mdmon.c

   1 /*
   2  * mdmon - monitor external metadata arrays
   3  *
   4  * Copyright (C) 2007-2009 Neil Brown <neilb@suse.de>
   5  * Copyright (C) 2007-2009 Intel Corporation
   6  *
   7  * This program is free software; you can redistribute it and/or modify it
   8  * under the terms and conditions of the GNU General Public License,
   9  * version 2, as published by the Free Software Foundation.
  10  *
  11  * This program is distributed in the hope it will be useful, but WITHOUT
  12  * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
  13  * FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License for
  14  * more details.
  15  *
  16  * You should have received a copy of the GNU General Public License along with
  17  * this program; if not, write to the Free Software Foundation, Inc.,
  18  * 51 Franklin St - Fifth Floor, Boston, MA 02110-1301 USA.
  19  */
  20
  21 /*
  22  * md array manager.
  23  * When md arrays have user-space managed metadata, this is the program
  24  * that does the managing.
  25  *
  26  * Given one argument: the name of the array (e.g. /dev/md0) that is
  27  * the container.
  28  * We fork off a helper that runs high priority and mlocked.  It responds to
  29  * device failures and other events that might stop writeout, or that are
  30  * trivial to deal with.
  31  * The main thread then watches for new arrays being created in the container
  32  * and starts monitoring them too ... along with a few other tasks.
  33  *
  34  * The main thread communicates with the priority thread by writing over
  35  * a pipe.
  36  * Separate programs can communicate with the main thread via Unix-domain
  37  * socket.
  38  * The two threads share address space and open file table.
  39  *
  40  */
  41
  42 #ifndef _GNU_SOURCE
  43 #define _GNU_SOURCE
  44 #endif
  45
  46 #include        <unistd.h>
  47 #include        <stdlib.h>
  48 #include        <sys/types.h>
  49 #include        <sys/stat.h>
  50 #include        <sys/socket.h>
  51 #include        <sys/un.h>
  52 #include        <sys/mman.h>
  53 #include        <sys/syscall.h>
  54 #include        <sys/wait.h>
  55 #include        <stdio.h>
  56 #include        <errno.h>
  57 #include        <string.h>
  58 #include        <fcntl.h>
  59 #include        <dirent.h>
  60 #ifdef USE_PTHREADS
  61 #include        <pthread.h>
  62 #else
  63 #include        <sched.h>
  64 #endif
  65
  66 #include        "mdadm.h"
  67 #include        "mdmon.h"
  68 #include        "xmalloc.h"
  69
  70 char const Name[] = "mdmon";
  71
  72 struct active_array *discard_this;
  73 struct active_array *pending_discard;
  74
  75 int mon_tid, mgr_tid;
  76
  77 int sigterm;
  78
  79 #ifdef USE_PTHREADS
  80 static void *run_child(void *v)
  81 {
  82         struct supertype *c = v;
  83
  84         mon_tid = syscall(SYS_gettid);
  85         do_monitor(c);
  86         return 0;
  87 }
  88
  89 static int clone_monitor(struct supertype *container)
  90 {
  91         pthread_attr_t attr;
  92         pthread_t thread;
  93         int rc;
  94
  95         mon_tid = -1;
  96         pthread_attr_init(&attr);
  97         pthread_attr_setstacksize(&attr, 4096);
  98         pthread_attr_setdetachstate(&attr, PTHREAD_CREATE_DETACHED);
  99         rc = pthread_create(&thread, &attr, run_child, container);
 100         if (rc)
 101                 return rc;
 102         while (mon_tid == -1)
 103                 sleep_for(0, USEC_TO_NSEC(10), true);
 104         pthread_attr_destroy(&attr);
 105
 106         mgr_tid = syscall(SYS_gettid);
 107
 108         return mon_tid;
 109 }
 110 #else /* USE_PTHREADS */
 111 static int run_child(void *v)
 112 {
 113         struct supertype *c = v;
 114
 115         do_monitor(c);
 116         return 0;
 117 }
 118
 119 #ifdef __ia64__
 120 int __clone2(int (*fn)(void *),
 121             void *child_stack_base, size_t stack_size,
 122             int flags, void *arg, ...
