fscache: Netfs function for cleanup post readpages
[pandora-kernel.git] / fs / fscache / page.c
1 /* Cache page management and data I/O routines
2  *
3  * Copyright (C) 2004-2008 Red Hat, Inc. All Rights Reserved.
4  * Written by David Howells (dhowells@redhat.com)
5  *
6  * This program is free software; you can redistribute it and/or
7  * modify it under the terms of the GNU General Public License
8  * as published by the Free Software Foundation; either version
9  * 2 of the License, or (at your option) any later version.
10  */
11
12 #define FSCACHE_DEBUG_LEVEL PAGE
13 #include <linux/module.h>
14 #include <linux/fscache-cache.h>
15 #include <linux/buffer_head.h>
16 #include <linux/pagevec.h>
17 #include <linux/slab.h>
18 #include "internal.h"
19
20 /*
21  * check to see if a page is being written to the cache
22  */
23 bool __fscache_check_page_write(struct fscache_cookie *cookie, struct page *page)
24 {
25         void *val;
26
27         rcu_read_lock();
28         val = radix_tree_lookup(&cookie->stores, page->index);
29         rcu_read_unlock();
30
31         return val != NULL;
32 }
33 EXPORT_SYMBOL(__fscache_check_page_write);
34
35 /*
36  * wait for a page to finish being written to the cache
37  */
38 void __fscache_wait_on_page_write(struct fscache_cookie *cookie, struct page *page)
39 {
40         wait_queue_head_t *wq = bit_waitqueue(&cookie->flags, 0);
41
42         wait_event(*wq, !__fscache_check_page_write(cookie, page));
43 }
44 EXPORT_SYMBOL(__fscache_wait_on_page_write);
45
46 /*
47  * decide whether a page can be released, possibly by cancelling a store to it
48  * - we're allowed to sleep if __GFP_WAIT is flagged
49  */
50 bool __fscache_maybe_release_page(struct fscache_cookie *cookie,
51                                   struct page *page,
52                                   gfp_t gfp)
53 {
54         struct page *xpage;
55         void *val;
56
57         _enter("%p,%p,%x", cookie, page, gfp);
58
59 try_again:
60         rcu_read_lock();
61         val = radix_tree_lookup(&cookie->stores, page->index);
62         if (!val) {
63                 rcu_read_unlock();
64                 fscache_stat(&fscache_n_store_vmscan_not_storing);
65                 __fscache_uncache_page(cookie, page);
66                 return true;
67         }
68
69         /* see if the page is actually undergoing storage - if so we can't get
70          * rid of it till the cache has finished with it */
71         if (radix_tree_tag_get(&cookie->stores, page->index,
72                                FSCACHE_COOKIE_STORING_TAG)) {
73                 rcu_read_unlock();
74                 goto page_busy;
75         }
76
77         /* the page is pending storage, so we attempt to cancel the store and
78          * discard the store request so that the page can be reclaimed */
79         spin_lock(&cookie->stores_lock);
80         rcu_read_unlock();
81
82         if (radix_tree_tag_get(&cookie->stores, page->index,
83                                FSCACHE_COOKIE_STORING_TAG)) {
84                 /* the page started to undergo storage whilst we were looking,
85                  * so now we can only wait or return */
86                 spin_unlock(&cookie->stores_lock);
87                 goto page_busy;
88         }
89
90         xpage = radix_tree_delete(&cookie->stores, page->index);
91         spin_unlock(&cookie->stores_lock);
92
93         if (xpage) {
94                 fscache_stat(&fscache_n_store_vmscan_cancelled);
95                 fscache_stat(&fscache_n_store_radix_deletes);
96                 ASSERTCMP(xpage, ==, page);
97         } else {
98                 fscache_stat(&fscache_n_store_vmscan_gone);
99         }
100
101         wake_up_bit(&cookie->flags, 0);
102         if (xpage)
103                 page_cache_release(xpage);
104         __fscache_uncache_page(cookie, page);
