fs/afs/server.c

   1 /* AFS server record management
   2  *
   3  * Copyright (C) 2002, 2007 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 License
   8  * as published by the Free Software Foundation; either version
   9  * 2 of the License, or (at your option) any later version.
  10  */
  11
  12 #include <linux/sched.h>
  13 #include <linux/slab.h>
  14 #include "afs_fs.h"
  15 #include "internal.h"
  16 #include "protocol_yfs.h"
  17
  18 static unsigned afs_server_gc_delay = 10;       /* Server record timeout in seconds */
  19 static unsigned afs_server_update_delay = 30;   /* Time till VLDB recheck in secs */
  20
  21 static void afs_inc_servers_outstanding(struct afs_net *net)
  22 {
  23         atomic_inc(&net->servers_outstanding);
  24 }
  25
  26 static void afs_dec_servers_outstanding(struct afs_net *net)
  27 {
  28         if (atomic_dec_and_test(&net->servers_outstanding))
  29                 wake_up_var(&net->servers_outstanding);
  30 }
  31
  32 /*
  33  * Find a server by one of its addresses.
  34  */
  35 struct afs_server *afs_find_server(struct afs_net *net,
  36                                    const struct sockaddr_rxrpc *srx)
  37 {
  38         const struct sockaddr_in6 *a = &srx->transport.sin6, *b;
  39         const struct afs_addr_list *alist;
  40         struct afs_server *server = NULL;
  41         unsigned int i;
  42         bool ipv6 = true;
  43         int seq = 0, diff;
  44
  45         if (srx->transport.sin6.sin6_addr.s6_addr32[0] == 0 ||
  46             srx->transport.sin6.sin6_addr.s6_addr32[1] == 0 ||
  47             srx->transport.sin6.sin6_addr.s6_addr32[2] == htonl(0xffff))
  48                 ipv6 = false;
  49
  50         rcu_read_lock();
  51
  52         do {
  53                 if (server)
  54                         afs_put_server(net, server);
  55                 server = NULL;
  56                 read_seqbegin_or_lock(&net->fs_addr_lock, &seq);
  57
  58                 if (ipv6) {
  59                         hlist_for_each_entry_rcu(server, &net->fs_addresses6, addr6_link) {
  60                                 alist = rcu_dereference(server->addresses);
  61                                 for (i = alist->nr_ipv4; i < alist->nr_addrs; i++) {
  62                                         b = &alist->addrs[i].transport.sin6;
  63                                         diff = ((u16 __force)a->sin6_port -
  64                                                 (u16 __force)b->sin6_port);
  65                                         if (diff == 0)
  66                                                 diff = memcmp(&a->sin6_addr,
  67                                                               &b->sin6_addr,
  68                                                               sizeof(struct in6_addr));
  69                                         if (diff == 0)
  70                                                 goto found;
  71                                 }
  72                         }
  73                 } else {
  74                         hlist_for_each_entry_rcu(server, &net->fs_addresses4, addr4_link) {
  75                                 alist = rcu_dereference(server->addresses);
  76                                 for (i = 0; i < alist->nr_ipv4; i++) {
  77                                         b = &alist->addrs[i].transport.sin6;
  78                                         diff = ((u16 __force)a->sin6_port -
  79                                                 (u16 __force)b->sin6_port);
  80                                         if (diff == 0)
  81                                                 diff = ((u32 __force)a->sin6_addr.s6_addr32[3] -
  82                                                         (u32 __force)b->sin6_addr.s6_addr32[3]);
  83                                         if (diff == 0)
  84                                                 goto found;
  85                                 }
  86                         }
  87                 }
  88
  89                 server = NULL;
  90         found:
  91                 if (server && !atomic_inc_not_zero(&server->usage))
  92                         server = NULL;
  93
  94         } while (need_seqretry(&net->fs_addr_lock, seq));
  95
  96         done_seqretry(&net->fs_addr_lock, seq);
  97
  98         rcu_read_unlock();
  99         return server;
 100 }
 101
 102 /*
 103  * Look up a server by its UUID
 104  */
 105 struct afs_server *afs_find_server_by_uuid(struct afs_net *net, const uuid_t *uuid)
 106 {
 107         struct afs_server *server = NULL;
 108         struct rb_node *p;
 109         int diff, seq = 0;
 110
 111         _enter("%pU", uuid);
 112
 113         do {
 114                 /* Unfortunately, rbtree walking doesn't give reliable results
 115                  * under just the RCU read lock, so we have to check for
 116                  * changes.
