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                 return;
 189
 190         kill(pid, SIGTERM);
 191
 192         if (sock < 0)
 193                 return;
 194
 195         /* Wait for monitor to exit by reading from the socket, after
 196          * clearing the non-blocking flag */
 197         fl = fcntl(sock, F_GETFL, 0);
 198         fl &= ~O_NONBLOCK;
 199         fcntl(sock, F_SETFL, fl);
 200         n = read(sock, buf, 100);
 201         /* Ignore result, it is just the wait that
 202          * matters
 203          */
 204 }
 205
 206 void remove_pidfile(char *devname)
 207 {
 208         char buf[100];
 209
 210         sprintf(buf, "%s/%s.pid", MDMON_DIR, devname);
 211         unlink(buf);
 212         sprintf(buf, "%s/%s.sock", MDMON_DIR, devname);
 213         unlink(buf);
 214 }
 215
 216 static int make_control_sock(char *devname)
 217 {
 218         char path[100];
 219         int sfd;
 220         long fl;
 221         struct sockaddr_un addr;
 222
 223         if (sigterm)
 224                 return -1;
 225
 226         sprintf(path, "%s/%s.sock", MDMON_DIR, devname);
 227         unlink(path);
 228         sfd = socket(PF_LOCAL, SOCK_STREAM, 0);
 229         if (sfd < 0)
 230                 return -1;
 231
 232         addr.sun_family = PF_LOCAL;
 233         strcpy(addr.sun_path, path);
 234         if (bind(sfd, &addr, sizeof(addr)) < 0) {
 235                 close(sfd);
 236                 return -1;
 237         }
 238         listen(sfd, 10);
 239         fl = fcntl(sfd, F_GETFL, 0);
 240         fl |= O_NONBLOCK;
 241         fcntl(sfd, F_SETFL, fl);
 242         return sfd;
 243 }
 244
 245 static void term(int sig)
 246 {
 247         sigterm = 1;
 248 }
 249
 250 static void wake_me(int sig)
 251 {
 252
 253 }
 254
 255 /* if we are debugging and starting mdmon by hand then don't fork */
 256 static int do_fork(void)
 257 {
 258         #ifdef DEBUG
 259         if (check_env("MDADM_NO_MDMON"))
 260                 return 0;
 261         #endif
 262
 263         return 1;
 264 }
 265
 266 void usage(void)
 267 {
 268         fprintf(stderr,
 269 "Usage: mdmon [options] CONTAINER\n"
 270 "\n"
 271 "Options are:\n"
 272 "  --help        -h   : This message\n"
 273 "  --all              : All devices\n"
 274 "  --takeover    -t   : Takeover container\n"
 275 "  --offroot          : Set first character of argv[0] to @ to indicate the\n"
 276 "                       application was launched from initrd/initramfs and\n"
 277 "                       should not be shutdown by systemd as part of the\n"
 278 "                       regular shutdown process.\n"
 279 );
 280         exit(2);
 281 }
 282
 283 static int mdmon(char *devname, int devnum, int must_fork, int takeover);
 284
 285 int main(int argc, char *argv[])
 286 {
 287         char *container_name = NULL;
 288         int devnum;
 289         char *devname;
 290         int status = 0;
 291         int opt;
 292         int all = 0;
 293         int takeover = 0;
 294         static struct option options[] = {
 295                 {"all", 0, NULL, 'a'},
 296                 {"takeover", 0, NULL, 't'},
 297                 {"help", 0, NULL, 'h'},
 298                 {"offroot", 0, NULL, OffRootOpt},
 299                 {NULL, 0, NULL, 0}
 300         };
 301
 302         while ((opt = getopt_long(argc, argv, "th", options, NULL)) != -1) {
 303                 switch (opt) {
 304                 case 'a':
 305                         container_name = argv[optind-1];
 306                         all = 1;
 307                         break;
 308                 case 't':
 309                         container_name = optarg;
 310                         takeover = 1;
 311                         break;
 312                 case OffRootOpt:
 313                         argv[0][0] = '@';
 314                         break;
 315                 case 'h':
 316                 default:
 317                         usage();
 318                         break;
 319                 }
 320         }
 321
 322         if (all == 0 && container_name == NULL) {
 323                 if (argv[optind])
 324                         container_name = argv[optind];
 325         }
 326
 327         if (container_name == NULL)
 328                 usage();
 329
 330         if (argc - optind > 1)
 331                 usage();
 332
 333         if (strcmp(container_name, "/proc/mdstat") == 0)
 334                 all = 1;
 335
 336         if (all) {
 337                 struct mdstat_ent *mdstat, *e;
 338                 int container_len = strlen(container_name);
 339
 340                 /* launch an mdmon instance for each container found */
 341                 mdstat = mdstat_read(0, 0);
 342                 for (e = mdstat; e; e = e->next) {
 343                         if (e->metadata_version &&
 344                             strncmp(e->metadata_version, "external:", 9) == 0 &&
 345                             !is_subarray(&e->metadata_version[9])) {
 346                                 devname = devnum2devname(e->devnum);
 347                                 /* update cmdline so this mdmon instance can be
 348                                  * distinguished from others in a call to ps(1)
 349                                  */
