fs/afs/rotate.c

   1 /* Handle fileserver selection and rotation.
   2  *
   3  * Copyright (C) 2017 Red Hat, Inc. All Rights Reserved.
   4  * Written by David Howells (dhowells@redhat.com)
   5  *
   6  * This program is free software; you can redistribute it and/or
   7  * modify it under the terms of the GNU General Public Licence
   8  * as published by the Free Software Foundation; either version
   9  * 2 of the Licence, or (at your option) any later version.
  10  */
  11
  12 #include <linux/kernel.h>
  13 #include <linux/slab.h>
  14 #include <linux/fs.h>
  15 #include <linux/sched.h>
  16 #include <linux/delay.h>
  17 #include <linux/sched/signal.h>
  18 #include "internal.h"
  19 #include "afs_fs.h"
  20
  21 /*
  22  * Initialise a filesystem server cursor for iterating over FS servers.
  23  */
  24 static void afs_init_fs_cursor(struct afs_fs_cursor *fc, struct afs_vnode *vnode)
  25 {
  26         memset(fc, 0, sizeof(*fc));
  27 }
  28
  29 /*
  30  * Begin an operation on the fileserver.
  31  *
  32  * Fileserver operations are serialised on the server by vnode, so we serialise
  33  * them here also using the io_lock.
  34  */
  35 bool afs_begin_vnode_operation(struct afs_fs_cursor *fc, struct afs_vnode *vnode,
  36                                struct key *key)
  37 {
  38         afs_init_fs_cursor(fc, vnode);
  39         fc->vnode = vnode;
  40         fc->key = key;
  41         fc->ac.error = SHRT_MAX;
  42
  43         if (mutex_lock_interruptible(&vnode->io_lock) < 0) {
  44                 fc->ac.error = -EINTR;
  45                 fc->flags |= AFS_FS_CURSOR_STOP;
  46                 return false;
  47         }
  48
  49         if (vnode->lock_state != AFS_VNODE_LOCK_NONE)
  50                 fc->flags |= AFS_FS_CURSOR_CUR_ONLY;
  51         return true;
  52 }
  53
  54 /*
  55  * Begin iteration through a server list, starting with the vnode's last used
  56  * server if possible, or the last recorded good server if not.
  57  */
  58 static bool afs_start_fs_iteration(struct afs_fs_cursor *fc,
  59                                    struct afs_vnode *vnode)
  60 {
  61         struct afs_cb_interest *cbi;
  62         int i;
  63
  64         read_lock(&vnode->volume->servers_lock);
  65         fc->server_list = afs_get_serverlist(vnode->volume->servers);
  66         read_unlock(&vnode->volume->servers_lock);
  67
  68         cbi = vnode->cb_interest;
  69         if (cbi) {
  70                 /* See if the vnode's preferred record is still available */
  71                 for (i = 0; i < fc->server_list->nr_servers; i++) {
  72                         if (fc->server_list->servers[i].cb_interest == cbi) {
  73                                 fc->start = i;
  74                                 goto found_interest;
  75                         }
  76                 }
  77
  78                 /* If we have a lock outstanding on a server that's no longer
  79                  * serving this vnode, then we can't switch to another server
  80                  * and have to return an error.
  81                  */
  82                 if (fc->flags & AFS_FS_CURSOR_CUR_ONLY) {
  83                         fc->ac.error = -ESTALE;
  84                         return false;
  85                 }
  86
  87                 /* Note that the callback promise is effectively broken */
  88                 write_seqlock(&vnode->cb_lock);
  89                 ASSERTCMP(cbi, ==, vnode->cb_interest);
  90                 vnode->cb_interest = NULL;
  91                 if (test_and_clear_bit(AFS_VNODE_CB_PROMISED, &vnode->flags))
  92                         vnode->cb_break++;
  93                 write_sequnlock(&vnode->cb_lock);
  94
  95                 afs_put_cb_interest(afs_v2net(vnode), cbi);
  96                 cbi = NULL;
  97         } else {
  98                 fc->start = READ_ONCE(fc->server_list->index);
  99         }
 100
 101 found_interest:
 102         fc->index = fc->start;
 103         return true;
 104 }
 105
 106 /*
 107  * Post volume busy note.
