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  * Begin an operation on the fileserver.
  23  *
  24  * Fileserver operations are serialised on the server by vnode, so we serialise
  25  * them here also using the io_lock.
  26  */
  27 bool afs_begin_vnode_operation(struct afs_fs_cursor *fc, struct afs_vnode *vnode,
  28                                struct key *key, bool intr)
  29 {
  30         memset(fc, 0, sizeof(*fc));
  31         fc->vnode = vnode;
  32         fc->key = key;
  33         fc->ac.error = SHRT_MAX;
  34         fc->error = -EDESTADDRREQ;
  35
  36         if (intr) {
  37                 fc->flags |= AFS_FS_CURSOR_INTR;
  38                 if (mutex_lock_interruptible(&vnode->io_lock) < 0) {
  39                         fc->error = -EINTR;
  40                         fc->flags |= AFS_FS_CURSOR_STOP;
  41                         return false;
  42                 }
  43         } else {
  44                 mutex_lock(&vnode->io_lock);
  45         }
  46
  47         if (vnode->lock_state != AFS_VNODE_LOCK_NONE)
  48                 fc->flags |= AFS_FS_CURSOR_CUR_ONLY;
  49         return true;
  50 }
  51
  52 /*
  53  * Begin iteration through a server list, starting with the vnode's last used
  54  * server if possible, or the last recorded good server if not.
  55  */
  56 static bool afs_start_fs_iteration(struct afs_fs_cursor *fc,
  57                                    struct afs_vnode *vnode)
  58 {
  59         struct afs_cb_interest *cbi;
  60         int i;
  61
  62         read_lock(&vnode->volume->servers_lock);
  63         fc->server_list = afs_get_serverlist(vnode->volume->servers);
  64         read_unlock(&vnode->volume->servers_lock);
  65
  66         fc->untried = (1UL << fc->server_list->nr_servers) - 1;
  67         fc->index = READ_ONCE(fc->server_list->preferred);
  68
  69         cbi = rcu_dereference_protected(vnode->cb_interest,
  70                                         lockdep_is_held(&vnode->io_lock));
  71         if (cbi) {
  72                 /* See if the vnode's preferred record is still available */
  73                 for (i = 0; i < fc->server_list->nr_servers; i++) {
  74                         if (fc->server_list->servers[i].cb_interest == cbi) {
  75                                 fc->index = i;
  76                                 goto found_interest;
  77                         }
  78                 }
  79
  80                 /* If we have a lock outstanding on a server that's no longer
  81                  * serving this vnode, then we can't switch to another server
  82                  * and have to return an error.
  83                  */
  84                 if (fc->flags & AFS_FS_CURSOR_CUR_ONLY) {
  85                         fc->error = -ESTALE;
  86                         return false;
  87                 }
  88
  89                 /* Note that the callback promise is effectively broken */
  90                 write_seqlock(&vnode->cb_lock);
  91                 ASSERTCMP(cbi, ==, rcu_access_pointer(vnode->cb_interest));
  92                 rcu_assign_pointer(vnode->cb_interest, NULL);
  93                 if (test_and_clear_bit(AFS_VNODE_CB_PROMISED, &vnode->flags))
  94                         vnode->cb_break++;
  95                 write_sequnlock(&vnode->cb_lock);
  96
  97                 afs_put_cb_interest(afs_v2net(vnode), cbi);
  98                 cbi = NULL;
  99         }
 100
 101 found_interest:
 102         return true;
 103 }
 104
 105 /*
 106  * Post volume busy note.
 107  */
 108 static void afs_busy(struct afs_volume *volume, u32 abort_code)
 109 {
 110         const char *m;
 111
 112         switch (abort_code) {
 113         case VOFFLINE:          m = "offline";          break;
 114         case VRESTARTING:       m = "restarting";       break;
 115         case VSALVAGING:        m = "being salvaged";   break;
 116         default:                m = "busy";             break;
 117         }
 118
 119         pr_notice("kAFS: Volume %llu '%s' is %s\n", volume->vid, volume->name, m);
 120 }
 121
 122 /*
 123  * Sleep and retry the operation to the same fileserver.
