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 make_pidfile(char *devname)
 117 {
 118         char path[100];
 119         char pid[10];
 120         int fd;
 121         int n;
 122
 123         sprintf(path, "%s/%s.pid", pid_dir, devname);
 124
 125         fd = open(path, O_RDWR|O_CREAT|O_EXCL, 0600);
 126         if (fd < 0)
 127                 return -errno;
 128         sprintf(pid, "%d\n", getpid());
 129         n = write(fd, pid, strlen(pid));
 130         close(fd);
 131         if (n < 0)
 132                 return -errno;
 133         return 0;
 134 }
 135
 136 int is_container_member(struct mdstat_ent *mdstat, char *container)
 137 {
 138         if (mdstat->metadata_version == NULL ||
 139             strncmp(mdstat->metadata_version, "external:", 9) != 0 ||
 140             !is_subarray(mdstat->metadata_version+9) ||
 141             strncmp(mdstat->metadata_version+10, container, strlen(container)) != 0 ||
 142             mdstat->metadata_version[10+strlen(container)] != '/')
 143                 return 0;
 144
 145         return 1;
 146 }
 147
 148 static void try_kill_monitor(pid_t pid, char *devname, int sock)
 149 {
 150         char buf[100];
 151         int fd;
 152         int n;
 153         long fl;
 154
 155         /* first rule of survival... don't off yourself */
 156         if (pid == getpid())
 157                 return;
 158
 159         /* kill this process if it is mdmon */
 160         sprintf(buf, "/proc/%lu/cmdline", (unsigned long) pid);
 161         fd = open(buf, O_RDONLY);
 162         if (fd < 0)
 163                 return;
 164
 165         n = read(fd, buf, sizeof(buf)-1);
 166         buf[sizeof(buf)-1] = 0;
 167         close(fd);
 168
 169         if (n < 0 || !strstr(buf, "mdmon"))
 170                 return;
 171
 172         kill(pid, SIGTERM);
 173
 174         /* Wait for monitor to exit by reading from the socket, after
 175          * clearing the non-blocking flag */
 176         fl = fcntl(sock, F_GETFL, 0);
 177         fl &= ~O_NONBLOCK;
 178         fcntl(sock, F_SETFL, fl);
 179         read(sock, buf, 100);
 180 }
 181
 182 void remove_pidfile(char *devname)
 183 {
 184         char buf[100];
 185
 186         sprintf(buf, "%s/%s.pid", pid_dir, devname);
 187         unlink(buf);
 188         sprintf(buf, "%s/%s.sock", pid_dir, devname);
 189         unlink(buf);
 190         if (strcmp(pid_dir, ALT_RUN) == 0)
 191                 /* try to clean up when we are finished with this dir */
 192                 rmdir(pid_dir);
 193 }
 194
 195 static int make_control_sock(char *devname)
 196 {
 197         char path[100];
 198         int sfd;
 199         long fl;
 200         struct sockaddr_un addr;
 201
 202         if (sigterm)
 203                 return -1;
 204
 205         sprintf(path, "%s/%s.sock", pid_dir, devname);
 206         unlink(path);
 207         sfd = socket(PF_LOCAL, SOCK_STREAM, 0);
 208         if (sfd < 0)
 209                 return -1;
 210
 211         addr.sun_family = PF_LOCAL;
 212         strcpy(addr.sun_path, path);
 213         if (bind(sfd, &addr, sizeof(addr)) < 0) {
 214                 close(sfd);
 215                 return -1;
 216         }
 217         listen(sfd, 10);
 218         fl = fcntl(sfd, F_GETFL, 0);
 219         fl |= O_NONBLOCK;
 220         fcntl(sfd, F_SETFL, fl);
 221         return sfd;
 222 }
 223
 224 static void term(int sig)
 225 {
 226         sigterm = 1;
 227 }
 228
 229 static void wake_me(int sig)
 230 {
 231
 232 }
 233
 234 /* if we are debugging and starting mdmon by hand then don't fork */
 235 static int do_fork(void)
 236 {
 237         #ifdef DEBUG
 238         if (check_env("MDADM_NO_MDMON"))
 239                 return 0;
 240         #endif
 241
 242         return 1;
 243 }
 244
 245 void usage(void)
 246 {
 247         fprintf(stderr, "Usage: mdmon /device/name/for/container [target_dir]\n");
 248         exit(2);
 249 }
 250
 251 static int mdmon(char *devname, int devnum, int must_fork, int takeover);
 252
 253 int main(int argc, char *argv[])
 254 {
 255         char *container_name = NULL;
 256         int devnum;
 257         char *devname;
 258         int status = 0;
 259         int arg;
 260         int all = 0;
 261         int takeover = 0;
 262
 263         for (arg = 1; arg < argc; arg++) {
 264                 if (strcmp(argv[arg], "--all") == 0 ||
 265                     strcmp(argv[arg], "/proc/mdstat") == 0)
 266                         all = 1;
 267                 else if (strcmp(argv[arg], "--takeover") == 0)
 268                         takeover = 1;
 269                 else if (container_name == NULL)
 270                         container_name = argv[arg];
 271                 else
 272                         usage();
 273         }
 274
 275         if (all) {
 276                 struct mdstat_ent *mdstat, *e;
 277
 278                 /* launch an mdmon instance for each container found */
 279                 mdstat = mdstat_read(0, 0);
 280                 for (e = mdstat; e; e = e->next) {
 281                         if (strncmp(e->metadata_version, "external:", 9) == 0 &&
 282                             !is_subarray(&e->metadata_version[9])) {
 283                                 devname = devnum2devname(e->devnum);
 284                                 /* update cmdline so this mdmon instance can be
 285                                  * distinguished from others in a call to ps(1)
 286                                  */
 287                                 if (strlen(devname) <= strlen(container_name)) {
