fs/afs/rotate.c

   1 // SPDX-License-Identifier: GPL-2.0-or-later
   2 /* Handle fileserver selection and rotation.
   3  *
   4  * Copyright (C) 2017 Red Hat, Inc. All Rights Reserved.
   5  * Written by David Howells (dhowells@redhat.com)
   6  */
   7
   8 #include <linux/kernel.h>
   9 #include <linux/slab.h>
  10 #include <linux/fs.h>
  11 #include <linux/sched.h>
  12 #include <linux/delay.h>
  13 #include <linux/sched/signal.h>
  14 #include "internal.h"
  15 #include "afs_fs.h"
  16
  17 /*
  18  * Begin iteration through a server list, starting with the vnode's last used
  19  * server if possible, or the last recorded good server if not.
  20  */
  21 static bool afs_start_fs_iteration(struct afs_operation *op,
  22                                    struct afs_vnode *vnode)
  23 {
  24         struct afs_server *server;
  25         void *cb_server;
  26         int i;
  27
  28         read_lock(&op->volume->servers_lock);
  29         op->server_list = afs_get_serverlist(
  30                 rcu_dereference_protected(op->volume->servers,
  31                                           lockdep_is_held(&op->volume->servers_lock)));
  32         read_unlock(&op->volume->servers_lock);
  33
  34         op->untried = (1UL << op->server_list->nr_servers) - 1;
  35         op->index = READ_ONCE(op->server_list->preferred);
  36
  37         cb_server = vnode->cb_server;
  38         if (cb_server) {
  39                 /* See if the vnode's preferred record is still available */
  40                 for (i = 0; i < op->server_list->nr_servers; i++) {
  41                         server = op->server_list->servers[i].server;
  42                         if (server == cb_server) {
  43                                 op->index = i;
  44                                 goto found_interest;
  45                         }
  46                 }
  47
  48                 /* If we have a lock outstanding on a server that's no longer
  49                  * serving this vnode, then we can't switch to another server
  50                  * and have to return an error.
  51                  */
  52                 if (op->flags & AFS_OPERATION_CUR_ONLY) {
  53                         op->error = -ESTALE;
  54                         return false;
  55                 }
  56
  57                 /* Note that the callback promise is effectively broken */
  58                 write_seqlock(&vnode->cb_lock);
  59                 ASSERTCMP(cb_server, ==, vnode->cb_server);
  60                 vnode->cb_server = NULL;
  61                 if (test_and_clear_bit(AFS_VNODE_CB_PROMISED, &vnode->flags))
  62                         vnode->cb_break++;
  63                 write_sequnlock(&vnode->cb_lock);
  64         }
  65
  66 found_interest:
  67         return true;
  68 }
  69
  70 /*
  71  * Post volume busy note.
  72  */
  73 static void afs_busy(struct afs_volume *volume, u32 abort_code)
  74 {
  75         const char *m;
  76
  77         switch (abort_code) {
  78         case VOFFLINE:          m = "offline";          break;
  79         case VRESTARTING:       m = "restarting";       break;
  80         case VSALVAGING:        m = "being salvaged";   break;
  81         default:                m = "busy";             break;
  82         }
  83
  84         pr_notice("kAFS: Volume %llu '%s' is %s\n", volume->vid, volume->name, m);
  85 }
  86
  87 /*
  88  * Sleep and retry the operation to the same fileserver.
  89  */
  90 static bool afs_sleep_and_retry(struct afs_operation *op)
  91 {
  92         if (!(op->flags & AFS_OPERATION_UNINTR)) {
  93                 msleep_interruptible(1000);
  94                 if (signal_pending(current)) {
  95                         op->error = -ERESTARTSYS;
  96                         return false;
  97                 }
  98         } else {
  99                 msleep(1000);
 100         }
 101
 102         return true;
 103 }
 104
 105 /*
 106  * Select the fileserver to use.  May be called multiple times to rotate
 107  * through the fileservers.