 124  */
 125 static bool afs_sleep_and_retry(struct afs_fs_cursor *fc)
 126 {
 127         if (fc->flags & AFS_FS_CURSOR_INTR) {
 128                 msleep_interruptible(1000);
 129                 if (signal_pending(current)) {
 130                         fc->error = -ERESTARTSYS;
 131                         return false;
 132                 }
 133         } else {
 134                 msleep(1000);
 135         }
 136
 137         return true;
 138 }
 139
 140 /*
 141  * Select the fileserver to use.  May be called multiple times to rotate
 142  * through the fileservers.
 143  */
 144 bool afs_select_fileserver(struct afs_fs_cursor *fc)
 145 {
 146         struct afs_addr_list *alist;
 147         struct afs_server *server;
 148         struct afs_vnode *vnode = fc->vnode;
 149         struct afs_error e;
 150         u32 rtt;
 151         int error = fc->ac.error, i;
 152
 153         _enter("%lx[%d],%lx[%d],%d,%d",
 154                fc->untried, fc->index,
 155                fc->ac.tried, fc->ac.index,
 156                error, fc->ac.abort_code);
 157
 158         if (fc->flags & AFS_FS_CURSOR_STOP) {
 159                 _leave(" = f [stopped]");
 160                 return false;
 161         }
 162
 163         fc->nr_iterations++;
 164
 165         /* Evaluate the result of the previous operation, if there was one. */
 166         switch (error) {
 167         case SHRT_MAX:
 168                 goto start;
 169
 170         case 0:
 171         default:
 172                 /* Success or local failure.  Stop. */
 173                 fc->error = error;
 174                 fc->flags |= AFS_FS_CURSOR_STOP;
 175                 _leave(" = f [okay/local %d]", error);
 176                 return false;
 177
 178         case -ECONNABORTED:
 179                 /* The far side rejected the operation on some grounds.  This
 180                  * might involve the server being busy or the volume having been moved.
 181                  */
 182                 switch (fc->ac.abort_code) {
 183                 case VNOVOL:
 184                         /* This fileserver doesn't know about the volume.
 185                          * - May indicate that the VL is wrong - retry once and compare
 186                          *   the results.
 187                          * - May indicate that the fileserver couldn't attach to the vol.
 188                          */
 189                         if (fc->flags & AFS_FS_CURSOR_VNOVOL) {
 190                                 fc->error = -EREMOTEIO;
 191                                 goto next_server;
 192                         }
 193
 194                         write_lock(&vnode->volume->servers_lock);
 195                         fc->server_list->vnovol_mask |= 1 << fc->index;
 196                         write_unlock(&vnode->volume->servers_lock);
 197
 198                         set_bit(AFS_VOLUME_NEEDS_UPDATE, &vnode->volume->flags);
 199                         error = afs_check_volume_status(vnode->volume, fc->key);
 200                         if (error < 0)
 201                                 goto failed_set_error;
 202
 203                         if (test_bit(AFS_VOLUME_DELETED, &vnode->volume->flags)) {
 204                                 fc->error = -ENOMEDIUM;
 205                                 goto failed;
 206                         }
 207
 208                         /* If the server list didn't change, then assume that
 209                          * it's the fileserver having trouble.
