fs/fscache/object.c

   1 /* FS-Cache object state machine handler
   2  *
   3  * Copyright (C) 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  * See Documentation/filesystems/caching/object.txt for a description of the
  12  * object state machine and the in-kernel representations.
  13  */
  14
  15 #define FSCACHE_DEBUG_LEVEL COOKIE
  16 #include <linux/module.h>
  17 #include <linux/slab.h>
  18 #include <linux/prefetch.h>
  19 #include "internal.h"
  20
  21 static const struct fscache_state *fscache_abort_initialisation(struct fscache_object *, int);
  22 static const struct fscache_state *fscache_kill_dependents(struct fscache_object *, int);
  23 static const struct fscache_state *fscache_drop_object(struct fscache_object *, int);
  24 static const struct fscache_state *fscache_initialise_object(struct fscache_object *, int);
  25 static const struct fscache_state *fscache_invalidate_object(struct fscache_object *, int);
  26 static const struct fscache_state *fscache_jumpstart_dependents(struct fscache_object *, int);
  27 static const struct fscache_state *fscache_kill_object(struct fscache_object *, int);
  28 static const struct fscache_state *fscache_lookup_failure(struct fscache_object *, int);
  29 static const struct fscache_state *fscache_look_up_object(struct fscache_object *, int);
  30 static const struct fscache_state *fscache_object_available(struct fscache_object *, int);
  31 static const struct fscache_state *fscache_parent_ready(struct fscache_object *, int);
  32 static const struct fscache_state *fscache_update_object(struct fscache_object *, int);
  33 static const struct fscache_state *fscache_object_dead(struct fscache_object *, int);
  34
  35 #define __STATE_NAME(n) fscache_osm_##n
  36 #define STATE(n) (&__STATE_NAME(n))
  37
  38 /*
  39  * Define a work state.  Work states are execution states.  No event processing
  40  * is performed by them.  The function attached to a work state returns a
  41  * pointer indicating the next state to which the state machine should
  42  * transition.  Returning NO_TRANSIT repeats the current state, but goes back
  43  * to the scheduler first.
  44  */
  45 #define WORK_STATE(n, sn, f) \
  46         const struct fscache_state __STATE_NAME(n) = {                  \
  47                 .name = #n,                                             \
  48                 .short_name = sn,                                       \
  49                 .work = f                                               \
  50         }
  51
  52 /*
  53  * Returns from work states.
  54  */
  55 #define transit_to(state) ({ prefetch(&STATE(state)->work); STATE(state); })
  56
  57 #define NO_TRANSIT ((struct fscache_state *)NULL)
  58
  59 /*
  60  * Define a wait state.  Wait states are event processing states.  No execution
  61  * is performed by them.  Wait states are just tables of "if event X occurs,
  62  * clear it and transition to state Y".  The dispatcher returns to the
  63  * scheduler if none of the events in which the wait state has an interest are
  64  * currently pending.
  65  */
  66 #define WAIT_STATE(n, sn, ...) \
  67         const struct fscache_state __STATE_NAME(n) = {                  \
  68                 .name = #n,                                             \
  69                 .short_name = sn,                                       \
  70                 .work = NULL,                                           \
  71                 .transitions = { __VA_ARGS__, { 0, NULL } }             \
  72         }
  73
  74 #define TRANSIT_TO(state, emask) \
  75         { .events = (emask), .transit_to = STATE(state) }
  76
  77 /*
  78  * The object state machine.
  79  */
  80 static WORK_STATE(INIT_OBJECT,          "INIT", fscache_initialise_object);
  81 static WORK_STATE(PARENT_READY,         "PRDY", fscache_parent_ready);
  82 static WORK_STATE(ABORT_INIT,           "ABRT", fscache_abort_initialisation);
  83 static WORK_STATE(LOOK_UP_OBJECT,       "LOOK", fscache_look_up_object);
  84 static WORK_STATE(CREATE_OBJECT,        "CRTO", fscache_look_up_object);
  85 static WORK_STATE(OBJECT_AVAILABLE,     "AVBL", fscache_object_available);
  86 static WORK_STATE(JUMPSTART_DEPS,       "JUMP", fscache_jumpstart_dependents);
  87
  88 static WORK_STATE(INVALIDATE_OBJECT,    "INVL", fscache_invalidate_object);
  89 static WORK_STATE(UPDATE_OBJECT,        "UPDT", fscache_update_object);
  90
  91 static WORK_STATE(LOOKUP_FAILURE,       "LCFL", fscache_lookup_failure);
  92 static WORK_STATE(KILL_OBJECT,          "KILL", fscache_kill_object);
  93 static WORK_STATE(KILL_DEPENDENTS,      "KDEP", fscache_kill_dependents);
  94 static WORK_STATE(DROP_OBJECT,          "DROP", fscache_drop_object);
  95 static WORK_STATE(OBJECT_DEAD,          "DEAD", fscache_object_dead);
  96
  97 static WAIT_STATE(WAIT_FOR_INIT,        "?INI",
  98                   TRANSIT_TO(INIT_OBJECT,       1 << FSCACHE_OBJECT_EV_NEW_CHILD));
  99
 100 static WAIT_STATE(WAIT_FOR_PARENT,      "?PRN",
 101                   TRANSIT_TO(PARENT_READY,      1 << FSCACHE_OBJECT_EV_PARENT_READY));
 102
 103 static WAIT_STATE(WAIT_FOR_CMD,         "?CMD",
 104                   TRANSIT_TO(INVALIDATE_OBJECT, 1 << FSCACHE_OBJECT_EV_INVALIDATE),
 105                   TRANSIT_TO(UPDATE_OBJECT,     1 << FSCACHE_OBJECT_EV_UPDATE),
 106                   TRANSIT_TO(JUMPSTART_DEPS,    1 << FSCACHE_OBJECT_EV_NEW_CHILD));
 107
 108 static WAIT_STATE(WAIT_FOR_CLEARANCE,   "?CLR",
 109                   TRANSIT_TO(KILL_OBJECT,       1 << FSCACHE_OBJECT_EV_CLEARED));
 110
 111 /*
 112  * Out-of-band event transition tables.  These are for handling unexpected
 113  * events, such as an I/O error.  If an OOB event occurs, the state machine
 114  * clears and disables the event and forces a transition to the nominated work
 115  * state (acurrently executing work states will complete first).
