mdmon.c

   1 /*
   2  * mdmon - monitor external metadata arrays
   3  *
   4  * Copyright (C) 2007-2008 Neil Brown <neilb@suse.de>
   5  * Copyright (C) 2007-2008 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
 117 int make_pidfile(char *devname, int o_excl)
 118 {
 119         char path[100];
 120         char pid[10];
 121         int fd;
 122         int n;
 123
 124         if (sigterm)
 125                 return -1;
 126
 127         sprintf(path, "/var/run/mdadm/%s.pid", devname);
 128
 129         fd = open(path, O_RDWR|O_CREAT|o_excl, 0600);
 130         if (fd < 0)
 131                 return -errno;
 132         sprintf(pid, "%d\n", getpid());
 133         n = write(fd, pid, strlen(pid));
 134         close(fd);
 135         if (n < 0)
 136                 return -errno;
 137         return 0;
 138 }
 139
 140 int is_container_member(struct mdstat_ent *mdstat, char *container)
 141 {
 142         if (mdstat->metadata_version == NULL ||
 143             strncmp(mdstat->metadata_version, "external:", 9) != 0 ||
 144             !is_subarray(mdstat->metadata_version+9) ||
 145             strncmp(mdstat->metadata_version+10, container, strlen(container)) != 0 ||
 146             mdstat->metadata_version[10+strlen(container)] != '/')
 147                 return 0;
 148
 149         return 1;
 150 }
 151
 152 void remove_pidfile(char *devname);
 153 static void try_kill_monitor(char *devname)
 154 {
 155         char buf[100];
 156         int fd;
 157         pid_t pid;
 158         struct mdstat_ent *mdstat;
 159
 160         sprintf(buf, "/var/run/mdadm/%s.pid", devname);
 161         fd = open(buf, O_RDONLY);
 162         if (fd < 0)
 163                 return;
 164
 165         if (read(fd, buf, sizeof(buf)) < 0) {
 166                 close(fd);
 167                 return;
 168         }
 169
 170         close(fd);
 171         pid = strtoul(buf, NULL, 10);
 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         if (read(fd, buf, sizeof(buf)) < 0) {
 184                 close(fd);
 185                 return;
 186         }
 187
 188         if (!strstr(buf, "mdmon"))
 189                 return;
 190
 191         kill(pid, SIGTERM);
 192
 193         mdstat = mdstat_read(0, 0);
 194         for ( ; mdstat; mdstat = mdstat->next)
 195                 if (is_container_member(mdstat, devname)) {
 196                         sprintf(buf, "/dev/%s", mdstat->dev);
 197                         WaitClean(buf, 0);
 198                 }
 199         free_mdstat(mdstat);
 200         remove_pidfile(devname);
 201 }
 202
 203 void remove_pidfile(char *devname)
 204 {
 205         char buf[100];
 206
 207         if (sigterm)
 208                 return;
 209
 210         sprintf(buf, "/var/run/mdadm/%s.pid", devname);
 211         unlink(buf);
 212         sprintf(buf, "/var/run/mdadm/%s.sock", devname);
 213         unlink(buf);
 214 }
 215
 216 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, "/var/run/mdadm/%s.sock", 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 int socket_hup_requested;
 246 static void hup(int sig)
 247 {
 248         socket_hup_requested = 1;
 249 }
 250
 251 static void term(int sig)
 252 {
 253         sigterm = 1;
 254 }
 255
 256 static void wake_me(int sig)
 257 {
 258
 259 }
 260
 261 /* if we are debugging and starting mdmon by hand then don't fork */
 262 static int do_fork(void)
 263 {
 264         #ifdef DEBUG
 265         if (check_env("MDADM_NO_MDMON"))
 266                 return 0;
 267         #endif
 268
 269         return 1;
 270 }
 271
 272 void usage(void)
 273 {
 274         fprintf(stderr, "Usage: mdmon /device/name/for/container [target_dir]\n");
 275         exit(2);
 276 }
 277
 278 int mdmon(char *devname, int devnum, int scan, char *switchroot);
 279
 280 int main(int argc, char *argv[])
 281 {
 282         char *container_name = NULL;
 283         char *switchroot = NULL;
 284         int devnum;
 285         char *devname;
 286         int scan = 0;
 287         int status = 0;
 288
 289         switch (argc) {
 290         case 3:
 291                 switchroot = argv[2];
 292         case 2:
 293                 container_name = argv[1];
 294                 break;
 295         default:
 296                 usage();
 297         }
 298
 299         if (strcmp(container_name, "/proc/mdstat") == 0) {
 300                 struct mdstat_ent *mdstat, *e;
 301
 302                 /* launch an mdmon instance for each container found */
 303                 scan = 1;
 304                 mdstat = mdstat_read(0, 0);
 305                 for (e = mdstat; e; e = e->next) {
 306                         if (strncmp(e->metadata_version, "external:", 9) == 0 &&
 307                             !is_subarray(&e->metadata_version[9])) {
 308                                 devname = devnum2devname(e->devnum);
 309                                 /* update cmdline so this mdmon instance can be
 310                                  * distinguished from others in a call to ps(1)
 311                                  */
 312                                 if (strlen(devname) <= strlen(container_name)) {
 313                                         memset(container_name, 0, strlen(container_name));
 314                                         sprintf(container_name, "%s", devname);
