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