105         return true;
106
107 page_busy:
108         /* We will wait here if we're allowed to, but that could deadlock the
109          * allocator as the work threads writing to the cache may all end up
110          * sleeping on memory allocation, so we may need to impose a timeout
111          * too. */
112         if (!(gfp & __GFP_WAIT) || !(gfp & __GFP_FS)) {
113                 fscache_stat(&fscache_n_store_vmscan_busy);
114                 return false;
115         }
116
117         fscache_stat(&fscache_n_store_vmscan_wait);
118         __fscache_wait_on_page_write(cookie, page);
119         gfp &= ~__GFP_WAIT;
120         goto try_again;
121 }
122 EXPORT_SYMBOL(__fscache_maybe_release_page);
123
124 /*
125  * note that a page has finished being written to the cache
126  */
127 static void fscache_end_page_write(struct fscache_object *object,
128                                    struct page *page)
129 {
130         struct fscache_cookie *cookie;
131         struct page *xpage = NULL;
132
133         spin_lock(&object->lock);
134         cookie = object->cookie;
135         if (cookie) {
136                 /* delete the page from the tree if it is now no longer
137                  * pending */
138                 spin_lock(&cookie->stores_lock);
139                 radix_tree_tag_clear(&cookie->stores, page->index,
140                                      FSCACHE_COOKIE_STORING_TAG);
141                 if (!radix_tree_tag_get(&cookie->stores, page->index,
142                                         FSCACHE_COOKIE_PENDING_TAG)) {
143                         fscache_stat(&fscache_n_store_radix_deletes);
144                         xpage = radix_tree_delete(&cookie->stores, page->index);
145                 }
146                 spin_unlock(&cookie->stores_lock);
147                 wake_up_bit(&cookie->flags, 0);
148         }
149         spin_unlock(&object->lock);
150         if (xpage)
151                 page_cache_release(xpage);
152 }
153
154 /*
155  * actually apply the changed attributes to a cache object
156  */
157 static void fscache_attr_changed_op(struct fscache_operation *op)
158 {
159         struct fscache_object *object = op->object;
160         int ret;
161
162         _enter("{OBJ%x OP%x}", object->debug_id, op->debug_id);
163
164         fscache_stat(&fscache_n_attr_changed_calls);
165
166         if (fscache_object_is_active(object) &&
167             fscache_use_cookie(object)) {
168                 fscache_stat(&fscache_n_cop_attr_changed);
169                 ret = object->cache->ops->attr_changed(object);
170                 fscache_stat_d(&fscache_n_cop_attr_changed);
171                 fscache_unuse_cookie(object);
172                 if (ret < 0)
173                         fscache_abort_object(object);
174         }
175
176         fscache_op_complete(op, true);
177         _leave("");
178 }
179
180 /*
181  * notification that the attributes on an object have changed
182  */
183 int __fscache_attr_changed(struct fscache_cookie *cookie)
184 {
185         struct fscache_operation *op;
186         struct fscache_object *object;
187
188         _enter("%p", cookie);
189
190         ASSERTCMP(cookie->def->type, !=, FSCACHE_COOKIE_TYPE_INDEX);
191
192         fscache_stat(&fscache_n_attr_changed);
193
194         op = kzalloc(sizeof(*op), GFP_KERNEL);
195         if (!op) {
196                 fscache_stat(&fscache_n_attr_changed_nomem);
197                 _leave(" = -ENOMEM");
198                 return -ENOMEM;
199         }
200
201         fscache_operation_init(op, fscache_attr_changed_op, NULL);
202         op->flags = FSCACHE_OP_ASYNC | (1 << FSCACHE_OP_EXCLUSIVE);
203
204         spin_lock(&cookie->lock);
205
206         if (hlist_empty(&cookie->backing_objects))
207                 goto nobufs;
208         object = hlist_entry(cookie->backing_objects.first,
