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                 sleep_for(0, USEC_TO_NSEC(10), true);
 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                 sleep_for(0, MSEC_TO_NSEC(200), true);
 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         snprintf(addr.sun_path, sizeof(addr.sun_path), "%s", 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 bool is_duplicate_opt(const int opt, const int set_val, const char *long_name)
 292 {
 293         if (opt == set_val) {
 294                 pr_err("--%s option duplicated!\n", long_name);
 295                 return true;
 296         }
 297         return false;
 298 }
 299
 300 static int mdmon(char *devnm, int must_fork, int takeover);
 301
 302 int main(int argc, char *argv[])
 303 {
 304         char *container_name = NULL;
 305         int status = 0;
 306         int opt;
 307         int all = 0;
 308         int takeover = 0;
 309         int dofork = 1;
 310         int mdfd = -1;
 311         bool help = false;
 312         static struct option options[] = {
 313                 {"all", 0, NULL, 'a'},
 314                 {"takeover", 0, NULL, 't'},
 315                 {"help", 0, NULL, 'h'},
 316                 {"offroot", 0, NULL, OffRootOpt},
 317                 {"foreground", 0, NULL, 'F'},
 318                 {NULL, 0, NULL, 0}
 319         };
 320
 321         /*
 322          * mdmon should never complain due to lack of a platform,
 323          * that is mdadm's job if at all.
 324          */
 325         imsm_set_no_platform(1);
 326
 327         while ((opt = getopt_long(argc, argv, "thaF", options, NULL)) != -1) {
 328                 switch (opt) {
 329                 case 'a':
 330                         if (is_duplicate_opt(all, 1, "all"))
 331                                 exit(1);
 332                         container_name = argv[optind-1];
 333                         all = 1;
 334                         break;
 335                 case 't':
 336                         if (is_duplicate_opt(takeover, 1, "takeover"))
 337                                 exit(1);
 338                         takeover = 1;
 339                         break;
 340                 case 'F':
 341                         if (is_duplicate_opt(dofork, 0, "foreground"))
 342                                 exit(1);
 343                         dofork = 0;
 344                         break;
 345                 case OffRootOpt:
 346                         if (is_duplicate_opt(argv[0][0], '@', "offroot"))
 347                                 exit(1);
 348                         argv[0][0] = '@';
 349                         break;
 350                 case 'h':
 351                         if (is_duplicate_opt(help, true, "help"))
 352                                 exit(1);
 353                         help = true;
 354                         break;
 355                 default:
 356                         usage();
 357                         break;
 358                 }
 359         }
 360
 361         if (in_initrd()) {
 362                 /*
 363                  * set first char of argv[0] to @. This is used by
 364                  * systemd to signal that the task was launched from
 365                  * initrd/initramfs and should be preserved during shutdown
 366                  */
 367                 argv[0][0] = '@';
 368         }
 369
 370         if (!all && argv[optind]) {
 371                 static const char prefix[] = "initrd/";
 372                 container_name = argv[optind];
 373                 if (strncmp(container_name, prefix,
 374                             sizeof(prefix) - 1) == 0)
 375                         container_name += sizeof(prefix)-1;
 376                 container_name = get_md_name(container_name);
 377                 if (!container_name)
 378                         return 1;
 379         }
 380
 381         if (container_name == NULL || argc - optind > 1)
 382                 usage();
 383
 384         if (strcmp(container_name, "/proc/mdstat") == 0)
 385                 all = 1;
 386
 387         if (help)
 388                 usage();
 389
 390         if (all) {
 391                 struct mdstat_ent *mdstat, *e;
 392                 int container_len = strnlen(container_name, MD_NAME_MAX);
 393
 394                 /* launch an mdmon instance for each container found */
 395                 mdstat = mdstat_read(0, 0);
 396                 for (e = mdstat; e; e = e->next) {
 397                         if (e->metadata_version &&
 398                             strncmp(e->metadata_version, "external:", 9) == 0 &&
