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
  69 char const Name[] = "mdmon";
  70
  71 struct active_array *discard_this;
  72 struct active_array *pending_discard;
  73
  74 int mon_tid, mgr_tid;
  75
  76 int sigterm;
  77
  78 #ifdef USE_PTHREADS
  79 static void *run_child(void *v)
  80 {
  81         struct supertype *c = v;
  82
  83         mon_tid = syscall(SYS_gettid);
  84         do_monitor(c);
  85         return 0;
  86 }
  87
  88 static int clone_monitor(struct supertype *container)
  89 {
  90         pthread_attr_t attr;
  91         pthread_t thread;
  92         int rc;
  93
  94         mon_tid = -1;
  95         pthread_attr_init(&attr);
  96         pthread_attr_setstacksize(&attr, 4096);
  97         pthread_attr_setdetachstate(&attr, PTHREAD_CREATE_DETACHED);
  98         rc = pthread_create(&thread, &attr, run_child, container);
  99         if (rc)
 100                 return rc;
 101         while (mon_tid == -1)
 102                 usleep(10);
 103         pthread_attr_destroy(&attr);
 104
 105         mgr_tid = syscall(SYS_gettid);
 106
 107         return mon_tid;
 108 }
 109 #else /* USE_PTHREADS */
 110 static int run_child(void *v)
 111 {
 112         struct supertype *c = v;
 113
 114         do_monitor(c);
 115         return 0;
 116 }
 117
 118 #ifdef __ia64__
 119 int __clone2(int (*fn)(void *),
 120             void *child_stack_base, size_t stack_size,
 121             int flags, void *arg, ...
 122          /* pid_t *pid, struct user_desc *tls, pid_t *ctid */ );
 123 #endif
 124 static int clone_monitor(struct supertype *container)
 125 {
 126         static char stack[4096];
 127
 128 #ifdef __ia64__
 129         mon_tid = __clone2(run_child, stack, sizeof(stack),
 130                    CLONE_FS|CLONE_FILES|CLONE_VM|CLONE_SIGHAND|CLONE_THREAD,
 131                    container);
 132 #else
 133         mon_tid = clone(run_child, stack+4096-64,
 134                    CLONE_FS|CLONE_FILES|CLONE_VM|CLONE_SIGHAND|CLONE_THREAD,
 135                    container);
 136 #endif
 137
 138         mgr_tid = syscall(SYS_gettid);
 139
 140         return mon_tid;
 141 }
 142 #endif /* USE_PTHREADS */
 143
 144 static int make_pidfile(char *devname)
 145 {
 146         char path[100];
 147         char pid[10];
 148         int fd;
 149         int n;
 150
 151         if (mkdir(MDMON_DIR, 0755) < 0 &&
 152             errno != EEXIST)
 153                 return -errno;
 154         sprintf(path, "%s/%s.pid", MDMON_DIR, devname);
 155
 156         fd = open(path, O_RDWR|O_CREAT|O_EXCL, 0600);
 157         if (fd < 0)
 158                 return -errno;
 159         sprintf(pid, "%d\n", getpid());
 160         n = write(fd, pid, strlen(pid));
 161         close(fd);
 162         if (n < 0)
 163                 return -errno;
 164         return 0;
 165 }
 166
 167 static void try_kill_monitor(pid_t pid, char *devname, int sock)
 168 {
 169         char buf[100];
 170         int fd;
 171         int n;
 172         long fl;
 173         int rv;
 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
 205         /* If there is I/O going on it might took some time to get to
 206          * clean state. Wait for monitor to exit fully to avoid races.