 108  */
 109 static void afs_busy(struct afs_volume *volume, u32 abort_code)
 110 {
 111         const char *m;
 112
 113         switch (abort_code) {
 114         case VOFFLINE:          m = "offline";          break;
 115         case VRESTARTING:       m = "restarting";       break;
 116         case VSALVAGING:        m = "being salvaged";   break;
 117         default:                m = "busy";             break;
 118         }
 119
 120         pr_notice("kAFS: Volume %u '%s' is %s\n", volume->vid, volume->name, m);
 121 }
 122
 123 /*
 124  * Sleep and retry the operation to the same fileserver.
 125  */
 126 static bool afs_sleep_and_retry(struct afs_fs_cursor *fc)
 127 {
 128         msleep_interruptible(1000);
 129         if (signal_pending(current)) {
 130                 fc->ac.error = -ERESTARTSYS;
 131                 return false;
 132         }
 133
 134         return true;
 135 }
 136
 137 /*
 138  * Select the fileserver to use.  May be called multiple times to rotate
 139  * through the fileservers.
 140  */
 141 bool afs_select_fileserver(struct afs_fs_cursor *fc)
 142 {
 143         struct afs_addr_list *alist;
 144         struct afs_server *server;
 145         struct afs_vnode *vnode = fc->vnode;
 146
 147         _enter("%u/%u,%u/%u,%d,%d",
 148                fc->index, fc->start,
 149                fc->ac.index, fc->ac.start,
 150                fc->ac.error, fc->ac.abort_code);
 151
 152         if (fc->flags & AFS_FS_CURSOR_STOP) {
 153                 _leave(" = f [stopped]");
 154                 return false;
 155         }
 156
 157         /* Evaluate the result of the previous operation, if there was one. */
 158         switch (fc->ac.error) {
 159         case SHRT_MAX:
 160                 goto start;
 161
 162         case 0:
 163         default:
 164                 /* Success or local failure.  Stop. */
 165                 fc->flags |= AFS_FS_CURSOR_STOP;
 166                 _leave(" = f [okay/local %d]", fc->ac.error);
 167                 return false;
 168
 169         case -ECONNABORTED:
 170                 /* The far side rejected the operation on some grounds.  This
 171                  * might involve the server being busy or the volume having been moved.
 172                  */
 173                 switch (fc->ac.abort_code) {
 174                 case VNOVOL:
 175                         /* This fileserver doesn't know about the volume.
 176                          * - May indicate that the VL is wrong - retry once and compare
 177                          *   the results.
 178                          * - May indicate that the fileserver couldn't attach to the vol.
 179                          */
 180                         if (fc->flags & AFS_FS_CURSOR_VNOVOL) {
 181                                 fc->ac.error = -EREMOTEIO;
 182                                 goto next_server;
 183                         }
 184
 185                         write_lock(&vnode->volume->servers_lock);
 186                         fc->server_list->vnovol_mask |= 1 << fc->index;
 187                         write_unlock(&vnode->volume->servers_lock);
 188
 189                         set_bit(AFS_VOLUME_NEEDS_UPDATE, &vnode->volume->flags);
 190                         fc->ac.error = afs_check_volume_status(vnode->volume, fc->key);
 191                         if (fc->ac.error < 0)
 192                                 goto failed;
 193
 194                         if (test_bit(AFS_VOLUME_DELETED, &vnode->volume->flags)) {
 195                                 fc->ac.error = -ENOMEDIUM;
 196                                 goto failed;
 197                         }
 198
 199                         /* If the server list didn't change, then assume that
 200                          * it's the fileserver having trouble.
