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
  80 static void free_aa(struct active_array *aa)
  81 {
  82         /* Note that this doesn't close fds, as they may be in used
  83          * by a clone.  Use close_aa for that.
  84          */
  85         while (aa->info.devs) {
  86                 struct mdinfo *d = aa->info.devs;
  87                 aa->info.devs = d->next;
  88                 free(d);
  89         }
  90         free(aa);
  91 }
  92
  93 static void replace_array(struct supertype *container,
  94                           struct active_array *old,
  95                           struct active_array *new)
  96 {
  97         /* To replace an array, we add it to the top of the list
  98          * marked with ->replaces to point to the original.
  99          * 'monitor' will take the original out of the list
 100          * and put it on 'discard_this'.  We take it from there
 101          * and discard it.
 102          */
 103
 104         while (pending_discard) {
 105                 while (discard_this == NULL)
 106                         sleep(1);
 107                 if (discard_this != pending_discard)
 108                         abort();
 109                 discard_this->next = NULL;
 110                 free_aa(discard_this);
 111                 discard_this = NULL;
 112                 pending_discard = NULL;
 113         }
 114         pending_discard = old;
 115         new->replaces = old;
 116         new->next = container->arrays;
 117         container->arrays = new;
 118 }
 119
 120
 121 static void manage_container(struct mdstat_ent *mdstat,
 122                              struct supertype *container)
 123 {
 124         /* The only thing of interest here is if a new device
 125          * has been added to the container.  We add it to the
 126          * array ignoring any metadata on it.
 127          * FIXME should we look for compatible metadata and take hints
 128          * about spare assignment.... probably not.
 129          *
 130          */
 131         if (mdstat->devcnt != container->devcnt) {
 132                 /* read /sys/block/NAME/md/dev-??/block/dev to find out
 133                  * what is there, and compare with container->info.devs
 134                  * To see what is removed and what is added.
 135                  * These need to be remove from, or added to, the array
 136                  */
 137                 // FIXME
 138                 container->devcnt = mdstat->devcnt;
 139         }
 140 }
 141
 142 static void manage_member(struct mdstat_ent *mdstat,
 143                           struct active_array *a)
 144 {
 145         /* Compare mdstat info with known state of member array.
 146          * We do not need to look for device state changes here, that
 147          * is dealt with by the monitor.
 148          *
 149          * We just look for changes which suggest that a reshape is
 150          * being requested.
 151          * Unfortunately decreases in raid_disks don't show up in
 152          * mdstat until the reshape completes FIXME.
 153          */
 154         // FIXME
 155         a->info.array.raid_disks = mdstat->raid_disks;
 156         a->info.array.chunk_size = mdstat->chunk_size;
 157         // MORE
 158
 159 }
 160
 161 static void write_wakeup(struct supertype *c)
 162 {
 163         write(c->pipe[1], "PING", 4);
 164 }
 165
 166 static void manage_new(struct mdstat_ent *mdstat,
 167                        struct supertype *container)
 168 {
 169         /* A new array has appeared in this container.
 170          * Hopefully it is already recorded in the metadata.
 171          * Check, then create the new array to report it to
 172          * the monitor.
 173          */
 174
 175         struct active_array *new;
 176         struct mdinfo *mdi, *di;
 177         char *n;
 178         int inst;
 179         int i;
 180
 181         new = malloc(sizeof(*new));
 182
 183         new->devnum = mdstat->devnum;
 184
 185         new->prev_state = new->curr_state = new->next_state = inactive;
 186         new->prev_action= new->curr_action= new->next_action= idle;
 187
 188         new->container = container;
 189
 190         n = &mdstat->metadata_version[10+strlen(container->devname)+1];
 191         inst = atoi(n);
 192         if (inst < 0)
 193                 abort();//FIXME
 194
 195         mdi = sysfs_read(-1, new->devnum,
 196                          GET_LEVEL|GET_CHUNK|GET_DISKS|
 197                          GET_DEVS|GET_OFFSET|GET_SIZE|GET_STATE);
 198         if (!mdi) {
 199                 /* Eeek. Cannot monitor this array.