 117                  */
 118                 if (server)
 119                         afs_put_server(net, server);
 120                 server = NULL;
 121
 122                 read_seqbegin_or_lock(&net->fs_lock, &seq);
 123
 124                 p = net->fs_servers.rb_node;
 125                 while (p) {
 126                         server = rb_entry(p, struct afs_server, uuid_rb);
 127
 128                         diff = memcmp(uuid, &server->uuid, sizeof(*uuid));
 129                         if (diff < 0) {
 130                                 p = p->rb_left;
 131                         } else if (diff > 0) {
 132                                 p = p->rb_right;
 133                         } else {
 134                                 afs_get_server(server);
 135                                 break;
 136                         }
 137
 138                         server = NULL;
 139                 }
 140         } while (need_seqretry(&net->fs_lock, seq));
 141
 142         done_seqretry(&net->fs_lock, seq);
 143
 144         _leave(" = %p", server);
 145         return server;
 146 }
 147
 148 /*
 149  * Install a server record in the namespace tree
 150  */
 151 static struct afs_server *afs_install_server(struct afs_net *net,
 152                                              struct afs_server *candidate)
 153 {
 154         const struct afs_addr_list *alist;
 155         struct afs_server *server;
 156         struct rb_node **pp, *p;
 157         int ret = -EEXIST, diff;
 158
 159         _enter("%p", candidate);
 160
 161         write_seqlock(&net->fs_lock);
 162
 163         /* Firstly install the server in the UUID lookup tree */
 164         pp = &net->fs_servers.rb_node;
 165         p = NULL;
 166         while (*pp) {
 167                 p = *pp;
 168                 _debug("- consider %p", p);
 169                 server = rb_entry(p, struct afs_server, uuid_rb);
 170                 diff = memcmp(&candidate->uuid, &server->uuid, sizeof(uuid_t));
 171                 if (diff < 0)
 172                         pp = &(*pp)->rb_left;
 173                 else if (diff > 0)
 174                         pp = &(*pp)->rb_right;
 175                 else
 176                         goto exists;
 177         }
 178
 179         server = candidate;
 180         rb_link_node(&server->uuid_rb, p, pp);
 181         rb_insert_color(&server->uuid_rb, &net->fs_servers);
 182         hlist_add_head_rcu(&server->proc_link, &net->fs_proc);
 183
 184         write_seqlock(&net->fs_addr_lock);
 185         alist = rcu_dereference_protected(server->addresses,
 186                                           lockdep_is_held(&net->fs_addr_lock.lock));
 187
 188         /* Secondly, if the server has any IPv4 and/or IPv6 addresses, install
 189          * it in the IPv4 and/or IPv6 reverse-map lists.
 190          *
 191          * TODO: For speed we want to use something other than a flat list
 192          * here; even sorting the list in terms of lowest address would help a
 193          * bit, but anything we might want to do gets messy and memory
 194          * intensive.
 195          */
 196         if (alist->nr_ipv4 > 0)
 197                 hlist_add_head_rcu(&server->addr4_link, &net->fs_addresses4);
 198         if (alist->nr_addrs > alist->nr_ipv4)
 199                 hlist_add_head_rcu(&server->addr6_link, &net->fs_addresses6);
 200
 201         write_sequnlock(&net->fs_addr_lock);
 202         ret = 0;
 203
 204 exists:
 205         afs_get_server(server);
 206         write_sequnlock(&net->fs_lock);
 207         return server;
 208 }
 209
 210 /*
 211  * allocate a new server record
 212  */
 213 static struct afs_server *afs_alloc_server(struct afs_net *net,
 214                                            const uuid_t *uuid,
 215                                            struct afs_addr_list *alist)
 216 {
 217         struct afs_server *server;
 218
 219         _enter("");
 220
