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
  62 #include        <sched.h>
  63
  64 #include        "mdadm.h"
  65 #include        "mdmon.h"
  66
  67 struct active_array *discard_this;
  68 struct active_array *pending_discard;
  69
  70 int mon_tid, mgr_tid;
  71
  72 int sigterm;
  73
  74 int run_child(void *v)
  75 {
  76         struct supertype *c = v;
  77
  78         do_monitor(c);
  79         return 0;
  80 }
  81
  82 #ifdef __ia64__
  83 int __clone2(int (*fn)(void *),
  84             void *child_stack_base, size_t stack_size,
  85             int flags, void *arg, ...
  86          /* pid_t *pid, struct user_desc *tls, pid_t *ctid */ );
  87 #endif
  88  int clone_monitor(struct supertype *container)
  89 {
  90         static char stack[4096];
  91
  92 #ifdef __ia64__
  93         mon_tid = __clone2(run_child, stack, sizeof(stack),
  94                    CLONE_FS|CLONE_FILES|CLONE_VM|CLONE_SIGHAND|CLONE_THREAD,
  95                    container);
  96 #else
  97         mon_tid = clone(run_child, stack+4096-64,
  98                    CLONE_FS|CLONE_FILES|CLONE_VM|CLONE_SIGHAND|CLONE_THREAD,
  99                    container);
 100 #endif
 101
 102         mgr_tid = syscall(SYS_gettid);
 103
 104         return mon_tid;
 105 }
 106
 107 static struct superswitch *find_metadata_methods(char *vers)
 108 {
 109         if (strcmp(vers, "ddf") == 0)
 110                 return &super_ddf;
 111         if (strcmp(vers, "imsm") == 0)
 112                 return &super_imsm;
 113         return NULL;
 114 }
 115
 116 static int test_pidfile(char *devname)
 117 {
 118         char path[100];
 119         struct stat st;
 120
 121         sprintf(path, "/var/run/mdadm/%s.pid", devname);
 122         return stat(path, &st);
 123 }
 124
 125 int make_pidfile(char *devname, int o_excl)
 126 {
 127         char path[100];
 128         char pid[10];
 129         int fd;
 130         int n;
 131
 132         if (sigterm)
 133                 return -1;
 134
 135         sprintf(path, "/var/run/mdadm/%s.pid", devname);
 136
 137         fd = open(path, O_RDWR|O_CREAT|o_excl, 0600);
 138         if (fd < 0)
 139                 return -errno;
 140         sprintf(pid, "%d\n", getpid());
 141         n = write(fd, pid, strlen(pid));
 142         close(fd);
 143         if (n < 0)
 144                 return -errno;
 145         return 0;
 146 }
 147
 148 int is_container_member(struct mdstat_ent *mdstat, char *container)
 149 {
 150         if (mdstat->metadata_version == NULL ||
 151             strncmp(mdstat->metadata_version, "external:", 9) != 0 ||
 152             !is_subarray(mdstat->metadata_version+9) ||
 153             strncmp(mdstat->metadata_version+10, container, strlen(container)) != 0 ||
 154             mdstat->metadata_version[10+strlen(container)] != '/')
 155                 return 0;
 156
 157         return 1;
 158 }
 159
 160 pid_t devname2mdmon(char *devname)
 161 {
 162         char buf[100];
 163         pid_t pid = -1;
 164         int fd;
 165
 166         sprintf(buf, "/var/run/mdadm/%s.pid", devname);
 167         fd = open(buf, O_RDONLY|O_NOATIME);
 168         if (fd < 0)
 169                 return -1;
 170
 171         if (read(fd, buf, sizeof(buf)) > 0)
 172                 sscanf(buf, "%d\n", &pid);
 173         close(fd);
 174
 175         return pid;
 176 }
 177
 178 static void try_kill_monitor(pid_t pid, char *devname, int sock)
 179 {
 180         char buf[100];
 181         int fd;
 182         int n;
 183         long fl;
 184
 185         /* first rule of survival... don't off yourself */
 186         if (pid == getpid())
 187                 return;
 188
 189         /* kill this process if it is mdmon */
 190         sprintf(buf, "/proc/%lu/cmdline", (unsigned long) pid);
 191         fd = open(buf, O_RDONLY);
 192         if (fd < 0)
 193                 return;
 194
 195         n = read(fd, buf, sizeof(buf)-1);
 196         buf[sizeof(buf)-1] = 0;
 197         close(fd);
 198
 199         if (n < 0 || !strstr(buf, "mdmon"))
 200                 return;
 201
 202         kill(pid, SIGTERM);
 203
 204         /* Wait for monitor to exit by reading from the socket, after
 205          * clearing the non-blocking flag */
