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         int rv;
 175
 176         /* first rule of survival... don't off yourself */
 177         if (pid == getpid())
 178                 return;
 179
 180         /* kill this process if it is mdmon */
 181         sprintf(buf, "/proc/%lu/cmdline", (unsigned long) pid);
 182         fd = open(buf, O_RDONLY);
 183         if (fd < 0)
 184                 return;
 185
 186         n = read(fd, buf, sizeof(buf)-1);
 187         buf[sizeof(buf)-1] = 0;
 188         close(fd);
 189
 190         if (n < 0 || !(strstr(buf, "mdmon") ||
 191                        strstr(buf, "@dmon")))
 192                 return;
 193
 194         kill(pid, SIGTERM);
 195
 196         if (sock < 0)
 197                 return;
 198
 199         /* Wait for monitor to exit by reading from the socket, after
 200          * clearing the non-blocking flag */
 201         fl = fcntl(sock, F_GETFL, 0);
 202         fl &= ~O_NONBLOCK;
 203         fcntl(sock, F_SETFL, fl);
 204         n = read(sock, buf, 100);
 205
 206         /* If there is I/O going on it might took some time to get to
 207          * clean state. Wait for monitor to exit fully to avoid races.
 208          * Ping it with SIGUSR1 in case that it is sleeping  */
 209         for (n = 0; n < 25; n++) {
 210                 rv = kill(pid, SIGUSR1);
 211                 if (rv < 0)
 212                         break;
 213                 usleep(200000);
 214         }
 215 }
 216
 217 void remove_pidfile(char *devname)
 218 {
 219         char buf[100];
 220
 221         sprintf(buf, "%s/%s.pid", MDMON_DIR, devname);
 222         unlink(buf);
 223         sprintf(buf, "%s/%s.sock", MDMON_DIR, devname);
 224         unlink(buf);
 225 }
 226
 227 static int make_control_sock(char *devname)
 228 {
 229         char path[100];
 230         int sfd;
 231         long fl;
 232         struct sockaddr_un addr;
 233
 234         if (sigterm)
 235                 return -1;
 236
 237         sprintf(path, "%s/%s.sock", MDMON_DIR, devname);
 238         unlink(path);
 239         sfd = socket(PF_LOCAL, SOCK_STREAM, 0);
 240         if (sfd < 0)
 241                 return -1;
 242
 243         addr.sun_family = PF_LOCAL;
 244         strcpy(addr.sun_path, path);
 245         umask(077); /* ensure no world write access */
 246         if (bind(sfd, (struct sockaddr*)&addr, sizeof(addr)) < 0) {
 247                 close(sfd);
 248                 return -1;
 249         }
 250         listen(sfd, 10);
 251         fl = fcntl(sfd, F_GETFL, 0);
 252         fl |= O_NONBLOCK;
 253         fcntl(sfd, F_SETFL, fl);
 254         return sfd;
 255 }
 256
 257 static void term(int sig)
 258 {
 259         sigterm = 1;
 260 }
 261
 262 static void wake_me(int sig)
 263 {
 264
 265 }
 266
 267 /* if we are debugging and starting mdmon by hand then don't fork */
 268 static int do_fork(void)
 269 {
 270         #ifdef DEBUG
 271         if (check_env("MDADM_NO_MDMON"))
 272                 return 0;
 273         #endif
 274
 275         return 1;
 276 }
 277
 278 void usage(void)
 279 {
 280         fprintf(stderr,
 281 "Usage: mdmon [options] CONTAINER\n"
 282 "\n"
 283 "Options are:\n"
 284 "  --help        -h   : This message\n"
 285 "  --all         -a   : All devices\n"
 286 "  --foreground  -F   : Run in foreground (do not fork)\n"
 287 "  --takeover    -t   : Takeover container\n"
 288 );
 289         exit(2);
 290 }
 291
 292 static int mdmon(char *devnm, int must_fork, int takeover);
 293
 294 int main(int argc, char *argv[])
 295 {
 296         char *container_name = NULL;
 297         char *devnm = NULL;
 298         int status = 0;
 299         int opt;
 300         int all = 0;
 301         int takeover = 0;
 302         int dofork = 1;
 303         static struct option options[] = {
 304                 {"all", 0, NULL, 'a'},
 305                 {"takeover", 0, NULL, 't'},
 306                 {"help", 0, NULL, 'h'},
 307                 {"offroot", 0, NULL, OffRootOpt},
 308                 {"foreground", 0, NULL, 'F'},
 309                 {NULL, 0, NULL, 0}
 310         };
 311
 312         if (in_initrd()) {
 313                 /*
 314                  * set first char of argv[0] to @. This is used by
 315                  * systemd to signal that the task was launched from
 316                  * initrd/initramfs and should be preserved during shutdown
 317                  */
 318                 argv[0][0] = '@';
 319         }
 320
 321         while ((opt = getopt_long(argc, argv, "thaF", options, NULL)) != -1) {
 322                 switch (opt) {
 323                 case 'a':
 324                         container_name = argv[optind-1];
 325                         all = 1;
 326                         break;
 327                 case 't':
 328                         takeover = 1;
 329                         break;
 330                 case 'F':
 331                         dofork = 0;
 332                         break;
 333                 case OffRootOpt:
 334                         argv[0][0] = '@';
 335                         break;
 336                 case 'h':
 337                 default:
 338                         usage();
 339                         break;
 340                 }
 341         }
 342
 343         if (all == 0 && container_name == NULL) {
 344                 if (argv[optind])
 345                         container_name = argv[optind];
 346         }
 347
