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 struct md_generic_cmd *active_cmd;
  48
  49 int run_child(void *v)
  50 {
  51         struct supertype *c = v;
  52         do_monitor(c);
  53         return 0;
  54 }
  55
  56 int clone_monitor(struct supertype *container)
  57 {
  58         static char stack[4096];
  59         int rv;
  60
  61         rv = pipe(container->mgr_pipe);
  62         if (rv < 0)
  63                 return rv;
  64         rv = pipe(container->mon_pipe);
  65         if (rv < 0)
  66                 goto err_mon_pipe;
  67
  68         rv = clone(run_child, stack+4096-64,
  69                    CLONE_FS|CLONE_FILES|CLONE_VM|CLONE_SIGHAND|CLONE_THREAD,
  70                    container);
  71         if (rv < 0)
  72                 goto err_clone;
  73         else
  74                 return rv;
  75
  76  err_clone:
  77         close(container->mon_pipe[0]);
  78         close(container->mon_pipe[1]);
  79  err_mon_pipe:
  80         close(container->mgr_pipe[0]);
  81         close(container->mgr_pipe[1]);
  82
  83         return rv;
  84 }
  85
  86 static struct superswitch *find_metadata_methods(char *vers)
  87 {
  88         if (strcmp(vers, "ddf") == 0)
  89                 return &super_ddf;
  90         return NULL;
  91 }
  92
  93
  94 static int make_pidfile(char *devname, int o_excl)
  95 {
  96         char path[100];
  97         char pid[10];
  98         int fd;
  99         sprintf(path, "/var/run/mdadm/%s.pid", devname);
 100
 101         fd = open(path, O_RDWR|O_CREAT|o_excl, 0600);
 102         if (fd < 0)
 103                 return -1;
 104         sprintf(pid, "%d\n", getpid());
 105         write(fd, pid, strlen(pid));
 106         close(fd);
 107         return 0;
 108 }
 109
 110 static void try_kill_monitor(char *devname)
 111 {
 112         char buf[100];
 113         int fd;
 114         pid_t pid;
 115
 116         sprintf(buf, "/var/run/mdadm/%s.pid", devname);
 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         close(fd);
 127         pid = strtoul(buf, NULL, 10);
 128
 129         /* kill this process if it is mdmon */
 130         sprintf(buf, "/proc/%lu/cmdline", (unsigned long) pid);
 131         fd = open(buf, O_RDONLY);
 132         if (fd < 0)
 133                 return;
 134
 135         if (read(fd, buf, sizeof(buf)) < 0) {
 136                 close(fd);
 137                 return;
 138         }
 139
 140         if (strstr(buf, "mdmon") != NULL)
 141                 kill(pid, SIGTERM);
 142 }
 143
 144 static int make_control_sock(char *devname)
 145 {
 146         char path[100];
 147         int sfd;
 148         long fl;
 149         struct sockaddr_un addr;
 150
 151         sprintf(path, "/var/run/mdadm/%s.sock", devname);
 152         unlink(path);
 153         sfd = socket(PF_LOCAL, SOCK_STREAM, 0);
 154         if (sfd < 0)
 155                 return -1;
 156
 157         addr.sun_family = PF_LOCAL;
 158         strcpy(addr.sun_path, path);
 159         if (bind(sfd, &addr, sizeof(addr)) < 0) {
 160                 close(sfd);
 161                 return -1;
 162         }
 163         listen(sfd, 10);
 164         fl = fcntl(sfd, F_GETFL, 0);
 165         fl |= O_NONBLOCK;
 166         fcntl(sfd, F_SETFL, fl);
 167         return sfd;
 168 }
 169
 170 int main(int argc, char *argv[])
 171 {
 172         int mdfd;
 173         struct mdinfo *mdi, *di;
 174         struct supertype *container;
 175         if (argc != 2) {
 176                 fprintf(stderr, "Usage: md-manage /device/name/for/container\n");
 177                 exit(2);
 178         }
 179         mdfd = open(argv[1], O_RDWR);
 180         if (mdfd < 0) {
 181                 fprintf(stderr, "md-manage: %s: %s\n", argv[1],
 182                         strerror(errno));
 183                 exit(1);
