fs/afs/fs_probe.c

   1 // SPDX-License-Identifier: GPL-2.0-or-later
   2 /* AFS fileserver probing
   3  *
   4  * Copyright (C) 2018, 2020 Red Hat, Inc. All Rights Reserved.
   5  * Written by David Howells ([email protected])
   6  */
   7
   8 #include <linux/sched.h>
   9 #include <linux/slab.h>
  10 #include "afs_fs.h"
  11 #include "internal.h"
  12 #include "protocol_yfs.h"
  13
  14 static unsigned int afs_fs_probe_fast_poll_interval = 30 * HZ;
  15 static unsigned int afs_fs_probe_slow_poll_interval = 5 * 60 * HZ;
  16
  17 /*
  18  * Start the probe polling timer.  We have to supply it with an inc on the
  19  * outstanding server count.
  20  */
  21 static void afs_schedule_fs_probe(struct afs_net *net,
  22                                   struct afs_server *server, bool fast)
  23 {
  24         unsigned long atj;
  25
  26         if (!net->live)
  27                 return;
  28
  29         atj = server->probed_at;
  30         atj += fast ? afs_fs_probe_fast_poll_interval : afs_fs_probe_slow_poll_interval;
  31
  32         afs_inc_servers_outstanding(net);
  33         if (timer_reduce(&net->fs_probe_timer, atj))
  34                 afs_dec_servers_outstanding(net);
  35 }
  36
  37 /*
  38  * Handle the completion of a set of probes.
  39  */
  40 static void afs_finished_fs_probe(struct afs_net *net, struct afs_server *server)
  41 {
  42         bool responded = server->probe.responded;
  43
  44         write_seqlock(&net->fs_lock);
  45         if (responded) {
  46                 list_add_tail(&server->probe_link, &net->fs_probe_slow);
  47         } else {
  48                 server->rtt = UINT_MAX;
  49                 clear_bit(AFS_SERVER_FL_RESPONDING, &server->flags);
  50                 list_add_tail(&server->probe_link, &net->fs_probe_fast);
  51         }
  52         write_sequnlock(&net->fs_lock);
  53
  54         afs_schedule_fs_probe(net, server, !responded);
  55 }
  56
  57 /*
  58  * Handle the completion of a probe.
  59  */
  60 static void afs_done_one_fs_probe(struct afs_net *net, struct afs_server *server)
  61 {
  62         _enter("");
  63
  64         if (atomic_dec_and_test(&server->probe_outstanding))
  65                 afs_finished_fs_probe(net, server);
  66
  67         wake_up_all(&server->probe_wq);
  68 }
  69
  70 /*
  71  * Handle inability to send a probe due to ENOMEM when trying to allocate a
  72  * call struct.
  73  */
  74 static void afs_fs_probe_not_done(struct afs_net *net,
  75                                   struct afs_server *server,
  76                                   struct afs_addr_cursor *ac)
  77 {
  78         struct afs_addr_list *alist = ac->alist;
  79         unsigned int index = ac->index;
  80
  81         _enter("");
  82
  83         trace_afs_io_error(0, -ENOMEM, afs_io_error_fs_probe_fail);
  84         spin_lock(&server->probe_lock);
  85
  86         server->probe.local_failure = true;
  87         if (server->probe.error == 0)
  88                 server->probe.error = -ENOMEM;
  89
  90         set_bit(index, &alist->failed);
  91
  92         spin_unlock(&server->probe_lock);
  93         return afs_done_one_fs_probe(net, server);
  94 }
  95
  96 /*
  97  * Process the result of probing a fileserver.  This is called after successful
  98  * or failed delivery of an FS.GetCapabilities operation.
