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        <signal.h>
  60 #include        <dirent.h>
  61 #ifdef USE_PTHREADS
  62 #include        <pthread.h>
  63 #else
  64 #include        <sched.h>
  65 #endif
  66
  67 #include        "mdadm.h"
  68 #include        "mdmon.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                 usleep(10);
 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
 175         /* first rule of survival... don't off yourself */
 176         if (pid == getpid())
 177                 return;
 178
 179         /* kill this process if it is mdmon */
 180         sprintf(buf, "/proc/%lu/cmdline", (unsigned long) pid);
 181         fd = open(buf, O_RDONLY);
 182         if (fd < 0)
 183                 return;
 184
 185         n = read(fd, buf, sizeof(buf)-1);
 186         buf[sizeof(buf)-1] = 0;
 187         close(fd);
 188
 189         if (n < 0 || !(strstr(buf, "mdmon") ||
 190                        strstr(buf, "@dmon")))
 191                 return;
 192
 193         kill(pid, SIGTERM);
 194
 195         if (sock < 0)
 196                 return;
 197
 198         /* Wait for monitor to exit by reading from the socket, after
 199          * clearing the non-blocking flag */
 200         fl = fcntl(sock, F_GETFL, 0);
 201         fl &= ~O_NONBLOCK;
 202         fcntl(sock, F_SETFL, fl);
 203         n = read(sock, buf, 100);
 204         /* Ignore result, it is just the wait that
 205          * matters
 206          */
 207 }
 208
 209 void remove_pidfile(char *devname)
 210 {
 211         char buf[100];
 212
 213         sprintf(buf, "%s/%s.pid", MDMON_DIR, devname);
 214         unlink(buf);
 215         sprintf(buf, "%s/%s.sock", MDMON_DIR, devname);
 216         unlink(buf);
 217 }
 218
 219 static int make_control_sock(char *devname)
 220 {
 221         char path[100];
 222         int sfd;
 223         long fl;
 224         struct sockaddr_un addr;
 225
 226         if (sigterm)
 227                 return -1;
 228
 229         sprintf(path, "%s/%s.sock", MDMON_DIR, devname);
 230         unlink(path);
 231         sfd = socket(PF_LOCAL, SOCK_STREAM, 0);
 232         if (sfd < 0)
 233                 return -1;
 234
 235         addr.sun_family = PF_LOCAL;
 236         strcpy(addr.sun_path, path);
 237         umask(077); /* ensure no world write access */
 238         if (bind(sfd, (struct sockaddr*)&addr, sizeof(addr)) < 0) {
 239                 close(sfd);
 240                 return -1;
 241         }
 242         listen(sfd, 10);
 243         fl = fcntl(sfd, F_GETFL, 0);
 244         fl |= O_NONBLOCK;
 245         fcntl(sfd, F_SETFL, fl);
 246         return sfd;
 247 }
 248
 249 static void term(int sig)
 250 {
 251         sigterm = 1;
 252 }
 253
 254 static void wake_me(int sig)
 255 {
 256
 257 }
 258
 259 /* if we are debugging and starting mdmon by hand then don't fork */
 260 static int do_fork(void)
 261 {
 262         #ifdef DEBUG
 263         if (check_env("MDADM_NO_MDMON"))
 264                 return 0;
 265         #endif
 266
 267         return 1;
 268 }
 269
 270 void usage(void)
 271 {
 272         fprintf(stderr,
 273 "Usage: mdmon [options] CONTAINER\n"
 274 "\n"
 275 "Options are:\n"
 276 "  --help        -h   : This message\n"
 277 "  --all         -a   : All devices\n"
 278 "  --foreground  -F   : Run in foreground (do not fork)\n"
 279 "  --takeover    -t   : Takeover container\n"
 280 );
 281         exit(2);
 282 }
 283
 284 static int mdmon(char *devnm, int must_fork, int takeover);
 285
 286 int main(int argc, char *argv[])
 287 {
 288         char *container_name = NULL;
 289         char *devnm = NULL;
 290         int status = 0;
 291         int opt;
 292         int all = 0;
 293         int takeover = 0;
 294         int dofork = 1;
 295         static struct option options[] = {
 296                 {"all", 0, NULL, 'a'},
 297                 {"takeover", 0, NULL, 't'},
 298                 {"help", 0, NULL, 'h'},
 299                 {"offroot", 0, NULL, OffRootOpt},
 300                 {"foreground", 0, NULL, 'F'},
 301                 {NULL, 0, NULL, 0}
 302         };
 303
 304         if (in_initrd()) {
 305                 /*
 306                  * set first char of argv[0] to @. This is used by
 307                  * systemd to signal that the task was launched from
 308                  * initrd/initramfs and should be preserved during shutdown
 309                  */
 310                 argv[0][0] = '@';
 311         }
 312
 313         while ((opt = getopt_long(argc, argv, "thaF", options, NULL)) != -1) {
 314                 switch (opt) {
 315                 case 'a':
 316                         container_name = argv[optind-1];
 317                         all = 1;
 318                         break;
 319                 case 't':
 320                         takeover = 1;
 321                         break;
 322                 case 'F':
 323                         dofork = 0;
 324                         break;
 325                 case OffRootOpt:
 326                         argv[0][0] = '@';
 327                         break;
 328                 case 'h':
 329                 default:
 330                         usage();
 331                         break;
 332                 }
 333         }
 334
 335         if (all == 0 && container_name == NULL) {
 336                 if (argv[optind])
 337                         container_name = argv[optind];
 338         }
 339