 315                                 }
 316                                 status |= mdmon(devname, e->devnum, scan,
 317                                                 switchroot);
 318                         }
 319                 }
 320                 free_mdstat(mdstat);
 321
 322                 return status;
 323         } else if (strncmp(container_name, "md", 2) == 0) {
 324                 devnum = devname2devnum(container_name);
 325                 devname = devnum2devname(devnum);
 326                 if (strcmp(container_name, devname) != 0)
 327                         devname = NULL;
 328         } else {
 329                 struct stat st;
 330
 331                 devnum = NoMdDev;
 332                 if (stat(container_name, &st) == 0)
 333                         devnum = stat2devnum(&st);
 334                 if (devnum == NoMdDev)
 335                         devname = NULL;
 336                 else
 337                         devname = devnum2devname(devnum);
 338         }
 339
 340         if (!devname) {
 341                 fprintf(stderr, "mdmon: %s is not a valid md device name\n",
 342                         container_name);
 343                 exit(1);
 344         }
 345         return mdmon(devname, devnum, scan, switchroot);
 346 }
 347
 348 int mdmon(char *devname, int devnum, int scan, char *switchroot)
 349 {
 350         int mdfd;
 351         struct mdinfo *mdi, *di;
 352         struct supertype *container;
 353         sigset_t set;
 354         struct sigaction act;
 355         int pfd[2];
 356         int status;
 357         int ignore;
 358
 359         dprintf("starting mdmon for %s in %s\n",
 360                 devname, switchroot ? : "/");
 361         mdfd = open_dev(devnum);
 362         if (mdfd < 0) {
 363                 fprintf(stderr, "mdmon: %s: %s\n", devname,
 364                         strerror(errno));
 365                 return 1;
 366         }
 367         if (md_get_version(mdfd) < 0) {
 368                 fprintf(stderr, "mdmon: %s: Not an md device\n",
 369                         devname);
 370                 return 1;
 371         }
 372
 373         /* Fork, and have the child tell us when they are ready */
 374         if (do_fork() || scan) {
 375                 if (pipe(pfd) != 0) {
 376                         fprintf(stderr, "mdmon: failed to create pipe\n");
 377                         return 1;
 378                 }
 379                 switch(fork()) {
 380                 case -1:
 381                         fprintf(stderr, "mdmon: failed to fork: %s\n",
 382                                 strerror(errno));
 383                         return 1;
 384                 case 0: /* child */
 385                         close(pfd[0]);
 386                         break;
 387                 default: /* parent */
 388                         close(pfd[1]);
 389                         if (read(pfd[0], &status, sizeof(status)) != sizeof(status)) {
 390                                 wait(&status);
 391                                 status = WEXITSTATUS(status);
 392                         }
 393                         return status;
 394                 }
 395         } else
 396                 pfd[0] = pfd[1] = -1;
 397
 398         container = malloc(sizeof(*container));
 399         container->devnum = devnum;
 400         container->devname = devname;
 401         container->arrays = NULL;
 402         container->subarray[0] = 0;
 403
 404         if (!container->devname) {
 405                 fprintf(stderr, "mdmon: failed to allocate container name string\n");
 406                 exit(3);
 407         }
 408
 409         mdi = sysfs_read(mdfd, container->devnum,
 410                          GET_VERSION|GET_LEVEL|GET_DEVS|SKIP_GONE_DEVS);
 411
 412         if (!mdi) {
 413                 fprintf(stderr, "mdmon: failed to load sysfs info for %s\n",
 414                         container->devname);
 415                 exit(3);
 416         }
 417         if (mdi->array.level != UnSet) {
 418                 fprintf(stderr, "mdmon: %s is not a container - cannot monitor\n",
 419                         devname);
 420                 exit(3);
 421         }
 422         if (mdi->array.major_version != -1 ||
 423             mdi->array.minor_version != -2) {
 424                 fprintf(stderr, "mdmon: %s does not use external metadata - cannot monitor\n",
 425                         devname);
 426                 exit(3);
 427         }
 428
 429         container->ss = find_metadata_methods(mdi->text_version);
 430         if (container->ss == NULL) {
 431                 fprintf(stderr, "mdmon: %s uses unknown metadata: %s\n",
 432                         devname, mdi->text_version);
 433                 exit(3);
 434         }
 435
 436         container->devs = NULL;
 437         for (di = mdi->devs; di; di = di->next) {
 438                 struct mdinfo *cd = malloc(sizeof(*cd));
 439                 *cd = *di;
 440                 cd->next = container->devs;
 441                 container->devs = cd;
 442         }
 443         sysfs_free(mdi);
 444
 445         /* SIGUSR is sent between parent and child.  So both block it
 446          * and enable it only with pselect.