209                              struct fscache_object, cookie_link);
210
211         if (fscache_submit_exclusive_op(object, op) < 0)
212                 goto nobufs;
213         spin_unlock(&cookie->lock);
214         fscache_stat(&fscache_n_attr_changed_ok);
215         fscache_put_operation(op);
216         _leave(" = 0");
217         return 0;
218
219 nobufs:
220         spin_unlock(&cookie->lock);
221         kfree(op);
222         fscache_stat(&fscache_n_attr_changed_nobufs);
223         _leave(" = %d", -ENOBUFS);
224         return -ENOBUFS;
225 }
226 EXPORT_SYMBOL(__fscache_attr_changed);
227
228 /*
229  * release a retrieval op reference
230  */
231 static void fscache_release_retrieval_op(struct fscache_operation *_op)
232 {
233         struct fscache_retrieval *op =
234                 container_of(_op, struct fscache_retrieval, op);
235
236         _enter("{OP%x}", op->op.debug_id);
237
238         ASSERTCMP(atomic_read(&op->n_pages), ==, 0);
239
240         fscache_hist(fscache_retrieval_histogram, op->start_time);
241         if (op->context)
242                 fscache_put_context(op->op.object->cookie, op->context);
243
244         _leave("");
245 }
246
247 /*
248  * allocate a retrieval op
249  */
250 static struct fscache_retrieval *fscache_alloc_retrieval(
251         struct fscache_cookie *cookie,
252         struct address_space *mapping,
253         fscache_rw_complete_t end_io_func,
254         void *context)
255 {
256         struct fscache_retrieval *op;
257
258         /* allocate a retrieval operation and attempt to submit it */
259         op = kzalloc(sizeof(*op), GFP_NOIO);
260         if (!op) {
261                 fscache_stat(&fscache_n_retrievals_nomem);
262                 return NULL;
263         }
264
265         fscache_operation_init(&op->op, NULL, fscache_release_retrieval_op);
266         atomic_inc(&cookie->n_active);
267         op->op.flags    = FSCACHE_OP_MYTHREAD |
268                 (1UL << FSCACHE_OP_WAITING) |
269                 (1UL << FSCACHE_OP_UNUSE_COOKIE);
270         op->mapping     = mapping;
271         op->end_io_func = end_io_func;
272         op->context     = context;
273         op->start_time  = jiffies;
274         INIT_LIST_HEAD(&op->to_do);
275         return op;
276 }
277
278 /*
279  * wait for a deferred lookup to complete
280  */
281 int fscache_wait_for_deferred_lookup(struct fscache_cookie *cookie)
282 {
283         unsigned long jif;
284
285         _enter("");
286
287         if (!test_bit(FSCACHE_COOKIE_LOOKING_UP, &cookie->flags)) {
288                 _leave(" = 0 [imm]");
289                 return 0;
290         }
291
292         fscache_stat(&fscache_n_retrievals_wait);
293
294         jif = jiffies;
295         if (wait_on_bit(&cookie->flags, FSCACHE_COOKIE_LOOKING_UP,
296                         fscache_wait_bit_interruptible,
297                         TASK_INTERRUPTIBLE) != 0) {
298                 fscache_stat(&fscache_n_retrievals_intr);
299                 _leave(" = -ERESTARTSYS");
300                 return -ERESTARTSYS;
301         }
302
303         ASSERT(!test_bit(FSCACHE_COOKIE_LOOKING_UP, &cookie->flags));
304
305         smp_rmb();
306         fscache_hist(fscache_retrieval_delay_histogram, jif);
307         _leave(" = 0 [dly]");
308         return 0;
309 }
310
311 /*
312  * Handle cancellation of a pending retrieval op
313  */
314 static void fscache_do_cancel_retrieval(struct fscache_operation *_op)
315 {
316         struct fscache_retrieval *op =
317                 container_of(_op, struct fscache_retrieval, op);
318
319         atomic_set(&op->n_pages, 0);
320 }
321
322 /*
323  * wait for an object to become active (or dead)
324  */
325 int fscache_wait_for_operation_activation(struct fscache_object *object,
326                                           struct fscache_operation *op,