  99  */
 100 void afs_fileserver_probe_result(struct afs_call *call)
 101 {
 102         struct afs_addr_list *alist = call->alist;
 103         struct afs_server *server = call->server;
 104         unsigned int index = call->addr_ix;
 105         unsigned int rtt_us = 0;
 106         int ret = call->error;
 107
 108         _enter("%pU,%u", &server->uuid, index);
 109
 110         spin_lock(&server->probe_lock);
 111
 112         switch (ret) {
 113         case 0:
 114                 server->probe.error = 0;
 115                 goto responded;
 116         case -ECONNABORTED:
 117                 if (!server->probe.responded) {
 118                         server->probe.abort_code = call->abort_code;
 119                         server->probe.error = ret;
 120                 }
 121                 goto responded;
 122         case -ENOMEM:
 123         case -ENONET:
 124                 clear_bit(index, &alist->responded);
 125                 server->probe.local_failure = true;
 126                 trace_afs_io_error(call->debug_id, ret, afs_io_error_fs_probe_fail);
 127                 goto out;
 128         case -ECONNRESET: /* Responded, but call expired. */
 129         case -ERFKILL:
 130         case -EADDRNOTAVAIL:
 131         case -ENETUNREACH:
 132         case -EHOSTUNREACH:
 133         case -EHOSTDOWN:
 134         case -ECONNREFUSED:
 135         case -ETIMEDOUT:
 136         case -ETIME:
 137         default:
 138                 clear_bit(index, &alist->responded);
 139                 set_bit(index, &alist->failed);
 140                 if (!server->probe.responded &&
 141                     (server->probe.error == 0 ||
 142                      server->probe.error == -ETIMEDOUT ||
 143                      server->probe.error == -ETIME))
 144                         server->probe.error = ret;
 145                 trace_afs_io_error(call->debug_id, ret, afs_io_error_fs_probe_fail);
 146                 goto out;
 147         }
 148
 149 responded:
 150         clear_bit(index, &alist->failed);
 151
 152         if (call->service_id == YFS_FS_SERVICE) {
 153                 server->probe.is_yfs = true;
 154                 set_bit(AFS_SERVER_FL_IS_YFS, &server->flags);
 155                 alist->addrs[index].srx_service = call->service_id;
 156         } else {
 157                 server->probe.not_yfs = true;
 158                 if (!server->probe.is_yfs) {
 159                         clear_bit(AFS_SERVER_FL_IS_YFS, &server->flags);
 160                         alist->addrs[index].srx_service = call->service_id;
 161                 }
 162         }
 163
 164         if (rxrpc_kernel_get_srtt(call->net->socket, call->rxcall, &rtt_us) &&
 165             rtt_us < server->probe.rtt) {
 166                 server->probe.rtt = rtt_us;
 167                 server->rtt = rtt_us;
 168                 alist->preferred = index;
 169         }
 170
 171         smp_wmb(); /* Set rtt before responded. */
 172         server->probe.responded = true;
 173         set_bit(index, &alist->responded);
 174         set_bit(AFS_SERVER_FL_RESPONDING, &server->flags);
 175 out:
 176         spin_unlock(&server->probe_lock);
 177
 178         _debug("probe %pU [%u] %pISpc rtt=%u ret=%d",
 179                &server->uuid, index, &alist->addrs[index].transport,
 180                rtt_us, ret);
 181
 182         return afs_done_one_fs_probe(call->net, server);
 183 }
 184
 185 /*
 186  * Probe one or all of a fileserver's addresses to find out the best route and
 187  * to query its capabilities.
 188  */
 189 void afs_fs_probe_fileserver(struct afs_net *net, struct afs_server *server,
 190                              struct key *key, bool all)
 191 {
 192         struct afs_addr_cursor ac = {
 193                 .index = 0,
 194         };
 195
 196         _enter("%pU", &server->uuid);
 197
 198         read_lock(&server->fs_lock);
 199         ac.alist = rcu_dereference_protected(server->addresses,
 200                                              lockdep_is_held(&server->fs_lock));
 201         afs_get_addrlist(ac.alist);
 202         read_unlock(&server->fs_lock);
 203
 204         server->probed_at = jiffies;
 205         atomic_set(&server->probe_outstanding, all ? ac.alist->nr_addrs : 1);
 206         memset(&server->probe, 0, sizeof(server->probe));
 207         server->probe.rtt = UINT_MAX;
 208
 209         ac.index = ac.alist->preferred;
 210         if (ac.index < 0 || ac.index >= ac.alist->nr_addrs)
 211                 all = true;
 212
 213         if (all) {
 214                 for (ac.index = 0; ac.index < ac.alist->nr_addrs; ac.index++)
 215                         if (!afs_fs_get_capabilities(net, server, &ac, key))
 216                                 afs_fs_probe_not_done(net, server, &ac);
 217         } else {
 218                 if (!afs_fs_get_capabilities(net, server, &ac, key))
 219                         afs_fs_probe_not_done(net, server, &ac);
 220         }
 221
 222         afs_put_addrlist(ac.alist);
 223 }
 224
 225 /*
 226  * Wait for the first as-yet untried fileserver to respond.
