managemon.c

   1
   2 /*
   3  * The management thread for monitoring active md arrays.
   4  * This thread does things which might block such as memory
   5  * allocation.
   6  * In particular:
   7  *
   8  * - Find out about new arrays in this container.
   9  *   Allocate the data structures and open the files.
  10  *
  11  *   For this we watch /proc/mdstat and find new arrays with
  12  *   metadata type that confirms sharing. e.g. "md4"
  13  *   When we find a new array we slip it into the list of
  14  *   arrays and signal 'monitor' by writing to a pipe.
  15  *
  16  * - Respond to reshape requests by allocating new data structures
  17  *   and opening new files.
  18  *
  19  *   These come as a change to raid_disks.  We allocate a new
  20  *   version of the data structures and slip it into the list.
  21  *   'monitor' will notice and release the old version.
  22  *   Changes to level, chunksize, layout.. do not need re-allocation.
  23  *   Reductions in raid_disks don't really either, but we handle
  24  *   them the same way for consistency.
  25  *
  26  * - When a device is added to the container, we add it to the metadata
  27  *   as a spare.
  28  *
  29  * - assist with activating spares by opening relevant sysfs file.
  30  *
  31  * - Pass on metadata updates from external programs such as
  32  *   mdadm creating a new array.
  33  *
  34  *   This is most-messy.
  35  *   It might involve adding a new array or changing the status of
  36  *   a spare, or any reconfig that the kernel doesn't get involved in.
  37  *
  38  *   The required updates are received via a named pipe.  There will
  39  *   be one named pipe for each container. Each message contains a
  40  *   sync marker: 0x5a5aa5a5, A byte count, and the message.  This is
  41  *   passed to the metadata handler which will interpret and process it.
  42  *   For 'DDF' messages are internal data blocks with the leading
  43  *   'magic number' signifying what sort of data it is.
  44  *
  45  */
  46
  47 /*
  48  * We select on /proc/mdstat and the named pipe.
  49  * We create new arrays or updated version of arrays and slip
  50  * them into the head of the list, then signal 'monitor' via a pipe write.
  51  * 'monitor' will notice and place the old array on a return list.
  52  * Metadata updates are placed on a queue just like they arrive
  53  * from the named pipe.
  54  *
  55  * When new arrays are found based on correct metadata string, we
  56  * need to identify them with an entry in the metadata.  Maybe we require
  57  * the metadata to be mdX/NN  when NN is the index into an appropriate table.
  58  *
  59  */
  60
  61 /*
  62  * List of tasks:
  63  * - Watch for spares to be added to the container, and write updated
  64  *   metadata to them.
  65  * - Watch for new arrays using this container, confirm they match metadata
  66  *   and if so, start monitoring them
  67  * - Watch for spares being added to monitored arrays.  This shouldn't
  68  *   happen, as we should do all the adding.  Just remove them.
  69  * - Watch for change in raid-disks, chunk-size, etc.  Update metadata and
  70  *   start a reshape.
  71  */
  72 #ifndef _GNU_SOURCE
  73 #define _GNU_SOURCE
  74 #endif
  75 #include        "mdadm.h"
  76 #include        "mdmon.h"
  77 #include        <sys/socket.h>
  78
  79 static void close_aa(struct active_array *aa)
  80 {
  81         struct mdinfo *d;
  82
  83         for (d = aa->info.devs; d; d = d->next)
  84                 close(d->state_fd);
  85
  86         close(aa->action_fd);
  87         close(aa->info.state_fd);
  88         close(aa->resync_start_fd);
  89         close(aa->sync_pos_fd);
  90 }
  91
  92 static void free_aa(struct active_array *aa)
  93 {
  94         /* Note that this doesn't close fds if they are being used
  95          * by a clone.  ->container will be set for a clone
  96          */
  97         if (!aa->container)
  98                 close_aa(aa);
  99         while (aa->info.devs) {
 100                 struct mdinfo *d = aa->info.devs;
 101                 aa->info.devs = d->next;
 102                 free(d);
 103         }
 104         free(aa);
 105 }
 106
 107 static void write_wakeup(struct supertype *c)
 108 {
 109         static struct md_generic_cmd cmd = { .action = md_action_ping_monitor };
 110         int err;
 111
 112         active_cmd = &cmd;
 113
 114         /* send the monitor thread a pointer to the ping action */
 115         write(c->mgr_pipe[1], &err, 1);
 116         read(c->mon_pipe[0], &err, 1);
 117 }
 118
 119 static void replace_array(struct supertype *container,
 120                           struct active_array *old,
 121                           struct active_array *new)
 122 {
 123         /* To replace an array, we add it to the top of the list
 124          * marked with ->replaces to point to the original.