327                                           atomic_t *stat_op_waits,
328                                           atomic_t *stat_object_dead,
329                                           void (*do_cancel)(struct fscache_operation *))
330 {
331         int ret;
332
333         if (!test_bit(FSCACHE_OP_WAITING, &op->flags))
334                 goto check_if_dead;
335
336         _debug(">>> WT");
337         if (stat_op_waits)
338                 fscache_stat(stat_op_waits);
339         if (wait_on_bit(&op->flags, FSCACHE_OP_WAITING,
340                         fscache_wait_bit_interruptible,
341                         TASK_INTERRUPTIBLE) != 0) {
342                 ret = fscache_cancel_op(op, do_cancel);
343                 if (ret == 0)
344                         return -ERESTARTSYS;
345
346                 /* it's been removed from the pending queue by another party,
347                  * so we should get to run shortly */
348                 wait_on_bit(&op->flags, FSCACHE_OP_WAITING,
349                             fscache_wait_bit, TASK_UNINTERRUPTIBLE);
350         }
351         _debug("<<< GO");
352
353 check_if_dead:
354         if (op->state == FSCACHE_OP_ST_CANCELLED) {
355                 if (stat_object_dead)
356                         fscache_stat(stat_object_dead);
357                 _leave(" = -ENOBUFS [cancelled]");
358                 return -ENOBUFS;
359         }
360         if (unlikely(fscache_object_is_dead(object))) {
361                 pr_err("%s() = -ENOBUFS [obj dead %d]\n", __func__, op->state);
362                 fscache_cancel_op(op, do_cancel);
363                 if (stat_object_dead)
364                         fscache_stat(stat_object_dead);
365                 return -ENOBUFS;
366         }
367         return 0;
368 }
369
370 /*
371  * read a page from the cache or allocate a block in which to store it
372  * - we return:
373  *   -ENOMEM    - out of memory, nothing done
374  *   -ERESTARTSYS - interrupted
375  *   -ENOBUFS   - no backing object available in which to cache the block
376  *   -ENODATA   - no data available in the backing object for this block
377  *   0          - dispatched a read - it'll call end_io_func() when finished
378  */
379 int __fscache_read_or_alloc_page(struct fscache_cookie *cookie,
380                                  struct page *page,
381                                  fscache_rw_complete_t end_io_func,
382                                  void *context,
383                                  gfp_t gfp)
384 {
385         struct fscache_retrieval *op;
386         struct fscache_object *object;
387         int ret;
388
389         _enter("%p,%p,,,", cookie, page);
390
391         fscache_stat(&fscache_n_retrievals);
392
393         if (hlist_empty(&cookie->backing_objects))
394                 goto nobufs;
395
396         if (test_bit(FSCACHE_COOKIE_INVALIDATING, &cookie->flags)) {
397                 _leave(" = -ENOBUFS [invalidating]");
398                 return -ENOBUFS;
399         }
400
401         ASSERTCMP(cookie->def->type, !=, FSCACHE_COOKIE_TYPE_INDEX);
402         ASSERTCMP(page, !=, NULL);
403
404         if (fscache_wait_for_deferred_lookup(cookie) < 0)
405                 return -ERESTARTSYS;
406
407         op = fscache_alloc_retrieval(cookie, page->mapping,
408                                      end_io_func,context);
409         if (!op) {
410                 _leave(" = -ENOMEM");
411                 return -ENOMEM;
412         }
413         atomic_set(&op->n_pages, 1);
414
415         spin_lock(&cookie->lock);
416
417         if (hlist_empty(&cookie->backing_objects))
418                 goto nobufs_unlock;
419         object = hlist_entry(cookie->backing_objects.first,
420                              struct fscache_object, cookie_link);
421
422         ASSERT(test_bit(FSCACHE_OBJECT_IS_LOOKED_UP, &object->flags));
