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 struct active_array *discard_this;
  71 struct active_array *pending_discard;
  72
  73 int mon_tid, mgr_tid;
  74
  75 int sigterm;
  76
  77 #ifdef USE_PTHREADS
  78 static void *run_child(void *v)
  79 {
  80         struct supertype *c = v;
  81
  82         mon_tid = syscall(SYS_gettid);
  83         do_monitor(c);
  84         return 0;
  85 }
  86
  87 static int clone_monitor(struct supertype *container)
  88 {
  89         pthread_attr_t attr;
  90         pthread_t thread;
  91         int rc;
  92
  93         mon_tid = -1;
  94         pthread_attr_init(&attr);
  95         pthread_attr_setstacksize(&attr, 4096);
  96         pthread_attr_setdetachstate(&attr, PTHREAD_CREATE_DETACHED);
  97         rc = pthread_create(&thread, &attr, run_child, container);
  98         if (rc)
  99                 return rc;
 100         while (mon_tid == -1)
 101                 usleep(10);
 102         pthread_attr_destroy(&attr);
 103
 104         mgr_tid = syscall(SYS_gettid);
 105
 106         return mon_tid;
 107 }
 108 #else /* USE_PTHREADS */
 109 static int run_child(void *v)
 110 {
 111         struct supertype *c = v;
 112
 113         do_monitor(c);
 114         return 0;
 115 }
 116
 117 #ifdef __ia64__
 118 int __clone2(int (*fn)(void *),
 119             void *child_stack_base, size_t stack_size,
 120             int flags, void *arg, ...
 121          /* pid_t *pid, struct user_desc *tls, pid_t *ctid */ );
 122 #endif
 123 static int clone_monitor(struct supertype *container)
 124 {
 125         static char stack[4096];
 126
 127 #ifdef __ia64__
 128         mon_tid = __clone2(run_child, stack, sizeof(stack),
 129                    CLONE_FS|CLONE_FILES|CLONE_VM|CLONE_SIGHAND|CLONE_THREAD,
 130                    container);
 131 #else
 132         mon_tid = clone(run_child, stack+4096-64,
 133                    CLONE_FS|CLONE_FILES|CLONE_VM|CLONE_SIGHAND|CLONE_THREAD,
 134                    container);
 135 #endif
 136
 137         mgr_tid = syscall(SYS_gettid);
 138
 139         return mon_tid;
 140 }
 141 #endif /* USE_PTHREADS */
 142
 143 static int make_pidfile(char *devname)
 144 {
 145         char path[100];
 146         char pid[10];
 147         int fd;
 148         int n;
 149
 150         if (mkdir(MDMON_DIR, 0755) < 0 &&
 151             errno != EEXIST)
 152                 return -errno;
 153         sprintf(path, "%s/%s.pid", MDMON_DIR, devname);
 154
 155         fd = open(path, O_RDWR|O_CREAT|O_EXCL, 0600);
 156         if (fd < 0)
 157                 return -errno;
 158         sprintf(pid, "%d\n", getpid());
 159         n = write(fd, pid, strlen(pid));
 160         close(fd);
 161         if (n < 0)
 162                 return -errno;
 163         return 0;
 164 }
 165
 166 static void try_kill_monitor(pid_t pid, char *devname, int sock)
 167 {
 168         char buf[100];
 169         int fd;
 170         int n;
 171         long fl;
 172
 173         /* first rule of survival... don't off yourself */
 174         if (pid == getpid())
 175                 return;
 176
 177         /* kill this process if it is mdmon */
 178         sprintf(buf, "/proc/%lu/cmdline", (unsigned long) pid);
 179         fd = open(buf, O_RDONLY);
 180         if (fd < 0)
 181                 return;
 182
 183         n = read(fd, buf, sizeof(buf)-1);
 184         buf[sizeof(buf)-1] = 0;
 185         close(fd);
 186
 187         if (n < 0 || !(strstr(buf, "mdmon") ||
 188                        strstr(buf, "@dmon")))
 189                 return;
 190
 191         kill(pid, SIGTERM);
 192
 193         if (sock < 0)
 194                 return;
 195
 196         /* Wait for monitor to exit by reading from the socket, after
 197          * clearing the non-blocking flag */
 198         fl = fcntl(sock, F_GETFL, 0);
 199         fl &= ~O_NONBLOCK;
 200         fcntl(sock, F_SETFL, fl);
 201         n = read(sock, buf, 100);
 202         /* Ignore result, it is just the wait that
 203          * matters
 204          */
 205 }
 206
 207 void remove_pidfile(char *devname)
 208 {
 209         char buf[100];
 210
 211         sprintf(buf, "%s/%s.pid", MDMON_DIR, devname);
 212         unlink(buf);
 213         sprintf(buf, "%s/%s.sock", MDMON_DIR, devname);
 214         unlink(buf);
 215 }
 216
 217 static int make_control_sock(char *devname)
 218 {
 219         char path[100];
 220         int sfd;
 221         long fl;
 222         struct sockaddr_un addr;
 223
 224         if (sigterm)