 116  *
 117  * In such a situation, object->state remembers the state the machine should
 118  * have been in/gone to and returning NO_TRANSIT returns to that.
 119  */
 120 static const struct fscache_transition fscache_osm_init_oob[] = {
 121            TRANSIT_TO(ABORT_INIT,
 122                       (1 << FSCACHE_OBJECT_EV_ERROR) |
 123                       (1 << FSCACHE_OBJECT_EV_KILL)),
 124            { 0, NULL }
 125 };
 126
 127 static const struct fscache_transition fscache_osm_lookup_oob[] = {
 128            TRANSIT_TO(LOOKUP_FAILURE,
 129                       (1 << FSCACHE_OBJECT_EV_ERROR) |
 130                       (1 << FSCACHE_OBJECT_EV_KILL)),
 131            { 0, NULL }
 132 };
 133
 134 static const struct fscache_transition fscache_osm_run_oob[] = {
 135            TRANSIT_TO(KILL_OBJECT,
 136                       (1 << FSCACHE_OBJECT_EV_ERROR) |
 137                       (1 << FSCACHE_OBJECT_EV_KILL)),
 138            { 0, NULL }
 139 };
 140
 141 static int  fscache_get_object(struct fscache_object *,
 142                                enum fscache_obj_ref_trace);
 143 static void fscache_put_object(struct fscache_object *,
 144                                enum fscache_obj_ref_trace);
 145 static bool fscache_enqueue_dependents(struct fscache_object *, int);
 146 static void fscache_dequeue_object(struct fscache_object *);
 147 static void fscache_update_aux_data(struct fscache_object *);
 148
 149 /*
 150  * we need to notify the parent when an op completes that we had outstanding
 151  * upon it
 152  */
 153 static inline void fscache_done_parent_op(struct fscache_object *object)
 154 {
 155         struct fscache_object *parent = object->parent;
 156
 157         _enter("OBJ%x {OBJ%x,%x}",
 158                object->debug_id, parent->debug_id, parent->n_ops);
 159
 160         spin_lock_nested(&parent->lock, 1);
 161         parent->n_obj_ops--;
 162         parent->n_ops--;
 163         if (parent->n_ops == 0)
 164                 fscache_raise_event(parent, FSCACHE_OBJECT_EV_CLEARED);
 165         spin_unlock(&parent->lock);
 166 }
 167
 168 /*
 169  * Object state machine dispatcher.
 170  */
 171 static void fscache_object_sm_dispatcher(struct fscache_object *object)
 172 {
 173         const struct fscache_transition *t;
 174         const struct fscache_state *state, *new_state;
 175         unsigned long events, event_mask;
 176         bool oob;
 177         int event = -1;
 178
 179         ASSERT(object != NULL);
 180
 181         _enter("{OBJ%x,%s,%lx}",
 182                object->debug_id, object->state->name, object->events);
 183
 184         event_mask = object->event_mask;
 185 restart:
 186         object->event_mask = 0; /* Mask normal event handling */
 187         state = object->state;
 188 restart_masked:
 189         events = object->events;
 190
 191         /* Handle any out-of-band events (typically an error) */
 192         if (events & object->oob_event_mask) {
 193                 _debug("{OBJ%x} oob %lx",
 194                        object->debug_id, events & object->oob_event_mask);
 195                 oob = true;
 196                 for (t = object->oob_table; t->events; t++) {
 197                         if (events & t->events) {
 198                                 state = t->transit_to;
 199                                 ASSERT(state->work != NULL);
 200                                 event = fls(events & t->events) - 1;
 201                                 __clear_bit(event, &object->oob_event_mask);
 202                                 clear_bit(event, &object->events);
 203                                 goto execute_work_state;
 204                         }
 205                 }
 206         }
 207         oob = false;
 208
 209         /* Wait states are just transition tables */
 210         if (!state->work) {
 211                 if (events & event_mask) {
 212                         for (t = state->transitions; t->events; t++) {
 213                                 if (events & t->events) {
 214                                         new_state = t->transit_to;
 215                                         event = fls(events & t->events) - 1;
 216                                         trace_fscache_osm(object, state,
 217                                                           true, false, event);
 218                                         clear_bit(event, &object->events);
 219                                         _debug("{OBJ%x} ev %d: %s -> %s",
 220                                                object->debug_id, event,
 221                                                state->name, new_state->name);
 222                                         object->state = state = new_state;
 223                                         goto execute_work_state;
 224                                 }
 225                         }
 226
 227                         /* The event mask didn't include all the tabled bits */
 228                         BUG();
 229                 }
 230                 /* Randomly woke up */
 231                 goto unmask_events;
 232         }
 233
 234 execute_work_state:
 235         _debug("{OBJ%x} exec %s", object->debug_id, state->name);
 236
 237         trace_fscache_osm(object, state, false, oob, event);
 238         new_state = state->work(object, event);
 239         event = -1;
 240         if (new_state == NO_TRANSIT) {
 241                 _debug("{OBJ%x} %s notrans", object->debug_id, state->name);
 242                 if (unlikely(state == STATE(OBJECT_DEAD))) {
 243                         _leave(" [dead]");
 244                         return;
 245                 }
 246                 fscache_enqueue_object(object);
 247                 event_mask = object->oob_event_mask;
 248                 goto unmask_events;
 249         }
 250
 251         _debug("{OBJ%x} %s -> %s",
 252                object->debug_id, state->name, new_state->name);
 253         object->state = state = new_state;
 254
 255         if (state->work) {
 256                 if (unlikely(state == STATE(OBJECT_DEAD))) {
 257                         _leave(" [dead]");
 258                         return;
 259                 }
 260                 goto restart_masked;
 261         }
 262
 263         /* Transited to wait state */
 264         event_mask = object->oob_event_mask;
 265         for (t = state->transitions; t->events; t++)
 266                 event_mask |= t->events;
 267
 268 unmask_events:
 269         object->event_mask = event_mask;
 270         smp_mb();
 271         events = object->events;
 272         if (events & event_mask)
 273                 goto restart;
 274         _leave(" [msk %lx]", event_mask);
 275 }
 276
 277 /*
 278  * execute an object
 279  */
 280 static void fscache_object_work_func(struct work_struct *work)
 281 {
 282         struct fscache_object *object =
 283                 container_of(work, struct fscache_object, work);
 284         unsigned long start;
 285
 286         _enter("{OBJ%x}", object->debug_id);
 287
 288         start = jiffies;
 289         fscache_object_sm_dispatcher(object);
 290         fscache_hist(fscache_objs_histogram, start);
 291         fscache_put_object(object, fscache_obj_put_work);
 292 }
 293
 294 /**
 295  * fscache_object_init - Initialise a cache object description
 296  * @object: Object description
 297  * @cookie: Cookie object will be attached to
 298  * @cache: Cache in which backing object will be found
 299  *
 300  * Initialise a cache object description to its basic values.
