Merge branch 'for-linus' of ssh://gitolite.kernel.org/pub/scm/linux/kernel/git/jikos...
[sfrench/cifs-2.6.git] / 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 static void fscache_put_object(struct fscache_object *);
143 static bool fscache_enqueue_dependents(struct fscache_object *, int);
144 static void fscache_dequeue_object(struct fscache_object *);
145
146 /*
147  * we need to notify the parent when an op completes that we had outstanding
148  * upon it
149  */
150 static inline void fscache_done_parent_op(struct fscache_object *object)
151 {
152         struct fscache_object *parent = object->parent;
153
154         _enter("OBJ%x {OBJ%x,%x}",
155                object->debug_id, parent->debug_id, parent->n_ops);
156
157         spin_lock_nested(&parent->lock, 1);
158         parent->n_obj_ops--;
159         parent->n_ops--;
160         if (parent->n_ops == 0)
161                 fscache_raise_event(parent, FSCACHE_OBJECT_EV_CLEARED);
162         spin_unlock(&parent->lock);
163 }
164
165 /*
166  * Object state machine dispatcher.
167  */
168 static void fscache_object_sm_dispatcher(struct fscache_object *object)
169 {
170         const struct fscache_transition *t;
171         const struct fscache_state *state, *new_state;
172         unsigned long events, event_mask;
173         int event = -1;
174
175         ASSERT(object != NULL);
176
177         _enter("{OBJ%x,%s,%lx}",
178                object->debug_id, object->state->name, object->events);
179
180         event_mask = object->event_mask;
181 restart:
182         object->event_mask = 0; /* Mask normal event handling */
183         state = object->state;
184 restart_masked:
185         events = object->events;
186
187         /* Handle any out-of-band events (typically an error) */
188         if (events & object->oob_event_mask) {
189                 _debug("{OBJ%x} oob %lx",
190                        object->debug_id, events & object->oob_event_mask);
191                 for (t = object->oob_table; t->events; t++) {
192                         if (events & t->events) {
193                                 state = t->transit_to;
194                                 ASSERT(state->work != NULL);
195                                 event = fls(events & t->events) - 1;
196                                 __clear_bit(event, &object->oob_event_mask);
197                                 clear_bit(event, &object->events);
198                                 goto execute_work_state;
199                         }
200                 }
201         }
202
203         /* Wait states are just transition tables */
204         if (!state->work) {
205                 if (events & event_mask) {
206                         for (t = state->transitions; t->events; t++) {
207                                 if (events & t->events) {
208                                         new_state = t->transit_to;
209                                         event = fls(events & t->events) - 1;
210                                         clear_bit(event, &object->events);
211                                         _debug("{OBJ%x} ev %d: %s -> %s",
212                                                object->debug_id, event,
213                                                state->name, new_state->name);
214                                         object->state = state = new_state;
215                                         goto execute_work_state;
216                                 }
217                         }
218
219                         /* The event mask didn't include all the tabled bits */
220                         BUG();
221                 }
222                 /* Randomly woke up */
223                 goto unmask_events;
224         }
225
226 execute_work_state:
227         _debug("{OBJ%x} exec %s", object->debug_id, state->name);
228
229         new_state = state->work(object, event);
230         event = -1;
231         if (new_state == NO_TRANSIT) {
232                 _debug("{OBJ%x} %s notrans", object->debug_id, state->name);
233                 if (unlikely(state == STATE(OBJECT_DEAD))) {
234                         _leave(" [dead]");
235                         return;
236                 }
237                 fscache_enqueue_object(object);
238                 event_mask = object->oob_event_mask;
239                 goto unmask_events;
240         }
241
242         _debug("{OBJ%x} %s -> %s",
243                object->debug_id, state->name, new_state->name);
244         object->state = state = new_state;
245
246         if (state->work) {
247                 if (unlikely(state == STATE(OBJECT_DEAD))) {
248                         _leave(" [dead]");
249                         return;
250                 }
251                 goto restart_masked;
252         }
253
254         /* Transited to wait state */
255         event_mask = object->oob_event_mask;
256         for (t = state->transitions; t->events; t++)
257                 event_mask |= t->events;
258
259 unmask_events:
260         object->event_mask = event_mask;
261         smp_mb();
262         events = object->events;
263         if (events & event_mask)
264                 goto restart;
265         _leave(" [msk %lx]", event_mask);
266 }
267
268 /*
269  * execute an object
270  */
271 static void fscache_object_work_func(struct work_struct *work)
272 {
273         struct fscache_object *object =
274                 container_of(work, struct fscache_object, work);
275         unsigned long start;
276
277         _enter("{OBJ%x}", object->debug_id);
278
279         start = jiffies;
280         fscache_object_sm_dispatcher(object);
281         fscache_hist(fscache_objs_histogram, start);
282         fscache_put_object(object);
283 }
284
285 /**
286  * fscache_object_init - Initialise a cache object description
287  * @object: Object description
288  * @cookie: Cookie object will be attached to
289  * @cache: Cache in which backing object will be found
290  *
291  * Initialise a cache object description to its basic values.