 210                          */
 211                         if (vnode->volume->servers == fc->server_list) {
 212                                 fc->error = -EREMOTEIO;
 213                                 goto next_server;
 214                         }
 215
 216                         /* Try again */
 217                         fc->flags |= AFS_FS_CURSOR_VNOVOL;
 218                         _leave(" = t [vnovol]");
 219                         return true;
 220
 221                 case VSALVAGE: /* TODO: Should this return an error or iterate? */
 222                 case VVOLEXISTS:
 223                 case VNOSERVICE:
 224                 case VONLINE:
 225                 case VDISKFULL:
 226                 case VOVERQUOTA:
 227                         fc->error = afs_abort_to_error(fc->ac.abort_code);
 228                         goto next_server;
 229
 230                 case VOFFLINE:
 231                         if (!test_and_set_bit(AFS_VOLUME_OFFLINE, &vnode->volume->flags)) {
 232                                 afs_busy(vnode->volume, fc->ac.abort_code);
 233                                 clear_bit(AFS_VOLUME_BUSY, &vnode->volume->flags);
 234                         }
 235                         if (fc->flags & AFS_FS_CURSOR_NO_VSLEEP) {
 236                                 fc->error = -EADV;
 237                                 goto failed;
 238                         }
 239                         if (fc->flags & AFS_FS_CURSOR_CUR_ONLY) {
 240                                 fc->error = -ESTALE;
 241                                 goto failed;
 242                         }
 243                         goto busy;
 244
 245                 case VSALVAGING:
 246                 case VRESTARTING:
 247                 case VBUSY:
 248                         /* Retry after going round all the servers unless we
 249                          * have a file lock we need to maintain.
 250                          */
 251                         if (fc->flags & AFS_FS_CURSOR_NO_VSLEEP) {
 252                                 fc->error = -EBUSY;
 253                                 goto failed;
 254                         }
 255                         if (!test_and_set_bit(AFS_VOLUME_BUSY, &vnode->volume->flags)) {
 256                                 afs_busy(vnode->volume, fc->ac.abort_code);
 257                                 clear_bit(AFS_VOLUME_OFFLINE, &vnode->volume->flags);
 258                         }
 259                 busy:
 260                         if (fc->flags & AFS_FS_CURSOR_CUR_ONLY) {
 261                                 if (!afs_sleep_and_retry(fc))
 262                                         goto failed;
 263
 264                                  /* Retry with same server & address */
 265                                 _leave(" = t [vbusy]");
 266                                 return true;
 267                         }
 268
 269                         fc->flags |= AFS_FS_CURSOR_VBUSY;
 270                         goto next_server;
 271
 272                 case VMOVED:
 273                         /* The volume migrated to another server.  We consider
 274                          * consider all locks and callbacks broken and request
 275                          * an update from the VLDB.
 276                          *
 277                          * We also limit the number of VMOVED hops we will
 278                          * honour, just in case someone sets up a loop.
 279                          */
 280                         if (fc->flags & AFS_FS_CURSOR_VMOVED) {
 281                                 fc->error = -EREMOTEIO;
 282                                 goto failed;
 283                         }
 284                         fc->flags |= AFS_FS_CURSOR_VMOVED;
 285
 286                         set_bit(AFS_VOLUME_WAIT, &vnode->volume->flags);
 287                         set_bit(AFS_VOLUME_NEEDS_UPDATE, &vnode->volume->flags);
 288                         error = afs_check_volume_status(vnode->volume, fc->key);
 289                         if (error < 0)
 290                                 goto failed_set_error;
 291
 292                         /* If the server list didn't change, then the VLDB is
 293                          * out of sync with the fileservers.  This is hopefully
 294                          * a temporary condition, however, so we don't want to
 295                          * permanently block access to the file.
 296                          *
 297                          * TODO: Try other fileservers if we can.
 298                          *
 299                          * TODO: Retry a few times with sleeps.
 300                          */
 301                         if (vnode->volume->servers == fc->server_list) {
 302                                 fc->error = -ENOMEDIUM;
 303                                 goto failed;
 304                         }
 305
 306                         goto restart_from_beginning;
 307
 308                 default:
 309                         clear_bit(AFS_VOLUME_OFFLINE, &vnode->volume->flags);
 310                         clear_bit(AFS_VOLUME_BUSY, &vnode->volume->flags);
 311                         fc->error = afs_abort_to_error(fc->ac.abort_code);
 312                         goto failed;
 313                 }
 314
 315         case -ETIMEDOUT:
 316         case -ETIME:
 317                 if (fc->error != -EDESTADDRREQ)
 318                         goto iterate_address;
 319                 /* Fall through */
 320         case -ERFKILL:
 321         case -EADDRNOTAVAIL:
 322         case -ENETUNREACH:
 323         case -EHOSTUNREACH:
 324         case -EHOSTDOWN:
 325         case -ECONNREFUSED:
 326                 _debug("no conn");
 327                 fc->error = error;
 328                 goto iterate_address;
 329
 330         case -ECONNRESET:
 331                 _debug("call reset");
 332                 fc->error = error;
 333                 goto failed;
 334         }
 335
 336 restart_from_beginning:
 337         _debug("restart");
 338         afs_end_cursor(&fc->ac);
 339         afs_put_cb_interest(afs_v2net(vnode), fc->cbi);
 340         fc->cbi = NULL;
 341         afs_put_serverlist(afs_v2net(vnode), fc->server_list);
 342         fc->server_list = NULL;
 343 start:
 344         _debug("start");
 345         /* See if we need to do an update of the volume record.  Note that the
 346          * volume may have moved or even have been deleted.
