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         do_monitor(c);
  52         return 0;
  53 }
  54
  55 int clone_monitor(struct supertype *container)
  56 {
  57         static char stack[4096];
  58         int rv;
  59
  60         rv = pipe(container->mgr_pipe);
  61         if (rv < 0)
  62                 return rv;
  63         rv = pipe(container->mon_pipe);
  64         if (rv < 0)
  65                 goto err_mon_pipe;
  66
  67         rv = clone(run_child, stack+4096-64,
  68                    CLONE_FS|CLONE_FILES|CLONE_VM|CLONE_SIGHAND|CLONE_THREAD,
  69                    container);
  70         if (rv < 0)
  71                 goto err_clone;
  72         else
  73                 return rv;
  74
  75  err_clone:
  76         close(container->mon_pipe[0]);
  77         close(container->mon_pipe[1]);
  78  err_mon_pipe:
  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         return NULL;
  90 }
  91
  92
  93 static int make_pidfile(char *devname, int o_excl)
  94 {
  95         char path[100];
  96         char pid[10];
  97         int fd;
  98         sprintf(path, "/var/run/mdadm/%s.pid", devname);
  99
 100         fd = open(path, O_RDWR|O_CREAT|o_excl, 0600);
 101         if (fd < 0)
 102                 return -1;
 103         sprintf(pid, "%d\n", getpid());
 104         write(fd, pid, strlen(pid));
 105         close(fd);
 106         return 0;
 107 }
 108
 109 static void try_kill_monitor(char *devname)
 110 {
 111         char buf[100];
 112         int fd;
 113         pid_t pid;
 114
 115         sprintf(buf, "/var/run/mdadm/%s.pid", devname);
 116         fd = open(buf, O_RDONLY);
 117         if (fd < 0)
 118                 return;
 119
 120         if (read(fd, buf, sizeof(buf)) < 0) {
 121                 close(fd);
 122                 return;
 123         }
 124
 125         close(fd);
 126         pid = strtoul(buf, NULL, 10);
 127
 128         /* kill this process if it is mdmon */
 129         sprintf(buf, "/proc/%lu/cmdline", (unsigned long) pid);
 130         fd = open(buf, O_RDONLY);
 131         if (fd < 0)
 132                 return;
 133
 134         if (read(fd, buf, sizeof(buf)) < 0) {
 135                 close(fd);
 136                 return;
 137         }
 138
 139         if (strstr(buf, "mdmon") != NULL)
 140                 kill(pid, SIGTERM);
 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 int main(int argc, char *argv[])
 170 {
 171         int mdfd;
 172         struct mdinfo *mdi, *di;
 173         struct supertype *container;
 174         if (argc != 2) {
 175                 fprintf(stderr, "Usage: md-manage /device/name/for/container\n");
 176                 exit(2);
 177         }
 178         mdfd = open(argv[1], O_RDWR);
 179         if (mdfd < 0) {
 180                 fprintf(stderr, "md-manage: %s: %s\n", argv[1],
 181                         strerror(errno));
 182                 exit(1);
 183         }
 184         if (md_get_version(mdfd) < 0) {
 185                 fprintf(stderr, "md-manage: %s: Not an md device\n",
 186                         argv[1]);
 187                 exit(1);
 188         }
 189
 190         /* hopefully it is a container - we'll check later */
 191
 192         container = malloc(sizeof(*container));
 193         container->devfd = mdfd;
 194         container->devnum = fd2devnum(mdfd);
 195         container->devname = devnum2devname(container->devnum);
 196
 197         /* If this fails, we hope it already exists */
 198         mkdir("/var/run/mdadm", 0600);
 199         /* pid file lives in /var/run/mdadm/mdXX.pid */
 200         if (make_pidfile(container->devname, O_EXCL) < 0) {
 201                 if (ping_monitor(container->devname) == 0) {
 202                         fprintf(stderr, "mdmon: %s already managed\n",
 203                                 container->devname);
 204                         exit(3);
 205                 } else {
 206                         /* cleanup the old monitor, this one is taking over */
 207                         try_kill_monitor(container->devname);
 208                         if (make_pidfile(container->devname, 0) < 0) {
 209                                 fprintf(stderr, "mdmon: %s Cannot create pidfile\n",
 210                                         container->devname);
 211                                 exit(3);
 212                         }
 213                 }
 214         }
 215
 216         container->sock = make_control_sock(container->devname);
 217         if (container->sock < 0) {
 218                 fprintf(stderr, "mdmon: Cannot create socket in /var/run/mdadm\n");
 219                 exit(3);
 220         }
 221         container->arrays = NULL;
 222
 223         mdi = sysfs_read(mdfd, container->devnum,
 224                          GET_VERSION|GET_LEVEL|GET_DEVS);
 225
 226         if (!mdi) {
 227                 fprintf(stderr, "mdmon: failed to load sysfs info for %s\n",
 228                         container->devname);
 229                 exit(3);
 230         }
 231         if (mdi->array.level != UnSet) {
 232                 fprintf(stderr, "mdmon: %s is not a container - cannot monitor\n",
 233                         argv[1]);
 234                 exit(3);
 235         }
 236         if (mdi->array.major_version != -1 ||
 237             mdi->array.minor_version != -2) {
 238                 fprintf(stderr, "mdmon: %s does not use external metadata - cannot monitor\n",
 239                         argv[1]);
 240                 exit(3);
 241         }
 242
 243         container->ss = find_metadata_methods(mdi->text_version);
 244         if (container->ss == NULL) {
 245                 fprintf(stderr, "mdmon: %s uses unknown metadata: %s\n",
 246                         argv[1], mdi->text_version);
 247                 exit(3);
 248         }
 249
 250         container->devs = NULL;
 251         for (di = mdi->devs; di; di = di->next) {
 252                 struct mdinfo *cd = malloc(sizeof(*cd));
 253                 cd = di;
 254                 cd->next = container->devs;
 255                 container->devs = cd;
 256         }
 257         sysfs_free(mdi);
 258
 259
 260         if (container->ss->load_super(container, mdfd, argv[1])) {
 261                 fprintf(stderr, "mdmon: Cannot load metadata for %s\n",
 262                         argv[1]);
 263                 exit(3);
 264         }
 265
 266
 267         mlockall(MCL_FUTURE);
 268
 269         if (clone_monitor(container) < 0) {
 270                 fprintf(stderr, "md-manage: failed to start monitor process: %s\n",
 271                         strerror(errno));
 272                 exit(2);
 273         }
 274
 275         do_manager(container);
 276
 277         exit(0);
 278 }