 225                 return -1;
 226
 227         sprintf(path, "%s/%s.sock", MDMON_DIR, devname);
 228         unlink(path);
 229         sfd = socket(PF_LOCAL, SOCK_STREAM, 0);
 230         if (sfd < 0)
 231                 return -1;
 232
 233         addr.sun_family = PF_LOCAL;
 234         strcpy(addr.sun_path, path);
 235         if (bind(sfd, &addr, sizeof(addr)) < 0) {
 236                 close(sfd);
 237                 return -1;
 238         }
 239         listen(sfd, 10);
 240         fl = fcntl(sfd, F_GETFL, 0);
 241         fl |= O_NONBLOCK;
 242         fcntl(sfd, F_SETFL, fl);
 243         return sfd;
 244 }
 245
 246 static void term(int sig)
 247 {
 248         sigterm = 1;
 249 }
 250
 251 static void wake_me(int sig)
 252 {
 253
 254 }
 255
 256 /* if we are debugging and starting mdmon by hand then don't fork */
 257 static int do_fork(void)
 258 {
 259         #ifdef DEBUG
 260         if (check_env("MDADM_NO_MDMON"))
 261                 return 0;
 262         #endif
 263
 264         return 1;
 265 }
 266
 267 void usage(void)
 268 {
 269         fprintf(stderr,
 270 "Usage: mdmon [options] CONTAINER\n"
 271 "\n"
 272 "Options are:\n"
 273 "  --help        -h   : This message\n"
 274 "  --all         -a   : All devices\n"
 275 "  --foreground  -F   : Run in foreground (do not fork)\n"
 276 "  --takeover    -t   : Takeover container\n"
 277 );
 278         exit(2);
 279 }
 280
 281 static int mdmon(char *devnm, int must_fork, int takeover);
 282
 283 int main(int argc, char *argv[])
 284 {
 285         char *container_name = NULL;
 286         char *devnm = NULL;
 287         int status = 0;
 288         int opt;
 289         int all = 0;
 290         int takeover = 0;
 291         int dofork = 1;
 292         static struct option options[] = {
 293                 {"all", 0, NULL, 'a'},
 294                 {"takeover", 0, NULL, 't'},
 295                 {"help", 0, NULL, 'h'},
 296                 {"offroot", 0, NULL, OffRootOpt},
 297                 {"foreground", 0, NULL, 'F'},
 298                 {NULL, 0, NULL, 0}
 299         };
 300
 301         if (in_initrd()) {
 302                 /*
 303                  * set first char of argv[0] to @. This is used by
 304                  * systemd to signal that the task was launched from
 305                  * initrd/initramfs and should be preserved during shutdown
 306                  */
 307                 argv[0][0] = '@';
 308         }
 309
 310         while ((opt = getopt_long(argc, argv, "thaF", options, NULL)) != -1) {
 311                 switch (opt) {
 312                 case 'a':
 313                         container_name = argv[optind-1];
 314                         all = 1;
 315                         break;
 316                 case 't':
 317                         takeover = 1;
 318                         break;
 319                 case 'F':
 320                         dofork = 0;
 321                         break;
 322                 case OffRootOpt:
 323                         argv[0][0] = '@';
 324                         break;
 325                 case 'h':
 326                 default:
 327                         usage();
 328                         break;
 329                 }
 330         }
 331
 332         if (all == 0 && container_name == NULL) {
 333                 if (argv[optind])
 334                         container_name = argv[optind];
 335         }
 336
 337         if (container_name == NULL)
 338                 usage();
 339
 340         if (argc - optind > 1)
 341                 usage();
 342
 343         if (strcmp(container_name, "/proc/mdstat") == 0)
 344                 all = 1;
 345
 346         if (all) {
 347                 struct mdstat_ent *mdstat, *e;
 348                 int container_len = strlen(container_name);
 349
 350                 /* launch an mdmon instance for each container found */
 351                 mdstat = mdstat_read(0, 0);
 352                 for (e = mdstat; e; e = e->next) {
 353                         if (e->metadata_version &&
 354                             strncmp(e->metadata_version, "external:", 9) == 0 &&
 355                             !is_subarray(&e->metadata_version[9])) {
 356                                 /* update cmdline so this mdmon instance can be
 357                                  * distinguished from others in a call to ps(1)
 358                                  */
 359                                 if (strlen(e->devnm) <= (unsigned)container_len) {