 201                          */
 202                         if (vnode->volume->servers == fc->server_list) {
 203                                 fc->ac.error = -EREMOTEIO;
 204                                 goto next_server;
 205                         }
 206
 207                         /* Try again */
 208                         fc->flags |= AFS_FS_CURSOR_VNOVOL;
 209                         _leave(" = t [vnovol]");
 210                         return true;
 211
 212                 case VSALVAGE: /* TODO: Should this return an error or iterate? */
 213                 case VVOLEXISTS:
 214                 case VNOSERVICE:
 215                 case VONLINE:
 216                 case VDISKFULL:
 217                 case VOVERQUOTA:
 218                         fc->ac.error = afs_abort_to_error(fc->ac.abort_code);
 219                         goto next_server;
 220
 221                 case VOFFLINE:
 222                         if (!test_and_set_bit(AFS_VOLUME_OFFLINE, &vnode->volume->flags)) {
 223                                 afs_busy(vnode->volume, fc->ac.abort_code);
 224                                 clear_bit(AFS_VOLUME_BUSY, &vnode->volume->flags);
 225                         }
 226                         if (fc->flags & AFS_FS_CURSOR_NO_VSLEEP) {
 227                                 fc->ac.error = -EADV;
 228                                 goto failed;
 229                         }
 230                         if (fc->flags & AFS_FS_CURSOR_CUR_ONLY) {
 231                                 fc->ac.error = -ESTALE;
 232                                 goto failed;
 233                         }
 234                         goto busy;
 235
 236                 case VSALVAGING:
 237                 case VRESTARTING:
 238                 case VBUSY:
 239                         /* Retry after going round all the servers unless we
 240                          * have a file lock we need to maintain.
 241                          */
 242                         if (fc->flags & AFS_FS_CURSOR_NO_VSLEEP) {
 243                                 fc->ac.error = -EBUSY;
 244                                 goto failed;
 245                         }
 246                         if (!test_and_set_bit(AFS_VOLUME_BUSY, &vnode->volume->flags)) {
 247                                 afs_busy(vnode->volume, fc->ac.abort_code);
 248                                 clear_bit(AFS_VOLUME_OFFLINE, &vnode->volume->flags);
 249                         }
 250                 busy:
 251                         if (fc->flags & AFS_FS_CURSOR_CUR_ONLY) {
 252                                 if (!afs_sleep_and_retry(fc))
 253                                         goto failed;
 254
 255                                  /* Retry with same server & address */
 256                                 _leave(" = t [vbusy]");
 257                                 return true;
 258                         }
 259
 260                         fc->flags |= AFS_FS_CURSOR_VBUSY;
 261                         goto next_server;
 262
 263                 case VMOVED:
 264                         /* The volume migrated to another server.  We consider
 265                          * consider all locks and callbacks broken and request
 266                          * an update from the VLDB.
 267                          *
 268                          * We also limit the number of VMOVED hops we will
 269                          * honour, just in case someone sets up a loop.
 270                          */
 271                         if (fc->flags & AFS_FS_CURSOR_VMOVED) {
 272                                 fc->ac.error = -EREMOTEIO;
 273                                 goto failed;
 274                         }
 275                         fc->flags |= AFS_FS_CURSOR_VMOVED;
 276
 277                         set_bit(AFS_VOLUME_WAIT, &vnode->volume->flags);
 278                         set_bit(AFS_VOLUME_NEEDS_UPDATE, &vnode->volume->flags);
 279                         fc->ac.error = afs_check_volume_status(vnode->volume, fc->key);
 280                         if (fc->ac.error < 0)
 281                                 goto failed;
 282
 283                         /* If the server list didn't change, then the VLDB is
 284                          * out of sync with the fileservers.  This is hopefully
 285                          * a temporary condition, however, so we don't want to
 286                          * permanently block access to the file.
 287                          *
 288                          * TODO: Try other fileservers if we can.
 289                          *
 290                          * TODO: Retry a few times with sleeps.
