_time;
 569         int timeout = 30;
 570         LIST_HEAD(uevlisten_tmp);
 571
 572         /*
 573          * Queue uevents for service by dedicated thread so that the uevent
 574          * listening thread does not block on multipathd locks (vecs->lock)
 575          * thereby not getting to empty the socket's receive buffer queue
 576          * often enough.
 577          */
 578         if (!udev) {
 579                 condlog(1, "no udev context");
 580                 return 1;
 581         }
 582         udev_ref(udev);
 583         pthread_cleanup_push(uevent_cleanup, udev);
 584
 585         monitor = udev_monitor_new_from_netlink(udev, "udev");
 586         if (!monitor) {
 587                 condlog(2, "failed to create udev monitor");
 588                 goto out_udev;
 589         }
 590         pthread_cleanup_push(monitor_cleanup, monitor);
 591 #ifdef LIBUDEV_API_RECVBUF
 592         if (udev_monitor_set_receive_buffer_size(monitor, 128 * 1024 * 1024) < 0)
 593                 condlog(2, "failed to increase buffer size");
 594 #endif
 595         fd = udev_monitor_get_fd(monitor);
 596         if (fd < 0) {
 597                 condlog(2, "failed to get monitor fd");
 598                 goto out;
 599         }
 600         socket_flags = fcntl(fd, F_GETFL);
 601         if (socket_flags < 0) {
 602                 condlog(2, "failed to get monitor socket flags : %s",
 603                         strerror(errno));
 604                 goto out;
 605         }
 606         if (fcntl(fd, F_SETFL, socket_flags & ~O_NONBLOCK) < 0) {
 607                 condlog(2, "failed to set monitor socket flags : %s",
 608                         strerror(errno));
 609                 goto out;
 610         }
 611         err = udev_monitor_filter_add_match_subsystem_devtype(monitor, "block",
 612                                                               "disk");
 613         if (err)
 614                 condlog(2, "failed to create filter : %s", strerror(-err));
 615         err = udev_monitor_enable_receiving(monitor);
 616         if (err) {
 617                 condlog(2, "failed to enable receiving : %s", strerror(-err));
 618                 goto out;
 619         }
 620
 621         events = 0;
 622         gettimeofday(&start_time, NULL);
 623         pthread_cleanup_push(cleanup_global_uevq, NULL);
 624         pthread_cleanup_push(cleanup_uevq, &uevlisten_tmp);
 625         while (1) {
 626                 struct uevent *uev;
 627                 struct udev_device *dev;
 628                 struct pollfd ev_poll;
 629                 int poll_timeout;
 630                 int fdcount;
 631
 632                 memset(&ev_poll, 0, sizeof(struct pollfd));
 633                 ev_poll.fd = fd;
 634                 ev_poll.events = POLLIN;
 635                 poll_timeout = timeout * 1000;
 636                 errno = 0;
 637                 fdcount = poll(&ev_poll, 1, poll_timeout);
 638                 if (fdcount > 0 && ev_poll.revents & POLLIN) {
 639                         timeout = uevent_burst(&start_time, events + 1) ? 1 : 0;
 640                         dev = udev_monitor_receive_device(monitor);
 641                         if (!dev) {
 642                                 condlog(0, "failed getting udev device");
 643                                 continue;
 644                         }
 645                         uev = uevent_from_udev_device(dev);
 646                         if (!uev)
 647                                 continue;
 648                         list_add_tail(&uev->node, &uevlisten_tmp);
 649                         events++;
 650                         continue;
 651                 }
 652                 if (fdcount < 0) {
 653                         if (errno == EINTR)
 654                                 continue;
 655
 656                         condlog(0, "error receiving "
 657                                 "uevent message: %m");
 658                         err = -errno;
 659                         break;
 660                 }
 661                 if (!list_empty(&uevlisten_tmp)) {
 662                         /*
 663                          * Queue uevents and poke service pthread.
 664                          */
 665                         condlog(3, "Forwarding %d uevents", events);
 666                         pthread_mutex_lock(uevq_lockp);
 667                         list_splice_tail_init(&uevlisten_tmp, &uevq);
 668                         pthread_cond_signal(uev_condp);
 669                         pthread_mutex_unlock(uevq_lockp);
 670                         events = 0;
 671                 }
 672                 gettimeofday(&start_time, NULL);
 673                 timeout = 30;
 674         }
 675         pthread_cleanup_pop(1);
 676         pthread_cleanup_pop(1);
 677 out:
 678         pthread_cleanup_pop(1);
 679 out_udev:
 680         pthread_cleanup_pop(1);
 681         return err;
 682 }
 683
 684 char *uevent_get_dm_str(const struct uevent *uev, char *attr)
 685 {
 686         const char *tmp = uevent_get_env_var(uev, attr);
 687
 688         if (tmp == NULL)
 689                 return NULL;
 690         return strdup(tmp);
 691 }
 692
 693 bool uevent_is_mpath(const struct uevent *uev)
 694 {
 695         const char *uuid = uevent_get_env_var(uev, "DM_UUID");
 696
 697         if (uuid == NULL)
 698                 return false;
 699         if (strncmp(uuid, UUID_PREFIX, UUID_PREFIX_LEN))
 700                 return false;
 701         return uuid[UUID_PREFIX_LEN] != '\0';
 702 }