 123          /* pid_t *pid, struct user_desc *tls, pid_t *ctid */ );
 124 #endif
 125 static int clone_monitor(struct supertype *container)
 126 {
 127         static char stack[4096];
 128
 129 #ifdef __ia64__
 130         mon_tid = __clone2(run_child, stack, sizeof(stack),
 131                    CLONE_FS|CLONE_FILES|CLONE_VM|CLONE_SIGHAND|CLONE_THREAD,
 132                    container);
 133 #else
 134         mon_tid = clone(run_child, stack+4096-64,
 135                    CLONE_FS|CLONE_FILES|CLONE_VM|CLONE_SIGHAND|CLONE_THREAD,
 136                    container);
 137 #endif
 138
 139         mgr_tid = syscall(SYS_gettid);
 140
 141         return mon_tid;
 142 }
 143 #endif /* USE_PTHREADS */
 144
 145 static int make_pidfile(char *devname)
 146 {
 147         char path[100];
 148         char pid[10];
 149         int fd;
 150         int n;
 151
 152         if (mkdir(MDMON_DIR, 0755) < 0 &&
 153             errno != EEXIST)
 154                 return -errno;
 155         sprintf(path, "%s/%s.pid", MDMON_DIR, devname);
 156
 157         fd = open(path, O_RDWR|O_CREAT|O_EXCL, 0600);
 158         if (fd < 0)
 159                 return -errno;
 160         sprintf(pid, "%d\n", getpid());
 161         n = write(fd, pid, strlen(pid));
 162         close(fd);
 163         if (n < 0)
 164                 return -errno;
 165         return 0;
 166 }
 167
 168 static void try_kill_monitor(pid_t pid, char *devname, int sock)
 169 {
 170         char buf[100];
 171         int fd;
 172         int n;
 173         long fl;
 174         int rv;
 175
 176         /* first rule of survival... don't off yourself */
 177         if (pid == getpid())
 178                 return;
 179
 180         /* kill this process if it is mdmon */
 181         sprintf(buf, "/proc/%lu/cmdline", (unsigned long) pid);
 182         fd = open(buf, O_RDONLY);
 183         if (fd < 0)
 184                 return;
 185
 186         n = read(fd, buf, sizeof(buf)-1);
 187         buf[sizeof(buf)-1] = 0;
 188         close(fd);
 189
 190         if (n < 0 || !(strstr(buf, "mdmon") ||
 191                        strstr(buf, "@dmon")))
 192                 return;
 193
 194         kill(pid, SIGTERM);
 195
 196         if (sock < 0)
 197                 return;
 198
 199         /* Wait for monitor to exit by reading from the socket, after
 200          * clearing the non-blocking flag */
 201         fl = fcntl(sock, F_GETFL, 0);
 202         if (fl < 0)
 203                 return;
 204
 205         fl &= ~O_NONBLOCK;
 206         if (fcntl(sock, F_SETFL, fl) < 0)
 207                 return;
 208         n = read(sock, buf, 100);
 209
 210         /* If there is I/O going on it might took some time to get to
 211          * clean state. Wait for monitor to exit fully to avoid races.
 212          * Ping it with SIGUSR1 in case that it is sleeping  */
 213         for (n = 0; n < 25; n++) {
 214                 rv = kill(pid, SIGUSR1);
 215                 if (rv < 0)
 216                         break;
 217                 sleep_for(0, MSEC_TO_NSEC(200), true);
 218         }
 219 }
 220
 221 void remove_pidfile(char *devname)
 222 {
 223         char buf[100];
 224
 225         sprintf(buf, "%s/%s.pid", MDMON_DIR, devname);
 226         unlink(buf);
 227         sprintf(buf, "%s/%s.sock", MDMON_DIR, devname);
 228         unlink(buf);
 229 }
 230
 231 static int make_control_sock(char *devname)
 232 {
 233         char path[100];
 234         int sfd;
 235         long fl;
 236         struct sockaddr_un addr;
 237
 238         if (sigterm)
 239                 return -1;
 240
 241         sprintf(path, "%s/%s.sock", MDMON_DIR, devname);
 242         unlink(path);
 243         sfd = socket(PF_LOCAL, SOCK_STREAM, 0);
 244         if (sfd < 0)
 245                 return -1;
 246
 247         addr.sun_family = PF_LOCAL;
 248         snprintf(addr.sun_path, sizeof(addr.sun_path), "%s", path);
 249         umask(077); /* ensure no world write access */
 250         if (bind(sfd, (struct sockaddr*)&addr, sizeof(addr)) < 0) {
 251                 close(sfd);
 252                 return -1;
 253         }
 254         listen(sfd, 10);
 255         fl = fcntl(sfd, F_GETFL, 0);
 256         fl |= O_NONBLOCK;
 257         if (fcntl(sfd, F_SETFL, fl) < 0) {
 258                 close_fd(&sfd);
 259                 return -1;
 260         }
 261         return sfd;
 262 }
 263
 264 static void term(int sig)
 265 {
 266         sigterm = 1;
 267 }
 268
 269 static void wake_me(int sig)
 270 {
 271
 272 }
 273
 274 /* if we are debugging and starting mdmon by hand then don't fork */
 275 static int do_fork(void)
 276 {
 277         #ifdef DEBUG
