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)
  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 (mutex_lock_interruptible(&vnode->io_lock) < 0) {
  37                 fc->error = -EINTR;
  38                 fc->flags |= AFS_FS_CURSOR_STOP;
  39                 return false;
  40         }
  41
  42         if (vnode->lock_state != AFS_VNODE_LOCK_NONE)
  43                 fc->flags |= AFS_FS_CURSOR_CUR_ONLY;
  44         return true;
  45 }
  46
  47 /*
  48  * Begin iteration through a server list, starting with the vnode's last used
  49  * server if possible, or the last recorded good server if not.
  50  */
  51 static bool afs_start_fs_iteration(struct afs_fs_cursor *fc,
  52                                    struct afs_vnode *vnode)
  53 {
  54         struct afs_cb_interest *cbi;
  55         int i;
  56
  57         read_lock(&vnode->volume->servers_lock);
  58         fc->server_list = afs_get_serverlist(vnode->volume->servers);
  59         read_unlock(&vnode->volume->servers_lock);
  60
  61         fc->untried = (1UL << fc->server_list->nr_servers) - 1;
  62         fc->index = READ_ONCE(fc->server_list->preferred);
  63
  64         cbi = vnode->cb_interest;
  65         if (cbi) {
  66                 /* See if the vnode's preferred record is still available */
  67                 for (i = 0; i < fc->server_list->nr_servers; i++) {
  68                         if (fc->server_list->servers[i].cb_interest == cbi) {
  69                                 fc->index = i;
  70                                 goto found_interest;
  71                         }
  72                 }
  73
  74                 /* If we have a lock outstanding on a server that's no longer
  75                  * serving this vnode, then we can't switch to another server
  76                  * and have to return an error.
  77                  */
  78                 if (fc->flags & AFS_FS_CURSOR_CUR_ONLY) {
  79                         fc->error = -ESTALE;
  80                         return false;
  81                 }
  82
  83                 /* Note that the callback promise is effectively broken */
  84                 write_seqlock(&vnode->cb_lock);
  85                 ASSERTCMP(cbi, ==, vnode->cb_interest);
  86                 vnode->cb_interest = NULL;
  87                 if (test_and_clear_bit(AFS_VNODE_CB_PROMISED, &vnode->flags))
  88                         vnode->cb_break++;
  89                 write_sequnlock(&vnode->cb_lock);
  90
  91                 afs_put_cb_interest(afs_v2net(vnode), cbi);
  92                 cbi = NULL;
  93         }
  94
  95 found_interest:
  96         return true;
  97 }
  98
  99 /*
 100  * Post volume busy note.
 101  */
 102 static void afs_busy(struct afs_volume *volume, u32 abort_code)
 103 {
 104         const char *m;
 105
 106         switch (abort_code) {
 107         case VOFFLINE:          m = "offline";          break;
 108         case VRESTARTING:       m = "restarting";       break;
 109         case VSALVAGING:        m = "being salvaged";   break;
 110         default:                m = "busy";             break;
 111         }
 112
 113         pr_notice("kAFS: Volume %llu '%s' is %s\n", volume->vid, volume->name, m);
 114 }
 115
 116 /*
 117  * Sleep and retry the operation to the same fileserver.
 118  */
 119 static bool afs_sleep_and_retry(struct afs_fs_cursor *fc)
 120 {
 121         msleep_interruptible(1000);
 122         if (signal_pending(current)) {
 123                 fc->error = -ERESTARTSYS;
 124                 return false;
 125         }
 126
 127         return true;
 128 }
 129
 130 /*
 131  * Select the fileserver to use.  May be called multiple times to rotate
 132  * through the fileservers.