 350                                 if (strlen(devname) <= (unsigned)container_len) {
 351                                         memset(container_name, 0, container_len);
 352                                         sprintf(container_name, "%s", devname);
 353                                 }
 354                                 status |= mdmon(devname, e->devnum, 1,
 355                                                 takeover);
 356                         }
 357                 }
 358                 free_mdstat(mdstat);
 359
 360                 return status;
 361         } else if (strncmp(container_name, "md", 2) == 0) {
 362                 devnum = devname2devnum(container_name);
 363                 devname = devnum2devname(devnum);
 364                 if (strcmp(container_name, devname) != 0)
 365                         devname = NULL;
 366         } else {
 367                 struct stat st;
 368
 369                 devnum = NoMdDev;
 370                 if (stat(container_name, &st) == 0)
 371                         devnum = stat2devnum(&st);
 372                 if (devnum == NoMdDev)
 373                         devname = NULL;
 374                 else
 375                         devname = devnum2devname(devnum);
 376         }
 377
 378         if (!devname) {
 379                 fprintf(stderr, "mdmon: %s is not a valid md device name\n",
 380                         container_name);
 381                 exit(1);
 382         }
 383         return mdmon(devname, devnum, do_fork(), takeover);
 384 }
 385
 386 static int mdmon(char *devname, int devnum, int must_fork, int takeover)
 387 {
 388         int mdfd;
 389         struct mdinfo *mdi, *di;
 390         struct supertype *container;
 391         sigset_t set;
 392         struct sigaction act;
 393         int pfd[2];
 394         int status;
 395         int ignore;
 396         pid_t victim = -1;
 397         int victim_sock = -1;
 398
 399         dprintf("starting mdmon for %s\n", devname);
 400
 401         mdfd = open_dev(devnum);
 402         if (mdfd < 0) {
 403                 fprintf(stderr, "mdmon: %s: %s\n", devname,
 404                         strerror(errno));
 405                 return 1;
 406         }
 407         if (md_get_version(mdfd) < 0) {
 408                 fprintf(stderr, "mdmon: %s: Not an md device\n",
 409                         devname);
 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                         fprintf(stderr, "mdmon: failed to create pipe\n");
 417                         return 1;
 418                 }
 419                 switch(fork()) {
 420                 case -1:
 421                         fprintf(stderr, "mdmon: failed to fork: %s\n",
 422                                 strerror(errno));
 423                         return 1;
 424                 case 0: /* child */
 425                         close(pfd[0]);
 426                         break;
 427                 default: /* parent */
 428                         close(pfd[1]);
 429                         if (read(pfd[0], &status, sizeof(status)) != sizeof(status)) {
 430                                 wait(&status);
 431                                 status = WEXITSTATUS(status);
 432                         }
 433                         return status;
 434                 }
 435         } else
 436                 pfd[0] = pfd[1] = -1;
 437
 438         container = calloc(1, sizeof(*container));
 439         container->devnum = devnum;
 440         container->devname = devname;
 441         container->arrays = NULL;
 442         container->sock = -1;
 443
 444         if (!container->devname) {
 445                 fprintf(stderr, "mdmon: failed to allocate container name string\n");
 446                 exit(3);
 447         }
 448
 449         mdi = sysfs_read(mdfd, container->devnum, GET_VERSION|GET_LEVEL|GET_DEVS);
 450
 451         if (!mdi) {
 452                 fprintf(stderr, "mdmon: failed to load sysfs info for %s\n",
 453                         container->devname);
 454                 exit(3);
 455         }
 456         if (mdi->array.level != UnSet) {
 457                 fprintf(stderr, "mdmon: %s is not a container - cannot monitor\n",
 458                         devname);
 459                 exit(3);
 460         }
 461         if (mdi->array.major_version != -1 ||
 462             mdi->array.minor_version != -2) {
 463                 fprintf(stderr, "mdmon: %s does not use external metadata - cannot monitor\n",
 464                         devname);
 465                 exit(3);
 466         }
 467
 468         container->ss = version_to_superswitch(mdi->text_version);
 469         if (container->ss == NULL) {
 470                 fprintf(stderr, "mdmon: %s uses unsupported metadata: %s\n",
 471                         devname, mdi->text_version);
 472                 exit(3);
 473         }
 474
 475         container->devs = NULL;
 476         for (di = mdi->devs; di; di = di->next) {
 477                 struct mdinfo *cd = malloc(sizeof(*cd));
 478                 *cd = *di;
 479                 cd->next = container->devs;
 480                 container->devs = cd;
 481         }
 482         sysfs_free(mdi);
 483
 484         /* SIGUSR is sent between parent and child.  So both block it
 485          * and enable it only with pselect.
