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 *array_list;
  45 struct active_array *discard_this;
  46 struct active_array *pending_discard;
  47
  48 int run_child(void *v)
  49 {
  50         struct supertype *c = v;
  51         do_monitor(c);
  52         return 0;
  53 }
  54
  55 int clone_monitor(struct supertype *container)
  56 {
  57         int pfd[2];
  58         static char stack[4096];
  59         int rv;
  60
  61         pipe(container->pipe);
  62
  63         rv = clone(run_child, stack+4096-64,
  64                    CLONE_FS|CLONE_FILES|CLONE_VM|CLONE_SIGHAND|CLONE_THREAD,
  65                    container);
  66
  67         if (rv < 0)
  68                 return rv;
  69         return pfd[1];
  70 }
  71
  72 static struct superswitch *find_metadata_methods(char *vers)
  73 {
  74         if (strcmp(vers, "ddf") == 0)
  75                 return &super_ddf;
  76         return NULL;
  77 }
  78
  79
  80 static int make_pidfile(char *devname, int o_excl)
  81 {
  82         char path[100];
  83         char pid[10];
  84         int fd;
  85         sprintf(path, "/var/run/mdadm/%s.pid", devname);
  86
  87         fd = open(path, O_RDWR|O_CREAT|o_excl, 0600);
  88         if (fd < 0)
  89                 return -1;
  90         sprintf(pid, "%d\n", getpid());
  91         write(fd, pid, strlen(pid));
  92         close(fd);
  93         return 0;
  94 }
  95
  96 static void try_kill_monitor(char *devname)
  97 {
  98         char buf[100];
  99         int fd;
 100         pid_t pid;
 101
 102         sprintf(buf, "/var/run/mdadm/%s.pid", devname);
 103         fd = open(buf, O_RDONLY);
 104         if (fd < 0)
 105                 return;
 106
 107         if (read(fd, buf, sizeof(buf)) < 0) {
 108                 close(fd);
 109                 return;
 110         }
 111
 112         close(fd);
 113         pid = strtoul(buf, NULL, 10);
 114
 115         /* kill this process if it is mdmon */
 116         sprintf(buf, "/proc/%lu/cmdline", (unsigned long) pid);
 117         fd = open(buf, O_RDONLY);
 118         if (fd < 0)
 119                 return;
 120
 121         if (read(fd, buf, sizeof(buf)) < 0) {
 122                 close(fd);
 123                 return;
 124         }
 125
 126         if (strstr(buf, "mdmon") != NULL)
 127                 kill(pid, SIGTERM);
 128 }
 129
 130 static int make_control_sock(char *devname)
 131 {
 132         char path[100];
 133         int sfd;
 134         long fl;
 135         struct sockaddr_un addr;
 136
 137         sprintf(path, "/var/run/mdadm/%s.sock", devname);
 138         unlink(path);
 139         sfd = socket(PF_LOCAL, SOCK_STREAM, 0);
 140         if (sfd < 0)
 141                 return -1;
 142
 143         addr.sun_family = PF_LOCAL;
 144         strcpy(addr.sun_path, path);
 145         if (bind(sfd, &addr, sizeof(addr)) < 0) {
 146                 close(sfd);
 147                 return -1;
 148         }
 149         listen(sfd, 10);
 150         fl = fcntl(sfd, F_GETFL, 0);
 151         fl |= O_NONBLOCK;
 152         fcntl(sfd, F_SETFL, fl);
 153         return sfd;
 154 }
 155
 156 int main(int argc, char *argv[])
 157 {
 158         int mdfd;
 159         int pipefd;
 160         struct mdinfo *mdi, *di;
 161         struct supertype *container;
 162         if (argc != 2) {
 163                 fprintf(stderr, "Usage: md-manage /device/name/for/container\n");
 164                 exit(2);
 165         }
 166         mdfd = open(argv[1], O_RDWR);
 167         if (mdfd < 0) {
 168                 fprintf(stderr, "md-manage: %s: %s\n", argv[1],
 169                         strerror(errno));
 170                 exit(1);
 171         }
 172         if (md_get_version(mdfd) < 0) {
 173                 fprintf(stderr, "md-manage: %s: Not an md device\n",
 174                         argv[1]);
 175                 exit(1);
 176         }
 177
 178         /* hopefully it is a container - we'll check later */
 179
 180         container = malloc(sizeof(*container));
 181         container->devfd = mdfd;
 182         container->devnum = fd2devnum(mdfd);
 183         container->devname = devnum2devname(container->devnum);
 184
 185         /* If this fails, we hope it already exists */
 186         mkdir("/var/run/mdadm", 0600);
 187         /* pid file lives in /var/run/mdadm/mdXX.pid */
 188         if (make_pidfile(container->devname, O_EXCL) < 0) {
 189                 if (ping_monitor(container->devname) == 0) {
 190                         fprintf(stderr, "mdmon: %s already managed\n",
 191                                 container->devname);
 192                         exit(3);
 193                 } else {
 194                         /* cleanup the old monitor, this one is taking over */
 195                         try_kill_monitor(container->devname);
 196                         if (make_pidfile(container->devname, 0) < 0) {
 197                                 fprintf(stderr, "mdmon: %s Cannot create pidfile\n",
 198                                         container->devname);
 199                                 exit(3);
 200                         }
 201                 }
 202         }
 203
 204         container->sock = make_control_sock(container->devname);
 205         if (container->sock < 0) {
 206                 fprintf(stderr, "mdmon: Cannot create socket in /var/run/mdadm\n");
 207                 exit(3);
 208         }
 209         container->arrays = NULL;
 210
 211         mdi = sysfs_read(mdfd, container->devnum,
 212                          GET_VERSION|GET_LEVEL|GET_DEVS);
 213
 214         if (!mdi) {
 215                 fprintf(stderr, "mdmon: failed to load sysfs info for %s\n",
 216                         container->devname);
 217                 exit(3);
 218         }
 219         if (mdi->array.level != UnSet) {
 220                 fprintf(stderr, "mdmon: %s is not a container - cannot monitor\n",
 221                         argv[1]);
 222                 exit(3);
 223         }
 224         if (mdi->array.major_version != -1 ||
 225             mdi->array.minor_version != -2) {
 226                 fprintf(stderr, "mdmon: %s does not use external metadata - cannot monitor\n",
 227                         argv[1]);
 228                 exit(3);
 229         }
 230
 231         container->ss = find_metadata_methods(mdi->text_version);
 232         if (container->ss == NULL) {
 233                 fprintf(stderr, "mdmon: %s uses unknown metadata: %s\n",
 234                         argv[1], mdi->text_version);
 235                 exit(3);
 236         }
 237
 238         container->devs = NULL;
 239         for (di = mdi->devs; di; di = di->next) {
 240                 struct mdinfo *cd = malloc(sizeof(*cd));
 241                 cd = di;
 242                 cd->next = container->devs;
 243                 container->devs = cd;
 244         }
 245         sysfs_free(mdi);
 246
 247
 248         if (container->ss->load_super(container, mdfd, argv[1])) {
 249                 fprintf(stderr, "mdmon: Cannot load metadata for %s\n",
 250                         argv[1]);
 251                 exit(3);
 252         }
 253
 254
 255         mlockall(MCL_FUTURE);
 256
 257         pipefd = clone_monitor(container);
 258         if (pipefd < 0) {
 259                 fprintf(stderr, "md-manage: failed to start monitor process: %s\n",
 260                         strerror(errno));
 261                 exit(2);
 262         }
 263
 264         do_manager(container);
 265
 266         exit(0);
 267 }