 301  *
 302  * See Documentation/filesystems/caching/backend-api.txt for a complete
 303  * description.
 304  */
 305 void fscache_object_init(struct fscache_object *object,
 306                          struct fscache_cookie *cookie,
 307                          struct fscache_cache *cache)
 308 {
 309         const struct fscache_transition *t;
 310
 311         atomic_inc(&cache->object_count);
 312
 313         object->state = STATE(WAIT_FOR_INIT);
 314         object->oob_table = fscache_osm_init_oob;
 315         object->flags = 1 << FSCACHE_OBJECT_IS_LIVE;
 316         spin_lock_init(&object->lock);
 317         INIT_LIST_HEAD(&object->cache_link);
 318         INIT_HLIST_NODE(&object->cookie_link);
 319         INIT_WORK(&object->work, fscache_object_work_func);
 320         INIT_LIST_HEAD(&object->dependents);
 321         INIT_LIST_HEAD(&object->dep_link);
 322         INIT_LIST_HEAD(&object->pending_ops);
 323         object->n_children = 0;
 324         object->n_ops = object->n_in_progress = object->n_exclusive = 0;
 325         object->events = 0;
 326         object->store_limit = 0;
 327         object->store_limit_l = 0;
 328         object->cache = cache;
 329         object->cookie = cookie;
 330         fscache_cookie_get(cookie, fscache_cookie_get_attach_object);
 331         object->parent = NULL;
 332 #ifdef CONFIG_FSCACHE_OBJECT_LIST
 333         RB_CLEAR_NODE(&object->objlist_link);
 334 #endif
 335
 336         object->oob_event_mask = 0;
 337         for (t = object->oob_table; t->events; t++)
 338                 object->oob_event_mask |= t->events;
 339         object->event_mask = object->oob_event_mask;
 340         for (t = object->state->transitions; t->events; t++)
 341                 object->event_mask |= t->events;
 342 }
 343 EXPORT_SYMBOL(fscache_object_init);
 344
 345 /*
 346  * Mark the object as no longer being live, making sure that we synchronise
 347  * against op submission.
 348  */
 349 static inline void fscache_mark_object_dead(struct fscache_object *object)
 350 {
 351         spin_lock(&object->lock);
 352         clear_bit(FSCACHE_OBJECT_IS_LIVE, &object->flags);
 353         spin_unlock(&object->lock);
 354 }
 355
 356 /*
 357  * Abort object initialisation before we start it.
 358  */
 359 static const struct fscache_state *fscache_abort_initialisation(struct fscache_object *object,
 360                                                                 int event)
 361 {
 362         _enter("{OBJ%x},%d", object->debug_id, event);
 363
 364         object->oob_event_mask = 0;
 365         fscache_dequeue_object(object);
 366         return transit_to(KILL_OBJECT);
 367 }
 368
 369 /*
 370  * initialise an object
 371  * - check the specified object's parent to see if we can make use of it
 372  *   immediately to do a creation
 373  * - we may need to start the process of creating a parent and we need to wait
 374  *   for the parent's lookup and creation to complete if it's not there yet
 375  */
 376 static const struct fscache_state *fscache_initialise_object(struct fscache_object *object,
 377                                                              int event)
 378 {
 379         struct fscache_object *parent;
 380         bool success;
 381
 382         _enter("{OBJ%x},%d", object->debug_id, event);
 383
 384         ASSERT(list_empty(&object->dep_link));
 385
 386         parent = object->parent;
 387         if (!parent) {
 388                 _leave(" [no parent]");
 389                 return transit_to(DROP_OBJECT);
 390         }
 391
 392         _debug("parent: %s of:%lx", parent->state->name, parent->flags);
 393
 394         if (fscache_object_is_dying(parent)) {
 395                 _leave(" [bad parent]");
 396                 return transit_to(DROP_OBJECT);
 397         }
 398
 399         if (fscache_object_is_available(parent)) {
 400                 _leave(" [ready]");
 401                 return transit_to(PARENT_READY);
 402         }
 403
 404         _debug("wait");
 405
 406         spin_lock(&parent->lock);
 407         fscache_stat(&fscache_n_cop_grab_object);
 408         success = false;
 409         if (fscache_object_is_live(parent) &&
 410             object->cache->ops->grab_object(object, fscache_obj_get_add_to_deps)) {
 411                 list_add(&object->dep_link, &parent->dependents);
 412                 success = true;
 413         }
 414         fscache_stat_d(&fscache_n_cop_grab_object);
 415         spin_unlock(&parent->lock);
 416         if (!success) {
 417                 _leave(" [grab failed]");
 418                 return transit_to(DROP_OBJECT);
 419         }
 420
 421         /* fscache_acquire_non_index_cookie() uses this
 422          * to wake the chain up */
 423         fscache_raise_event(parent, FSCACHE_OBJECT_EV_NEW_CHILD);
 424         _leave(" [wait]");
 425         return transit_to(WAIT_FOR_PARENT);
 426 }
 427
 428 /*
 429  * Once the parent object is ready, we should kick off our lookup op.