 399                             !is_subarray(&e->metadata_version[9])) {
 400                                 /* update cmdline so this mdmon instance can be
 401                                  * distinguished from others in a call to ps(1)
 402                                  */
 403                                 if (strlen(e->devnm) <= (unsigned)container_len) {
 404                                         memset(container_name, 0, container_len);
 405                                         sprintf(container_name, "%s", e->devnm);
 406                                 }
 407                                 status |= mdmon(e->devnm, 1, takeover);
 408                         }
 409                 }
 410                 free_mdstat(mdstat);
 411
 412                 return status;
 413         }
 414
 415         mdfd = open_mddev(container_name, 0);
 416         if (is_fd_valid(mdfd)) {
 417                 char *devnm = fd2devnm(mdfd);
 418
 419                 close(mdfd);
 420
 421                 if (devnm)
 422                         return mdmon(devnm, dofork && do_fork(), takeover);
 423         }
 424
 425         pr_err("%s is not a valid md device name\n", container_name);
 426         return 1;
 427 }
 428
 429 static int mdmon(char *devnm, int must_fork, int takeover)
 430 {
 431         int mdfd;
 432         struct mdinfo *mdi, *di;
 433         struct supertype *container;
 434         sigset_t set;
 435         struct sigaction act;
 436         int pfd[2];
 437         int status;
 438         int ignore;
 439         pid_t victim = -1;
 440         int victim_sock = -1;
 441
 442         dprintf("starting mdmon for %s\n", devnm);
 443
 444         mdfd = open_dev(devnm);
 445         if (mdfd < 0) {
 446                 pr_err("%s: %s\n", devnm, strerror(errno));
 447                 return 1;
 448         }
 449
 450         /* Fork, and have the child tell us when they are ready */
 451         if (must_fork) {
 452                 if (pipe(pfd) != 0) {
 453                         pr_err("failed to create pipe\n");
 454                         return 1;
 455                 }
 456                 switch(fork()) {
 457                 case -1:
 458                         pr_err("failed to fork: %s\n", strerror(errno));
 459                         return 1;
 460                 case 0: /* child */
 461                         close(pfd[0]);
 462                         break;
 463                 default: /* parent */
 464                         close(pfd[1]);
 465                         if (read(pfd[0], &status, sizeof(status)) != sizeof(status)) {
 466                                 wait(&status);
 467                                 status = WEXITSTATUS(status);
 468                         }
 469                         close(pfd[0]);
 470                         return status;
 471                 }
 472         } else
 473                 pfd[0] = pfd[1] = -1;
 474
 475         container = xcalloc(1, sizeof(*container));
 476         snprintf(container->devnm, MD_NAME_MAX, "%s", devnm);
 477         container->arrays = NULL;
 478         container->sock = -1;
 479
 480         mdi = sysfs_read(mdfd, container->devnm, GET_VERSION|GET_LEVEL|GET_DEVS);
 481
 482         if (!mdi) {
 483                 pr_err("failed to load sysfs info for %s\n", container->devnm);
 484                 exit(3);
 485         }
 486         if (mdi->array.level != UnSet) {
 487                 pr_err("%s is not a container - cannot monitor\n", devnm);
 488                 exit(3);
 489         }
 490         if (mdi->array.major_version != -1 ||
 491             mdi->array.minor_version != -2) {
 492                 pr_err("%s does not use external metadata - cannot monitor\n",
 493                         devnm);
 494                 exit(3);
 495         }
 496
 497         container->ss = version_to_superswitch(mdi->text_version);
 498         if (container->ss == NULL) {
 499                 pr_err("%s uses unsupported metadata: %s\n",
 500                         devnm, mdi->text_version);
 501                 exit(3);
 502         }
 503
 504         container->devs = NULL;
 505         for (di = mdi->devs; di; di = di->next) {
 506                 struct mdinfo *cd = xmalloc(sizeof(*cd));
 507                 *cd = *di;
 508                 cd->next = container->devs;
 509                 container->devs = cd;
 510         }
 511         sysfs_free(mdi);
 512
 513         /* SIGUSR is sent between parent and child.  So both block it
 514          * and enable it only with pselect.