 133  */
 134 bool afs_select_fileserver(struct afs_fs_cursor *fc)
 135 {
 136         struct afs_addr_list *alist;
 137         struct afs_server *server;
 138         struct afs_vnode *vnode = fc->vnode;
 139         u32 rtt, abort_code;
 140         int error = fc->ac.error, i;
 141
 142         _enter("%lx[%d],%lx[%d],%d,%d",
 143                fc->untried, fc->index,
 144                fc->ac.tried, fc->ac.index,
 145                error, fc->ac.abort_code);
 146
 147         if (fc->flags & AFS_FS_CURSOR_STOP) {
 148                 _leave(" = f [stopped]");
 149                 return false;
 150         }
 151
 152         fc->nr_iterations++;
 153
 154         /* Evaluate the result of the previous operation, if there was one. */
 155         switch (error) {
 156         case SHRT_MAX:
 157                 goto start;
 158
 159         case 0:
 160         default:
 161                 /* Success or local failure.  Stop. */
 162                 fc->error = error;
 163                 fc->flags |= AFS_FS_CURSOR_STOP;
 164                 _leave(" = f [okay/local %d]", error);
 165                 return false;
 166
 167         case -ECONNABORTED:
 168                 /* The far side rejected the operation on some grounds.  This
 169                  * might involve the server being busy or the volume having been moved.
 170                  */
 171                 switch (fc->ac.abort_code) {
 172                 case VNOVOL:
 173                         /* This fileserver doesn't know about the volume.
 174                          * - May indicate that the VL is wrong - retry once and compare
 175                          *   the results.
 176                          * - May indicate that the fileserver couldn't attach to the vol.
 177                          */
 178                         if (fc->flags & AFS_FS_CURSOR_VNOVOL) {
 179                                 fc->error = -EREMOTEIO;
 180                                 goto next_server;
 181                         }
 182
 183                         write_lock(&vnode->volume->servers_lock);
 184                         fc->server_list->vnovol_mask |= 1 << fc->index;
 185                         write_unlock(&vnode->volume->servers_lock);
 186
 187                         set_bit(AFS_VOLUME_NEEDS_UPDATE, &vnode->volume->flags);
 188                         error = afs_check_volume_status(vnode->volume, fc->key);
 189                         if (error < 0)
 190                                 goto failed_set_error;
 191
 192                         if (test_bit(AFS_VOLUME_DELETED, &vnode->volume->flags)) {
 193                                 fc->error = -ENOMEDIUM;
 194                                 goto failed;
 195                         }
 196
 197                         /* If the server list didn't change, then assume that
 198                          * it's the fileserver having trouble.
 199                          */
 200                         if (vnode->volume->servers == fc->server_list) {
 201                                 fc->error = -EREMOTEIO;
 202                                 goto next_server;
 203                         }
 204
 205                         /* Try again */
 206                         fc->flags |= AFS_FS_CURSOR_VNOVOL;
 207                         _leave(" = t [vnovol]");
 208                         return true;
 209
 210                 case VSALVAGE: /* TODO: Should this return an error or iterate? */
 211                 case VVOLEXISTS:
 212                 case VNOSERVICE:
 213                 case VONLINE:
 214                 case VDISKFULL:
 215                 case VOVERQUOTA:
 216                         fc->error = afs_abort_to_error(fc->ac.abort_code);
 217                         goto next_server;
 218
 219                 case VOFFLINE:
 220                         if (!test_and_set_bit(AFS_VOLUME_OFFLINE, &vnode->volume->flags)) {
 221                                 afs_busy(vnode->volume, fc->ac.abort_code);
 222                                 clear_bit(AFS_VOLUME_BUSY, &vnode->volume->flags);
 223                         }
 224                         if (fc->flags & AFS_FS_CURSOR_NO_VSLEEP) {
 225                                 fc->error = -EADV;
 226                                 goto failed;
 227                         }
 228                         if (fc->flags & AFS_FS_CURSOR_CUR_ONLY) {
 229                                 fc->error = -ESTALE;
 230                                 goto failed;
 231                         }
 232                         goto busy;
 233
 234                 case VSALVAGING:
 235                 case VRESTARTING:
 236                 case VBUSY:
 237                         /* Retry after going round all the servers unless we
 238                          * have a file lock we need to maintain.