 288                                         memset(container_name, 0, strlen(container_name));
 289                                         sprintf(container_name, "%s", devname);
 290                                 }
 291                                 status |= mdmon(devname, e->devnum, 1,
 292                                                 takeover);
 293                         }
 294                 }
 295                 free_mdstat(mdstat);
 296
 297                 return status;
 298         } else if (strncmp(container_name, "md", 2) == 0) {
 299                 devnum = devname2devnum(container_name);
 300                 devname = devnum2devname(devnum);
 301                 if (strcmp(container_name, devname) != 0)
 302                         devname = NULL;
 303         } else {
 304                 struct stat st;
 305
 306                 devnum = NoMdDev;
 307                 if (stat(container_name, &st) == 0)
 308                         devnum = stat2devnum(&st);
 309                 if (devnum == NoMdDev)
 310                         devname = NULL;
 311                 else
 312                         devname = devnum2devname(devnum);
 313         }
 314
 315         if (!devname) {
 316                 fprintf(stderr, "mdmon: %s is not a valid md device name\n",
 317                         container_name);
 318                 exit(1);
 319         }
 320         return mdmon(devname, devnum, do_fork(), takeover);
 321 }
 322
 323 static int mdmon(char *devname, int devnum, int must_fork, int takeover)
 324 {
 325         int mdfd;
 326         struct mdinfo *mdi, *di;
 327         struct supertype *container;
 328         sigset_t set;
 329         struct sigaction act;
 330         int pfd[2];
 331         int status;
 332         int ignore;
 333         pid_t victim = -1;
 334         int victim_sock = -1;
 335
 336         dprintf("starting mdmon for %s\n", devname);
 337
 338         mdfd = open_dev(devnum);
 339         if (mdfd < 0) {
 340                 fprintf(stderr, "mdmon: %s: %s\n", devname,
 341                         strerror(errno));
 342                 return 1;
 343         }
 344         if (md_get_version(mdfd) < 0) {
 345                 fprintf(stderr, "mdmon: %s: Not an md device\n",
 346                         devname);
 347                 return 1;
 348         }
 349
 350         /* Fork, and have the child tell us when they are ready */
 351         if (must_fork) {
 352                 if (pipe(pfd) != 0) {
 353                         fprintf(stderr, "mdmon: failed to create pipe\n");
 354                         return 1;
 355                 }
 356                 switch(fork()) {
 357                 case -1:
 358                         fprintf(stderr, "mdmon: failed to fork: %s\n",
 359                                 strerror(errno));
 360                         return 1;
 361                 case 0: /* child */
 362                         close(pfd[0]);
 363                         break;
 364                 default: /* parent */
 365                         close(pfd[1]);
 366                         if (read(pfd[0], &status, sizeof(status)) != sizeof(status)) {
 367                                 wait(&status);
 368                                 status = WEXITSTATUS(status);
 369                         }
 370                         return status;
 371                 }
 372         } else
 373                 pfd[0] = pfd[1] = -1;
 374
 375         container = calloc(1, sizeof(*container));
 376         container->devnum = devnum;
 377         container->devname = devname;
 378         container->arrays = NULL;
 379         container->subarray[0] = 0;
 380         container->sock = -1;
 381
 382         if (!container->devname) {
 383                 fprintf(stderr, "mdmon: failed to allocate container name string\n");
 384                 exit(3);
 385         }
 386
 387         mdi = sysfs_read(mdfd, container->devnum,
 388                          GET_VERSION|GET_LEVEL|GET_DEVS|SKIP_GONE_DEVS);
 389
 390         if (!mdi) {
 391                 fprintf(stderr, "mdmon: failed to load sysfs info for %s\n",
 392                         container->devname);
 393                 exit(3);
 394         }
 395         if (mdi->array.level != UnSet) {
 396                 fprintf(stderr, "mdmon: %s is not a container - cannot monitor\n",
 397                         devname);
 398                 exit(3);
 399         }
 400         if (mdi->array.major_version != -1 ||
 401             mdi->array.minor_version != -2) {
 402                 fprintf(stderr, "mdmon: %s does not use external metadata - cannot monitor\n",
 403                         devname);
 404                 exit(3);
 405         }
 406
 407         container->ss = find_metadata_methods(mdi->text_version);
 408         if (container->ss == NULL) {
 409                 fprintf(stderr, "mdmon: %s uses unknown metadata: %s\n",
 410                         devname, mdi->text_version);
 411                 exit(3);
 412         }
 413
 414         container->devs = NULL;
 415         for (di = mdi->devs; di; di = di->next) {
 416                 struct mdinfo *cd = malloc(sizeof(*cd));
 417                 *cd = *di;
 418                 cd->next = container->devs;
 419                 container->devs = cd;
 420         }
 421         sysfs_free(mdi);
 422
 423         /* SIGUSR is sent between parent and child.  So both block it
 424          * and enable it only with pselect.