 278         if (check_env("MDADM_NO_MDMON"))
 279                 return 0;
 280         #endif
 281
 282         return 1;
 283 }
 284
 285 void usage(void)
 286 {
 287         fprintf(stderr,
 288 "Usage: mdmon [options] CONTAINER\n"
 289 "\n"
 290 "Options are:\n"
 291 "  --help        -h   : This message\n"
 292 "  --all         -a   : All devices\n"
 293 "  --foreground  -F   : Run in foreground (do not fork)\n"
 294 "  --takeover    -t   : Takeover container\n"
 295 );
 296         exit(2);
 297 }
 298
 299 static bool is_duplicate_opt(const int opt, const int set_val, const char *long_name)
 300 {
 301         if (opt == set_val) {
 302                 pr_err("--%s option duplicated!\n", long_name);
 303                 return true;
 304         }
 305         return false;
 306 }
 307
 308 static int mdmon(char *devnm, int must_fork, int takeover);
 309
 310 int main(int argc, char *argv[])
 311 {
 312         char *container_name = NULL;
 313         int status = 0;
 314         int opt;
 315         int all = 0;
 316         int takeover = 0;
 317         int dofork = 1;
 318         int mdfd = -1;
 319         bool help = false;
 320         static struct option options[] = {
 321                 {"all", 0, NULL, 'a'},
 322                 {"takeover", 0, NULL, 't'},
 323                 {"help", 0, NULL, 'h'},
 324                 {"offroot", 0, NULL, OffRootOpt},
 325                 {"foreground", 0, NULL, 'F'},
 326                 {NULL, 0, NULL, 0}
 327         };
 328
 329         /*
 330          * mdmon should never complain due to lack of a platform,
 331          * that is mdadm's job if at all.
 332          */
 333         imsm_set_no_platform(1);
 334
 335         while ((opt = getopt_long(argc, argv, "thaF", options, NULL)) != -1) {
 336                 switch (opt) {
 337                 case 'a':
 338                         if (is_duplicate_opt(all, 1, "all"))
 339                                 exit(1);
 340                         container_name = argv[optind-1];
 341                         all = 1;
 342                         break;
 343                 case 't':
 344                         if (is_duplicate_opt(takeover, 1, "takeover"))
 345                                 exit(1);
 346                         takeover = 1;
 347                         break;
 348                 case 'F':
 349                         if (is_duplicate_opt(dofork, 0, "foreground"))
 350                                 exit(1);
 351                         dofork = 0;
 352                         break;
 353                 case OffRootOpt:
 354                         if (is_duplicate_opt(argv[0][0], '@', "offroot"))
 355                                 exit(1);
 356                         argv[0][0] = '@';
 357                         break;
 358                 case 'h':
 359                         if (is_duplicate_opt(help, true, "help"))
 360                                 exit(1);
 361                         help = true;
 362                         break;
 363                 default:
 364                         usage();
 365                         break;
 366                 }
 367         }
 368
 369         if (in_initrd()) {
 370                 /*
 371                  * set first char of argv[0] to @. This is used by
 372                  * systemd to signal that the task was launched from
 373                  * initrd/initramfs and should be preserved during shutdown
 374                  */
 375                 argv[0][0] = '@';
 376         }
 377
 378         if (!all && argv[optind]) {
 379                 static const char prefix[] = "initrd/";
 380                 container_name = argv[optind];
 381                 if (strncmp(container_name, prefix,
 382                             sizeof(prefix) - 1) == 0)
 383                         container_name += sizeof(prefix)-1;
 384                 container_name = get_md_name(container_name);
 385                 if (!container_name)
 386                         return 1;
 387         }
 388
 389         if (container_name == NULL || argc - optind > 1)
 390                 usage();
 391
 392         if (strcmp(container_name, "/proc/mdstat") == 0)
 393                 all = 1;
 394
 395         if (help)
 396                 usage();
 397
 398         if (all) {
 399                 struct mdstat_ent *mdstat, *e;
 400                 int container_len = strnlen(container_name, MD_NAME_MAX);
 401
 402                 /* launch an mdmon instance for each container found */
 403                 mdstat = mdstat_read(0, 0);
 404                 for (e = mdstat; e; e = e->next) {
 405                         if (is_mdstat_ent_external(e) && !is_mdstat_ent_subarray(e)) {