292  *
293  * See Documentation/filesystems/caching/backend-api.txt for a complete
294  * description.
295  */
296 void fscache_object_init(struct fscache_object *object,
297                          struct fscache_cookie *cookie,
298                          struct fscache_cache *cache)
299 {
300         const struct fscache_transition *t;
301
302         atomic_inc(&cache->object_count);
303
304         object->state = STATE(WAIT_FOR_INIT);
305         object->oob_table = fscache_osm_init_oob;
306         object->flags = 1 << FSCACHE_OBJECT_IS_LIVE;
307         spin_lock_init(&object->lock);
308         INIT_LIST_HEAD(&object->cache_link);
309         INIT_HLIST_NODE(&object->cookie_link);
310         INIT_WORK(&object->work, fscache_object_work_func);
311         INIT_LIST_HEAD(&object->dependents);
312         INIT_LIST_HEAD(&object->dep_link);
313         INIT_LIST_HEAD(&object->pending_ops);
314         object->n_children = 0;
315         object->n_ops = object->n_in_progress = object->n_exclusive = 0;
316         object->events = 0;
317         object->store_limit = 0;
318         object->store_limit_l = 0;
319         object->cache = cache;
320         object->cookie = cookie;
321         object->parent = NULL;
322 #ifdef CONFIG_FSCACHE_OBJECT_LIST
323         RB_CLEAR_NODE(&object->objlist_link);
324 #endif
325
326         object->oob_event_mask = 0;
327         for (t = object->oob_table; t->events; t++)
328                 object->oob_event_mask |= t->events;
329         object->event_mask = object->oob_event_mask;
330         for (t = object->state->transitions; t->events; t++)
331                 object->event_mask |= t->events;
332 }
333 EXPORT_SYMBOL(fscache_object_init);
334
335 /*
336  * Mark the object as no longer being live, making sure that we synchronise
337  * against op submission.
338  */
339 static inline void fscache_mark_object_dead(struct fscache_object *object)
340 {
341         spin_lock(&object->lock);
342         clear_bit(FSCACHE_OBJECT_IS_LIVE, &object->flags);
343         spin_unlock(&object->lock);
344 }
345
346 /*
347  * Abort object initialisation before we start it.
348  */
349 static const struct fscache_state *fscache_abort_initialisation(struct fscache_object *object,
350                                                                 int event)
351 {
352         _enter("{OBJ%x},%d", object->debug_id, event);
353
354         object->oob_event_mask = 0;
355         fscache_dequeue_object(object);
356         return transit_to(KILL_OBJECT);
357 }
358
359 /*
360  * initialise an object
361  * - check the specified object's parent to see if we can make use of it
362  *   immediately to do a creation
363  * - we may need to start the process of creating a parent and we need to wait
364  *   for the parent's lookup and creation to complete if it's not there yet
365  */
366 static const struct fscache_state *fscache_initialise_object(struct fscache_object *object,
367                                                              int event)
368 {
369         struct fscache_object *parent;
370         bool success;
371
372         _enter("{OBJ%x},%d", object->debug_id, event);
373
374         ASSERT(list_empty(&object->dep_link));
375
376         parent = object->parent;
377         if (!parent) {
378                 _leave(" [no parent]");
379                 return transit_to(DROP_OBJECT);
380         }
381
382         _debug("parent: %s of:%lx", parent->state->name, parent->flags);
383
384         if (fscache_object_is_dying(parent)) {
385                 _leave(" [bad parent]");
386                 return transit_to(DROP_OBJECT);
387         }
388
389         if (fscache_object_is_available(parent)) {
390                 _leave(" [ready]");
391                 return transit_to(PARENT_READY);
392         }
393
394         _debug("wait");
395
396         spin_lock(&parent->lock);
397         fscache_stat(&fscache_n_cop_grab_object);
398         success = false;
399         if (fscache_object_is_live(parent) &&
400             object->cache->ops->grab_object(object)) {
401                 list_add(&object->dep_link, &parent->dependents);
402                 success = true;
403         }
404         fscache_stat_d(&fscache_n_cop_grab_object);
405         spin_unlock(&parent->lock);
406         if (!success) {
407                 _leave(" [grab failed]");
408                 return transit_to(DROP_OBJECT);
409         }
410
411         /* fscache_acquire_non_index_cookie() uses this
412          * to wake the chain up */
413         fscache_raise_event(parent, FSCACHE_OBJECT_EV_NEW_CHILD);
414         _leave(" [wait]");
415         return transit_to(WAIT_FOR_PARENT);
416 }
417
418 /*
419  * Once the parent object is ready, we should kick off our lookup op.