 430  */
 431 static const struct fscache_state *fscache_parent_ready(struct fscache_object *object,
 432                                                         int event)
 433 {
 434         struct fscache_object *parent = object->parent;
 435
 436         _enter("{OBJ%x},%d", object->debug_id, event);
 437
 438         ASSERT(parent != NULL);
 439
 440         spin_lock(&parent->lock);
 441         parent->n_ops++;
 442         parent->n_obj_ops++;
 443         object->lookup_jif = jiffies;
 444         spin_unlock(&parent->lock);
 445
 446         _leave("");
 447         return transit_to(LOOK_UP_OBJECT);
 448 }
 449
 450 /*
 451  * look an object up in the cache from which it was allocated
 452  * - we hold an "access lock" on the parent object, so the parent object cannot
 453  *   be withdrawn by either party till we've finished
 454  */
 455 static const struct fscache_state *fscache_look_up_object(struct fscache_object *object,
 456                                                           int event)
 457 {
 458         struct fscache_cookie *cookie = object->cookie;
 459         struct fscache_object *parent = object->parent;
 460         int ret;
 461
 462         _enter("{OBJ%x},%d", object->debug_id, event);
 463
 464         object->oob_table = fscache_osm_lookup_oob;
 465
 466         ASSERT(parent != NULL);
 467         ASSERTCMP(parent->n_ops, >, 0);
 468         ASSERTCMP(parent->n_obj_ops, >, 0);
 469
 470         /* make sure the parent is still available */
 471         ASSERT(fscache_object_is_available(parent));
 472
 473         if (fscache_object_is_dying(parent) ||
 474             test_bit(FSCACHE_IOERROR, &object->cache->flags) ||
 475             !fscache_use_cookie(object)) {
 476                 _leave(" [unavailable]");
 477                 return transit_to(LOOKUP_FAILURE);
 478         }
 479
 480         _debug("LOOKUP \"%s\" in \"%s\"",
 481                cookie->def->name, object->cache->tag->name);
 482
 483         fscache_stat(&fscache_n_object_lookups);
 484         fscache_stat(&fscache_n_cop_lookup_object);
 485         ret = object->cache->ops->lookup_object(object);
 486         fscache_stat_d(&fscache_n_cop_lookup_object);
 487
 488         fscache_unuse_cookie(object);
 489
 490         if (ret == -ETIMEDOUT) {
 491                 /* probably stuck behind another object, so move this one to
 492                  * the back of the queue */
 493                 fscache_stat(&fscache_n_object_lookups_timed_out);
 494                 _leave(" [timeout]");
 495                 return NO_TRANSIT;
 496         }
 497
 498         if (ret < 0) {
 499                 _leave(" [error]");
 500                 return transit_to(LOOKUP_FAILURE);
 501         }
 502
 503         _leave(" [ok]");
 504         return transit_to(OBJECT_AVAILABLE);
 505 }
 506
 507 /**
 508  * fscache_object_lookup_negative - Note negative cookie lookup
 509  * @object: Object pointing to cookie to mark
 510  *
 511  * Note negative lookup, permitting those waiting to read data from an already
 512  * existing backing object to continue as there's no data for them to read.
 513  */
 514 void fscache_object_lookup_negative(struct fscache_object *object)
 515 {
 516         struct fscache_cookie *cookie = object->cookie;
 517
 518         _enter("{OBJ%x,%s}", object->debug_id, object->state->name);
 519
 520         if (!test_and_set_bit(FSCACHE_OBJECT_IS_LOOKED_UP, &object->flags)) {
 521                 fscache_stat(&fscache_n_object_lookups_negative);
 522
 523                 /* Allow write requests to begin stacking up and read requests to begin
 524                  * returning ENODATA.
 525                  */
 526                 set_bit(FSCACHE_COOKIE_NO_DATA_YET, &cookie->flags);
 527                 clear_bit(FSCACHE_COOKIE_UNAVAILABLE, &cookie->flags);
 528
 529                 _debug("wake up lookup %p", &cookie->flags);
 530                 clear_bit_unlock(FSCACHE_COOKIE_LOOKING_UP, &cookie->flags);
 531                 wake_up_bit(&cookie->flags, FSCACHE_COOKIE_LOOKING_UP);
 532         }
 533         _leave("");
 534 }
 535 EXPORT_SYMBOL(fscache_object_lookup_negative);
 536
 537 /**
 538  * fscache_obtained_object - Note successful object lookup or creation
 539  * @object: Object pointing to cookie to mark
 540  *
 541  * Note successful lookup and/or creation, permitting those waiting to write
 542  * data to a backing object to continue.
 543  *
 544  * Note that after calling this, an object's cookie may be relinquished by the
 545  * netfs, and so must be accessed with object lock held.