 425          */
 426         sigemptyset(&set);
 427         sigaddset(&set, SIGUSR1);
 428         sigaddset(&set, SIGTERM);
 429         sigprocmask(SIG_BLOCK, &set, NULL);
 430         act.sa_handler = wake_me;
 431         act.sa_flags = 0;
 432         sigaction(SIGUSR1, &act, NULL);
 433         act.sa_handler = term;
 434         sigaction(SIGTERM, &act, NULL);
 435         act.sa_handler = SIG_IGN;
 436         sigaction(SIGPIPE, &act, NULL);
 437
 438         if (takeover) {
 439                 pid_dir = VAR_RUN;
 440                 victim = mdmon_pid(container->devnum);
 441                 if (victim < 0) {
 442                         pid_dir = ALT_RUN;
 443                         victim = mdmon_pid(container->devnum);
 444                 }
 445                 if (victim >= 0)
 446                         victim_sock = connect_monitor(container->devname);
 447         }
 448
 449         ignore = chdir("/");
 450         if (victim < 0) {
 451                 if (ping_monitor(container->devname) == 0) {
 452                         fprintf(stderr, "mdmon: %s already managed\n",
 453                                 container->devname);
 454                         exit(3);
 455                 }
 456                 /* if there is a pid file, kill whoever is there just in case */
 457                 victim = mdmon_pid(container->devnum);
 458         }
 459         if (container->ss->load_super(container, mdfd, devname)) {
 460                 fprintf(stderr, "mdmon: Cannot load metadata for %s\n",
 461                         devname);
 462                 exit(3);
 463         }
 464         close(mdfd);
 465
 466         /* Ok, this is close enough.  We can say goodbye to our parent now.
 467          */
 468         if (victim > 0)
 469                 remove_pidfile(devname);
 470         if (mkdir(VAR_RUN, 0600) >= 0 || errno == EEXIST)
 471                 pid_dir = VAR_RUN;
 472         else if (mkdir(ALT_RUN, 0600) >= 0 || errno == EEXIST)
 473                 pid_dir = ALT_RUN;
 474         else {
 475                 fprintf(stderr, "mdmon: Neither %s nor %s are writable\n"
 476                         "       cannot create .pid or .sock files.  Aborting\n",
 477                         VAR_RUN, ALT_RUN);
 478                 exit(3);
 479         }
 480         if (make_pidfile(devname) < 0) {
 481                 fprintf(stderr, "mdmon: Cannot create pid file in %s - aborting.\n",
 482                         pid_dir);
 483                 exit(3);
 484         }
 485         container->sock = make_control_sock(devname);
 486
 487         status = 0;
 488         if (write(pfd[1], &status, sizeof(status)) < 0)
 489                 fprintf(stderr, "mdmon: failed to notify our parent: %d\n",
 490                         getppid());
 491         close(pfd[1]);
 492
 493         mlockall(MCL_CURRENT | MCL_FUTURE);
 494
 495         if (clone_monitor(container) < 0) {
 496                 fprintf(stderr, "mdmon: failed to start monitor process: %s\n",
 497                         strerror(errno));
 498                 exit(2);
 499         }
 500
 501         if (victim > 0) {
 502                 try_kill_monitor(victim, container->devname, victim_sock);
 503                 close(victim_sock);
 504         }
 505
 506         setsid();
 507         close(0);
 508         open("/dev/null", O_RDWR);
 509         close(1);
 510         ignore = dup(0);
 511 #ifndef DEBUG
 512         close(2);
 513         ignore = dup(0);
 514 #endif
 515
 516         do_manager(container);
 517
 518         exit(0);
 519 }