 125          * 'monitor' will take the original out of the list
 126          * and put it on 'discard_this'.  We take it from there
 127          * and discard it.
 128          */
 129
 130         while (pending_discard) {
 131                 while (discard_this == NULL)
 132                         sleep(1);
 133                 if (discard_this != pending_discard)
 134                         abort();
 135                 discard_this->next = NULL;
 136                 free_aa(discard_this);
 137                 discard_this = NULL;
 138                 pending_discard = NULL;
 139         }
 140         pending_discard = old;
 141         new->replaces = old;
 142         new->next = container->arrays;
 143         container->arrays = new;
 144         write_wakeup(container);
 145 }
 146
 147
 148 static void manage_container(struct mdstat_ent *mdstat,
 149                              struct supertype *container)
 150 {
 151         /* The only thing of interest here is if a new device
 152          * has been added to the container.  We add it to the
 153          * array ignoring any metadata on it.
 154          * FIXME should we look for compatible metadata and take hints
 155          * about spare assignment.... probably not.
 156          *
 157          */
 158         if (mdstat->devcnt != container->devcnt) {
 159                 /* read /sys/block/NAME/md/dev-??/block/dev to find out
 160                  * what is there, and compare with container->info.devs
 161                  * To see what is removed and what is added.
 162                  * These need to be remove from, or added to, the array
 163                  */
 164                 // FIXME
 165                 container->devcnt = mdstat->devcnt;
 166         }
 167 }
 168
 169 static void manage_member(struct mdstat_ent *mdstat,
 170                           struct active_array *a)
 171 {
 172         /* Compare mdstat info with known state of member array.
 173          * We do not need to look for device state changes here, that
 174          * is dealt with by the monitor.
 175          *
 176          * We just look for changes which suggest that a reshape is
 177          * being requested.
 178          * Unfortunately decreases in raid_disks don't show up in
 179          * mdstat until the reshape completes FIXME.
 180          */
 181         // FIXME
 182         a->info.array.raid_disks = mdstat->raid_disks;
 183         a->info.array.chunk_size = mdstat->chunk_size;
 184         // MORE
 185
 186 }
 187
 188 static void manage_new(struct mdstat_ent *mdstat,
 189                        struct supertype *container,
 190                        struct active_array *victim)
 191 {
 192         /* A new array has appeared in this container.
 193          * Hopefully it is already recorded in the metadata.
 194          * Check, then create the new array to report it to
 195          * the monitor.
 196          */
 197
 198         struct active_array *new;
 199         struct mdinfo *mdi, *di;
 200         char *n;
 201         int inst;
 202         int i;
 203
 204         new = malloc(sizeof(*new));
 205
 206         memset(new, 0, sizeof(*new));
 207
 208         new->devnum = mdstat->devnum;
 209
 210         new->prev_state = new->curr_state = new->next_state = inactive;
 211         new->prev_action= new->curr_action= new->next_action= idle;
 212
 213         new->container = container;
 214
 215         n = &mdstat->metadata_version[10+strlen(container->devname)+1];
 216         inst = atoi(n);
 217         if (inst < 0)
 218                 abort();//FIXME
 219
 220         mdi = sysfs_read(-1, new->devnum,
 221                          GET_LEVEL|GET_CHUNK|GET_DISKS|
 222                          GET_DEVS|GET_OFFSET|GET_SIZE|GET_STATE);
 223         if (!mdi) {
 224                 /* Eeek. Cannot monitor this array.
 225                  * Mark it to be ignored by setting container to NULL
 226                  */
 227                 new->container = NULL;
 228                 replace_array(container, victim, new);
 229                 return;
 230         }
 231
 232         new->info.array = mdi->array;
 233
 234         for (i = 0; i < new->info.array.raid_disks; i++) {
 235                 struct mdinfo *newd = malloc(sizeof(*newd));
 236
 237                 for (di = mdi->devs; di; di = di->next)
 238                         if (i == di->disk.raid_disk)
 239                                 break;
 240
 241                 if (di) {
 242                         memcpy(newd, di, sizeof(*newd));
 243
 244                         sprintf(newd->sys_name, "rd%d", i);
 245
 246                         newd->state_fd = sysfs_open(new->devnum,
 247                                                     newd->sys_name,
 248                                                     "state");
 249
 250                         newd->prev_state = read_dev_state(newd->state_fd);
 251                         newd->curr_state = newd->curr_state;
 252                 } else {
 253                         newd->state_fd = -1;
 254                 }
 255                 newd->next = new->info.devs;
 256                 new->info.devs = newd;
 257         }
 258         new->action_fd = sysfs_open(new->devnum, NULL, "sync_action");
 259         new->info.state_fd = sysfs_open(new->devnum, NULL, "array_state");
 260         new->resync_start_fd = sysfs_open(new->devnum, NULL, "resync_start");
 261         new->sync_pos_fd = sysfs_open(new->devnum, NULL, "sync_completed");
 262         new->sync_pos = 0;
 263
 264         sysfs_free(mdi);
 265         // finds and compares.