 546  */
 547 void fscache_obtained_object(struct fscache_object *object)
 548 {
 549         struct fscache_cookie *cookie = object->cookie;
 550
 551         _enter("{OBJ%x,%s}", object->debug_id, object->state->name);
 552
 553         /* if we were still looking up, then we must have a positive lookup
 554          * result, in which case there may be data available */
 555         if (!test_and_set_bit(FSCACHE_OBJECT_IS_LOOKED_UP, &object->flags)) {
 556                 fscache_stat(&fscache_n_object_lookups_positive);
 557
 558                 /* We do (presumably) have data */
 559                 clear_bit_unlock(FSCACHE_COOKIE_NO_DATA_YET, &cookie->flags);
 560                 clear_bit(FSCACHE_COOKIE_UNAVAILABLE, &cookie->flags);
 561
 562                 /* Allow write requests to begin stacking up and read requests
 563                  * to begin shovelling data.
 564                  */
 565                 clear_bit_unlock(FSCACHE_COOKIE_LOOKING_UP, &cookie->flags);
 566                 wake_up_bit(&cookie->flags, FSCACHE_COOKIE_LOOKING_UP);
 567         } else {
 568                 fscache_stat(&fscache_n_object_created);
 569         }
 570
 571         set_bit(FSCACHE_OBJECT_IS_AVAILABLE, &object->flags);
 572         _leave("");
 573 }
 574 EXPORT_SYMBOL(fscache_obtained_object);
 575
 576 /*
 577  * handle an object that has just become available
 578  */
 579 static const struct fscache_state *fscache_object_available(struct fscache_object *object,
 580                                                             int event)
 581 {
 582         _enter("{OBJ%x},%d", object->debug_id, event);
 583
 584         object->oob_table = fscache_osm_run_oob;
 585
 586         spin_lock(&object->lock);
 587
 588         fscache_done_parent_op(object);
 589         if (object->n_in_progress == 0) {
 590                 if (object->n_ops > 0) {
 591                         ASSERTCMP(object->n_ops, >=, object->n_obj_ops);
 592                         fscache_start_operations(object);
 593                 } else {
 594                         ASSERT(list_empty(&object->pending_ops));
 595                 }
 596         }
 597         spin_unlock(&object->lock);
 598
 599         fscache_stat(&fscache_n_cop_lookup_complete);
 600         object->cache->ops->lookup_complete(object);
 601         fscache_stat_d(&fscache_n_cop_lookup_complete);
 602
 603         fscache_hist(fscache_obj_instantiate_histogram, object->lookup_jif);
 604         fscache_stat(&fscache_n_object_avail);
 605
 606         _leave("");
 607         return transit_to(JUMPSTART_DEPS);
 608 }
 609
 610 /*
 611  * Wake up this object's dependent objects now that we've become available.
 612  */
 613 static const struct fscache_state *fscache_jumpstart_dependents(struct fscache_object *object,
 614                                                                 int event)
 615 {
 616         _enter("{OBJ%x},%d", object->debug_id, event);
 617
 618         if (!fscache_enqueue_dependents(object, FSCACHE_OBJECT_EV_PARENT_READY))
 619                 return NO_TRANSIT; /* Not finished; requeue */
 620         return transit_to(WAIT_FOR_CMD);
 621 }
 622
 623 /*
 624  * Handle lookup or creation failute.
 625  */
 626 static const struct fscache_state *fscache_lookup_failure(struct fscache_object *object,
 627                                                           int event)
 628 {
 629         struct fscache_cookie *cookie;
 630
 631         _enter("{OBJ%x},%d", object->debug_id, event);
 632
 633         object->oob_event_mask = 0;
 634
 635         fscache_stat(&fscache_n_cop_lookup_complete);
 636         object->cache->ops->lookup_complete(object);
 637         fscache_stat_d(&fscache_n_cop_lookup_complete);
 638
 639         set_bit(FSCACHE_OBJECT_KILLED_BY_CACHE, &object->flags);
 640
 641         cookie = object->cookie;
 642         set_bit(FSCACHE_COOKIE_UNAVAILABLE, &cookie->flags);
 643         if (test_and_clear_bit(FSCACHE_COOKIE_LOOKING_UP, &cookie->flags))
 644                 wake_up_bit(&cookie->flags, FSCACHE_COOKIE_LOOKING_UP);
 645
 646         fscache_done_parent_op(object);
 647         return transit_to(KILL_OBJECT);
 648 }
 649
 650 /*
 651  * Wait for completion of all active operations on this object and the death of
 652  * all child objects of this object.
 653  */
 654 static const struct fscache_state *fscache_kill_object(struct fscache_object *object,
 655                                                        int event)
 656 {
 657         _enter("{OBJ%x,%d,%d},%d",
 658                object->debug_id, object->n_ops, object->n_children, event);
 659
 660         fscache_mark_object_dead(object);
 661         object->oob_event_mask = 0;
 662
 663         if (test_bit(FSCACHE_OBJECT_RETIRED, &object->flags)) {
 664                 /* Reject any new read/write ops and abort any that are pending. */
 665                 clear_bit(FSCACHE_OBJECT_PENDING_WRITE, &object->flags);
 666                 fscache_cancel_all_ops(object);
 667         }
 668
 669         if (list_empty(&object->dependents) &&
 670             object->n_ops == 0 &&
 671             object->n_children == 0)
 672                 return transit_to(DROP_OBJECT);
 673
 674         if (object->n_in_progress == 0) {
 675                 spin_lock(&object->lock);
 676                 if (object->n_ops > 0 && object->n_in_progress == 0)
 677                         fscache_start_operations(object);
 678                 spin_unlock(&object->lock);
 679         }
 680
 681         if (!list_empty(&object->dependents))
 682                 return transit_to(KILL_DEPENDENTS);
 683
 684         return transit_to(WAIT_FOR_CLEARANCE);
 685 }
 686
 687 /*
 688  * Kill dependent objects.