 239                          */
 240                         if (fc->flags & AFS_FS_CURSOR_NO_VSLEEP) {
 241                                 fc->error = -EBUSY;
 242                                 goto failed;
 243                         }
 244                         if (!test_and_set_bit(AFS_VOLUME_BUSY, &vnode->volume->flags)) {
 245                                 afs_busy(vnode->volume, fc->ac.abort_code);
 246                                 clear_bit(AFS_VOLUME_OFFLINE, &vnode->volume->flags);
 247                         }
 248                 busy:
 249                         if (fc->flags & AFS_FS_CURSOR_CUR_ONLY) {
 250                                 if (!afs_sleep_and_retry(fc))
 251                                         goto failed;
 252
 253                                  /* Retry with same server & address */
 254                                 _leave(" = t [vbusy]");
 255                                 return true;
 256                         }
 257
 258                         fc->flags |= AFS_FS_CURSOR_VBUSY;
 259                         goto next_server;
 260
 261                 case VMOVED:
 262                         /* The volume migrated to another server.  We consider
 263                          * consider all locks and callbacks broken and request
 264                          * an update from the VLDB.
 265                          *
 266                          * We also limit the number of VMOVED hops we will
 267                          * honour, just in case someone sets up a loop.
 268                          */
 269                         if (fc->flags & AFS_FS_CURSOR_VMOVED) {
 270                                 fc->error = -EREMOTEIO;
 271                                 goto failed;
 272                         }
 273                         fc->flags |= AFS_FS_CURSOR_VMOVED;
 274
 275                         set_bit(AFS_VOLUME_WAIT, &vnode->volume->flags);
 276                         set_bit(AFS_VOLUME_NEEDS_UPDATE, &vnode->volume->flags);
 277                         error = afs_check_volume_status(vnode->volume, fc->key);
 278                         if (error < 0)
 279                                 goto failed_set_error;
 280
 281                         /* If the server list didn't change, then the VLDB is
 282                          * out of sync with the fileservers.  This is hopefully
 283                          * a temporary condition, however, so we don't want to
 284                          * permanently block access to the file.
 285                          *
 286                          * TODO: Try other fileservers if we can.
 287                          *
 288                          * TODO: Retry a few times with sleeps.
 289                          */
 290                         if (vnode->volume->servers == fc->server_list) {
 291                                 fc->error = -ENOMEDIUM;
 292                                 goto failed;
 293                         }
 294
 295                         goto restart_from_beginning;
 296
 297                 default:
 298                         clear_bit(AFS_VOLUME_OFFLINE, &vnode->volume->flags);
 299                         clear_bit(AFS_VOLUME_BUSY, &vnode->volume->flags);
 300                         fc->error = afs_abort_to_error(fc->ac.abort_code);
 301                         goto failed;
 302                 }
 303
 304         case -ETIMEDOUT:
 305         case -ETIME:
 306                 if (fc->error != -EDESTADDRREQ)
 307                         goto iterate_address;
 308                 /* Fall through */
 309         case -ENETUNREACH:
 310         case -EHOSTUNREACH:
 311         case -ECONNREFUSED:
 312                 _debug("no conn");
 313                 fc->error = error;
 314                 goto iterate_address;
 315
 316         case -ECONNRESET:
 317                 _debug("call reset");
 318                 fc->error = error;
 319                 goto failed;
 320         }
 321
 322 restart_from_beginning:
 323         _debug("restart");
 324         afs_end_cursor(&fc->ac);
 325         afs_put_cb_interest(afs_v2net(vnode), fc->cbi);
 326         fc->cbi = NULL;
 327         afs_put_serverlist(afs_v2net(vnode), fc->server_list);
 328         fc->server_list = NULL;
 329 start:
 330         _debug("start");
 331         /* See if we need to do an update of the volume record.  Note that the
 332          * volume may have moved or even have been deleted.