 200                  * Mark it to be ignored by setting container to NULL
 201                  */
 202                 new->container = NULL;
 203                 replace_array(container, NULL, new);
 204                 return;
 205         }
 206
 207         new->info.array = mdi->array;
 208
 209         for (i = 0; i < new->info.array.raid_disks; i++) {
 210                 struct mdinfo *newd = malloc(sizeof(*newd));
 211
 212                 for (di = mdi->devs; di; di = di->next)
 213                         if (i == di->disk.raid_disk)
 214                                 break;
 215
 216                 if (di) {
 217                         memcpy(newd, di, sizeof(*newd));
 218
 219                         sprintf(newd->sys_name, "rd%d", i);
 220
 221                         newd->state_fd = sysfs_open(new->devnum,
 222                                                     newd->sys_name,
 223                                                     "state");
 224
 225                         newd->prev_state = read_dev_state(newd->state_fd);
 226                         newd->curr_state = newd->curr_state;
 227                 } else {
 228                         newd->state_fd = -1;
 229                 }
 230                 newd->next = new->info.devs;
 231                 new->info.devs = newd;
 232         }
 233         new->action_fd = sysfs_open(new->devnum, NULL, "sync_action");
 234         new->info.state_fd = sysfs_open(new->devnum, NULL, "array_state");
 235         new->resync_start_fd = sysfs_open(new->devnum, NULL, "resync_start");
 236         new->sync_pos_fd = sysfs_open(new->devnum, NULL, "sync_completed");
 237         new->sync_pos = 0;
 238
 239         // finds and compares.
 240         if (container->ss->open_new(container, new, inst) < 0) {
 241                 // FIXME close all those files
 242                 new->container = NULL;
 243                 replace_array(container, NULL, new);
 244                 return;
 245         }
 246         replace_array(container, NULL, new);
 247         write_wakeup(container);
 248         return;
 249 }
 250
 251 void manage(struct mdstat_ent *mdstat, struct active_array *aa,
 252             struct supertype *container)
 253 {
 254         /* We have just read mdstat and need to compare it with
 255          * the known active arrays.
 256          * Arrays with the wrong metadata are ignored.
 257          */
 258
 259         for ( ; mdstat ; mdstat = mdstat->next) {
 260                 struct active_array *a;
 261                 if (mdstat->devnum == container->devnum) {
 262                         manage_container(mdstat, container);
 263                         continue;
 264                 }
 265                 if (mdstat->metadata_version == NULL ||
 266                     strncmp(mdstat->metadata_version, "external:/", 10) != 0 ||
 267                     strncmp(mdstat->metadata_version+10, container->devname,
 268                             strlen(container->devname)) != 0 ||
 269                     mdstat->metadata_version[10+strlen(container->devname)]
 270                       != '/')
 271                         /* Not for this array */
 272                         continue;
 273                 /* Looks like a member of this container */
 274                 for (a = aa; a; a = a->next) {
 275                         if (mdstat->devnum == a->devnum) {
 276                                 if (a->container)
 277                                         manage_member(mdstat, a);
 278                                 break;
 279                         }
 280                 }
 281                 if (a == NULL)
 282                         manage_new(mdstat, container);
 283         }
 284 }
 285
 286 void read_sock(int pfd)
 287 {
 288         int fd;
 289
 290         // FIXME set non-blocking
 291         fd = accept(pfd, NULL, NULL);
 292         if (fd < 0)
 293                 return;
 294         // FIXME do something useful
 295         close(fd);
 296 }
 297 void do_manager(struct supertype *container)
 298 {
 299         struct mdstat_ent *mdstat;
 300
 301         do {
 302                 mdstat = mdstat_read(1, 0);
 303
 304                 manage(mdstat, array_list, container);
 305
 306                 read_sock(container->sock);
 307
 308                 mdstat_wait_fd(container->sock);
 309         } while(1);
 310 }