 689  */
 690 static const struct fscache_state *fscache_kill_dependents(struct fscache_object *object,
 691                                                            int event)
 692 {
 693         _enter("{OBJ%x},%d", object->debug_id, event);
 694
 695         if (!fscache_enqueue_dependents(object, FSCACHE_OBJECT_EV_KILL))
 696                 return NO_TRANSIT; /* Not finished */
 697         return transit_to(WAIT_FOR_CLEARANCE);
 698 }
 699
 700 /*
 701  * Drop an object's attachments
 702  */
 703 static const struct fscache_state *fscache_drop_object(struct fscache_object *object,
 704                                                        int event)
 705 {
 706         struct fscache_object *parent = object->parent;
 707         struct fscache_cookie *cookie = object->cookie;
 708         struct fscache_cache *cache = object->cache;
 709         bool awaken = false;
 710
 711         _enter("{OBJ%x,%d},%d", object->debug_id, object->n_children, event);
 712
 713         ASSERT(cookie != NULL);
 714         ASSERT(!hlist_unhashed(&object->cookie_link));
 715
 716         if (test_bit(FSCACHE_COOKIE_AUX_UPDATED, &cookie->flags)) {
 717                 _debug("final update");
 718                 fscache_update_aux_data(object);
 719         }
 720
 721         /* Make sure the cookie no longer points here and that the netfs isn't
 722          * waiting for us.
 723          */
 724         spin_lock(&cookie->lock);
 725         hlist_del_init(&object->cookie_link);
 726         if (hlist_empty(&cookie->backing_objects) &&
 727             test_and_clear_bit(FSCACHE_COOKIE_INVALIDATING, &cookie->flags))
 728                 awaken = true;
 729         spin_unlock(&cookie->lock);
 730
 731         if (awaken)
 732                 wake_up_bit(&cookie->flags, FSCACHE_COOKIE_INVALIDATING);
 733         if (test_and_clear_bit(FSCACHE_COOKIE_LOOKING_UP, &cookie->flags))
 734                 wake_up_bit(&cookie->flags, FSCACHE_COOKIE_LOOKING_UP);
 735
 736
 737         /* Prevent a race with our last child, which has to signal EV_CLEARED
 738          * before dropping our spinlock.
 739          */
 740         spin_lock(&object->lock);
 741         spin_unlock(&object->lock);
 742
 743         /* Discard from the cache's collection of objects */
 744         spin_lock(&cache->object_list_lock);
 745         list_del_init(&object->cache_link);
 746         spin_unlock(&cache->object_list_lock);
 747
 748         fscache_stat(&fscache_n_cop_drop_object);
 749         cache->ops->drop_object(object);
 750         fscache_stat_d(&fscache_n_cop_drop_object);
 751
 752         /* The parent object wants to know when all it dependents have gone */
 753         if (parent) {
 754                 _debug("release parent OBJ%x {%d}",
 755                        parent->debug_id, parent->n_children);
 756
 757                 spin_lock(&parent->lock);
 758                 parent->n_children--;
 759                 if (parent->n_children == 0)
 760                         fscache_raise_event(parent, FSCACHE_OBJECT_EV_CLEARED);
 761                 spin_unlock(&parent->lock);
 762                 object->parent = NULL;
 763         }
 764
 765         /* this just shifts the object release to the work processor */
 766         fscache_put_object(object, fscache_obj_put_drop_obj);
 767         fscache_stat(&fscache_n_object_dead);
 768
 769         _leave("");
 770         return transit_to(OBJECT_DEAD);
 771 }
 772
 773 /*
 774  * get a ref on an object
 775  */
 776 static int fscache_get_object(struct fscache_object *object,
 777                               enum fscache_obj_ref_trace why)
 778 {
 779         int ret;
 780
 781         fscache_stat(&fscache_n_cop_grab_object);
 782         ret = object->cache->ops->grab_object(object, why) ? 0 : -EAGAIN;
 783         fscache_stat_d(&fscache_n_cop_grab_object);
 784         return ret;
 785 }
 786
 787 /*
 788  * Discard a ref on an object
 789  */
 790 static void fscache_put_object(struct fscache_object *object,
 791                                enum fscache_obj_ref_trace why)
 792 {
 793         fscache_stat(&fscache_n_cop_put_object);
 794         object->cache->ops->put_object(object, why);
 795         fscache_stat_d(&fscache_n_cop_put_object);
 796 }
 797
 798 /**
 799  * fscache_object_destroy - Note that a cache object is about to be destroyed
 800  * @object: The object to be destroyed
 801  *
 802  * Note the imminent destruction and deallocation of a cache object record.
 803  */
 804 void fscache_object_destroy(struct fscache_object *object)
 805 {
 806         fscache_objlist_remove(object);
 807
 808         /* We can get rid of the cookie now */
 809         fscache_cookie_put(object->cookie, fscache_cookie_put_object);
 810         object->cookie = NULL;
 811 }
 812 EXPORT_SYMBOL(fscache_object_destroy);
 813
 814 /*
 815  * enqueue an object for metadata-type processing
 816  */
 817 void fscache_enqueue_object(struct fscache_object *object)
 818 {
 819         _enter("{OBJ%x}", object->debug_id);
 820
 821         if (fscache_get_object(object, fscache_obj_get_queue) >= 0) {
 822                 wait_queue_head_t *cong_wq =
 823                         &get_cpu_var(fscache_object_cong_wait);
 824
 825                 if (queue_work(fscache_object_wq, &object->work)) {
 826                         if (fscache_object_congested())
 827                                 wake_up(cong_wq);
 828                 } else
 829                         fscache_put_object(object, fscache_obj_put_queue);
 830
 831                 put_cpu_var(fscache_object_cong_wait);
 832         }
 833 }
 834
 835 /**
 836  * fscache_object_sleep_till_congested - Sleep until object wq is congested
 837  * @timeoutp: Scheduler sleep timeout
 838  *
 839  * Allow an object handler to sleep until the object workqueue is congested.
