mdmon.c

   1
   2 /*
   3  * md array manager.
   4  * When md arrays have user-space managed metadata, this is the program
   5  * that does the managing.
   6  *
   7  * Given one argument: the name of the array (e.g. /dev/md0) that is
   8  * the container.
   9  * We fork off a helper that runs high priority and mlocked.  It responds to
  10  * device failures and other events that might stop writeout, or that are
  11  * trivial to deal with.
  12  * The main thread then watches for new arrays being created in the container
  13  * and starts monitoring them too ... along with a few other tasks.
  14  *
  15  * The main thread communicates with the priority thread by writing over
  16  * a pipe.
  17  * Separate programs can communicate with the main thread via Unix-domain
  18  * socket.
  19  * The two threads share address space and open file table.
  20  *
  21  */
  22
  23 #ifndef _GNU_SOURCE
  24 #define _GNU_SOURCE
  25 #endif
  26
  27 #include        <unistd.h>
  28 #include        <stdlib.h>
  29 #include        <sys/stat.h>
  30 #include        <sys/socket.h>
  31 #include        <sys/un.h>
  32 #include        <sys/mman.h>
  33 #include        <stdio.h>
  34 #include        <errno.h>
  35 #include        <string.h>
  36 #include        <fcntl.h>
  37 #include        <signal.h>
  38
  39 #include        <sched.h>
  40
  41 #include        "mdadm.h"
  42 #include        "mdmon.h"
  43
  44 struct active_array *discard_this;
  45 struct active_array *pending_discard;
  46 struct md_generic_cmd *active_cmd;
  47
  48 int run_child(void *v)
  49 {
  50         struct supertype *c = v;
  51         sigset_t set;
  52         /* SIGUSR is sent from child to parent,  So child must block it */
  53         sigemptyset(&set);
  54         sigaddset(&set, SIGUSR1);
  55         sigprocmask(SIG_BLOCK, &set, NULL);
  56
  57         do_monitor(c);
  58         return 0;
  59 }
  60
  61 int clone_monitor(struct supertype *container)
  62 {
  63         static char stack[4096];
  64         int rv;
  65
  66         rv = pipe(container->mgr_pipe);
  67         if (rv < 0)
  68                 return rv;
  69         rv = pipe(container->mon_pipe);
  70         if (rv < 0)
  71                 goto err_mon_pipe;
  72
  73         rv = clone(run_child, stack+4096-64,
  74                    CLONE_FS|CLONE_FILES|CLONE_VM|CLONE_SIGHAND|CLONE_THREAD,
  75                    container);
  76         if (rv < 0)
  77                 goto err_clone;
  78         else
  79                 return rv;
  80
  81  err_clone:
  82         close(container->mon_pipe[0]);
  83         close(container->mon_pipe[1]);
  84  err_mon_pipe:
  85         close(container->mgr_pipe[0]);
  86         close(container->mgr_pipe[1]);
  87
  88         return rv;
  89 }
  90
  91 static struct superswitch *find_metadata_methods(char *vers)
  92 {
  93         if (strcmp(vers, "ddf") == 0)
  94                 return &super_ddf;
  95         if (strcmp(vers, "imsm") == 0)
  96                 return &super_imsm;
  97         return NULL;
  98 }
  99
 100
 101 static int make_pidfile(char *devname, int o_excl)
 102 {
 103         char path[100];
 104         char pid[10];
 105         int fd;
 106         sprintf(path, "/var/run/mdadm/%s.pid", devname);
 107
 108         fd = open(path, O_RDWR|O_CREAT|o_excl, 0600);
 109         if (fd < 0)
 110                 return -1;
 111         sprintf(pid, "%d\n", getpid());
 112         write(fd, pid, strlen(pid));
 113         close(fd);
 114         return 0;
 115 }
 116
 117 static void try_kill_monitor(char *devname)
 118 {
 119         char buf[100];
 120         int fd;
 121         pid_t pid;
 122
 123         sprintf(buf, "/var/run/mdadm/%s.pid", devname);
 124         fd = open(buf, O_RDONLY);
 125         if (fd < 0)
 126                 return;
 127
 128         if (read(fd, buf, sizeof(buf)) < 0) {
 129                 close(fd);
 130                 return;
 131         }
 132
 133         close(fd);
 134         pid = strtoul(buf, NULL, 10);
 135
 136         /* kill this process if it is mdmon */
 137         sprintf(buf, "/proc/%lu/cmdline", (unsigned long) pid);
 138         fd = open(buf, O_RDONLY);
 139         if (fd < 0)
 140                 return;
 141
 142         if (read(fd, buf, sizeof(buf)) < 0) {
 143                 close(fd);
 144                 return;
 145         }
 146
 147         if (strstr(buf, "mdmon") != NULL)
 148                 kill(pid, SIGTERM);
 149 }
 150
 151 void remove_pidfile(char *devname)
 152 {
 153         char buf[100];
 154
 155         sprintf(buf, "/var/run/mdadm/%s.pid", devname);
 156         unlink(buf);
 157 }
 158
 159 static int make_control_sock(char *devname)
 160 {
 161         char path[100];
 162         int sfd;
 163         long fl;
 164         struct sockaddr_un addr;
 165
 166         sprintf(path, "/var/run/mdadm/%s.sock", devname);
 167         unlink(path);