 207          * Ping it with SIGUSR1 in case that it is sleeping  */
 208         for (n = 0; n < 25; n++) {
 209                 rv = kill(pid, SIGUSR1);
 210                 if (rv < 0)
 211                         break;
 212                 usleep(200000);
 213         }
 214 }
 215
 216 void remove_pidfile(char *devname)
 217 {
 218         char buf[100];
 219
 220         sprintf(buf, "%s/%s.pid", MDMON_DIR, devname);
 221         unlink(buf);
 222         sprintf(buf, "%s/%s.sock", MDMON_DIR, devname);
 223         unlink(buf);
 224 }
 225
 226 static int make_control_sock(char *devname)
 227 {
 228         char path[100];
 229         int sfd;
 230         long fl;
 231         struct sockaddr_un addr;
 232
 233         if (sigterm)
 234                 return -1;
 235
 236         sprintf(path, "%s/%s.sock", MDMON_DIR, devname);
 237         unlink(path);
 238         sfd = socket(PF_LOCAL, SOCK_STREAM, 0);
 239         if (sfd < 0)
 240                 return -1;
 241
 242         addr.sun_family = PF_LOCAL;
 243         strcpy(addr.sun_path, path);
 244         umask(077); /* ensure no world write access */
 245         if (bind(sfd, (struct sockaddr*)&addr, sizeof(addr)) < 0) {
 246                 close(sfd);
 247                 return -1;
 248         }
 249         listen(sfd, 10);
 250         fl = fcntl(sfd, F_GETFL, 0);
 251         fl |= O_NONBLOCK;
 252         fcntl(sfd, F_SETFL, fl);
 253         return sfd;
 254 }
 255
 256 static void term(int sig)
 257 {
 258         sigterm = 1;
 259 }
 260
 261 static void wake_me(int sig)
 262 {
 263
 264 }
 265
 266 /* if we are debugging and starting mdmon by hand then don't fork */
 267 static int do_fork(void)
 268 {
 269         #ifdef DEBUG
 270         if (check_env("MDADM_NO_MDMON"))
 271                 return 0;
 272         #endif
 273
 274         return 1;
 275 }
 276
 277 void usage(void)
 278 {
 279         fprintf(stderr,
 280 "Usage: mdmon [options] CONTAINER\n"
 281 "\n"
 282 "Options are:\n"
 283 "  --help        -h   : This message\n"
 284 "  --all         -a   : All devices\n"
 285 "  --foreground  -F   : Run in foreground (do not fork)\n"
 286 "  --takeover    -t   : Takeover container\n"
 287 );
 288         exit(2);
 289 }
 290
 291 static int mdmon(char *devnm, int must_fork, int takeover);
 292
 293 int main(int argc, char *argv[])
 294 {
 295         char *container_name = NULL;
 296         char *devnm = NULL;
 297         int status = 0;
 298         int opt;
 299         int all = 0;
 300         int takeover = 0;
 301         int dofork = 1;
 302         static struct option options[] = {
 303                 {"all", 0, NULL, 'a'},
 304                 {"takeover", 0, NULL, 't'},
 305                 {"help", 0, NULL, 'h'},
 306                 {"offroot", 0, NULL, OffRootOpt},
 307                 {"foreground", 0, NULL, 'F'},
 308                 {NULL, 0, NULL, 0}
 309         };
 310
 311         if (in_initrd()) {
 312                 /*
 313                  * set first char of argv[0] to @. This is used by
 314                  * systemd to signal that the task was launched from
 315                  * initrd/initramfs and should be preserved during shutdown
 316                  */
 317                 argv[0][0] = '@';
 318         }
 319
 320         while ((opt = getopt_long(argc, argv, "thaF", options, NULL)) != -1) {
 321                 switch (opt) {
 322                 case 'a':
 323                         container_name = argv[optind-1];
 324                         all = 1;
 325                         break;
 326                 case 't':
 327                         takeover = 1;
 328                         break;
 329                 case 'F':
 330                         dofork = 0;
 331                         break;
 332                 case OffRootOpt:
 333                         argv[0][0] = '@';
 334                         break;
 335                 case 'h':
 336                 default:
 337                         usage();
 338                         break;
 339                 }
 340         }
 341
 342         if (all == 0 && container_name == NULL) {
 343                 if (argv[optind])
 344                         container_name = argv[optind];
 345         }
 346
 347         if (container_name == NULL)
 348                 usage();
 349
 350         if (argc - optind > 1)
 351                 usage();
 352
 353         if (strcmp(container_name, "/proc/mdstat") == 0)
 354                 all = 1;
 355
 356         if (all) {
 357                 struct mdstat_ent *mdstat, *e;
 358                 int container_len = strlen(container_name);
 359
 360                 /* launch an mdmon instance for each container found */
 361                 mdstat = mdstat_read(0, 0);
 362                 for (e = mdstat; e; e = e->next) {
 363                         if (e->metadata_version &&
 364                             strncmp(e->metadata_version, "external:", 9) == 0 &&
 365                             !is_subarray(&e->metadata_version[9])) {
 366                                 /* update cmdline so this mdmon instance can be
 367                                  * distinguished from others in a call to ps(1)
 368                                  */
 369                                 if (strlen(e->devnm) <= (unsigned)container_len) {
 370                                         memset(container_name, 0, container_len);
 371                                         sprintf(container_name, "%s", e->devnm);
 372                                 }
 373                                 status |= mdmon(e->devnm, 1, takeover);
 374                         }
 375                 }
 376                 free_mdstat(mdstat);
 377
 378                 return status;
 379         } else if (strncmp(container_name, "md", 2) == 0) {
 380                 int id = devnm2devid(container_name);
 381                 if (id)
 382                         devnm = container_name;
 383         } else {
 384                 struct stat st;
 385
 386                 if (stat(container_name, &st) == 0)