 333          */
 334         error = afs_check_volume_status(vnode->volume, fc->key);
 335         if (error < 0)
 336                 goto failed_set_error;
 337
 338         if (!afs_start_fs_iteration(fc, vnode))
 339                 goto failed;
 340
 341         _debug("__ VOL %llx __", vnode->volume->vid);
 342         error = afs_probe_fileservers(afs_v2net(vnode), fc->key, fc->server_list);
 343         if (error < 0)
 344                 goto failed_set_error;
 345
 346 pick_server:
 347         _debug("pick [%lx]", fc->untried);
 348
 349         error = afs_wait_for_fs_probes(fc->server_list, fc->untried);
 350         if (error < 0)
 351                 goto failed_set_error;
 352
 353         /* Pick the untried server with the lowest RTT.  If we have outstanding
 354          * callbacks, we stick with the server we're already using if we can.
 355          */
 356         if (fc->cbi) {
 357                 _debug("cbi %u", fc->index);
 358                 if (test_bit(fc->index, &fc->untried))
 359                         goto selected_server;
 360                 afs_put_cb_interest(afs_v2net(vnode), fc->cbi);
 361                 fc->cbi = NULL;
 362                 _debug("nocbi");
 363         }
 364
 365         fc->index = -1;
 366         rtt = U32_MAX;
 367         for (i = 0; i < fc->server_list->nr_servers; i++) {
 368                 struct afs_server *s = fc->server_list->servers[i].server;
 369
 370                 if (!test_bit(i, &fc->untried) || !s->probe.responded)
 371                         continue;
 372                 if (s->probe.rtt < rtt) {
 373                         fc->index = i;
 374                         rtt = s->probe.rtt;
 375                 }
 376         }
 377
 378         if (fc->index == -1)
 379                 goto no_more_servers;
 380
 381 selected_server:
 382         _debug("use %d", fc->index);
 383         __clear_bit(fc->index, &fc->untried);
 384
 385         /* We're starting on a different fileserver from the list.  We need to
 386          * check it, create a callback intercept, find its address list and
 387          * probe its capabilities before we use it.
 388          */
 389         ASSERTCMP(fc->ac.alist, ==, NULL);
 390         server = fc->server_list->servers[fc->index].server;
 391
 392         if (!afs_check_server_record(fc, server))
 393                 goto failed;
 394
 395         _debug("USING SERVER: %pU", &server->uuid);
 396
 397         /* Make sure we've got a callback interest record for this server.  We
 398          * have to link it in before we send the request as we can be sent a
 399          * break request before we've finished decoding the reply and
 400          * installing the vnode.
 401          */
 402         error = afs_register_server_cb_interest(vnode, fc->server_list,
 403                                                 fc->index);
 404         if (error < 0)
 405                 goto failed_set_error;
 406
 407         fc->cbi = afs_get_cb_interest(vnode->cb_interest);
 408
 409         read_lock(&server->fs_lock);
 410         alist = rcu_dereference_protected(server->addresses,
 411                                           lockdep_is_held(&server->fs_lock));
 412         afs_get_addrlist(alist);
 413         read_unlock(&server->fs_lock);
 414
 415         memset(&fc->ac, 0, sizeof(fc->ac));
 416
 417         if (!fc->ac.alist)
 418                 fc->ac.alist = alist;
 419         else
 420                 afs_put_addrlist(alist);
 421
 422         fc->ac.index = -1;
 423
 424 iterate_address:
 425         ASSERT(fc->ac.alist);
 426         /* Iterate over the current server's address list to try and find an
 427          * address on which it will respond to us.
 428          */
 429         if (!afs_iterate_addresses(&fc->ac))
 430                 goto next_server;
 431
 432         _debug("address [%u] %u/%u", fc->index, fc->ac.index, fc->ac.alist->nr_addrs);
 433
 434         _leave(" = t");
 435         return true;
 436
 437 next_server:
 438         _debug("next");
 439         afs_end_cursor(&fc->ac);
 440         goto pick_server;
 441
 442 no_more_servers:
 443         /* That's all the servers poked to no good effect.  Try again if some
 444          * of them were busy.