 406                                 /* update cmdline so this mdmon instance can be
 407                                  * distinguished from others in a call to ps(1)
 408                                  */
 409                                 if (strlen(e->devnm) <= (unsigned)container_len) {
 410                                         memset(container_name, 0, container_len);
 411                                         sprintf(container_name, "%s", e->devnm);
 412                                 }
 413                                 status |= mdmon(e->devnm, 1, takeover);
 414                         }
 415                 }
 416                 free_mdstat(mdstat);
 417
 418                 return status;
 419         }
 420
 421         mdfd = open_mddev(container_name, 0);
 422         if (is_fd_valid(mdfd)) {
 423                 char *devnm = fd2devnm(mdfd);
 424
 425                 close(mdfd);
 426
 427                 if (devnm)
 428                         return mdmon(devnm, dofork && do_fork(), takeover);
 429         }
 430
 431         pr_err("%s is not a valid md device name\n", container_name);
 432         return 1;
 433 }
 434
 435 static int mdmon(char *devnm, int must_fork, int takeover)
 436 {
 437         int mdfd;
 438         struct mdinfo *mdi, *di;
 439         struct supertype *container;
 440         sigset_t set;
 441         struct sigaction act;
 442         int pfd[2];
 443         int status;
 444         int ignore;
 445         pid_t victim = -1;
 446         int victim_sock = -1;
 447
 448         dprintf("starting mdmon for %s\n", devnm);
 449
 450         mdfd = open_dev(devnm);
 451         if (mdfd < 0) {
 452                 pr_err("%s: %s\n", devnm, strerror(errno));
 453                 return 1;
 454         }
 455
 456         /* Fork, and have the child tell us when they are ready */
 457         if (must_fork) {
 458                 if (pipe(pfd) != 0) {
 459                         pr_err("failed to create pipe\n");
 460                         close_fd(&mdfd);
 461                         return 1;
 462                 }
 463                 switch(fork()) {
 464                 case -1:
 465                         pr_err("failed to fork: %s\n", strerror(errno));
 466                         close_fd(&mdfd);
 467                         return 1;
 468                 case 0: /* child */
 469                         close_fd(&pfd[0]);
 470                         break;
 471                 default: /* parent */
 472                         close_fd(&pfd[1]);
 473                         if (read(pfd[0], &status, sizeof(status)) != sizeof(status)) {
 474                                 wait(&status);
 475                                 status = WEXITSTATUS(status);
 476                         }
 477                         close_fd(&pfd[0]);
 478                         close_fd(&mdfd);
 479                         return status;
 480                 }
 481         } else
 482                 pfd[0] = pfd[1] = -1;
 483
 484         container = xcalloc(1, sizeof(*container));
 485         snprintf(container->devnm, MD_NAME_MAX, "%s", devnm);
 486         container->arrays = NULL;
 487         container->sock = -1;
 488
 489         mdi = sysfs_read(mdfd, container->devnm, GET_VERSION|GET_LEVEL|GET_DEVS);
 490
 491         if (!mdi) {
 492                 pr_err("failed to load sysfs info for %s\n", container->devnm);
 493                 exit(3);
 494         }
 495         if (mdi->array.level != UnSet) {
 496                 pr_err("%s is not a container - cannot monitor\n", devnm);
 497                 exit(3);
 498         }
 499         if (mdi->array.major_version != -1 ||
 500             mdi->array.minor_version != -2) {
 501                 pr_err("%s does not use external metadata - cannot monitor\n",
 502                         devnm);
 503                 exit(3);
 504         }
 505
 506         container->ss = version_to_superswitch(mdi->text_version);
 507         if (container->ss == NULL) {
 508                 pr_err("%s uses unsupported metadata: %s\n",
 509                         devnm, mdi->text_version);
 510                 exit(3);
 511         }
 512
 513         container->devs = NULL;
 514         for (di = mdi->devs; di; di = di->next) {
 515                 struct mdinfo *cd = xmalloc(sizeof(*cd));
 516                 *cd = *di;
 517                 cd->next = container->devs;
 518                 container->devs = cd;
 519         }
 520         sysfs_free(mdi);
 521
 522         /* SIGUSR is sent between parent and child.  So both block it
 523          * and enable it only with pselect.
