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
  70         rv = clone(run_child, stack+4096-64,
  71                    CLONE_FS|CLONE_FILES|CLONE_VM|CLONE_SIGHAND|CLONE_THREAD,
  72                    container);
  73         if (rv < 0)
  74                 goto err_clone;
  75         else
  76                 return rv;
  77
  78  err_clone:
  79         close(container->mgr_pipe[0]);
  80         close(container->mgr_pipe[1]);
  81
  82         return rv;
  83 }
  84
  85 static struct superswitch *find_metadata_methods(char *vers)
  86 {
  87         if (strcmp(vers, "ddf") == 0)
  88                 return &super_ddf;
  89         if (strcmp(vers, "imsm") == 0)
  90                 return &super_imsm;
  91         return NULL;
  92 }
  93
  94
  95 static int make_pidfile(char *devname, int o_excl)
  96 {
  97         char path[100];
  98         char pid[10];
  99         int fd;
 100         sprintf(path, "/var/run/mdadm/%s.pid", devname);
 101
 102         fd = open(path, O_RDWR|O_CREAT|o_excl, 0600);
 103         if (fd < 0)
 104                 return -1;
 105         sprintf(pid, "%d\n", getpid());
 106         write(fd, pid, strlen(pid));
 107         close(fd);
 108         return 0;
 109 }
 110
 111 static void try_kill_monitor(char *devname)
 112 {
 113         char buf[100];
 114         int fd;
 115         pid_t pid;
 116
 117         sprintf(buf, "/var/run/mdadm/%s.pid", devname);
 118         fd = open(buf, O_RDONLY);
 119         if (fd < 0)
 120                 return;
 121
 122         if (read(fd, buf, sizeof(buf)) < 0) {
 123                 close(fd);
 124                 return;
 125         }
 126
 127         close(fd);
 128         pid = strtoul(buf, NULL, 10);
 129
 130         /* kill this process if it is mdmon */
 131         sprintf(buf, "/proc/%lu/cmdline", (unsigned long) pid);
 132         fd = open(buf, O_RDONLY);
 133         if (fd < 0)
 134                 return;
 135
 136         if (read(fd, buf, sizeof(buf)) < 0) {
 137                 close(fd);
 138                 return;
 139         }
 140
 141         if (strstr(buf, "mdmon") != NULL)
 142                 kill(pid, SIGTERM);
 143 }
 144
 145 void remove_pidfile(char *devname)
 146 {
 147         char buf[100];
 148
 149         sprintf(buf, "/var/run/mdadm/%s.pid", devname);
 150         unlink(buf);
 151 }
 152
 153 static int make_control_sock(char *devname)
 154 {
 155         char path[100];
 156         int sfd;
 157         long fl;
 158         struct sockaddr_un addr;
 159
 160         sprintf(path, "/var/run/mdadm/%s.sock", devname);
 161         unlink(path);
 162         sfd = socket(PF_LOCAL, SOCK_STREAM, 0);
 163         if (sfd < 0)
 164                 return -1;
 165
 166         addr.sun_family = PF_LOCAL;
 167         strcpy(addr.sun_path, path);
 168         if (bind(sfd, &addr, sizeof(addr)) < 0) {
 169                 close(sfd);
 170                 return -1;
 171         }
 172         listen(sfd, 10);
 173         fl = fcntl(sfd, F_GETFL, 0);
 174         fl |= O_NONBLOCK;
 175         fcntl(sfd, F_SETFL, fl);
 176         return sfd;
 177 }
 178
 179 int main(int argc, char *argv[])
 180 {
 181         int mdfd;
 182         struct mdinfo *mdi, *di;
 183         struct supertype *container;
 184         if (argc != 2) {
 185                 fprintf(stderr, "Usage: md-manage /device/name/for/container\n");
 186                 exit(2);
 187         }
 188         mdfd = open(argv[1], O_RDWR);
 189         if (mdfd < 0) {
 190                 fprintf(stderr, "md-manage: %s: %s\n", argv[1],
 191                         strerror(errno));
 192                 exit(1);
 193         }
 194         if (md_get_version(mdfd) < 0) {
 195                 fprintf(stderr, "md-manage: %s: Not an md device\n",
 196                         argv[1]);
 197                 exit(1);
 198         }
 199
 200         /* hopefully it is a container - we'll check later */
 201
 202         container = malloc(sizeof(*container));
 203         container->devnum = fd2devnum(mdfd);
 204         container->devname = devnum2devname(container->devnum);
 205         container->device_name = argv[1];
 206
 207         /* If this fails, we hope it already exists */
 208         mkdir("/var/run/mdadm", 0600);
 209         /* pid file lives in /var/run/mdadm/mdXX.pid */
 210         if (make_pidfile(container->devname, O_EXCL) < 0) {
 211                 if (ping_monitor(container->devname) == 0) {
 212                         fprintf(stderr, "mdmon: %s already managed\n",
 213                                 container->devname);
 214                         exit(3);
 215                 } else {
 216                         /* cleanup the old monitor, this one is taking over */
 217                         try_kill_monitor(container->devname);
 218                         if (make_pidfile(container->devname, 0) < 0) {
 219                                 fprintf(stderr, "mdmon: %s Cannot create pidfile\n",
 220                                         container->devname);
 221                                 exit(3);
 222                         }
 223                 }
 224         }
 225
 226         container->sock = make_control_sock(container->devname);
 227         if (container->sock < 0) {
 228                 fprintf(stderr, "mdmon: Cannot create socket in /var/run/mdadm\n");
 229                 exit(3);
 230         }
 231         container->arrays = NULL;
 232
 233         mdi = sysfs_read(mdfd, container->devnum,
 234                          GET_VERSION|GET_LEVEL|GET_DEVS);
 235
 236         if (!mdi) {
 237                 fprintf(stderr, "mdmon: failed to load sysfs info for %s\n",
 238                         container->devname);
 239                 exit(3);
 240         }
 241         if (mdi->array.level != UnSet) {
 242                 fprintf(stderr, "mdmon: %s is not a container - cannot monitor\n",
 243                         argv[1]);
 244                 exit(3);
 245         }
 246         if (mdi->array.major_version != -1 ||
 247             mdi->array.minor_version != -2) {
 248                 fprintf(stderr, "mdmon: %s does not use external metadata - cannot monitor\n",
 249                         argv[1]);
 250                 exit(3);
 251         }
 252
 253         container->ss = find_metadata_methods(mdi->text_version);
 254         if (container->ss == NULL) {
 255                 fprintf(stderr, "mdmon: %s uses unknown metadata: %s\n",
 256                         argv[1], mdi->text_version);
 257                 exit(3);
 258         }
 259
 260         container->devs = NULL;
 261         for (di = mdi->devs; di; di = di->next) {
 262                 struct mdinfo *cd = malloc(sizeof(*cd));
 263                 cd = di;
 264                 cd->next = container->devs;
 265                 container->devs = cd;
 266         }
 267         sysfs_free(mdi);
 268
 269
 270         if (container->ss->load_super(container, mdfd, argv[1])) {
 271                 fprintf(stderr, "mdmon: Cannot load metadata for %s\n",
 272                         argv[1]);
 273                 exit(3);
 274         }
 275         close(mdfd);
 276         close(mdfd);
 277
 278         mlockall(MCL_FUTURE);
 279
 280         if (clone_monitor(container) < 0) {
 281                 fprintf(stderr, "md-manage: failed to start monitor process: %s\n",
 282                         strerror(errno));
 283                 exit(2);
 284         }
 285
 286         do_manager(container);
 287
 288         exit(0);
 289 }