 348         if (container_name == NULL)
 349                 usage();
 350
 351         if (argc - optind > 1)
 352                 usage();
 353
 354         if (strcmp(container_name, "/proc/mdstat") == 0)
 355                 all = 1;
 356
 357         if (all) {
 358                 struct mdstat_ent *mdstat, *e;
 359                 int container_len = strlen(container_name);
 360
 361                 /* launch an mdmon instance for each container found */
 362                 mdstat = mdstat_read(0, 0);
 363                 for (e = mdstat; e; e = e->next) {
 364                         if (e->metadata_version &&
 365                             strncmp(e->metadata_version, "external:", 9) == 0 &&
 366                             !is_subarray(&e->metadata_version[9])) {
 367                                 /* update cmdline so this mdmon instance can be
 368                                  * distinguished from others in a call to ps(1)
 369                                  */
 370                                 if (strlen(e->devnm) <= (unsigned)container_len) {
 371                                         memset(container_name, 0, container_len);
 372                                         sprintf(container_name, "%s", e->devnm);
 373                                 }
 374                                 status |= mdmon(e->devnm, 1, takeover);
 375                         }
 376                 }
 377                 free_mdstat(mdstat);
 378
 379                 return status;
 380         } else if (strncmp(container_name, "md", 2) == 0) {
 381                 int id = devnm2devid(container_name);
 382                 if (id)
 383                         devnm = container_name;
 384         } else {
 385                 struct stat st;
 386
 387                 if (stat(container_name, &st) == 0)
 388                         devnm = xstrdup(stat2devnm(&st));
 389         }
 390
 391         if (!devnm) {
 392                 pr_err("%s is not a valid md device name\n",
 393                         container_name);
 394                 exit(1);
 395         }
 396         return mdmon(devnm, dofork && do_fork(), takeover);
 397 }
 398
 399 static int mdmon(char *devnm, int must_fork, int takeover)
 400 {
 401         int mdfd;
 402         struct mdinfo *mdi, *di;
 403         struct supertype *container;
 404         sigset_t set;
 405         struct sigaction act;
 406         int pfd[2];
 407         int status;
 408         int ignore;
 409         pid_t victim = -1;
 410         int victim_sock = -1;
 411
 412         dprintf("starting mdmon for %s\n", devnm);
 413
 414         mdfd = open_dev(devnm);
 415         if (mdfd < 0) {
 416                 pr_err("%s: %s\n", devnm, strerror(errno));
 417                 return 1;
 418         }
 419
 420         /* Fork, and have the child tell us when they are ready */
 421         if (must_fork) {
 422                 if (pipe(pfd) != 0) {
 423                         pr_err("failed to create pipe\n");
 424                         return 1;
 425                 }
 426                 switch(fork()) {
 427                 case -1:
 428                         pr_err("failed to fork: %s\n", strerror(errno));
 429                         return 1;
 430                 case 0: /* child */
 431                         close(pfd[0]);
 432                         break;
 433                 default: /* parent */
 434                         close(pfd[1]);
 435                         if (read(pfd[0], &status, sizeof(status)) != sizeof(status)) {
 436                                 wait(&status);
 437                                 status = WEXITSTATUS(status);
 438                         }
 439                         close(pfd[0]);
 440                         return status;
 441                 }
 442         } else
 443                 pfd[0] = pfd[1] = -1;
 444
 445         container = xcalloc(1, sizeof(*container));
 446         strcpy(container->devnm, devnm);
 447         container->arrays = NULL;
 448         container->sock = -1;
 449
 450         mdi = sysfs_read(mdfd, container->devnm, GET_VERSION|GET_LEVEL|GET_DEVS);
 451
 452         if (!mdi) {
 453                 pr_err("failed to load sysfs info for %s\n", container->devnm);
 454                 exit(3);
 455         }
 456         if (mdi->array.level != UnSet) {
 457                 pr_err("%s is not a container - cannot monitor\n", devnm);
 458                 exit(3);
 459         }
 460         if (mdi->array.major_version != -1 ||
 461             mdi->array.minor_version != -2) {
 462                 pr_err("%s does not use external metadata - cannot monitor\n",
 463                         devnm);
 464                 exit(3);
 465         }
 466
 467         container->ss = version_to_superswitch(mdi->text_version);
 468         if (container->ss == NULL) {
 469                 pr_err("%s uses unsupported metadata: %s\n",
 470                         devnm, mdi->text_version);
 471                 exit(3);
 472         }
 473
 474         container->devs = NULL;
 475         for (di = mdi->devs; di; di = di->next) {
 476                 struct mdinfo *cd = xmalloc(sizeof(*cd));
 477                 *cd = *di;
 478                 cd->next = container->devs;
 479                 container->devs = cd;
 480         }
 481         sysfs_free(mdi);
 482
 483         /* SIGUSR is sent between parent and child.  So both block it
 484          * and enable it only with pselect.