 266         if (container->ss->open_new(container, new, inst) < 0) {
 267                 // FIXME close all those files
 268                 new->container = NULL;
 269                 replace_array(container, victim, new);
 270                 return;
 271         }
 272         replace_array(container, victim, new);
 273         return;
 274 }
 275
 276 void manage(struct mdstat_ent *mdstat, struct supertype *container)
 277 {
 278         /* We have just read mdstat and need to compare it with
 279          * the known active arrays.
 280          * Arrays with the wrong metadata are ignored.
 281          */
 282
 283         for ( ; mdstat ; mdstat = mdstat->next) {
 284                 struct active_array *a;
 285                 if (mdstat->devnum == container->devnum) {
 286                         manage_container(mdstat, container);
 287                         continue;
 288                 }
 289                 if (mdstat->metadata_version == NULL ||
 290                     strncmp(mdstat->metadata_version, "external:/", 10) != 0 ||
 291                     strncmp(mdstat->metadata_version+10, container->devname,
 292                             strlen(container->devname)) != 0 ||
 293                     mdstat->metadata_version[10+strlen(container->devname)]
 294                       != '/')
 295                         /* Not for this array */
 296                         continue;
 297                 /* Looks like a member of this container */
 298                 for (a = container->arrays; a; a = a->next) {
 299                         if (mdstat->devnum == a->devnum) {
 300                                 if (a->container)
 301                                         manage_member(mdstat, a);
 302                                 break;
 303                         }
 304                 }
 305                 if (a == NULL || !a->container)
 306                         manage_new(mdstat, container, a);
 307         }
 308 }
 309
 310 static int handle_message(struct supertype *container, struct md_message *msg)
 311 {
 312         int err;
 313         struct md_generic_cmd *cmd = msg->buf;
 314
 315         if (!cmd)
 316                 return 0;
 317
 318         switch (cmd->action) {
 319         case md_action_remove_device:
 320
 321                 /* forward to the monitor */
 322                 active_cmd = cmd;
 323                 write(container->mgr_pipe[1], &err, 1);
 324                 read(container->mon_pipe[0], &err, 1);
 325                 return err;
 326
 327         default:
 328                 return -1;
 329         }
 330 }
 331
 332 void read_sock(struct supertype *container)
 333 {
 334         int fd;
 335         struct md_message msg;
 336         int terminate = 0;
 337         long fl;
 338         int tmo = 3; /* 3 second timeout before hanging up the socket */
 339
 340         fd = accept(container->sock, NULL, NULL);
 341         if (fd < 0)
 342                 return;
 343
 344         fl = fcntl(fd, F_GETFL, 0);
 345         fl |= O_NONBLOCK;
 346         fcntl(fd, F_SETFL, fl);
 347
 348         do {
 349                 int err;
 350
 351                 msg.buf = NULL;
 352
 353                 /* read and validate the message */
 354                 if (receive_message(fd, &msg, tmo) == 0) {
 355                         err = handle_message(container, &msg);
 356                         if (!err)
 357                                 ack(fd, msg.seq, tmo);
 358                         else
 359                                 nack(fd, err, tmo);
 360                 } else {
 361                         terminate = 1;
 362                         nack(fd, -1, tmo);
 363                 }
 364
 365                 if (msg.buf)
 366                         free(msg.buf);
 367         } while (!terminate);
 368
 369         close(fd);
 370 }
 371 void do_manager(struct supertype *container)
 372 {
 373         struct mdstat_ent *mdstat;
 374
 375         do {
 376                 mdstat = mdstat_read(1, 0);
 377
 378                 manage(mdstat, container);
 379
 380                 read_sock(container);
 381
 382                 free_mdstat(mdstat);
 383
 384                 mdstat_wait_fd(container->sock);
 385         } while(1);
 386 }