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         n = read(sock, buf, 100);
 180         /* Ignore result, it is just the wait that
 181          * matters
 182          */
 183 }
 184
 185 void remove_pidfile(char *devname)
 186 {
 187         char buf[100];
 188
 189         sprintf(buf, "%s/%s.pid", pid_dir, devname);
 190         unlink(buf);
 191         sprintf(buf, "%s/%s.sock", pid_dir, devname);
 192         unlink(buf);
 193         if (strcmp(pid_dir, ALT_RUN) == 0)
 194                 /* try to clean up when we are finished with this dir */
 195                 rmdir(pid_dir);
 196 }
 197
 198 static int make_control_sock(char *devname)
 199 {
 200         char path[100];
 201         int sfd;
 202         long fl;
 203         struct sockaddr_un addr;
 204
 205         if (sigterm)
 206                 return -1;
 207
 208         sprintf(path, "%s/%s.sock", pid_dir, devname);
 209         unlink(path);
 210         sfd = socket(PF_LOCAL, SOCK_STREAM, 0);
 211         if (sfd < 0)
 212                 return -1;
 213
 214         addr.sun_family = PF_LOCAL;
 215         strcpy(addr.sun_path, path);
 216         if (bind(sfd, &addr, sizeof(addr)) < 0) {
 217                 close(sfd);
 218                 return -1;
 219         }
 220         listen(sfd, 10);
 221         fl = fcntl(sfd, F_GETFL, 0);
 222         fl |= O_NONBLOCK;
 223         fcntl(sfd, F_SETFL, fl);
 224         return sfd;
 225 }
 226
 227 static void term(int sig)
 228 {
 229         sigterm = 1;
 230 }
 231
 232 static void wake_me(int sig)
 233 {
 234
 235 }
 236
 237 /* if we are debugging and starting mdmon by hand then don't fork */
 238 static int do_fork(void)
 239 {
 240         #ifdef DEBUG
 241         if (check_env("MDADM_NO_MDMON"))
 242                 return 0;
 243         #endif
 244
 245         return 1;
 246 }
 247
 248 void usage(void)
 249 {
 250         fprintf(stderr, "Usage: mdmon /device/name/for/container [target_dir]\n");
 251         exit(2);
 252 }
 253
 254 static int mdmon(char *devname, int devnum, int must_fork, int takeover);
 255
 256 int main(int argc, char *argv[])
 257 {
 258         char *container_name = NULL;
 259         int devnum;
 260         char *devname;
 261         int status = 0;
 262         int arg;
 263         int all = 0;
 264         int takeover = 0;
 265
 266         for (arg = 1; arg < argc; arg++) {
 267                 if (strcmp(argv[arg], "--all") == 0 ||
 268                     strcmp(argv[arg], "/proc/mdstat") == 0)
 269                         all = 1;
 270                 else if (strcmp(argv[arg], "--takeover") == 0)
 271                         takeover = 1;
 272                 else if (container_name == NULL)
 273                         container_name = argv[arg];
 274                 else
 275                         usage();
 276         }
 277
 278         if (all) {
 279                 struct mdstat_ent *mdstat, *e;
 280
 281                 /* launch an mdmon instance for each container found */
 282                 mdstat = mdstat_read(0, 0);
 283                 for (e = mdstat; e; e = e->next) {
 284                         if (strncmp(e->metadata_version, "external:", 9) == 0 &&
 285                             !is_subarray(&e->metadata_version[9])) {
 286                                 devname = devnum2devname(e->devnum);
 287                                 /* update cmdline so this mdmon instance can be
 288                                  * distinguished from others in a call to ps(1)
 289                                  */
 290                                 if (strlen(devname) <= strlen(container_name)) {
 291                                         memset(container_name, 0, strlen(container_name));
 292                                         sprintf(container_name, "%s", devname);
 293                                 }
 294                                 status |= mdmon(devname, e->devnum, 1,
 295                                                 takeover);
 296                         }
 297                 }
 298                 free_mdstat(mdstat);
 299
 300                 return status;
 301         } else if (strncmp(container_name, "md", 2) == 0) {
 302                 devnum = devname2devnum(container_name);
 303                 devname = devnum2devname(devnum);
 304                 if (strcmp(container_name, devname) != 0)
 305                         devname = NULL;
 306         } else {
 307                 struct stat st;
 308
 309                 devnum = NoMdDev;
 310                 if (stat(container_name, &st) == 0)