 485          */
 486         sigemptyset(&set);
 487         sigaddset(&set, SIGUSR1);
 488         sigaddset(&set, SIGTERM);
 489         sigprocmask(SIG_BLOCK, &set, NULL);
 490         act.sa_handler = wake_me;
 491         act.sa_flags = 0;
 492         sigaction(SIGUSR1, &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         victim = mdmon_pid(container->devnm);
 499         if (victim >= 0)
 500                 victim_sock = connect_monitor(container->devnm);
 501
 502         ignore = chdir("/");
 503         if (!takeover && victim > 0 && victim_sock >= 0) {
 504                 if (fping_monitor(victim_sock) == 0) {
 505                         pr_err("%s already managed\n", container->devnm);
 506                         exit(3);
 507                 }
 508                 close(victim_sock);
 509                 victim_sock = -1;
 510         }
 511         if (container->ss->load_container(container, mdfd, devnm)) {
 512                 pr_err("Cannot load metadata for %s\n", devnm);
 513                 exit(3);
 514         }
 515         close(mdfd);
 516
 517         /* Ok, this is close enough.  We can say goodbye to our parent now.
 518          */
 519         if (victim > 0)
 520                 remove_pidfile(devnm);
 521         if (make_pidfile(devnm) < 0) {
 522                 exit(3);
 523         }
 524         container->sock = make_control_sock(devnm);
 525
 526         status = 0;
 527         if (pfd[1] >= 0) {
 528                 if (write(pfd[1], &status, sizeof(status)) < 0)
 529                         pr_err("failed to notify our parent: %d\n",
 530                                getppid());
 531                 close(pfd[1]);
 532         }
 533
 534         mlockall(MCL_CURRENT | MCL_FUTURE);
 535
 536         if (clone_monitor(container) < 0) {
 537                 pr_err("failed to start monitor process: %s\n",
 538                         strerror(errno));
 539                 exit(2);
 540         }
 541
 542         if (victim > 0) {
 543                 try_kill_monitor(victim, container->devnm, victim_sock);
 544                 if (victim_sock >= 0)
 545                         close(victim_sock);
 546         }
 547
 548         setsid();
 549         close(0);
 550         open("/dev/null", O_RDWR);
 551         close(1);
 552         ignore = dup(0);
 553 #ifndef DEBUG
 554         close(2);
 555         ignore = dup(0);
 556 #endif
 557
 558         /* This silliness is to stop the compiler complaining
 559          * that we ignore 'ignore'
 560          */
 561         if (ignore)
 562                 ignore++;
 563
 564         do_manager(container);
 565
 566         exit(0);
 567 }
 568
 569 /* Some stub functions so super-* can link with us */
 570 int child_monitor(int afd, struct mdinfo *sra, struct reshape *reshape,
 571                   struct supertype *st, unsigned long blocks,
 572                   int *fds, unsigned long long *offsets,
 573                   int dests, int *destfd, unsigned long long *destoffsets)
 574 {
 575         return 0;
 576 }
 577
 578 int restore_stripes(int *dest, unsigned long long *offsets,
 579                     int raid_disks, int chunk_size, int level, int layout,
 580                     int source, unsigned long long read_offset,
 581                     unsigned long long start, unsigned long long length,
 582                     char *src_buf)
 583 {
 584         return 1;
 585 }
 586
 587 int save_stripes(int *source, unsigned long long *offsets,
 588                  int raid_disks, int chunk_size, int level, int layout,
 589                  int nwrites, int *dest,
 590                  unsigned long long start, unsigned long long length,
 591                  char *buf)
 592 {
 593         return 0;
 594 }
 595
 596 struct superswitch super0 = {
 597         .name = "0.90",
 598 };
 599 struct superswitch super1 = {
 600         .name = "1.x",
 601 };