 486          */
 487         sigemptyset(&set);
 488         sigaddset(&set, SIGUSR1);
 489         sigaddset(&set, SIGTERM);
 490         sigprocmask(SIG_BLOCK, &set, NULL);
 491         act.sa_handler = wake_me;
 492         act.sa_flags = 0;
 493         sigaction(SIGUSR1, &act, NULL);
 494         act.sa_handler = term;
 495         sigaction(SIGTERM, &act, NULL);
 496         act.sa_handler = SIG_IGN;
 497         sigaction(SIGPIPE, &act, NULL);
 498
 499         victim = mdmon_pid(container->devnum);
 500         if (victim >= 0)
 501                 victim_sock = connect_monitor(container->devname);
 502
 503         ignore = chdir("/");
 504         if (!takeover && victim > 0 && victim_sock >= 0) {
 505                 if (fping_monitor(victim_sock) == 0) {
 506                         fprintf(stderr, "mdmon: %s already managed\n",
 507                                 container->devname);
 508                         exit(3);
 509                 }
 510                 close(victim_sock);
 511                 victim_sock = -1;
 512         }
 513         if (container->ss->load_container(container, mdfd, devname)) {
 514                 fprintf(stderr, "mdmon: Cannot load metadata for %s\n",
 515                         devname);
 516                 exit(3);
 517         }
 518         close(mdfd);
 519
 520         /* Ok, this is close enough.  We can say goodbye to our parent now.
 521          */
 522         if (victim > 0)
 523                 remove_pidfile(devname);
 524         if (make_pidfile(devname) < 0) {
 525                 exit(3);
 526         }
 527         container->sock = make_control_sock(devname);
 528
 529         status = 0;
 530         if (write(pfd[1], &status, sizeof(status)) < 0)
 531                 fprintf(stderr, "mdmon: failed to notify our parent: %d\n",
 532                         getppid());
 533         close(pfd[1]);
 534
 535         mlockall(MCL_CURRENT | MCL_FUTURE);
 536
 537         if (clone_monitor(container) < 0) {
 538                 fprintf(stderr, "mdmon: failed to start monitor process: %s\n",
 539                         strerror(errno));
 540                 exit(2);
 541         }
 542
 543         if (victim > 0) {
 544                 try_kill_monitor(victim, container->devname, victim_sock);
 545                 if (victim_sock >= 0)
 546                         close(victim_sock);
 547         }
 548
 549         setsid();
 550         close(0);
 551         open("/dev/null", O_RDWR);
 552         close(1);
 553         ignore = dup(0);
 554 #ifndef DEBUG
 555         close(2);
 556         ignore = dup(0);
 557 #endif
 558
 559         /* This silliness is to stop the compiler complaining
 560          * that we ignore 'ignore'
 561          */
 562         if (ignore)
 563                 ignore++;
 564
 565         do_manager(container);
 566
 567         exit(0);
 568 }
 569
 570 /* Some stub functions so super-* can link with us */
 571 int child_monitor(int afd, struct mdinfo *sra, struct reshape *reshape,
 572                   struct supertype *st, unsigned long blocks,
 573                   int *fds, unsigned long long *offsets,
 574                   int dests, int *destfd, unsigned long long *destoffsets)
 575 {
 576         return 0;
 577 }
 578
 579 int restore_stripes(int *dest, unsigned long long *offsets,
 580                     int raid_disks, int chunk_size, int level, int layout,
 581                     int source, unsigned long long read_offset,
 582                     unsigned long long start, unsigned long long length,
 583                     char *src_buf)
 584 {
 585         return 1;
 586 }
 587
 588 void abort_reshape(struct mdinfo *sra)
 589 {
 590         return;
 591 }
 592
 593 int save_stripes(int *source, unsigned long long *offsets,
 594                  int raid_disks, int chunk_size, int level, int layout,
 595                  int nwrites, int *dest,
 596                  unsigned long long start, unsigned long long length,
 597                  char *buf)
 598 {
 599         return 0;
 600 }