 221         server = kzalloc(sizeof(struct afs_server), GFP_KERNEL);
 222         if (!server)
 223                 goto enomem;
 224
 225         atomic_set(&server->usage, 1);
 226         RCU_INIT_POINTER(server->addresses, alist);
 227         server->addr_version = alist->version;
 228         server->uuid = *uuid;
 229         server->flags = (1UL << AFS_SERVER_FL_NEW);
 230         server->update_at = ktime_get_real_seconds() + afs_server_update_delay;
 231         rwlock_init(&server->fs_lock);
 232         INIT_HLIST_HEAD(&server->cb_volumes);
 233         rwlock_init(&server->cb_break_lock);
 234         init_waitqueue_head(&server->probe_wq);
 235         spin_lock_init(&server->probe_lock);
 236
 237         afs_inc_servers_outstanding(net);
 238         _leave(" = %p", server);
 239         return server;
 240
 241 enomem:
 242         _leave(" = NULL [nomem]");
 243         return NULL;
 244 }
 245
 246 /*
 247  * Look up an address record for a server
 248  */
 249 static struct afs_addr_list *afs_vl_lookup_addrs(struct afs_cell *cell,
 250                                                  struct key *key, const uuid_t *uuid)
 251 {
 252         struct afs_vl_cursor vc;
 253         struct afs_addr_list *alist = NULL;
 254         int ret;
 255
 256         ret = -ERESTARTSYS;
 257         if (afs_begin_vlserver_operation(&vc, cell, key)) {
 258                 while (afs_select_vlserver(&vc)) {
 259                         if (test_bit(AFS_VLSERVER_FL_IS_YFS, &vc.server->flags))
 260                                 alist = afs_yfsvl_get_endpoints(&vc, uuid);
 261                         else
 262                                 alist = afs_vl_get_addrs_u(&vc, uuid);
 263                 }
 264
 265                 ret = afs_end_vlserver_operation(&vc);
 266         }
 267
 268         return ret < 0 ? ERR_PTR(ret) : alist;
 269 }
 270
 271 /*
 272  * Get or create a fileserver record.
 273  */
 274 struct afs_server *afs_lookup_server(struct afs_cell *cell, struct key *key,
 275                                      const uuid_t *uuid)
 276 {
 277         struct afs_addr_list *alist;
 278         struct afs_server *server, *candidate;
 279
 280         _enter("%p,%pU", cell->net, uuid);
 281
 282         server = afs_find_server_by_uuid(cell->net, uuid);
 283         if (server)
 284                 return server;
 285
 286         alist = afs_vl_lookup_addrs(cell, key, uuid);
 287         if (IS_ERR(alist))
 288                 return ERR_CAST(alist);
 289
 290         candidate = afs_alloc_server(cell->net, uuid, alist);
 291         if (!candidate) {
 292                 afs_put_addrlist(alist);
 293                 return ERR_PTR(-ENOMEM);
 294         }
 295
 296         server = afs_install_server(cell->net, candidate);
 297         if (server != candidate) {
 298                 afs_put_addrlist(alist);
 299                 kfree(candidate);
 300         }
 301
 302         _leave(" = %p{%d}", server, atomic_read(&server->usage));
 303         return server;
 304 }
 305
 306 /*
 307  * Set the server timer to fire after a given delay, assuming it's not already
 308  * set for an earlier time.
 309  */
 310 static void afs_set_server_timer(struct afs_net *net, time64_t delay)
 311 {
 312         if (net->live) {
 313                 afs_inc_servers_outstanding(net);
 314                 if (timer_reduce(&net->fs_timer, jiffies + delay * HZ))
 315                         afs_dec_servers_outstanding(net);
 316         }
 317 }
 318
 319 /*
 320  * Server management timer.  We have an increment on fs_outstanding that we
 321  * need to pass along to the work item.
 322  */
 323 void afs_servers_timer(struct timer_list *timer)
 324 {
 325         struct afs_net *net = container_of(timer, struct afs_net, fs_timer);
 326
 327         _enter("");
 328         if (!queue_work(afs_wq, &net->fs_manager))
 329                 afs_dec_servers_outstanding(net);
 330 }
 331
 332 /*
 333  * Release a reference on a server record.