420  */
421 static const struct fscache_state *fscache_parent_ready(struct fscache_object *object,
422                                                         int event)
423 {
424         struct fscache_object *parent = object->parent;
425
426         _enter("{OBJ%x},%d", object->debug_id, event);
427
428         ASSERT(parent != NULL);
429
430         spin_lock(&parent->lock);
431         parent->n_ops++;
432         parent->n_obj_ops++;
433         object->lookup_jif = jiffies;
434         spin_unlock(&parent->lock);
435
436         _leave("");
437         return transit_to(LOOK_UP_OBJECT);
438 }
439
440 /*
441  * look an object up in the cache from which it was allocated
442  * - we hold an "access lock" on the parent object, so the parent object cannot
443  *   be withdrawn by either party till we've finished
444  */
445 static const struct fscache_state *fscache_look_up_object(struct fscache_object *object,
446                                                           int event)
447 {
448         struct fscache_cookie *cookie = object->cookie;
449         struct fscache_object *parent = object->parent;
450         int ret;
451
452         _enter("{OBJ%x},%d", object->debug_id, event);
453
454         object->oob_table = fscache_osm_lookup_oob;
455
456         ASSERT(parent != NULL);
457         ASSERTCMP(parent->n_ops, >, 0);
458         ASSERTCMP(parent->n_obj_ops, >, 0);
459
460         /* make sure the parent is still available */
461         ASSERT(fscache_object_is_available(parent));
462
463         if (fscache_object_is_dying(parent) ||
464             test_bit(FSCACHE_IOERROR, &object->cache->flags) ||
465             !fscache_use_cookie(object)) {
466                 _leave(" [unavailable]");
467                 return transit_to(LOOKUP_FAILURE);
468         }
469
470         _debug("LOOKUP \"%s\" in \"%s\"",
471                cookie->def->name, object->cache->tag->name);
472
473         fscache_stat(&fscache_n_object_lookups);
474         fscache_stat(&fscache_n_cop_lookup_object);
475         ret = object->cache->ops->lookup_object(object);
476         fscache_stat_d(&fscache_n_cop_lookup_object);
477
478         fscache_unuse_cookie(object);
479
480         if (ret == -ETIMEDOUT) {
481                 /* probably stuck behind another object, so move this one to
482                  * the back of the queue */
483                 fscache_stat(&fscache_n_object_lookups_timed_out);
484                 _leave(" [timeout]");
485                 return NO_TRANSIT;
486         }
487
488         if (ret < 0) {
489                 _leave(" [error]");
490                 return transit_to(LOOKUP_FAILURE);
491         }
492
493         _leave(" [ok]");
494         return transit_to(OBJECT_AVAILABLE);
495 }
496
497 /**
498  * fscache_object_lookup_negative - Note negative cookie lookup
499  * @object: Object pointing to cookie to mark
500  *
501  * Note negative lookup, permitting those waiting to read data from an already
502  * existing backing object to continue as there's no data for them to read.
503  */
504 void fscache_object_lookup_negative(struct fscache_object *object)
505 {
506         struct fscache_cookie *cookie = object->cookie;
507
508         _enter("{OBJ%x,%s}", object->debug_id, object->state->name);
509
510         if (!test_and_set_bit(FSCACHE_OBJECT_IS_LOOKED_UP, &object->flags)) {
511                 fscache_stat(&fscache_n_object_lookups_negative);
512
513                 /* Allow write requests to begin stacking up and read requests to begin
514                  * returning ENODATA.