 445          */
 446         if (fc->flags & AFS_FS_CURSOR_VBUSY)
 447                 goto restart_from_beginning;
 448
 449         abort_code = 0;
 450         error = -EDESTADDRREQ;
 451         for (i = 0; i < fc->server_list->nr_servers; i++) {
 452                 struct afs_server *s = fc->server_list->servers[i].server;
 453                 int probe_error = READ_ONCE(s->probe.error);
 454
 455                 switch (probe_error) {
 456                 case 0:
 457                         continue;
 458                 default:
 459                         if (error == -ETIMEDOUT ||
 460                             error == -ETIME)
 461                                 continue;
 462                 case -ETIMEDOUT:
 463                 case -ETIME:
 464                         if (error == -ENOMEM ||
 465                             error == -ENONET)
 466                                 continue;
 467                 case -ENOMEM:
 468                 case -ENONET:
 469                         if (error == -ENETUNREACH)
 470                                 continue;
 471                 case -ENETUNREACH:
 472                         if (error == -EHOSTUNREACH)
 473                                 continue;
 474                 case -EHOSTUNREACH:
 475                         if (error == -ECONNREFUSED)
 476                                 continue;
 477                 case -ECONNREFUSED:
 478                         if (error == -ECONNRESET)
 479                                 continue;
 480                 case -ECONNRESET: /* Responded, but call expired. */
 481                         if (error == -ECONNABORTED)
 482                                 continue;
 483                 case -ECONNABORTED:
 484                         abort_code = s->probe.abort_code;
 485                         error = probe_error;
 486                         continue;
 487                 }
 488         }
 489
 490         if (error == -ECONNABORTED)
 491                 error = afs_abort_to_error(abort_code);
 492
 493 failed_set_error:
 494         fc->error = error;
 495 failed:
 496         fc->flags |= AFS_FS_CURSOR_STOP;
 497         afs_end_cursor(&fc->ac);
 498         _leave(" = f [failed %d]", fc->error);
 499         return false;
 500 }
 501
 502 /*
 503  * Select the same fileserver we used for a vnode before and only that
 504  * fileserver.  We use this when we have a lock on that file, which is backed
 505  * only by the fileserver we obtained it from.
 506  */
 507 bool afs_select_current_fileserver(struct afs_fs_cursor *fc)
 508 {
 509         struct afs_vnode *vnode = fc->vnode;
 510         struct afs_cb_interest *cbi = vnode->cb_interest;
 511         struct afs_addr_list *alist;
 512         int error = fc->ac.error;
 513
 514         _enter("");
 515
 516         switch (error) {
 517         case SHRT_MAX:
 518                 if (!cbi) {
 519                         fc->error = -ESTALE;
 520                         fc->flags |= AFS_FS_CURSOR_STOP;
 521                         return false;
 522                 }
 523
 524                 fc->cbi = afs_get_cb_interest(vnode->cb_interest);
 525
 526                 read_lock(&cbi->server->fs_lock);
 527                 alist = rcu_dereference_protected(cbi->server->addresses,
 528                                                   lockdep_is_held(&cbi->server->fs_lock));
 529                 afs_get_addrlist(alist);
 530                 read_unlock(&cbi->server->fs_lock);
 531                 if (!alist) {
 532                         fc->error = -ESTALE;
 533                         fc->flags |= AFS_FS_CURSOR_STOP;
 534                         return false;
 535                 }
 536
 537                 memset(&fc->ac, 0, sizeof(fc->ac));
 538                 fc->ac.alist = alist;
 539                 fc->ac.index = -1;
 540                 goto iterate_address;
 541
 542         case 0:
 543         default:
 544                 /* Success or local failure.  Stop. */
 545                 fc->error = error;
 546                 fc->flags |= AFS_FS_CURSOR_STOP;
 547                 _leave(" = f [okay/local %d]", error);
 548                 return false;
 549
 550         case -ECONNABORTED:
 551                 fc->error = afs_abort_to_error(fc->ac.abort_code);
 552                 fc->flags |= AFS_FS_CURSOR_STOP;
 553                 _leave(" = f [abort]");
 554                 return false;
 555
 556         case -ENETUNREACH:
 557         case -EHOSTUNREACH:
 558         case -ECONNREFUSED:
 559         case -ETIMEDOUT:
 560         case -ETIME:
 561                 _debug("no conn");
 562                 fc->error = error;
 563                 goto iterate_address;
 564         }
 565
 566 iterate_address:
 567         /* Iterate over the current server's address list to try and find an
 568          * address on which it will respond to us.