 291                          */
 292                         if (vnode->volume->servers == fc->server_list) {
 293                                 fc->ac.error = -ENOMEDIUM;
 294                                 goto failed;
 295                         }
 296
 297                         goto restart_from_beginning;
 298
 299                 default:
 300                         clear_bit(AFS_VOLUME_OFFLINE, &vnode->volume->flags);
 301                         clear_bit(AFS_VOLUME_BUSY, &vnode->volume->flags);
 302                         fc->ac.error = afs_abort_to_error(fc->ac.abort_code);
 303                         goto failed;
 304                 }
 305
 306         case -ENETUNREACH:
 307         case -EHOSTUNREACH:
 308         case -ECONNREFUSED:
 309         case -ETIMEDOUT:
 310         case -ETIME:
 311                 _debug("no conn");
 312                 goto iterate_address;
 313         }
 314
 315 restart_from_beginning:
 316         _debug("restart");
 317         afs_end_cursor(&fc->ac);
 318         afs_put_cb_interest(afs_v2net(vnode), fc->cbi);
 319         fc->cbi = NULL;
 320         afs_put_serverlist(afs_v2net(vnode), fc->server_list);
 321         fc->server_list = NULL;
 322 start:
 323         _debug("start");
 324         /* See if we need to do an update of the volume record.  Note that the
 325          * volume may have moved or even have been deleted.
 326          */
 327         fc->ac.error = afs_check_volume_status(vnode->volume, fc->key);
 328         if (fc->ac.error < 0)
 329                 goto failed;
 330
 331         if (!afs_start_fs_iteration(fc, vnode))
 332                 goto failed;
 333
 334 use_server:
 335         _debug("use");
 336         /* We're starting on a different fileserver from the list.  We need to
 337          * check it, create a callback intercept, find its address list and
 338          * probe its capabilities before we use it.
 339          */
 340         ASSERTCMP(fc->ac.alist, ==, NULL);
 341         server = fc->server_list->servers[fc->index].server;
 342
 343         if (!afs_check_server_record(fc, server))
 344                 goto failed;
 345
 346         _debug("USING SERVER: %pU", &server->uuid);
 347
 348         /* Make sure we've got a callback interest record for this server.  We
 349          * have to link it in before we send the request as we can be sent a
 350          * break request before we've finished decoding the reply and
 351          * installing the vnode.
 352          */
 353         fc->ac.error = afs_register_server_cb_interest(vnode, fc->server_list,
 354                                                        fc->index);
 355         if (fc->ac.error < 0)
 356                 goto failed;
 357
 358         fc->cbi = afs_get_cb_interest(vnode->cb_interest);
 359
 360         read_lock(&server->fs_lock);
 361         alist = rcu_dereference_protected(server->addresses,
 362                                           lockdep_is_held(&server->fs_lock));
 363         afs_get_addrlist(alist);
 364         read_unlock(&server->fs_lock);
 365
 366         memset(&fc->ac, 0, sizeof(fc->ac));
 367
 368         /* Probe the current fileserver if we haven't done so yet. */
 369         if (!test_bit(AFS_SERVER_FL_PROBED, &server->flags)) {
 370                 fc->ac.alist = afs_get_addrlist(alist);
 371
 372                 if (!afs_probe_fileserver(fc)) {
 373                         switch (fc->ac.error) {
 374                         case -ENOMEM:
 375                         case -ERESTARTSYS:
 376                         case -EINTR:
 377                                 goto failed;
 378                         default:
 379                                 goto next_server;
 380                         }
 381                 }
 382         }
 383
 384         if (!fc->ac.alist)
 385                 fc->ac.alist = alist;
 386         else
 387                 afs_put_addrlist(alist);
 388
 389         fc->ac.start = READ_ONCE(alist->index);
 390         fc->ac.index = fc->ac.start;
 391
 392 iterate_address:
 393         ASSERT(fc->ac.alist);
 394         _debug("iterate %d/%d", fc->ac.index, fc->ac.alist->nr_addrs);
 395         /* Iterate over the current server's address list to try and find an
 396          * address on which it will respond to us.
 397          */
 398         if (!afs_iterate_addresses(&fc->ac))
 399                 goto next_server;
 400
 401         _leave(" = t");
 402         return true;
 403
 404 next_server:
 405         _debug("next");
 406         afs_end_cursor(&fc->ac);
 407         afs_put_cb_interest(afs_v2net(vnode), fc->cbi);
 408         fc->cbi = NULL;
 409         fc->index++;
 410         if (fc->index >= fc->server_list->nr_servers)
 411                 fc->index = 0;
 412         if (fc->index != fc->start)
 413                 goto use_server;
 414
 415         /* That's all the servers poked to no good effect.  Try again if some
 416          * of them were busy.