 108  */
 109 bool afs_select_fileserver(struct afs_operation *op)
 110 {
 111         struct afs_addr_list *alist;
 112         struct afs_server *server;
 113         struct afs_vnode *vnode = op->file[0].vnode;
 114         struct afs_error e;
 115         u32 rtt;
 116         int error = op->ac.error, i;
 117
 118         _enter("%lx[%d],%lx[%d],%d,%d",
 119                op->untried, op->index,
 120                op->ac.tried, op->ac.index,
 121                error, op->ac.abort_code);
 122
 123         if (op->flags & AFS_OPERATION_STOP) {
 124                 _leave(" = f [stopped]");
 125                 return false;
 126         }
 127
 128         op->nr_iterations++;
 129
 130         /* Evaluate the result of the previous operation, if there was one. */
 131         switch (error) {
 132         case SHRT_MAX:
 133                 goto start;
 134
 135         case 0:
 136         default:
 137                 /* Success or local failure.  Stop. */
 138                 op->error = error;
 139                 op->flags |= AFS_OPERATION_STOP;
 140                 _leave(" = f [okay/local %d]", error);
 141                 return false;
 142
 143         case -ECONNABORTED:
 144                 /* The far side rejected the operation on some grounds.  This
 145                  * might involve the server being busy or the volume having been moved.
 146                  */
 147                 switch (op->ac.abort_code) {
 148                 case VNOVOL:
 149                         /* This fileserver doesn't know about the volume.
 150                          * - May indicate that the VL is wrong - retry once and compare
 151                          *   the results.
 152                          * - May indicate that the fileserver couldn't attach to the vol.
 153                          */
 154                         if (op->flags & AFS_OPERATION_VNOVOL) {
 155                                 op->error = -EREMOTEIO;
 156                                 goto next_server;
 157                         }
 158
 159                         write_lock(&op->volume->servers_lock);
 160                         op->server_list->vnovol_mask |= 1 << op->index;
 161                         write_unlock(&op->volume->servers_lock);
 162
 163                         set_bit(AFS_VOLUME_NEEDS_UPDATE, &op->volume->flags);
 164                         error = afs_check_volume_status(op->volume, op);
 165                         if (error < 0)
 166                                 goto failed_set_error;
 167
 168                         if (test_bit(AFS_VOLUME_DELETED, &op->volume->flags)) {
 169                                 op->error = -ENOMEDIUM;
 170                                 goto failed;
 171                         }
 172
 173                         /* If the server list didn't change, then assume that
 174                          * it's the fileserver having trouble.
 175                          */
 176                         if (rcu_access_pointer(op->volume->servers) == op->server_list) {
 177                                 op->error = -EREMOTEIO;
 178                                 goto next_server;
 179                         }
 180
 181                         /* Try again */
 182                         op->flags |= AFS_OPERATION_VNOVOL;
 183                         _leave(" = t [vnovol]");
 184                         return true;
 185
 186                 case VSALVAGE: /* TODO: Should this return an error or iterate? */
 187                 case VVOLEXISTS:
 188                 case VNOSERVICE:
 189                 case VONLINE:
 190                 case VDISKFULL:
 191                 case VOVERQUOTA:
 192                         op->error = afs_abort_to_error(op->ac.abort_code);
 193                         goto next_server;
 194
 195                 case VOFFLINE:
 196                         if (!test_and_set_bit(AFS_VOLUME_OFFLINE, &op->volume->flags)) {
 197                                 afs_busy(op->volume, op->ac.abort_code);
 198                                 clear_bit(AFS_VOLUME_BUSY, &op->volume->flags);
 199                         }
 200                         if (op->flags & AFS_OPERATION_NO_VSLEEP) {
 201                                 op->error = -EADV;
 202                                 goto failed;
 203                         }
 204                         if (op->flags & AFS_OPERATION_CUR_ONLY) {
 205                                 op->error = -ESTALE;
 206                                 goto failed;
 207                         }
 208                         goto busy;
 209
 210                 case VSALVAGING:
 211                 case VRESTARTING:
 212                 case VBUSY:
 213                         /* Retry after going round all the servers unless we
 214                          * have a file lock we need to maintain.