515                  */
516                 set_bit(FSCACHE_COOKIE_NO_DATA_YET, &cookie->flags);
517                 clear_bit(FSCACHE_COOKIE_UNAVAILABLE, &cookie->flags);
518
519                 _debug("wake up lookup %p", &cookie->flags);
520                 clear_bit_unlock(FSCACHE_COOKIE_LOOKING_UP, &cookie->flags);
521                 wake_up_bit(&cookie->flags, FSCACHE_COOKIE_LOOKING_UP);
522         }
523         _leave("");
524 }
525 EXPORT_SYMBOL(fscache_object_lookup_negative);
526
527 /**
528  * fscache_obtained_object - Note successful object lookup or creation
529  * @object: Object pointing to cookie to mark
530  *
531  * Note successful lookup and/or creation, permitting those waiting to write
532  * data to a backing object to continue.
533  *
534  * Note that after calling this, an object's cookie may be relinquished by the
535  * netfs, and so must be accessed with object lock held.
536  */
537 void fscache_obtained_object(struct fscache_object *object)
538 {
539         struct fscache_cookie *cookie = object->cookie;
540
541         _enter("{OBJ%x,%s}", object->debug_id, object->state->name);
542
543         /* if we were still looking up, then we must have a positive lookup
544          * result, in which case there may be data available */
545         if (!test_and_set_bit(FSCACHE_OBJECT_IS_LOOKED_UP, &object->flags)) {
546                 fscache_stat(&fscache_n_object_lookups_positive);
547
548                 /* We do (presumably) have data */
549                 clear_bit_unlock(FSCACHE_COOKIE_NO_DATA_YET, &cookie->flags);
550                 clear_bit(FSCACHE_COOKIE_UNAVAILABLE, &cookie->flags);
551
552                 /* Allow write requests to begin stacking up and read requests
553                  * to begin shovelling data.
554                  */
555                 clear_bit_unlock(FSCACHE_COOKIE_LOOKING_UP, &cookie->flags);
556                 wake_up_bit(&cookie->flags, FSCACHE_COOKIE_LOOKING_UP);
557         } else {
558                 fscache_stat(&fscache_n_object_created);
559         }
560
561         set_bit(FSCACHE_OBJECT_IS_AVAILABLE, &object->flags);
562         _leave("");
563 }
564 EXPORT_SYMBOL(fscache_obtained_object);
565
566 /*
567  * handle an object that has just become available
568  */
569 static const struct fscache_state *fscache_object_available(struct fscache_object *object,
570                                                             int event)
571 {
572         _enter("{OBJ%x},%d", object->debug_id, event);
573
574         object->oob_table = fscache_osm_run_oob;
575
576         spin_lock(&object->lock);
577
578         fscache_done_parent_op(object);
579         if (object->n_in_progress == 0) {
580                 if (object->n_ops > 0) {
581                         ASSERTCMP(object->n_ops, >=, object->n_obj_ops);
582                         fscache_start_operations(object);
583                 } else {
584                         ASSERT(list_empty(&object->pending_ops));
585                 }
586         }
587         spin_unlock(&object->lock);
588
589         fscache_stat(&fscache_n_cop_lookup_complete);
590         object->cache->ops->lookup_complete(object);
591         fscache_stat_d(&fscache_n_cop_lookup_complete);
592
593         fscache_hist(fscache_obj_instantiate_histogram, object->lookup_jif);
594         fscache_stat(&fscache_n_object_avail);
595
596         _leave("");
597         return transit_to(JUMPSTART_DEPS);
598 }
599
600 /*
601  * Wake up this object's dependent objects now that we've become available.
602  */
603 static const struct fscache_state *fscache_jumpstart_dependents(struct fscache_object *object,
604                                                                 int event)
605 {
606         _enter("{OBJ%x},%d", object->debug_id, event);
607
608         if (!fscache_enqueue_dependents(object, FSCACHE_OBJECT_EV_PARENT_READY))
609                 return NO_TRANSIT; /* Not finished; requeue */
610         return transit_to(WAIT_FOR_CMD);
611 }
612
613 /*
614  * Handle lookup or creation failute.