423
424         atomic_inc(&object->n_reads);
425         __set_bit(FSCACHE_OP_DEC_READ_CNT, &op->op.flags);
426
427         if (fscache_submit_op(object, &op->op) < 0)
428                 goto nobufs_unlock_dec;
429         spin_unlock(&cookie->lock);
430
431         fscache_stat(&fscache_n_retrieval_ops);
432
433         /* pin the netfs read context in case we need to do the actual netfs
434          * read because we've encountered a cache read failure */
435         fscache_get_context(object->cookie, op->context);
436
437         /* we wait for the operation to become active, and then process it
438          * *here*, in this thread, and not in the thread pool */
439         ret = fscache_wait_for_operation_activation(
440                 object, &op->op,
441                 __fscache_stat(&fscache_n_retrieval_op_waits),
442                 __fscache_stat(&fscache_n_retrievals_object_dead),
443                 fscache_do_cancel_retrieval);
444         if (ret < 0)
445                 goto error;
446
447         /* ask the cache to honour the operation */
448         if (test_bit(FSCACHE_COOKIE_NO_DATA_YET, &object->cookie->flags)) {
449                 fscache_stat(&fscache_n_cop_allocate_page);
450                 ret = object->cache->ops->allocate_page(op, page, gfp);
451                 fscache_stat_d(&fscache_n_cop_allocate_page);
452                 if (ret == 0)
453                         ret = -ENODATA;
454         } else {
455                 fscache_stat(&fscache_n_cop_read_or_alloc_page);
456                 ret = object->cache->ops->read_or_alloc_page(op, page, gfp);
457                 fscache_stat_d(&fscache_n_cop_read_or_alloc_page);
458         }
459
460 error:
461         if (ret == -ENOMEM)
462                 fscache_stat(&fscache_n_retrievals_nomem);
463         else if (ret == -ERESTARTSYS)
464                 fscache_stat(&fscache_n_retrievals_intr);
465         else if (ret == -ENODATA)
466                 fscache_stat(&fscache_n_retrievals_nodata);
467         else if (ret < 0)
468                 fscache_stat(&fscache_n_retrievals_nobufs);
469         else
470                 fscache_stat(&fscache_n_retrievals_ok);
471
472         fscache_put_retrieval(op);
473         _leave(" = %d", ret);
474         return ret;
475
476 nobufs_unlock_dec:
477         atomic_dec(&object->n_reads);
478 nobufs_unlock:
479         spin_unlock(&cookie->lock);
480         atomic_dec(&cookie->n_active);
481         kfree(op);
482 nobufs:
483         fscache_stat(&fscache_n_retrievals_nobufs);
484         _leave(" = -ENOBUFS");
485         return -ENOBUFS;
486 }
487 EXPORT_SYMBOL(__fscache_read_or_alloc_page);
488
489 /*
490  * read a list of page from the cache or allocate a block in which to store
491  * them
492  * - we return:
493  *   -ENOMEM    - out of memory, some pages may be being read
494  *   -ERESTARTSYS - interrupted, some pages may be being read
495  *   -ENOBUFS   - no backing object or space available in which to cache any
496  *                pages not being read
497  *   -ENODATA   - no data available in the backing object for some or all of
498  *                the pages
499  *   0          - dispatched a read on all pages
500  *
501  * end_io_func() will be called for each page read from the cache as it is
502  * finishes being read
503  *
504  * any pages for which a read is dispatched will be removed from pages and
505  * nr_pages
506  */
507 int __fscache_read_or_alloc_pages(struct fscache_cookie *cookie,
508                                   struct address_space *mapping,
509                                   struct list_head *pages,
510                                   unsigned *nr_pages,
511                                   fscache_rw_complete_t end_io_func,
512                                   void *context,
513                                   gfp_t gfp)