 360                                         memset(container_name, 0, container_len);
 361                                         sprintf(container_name, "%s", e->devnm);
 362                                 }
 363                                 status |= mdmon(e->devnm, 1, takeover);
 364                         }
 365                 }
 366                 free_mdstat(mdstat);
 367
 368                 return status;
 369         } else if (strncmp(container_name, "md", 2) == 0) {
 370                 int id = devnm2devid(container_name);
 371                 if (id)
 372                         devnm = container_name;
 373         } else {
 374                 struct stat st;
 375
 376                 if (stat(container_name, &st) == 0)
 377                         devnm = xstrdup(stat2devnm(&st));
 378         }
 379
 380         if (!devnm) {
 381                 pr_err("%s is not a valid md device name\n",
 382                         container_name);
 383                 exit(1);
 384         }
 385         return mdmon(devnm, dofork && do_fork(), takeover);
 386 }
 387
 388 static int mdmon(char *devnm, int must_fork, int takeover)
 389 {
 390         int mdfd;
 391         struct mdinfo *mdi, *di;
 392         struct supertype *container;
 393         sigset_t set;
 394         struct sigaction act;
 395         int pfd[2];
 396         int status;
 397         int ignore;
 398         pid_t victim = -1;
 399         int victim_sock = -1;
 400
 401         dprintf("starting mdmon for %s\n", devnm);
 402
 403         mdfd = open_dev(devnm);
 404         if (mdfd < 0) {
 405                 pr_err("%s: %s\n", devnm, strerror(errno));
 406                 return 1;
 407         }
 408         if (md_get_version(mdfd) < 0) {
 409                 pr_err("%s: Not an md device\n", devnm);
 410                 return 1;
 411         }
 412
 413         /* Fork, and have the child tell us when they are ready */
 414         if (must_fork) {
 415                 if (pipe(pfd) != 0) {
 416                         pr_err("failed to create pipe\n");
 417                         return 1;
 418                 }
 419                 switch(fork()) {
 420                 case -1:
 421                         pr_err("failed to fork: %s\n", strerror(errno));
 422                         return 1;
 423                 case 0: /* child */
 424                         close(pfd[0]);
 425                         break;
 426                 default: /* parent */
 427                         close(pfd[1]);
 428                         if (read(pfd[0], &status, sizeof(status)) != sizeof(status)) {
 429                                 wait(&status);
 430                                 status = WEXITSTATUS(status);
 431                         }
 432                         close(pfd[0]);
 433                         return status;
 434                 }
 435         } else
 436                 pfd[0] = pfd[1] = -1;
 437
 438         container = xcalloc(1, sizeof(*container));
 439         strcpy(container->devnm, devnm);
 440         container->arrays = NULL;
 441         container->sock = -1;
 442
 443         mdi = sysfs_read(mdfd, container->devnm, GET_VERSION|GET_LEVEL|GET_DEVS);
 444
 445         if (!mdi) {
 446                 pr_err("failed to load sysfs info for %s\n", container->devnm);
 447                 exit(3);
 448         }
 449         if (mdi->array.level != UnSet) {
 450                 pr_err("%s is not a container - cannot monitor\n", devnm);
 451                 exit(3);
 452         }
 453         if (mdi->array.major_version != -1 ||
 454             mdi->array.minor_version != -2) {
 455                 pr_err("%s does not use external metadata - cannot monitor\n",
 456                         devnm);
 457                 exit(3);
 458         }
 459
 460         container->ss = version_to_superswitch(mdi->text_version);
 461         if (container->ss == NULL) {
 462                 pr_err("%s uses unsupported metadata: %s\n",
 463                         devnm, mdi->text_version);
 464                 exit(3);
 465         }
 466
 467         container->devs = NULL;
 468         for (di = mdi->devs; di; di = di->next) {
 469                 struct mdinfo *cd = xmalloc(sizeof(*cd));
 470                 *cd = *di;
 471                 cd->next = container->devs;
 472                 container->devs = cd;
 473         }
 474         sysfs_free(mdi);
 475
 476         /* SIGUSR is sent between parent and child.  So both block it
 477          * and enable it only with pselect.