 347          */
 348         error = afs_check_volume_status(vnode->volume, fc->key);
 349         if (error < 0)
 350                 goto failed_set_error;
 351
 352         if (!afs_start_fs_iteration(fc, vnode))
 353                 goto failed;
 354
 355         _debug("__ VOL %llx __", vnode->volume->vid);
 356         error = afs_probe_fileservers(afs_v2net(vnode), fc->key, fc->server_list);
 357         if (error < 0)
 358                 goto failed_set_error;
 359
 360 pick_server:
 361         _debug("pick [%lx]", fc->untried);
 362
 363         error = afs_wait_for_fs_probes(fc->server_list, fc->untried);
 364         if (error < 0)
 365                 goto failed_set_error;
 366
 367         /* Pick the untried server with the lowest RTT.  If we have outstanding
 368          * callbacks, we stick with the server we're already using if we can.
 369          */
 370         if (fc->cbi) {
 371                 _debug("cbi %u", fc->index);
 372                 if (test_bit(fc->index, &fc->untried))
 373                         goto selected_server;
 374                 afs_put_cb_interest(afs_v2net(vnode), fc->cbi);
 375                 fc->cbi = NULL;
 376                 _debug("nocbi");
 377         }
 378
 379         fc->index = -1;
 380         rtt = U32_MAX;
 381         for (i = 0; i < fc->server_list->nr_servers; i++) {
 382                 struct afs_server *s = fc->server_list->servers[i].server;
 383
 384                 if (!test_bit(i, &fc->untried) || !s->probe.responded)
 385                         continue;
 386                 if (s->probe.rtt < rtt) {
 387                         fc->index = i;
 388                         rtt = s->probe.rtt;
 389                 }
 390         }
 391
 392         if (fc->index == -1)
 393                 goto no_more_servers;
 394
 395 selected_server:
 396         _debug("use %d", fc->index);
 397         __clear_bit(fc->index, &fc->untried);
 398
 399         /* We're starting on a different fileserver from the list.  We need to
 400          * check it, create a callback intercept, find its address list and
 401          * probe its capabilities before we use it.
 402          */
 403         ASSERTCMP(fc->ac.alist, ==, NULL);
 404         server = fc->server_list->servers[fc->index].server;
 405
 406         if (!afs_check_server_record(fc, server))
 407                 goto failed;
 408
 409         _debug("USING SERVER: %pU", &server->uuid);
 410
 411         /* Make sure we've got a callback interest record for this server.  We
 412          * have to link it in before we send the request as we can be sent a
 413          * break request before we've finished decoding the reply and
 414          * installing the vnode.
 415          */
 416         error = afs_register_server_cb_interest(vnode, fc->server_list,
 417                                                 fc->index);
 418         if (error < 0)
 419                 goto failed_set_error;
 420
 421         fc->cbi = afs_get_cb_interest(
 422                 rcu_dereference_protected(vnode->cb_interest,
 423                                           lockdep_is_held(&vnode->io_lock)));
 424
 425         read_lock(&server->fs_lock);
 426         alist = rcu_dereference_protected(server->addresses,
 427                                           lockdep_is_held(&server->fs_lock));
 428         afs_get_addrlist(alist);
 429         read_unlock(&server->fs_lock);
 430
 431         memset(&fc->ac, 0, sizeof(fc->ac));
 432
 433         if (!fc->ac.alist)
 434                 fc->ac.alist = alist;
 435         else
 436                 afs_put_addrlist(alist);
 437
 438         fc->ac.index = -1;
 439
 440 iterate_address:
 441         ASSERT(fc->ac.alist);
 442         /* Iterate over the current server's address list to try and find an
 443          * address on which it will respond to us.