 206         fl = fcntl(sock, F_GETFL, 0);
 207         fl &= ~O_NONBLOCK;
 208         fcntl(sock, F_SETFL, fl);
 209         read(sock, buf, 100);
 210 }
 211
 212 void remove_pidfile(char *devname)
 213 {
 214         char buf[100];
 215
 216         if (sigterm)
 217                 return;
 218
 219         sprintf(buf, "/var/run/mdadm/%s.pid", devname);
 220         unlink(buf);
 221         sprintf(buf, "/var/run/mdadm/%s.sock", devname);
 222         unlink(buf);
 223 }
 224
 225 int make_control_sock(char *devname)
 226 {
 227         char path[100];
 228         int sfd;
 229         long fl;
 230         struct sockaddr_un addr;
 231
 232         if (sigterm)
 233                 return -1;
 234
 235         sprintf(path, "/var/run/mdadm/%s.sock", devname);
 236         unlink(path);
 237         sfd = socket(PF_LOCAL, SOCK_STREAM, 0);
 238         if (sfd < 0)
 239                 return -1;
 240
 241         addr.sun_family = PF_LOCAL;
 242         strcpy(addr.sun_path, path);
 243         if (bind(sfd, &addr, sizeof(addr)) < 0) {
 244                 close(sfd);
 245                 return -1;
 246         }
 247         listen(sfd, 10);
 248         fl = fcntl(sfd, F_GETFL, 0);
 249         fl |= O_NONBLOCK;
 250         fcntl(sfd, F_SETFL, fl);
 251         return sfd;
 252 }
 253
 254 int socket_hup_requested;
 255 static void hup(int sig)
 256 {
 257         socket_hup_requested = 1;
 258 }
 259
 260 static void term(int sig)
 261 {
 262         sigterm = 1;
 263 }
 264
 265 static void wake_me(int sig)
 266 {
 267
 268 }
 269
 270 /* if we are debugging and starting mdmon by hand then don't fork */
 271 static int do_fork(void)
 272 {
 273         #ifdef DEBUG
 274         if (check_env("MDADM_NO_MDMON"))
 275                 return 0;
 276         #endif
 277
 278         return 1;
 279 }
 280
 281 void usage(void)
 282 {
 283         fprintf(stderr, "Usage: mdmon /device/name/for/container [target_dir]\n");
 284         exit(2);
 285 }
 286
 287 static int mdmon(char *devname, int devnum, int must_fork, char *switchroot);
 288
 289 int main(int argc, char *argv[])
 290 {
 291         char *container_name = NULL;
 292         char *switchroot = NULL;
 293         int devnum;
 294         char *devname;
 295         int status = 0;
 296
 297         switch (argc) {
 298         case 3:
 299                 switchroot = argv[2];
 300         case 2:
 301                 container_name = argv[1];
 302                 break;
 303         default:
 304                 usage();
 305         }
 306
 307         if (strcmp(container_name, "/proc/mdstat") == 0) {
 308                 struct mdstat_ent *mdstat, *e;
 309
 310                 /* launch an mdmon instance for each container found */
 311                 mdstat = mdstat_read(0, 0);
 312                 for (e = mdstat; e; e = e->next) {
 313                         if (strncmp(e->metadata_version, "external:", 9) == 0 &&
 314                             !is_subarray(&e->metadata_version[9])) {
 315                                 devname = devnum2devname(e->devnum);
 316                                 /* update cmdline so this mdmon instance can be
 317                                  * distinguished from others in a call to ps(1)
 318                                  */
 319                                 if (strlen(devname) <= strlen(container_name)) {
 320                                         memset(container_name, 0, strlen(container_name));
 321                                         sprintf(container_name, "%s", devname);
 322                                 }
 323                                 status |= mdmon(devname, e->devnum, 1,
 324                                                 switchroot);
 325                         }
 326                 }
 327                 free_mdstat(mdstat);
 328
 329                 return status;
 330         } else if (strncmp(container_name, "md", 2) == 0) {
 331                 devnum = devname2devnum(container_name);
 332                 devname = devnum2devname(devnum);
 333                 if (strcmp(container_name, devname) != 0)
 334                         devname = NULL;
 335         } else {
 336                 struct stat st;
 337