514 {
515         struct fscache_retrieval *op;
516         struct fscache_object *object;
517         int ret;
518
519         _enter("%p,,%d,,,", cookie, *nr_pages);
520
521         fscache_stat(&fscache_n_retrievals);
522
523         if (hlist_empty(&cookie->backing_objects))
524                 goto nobufs;
525
526         if (test_bit(FSCACHE_COOKIE_INVALIDATING, &cookie->flags)) {
527                 _leave(" = -ENOBUFS [invalidating]");
528                 return -ENOBUFS;
529         }
530
531         ASSERTCMP(cookie->def->type, !=, FSCACHE_COOKIE_TYPE_INDEX);
532         ASSERTCMP(*nr_pages, >, 0);
533         ASSERT(!list_empty(pages));
534
535         if (fscache_wait_for_deferred_lookup(cookie) < 0)
536                 return -ERESTARTSYS;
537
538         op = fscache_alloc_retrieval(cookie, mapping, end_io_func, context);
539         if (!op)
540                 return -ENOMEM;
541         atomic_set(&op->n_pages, *nr_pages);
542
543         spin_lock(&cookie->lock);
544
545         if (hlist_empty(&cookie->backing_objects))
546                 goto nobufs_unlock;
547         object = hlist_entry(cookie->backing_objects.first,
548                              struct fscache_object, cookie_link);
549
550         atomic_inc(&object->n_reads);
551         __set_bit(FSCACHE_OP_DEC_READ_CNT, &op->op.flags);
552
553         if (fscache_submit_op(object, &op->op) < 0)
554                 goto nobufs_unlock_dec;
555         spin_unlock(&cookie->lock);
556
557         fscache_stat(&fscache_n_retrieval_ops);
558
559         /* pin the netfs read context in case we need to do the actual netfs
560          * read because we've encountered a cache read failure */
561         fscache_get_context(object->cookie, op->context);
562
563         /* we wait for the operation to become active, and then process it
564          * *here*, in this thread, and not in the thread pool */
565         ret = fscache_wait_for_operation_activation(
566                 object, &op->op,
567                 __fscache_stat(&fscache_n_retrieval_op_waits),
568                 __fscache_stat(&fscache_n_retrievals_object_dead),
569                 fscache_do_cancel_retrieval);
570         if (ret < 0)
571                 goto error;
572
573         /* ask the cache to honour the operation */
574         if (test_bit(FSCACHE_COOKIE_NO_DATA_YET, &object->cookie->flags)) {
575                 fscache_stat(&fscache_n_cop_allocate_pages);
576                 ret = object->cache->ops->allocate_pages(
577                         op, pages, nr_pages, gfp);
578                 fscache_stat_d(&fscache_n_cop_allocate_pages);
579         } else {
580                 fscache_stat(&fscache_n_cop_read_or_alloc_pages);
581                 ret = object->cache->ops->read_or_alloc_pages(
582                         op, pages, nr_pages, gfp);
583                 fscache_stat_d(&fscache_n_cop_read_or_alloc_pages);
584         }
585
586 error:
587         if (ret == -ENOMEM)
588                 fscache_stat(&fscache_n_retrievals_nomem);
589         else if (ret == -ERESTARTSYS)
590                 fscache_stat(&fscache_n_retrievals_intr);
591         else if (ret == -ENODATA)
592                 fscache_stat(&fscache_n_retrievals_nodata);
593         else if (ret < 0)
594                 fscache_stat(&fscache_n_retrievals_nobufs);
595         else
596                 fscache_stat(&fscache_n_retrievals_ok);
597
598         fscache_put_retrieval(op);
599         _leave(" = %d", ret);
600         return ret;
601
602 nobufs_unlock_dec:
603         atomic_dec(&object->n_reads);
604 nobufs_unlock:
605         spin_unlock(&cookie->lock);
606         atomic_dec(&cookie->n_active);
607         kfree(op);
608 nobufs:
609         fscache_stat(&fscache_n_retrievals_nobufs);
610         _leave(" = -ENOBUFS");
611         return -ENOBUFS;
612 }