 340         if (container_name == NULL)
 341                 usage();
 342
 343         if (argc - optind > 1)
 344                 usage();
 345
 346         if (strcmp(container_name, "/proc/mdstat") == 0)
 347                 all = 1;
 348
 349         if (all) {
 350                 struct mdstat_ent *mdstat, *e;
 351                 int container_len = strlen(container_name);
 352
 353                 /* launch an mdmon instance for each container found */
 354                 mdstat = mdstat_read(0, 0);
 355                 for (e = mdstat; e; e = e->next) {
 356                         if (e->metadata_version &&
 357                             strncmp(e->metadata_version, "external:", 9) == 0 &&
 358                             !is_subarray(&e->metadata_version[9])) {
 359                                 /* update cmdline so this mdmon instance can be
 360                                  * distinguished from others in a call to ps(1)
 361                                  */
 362                                 if (strlen(e->devnm) <= (unsigned)container_len) {
 363                                         memset(container_name, 0, container_len);
 364                                         sprintf(container_name, "%s", e->devnm);
 365                                 }
 366                                 status |= mdmon(e->devnm, 1, takeover);
 367                         }
 368                 }
 369                 free_mdstat(mdstat);
 370
 371                 return status;
 372         } else if (strncmp(container_name, "md", 2) == 0) {
 373                 int id = devnm2devid(container_name);
 374                 if (id)
 375                         devnm = container_name;
 376         } else {
 377                 struct stat st;
 378
 379                 if (stat(container_name, &st) == 0)
 380                         devnm = xstrdup(stat2devnm(&st));
 381         }
 382
 383         if (!devnm) {
 384                 pr_err("%s is not a valid md device name\n",
 385                         container_name);
 386                 exit(1);
 387         }
 388         return mdmon(devnm, dofork && do_fork(), takeover);
 389 }
 390
 391 static int mdmon(char *devnm, int must_fork, int takeover)
 392 {
 393         int mdfd;
 394         struct mdinfo *mdi, *di;
 395         struct supertype *container;
 396         sigset_t set;
 397         struct sigaction act;
 398         int pfd[2];
 399         int status;
 400         int ignore;
 401         pid_t victim = -1;
 402         int victim_sock = -1;
 403
 404         dprintf("starting mdmon for %s\n", devnm);
 405
 406         mdfd = open_dev(devnm);
 407         if (mdfd < 0) {
 408                 pr_err("%s: %s\n", devnm, strerror(errno));
 409                 return 1;
 410         }
 411
 412         /* Fork, and have the child tell us when they are ready */
 413         if (must_fork) {
 414                 if (pipe(pfd) != 0) {
 415                         pr_err("failed to create pipe\n");
 416                         return 1;
 417                 }
 418                 switch(fork()) {
 419                 case -1:
 420                         pr_err("failed to fork: %s\n", strerror(errno));
 421                         return 1;
 422                 case 0: /* child */
 423                         close(pfd[0]);
 424                         break;
 425                 default: /* parent */
 426                         close(pfd[1]);
 427                         if (read(pfd[0], &status, sizeof(status)) != sizeof(status)) {
 428                                 wait(&status);
 429                                 status = WEXITSTATUS(status);
 430                         }
 431                         close(pfd[0]);
 432                         return status;
 433                 }
 434         } else
 435                 pfd[0] = pfd[1] = -1;
 436
 437         container = xcalloc(1, sizeof(*container));
 438         strcpy(container->devnm, devnm);
 439         container->arrays = NULL;
 440         container->sock = -1;
 441
 442         mdi = sysfs_read(mdfd, container->devnm, GET_VERSION|GET_LEVEL|GET_DEVS);
 443
 444         if (!mdi) {
 445                 pr_err("failed to load sysfs info for %s\n", container->devnm);
 446                 exit(3);
 447         }
 448         if (mdi->array.level != UnSet) {
 449                 pr_err("%s is not a container - cannot monitor\n", devnm);
 450                 exit(3);
 451         }
 452         if (mdi->array.major_version != -1 ||
 453             mdi->array.minor_version != -2) {
 454                 pr_err("%s does not use external metadata - cannot monitor\n",
 455                         devnm);
 456                 exit(3);
 457         }
 458
 459         container->ss = version_to_superswitch(mdi->text_version);
 460         if (container->ss == NULL) {
 461                 pr_err("%s uses unsupported metadata: %s\n",
 462                         devnm, mdi->text_version);
 463                 exit(3);
 464         }
 465
 466         container->devs = NULL;
 467         for (di = mdi->devs; di; di = di->next) {
 468                 struct mdinfo *cd = xmalloc(sizeof(*cd));
 469                 *cd = *di;
 470                 cd->next = container->devs;
 471                 container->devs = cd;
 472         }
 473         sysfs_free(mdi);
 474
 475         /* SIGUSR is sent between parent and child.  So both block it
 476          * and enable it only with pselect.