 444          */
 445         if (!afs_iterate_addresses(&fc->ac))
 446                 goto next_server;
 447
 448         _debug("address [%u] %u/%u", fc->index, fc->ac.index, fc->ac.alist->nr_addrs);
 449
 450         _leave(" = t");
 451         return true;
 452
 453 next_server:
 454         _debug("next");
 455         afs_end_cursor(&fc->ac);
 456         goto pick_server;
 457
 458 no_more_servers:
 459         /* That's all the servers poked to no good effect.  Try again if some
 460          * of them were busy.
 461          */
 462         if (fc->flags & AFS_FS_CURSOR_VBUSY)
 463                 goto restart_from_beginning;
 464
 465         e.error = -EDESTADDRREQ;
 466         e.responded = false;
 467         for (i = 0; i < fc->server_list->nr_servers; i++) {
 468                 struct afs_server *s = fc->server_list->servers[i].server;
 469
 470                 afs_prioritise_error(&e, READ_ONCE(s->probe.error),
 471                                      s->probe.abort_code);
 472         }
 473
 474         error = e.error;
 475
 476 failed_set_error:
 477         fc->error = error;
 478 failed:
 479         fc->flags |= AFS_FS_CURSOR_STOP;
 480         afs_end_cursor(&fc->ac);
 481         _leave(" = f [failed %d]", fc->error);
 482         return false;
 483 }
 484
 485 /*
 486  * Select the same fileserver we used for a vnode before and only that
 487  * fileserver.  We use this when we have a lock on that file, which is backed
 488  * only by the fileserver we obtained it from.
 489  */
 490 bool afs_select_current_fileserver(struct afs_fs_cursor *fc)
 491 {
 492         struct afs_vnode *vnode = fc->vnode;
 493         struct afs_cb_interest *cbi;
 494         struct afs_addr_list *alist;
 495         int error = fc->ac.error;
 496
 497         _enter("");
 498
 499         cbi = rcu_dereference_protected(vnode->cb_interest,
 500                                         lockdep_is_held(&vnode->io_lock));
 501
 502         switch (error) {
 503         case SHRT_MAX:
 504                 if (!cbi) {
 505                         fc->error = -ESTALE;
 506                         fc->flags |= AFS_FS_CURSOR_STOP;
 507                         return false;
 508                 }
 509
 510                 fc->cbi = afs_get_cb_interest(cbi);
 511
 512                 read_lock(&cbi->server->fs_lock);
 513                 alist = rcu_dereference_protected(cbi->server->addresses,
 514                                                   lockdep_is_held(&cbi->server->fs_lock));
 515                 afs_get_addrlist(alist);
 516                 read_unlock(&cbi->server->fs_lock);
 517                 if (!alist) {
 518                         fc->error = -ESTALE;
 519                         fc->flags |= AFS_FS_CURSOR_STOP;
 520                         return false;
 521                 }
 522
 523                 memset(&fc->ac, 0, sizeof(fc->ac));
 524                 fc->ac.alist = alist;
 525                 fc->ac.index = -1;
 526                 goto iterate_address;
 527
 528         case 0:
 529         default:
 530                 /* Success or local failure.  Stop. */
 531                 fc->error = error;
 532                 fc->flags |= AFS_FS_CURSOR_STOP;
 533                 _leave(" = f [okay/local %d]", error);
 534                 return false;
 535
 536         case -ECONNABORTED:
 537                 fc->error = afs_abort_to_error(fc->ac.abort_code);
 538                 fc->flags |= AFS_FS_CURSOR_STOP;
 539                 _leave(" = f [abort]");
 540                 return false;
 541
 542         case -ERFKILL:
 543         case -EADDRNOTAVAIL:
 544         case -ENETUNREACH:
 545         case -EHOSTUNREACH:
 546         case -EHOSTDOWN:
 547         case -ECONNREFUSED:
 548         case -ETIMEDOUT:
 549         case -ETIME:
 550                 _debug("no conn");
 551                 fc->error = error;
 552                 goto iterate_address;
 553         }
 554
 555 iterate_address:
 556         /* Iterate over the current server's address list to try and find an
 557          * address on which it will respond to us.