615  */
616 static const struct fscache_state *fscache_lookup_failure(struct fscache_object *object,
617                                                           int event)
618 {
619         struct fscache_cookie *cookie;
620
621         _enter("{OBJ%x},%d", object->debug_id, event);
622
623         object->oob_event_mask = 0;
624
625         fscache_stat(&fscache_n_cop_lookup_complete);
626         object->cache->ops->lookup_complete(object);
627         fscache_stat_d(&fscache_n_cop_lookup_complete);
628
629         set_bit(FSCACHE_OBJECT_KILLED_BY_CACHE, &object->flags);
630
631         cookie = object->cookie;
632         set_bit(FSCACHE_COOKIE_UNAVAILABLE, &cookie->flags);
633         if (test_and_clear_bit(FSCACHE_COOKIE_LOOKING_UP, &cookie->flags))
634                 wake_up_bit(&cookie->flags, FSCACHE_COOKIE_LOOKING_UP);
635
636         fscache_done_parent_op(object);
637         return transit_to(KILL_OBJECT);
638 }
639
640 /*
641  * Wait for completion of all active operations on this object and the death of
642  * all child objects of this object.
643  */
644 static const struct fscache_state *fscache_kill_object(struct fscache_object *object,
645                                                        int event)
646 {
647         _enter("{OBJ%x,%d,%d},%d",
648                object->debug_id, object->n_ops, object->n_children, event);
649
650         fscache_mark_object_dead(object);
651         object->oob_event_mask = 0;
652
653         if (test_bit(FSCACHE_OBJECT_RETIRED, &object->flags)) {
654                 /* Reject any new read/write ops and abort any that are pending. */
655                 clear_bit(FSCACHE_OBJECT_PENDING_WRITE, &object->flags);
656                 fscache_cancel_all_ops(object);
657         }
658
659         if (list_empty(&object->dependents) &&
660             object->n_ops == 0 &&
661             object->n_children == 0)
662                 return transit_to(DROP_OBJECT);
663
664         if (object->n_in_progress == 0) {
665                 spin_lock(&object->lock);
666                 if (object->n_ops > 0 && object->n_in_progress == 0)
667                         fscache_start_operations(object);
668                 spin_unlock(&object->lock);
669         }
670
671         if (!list_empty(&object->dependents))
672                 return transit_to(KILL_DEPENDENTS);
673
674         return transit_to(WAIT_FOR_CLEARANCE);
675 }
676
677 /*
678  * Kill dependent objects.
679  */
680 static const struct fscache_state *fscache_kill_dependents(struct fscache_object *object,
681                                                            int event)
682 {
683         _enter("{OBJ%x},%d", object->debug_id, event);
684
685         if (!fscache_enqueue_dependents(object, FSCACHE_OBJECT_EV_KILL))
686                 return NO_TRANSIT; /* Not finished */
687         return transit_to(WAIT_FOR_CLEARANCE);
688 }
689
690 /*
691  * Drop an object's attachments
692  */
693 static const struct fscache_state *fscache_drop_object(struct fscache_object *object,
694                                                        int event)
695 {
696         struct fscache_object *parent = object->parent;
697         struct fscache_cookie *cookie = object->cookie;
698         struct fscache_cache *cache = object->cache;
699         bool awaken = false;
700
701         _enter("{OBJ%x,%d},%d", object->debug_id, object->n_children, event);
702
703         ASSERT(cookie != NULL);
704         ASSERT(!hlist_unhashed(&object->cookie_link));
705
706         /* Make sure the cookie no longer points here and that the netfs isn't
707          * waiting for us.
708          */
709         spin_lock(&cookie->lock);
710         hlist_del_init(&object->cookie_link);
711         if (hlist_empty(&cookie->backing_objects) &&
712             test_and_clear_bit(FSCACHE_COOKIE_INVALIDATING, &cookie->flags))
713                 awaken = true;
714         spin_unlock(&cookie->lock);
715
716         if (awaken)
717                 wake_up_bit(&cookie->flags, FSCACHE_COOKIE_INVALIDATING);
718
719         /* Prevent a race with our last child, which has to signal EV_CLEARED
720          * before dropping our spinlock.