 334  */
 335 void afs_put_server(struct afs_net *net, struct afs_server *server)
 336 {
 337         unsigned int usage;
 338
 339         if (!server)
 340                 return;
 341
 342         server->put_time = ktime_get_real_seconds();
 343
 344         usage = atomic_dec_return(&server->usage);
 345
 346         _enter("{%u}", usage);
 347
 348         if (likely(usage > 0))
 349                 return;
 350
 351         afs_set_server_timer(net, afs_server_gc_delay);
 352 }
 353
 354 static void afs_server_rcu(struct rcu_head *rcu)
 355 {
 356         struct afs_server *server = container_of(rcu, struct afs_server, rcu);
 357
 358         afs_put_addrlist(rcu_access_pointer(server->addresses));
 359         kfree(server);
 360 }
 361
 362 /*
 363  * destroy a dead server
 364  */
 365 static void afs_destroy_server(struct afs_net *net, struct afs_server *server)
 366 {
 367         struct afs_addr_list *alist = rcu_access_pointer(server->addresses);
 368         struct afs_addr_cursor ac = {
 369                 .alist  = alist,
 370                 .index  = alist->preferred,
 371                 .error  = 0,
 372         };
 373         _enter("%p", server);
 374
 375         if (test_bit(AFS_SERVER_FL_MAY_HAVE_CB, &server->flags))
 376                 afs_fs_give_up_all_callbacks(net, server, &ac, NULL);
 377
 378         wait_var_event(&server->probe_outstanding,
 379                        atomic_read(&server->probe_outstanding) == 0);
 380
 381         call_rcu(&server->rcu, afs_server_rcu);
 382         afs_dec_servers_outstanding(net);
 383 }
 384
 385 /*
 386  * Garbage collect any expired servers.
 387  */
 388 static void afs_gc_servers(struct afs_net *net, struct afs_server *gc_list)
 389 {
 390         struct afs_server *server;
 391         bool deleted;
 392         int usage;
 393
 394         while ((server = gc_list)) {
 395                 gc_list = server->gc_next;
 396
 397                 write_seqlock(&net->fs_lock);
 398                 usage = 1;
 399                 deleted = atomic_try_cmpxchg(&server->usage, &usage, 0);
 400                 if (deleted) {
 401                         rb_erase(&server->uuid_rb, &net->fs_servers);
 402                         hlist_del_rcu(&server->proc_link);
 403                 }
 404                 write_sequnlock(&net->fs_lock);
 405
 406                 if (deleted) {
 407                         write_seqlock(&net->fs_addr_lock);
 408                         if (!hlist_unhashed(&server->addr4_link))
 409                                 hlist_del_rcu(&server->addr4_link);
 410                         if (!hlist_unhashed(&server->addr6_link))
 411                                 hlist_del_rcu(&server->addr6_link);
 412                         write_sequnlock(&net->fs_addr_lock);
 413                         afs_destroy_server(net, server);
 414                 }
 415         }
 416 }
 417
 418 /*
 419  * Manage the records of servers known to be within a network namespace.  This
 420  * includes garbage collecting unused servers.
 421  *
 422  * Note also that we were given an increment on net->servers_outstanding by
 423  * whoever queued us that we need to deal with before returning.
 424  */
 425 void afs_manage_servers(struct work_struct *work)
 426 {
 427         struct afs_net *net = container_of(work, struct afs_net, fs_manager);
 428         struct afs_server *gc_list = NULL;
 429         struct rb_node *cursor;
 430         time64_t now = ktime_get_real_seconds(), next_manage = TIME64_MAX;
 431         bool purging = !net->live;
 432
 433         _enter("");
 434
 435         /* Trawl the server list looking for servers that have expired from
 436          * lack of use.
 437          */
 438         read_seqlock_excl(&net->fs_lock);
 439
 440         for (cursor = rb_first(&net->fs_servers); cursor; cursor = rb_next(cursor)) {
 441                 struct afs_server *server =
 442                         rb_entry(cursor, struct afs_server, uuid_rb);
 443                 int usage = atomic_read(&server->usage);
 444
 445                 _debug("manage %pU %u", &server->uuid, usage);
 446
 447                 ASSERTCMP(usage, >=, 1);
 448                 ASSERTIFCMP(purging, usage, ==, 1);
 449
 450                 if (usage == 1) {
 451                         time64_t expire_at = server->put_time;
 452
 453                         if (!test_bit(AFS_SERVER_FL_VL_FAIL, &server->flags) &&
 454                             !test_bit(AFS_SERVER_FL_NOT_FOUND, &server->flags))
 455                                 expire_at += afs_server_gc_delay;
 456                         if (purging || expire_at <= now) {
 457                                 server->gc_next = gc_list;
 458                                 gc_list = server;
 459                         } else if (expire_at < next_manage) {
 460                                 next_manage = expire_at;
 461                         }
 462                 }
 463         }
 464
 465         read_sequnlock_excl(&net->fs_lock);
 466
 467         /* Update the timer on the way out.  We have to pass an increment on
 468          * servers_outstanding in the namespace that we are in to the timer or
 469          * the work scheduler.