 215                          */
 216                         if (op->flags & AFS_OPERATION_NO_VSLEEP) {
 217                                 op->error = -EBUSY;
 218                                 goto failed;
 219                         }
 220                         if (!test_and_set_bit(AFS_VOLUME_BUSY, &op->volume->flags)) {
 221                                 afs_busy(op->volume, op->ac.abort_code);
 222                                 clear_bit(AFS_VOLUME_OFFLINE, &op->volume->flags);
 223                         }
 224                 busy:
 225                         if (op->flags & AFS_OPERATION_CUR_ONLY) {
 226                                 if (!afs_sleep_and_retry(op))
 227                                         goto failed;
 228
 229                                  /* Retry with same server & address */
 230                                 _leave(" = t [vbusy]");
 231                                 return true;
 232                         }
 233
 234                         op->flags |= AFS_OPERATION_VBUSY;
 235                         goto next_server;
 236
 237                 case VMOVED:
 238                         /* The volume migrated to another server.  We consider
 239                          * consider all locks and callbacks broken and request
 240                          * an update from the VLDB.
 241                          *
 242                          * We also limit the number of VMOVED hops we will
 243                          * honour, just in case someone sets up a loop.
 244                          */
 245                         if (op->flags & AFS_OPERATION_VMOVED) {
 246                                 op->error = -EREMOTEIO;
 247                                 goto failed;
 248                         }
 249                         op->flags |= AFS_OPERATION_VMOVED;
 250
 251                         set_bit(AFS_VOLUME_WAIT, &op->volume->flags);
 252                         set_bit(AFS_VOLUME_NEEDS_UPDATE, &op->volume->flags);
 253                         error = afs_check_volume_status(op->volume, op);
 254                         if (error < 0)
 255                                 goto failed_set_error;
 256
 257                         /* If the server list didn't change, then the VLDB is
 258                          * out of sync with the fileservers.  This is hopefully
 259                          * a temporary condition, however, so we don't want to
 260                          * permanently block access to the file.
 261                          *
 262                          * TODO: Try other fileservers if we can.
 263                          *
 264                          * TODO: Retry a few times with sleeps.
 265                          */
 266                         if (rcu_access_pointer(op->volume->servers) == op->server_list) {
 267                                 op->error = -ENOMEDIUM;
 268                                 goto failed;
 269                         }
 270
 271                         goto restart_from_beginning;
 272
 273                 default:
 274                         clear_bit(AFS_VOLUME_OFFLINE, &op->volume->flags);
 275                         clear_bit(AFS_VOLUME_BUSY, &op->volume->flags);
 276                         op->error = afs_abort_to_error(op->ac.abort_code);
 277                         goto failed;
 278                 }
 279
 280         case -ETIMEDOUT:
 281         case -ETIME:
 282                 if (op->error != -EDESTADDRREQ)
 283                         goto iterate_address;
 284                 fallthrough;
 285         case -ERFKILL:
 286         case -EADDRNOTAVAIL:
 287         case -ENETUNREACH:
 288         case -EHOSTUNREACH:
 289         case -EHOSTDOWN:
 290         case -ECONNREFUSED:
 291                 _debug("no conn");
 292                 op->error = error;
 293                 goto iterate_address;
 294
 295         case -ECONNRESET:
 296                 _debug("call reset");
 297                 op->error = error;
 298                 goto failed;
 299         }
 300
 301 restart_from_beginning:
 302         _debug("restart");
 303         afs_end_cursor(&op->ac);
 304         op->server = NULL;
 305         afs_put_serverlist(op->net, op->server_list);
 306         op->server_list = NULL;
 307 start:
 308         _debug("start");
 309         /* See if we need to do an update of the volume record.  Note that the
 310          * volume may have moved or even have been deleted.