 168         sfd = socket(PF_LOCAL, SOCK_STREAM, 0);
 169         if (sfd < 0)
 170                 return -1;
 171
 172         addr.sun_family = PF_LOCAL;
 173         strcpy(addr.sun_path, path);
 174         if (bind(sfd, &addr, sizeof(addr)) < 0) {
 175                 close(sfd);
 176                 return -1;
 177         }
 178         listen(sfd, 10);
 179         fl = fcntl(sfd, F_GETFL, 0);
 180         fl |= O_NONBLOCK;
 181         fcntl(sfd, F_SETFL, fl);
 182         return sfd;
 183 }
 184
 185 int main(int argc, char *argv[])
 186 {
 187         int mdfd;
 188         struct mdinfo *mdi, *di;
 189         struct supertype *container;
 190         if (argc != 2) {
 191                 fprintf(stderr, "Usage: md-manage /device/name/for/container\n");
 192                 exit(2);
 193         }
 194         mdfd = open(argv[1], O_RDWR);
 195         if (mdfd < 0) {
 196                 fprintf(stderr, "md-manage: %s: %s\n", argv[1],
 197                         strerror(errno));
 198                 exit(1);
 199         }
 200         if (md_get_version(mdfd) < 0) {
 201                 fprintf(stderr, "md-manage: %s: Not an md device\n",
 202                         argv[1]);
 203                 exit(1);
 204         }
 205
 206         /* hopefully it is a container - we'll check later */
 207
 208         container = malloc(sizeof(*container));
 209         container->devnum = fd2devnum(mdfd);
 210         container->devname = devnum2devname(container->devnum);
 211         container->device_name = argv[1];
 212
 213         /* If this fails, we hope it already exists */
 214         mkdir("/var/run/mdadm", 0600);
 215         /* pid file lives in /var/run/mdadm/mdXX.pid */
 216         if (make_pidfile(container->devname, O_EXCL) < 0) {
 217                 if (ping_monitor(container->devname) == 0) {
 218                         fprintf(stderr, "mdmon: %s already managed\n",
 219                                 container->devname);
 220                         exit(3);
 221                 } else {
 222                         /* cleanup the old monitor, this one is taking over */
 223                         try_kill_monitor(container->devname);
 224                         if (make_pidfile(container->devname, 0) < 0) {
 225                                 fprintf(stderr, "mdmon: %s Cannot create pidfile\n",
 226                                         container->devname);
 227                                 exit(3);
 228                         }
 229                 }
 230         }
 231
 232         container->sock = make_control_sock(container->devname);
 233         if (container->sock < 0) {
 234                 fprintf(stderr, "mdmon: Cannot create socket in /var/run/mdadm\n");
 235                 exit(3);
 236         }
 237         container->arrays = NULL;
 238
 239         mdi = sysfs_read(mdfd, container->devnum,
 240                          GET_VERSION|GET_LEVEL|GET_DEVS);
 241
 242         if (!mdi) {
 243                 fprintf(stderr, "mdmon: failed to load sysfs info for %s\n",
 244                         container->devname);
 245                 exit(3);
 246         }
 247         if (mdi->array.level != UnSet) {
 248                 fprintf(stderr, "mdmon: %s is not a container - cannot monitor\n",
 249                         argv[1]);
 250                 exit(3);
 251         }
 252         if (mdi->array.major_version != -1 ||
 253             mdi->array.minor_version != -2) {
 254                 fprintf(stderr, "mdmon: %s does not use external metadata - cannot monitor\n",
 255                         argv[1]);
 256                 exit(3);
 257         }
 258
 259         container->ss = find_metadata_methods(mdi->text_version);
 260         if (container->ss == NULL) {
 261                 fprintf(stderr, "mdmon: %s uses unknown metadata: %s\n",
 262                         argv[1], mdi->text_version);
 263                 exit(3);
 264         }
 265
 266         container->devs = NULL;
 267         for (di = mdi->devs; di; di = di->next) {
 268                 struct mdinfo *cd = malloc(sizeof(*cd));
 269                 cd = di;
 270                 cd->next = container->devs;
 271                 container->devs = cd;
 272         }
 273         sysfs_free(mdi);
 274
 275
 276         if (container->ss->load_super(container, mdfd, argv[1])) {
 277                 fprintf(stderr, "mdmon: Cannot load metadata for %s\n",
 278                         argv[1]);
 279                 exit(3);
 280         }
 281         close(mdfd);
 282         close(mdfd);
 283
 284         mlockall(MCL_FUTURE);
 285
 286         if (clone_monitor(container) < 0) {
 287                 fprintf(stderr, "md-manage: failed to start monitor process: %s\n",
 288                         strerror(errno));
 289                 exit(2);
 290         }
 291
 292         do_manager(container);
 293
 294         exit(0);
 295 }