 227  */
 228 int afs_wait_for_fs_probes(struct afs_server_list *slist, unsigned long untried)
 229 {
 230         struct wait_queue_entry *waits;
 231         struct afs_server *server;
 232         unsigned int rtt = UINT_MAX, rtt_s;
 233         bool have_responders = false;
 234         int pref = -1, i;
 235
 236         _enter("%u,%lx", slist->nr_servers, untried);
 237
 238         /* Only wait for servers that have a probe outstanding. */
 239         for (i = 0; i < slist->nr_servers; i++) {
 240                 if (test_bit(i, &untried)) {
 241                         server = slist->servers[i].server;
 242                         if (!atomic_read(&server->probe_outstanding))
 243                                 __clear_bit(i, &untried);
 244                         if (server->probe.responded)
 245                                 have_responders = true;
 246                 }
 247         }
 248         if (have_responders || !untried)
 249                 return 0;
 250
 251         waits = kmalloc(array_size(slist->nr_servers, sizeof(*waits)), GFP_KERNEL);
 252         if (!waits)
 253                 return -ENOMEM;
 254
 255         for (i = 0; i < slist->nr_servers; i++) {
 256                 if (test_bit(i, &untried)) {
 257                         server = slist->servers[i].server;
 258                         init_waitqueue_entry(&waits[i], current);
 259                         add_wait_queue(&server->probe_wq, &waits[i]);
 260                 }
 261         }
 262
 263         for (;;) {
 264                 bool still_probing = false;
 265
 266                 set_current_state(TASK_INTERRUPTIBLE);
 267                 for (i = 0; i < slist->nr_servers; i++) {
 268                         if (test_bit(i, &untried)) {
 269                                 server = slist->servers[i].server;
 270                                 if (server->probe.responded)
 271                                         goto stop;
 272                                 if (atomic_read(&server->probe_outstanding))
 273                                         still_probing = true;
 274                         }
 275                 }
 276
 277                 if (!still_probing || signal_pending(current))
 278                         goto stop;
 279                 schedule();
 280         }
 281
 282 stop:
 283         set_current_state(TASK_RUNNING);
 284
 285         for (i = 0; i < slist->nr_servers; i++) {
 286                 if (test_bit(i, &untried)) {
 287                         server = slist->servers[i].server;
 288                         rtt_s = READ_ONCE(server->rtt);
 289                         if (test_bit(AFS_SERVER_FL_RESPONDING, &server->flags) &&
 290                             rtt_s < rtt) {
 291                                 pref = i;
 292                                 rtt = rtt_s;
 293                         }
 294
 295                         remove_wait_queue(&server->probe_wq, &waits[i]);
 296                 }
 297         }
 298
 299         kfree(waits);
 300
 301         if (pref == -1 && signal_pending(current))
 302                 return -ERESTARTSYS;
 303
 304         if (pref >= 0)
 305                 slist->preferred = pref;
 306         return 0;
 307 }
 308
 309 /*
 310  * Probe timer.  We have an increment on fs_outstanding that we need to pass
 311  * along to the work item.
 312  */
 313 void afs_fs_probe_timer(struct timer_list *timer)
 314 {
 315         struct afs_net *net = container_of(timer, struct afs_net, fs_probe_timer);
 316
 317         if (!net->live || !queue_work(afs_wq, &net->fs_prober))
 318                 afs_dec_servers_outstanding(net);
 319 }
 320
 321 /*
 322  * Dispatch a probe to a server.
 323  */
 324 static void afs_dispatch_fs_probe(struct afs_net *net, struct afs_server *server, bool all)
 325         __releases(&net->fs_lock)
 326 {
 327         struct key *key = NULL;
 328
 329         /* We remove it from the queues here - it will be added back to
 330          * one of the queues on the completion of the probe.
 331          */
 332         list_del_init(&server->probe_link);
 333
 334         afs_get_server(server, afs_server_trace_get_probe);
 335         write_sequnlock(&net->fs_lock);
 336
 337         afs_fs_probe_fileserver(net, server, key, all);
 338         afs_put_server(net, server, afs_server_trace_put_probe);
 339 }
 340
 341 /*
 342  * Probe a server immediately without waiting for its due time to come
 343  * round.  This is used when all of the addresses have been tried.
 344  */
 345 void afs_probe_fileserver(struct afs_net *net, struct afs_server *server)
 346 {
 347         write_seqlock(&net->fs_lock);
 348         if (!list_empty(&server->probe_link))
 349                 return afs_dispatch_fs_probe(net, server, true);
 350         write_sequnlock(&net->fs_lock);
 351 }
 352
 353 /*
 354  * Probe dispatcher to regularly dispatch probes to keep NAT alive.