 417          */
 418         if (fc->flags & AFS_FS_CURSOR_VBUSY)
 419                 goto restart_from_beginning;
 420
 421         fc->ac.error = -EDESTADDRREQ;
 422         goto failed;
 423
 424 failed:
 425         fc->flags |= AFS_FS_CURSOR_STOP;
 426         afs_end_cursor(&fc->ac);
 427         _leave(" = f [failed %d]", fc->ac.error);
 428         return false;
 429 }
 430
 431 /*
 432  * Select the same fileserver we used for a vnode before and only that
 433  * fileserver.  We use this when we have a lock on that file, which is backed
 434  * only by the fileserver we obtained it from.
 435  */
 436 bool afs_select_current_fileserver(struct afs_fs_cursor *fc)
 437 {
 438         struct afs_vnode *vnode = fc->vnode;
 439         struct afs_cb_interest *cbi = vnode->cb_interest;
 440         struct afs_addr_list *alist;
 441
 442         _enter("");
 443
 444         switch (fc->ac.error) {
 445         case SHRT_MAX:
 446                 if (!cbi) {
 447                         fc->ac.error = -ESTALE;
 448                         fc->flags |= AFS_FS_CURSOR_STOP;
 449                         return false;
 450                 }
 451
 452                 fc->cbi = afs_get_cb_interest(vnode->cb_interest);
 453
 454                 read_lock(&cbi->server->fs_lock);
 455                 alist = rcu_dereference_protected(cbi->server->addresses,
 456                                                   lockdep_is_held(&cbi->server->fs_lock));
 457                 afs_get_addrlist(alist);
 458                 read_unlock(&cbi->server->fs_lock);
 459                 if (!alist) {
 460                         fc->ac.error = -ESTALE;
 461                         fc->flags |= AFS_FS_CURSOR_STOP;
 462                         return false;
 463                 }
 464
 465                 memset(&fc->ac, 0, sizeof(fc->ac));
 466                 fc->ac.alist = alist;
 467                 fc->ac.start = READ_ONCE(alist->index);
 468                 fc->ac.index = fc->ac.start;
 469                 goto iterate_address;
 470
 471         case 0:
 472         default:
 473                 /* Success or local failure.  Stop. */
 474                 fc->flags |= AFS_FS_CURSOR_STOP;
 475                 _leave(" = f [okay/local %d]", fc->ac.error);
 476                 return false;
 477
 478         case -ECONNABORTED:
 479                 fc->flags |= AFS_FS_CURSOR_STOP;
 480                 _leave(" = f [abort]");
 481                 return false;
 482
 483         case -ENETUNREACH:
 484         case -EHOSTUNREACH:
 485         case -ECONNREFUSED:
 486         case -ETIMEDOUT:
 487         case -ETIME:
 488                 _debug("no conn");
 489                 goto iterate_address;
 490         }
 491
 492 iterate_address:
 493         /* Iterate over the current server's address list to try and find an
 494          * address on which it will respond to us.
 495          */
 496         if (afs_iterate_addresses(&fc->ac)) {
 497                 _leave(" = t");
 498                 return true;
 499         }
 500
 501         afs_end_cursor(&fc->ac);
 502         return false;
 503 }
 504
 505 /*
 506  * Tidy up a filesystem cursor and unlock the vnode.
 507  */
 508 int afs_end_vnode_operation(struct afs_fs_cursor *fc)
 509 {
 510         struct afs_net *net = afs_v2net(fc->vnode);
 511         int ret;
 512
 513         mutex_unlock(&fc->vnode->io_lock);
 514
 515         afs_end_cursor(&fc->ac);
 516         afs_put_cb_interest(net, fc->cbi);
 517         afs_put_serverlist(net, fc->server_list);
 518
 519         ret = fc->ac.error;
 520         if (ret == -ECONNABORTED)
 521                 afs_abort_to_error(fc->ac.abort_code);
 522
 523         return fc->ac.error;
 524 }