 311                         devnum = stat2devnum(&st);
 312                 if (devnum == NoMdDev)
 313                         devname = NULL;
 314                 else
 315                         devname = devnum2devname(devnum);
 316         }
 317
 318         if (!devname) {
 319                 fprintf(stderr, "mdmon: %s is not a valid md device name\n",
 320                         container_name);
 321                 exit(1);
 322         }
 323         return mdmon(devname, devnum, do_fork(), takeover);
 324 }
 325
 326 static int mdmon(char *devname, int devnum, int must_fork, int takeover)
 327 {
 328         int mdfd;
 329         struct mdinfo *mdi, *di;
 330         struct supertype *container;
 331         sigset_t set;
 332         struct sigaction act;
 333         int pfd[2];
 334         int status;
 335         int ignore;
 336         pid_t victim = -1;
 337         int victim_sock = -1;
 338
 339         dprintf("starting mdmon for %s\n", devname);
 340
 341         mdfd = open_dev(devnum);
 342         if (mdfd < 0) {
 343                 fprintf(stderr, "mdmon: %s: %s\n", devname,
 344                         strerror(errno));
 345                 return 1;
 346         }
 347         if (md_get_version(mdfd) < 0) {
 348                 fprintf(stderr, "mdmon: %s: Not an md device\n",
 349                         devname);
 350                 return 1;
 351         }
 352
 353         /* Fork, and have the child tell us when they are ready */
 354         if (must_fork) {
 355                 if (pipe(pfd) != 0) {
 356                         fprintf(stderr, "mdmon: failed to create pipe\n");
 357                         return 1;
 358                 }
 359                 switch(fork()) {
 360                 case -1:
 361                         fprintf(stderr, "mdmon: failed to fork: %s\n",
 362                                 strerror(errno));
 363                         return 1;
 364                 case 0: /* child */
 365                         close(pfd[0]);
 366                         break;
 367                 default: /* parent */
 368                         close(pfd[1]);
 369                         if (read(pfd[0], &status, sizeof(status)) != sizeof(status)) {
 370                                 wait(&status);
 371                                 status = WEXITSTATUS(status);
 372                         }
 373                         return status;
 374                 }
 375         } else
 376                 pfd[0] = pfd[1] = -1;
 377
 378         container = calloc(1, sizeof(*container));
 379         container->devnum = devnum;
 380         container->devname = devname;
 381         container->arrays = NULL;
 382         container->subarray[0] = 0;
 383         container->sock = -1;
 384
 385         if (!container->devname) {
 386                 fprintf(stderr, "mdmon: failed to allocate container name string\n");
 387                 exit(3);
 388         }
 389
 390         mdi = sysfs_read(mdfd, container->devnum,
 391                          GET_VERSION|GET_LEVEL|GET_DEVS|SKIP_GONE_DEVS);
 392
 393         if (!mdi) {
 394                 fprintf(stderr, "mdmon: failed to load sysfs info for %s\n",
 395                         container->devname);
 396                 exit(3);
 397         }
 398         if (mdi->array.level != UnSet) {
 399                 fprintf(stderr, "mdmon: %s is not a container - cannot monitor\n",
 400                         devname);
 401                 exit(3);
 402         }
 403         if (mdi->array.major_version != -1 ||
 404             mdi->array.minor_version != -2) {
 405                 fprintf(stderr, "mdmon: %s does not use external metadata - cannot monitor\n",
 406                         devname);
 407                 exit(3);
 408         }
 409
 410         container->ss = find_metadata_methods(mdi->text_version);
 411         if (container->ss == NULL) {
 412                 fprintf(stderr, "mdmon: %s uses unknown metadata: %s\n",
 413                         devname, mdi->text_version);
 414                 exit(3);
 415         }
 416
 417         container->devs = NULL;
 418         for (di = mdi->devs; di; di = di->next) {
 419                 struct mdinfo *cd = malloc(sizeof(*cd));
 420                 *cd = *di;
 421                 cd->next = container->devs;
 422                 container->devs = cd;
 423         }
 424         sysfs_free(mdi);
 425
 426         /* SIGUSR is sent between parent and child.  So both block it
 427          * and enable it only with pselect.