 338                 devnum = NoMdDev;
 339                 if (stat(container_name, &st) == 0)
 340                         devnum = stat2devnum(&st);
 341                 if (devnum == NoMdDev)
 342                         devname = NULL;
 343                 else
 344                         devname = devnum2devname(devnum);
 345         }
 346
 347         if (!devname) {
 348                 fprintf(stderr, "mdmon: %s is not a valid md device name\n",
 349                         container_name);
 350                 exit(1);
 351         }
 352         return mdmon(devname, devnum, do_fork(), switchroot);
 353 }
 354
 355 static int mdmon(char *devname, int devnum, int must_fork, char *switchroot)
 356 {
 357         int mdfd;
 358         struct mdinfo *mdi, *di;
 359         struct supertype *container;
 360         sigset_t set;
 361         struct sigaction act;
 362         int pfd[2];
 363         int status;
 364         int ignore;
 365         pid_t victim = -1;
 366         int victim_sock = -1;
 367
 368         dprintf("starting mdmon for %s in %s\n",
 369                 devname, switchroot ? : "/");
 370
 371         /* try to spawn mdmon instances from the target file system */
 372         if (switchroot && strcmp(switchroot, "/") != 0) {
 373                 char path[1024];
 374                 pid_t pid;
 375
 376                 sprintf(path, "%s/sbin/mdmon", switchroot);
 377                 switch (fork()) {
 378                 case 0:
 379                         execl(path, "mdmon", devname, NULL);
 380                         exit(1);
 381                 case -1:
 382                         return 1;
 383                 default:
 384                         pid = wait(&status);
 385                         if (pid > -1 && WIFEXITED(status) &&
 386                             WEXITSTATUS(status) == 0)
 387                                 return 0;
 388                         else
 389                                 return 1;
 390                 }
 391         }
 392
 393         mdfd = open_dev(devnum);
 394         if (mdfd < 0) {
 395                 fprintf(stderr, "mdmon: %s: %s\n", devname,
 396                         strerror(errno));
 397                 return 1;
 398         }
 399         if (md_get_version(mdfd) < 0) {
 400                 fprintf(stderr, "mdmon: %s: Not an md device\n",
 401                         devname);
 402                 return 1;
 403         }
 404
 405         /* Fork, and have the child tell us when they are ready */
 406         if (must_fork) {
 407                 if (pipe(pfd) != 0) {
 408                         fprintf(stderr, "mdmon: failed to create pipe\n");
 409                         return 1;
 410                 }
 411                 switch(fork()) {
 412                 case -1:
 413                         fprintf(stderr, "mdmon: failed to fork: %s\n",
 414                                 strerror(errno));
 415                         return 1;
 416                 case 0: /* child */
 417                         close(pfd[0]);
 418                         break;
 419                 default: /* parent */
 420                         close(pfd[1]);
 421                         if (read(pfd[0], &status, sizeof(status)) != sizeof(status)) {
 422                                 wait(&status);
 423                                 status = WEXITSTATUS(status);
 424                         }
 425                         return status;
 426                 }
 427         } else
 428                 pfd[0] = pfd[1] = -1;
 429
 430         container = calloc(1, sizeof(*container));
 431         container->devnum = devnum;
 432         container->devname = devname;
 433         container->arrays = NULL;
 434         container->subarray[0] = 0;
 435         container->sock = -1;
 436
 437         if (!container->devname) {
 438                 fprintf(stderr, "mdmon: failed to allocate container name string\n");
 439                 exit(3);
 440         }
 441
 442         mdi = sysfs_read(mdfd, container->devnum,
 443                          GET_VERSION|GET_LEVEL|GET_DEVS|SKIP_GONE_DEVS);
 444
 445         if (!mdi) {
 446                 fprintf(stderr, "mdmon: failed to load sysfs info for %s\n",