 558          */
 559         if (afs_iterate_addresses(&fc->ac)) {
 560                 _leave(" = t");
 561                 return true;
 562         }
 563
 564         afs_end_cursor(&fc->ac);
 565         return false;
 566 }
 567
 568 /*
 569  * Dump cursor state in the case of the error being EDESTADDRREQ.
 570  */
 571 static void afs_dump_edestaddrreq(const struct afs_fs_cursor *fc)
 572 {
 573         static int count;
 574         int i;
 575
 576         if (!IS_ENABLED(CONFIG_AFS_DEBUG_CURSOR) || count > 3)
 577                 return;
 578         count++;
 579
 580         rcu_read_lock();
 581
 582         pr_notice("EDESTADDR occurred\n");
 583         pr_notice("FC: cbb=%x cbb2=%x fl=%hx err=%hd\n",
 584                   fc->cb_break, fc->cb_break_2, fc->flags, fc->error);
 585         pr_notice("FC: ut=%lx ix=%d ni=%u\n",
 586                   fc->untried, fc->index, fc->nr_iterations);
 587
 588         if (fc->server_list) {
 589                 const struct afs_server_list *sl = fc->server_list;
 590                 pr_notice("FC: SL nr=%u pr=%u vnov=%hx\n",
 591                           sl->nr_servers, sl->preferred, sl->vnovol_mask);
 592                 for (i = 0; i < sl->nr_servers; i++) {
 593                         const struct afs_server *s = sl->servers[i].server;
 594                         pr_notice("FC: server fl=%lx av=%u %pU\n",
 595                                   s->flags, s->addr_version, &s->uuid);
 596                         if (s->addresses) {
 597                                 const struct afs_addr_list *a =
 598                                         rcu_dereference(s->addresses);
 599                                 pr_notice("FC:  - av=%u nr=%u/%u/%u pr=%u\n",
 600                                           a->version,
 601                                           a->nr_ipv4, a->nr_addrs, a->max_addrs,
 602                                           a->preferred);
 603                                 pr_notice("FC:  - pr=%lx R=%lx F=%lx\n",
 604                                           a->probed, a->responded, a->failed);
 605                                 if (a == fc->ac.alist)
 606                                         pr_notice("FC:  - current\n");
 607                         }
 608                 }
 609         }
 610
 611         pr_notice("AC: t=%lx ax=%u ac=%d er=%d r=%u ni=%u\n",
 612                   fc->ac.tried, fc->ac.index, fc->ac.abort_code, fc->ac.error,
 613                   fc->ac.responded, fc->ac.nr_iterations);
 614         rcu_read_unlock();
 615 }
 616
 617 /*
 618  * Tidy up a filesystem cursor and unlock the vnode.
 619  */
 620 int afs_end_vnode_operation(struct afs_fs_cursor *fc)
 621 {
 622         struct afs_net *net = afs_v2net(fc->vnode);
 623
 624         if (fc->error == -EDESTADDRREQ ||
 625             fc->error == -EADDRNOTAVAIL ||
 626             fc->error == -ENETUNREACH ||
 627             fc->error == -EHOSTUNREACH)
 628                 afs_dump_edestaddrreq(fc);
 629
 630         mutex_unlock(&fc->vnode->io_lock);
 631
 632         afs_end_cursor(&fc->ac);
 633         afs_put_cb_interest(net, fc->cbi);
 634         afs_put_serverlist(net, fc->server_list);
 635
 636         if (fc->error == -ECONNABORTED)
 637                 fc->error = afs_abort_to_error(fc->ac.abort_code);
 638
 639         return fc->error;
 640 }