 184         }
 185         if (md_get_version(mdfd) < 0) {
 186                 fprintf(stderr, "md-manage: %s: Not an md device\n",
 187                         argv[1]);
 188                 exit(1);
 189         }
 190
 191         /* hopefully it is a container - we'll check later */
 192
 193         container = malloc(sizeof(*container));
 194         container->devfd = mdfd;
 195         container->devnum = fd2devnum(mdfd);
 196         container->devname = devnum2devname(container->devnum);
 197
 198         /* If this fails, we hope it already exists */
 199         mkdir("/var/run/mdadm", 0600);
 200         /* pid file lives in /var/run/mdadm/mdXX.pid */
 201         if (make_pidfile(container->devname, O_EXCL) < 0) {
 202                 if (ping_monitor(container->devname) == 0) {
 203                         fprintf(stderr, "mdmon: %s already managed\n",
 204                                 container->devname);
 205                         exit(3);
 206                 } else {
 207                         /* cleanup the old monitor, this one is taking over */
 208                         try_kill_monitor(container->devname);
 209                         if (make_pidfile(container->devname, 0) < 0) {
 210                                 fprintf(stderr, "mdmon: %s Cannot create pidfile\n",
 211                                         container->devname);
 212                                 exit(3);
 213                         }
 214                 }
 215         }
 216
 217         container->sock = make_control_sock(container->devname);
 218         if (container->sock < 0) {
 219                 fprintf(stderr, "mdmon: Cannot create socket in /var/run/mdadm\n");
 220                 exit(3);
 221         }
 222         container->arrays = NULL;
 223
 224         mdi = sysfs_read(mdfd, container->devnum,
 225                          GET_VERSION|GET_LEVEL|GET_DEVS);
 226
 227         if (!mdi) {
 228                 fprintf(stderr, "mdmon: failed to load sysfs info for %s\n",
 229                         container->devname);
 230                 exit(3);
 231         }
 232         if (mdi->array.level != UnSet) {
 233                 fprintf(stderr, "mdmon: %s is not a container - cannot monitor\n",
 234                         argv[1]);
 235                 exit(3);
 236         }
 237         if (mdi->array.major_version != -1 ||
 238             mdi->array.minor_version != -2) {
 239                 fprintf(stderr, "mdmon: %s does not use external metadata - cannot monitor\n",
 240                         argv[1]);
 241                 exit(3);
 242         }
 243
 244         container->ss = find_metadata_methods(mdi->text_version);
 245         if (container->ss == NULL) {
 246                 fprintf(stderr, "mdmon: %s uses unknown metadata: %s\n",
 247                         argv[1], mdi->text_version);
 248                 exit(3);
 249         }
 250
 251         container->devs = NULL;
 252         for (di = mdi->devs; di; di = di->next) {
 253                 struct mdinfo *cd = malloc(sizeof(*cd));
 254                 cd = di;
 255                 cd->next = container->devs;
 256                 container->devs = cd;
 257         }
 258         sysfs_free(mdi);
 259
 260
 261         if (container->ss->load_super(container, mdfd, argv[1])) {
 262                 fprintf(stderr, "mdmon: Cannot load metadata for %s\n",
 263                         argv[1]);
 264                 exit(3);
 265         }
 266
 267
 268         mlockall(MCL_FUTURE);
 269
 270         if (clone_monitor(container) < 0) {
 271                 fprintf(stderr, "md-manage: failed to start monitor process: %s\n",
 272                         strerror(errno));
 273                 exit(2);
 274         }
 275
 276         do_manager(container);
 277
 278         exit(0);
 279 }