 428          */
 429         sigemptyset(&set);
 430         sigaddset(&set, SIGUSR1);
 431         sigaddset(&set, SIGTERM);
 432         sigprocmask(SIG_BLOCK, &set, NULL);
 433         act.sa_handler = wake_me;
 434         act.sa_flags = 0;
 435         sigaction(SIGUSR1, &act, NULL);
 436         act.sa_handler = term;
 437         sigaction(SIGTERM, &act, NULL);
 438         act.sa_handler = SIG_IGN;
 439         sigaction(SIGPIPE, &act, NULL);
 440
 441         if (takeover) {
 442                 pid_dir = VAR_RUN;
 443                 victim = mdmon_pid(container->devnum);
 444                 if (victim < 0) {
 445                         pid_dir = ALT_RUN;
 446                         victim = mdmon_pid(container->devnum);
 447                 }
 448                 if (victim >= 0)
 449                         victim_sock = connect_monitor(container->devname);
 450         }
 451
 452         ignore = chdir("/");
 453         if (victim < 0) {
 454                 if (ping_monitor(container->devname) == 0) {
 455                         fprintf(stderr, "mdmon: %s already managed\n",
 456                                 container->devname);
 457                         exit(3);
 458                 }
 459                 /* if there is a pid file, kill whoever is there just in case */
 460                 victim = mdmon_pid(container->devnum);
 461         }
 462         if (container->ss->load_super(container, mdfd, devname)) {
 463                 fprintf(stderr, "mdmon: Cannot load metadata for %s\n",
 464                         devname);
 465                 exit(3);
 466         }
 467         close(mdfd);
 468
 469         /* Ok, this is close enough.  We can say goodbye to our parent now.
 470          */
 471         if (victim > 0)
 472                 remove_pidfile(devname);
 473         if (mkdir(VAR_RUN, 0600) >= 0 || errno == EEXIST)
 474                 pid_dir = VAR_RUN;
 475         else if (mkdir(ALT_RUN, 0600) >= 0 || errno == EEXIST)
 476                 pid_dir = ALT_RUN;
 477         else {
 478                 fprintf(stderr, "mdmon: Neither %s nor %s are writable\n"
 479                         "       cannot create .pid or .sock files.  Aborting\n",
 480                         VAR_RUN, ALT_RUN);
 481                 exit(3);
 482         }
 483         if (make_pidfile(devname) < 0) {
 484                 fprintf(stderr, "mdmon: Cannot create pid file in %s - aborting.\n",
 485                         pid_dir);
 486                 exit(3);
 487         }
 488         container->sock = make_control_sock(devname);
 489
 490         status = 0;
 491         if (write(pfd[1], &status, sizeof(status)) < 0)
 492                 fprintf(stderr, "mdmon: failed to notify our parent: %d\n",
 493                         getppid());
 494         close(pfd[1]);
 495
 496         mlockall(MCL_CURRENT | MCL_FUTURE);
 497
 498         if (clone_monitor(container) < 0) {
 499                 fprintf(stderr, "mdmon: failed to start monitor process: %s\n",
 500                         strerror(errno));
 501                 exit(2);
 502         }
 503
 504         if (victim > 0) {
 505                 try_kill_monitor(victim, container->devname, victim_sock);
 506                 close(victim_sock);
 507         }
 508
 509         setsid();
 510         close(0);
 511         open("/dev/null", O_RDWR);
 512         close(1);
 513         ignore = dup(0);
 514 #ifndef DEBUG
 515         close(2);
 516         ignore = dup(0);
 517 #endif
 518
 519         do_manager(container);
 520
 521         exit(0);
 522 }