 355  */
 356 void afs_fs_probe_dispatcher(struct work_struct *work)
 357 {
 358         struct afs_net *net = container_of(work, struct afs_net, fs_prober);
 359         struct afs_server *fast, *slow, *server;
 360         unsigned long nowj, timer_at, poll_at;
 361         bool first_pass = true, set_timer = false;
 362
 363         if (!net->live)
 364                 return;
 365
 366         _enter("");
 367
 368         if (list_empty(&net->fs_probe_fast) && list_empty(&net->fs_probe_slow)) {
 369                 _leave(" [none]");
 370                 return;
 371         }
 372
 373 again:
 374         write_seqlock(&net->fs_lock);
 375
 376         fast = slow = server = NULL;
 377         nowj = jiffies;
 378         timer_at = nowj + MAX_JIFFY_OFFSET;
 379
 380         if (!list_empty(&net->fs_probe_fast)) {
 381                 fast = list_first_entry(&net->fs_probe_fast, struct afs_server, probe_link);
 382                 poll_at = fast->probed_at + afs_fs_probe_fast_poll_interval;
 383                 if (time_before(nowj, poll_at)) {
 384                         timer_at = poll_at;
 385                         set_timer = true;
 386                         fast = NULL;
 387                 }
 388         }
 389
 390         if (!list_empty(&net->fs_probe_slow)) {
 391                 slow = list_first_entry(&net->fs_probe_slow, struct afs_server, probe_link);
 392                 poll_at = slow->probed_at + afs_fs_probe_slow_poll_interval;
 393                 if (time_before(nowj, poll_at)) {
 394                         if (time_before(poll_at, timer_at))
 395                             timer_at = poll_at;
 396                         set_timer = true;
 397                         slow = NULL;
 398                 }
 399         }
 400
 401         server = fast ?: slow;
 402         if (server)
 403                 _debug("probe %pU", &server->uuid);
 404
 405         if (server && (first_pass || !need_resched())) {
 406                 afs_dispatch_fs_probe(net, server, server == fast);
 407                 first_pass = false;
 408                 goto again;
 409         }
 410
 411         write_sequnlock(&net->fs_lock);
 412
 413         if (server) {
 414                 if (!queue_work(afs_wq, &net->fs_prober))
 415                         afs_dec_servers_outstanding(net);
 416                 _leave(" [requeue]");
 417         } else if (set_timer) {
 418                 if (timer_reduce(&net->fs_probe_timer, timer_at))
 419                         afs_dec_servers_outstanding(net);
 420                 _leave(" [timer]");
 421         } else {
 422                 afs_dec_servers_outstanding(net);
 423                 _leave(" [quiesce]");
 424         }
 425 }
 426
 427 /*
 428  * Wait for a probe on a particular fileserver to complete for 2s.
 429  */
 430 int afs_wait_for_one_fs_probe(struct afs_server *server, bool is_intr)
 431 {
 432         struct wait_queue_entry wait;
 433         unsigned long timo = 2 * HZ;
 434
 435         if (atomic_read(&server->probe_outstanding) == 0)
 436                 goto dont_wait;
 437
 438         init_wait_entry(&wait, 0);
 439         for (;;) {
 440                 prepare_to_wait_event(&server->probe_wq, &wait,
 441                                       is_intr ? TASK_INTERRUPTIBLE : TASK_UNINTERRUPTIBLE);
 442                 if (timo == 0 ||
 443                     server->probe.responded ||
 444                     atomic_read(&server->probe_outstanding) == 0 ||
 445                     (is_intr && signal_pending(current)))
 446                         break;
 447                 timo = schedule_timeout(timo);
 448         }
 449
 450         finish_wait(&server->probe_wq, &wait);
 451
 452 dont_wait:
 453         if (server->probe.responded)
 454                 return 0;
 455         if (is_intr && signal_pending(current))
 456                 return -ERESTARTSYS;
 457         if (timo == 0)
 458                 return -ETIME;
 459         return -EDESTADDRREQ;
 460 }
 461
 462 /*
 463  * Clean up the probing when the namespace is killed off.
 464  */
 465 void afs_fs_probe_cleanup(struct afs_net *net)
 466 {
 467         if (del_timer_sync(&net->fs_probe_timer))
 468                 afs_dec_servers_outstanding(net);
 469 }