 311          */
 312         error = afs_check_volume_status(op->volume, op);
 313         if (error < 0)
 314                 goto failed_set_error;
 315
 316         if (!afs_start_fs_iteration(op, vnode))
 317                 goto failed;
 318
 319         _debug("__ VOL %llx __", op->volume->vid);
 320
 321 pick_server:
 322         _debug("pick [%lx]", op->untried);
 323
 324         error = afs_wait_for_fs_probes(op->server_list, op->untried);
 325         if (error < 0)
 326                 goto failed_set_error;
 327
 328         /* Pick the untried server with the lowest RTT.  If we have outstanding
 329          * callbacks, we stick with the server we're already using if we can.
 330          */
 331         if (op->server) {
 332                 _debug("server %u", op->index);
 333                 if (test_bit(op->index, &op->untried))
 334                         goto selected_server;
 335                 op->server = NULL;
 336                 _debug("no server");
 337         }
 338
 339         op->index = -1;
 340         rtt = U32_MAX;
 341         for (i = 0; i < op->server_list->nr_servers; i++) {
 342                 struct afs_server *s = op->server_list->servers[i].server;
 343
 344                 if (!test_bit(i, &op->untried) ||
 345                     !test_bit(AFS_SERVER_FL_RESPONDING, &s->flags))
 346                         continue;
 347                 if (s->probe.rtt < rtt) {
 348                         op->index = i;
 349                         rtt = s->probe.rtt;
 350                 }
 351         }
 352
 353         if (op->index == -1)
 354                 goto no_more_servers;
 355
 356 selected_server:
 357         _debug("use %d", op->index);
 358         __clear_bit(op->index, &op->untried);
 359
 360         /* We're starting on a different fileserver from the list.  We need to
 361          * check it, create a callback intercept, find its address list and
 362          * probe its capabilities before we use it.
 363          */
 364         ASSERTCMP(op->ac.alist, ==, NULL);
 365         server = op->server_list->servers[op->index].server;
 366
 367         if (!afs_check_server_record(op, server))
 368                 goto failed;
 369
 370         _debug("USING SERVER: %pU", &server->uuid);
 371
 372         op->flags |= AFS_OPERATION_RETRY_SERVER;
 373         op->server = server;
 374         if (vnode->cb_server != server) {
 375                 vnode->cb_server = server;
 376                 vnode->cb_s_break = server->cb_s_break;
 377                 vnode->cb_fs_s_break = atomic_read(&server->cell->fs_s_break);
 378                 vnode->cb_v_break = vnode->volume->cb_v_break;
 379                 clear_bit(AFS_VNODE_CB_PROMISED, &vnode->flags);
 380         }
 381
 382         read_lock(&server->fs_lock);
 383         alist = rcu_dereference_protected(server->addresses,
 384                                           lockdep_is_held(&server->fs_lock));
 385         afs_get_addrlist(alist);
 386         read_unlock(&server->fs_lock);
 387
 388 retry_server:
 389         memset(&op->ac, 0, sizeof(op->ac));
 390
 391         if (!op->ac.alist)
 392                 op->ac.alist = alist;
 393         else
 394                 afs_put_addrlist(alist);
 395
 396         op->ac.index = -1;
 397
 398 iterate_address:
 399         ASSERT(op->ac.alist);
 400         /* Iterate over the current server's address list to try and find an
 401          * address on which it will respond to us.
 402          */
 403         if (!afs_iterate_addresses(&op->ac))
 404                 goto out_of_addresses;
 405
 406         _debug("address [%u] %u/%u %pISp",
 407                op->index, op->ac.index, op->ac.alist->nr_addrs,
 408                &op->ac.alist->addrs[op->ac.index].transport);
 409
 410         _leave(" = t");
 411         return true;
 412
 413 out_of_addresses:
 414         /* We've now had a failure to respond on all of a server's addresses -
 415          * immediately probe them again and consider retrying the server.