 840  *
 841  * The caller must set up a wake up event before calling this and must have set
 842  * the appropriate sleep mode (such as TASK_UNINTERRUPTIBLE) and tested its own
 843  * condition before calling this function as no test is made here.
 844  *
 845  * %true is returned if the object wq is congested, %false otherwise.
 846  */
 847 bool fscache_object_sleep_till_congested(signed long *timeoutp)
 848 {
 849         wait_queue_head_t *cong_wq = this_cpu_ptr(&fscache_object_cong_wait);
 850         DEFINE_WAIT(wait);
 851
 852         if (fscache_object_congested())
 853                 return true;
 854
 855         add_wait_queue_exclusive(cong_wq, &wait);
 856         if (!fscache_object_congested())
 857                 *timeoutp = schedule_timeout(*timeoutp);
 858         finish_wait(cong_wq, &wait);
 859
 860         return fscache_object_congested();
 861 }
 862 EXPORT_SYMBOL_GPL(fscache_object_sleep_till_congested);
 863
 864 /*
 865  * Enqueue the dependents of an object for metadata-type processing.
 866  *
 867  * If we don't manage to finish the list before the scheduler wants to run
 868  * again then return false immediately.  We return true if the list was
 869  * cleared.
 870  */
 871 static bool fscache_enqueue_dependents(struct fscache_object *object, int event)
 872 {
 873         struct fscache_object *dep;
 874         bool ret = true;
 875
 876         _enter("{OBJ%x}", object->debug_id);
 877
 878         if (list_empty(&object->dependents))
 879                 return true;
 880
 881         spin_lock(&object->lock);
 882
 883         while (!list_empty(&object->dependents)) {
 884                 dep = list_entry(object->dependents.next,
 885                                  struct fscache_object, dep_link);
 886                 list_del_init(&dep->dep_link);
 887
 888                 fscache_raise_event(dep, event);
 889                 fscache_put_object(dep, fscache_obj_put_enq_dep);
 890
 891                 if (!list_empty(&object->dependents) && need_resched()) {
 892                         ret = false;
 893                         break;
 894                 }
 895         }
 896
 897         spin_unlock(&object->lock);
 898         return ret;
 899 }
 900
 901 /*
 902  * remove an object from whatever queue it's waiting on
 903  */
 904 static void fscache_dequeue_object(struct fscache_object *object)
 905 {
 906         _enter("{OBJ%x}", object->debug_id);
 907
 908         if (!list_empty(&object->dep_link)) {
 909                 spin_lock(&object->parent->lock);
 910                 list_del_init(&object->dep_link);
 911                 spin_unlock(&object->parent->lock);
 912         }
 913
 914         _leave("");
 915 }
 916
 917 /**
 918  * fscache_check_aux - Ask the netfs whether an object on disk is still valid
 919  * @object: The object to ask about
 920  * @data: The auxiliary data for the object
 921  * @datalen: The size of the auxiliary data
 922  *
 923  * This function consults the netfs about the coherency state of an object.
 924  * The caller must be holding a ref on cookie->n_active (held by
 925  * fscache_look_up_object() on behalf of the cache backend during object lookup
 926  * and creation).
 927  */
 928 enum fscache_checkaux fscache_check_aux(struct fscache_object *object,
 929                                         const void *data, uint16_t datalen,
 930                                         loff_t object_size)
 931 {
 932         enum fscache_checkaux result;
 933
 934         if (!object->cookie->def->check_aux) {
 935                 fscache_stat(&fscache_n_checkaux_none);
 936                 return FSCACHE_CHECKAUX_OKAY;
 937         }
 938
 939         result = object->cookie->def->check_aux(object->cookie->netfs_data,
 940                                                 data, datalen, object_size);
 941         switch (result) {
 942                 /* entry okay as is */
 943         case FSCACHE_CHECKAUX_OKAY:
 944                 fscache_stat(&fscache_n_checkaux_okay);
 945                 break;
 946
 947                 /* entry requires update */
 948         case FSCACHE_CHECKAUX_NEEDS_UPDATE:
 949                 fscache_stat(&fscache_n_checkaux_update);
 950                 break;
 951
 952                 /* entry requires deletion */
 953         case FSCACHE_CHECKAUX_OBSOLETE:
 954                 fscache_stat(&fscache_n_checkaux_obsolete);
 955                 break;
 956
 957         default:
 958                 BUG();
 959         }
 960
 961         return result;
 962 }
 963 EXPORT_SYMBOL(fscache_check_aux);
 964
 965 /*
 966  * Asynchronously invalidate an object.
 967  */
 968 static const struct fscache_state *_fscache_invalidate_object(struct fscache_object *object,
 969                                                               int event)
 970 {
 971         struct fscache_operation *op;
 972         struct fscache_cookie *cookie = object->cookie;
 973
 974         _enter("{OBJ%x},%d", object->debug_id, event);
 975
 976         /* We're going to need the cookie.  If the cookie is not available then
 977          * retire the object instead.