 515          */
 516         sigemptyset(&set);
 517         sigaddset(&set, SIGUSR1);
 518         sigaddset(&set, SIGTERM);
 519         sigprocmask(SIG_BLOCK, &set, NULL);
 520         act.sa_handler = wake_me;
 521         act.sa_flags = 0;
 522         sigaction(SIGUSR1, &act, NULL);
 523         act.sa_handler = term;
 524         sigaction(SIGTERM, &act, NULL);
 525         act.sa_handler = SIG_IGN;
 526         sigaction(SIGPIPE, &act, NULL);
 527
 528         victim = mdmon_pid(container->devnm);
 529         if (victim >= 0)
 530                 victim_sock = connect_monitor(container->devnm);
 531
 532         ignore = chdir("/");
 533         if (!takeover && victim > 0 && victim_sock >= 0) {
 534                 if (fping_monitor(victim_sock) == 0) {
 535                         pr_err("%s already managed\n", container->devnm);
 536                         exit(3);
 537                 }
 538                 close(victim_sock);
 539                 victim_sock = -1;
 540         }
 541         if (container->ss->load_container(container, mdfd, devnm)) {
 542                 pr_err("Cannot load metadata for %s\n", devnm);
 543                 exit(3);
 544         }
 545         close(mdfd);
 546
 547         /* Ok, this is close enough.  We can say goodbye to our parent now.
 548          */
 549         if (victim > 0)
 550                 remove_pidfile(devnm);
 551         if (make_pidfile(devnm) < 0) {
 552                 exit(3);
 553         }
 554         container->sock = make_control_sock(devnm);
 555
 556         status = 0;
 557         if (pfd[1] >= 0) {
 558                 if (write(pfd[1], &status, sizeof(status)) < 0)
 559                         pr_err("failed to notify our parent: %d\n",
 560                                getppid());
 561                 close(pfd[1]);
 562         }
 563
 564         mlockall(MCL_CURRENT | MCL_FUTURE);
 565
 566         if (clone_monitor(container) < 0) {
 567                 pr_err("failed to start monitor process: %s\n",
 568                         strerror(errno));
 569                 exit(2);
 570         }
 571
 572         if (victim > 0) {
 573                 try_kill_monitor(victim, container->devnm, victim_sock);
 574                 if (victim_sock >= 0)
 575                         close(victim_sock);
 576         }
 577
 578         setsid();
 579         manage_fork_fds(0);
 580
 581         /* This silliness is to stop the compiler complaining
 582          * that we ignore 'ignore'
 583          */
 584         if (ignore)
 585                 ignore++;
 586
 587         do_manager(container);
 588
 589         exit(0);
 590 }
 591
 592 /* Some stub functions so super-* can link with us */
 593 int child_monitor(int afd, struct mdinfo *sra, struct reshape *reshape,
 594                   struct supertype *st, unsigned long blocks,
 595                   int *fds, unsigned long long *offsets,
 596                   int dests, int *destfd, unsigned long long *destoffsets)
 597 {
 598         return 0;
 599 }
 600
 601 int restore_stripes(int *dest, unsigned long long *offsets,
 602                     int raid_disks, int chunk_size, int level, int layout,
 603                     int source, unsigned long long read_offset,
 604                     unsigned long long start, unsigned long long length,
 605                     char *src_buf)
 606 {
 607         return 1;
 608 }
 609
 610 int save_stripes(int *source, unsigned long long *offsets,
 611                  int raid_disks, int chunk_size, int level, int layout,
 612                  int nwrites, int *dest,
 613                  unsigned long long start, unsigned long long length,
 614                  char *buf)
 615 {
 616         return 0;
 617 }
 618
 619 struct superswitch super0 = {
 620         .name = "0.90",
 621 };
 622 struct superswitch super1 = {
 623         .name = "1.x",
 624 };