 478          */
 479         sigemptyset(&set);
 480         sigaddset(&set, SIGUSR1);
 481         sigaddset(&set, SIGTERM);
 482         sigprocmask(SIG_BLOCK, &set, NULL);
 483         act.sa_handler = wake_me;
 484         act.sa_flags = 0;
 485         sigaction(SIGUSR1, &act, NULL);
 486         act.sa_handler = term;
 487         sigaction(SIGTERM, &act, NULL);
 488         act.sa_handler = SIG_IGN;
 489         sigaction(SIGPIPE, &act, NULL);
 490
 491         victim = mdmon_pid(container->devnm);
 492         if (victim >= 0)
 493                 victim_sock = connect_monitor(container->devnm);
 494
 495         ignore = chdir("/");
 496         if (!takeover && victim > 0 && victim_sock >= 0) {
 497                 if (fping_monitor(victim_sock) == 0) {
 498                         pr_err("%s already managed\n", container->devnm);
 499                         exit(3);
 500                 }
 501                 close(victim_sock);
 502                 victim_sock = -1;
 503         }
 504         if (container->ss->load_container(container, mdfd, devnm)) {
 505                 pr_err("Cannot load metadata for %s\n", devnm);
 506                 exit(3);
 507         }
 508         close(mdfd);
 509
 510         /* Ok, this is close enough.  We can say goodbye to our parent now.
 511          */
 512         if (victim > 0)
 513                 remove_pidfile(devnm);
 514         if (make_pidfile(devnm) < 0) {
 515                 exit(3);
 516         }
 517         container->sock = make_control_sock(devnm);
 518
 519         status = 0;
 520         if (pfd[1] >= 0) {
 521                 if (write(pfd[1], &status, sizeof(status)) < 0)
 522                         pr_err("failed to notify our parent: %d\n",
 523                                getppid());
 524                 close(pfd[1]);
 525         }
 526
 527         mlockall(MCL_CURRENT | MCL_FUTURE);
 528
 529         if (clone_monitor(container) < 0) {
 530                 pr_err("failed to start monitor process: %s\n",
 531                         strerror(errno));
 532                 exit(2);
 533         }
 534
 535         if (victim > 0) {
 536                 try_kill_monitor(victim, container->devnm, victim_sock);
 537                 if (victim_sock >= 0)
 538                         close(victim_sock);
 539         }
 540
 541         setsid();
 542         close(0);
 543         open("/dev/null", O_RDWR);
 544         close(1);
 545         ignore = dup(0);
 546 #ifndef DEBUG
 547         close(2);
 548         ignore = dup(0);
 549 #endif
 550
 551         /* This silliness is to stop the compiler complaining
 552          * that we ignore 'ignore'
 553          */
 554         if (ignore)
 555                 ignore++;
 556
 557         do_manager(container);
 558
 559         exit(0);
 560 }
 561
 562 /* Some stub functions so super-* can link with us */
 563 int child_monitor(int afd, struct mdinfo *sra, struct reshape *reshape,
 564                   struct supertype *st, unsigned long blocks,
 565                   int *fds, unsigned long long *offsets,
 566                   int dests, int *destfd, unsigned long long *destoffsets)
 567 {
 568         return 0;
 569 }
 570
 571 int restore_stripes(int *dest, unsigned long long *offsets,
 572                     int raid_disks, int chunk_size, int level, int layout,
 573                     int source, unsigned long long read_offset,
 574                     unsigned long long start, unsigned long long length,
 575                     char *src_buf)
 576 {
 577         return 1;
 578 }
 579
 580 void abort_reshape(struct mdinfo *sra)
 581 {
 582         return;
 583 }
 584
 585 int save_stripes(int *source, unsigned long long *offsets,
 586                  int raid_disks, int chunk_size, int level, int layout,
 587                  int nwrites, int *dest,
 588                  unsigned long long start, unsigned long long length,
 589                  char *buf)
 590 {
 591         return 0;
 592 }