 447                         container->devname);
 448                 exit(3);
 449         }
 450         if (mdi->array.level != UnSet) {
 451                 fprintf(stderr, "mdmon: %s is not a container - cannot monitor\n",
 452                         devname);
 453                 exit(3);
 454         }
 455         if (mdi->array.major_version != -1 ||
 456             mdi->array.minor_version != -2) {
 457                 fprintf(stderr, "mdmon: %s does not use external metadata - cannot monitor\n",
 458                         devname);
 459                 exit(3);
 460         }
 461
 462         container->ss = find_metadata_methods(mdi->text_version);
 463         if (container->ss == NULL) {
 464                 fprintf(stderr, "mdmon: %s uses unknown metadata: %s\n",
 465                         devname, mdi->text_version);
 466                 exit(3);
 467         }
 468
 469         container->devs = NULL;
 470         for (di = mdi->devs; di; di = di->next) {
 471                 struct mdinfo *cd = malloc(sizeof(*cd));
 472                 *cd = *di;
 473                 cd->next = container->devs;
 474                 container->devs = cd;
 475         }
 476         sysfs_free(mdi);
 477
 478         /* SIGUSR is sent between parent and child.  So both block it
 479          * and enable it only with pselect.
 480          */
 481         sigemptyset(&set);
 482         sigaddset(&set, SIGUSR1);
 483         sigaddset(&set, SIGHUP);
 484         sigaddset(&set, SIGALRM);
 485         sigaddset(&set, SIGTERM);
 486         sigprocmask(SIG_BLOCK, &set, NULL);
 487         act.sa_handler = wake_me;
 488         act.sa_flags = 0;
 489         sigaction(SIGUSR1, &act, NULL);
 490         sigaction(SIGALRM, &act, NULL);
 491         act.sa_handler = hup;
 492         sigaction(SIGHUP, &act, NULL);
 493         act.sa_handler = term;
 494         sigaction(SIGTERM, &act, NULL);
 495         act.sa_handler = SIG_IGN;
 496         sigaction(SIGPIPE, &act, NULL);
 497
 498         if (switchroot) {
 499                 /* we assume we assume that /sys /proc /dev are available in
 500                  * the new root
 501                  */
 502                 victim = devname2mdmon(container->devname);
 503                 victim_sock = connect_monitor(container->devname);
 504                 if (chroot(switchroot) != 0) {
 505                         fprintf(stderr, "mdmon: failed to chroot to '%s': %s\n",
 506                                 switchroot, strerror(errno));
 507                         exit(4);
 508                 }
 509         }
 510
 511         ignore = chdir("/");
 512         if (victim < 0 && test_pidfile(container->devname) == 0) {
 513                 if (ping_monitor(container->devname) == 0) {
 514                         fprintf(stderr, "mdmon: %s already managed\n",
 515                                 container->devname);
 516                         exit(3);
 517                 } else if (victim < 0)
 518                         victim = devname2mdmon(container->devname);
 519         }
 520         if (container->ss->load_super(container, mdfd, devname)) {
 521                 fprintf(stderr, "mdmon: Cannot load metadata for %s\n",
 522                         devname);
 523                 exit(3);
 524         }
 525         close(mdfd);
 526
 527         /* Ok, this is close enough.  We can say goodbye to our parent now.
 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         setsid();
 536         close(0);
 537         open("/dev/null", O_RDWR);
 538         close(1);
 539         ignore = dup(0);
 540 #ifndef DEBUG
 541         close(2);
 542         ignore = dup(0);
 543 #endif
 544
 545         mlockall(MCL_CURRENT | MCL_FUTURE);
 546
 547         if (clone_monitor(container) < 0) {
 548                 fprintf(stderr, "mdmon: failed to start monitor process: %s\n",
 549                         strerror(errno));
 550                 exit(2);
 551         }
 552
 553         if (victim > -1) {
 554                 try_kill_monitor(victim, container->devname, victim_sock);
 555                 close(victim_sock);
 556         }
 557         do_manager(container);
 558
 559         exit(0);
 560 }