 416          */
 417         afs_probe_fileserver(op->net, op->server);
 418         if (op->flags & AFS_OPERATION_RETRY_SERVER) {
 419                 alist = op->ac.alist;
 420                 error = afs_wait_for_one_fs_probe(
 421                         op->server, !(op->flags & AFS_OPERATION_UNINTR));
 422                 switch (error) {
 423                 case 0:
 424                         op->flags &= ~AFS_OPERATION_RETRY_SERVER;
 425                         goto retry_server;
 426                 case -ERESTARTSYS:
 427                         goto failed_set_error;
 428                 case -ETIME:
 429                 case -EDESTADDRREQ:
 430                         goto next_server;
 431                 }
 432         }
 433
 434 next_server:
 435         _debug("next");
 436         afs_end_cursor(&op->ac);
 437         goto pick_server;
 438
 439 no_more_servers:
 440         /* That's all the servers poked to no good effect.  Try again if some
 441          * of them were busy.
 442          */
 443         if (op->flags & AFS_OPERATION_VBUSY)
 444                 goto restart_from_beginning;
 445
 446         e.error = -EDESTADDRREQ;
 447         e.responded = false;
 448         for (i = 0; i < op->server_list->nr_servers; i++) {
 449                 struct afs_server *s = op->server_list->servers[i].server;
 450
 451                 afs_prioritise_error(&e, READ_ONCE(s->probe.error),
 452                                      s->probe.abort_code);
 453         }
 454
 455         error = e.error;
 456
 457 failed_set_error:
 458         op->error = error;
 459 failed:
 460         op->flags |= AFS_OPERATION_STOP;
 461         afs_end_cursor(&op->ac);
 462         _leave(" = f [failed %d]", op->error);
 463         return false;
 464 }
 465
 466 /*
 467  * Dump cursor state in the case of the error being EDESTADDRREQ.
 468  */
 469 void afs_dump_edestaddrreq(const struct afs_operation *op)
 470 {
 471         static int count;
 472         int i;
 473
 474         if (!IS_ENABLED(CONFIG_AFS_DEBUG_CURSOR) || count > 3)
 475                 return;
 476         count++;
 477
 478         rcu_read_lock();
 479
 480         pr_notice("EDESTADDR occurred\n");
 481         pr_notice("FC: cbb=%x cbb2=%x fl=%x err=%hd\n",
 482                   op->file[0].cb_break_before,
 483                   op->file[1].cb_break_before, op->flags, op->error);
 484         pr_notice("FC: ut=%lx ix=%d ni=%u\n",
 485                   op->untried, op->index, op->nr_iterations);
 486
 487         if (op->server_list) {
 488                 const struct afs_server_list *sl = op->server_list;
 489                 pr_notice("FC: SL nr=%u pr=%u vnov=%hx\n",
 490                           sl->nr_servers, sl->preferred, sl->vnovol_mask);
 491                 for (i = 0; i < sl->nr_servers; i++) {
 492                         const struct afs_server *s = sl->servers[i].server;
 493                         pr_notice("FC: server fl=%lx av=%u %pU\n",
 494                                   s->flags, s->addr_version, &s->uuid);
 495                         if (s->addresses) {
 496                                 const struct afs_addr_list *a =
 497                                         rcu_dereference(s->addresses);
 498                                 pr_notice("FC:  - av=%u nr=%u/%u/%u pr=%u\n",
 499                                           a->version,
 500                                           a->nr_ipv4, a->nr_addrs, a->max_addrs,
 501                                           a->preferred);
 502                                 pr_notice("FC:  - R=%lx F=%lx\n",
 503                                           a->responded, a->failed);
 504                                 if (a == op->ac.alist)
 505                                         pr_notice("FC:  - current\n");
 506                         }
 507                 }
 508         }
 509
 510         pr_notice("AC: t=%lx ax=%u ac=%d er=%d r=%u ni=%u\n",
 511                   op->ac.tried, op->ac.index, op->ac.abort_code, op->ac.error,
 512                   op->ac.responded, op->ac.nr_iterations);
 513         rcu_read_unlock();
 514 }