721          */
722         spin_lock(&object->lock);
723         spin_unlock(&object->lock);
724
725         /* Discard from the cache's collection of objects */
726         spin_lock(&cache->object_list_lock);
727         list_del_init(&object->cache_link);
728         spin_unlock(&cache->object_list_lock);
729
730         fscache_stat(&fscache_n_cop_drop_object);
731         cache->ops->drop_object(object);
732         fscache_stat_d(&fscache_n_cop_drop_object);
733
734         /* The parent object wants to know when all it dependents have gone */
735         if (parent) {
736                 _debug("release parent OBJ%x {%d}",
737                        parent->debug_id, parent->n_children);
738
739                 spin_lock(&parent->lock);
740                 parent->n_children--;
741                 if (parent->n_children == 0)
742                         fscache_raise_event(parent, FSCACHE_OBJECT_EV_CLEARED);
743                 spin_unlock(&parent->lock);
744                 object->parent = NULL;
745         }
746
747         /* this just shifts the object release to the work processor */
748         fscache_put_object(object);
749         fscache_stat(&fscache_n_object_dead);
750
751         _leave("");
752         return transit_to(OBJECT_DEAD);
753 }
754
755 /*
756  * get a ref on an object
757  */
758 static int fscache_get_object(struct fscache_object *object)
759 {
760         int ret;
761
762         fscache_stat(&fscache_n_cop_grab_object);
763         ret = object->cache->ops->grab_object(object) ? 0 : -EAGAIN;
764         fscache_stat_d(&fscache_n_cop_grab_object);
765         return ret;
766 }
767
768 /*
769  * Discard a ref on an object
770  */
771 static void fscache_put_object(struct fscache_object *object)
772 {
773         fscache_stat(&fscache_n_cop_put_object);
774         object->cache->ops->put_object(object);
775         fscache_stat_d(&fscache_n_cop_put_object);
776 }
777
778 /**
779  * fscache_object_destroy - Note that a cache object is about to be destroyed
780  * @object: The object to be destroyed
781  *
782  * Note the imminent destruction and deallocation of a cache object record.
783  */
784 void fscache_object_destroy(struct fscache_object *object)
785 {
786         fscache_objlist_remove(object);
787
788         /* We can get rid of the cookie now */
789         fscache_cookie_put(object->cookie);
790         object->cookie = NULL;
791 }
792 EXPORT_SYMBOL(fscache_object_destroy);
793
794 /*
795  * enqueue an object for metadata-type processing
796  */
797 void fscache_enqueue_object(struct fscache_object *object)
798 {
799         _enter("{OBJ%x}", object->debug_id);
800
801         if (fscache_get_object(object) >= 0) {
802                 wait_queue_head_t *cong_wq =
803                         &get_cpu_var(fscache_object_cong_wait);
804
805                 if (queue_work(fscache_object_wq, &object->work)) {
806                         if (fscache_object_congested())
807                                 wake_up(cong_wq);
808                 } else
809                         fscache_put_object(object);
810
811                 put_cpu_var(fscache_object_cong_wait);
812         }
813 }
814
815 /**
816  * fscache_object_sleep_till_congested - Sleep until object wq is congested
817  * @timeoutp: Scheduler sleep timeout
818  *
819  * Allow an object handler to sleep until the object workqueue is congested.
820  *
821  * The caller must set up a wake up event before calling this and must have set
822  * the appropriate sleep mode (such as TASK_UNINTERRUPTIBLE) and tested its own
823  * condition before calling this function as no test is made here.
824  *
825  * %true is returned if the object wq is congested, %false otherwise.
826  */
827 bool fscache_object_sleep_till_congested(signed long *timeoutp)
828 {
829         wait_queue_head_t *cong_wq = this_cpu_ptr(&fscache_object_cong_wait);
830         DEFINE_WAIT(wait);
831
832         if (fscache_object_congested())
833                 return true;
834
835         add_wait_queue_exclusive(cong_wq, &wait);
836         if (!fscache_object_congested())
837                 *timeoutp = schedule_timeout(*timeoutp);
838         finish_wait(cong_wq, &wait);
839
840         return fscache_object_congested();
841 }
842 EXPORT_SYMBOL_GPL(fscache_object_sleep_till_congested);
843
844 /*
845  * Enqueue the dependents of an object for metadata-type processing.
846  *
847  * If we don't manage to finish the list before the scheduler wants to run
848  * again then return false immediately.  We return true if the list was
849  * cleared.