 569          */
 570         if (afs_iterate_addresses(&fc->ac)) {
 571                 _leave(" = t");
 572                 return true;
 573         }
 574
 575         afs_end_cursor(&fc->ac);
 576         return false;
 577 }
 578
 579 /*
 580  * Dump cursor state in the case of the error being EDESTADDRREQ.
 581  */
 582 static void afs_dump_edestaddrreq(const struct afs_fs_cursor *fc)
 583 {
 584         static int count;
 585         int i;
 586
 587         if (!IS_ENABLED(CONFIG_AFS_DEBUG_CURSOR) || count > 3)
 588                 return;
 589         count++;
 590
 591         rcu_read_lock();
 592
 593         pr_notice("EDESTADDR occurred\n");
 594         pr_notice("FC: cbb=%x cbb2=%x fl=%hx err=%hd\n",
 595                   fc->cb_break, fc->cb_break_2, fc->flags, fc->error);
 596         pr_notice("FC: ut=%lx ix=%d ni=%u\n",
 597                   fc->untried, fc->index, fc->nr_iterations);
 598
 599         if (fc->server_list) {
 600                 const struct afs_server_list *sl = fc->server_list;
 601                 pr_notice("FC: SL nr=%u pr=%u vnov=%hx\n",
 602                           sl->nr_servers, sl->preferred, sl->vnovol_mask);
 603                 for (i = 0; i < sl->nr_servers; i++) {
 604                         const struct afs_server *s = sl->servers[i].server;
 605                         pr_notice("FC: server fl=%lx av=%u %pU\n",
 606                                   s->flags, s->addr_version, &s->uuid);
 607                         if (s->addresses) {
 608                                 const struct afs_addr_list *a =
 609                                         rcu_dereference(s->addresses);
 610                                 pr_notice("FC:  - av=%u nr=%u/%u/%u pr=%u\n",
 611                                           a->version,
 612                                           a->nr_ipv4, a->nr_addrs, a->max_addrs,
 613                                           a->preferred);
 614                                 pr_notice("FC:  - pr=%lx R=%lx F=%lx\n",
 615                                           a->probed, a->responded, a->failed);
 616                                 if (a == fc->ac.alist)
 617                                         pr_notice("FC:  - current\n");
 618                         }
 619                 }
 620         }
 621
 622         pr_notice("AC: t=%lx ax=%u ac=%d er=%d r=%u ni=%u\n",
 623                   fc->ac.tried, fc->ac.index, fc->ac.abort_code, fc->ac.error,
 624                   fc->ac.responded, fc->ac.nr_iterations);
 625         rcu_read_unlock();
 626 }
 627
 628 /*
 629  * Tidy up a filesystem cursor and unlock the vnode.
 630  */
 631 int afs_end_vnode_operation(struct afs_fs_cursor *fc)
 632 {
 633         struct afs_net *net = afs_v2net(fc->vnode);
 634
 635         if (fc->error == -EDESTADDRREQ ||
 636             fc->error == -ENETUNREACH ||
 637             fc->error == -EHOSTUNREACH)
 638                 afs_dump_edestaddrreq(fc);
 639
 640         mutex_unlock(&fc->vnode->io_lock);
 641
 642         afs_end_cursor(&fc->ac);
 643         afs_put_cb_interest(net, fc->cbi);
 644         afs_put_serverlist(net, fc->server_list);
 645
 646         if (fc->error == -ECONNABORTED)
 647                 fc->error = afs_abort_to_error(fc->ac.abort_code);
 648
 649         return fc->error;
 650 }