 387                         devnm = xstrdup(stat2devnm(&st));
 388         }
 389
 390         if (!devnm) {
 391                 pr_err("%s is not a valid md device name\n",
 392                         container_name);
 393                 exit(1);
 394         }
 395         return mdmon(devnm, dofork && do_fork(), takeover);
 396 }
 397
 398 static int mdmon(char *devnm, int must_fork, int takeover)
 399 {
 400         int mdfd;
 401         struct mdinfo *mdi, *di;
 402         struct supertype *container;
 403         sigset_t set;
 404         struct sigaction act;
 405         int pfd[2];
 406         int status;
 407         int ignore;
 408         pid_t victim = -1;
 409         int victim_sock = -1;
 410
 411         dprintf("starting mdmon for %s\n", devnm);
 412
 413         mdfd = open_dev(devnm);
 414         if (mdfd < 0) {
 415                 pr_err("%s: %s\n", devnm, strerror(errno));
 416                 return 1;
 417         }
 418
 419         /* Fork, and have the child tell us when they are ready */
 420         if (must_fork) {
 421                 if (pipe(pfd) != 0) {
 422                         pr_err("failed to create pipe\n");
 423                         return 1;
 424                 }
 425                 switch(fork()) {
 426                 case -1:
 427                         pr_err("failed to fork: %s\n", strerror(errno));
 428                         return 1;
 429                 case 0: /* child */
 430                         close(pfd[0]);
 431                         break;
 432                 default: /* parent */
 433                         close(pfd[1]);
 434                         if (read(pfd[0], &status, sizeof(status)) != sizeof(status)) {
 435                                 wait(&status);
 436                                 status = WEXITSTATUS(status);
 437                         }
 438                         close(pfd[0]);
 439                         return status;
 440                 }
 441         } else
 442                 pfd[0] = pfd[1] = -1;
 443
 444         container = xcalloc(1, sizeof(*container));
 445         strcpy(container->devnm, devnm);
 446         container->arrays = NULL;
 447         container->sock = -1;
 448
 449         mdi = sysfs_read(mdfd, container->devnm, GET_VERSION|GET_LEVEL|GET_DEVS);
 450
 451         if (!mdi) {
 452                 pr_err("failed to load sysfs info for %s\n", container->devnm);
 453                 exit(3);
 454         }
 455         if (mdi->array.level != UnSet) {
 456                 pr_err("%s is not a container - cannot monitor\n", devnm);
 457                 exit(3);
 458         }
 459         if (mdi->array.major_version != -1 ||
 460             mdi->array.minor_version != -2) {
 461                 pr_err("%s does not use external metadata - cannot monitor\n",
 462                         devnm);
 463                 exit(3);
 464         }
 465
 466         container->ss = version_to_superswitch(mdi->text_version);
 467         if (container->ss == NULL) {
 468                 pr_err("%s uses unsupported metadata: %s\n",
 469                         devnm, mdi->text_version);
 470                 exit(3);
 471         }
 472
 473         container->devs = NULL;
 474         for (di = mdi->devs; di; di = di->next) {
 475                 struct mdinfo *cd = xmalloc(sizeof(*cd));
 476                 *cd = *di;
 477                 cd->next = container->devs;
 478                 container->devs = cd;
 479         }
 480         sysfs_free(mdi);
 481
 482         /* SIGUSR is sent between parent and child.  So both block it
 483          * and enable it only with pselect.
 484          */
 485         sigemptyset(&set);
 486         sigaddset(&set, SIGUSR1);
 487         sigaddset(&set, SIGTERM);
 488         sigprocmask(SIG_BLOCK, &set, NULL);
 489         act.sa_handler = wake_me;
 490         act.sa_flags = 0;
 491         sigaction(SIGUSR1, &act, NULL);
 492         act.sa_handler = term;
 493         sigaction(SIGTERM, &act, NULL);
 494         act.sa_handler = SIG_IGN;
 495         sigaction(SIGPIPE, &act, NULL);
 496
 497         victim = mdmon_pid(container->devnm);
 498         if (victim >= 0)
 499                 victim_sock = connect_monitor(container->devnm);
 500
 501         ignore = chdir("/");
 502         if (!takeover && victim > 0 && victim_sock >= 0) {
 503                 if (fping_monitor(victim_sock) == 0) {
 504                         pr_err("%s already managed\n", container->devnm);
 505                         exit(3);
 506                 }
 507                 close(victim_sock);
 508                 victim_sock = -1;
 509         }
 510         if (container->ss->load_container(container, mdfd, devnm)) {
 511                 pr_err("Cannot load metadata for %s\n", devnm);
 512                 exit(3);
 513         }
 514         close(mdfd);
 515
 516         /* Ok, this is close enough.  We can say goodbye to our parent now.
 517          */
 518         if (victim > 0)
 519                 remove_pidfile(devnm);
 520         if (make_pidfile(devnm) < 0) {
 521                 exit(3);
 522         }
 523         container->sock = make_control_sock(devnm);
 524
 525         status = 0;
 526         if (pfd[1] >= 0) {
 527                 if (write(pfd[1], &status, sizeof(status)) < 0)
 528                         pr_err("failed to notify our parent: %d\n",
 529                                getppid());
 530                 close(pfd[1]);
 531         }
 532
 533         mlockall(MCL_CURRENT | MCL_FUTURE);
 534
 535         if (clone_monitor(container) < 0) {
 536                 pr_err("failed to start monitor process: %s\n",
 537                         strerror(errno));
 538                 exit(2);
 539         }
 540
 541         if (victim > 0) {
 542                 try_kill_monitor(victim, container->devnm, victim_sock);
 543                 if (victim_sock >= 0)
 544                         close(victim_sock);
 545         }
 546
 547         setsid();
 548         manage_fork_fds(0);
 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 };