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