 477          */
 478         sigemptyset(&set);
 479         sigaddset(&set, SIGUSR1);
 480         sigaddset(&set, SIGTERM);
 481         sigprocmask(SIG_BLOCK, &set, NULL);
 482         act.sa_handler = wake_me;
 483         act.sa_flags = 0;
 484         sigaction(SIGUSR1, &act, NULL);
 485         act.sa_handler = term;
 486         sigaction(SIGTERM, &act, NULL);
 487         act.sa_handler = SIG_IGN;
 488         sigaction(SIGPIPE, &act, NULL);
 489
 490         victim = mdmon_pid(container->devnm);
 491         if (victim >= 0)
 492                 victim_sock = connect_monitor(container->devnm);
 493
 494         ignore = chdir("/");
 495         if (!takeover && victim > 0 && victim_sock >= 0) {
 496                 if (fping_monitor(victim_sock) == 0) {
 497                         pr_err("%s already managed\n", container->devnm);
 498                         exit(3);
 499                 }
 500                 close(victim_sock);
 501                 victim_sock = -1;
 502         }
 503         if (container->ss->load_container(container, mdfd, devnm)) {
 504                 pr_err("Cannot load metadata for %s\n", devnm);
 505                 exit(3);
 506         }
 507         close(mdfd);
 508
 509         /* Ok, this is close enough.  We can say goodbye to our parent now.
 510          */
 511         if (victim > 0)
 512                 remove_pidfile(devnm);
 513         if (make_pidfile(devnm) < 0) {
 514                 exit(3);
 515         }
 516         container->sock = make_control_sock(devnm);
 517
 518         status = 0;
 519         if (pfd[1] >= 0) {
 520                 if (write(pfd[1], &status, sizeof(status)) < 0)
 521                         pr_err("failed to notify our parent: %d\n",
 522                                getppid());
 523                 close(pfd[1]);
 524         }
 525
 526         mlockall(MCL_CURRENT | MCL_FUTURE);
 527
 528         if (clone_monitor(container) < 0) {
 529                 pr_err("failed to start monitor process: %s\n",
 530                         strerror(errno));
 531                 exit(2);
 532         }
 533
 534         if (victim > 0) {
 535                 try_kill_monitor(victim, container->devnm, victim_sock);
 536                 if (victim_sock >= 0)
 537                         close(victim_sock);
 538         }
 539
 540         setsid();
 541         close(0);
 542         open("/dev/null", O_RDWR);
 543         close(1);
 544         ignore = dup(0);
 545 #ifndef DEBUG
 546         close(2);
 547         ignore = dup(0);
 548 #endif
 549
 550         /* This silliness is to stop the compiler complaining
 551          * that we ignore 'ignore'
 552          */
 553         if (ignore)
 554                 ignore++;
 555
 556         do_manager(container);
 557
 558         exit(0);
 559 }
 560
 561 /* Some stub functions so super-* can link with us */
 562 int child_monitor(int afd, struct mdinfo *sra, struct reshape *reshape,
 563                   struct supertype *st, unsigned long blocks,
 564                   int *fds, unsigned long long *offsets,
 565                   int dests, int *destfd, unsigned long long *destoffsets)
 566 {
 567         return 0;
 568 }
 569
 570 int restore_stripes(int *dest, unsigned long long *offsets,
 571                     int raid_disks, int chunk_size, int level, int layout,
 572                     int source, unsigned long long read_offset,
 573                     unsigned long long start, unsigned long long length,
 574                     char *src_buf)
 575 {
 576         return 1;
 577 }
 578
 579 int save_stripes(int *source, unsigned long long *offsets,
 580                  int raid_disks, int chunk_size, int level, int layout,
 581                  int nwrites, int *dest,
 582                  unsigned long long start, unsigned long long length,
 583                  char *buf)
 584 {
 585         return 0;
 586 }
 587
 588 struct superswitch super0 = {
 589         .name = "0.90",
 590 };
 591 struct superswitch super1 = {
 592         .name = "1.x",
 593 };