850  */
851 static bool fscache_enqueue_dependents(struct fscache_object *object, int event)
852 {
853         struct fscache_object *dep;
854         bool ret = true;
855
856         _enter("{OBJ%x}", object->debug_id);
857
858         if (list_empty(&object->dependents))
859                 return true;
860
861         spin_lock(&object->lock);
862
863         while (!list_empty(&object->dependents)) {
864                 dep = list_entry(object->dependents.next,
865                                  struct fscache_object, dep_link);
866                 list_del_init(&dep->dep_link);
867
868                 fscache_raise_event(dep, event);
869                 fscache_put_object(dep);
870
871                 if (!list_empty(&object->dependents) && need_resched()) {
872                         ret = false;
873                         break;
874                 }
875         }
876
877         spin_unlock(&object->lock);
878         return ret;
879 }
880
881 /*
882  * remove an object from whatever queue it's waiting on
883  */
884 static void fscache_dequeue_object(struct fscache_object *object)
885 {
886         _enter("{OBJ%x}", object->debug_id);
887
888         if (!list_empty(&object->dep_link)) {
889                 spin_lock(&object->parent->lock);
890                 list_del_init(&object->dep_link);
891                 spin_unlock(&object->parent->lock);
892         }
893
894         _leave("");
895 }
896
897 /**
898  * fscache_check_aux - Ask the netfs whether an object on disk is still valid
899  * @object: The object to ask about
900  * @data: The auxiliary data for the object
901  * @datalen: The size of the auxiliary data
902  *
903  * This function consults the netfs about the coherency state of an object.
904  * The caller must be holding a ref on cookie->n_active (held by
905  * fscache_look_up_object() on behalf of the cache backend during object lookup
906  * and creation).
907  */
908 enum fscache_checkaux fscache_check_aux(struct fscache_object *object,
909                                         const void *data, uint16_t datalen)
910 {
911         enum fscache_checkaux result;
912
913         if (!object->cookie->def->check_aux) {
914                 fscache_stat(&fscache_n_checkaux_none);
915                 return FSCACHE_CHECKAUX_OKAY;
916         }
917
918         result = object->cookie->def->check_aux(object->cookie->netfs_data,
919                                                 data, datalen);
920         switch (result) {
921                 /* entry okay as is */
922         case FSCACHE_CHECKAUX_OKAY:
923                 fscache_stat(&fscache_n_checkaux_okay);
924                 break;
925
926                 /* entry requires update */
927         case FSCACHE_CHECKAUX_NEEDS_UPDATE:
928                 fscache_stat(&fscache_n_checkaux_update);
929                 break;
930
931                 /* entry requires deletion */
932         case FSCACHE_CHECKAUX_OBSOLETE:
933                 fscache_stat(&fscache_n_checkaux_obsolete);
934                 break;
935
936         default:
937                 BUG();
938         }
939
940         return result;
941 }
942 EXPORT_SYMBOL(fscache_check_aux);
943
944 /*
945  * Asynchronously invalidate an object.
946  */
947 static const struct fscache_state *_fscache_invalidate_object(struct fscache_object *object,
948                                                               int event)
949 {
950         struct fscache_operation *op;
951         struct fscache_cookie *cookie = object->cookie;
952
953         _enter("{OBJ%x},%d", object->debug_id, event);
954
955         /* We're going to need the cookie.  If the cookie is not available then
956          * retire the object instead.
957          */
958         if (!fscache_use_cookie(object)) {
959                 ASSERT(object->cookie->stores.rnode == NULL);
960                 set_bit(FSCACHE_OBJECT_RETIRED, &object->flags);
961                 _leave(" [no cookie]");
962                 return transit_to(KILL_OBJECT);
963         }
964
965         /* Reject any new read/write ops and abort any that are pending. */
966         fscache_invalidate_writes(cookie);
967         clear_bit(FSCACHE_OBJECT_PENDING_WRITE, &object->flags);
968         fscache_cancel_all_ops(object);
969
970         /* Now we have to wait for in-progress reads and writes */
971         op = kzalloc(sizeof(*op), GFP_KERNEL);
972         if (!op)
973                 goto nomem;
974
975         fscache_operation_init(op, object->cache->ops->invalidate_object,
976                                NULL, NULL);
977         op->flags = FSCACHE_OP_ASYNC |
978                 (1 << FSCACHE_OP_EXCLUSIVE) |
979                 (1 << FSCACHE_OP_UNUSE_COOKIE);
980
981         spin_lock(&cookie->lock);
982         if (fscache_submit_exclusive_op(object, op) < 0)
983                 goto submit_op_failed;
984         spin_unlock(&cookie->lock);
985         fscache_put_operation(op);
986
987         /* Once we've completed the invalidation, we know there will be no data
988          * stored in the cache and thus we can reinstate the data-check-skip
989          * optimisation.
990          */
991         set_bit(FSCACHE_COOKIE_NO_DATA_YET, &cookie->flags);
992
993         /* We can allow read and write requests to come in once again.  They'll
994          * queue up behind our exclusive invalidation operation.