 978          */
 979         if (!fscache_use_cookie(object)) {
 980                 ASSERT(radix_tree_empty(&object->cookie->stores));
 981                 set_bit(FSCACHE_OBJECT_RETIRED, &object->flags);
 982                 _leave(" [no cookie]");
 983                 return transit_to(KILL_OBJECT);
 984         }
 985
 986         /* Reject any new read/write ops and abort any that are pending. */
 987         fscache_invalidate_writes(cookie);
 988         clear_bit(FSCACHE_OBJECT_PENDING_WRITE, &object->flags);
 989         fscache_cancel_all_ops(object);
 990
 991         /* Now we have to wait for in-progress reads and writes */
 992         op = kzalloc(sizeof(*op), GFP_KERNEL);
 993         if (!op)
 994                 goto nomem;
 995
 996         fscache_operation_init(cookie, op, object->cache->ops->invalidate_object,
 997                                NULL, NULL);
 998         op->flags = FSCACHE_OP_ASYNC |
 999                 (1 << FSCACHE_OP_EXCLUSIVE) |
1000                 (1 << FSCACHE_OP_UNUSE_COOKIE);
1001         trace_fscache_page_op(cookie, NULL, op, fscache_page_op_invalidate);
1002
1003         spin_lock(&cookie->lock);
1004         if (fscache_submit_exclusive_op(object, op) < 0)
1005                 goto submit_op_failed;
1006         spin_unlock(&cookie->lock);
1007         fscache_put_operation(op);
1008
1009         /* Once we've completed the invalidation, we know there will be no data
1010          * stored in the cache and thus we can reinstate the data-check-skip
1011          * optimisation.
1012          */
1013         set_bit(FSCACHE_COOKIE_NO_DATA_YET, &cookie->flags);
1014
1015         /* We can allow read and write requests to come in once again.  They'll
1016          * queue up behind our exclusive invalidation operation.
1017          */
1018         if (test_and_clear_bit(FSCACHE_COOKIE_INVALIDATING, &cookie->flags))
1019                 wake_up_bit(&cookie->flags, FSCACHE_COOKIE_INVALIDATING);
1020         _leave(" [ok]");
1021         return transit_to(UPDATE_OBJECT);
1022
1023 nomem:
1024         fscache_mark_object_dead(object);
1025         fscache_unuse_cookie(object);
1026         _leave(" [ENOMEM]");
1027         return transit_to(KILL_OBJECT);
1028
1029 submit_op_failed:
1030         fscache_mark_object_dead(object);
1031         spin_unlock(&cookie->lock);
1032         fscache_unuse_cookie(object);
1033         kfree(op);
1034         _leave(" [EIO]");
1035         return transit_to(KILL_OBJECT);
1036 }
1037
1038 static const struct fscache_state *fscache_invalidate_object(struct fscache_object *object,
1039                                                              int event)
1040 {
1041         const struct fscache_state *s;
1042
1043         fscache_stat(&fscache_n_invalidates_run);
1044         fscache_stat(&fscache_n_cop_invalidate_object);
1045         s = _fscache_invalidate_object(object, event);
1046         fscache_stat_d(&fscache_n_cop_invalidate_object);
1047         return s;
1048 }
1049
1050 /*
1051  * Update auxiliary data.
1052  */
1053 static void fscache_update_aux_data(struct fscache_object *object)
1054 {
1055         fscache_stat(&fscache_n_updates_run);
1056         fscache_stat(&fscache_n_cop_update_object);
1057         object->cache->ops->update_object(object);
1058         fscache_stat_d(&fscache_n_cop_update_object);
1059 }
1060
1061 /*
1062  * Asynchronously update an object.
1063  */
1064 static const struct fscache_state *fscache_update_object(struct fscache_object *object,
1065                                                          int event)
1066 {
1067         _enter("{OBJ%x},%d", object->debug_id, event);
1068
1069         fscache_update_aux_data(object);
1070
1071         _leave("");
1072         return transit_to(WAIT_FOR_CMD);
1073 }
1074
1075 /**
1076  * fscache_object_retrying_stale - Note retrying stale object
1077  * @object: The object that will be retried
1078  *
1079  * Note that an object lookup found an on-disk object that was adjudged to be
1080  * stale and has been deleted.  The lookup will be retried.
1081  */
1082 void fscache_object_retrying_stale(struct fscache_object *object)
1083 {
1084         fscache_stat(&fscache_n_cache_no_space_reject);
1085 }
1086 EXPORT_SYMBOL(fscache_object_retrying_stale);
1087
1088 /**
1089  * fscache_object_mark_killed - Note that an object was killed
1090  * @object: The object that was culled
1091  * @why: The reason the object was killed.
1092  *
1093  * Note that an object was killed.  Returns true if the object was
1094  * already marked killed, false if it wasn't.
1095  */
1096 void fscache_object_mark_killed(struct fscache_object *object,
1097                                 enum fscache_why_object_killed why)
1098 {
1099         if (test_and_set_bit(FSCACHE_OBJECT_KILLED_BY_CACHE, &object->flags)) {
1100                 pr_err("Error: Object already killed by cache [%s]\n",
1101                        object->cache->identifier);
1102                 return;
1103         }
1104
1105         switch (why) {
1106         case FSCACHE_OBJECT_NO_SPACE:
1107                 fscache_stat(&fscache_n_cache_no_space_reject);
1108                 break;
1109         case FSCACHE_OBJECT_IS_STALE:
1110                 fscache_stat(&fscache_n_cache_stale_objects);
1111                 break;
1112         case FSCACHE_OBJECT_WAS_RETIRED:
1113                 fscache_stat(&fscache_n_cache_retired_objects);
1114                 break;
1115         case FSCACHE_OBJECT_WAS_CULLED:
1116                 fscache_stat(&fscache_n_cache_culled_objects);
1117                 break;
1118         }
1119 }
1120 EXPORT_SYMBOL(fscache_object_mark_killed);
1121
1122 /*
1123  * The object is dead.  We can get here if an object gets queued by an event
1124  * that would lead to its death (such as EV_KILL) when the dispatcher is
1125  * already running (and so can be requeued) but hasn't yet cleared the event
1126  * mask.
1127  */
1128 static const struct fscache_state *fscache_object_dead(struct fscache_object *object,
1129                                                        int event)
1130 {
1131         if (!test_and_set_bit(FSCACHE_OBJECT_RUN_AFTER_DEAD,
1132                               &object->flags))
1133                 return NO_TRANSIT;
1134
1135         WARN(true, "FS-Cache object redispatched after death");
1136         return NO_TRANSIT;
1137 }