 470          */
 471         if (!purging && next_manage < TIME64_MAX) {
 472                 now = ktime_get_real_seconds();
 473
 474                 if (next_manage - now <= 0) {
 475                         if (queue_work(afs_wq, &net->fs_manager))
 476                                 afs_inc_servers_outstanding(net);
 477                 } else {
 478                         afs_set_server_timer(net, next_manage - now);
 479                 }
 480         }
 481
 482         afs_gc_servers(net, gc_list);
 483
 484         afs_dec_servers_outstanding(net);
 485         _leave(" [%d]", atomic_read(&net->servers_outstanding));
 486 }
 487
 488 static void afs_queue_server_manager(struct afs_net *net)
 489 {
 490         afs_inc_servers_outstanding(net);
 491         if (!queue_work(afs_wq, &net->fs_manager))
 492                 afs_dec_servers_outstanding(net);
 493 }
 494
 495 /*
 496  * Purge list of servers.
 497  */
 498 void afs_purge_servers(struct afs_net *net)
 499 {
 500         _enter("");
 501
 502         if (del_timer_sync(&net->fs_timer))
 503                 atomic_dec(&net->servers_outstanding);
 504
 505         afs_queue_server_manager(net);
 506
 507         _debug("wait");
 508         wait_var_event(&net->servers_outstanding,
 509                        !atomic_read(&net->servers_outstanding));
 510         _leave("");
 511 }
 512
 513 /*
 514  * Get an update for a server's address list.
 515  */
 516 static noinline bool afs_update_server_record(struct afs_fs_cursor *fc, struct afs_server *server)
 517 {
 518         struct afs_addr_list *alist, *discard;
 519
 520         _enter("");
 521
 522         alist = afs_vl_lookup_addrs(fc->vnode->volume->cell, fc->key,
 523                                     &server->uuid);
 524         if (IS_ERR(alist)) {
 525                 fc->ac.error = PTR_ERR(alist);
 526                 _leave(" = f [%d]", fc->ac.error);
 527                 return false;
 528         }
 529
 530         discard = alist;
 531         if (server->addr_version != alist->version) {
 532                 write_lock(&server->fs_lock);
 533                 discard = rcu_dereference_protected(server->addresses,
 534                                                     lockdep_is_held(&server->fs_lock));
 535                 rcu_assign_pointer(server->addresses, alist);
 536                 server->addr_version = alist->version;
 537                 write_unlock(&server->fs_lock);
 538         }
 539
 540         server->update_at = ktime_get_real_seconds() + afs_server_update_delay;
 541         afs_put_addrlist(discard);
 542         _leave(" = t");
 543         return true;
 544 }
 545
 546 /*
 547  * See if a server's address list needs updating.
 548  */
 549 bool afs_check_server_record(struct afs_fs_cursor *fc, struct afs_server *server)
 550 {
 551         time64_t now = ktime_get_real_seconds();
 552         long diff;
 553         bool success;
 554         int ret, retries = 0;
 555
 556         _enter("");
 557
 558         ASSERT(server);
 559
 560 retry:
 561         diff = READ_ONCE(server->update_at) - now;
 562         if (diff > 0) {
 563                 _leave(" = t [not now %ld]", diff);
 564                 return true;
 565         }
 566
 567         if (!test_and_set_bit_lock(AFS_SERVER_FL_UPDATING, &server->flags)) {
 568                 success = afs_update_server_record(fc, server);
 569                 clear_bit_unlock(AFS_SERVER_FL_UPDATING, &server->flags);
 570                 wake_up_bit(&server->flags, AFS_SERVER_FL_UPDATING);
 571                 _leave(" = %d", success);
 572                 return success;
 573         }
 574
 575         ret = wait_on_bit(&server->flags, AFS_SERVER_FL_UPDATING,
 576                           TASK_INTERRUPTIBLE);
 577         if (ret == -ERESTARTSYS) {
 578                 fc->ac.error = ret;
 579                 _leave(" = f [intr]");
 580                 return false;
 581         }
 582
 583         retries++;
 584         if (retries == 4) {
 585                 _leave(" = f [stale]");
 586                 ret = -ESTALE;
 587                 return false;
 588         }
 589         goto retry;
 590 }