 447          */
 448         sigemptyset(&set);
 449         sigaddset(&set, SIGUSR1);
 450         sigaddset(&set, SIGHUP);
 451         sigaddset(&set, SIGALRM);
 452         sigaddset(&set, SIGTERM);
 453         sigprocmask(SIG_BLOCK, &set, NULL);
 454         act.sa_handler = wake_me;
 455         act.sa_flags = 0;
 456         sigaction(SIGUSR1, &act, NULL);
 457         sigaction(SIGALRM, &act, NULL);
 458         act.sa_handler = hup;
 459         sigaction(SIGHUP, &act, NULL);
 460         act.sa_handler = term;
 461         sigaction(SIGTERM, &act, NULL);
 462         act.sa_handler = SIG_IGN;
 463         sigaction(SIGPIPE, &act, NULL);
 464
 465         if (switchroot) {
 466                 /* we assume we assume that /sys /proc /dev are available in
 467                  * the new root (see nash:setuproot)
 468                  *
 469                  * kill any monitors in the current namespace and change
 470                  * to the new one
 471                  */
 472                 try_kill_monitor(container->devname);
 473                 if (chroot(switchroot) != 0) {
 474                         fprintf(stderr, "mdmon: failed to chroot to '%s': %s\n",
 475                                 switchroot, strerror(errno));
 476                         exit(4);
 477                 }
 478         }
 479
 480         /* If this fails, we hope it already exists
 481          * pid file lives in /var/run/mdadm/mdXX.pid
 482          */
 483         mkdir("/var", 0600);
 484         mkdir("/var/run", 0600);
 485         mkdir("/var/run/mdadm", 0600);
 486         ignore = chdir("/");
 487         if (make_pidfile(container->devname, O_EXCL) < 0) {
 488                 if (ping_monitor(container->devname) == 0) {
 489                         fprintf(stderr, "mdmon: %s already managed\n",
 490                                 container->devname);
 491                         exit(3);
 492                 } else {
 493                         int err;
 494
 495                         /* cleanup the old monitor, this one is taking over */
 496                         try_kill_monitor(container->devname);
 497                         err = make_pidfile(container->devname, 0);
 498                         if (err < 0) {
 499                                 fprintf(stderr, "mdmon: %s Cannot create pidfile\n",
 500                                         container->devname);
 501                                 if (err == -EROFS) {
 502                                         /* FIXME implement a mechanism to
 503                                          * prevent duplicate monitor instances
 504                                          */
 505                                         fprintf(stderr,
 506                                                 "mdmon: continuing on read-only file system\n");
 507                                 } else
 508                                         exit(3);
 509                         }
 510                 }
 511         }
 512         container->sock = make_control_sock(container->devname);
 513
 514         if (container->ss->load_super(container, mdfd, devname)) {
 515                 fprintf(stderr, "mdmon: Cannot load metadata for %s\n",
 516                         devname);
 517                 exit(3);
 518         }
 519         close(mdfd);
 520
 521         /* Ok, this is close enough.  We can say goodbye to our parent now.
 522          */
 523         status = 0;
 524         if (write(pfd[1], &status, sizeof(status)) < 0)
 525                 fprintf(stderr, "mdmon: failed to notify our parent: %d\n",
 526                         getppid());
 527         close(pfd[1]);
 528
 529         setsid();
 530         close(0);
 531         open("/dev/null", O_RDWR);
 532         close(1);
 533         ignore = dup(0);
 534 #ifndef DEBUG
 535         close(2);
 536         ignore = dup(0);
 537 #endif
 538
 539         mlockall(MCL_FUTURE);
 540
 541         if (clone_monitor(container) < 0) {
 542                 fprintf(stderr, "mdmon: failed to start monitor process: %s\n",
 543                         strerror(errno));
 544                 exit(2);
 545         }
 546
 547         do_manager(container);
 548
 549         exit(0);
 550 }