 524          */
 525         sigemptyset(&set);
 526         sigaddset(&set, SIGUSR1);
 527         sigaddset(&set, SIGTERM);
 528         sigprocmask(SIG_BLOCK, &set, NULL);
 529         act.sa_handler = wake_me;
 530         act.sa_flags = 0;
 531         sigaction(SIGUSR1, &act, NULL);
 532         act.sa_handler = term;
 533         sigaction(SIGTERM, &act, NULL);
 534         act.sa_handler = SIG_IGN;
 535         sigaction(SIGPIPE, &act, NULL);
 536
 537         victim = mdmon_pid(container->devnm);
 538         if (victim >= 0)
 539                 victim_sock = connect_monitor(container->devnm);
 540
 541         ignore = chdir("/");
 542         if (!takeover && victim > 0 && victim_sock >= 0) {
 543                 if (fping_monitor(victim_sock) == 0) {
 544                         pr_err("%s already managed\n", container->devnm);
 545                         exit(3);
 546                 }
 547                 close(victim_sock);
 548                 victim_sock = -1;
 549         }
 550         if (container->ss->load_container(container, mdfd, devnm)) {
 551                 pr_err("Cannot load metadata for %s\n", devnm);
 552                 exit(3);
 553         }
 554         close(mdfd);
 555
 556         /* Ok, this is close enough.  We can say goodbye to our parent now.
 557          */
 558         if (victim > 0)
 559                 remove_pidfile(devnm);
 560         if (make_pidfile(devnm) < 0) {
 561                 exit(3);
 562         }
 563         container->sock = make_control_sock(devnm);
 564
 565         status = 0;
 566         if (pfd[1] >= 0) {
 567                 if (write(pfd[1], &status, sizeof(status)) < 0)
 568                         pr_err("failed to notify our parent: %d\n",
 569                                getppid());
 570                 close(pfd[1]);
 571         }
 572
 573         mlockall(MCL_CURRENT | MCL_FUTURE);
 574
 575         if (clone_monitor(container) < 0) {
 576                 pr_err("failed to start monitor process: %s\n",
 577                         strerror(errno));
 578                 exit(2);
 579         }
 580
 581         if (victim > 0) {
 582                 try_kill_monitor(victim, container->devnm, victim_sock);
 583                 if (victim_sock >= 0)
 584                         close(victim_sock);
 585         }
 586
 587         setsid();
 588         manage_fork_fds(0);
 589
 590         /* This silliness is to stop the compiler complaining
 591          * that we ignore 'ignore'
 592          */
 593         if (ignore)
 594                 ignore++;
 595
 596         do_manager(container);
 597
 598         exit(0);
 599 }
 600
 601 /* Some stub functions so super-* can link with us */
 602 int child_monitor(int afd, struct mdinfo *sra, struct reshape *reshape,
 603                   struct supertype *st, unsigned long blocks,
 604                   int *fds, unsigned long long *offsets,
 605                   int dests, int *destfd, unsigned long long *destoffsets)
 606 {
 607         return 0;
 608 }
 609
 610 int restore_stripes(int *dest, unsigned long long *offsets,
 611                     int raid_disks, int chunk_size, int level, int layout,
 612                     int source, unsigned long long read_offset,
 613                     unsigned long long start, unsigned long long length,
 614                     char *src_buf)
 615 {
 616         return 1;
 617 }
 618
 619 int save_stripes(int *source, unsigned long long *offsets,
 620                  int raid_disks, int chunk_size, int level, int layout,
 621                  int nwrites, int *dest,
 622                  unsigned long long start, unsigned long long length,
 623                  char *buf)
 624 {
 625         return 0;
 626 }
 627
 628 struct superswitch super0 = {
 629         .name = "0.90",
 630 };
 631 struct superswitch super1 = {
 632         .name = "1.x",
 633 };