995          */
996         if (test_and_clear_bit(FSCACHE_COOKIE_INVALIDATING, &cookie->flags))
997                 wake_up_bit(&cookie->flags, FSCACHE_COOKIE_INVALIDATING);
998         _leave(" [ok]");
999         return transit_to(UPDATE_OBJECT);
1000
1001 nomem:
1002         fscache_mark_object_dead(object);
1003         fscache_unuse_cookie(object);
1004         _leave(" [ENOMEM]");
1005         return transit_to(KILL_OBJECT);
1006
1007 submit_op_failed:
1008         fscache_mark_object_dead(object);
1009         spin_unlock(&cookie->lock);
1010         fscache_unuse_cookie(object);
1011         kfree(op);
1012         _leave(" [EIO]");
1013         return transit_to(KILL_OBJECT);
1014 }
1015
1016 static const struct fscache_state *fscache_invalidate_object(struct fscache_object *object,
1017                                                              int event)
1018 {
1019         const struct fscache_state *s;
1020
1021         fscache_stat(&fscache_n_invalidates_run);
1022         fscache_stat(&fscache_n_cop_invalidate_object);
1023         s = _fscache_invalidate_object(object, event);
1024         fscache_stat_d(&fscache_n_cop_invalidate_object);
1025         return s;
1026 }
1027
1028 /*
1029  * Asynchronously update an object.
1030  */
1031 static const struct fscache_state *fscache_update_object(struct fscache_object *object,
1032                                                          int event)
1033 {
1034         _enter("{OBJ%x},%d", object->debug_id, event);
1035
1036         fscache_stat(&fscache_n_updates_run);
1037         fscache_stat(&fscache_n_cop_update_object);
1038         object->cache->ops->update_object(object);
1039         fscache_stat_d(&fscache_n_cop_update_object);
1040
1041         _leave("");
1042         return transit_to(WAIT_FOR_CMD);
1043 }
1044
1045 /**
1046  * fscache_object_retrying_stale - Note retrying stale object
1047  * @object: The object that will be retried
1048  *
1049  * Note that an object lookup found an on-disk object that was adjudged to be
1050  * stale and has been deleted.  The lookup will be retried.
1051  */
1052 void fscache_object_retrying_stale(struct fscache_object *object)
1053 {
1054         fscache_stat(&fscache_n_cache_no_space_reject);
1055 }
1056 EXPORT_SYMBOL(fscache_object_retrying_stale);
1057
1058 /**
1059  * fscache_object_mark_killed - Note that an object was killed
1060  * @object: The object that was culled
1061  * @why: The reason the object was killed.
1062  *
1063  * Note that an object was killed.  Returns true if the object was
1064  * already marked killed, false if it wasn't.
1065  */
1066 void fscache_object_mark_killed(struct fscache_object *object,
1067                                 enum fscache_why_object_killed why)
1068 {
1069         if (test_and_set_bit(FSCACHE_OBJECT_KILLED_BY_CACHE, &object->flags)) {
1070                 pr_err("Error: Object already killed by cache [%s]\n",
1071                        object->cache->identifier);
1072                 return;
1073         }
1074
1075         switch (why) {
1076         case FSCACHE_OBJECT_NO_SPACE:
1077                 fscache_stat(&fscache_n_cache_no_space_reject);
1078                 break;
1079         case FSCACHE_OBJECT_IS_STALE:
1080                 fscache_stat(&fscache_n_cache_stale_objects);
1081                 break;
1082         case FSCACHE_OBJECT_WAS_RETIRED:
1083                 fscache_stat(&fscache_n_cache_retired_objects);
1084                 break;
1085         case FSCACHE_OBJECT_WAS_CULLED:
1086                 fscache_stat(&fscache_n_cache_culled_objects);
1087                 break;
1088         }
1089 }
1090 EXPORT_SYMBOL(fscache_object_mark_killed);
1091
1092 /*
1093  * The object is dead.  We can get here if an object gets queued by an event
1094  * that would lead to its death (such as EV_KILL) when the dispatcher is
1095  * already running (and so can be requeued) but hasn't yet cleared the event
1096  * mask.
1097  */
1098 static const struct fscache_state *fscache_object_dead(struct fscache_object *object,
1099                                                        int event)
1100 {
1101         if (!test_and_set_bit(FSCACHE_OBJECT_RUN_AFTER_DEAD,
1102                               &object->flags))
1103                 return NO_TRANSIT;
1104
1105         WARN(true, "FS-Cache object redispatched after death");
1106         return NO_TRANSIT;
1107 }