613 EXPORT_SYMBOL(__fscache_read_or_alloc_pages);
614
615 /*
616  * allocate a block in the cache on which to store a page
617  * - we return:
618  *   -ENOMEM    - out of memory, nothing done
619  *   -ERESTARTSYS - interrupted
620  *   -ENOBUFS   - no backing object available in which to cache the block
621  *   0          - block allocated
622  */
623 int __fscache_alloc_page(struct fscache_cookie *cookie,
624                          struct page *page,
625                          gfp_t gfp)
626 {
627         struct fscache_retrieval *op;
628         struct fscache_object *object;
629         int ret;
630
631         _enter("%p,%p,,,", cookie, page);
632
633         fscache_stat(&fscache_n_allocs);
634
635         if (hlist_empty(&cookie->backing_objects))
636                 goto nobufs;
637
638         ASSERTCMP(cookie->def->type, !=, FSCACHE_COOKIE_TYPE_INDEX);
639         ASSERTCMP(page, !=, NULL);
640
641         if (test_bit(FSCACHE_COOKIE_INVALIDATING, &cookie->flags)) {
642                 _leave(" = -ENOBUFS [invalidating]");
643                 return -ENOBUFS;
644         }
645
646         if (fscache_wait_for_deferred_lookup(cookie) < 0)
647                 return -ERESTARTSYS;
648
649         op = fscache_alloc_retrieval(cookie, page->mapping, NULL, NULL);
650         if (!op)
651                 return -ENOMEM;
652         atomic_set(&op->n_pages, 1);
653
654         spin_lock(&cookie->lock);
655
656         if (hlist_empty(&cookie->backing_objects))
657                 goto nobufs_unlock;
658         object = hlist_entry(cookie->backing_objects.first,
659                              struct fscache_object, cookie_link);
660
661         if (fscache_submit_op(object, &op->op) < 0)
662                 goto nobufs_unlock;
663         spin_unlock(&cookie->lock);
664
665         fscache_stat(&fscache_n_alloc_ops);
666
667         ret = fscache_wait_for_operation_activation(
668                 object, &op->op,
669                 __fscache_stat(&fscache_n_alloc_op_waits),
670                 __fscache_stat(&fscache_n_allocs_object_dead),
671                 fscache_do_cancel_retrieval);
672         if (ret < 0)
673                 goto error;
674
675         /* ask the cache to honour the operation */
676         fscache_stat(&fscache_n_cop_allocate_page);
677         ret = object->cache->ops->allocate_page(op, page, gfp);
678         fscache_stat_d(&fscache_n_cop_allocate_page);
679
680 error:
681         if (ret == -ERESTARTSYS)
682                 fscache_stat(&fscache_n_allocs_intr);
683         else if (ret < 0)
684                 fscache_stat(&fscache_n_allocs_nobufs);
685         else
686                 fscache_stat(&fscache_n_allocs_ok);
687
688         fscache_put_retrieval(op);
689         _leave(" = %d", ret);
690         return ret;
691
692 nobufs_unlock:
693         spin_unlock(&cookie->lock);
694         atomic_dec(&cookie->n_active);
695         kfree(op);
696 nobufs:
697         fscache_stat(&fscache_n_allocs_nobufs);
698         _leave(" = -ENOBUFS");
699         return -ENOBUFS;
700 }
701 EXPORT_SYMBOL(__fscache_alloc_page);
702
703 /*
704  * Unmark pages allocate in the readahead code path (via:
705  * fscache_readpages_or_alloc) after delegating to the base filesystem
706  */
707 void __fscache_readpages_cancel(struct fscache_cookie *cookie,
708                                 struct list_head *pages)
709 {
710         struct page *page;
711
712         list_for_each_entry(page, pages, lru) {
713                 if (PageFsCache(page))
714                         __fscache_uncache_page(cookie, page);
715         }
716 }
717 EXPORT_SYMBOL(__fscache_readpages_cancel);
718
719 /*
720  * release a write op reference
721  */
722 static void fscache_release_write_op(struct fscache_operation *_op)
723 {
724         _enter("{OP%x}", _op->debug_id);
725 }
726
727 /*
728  * perform the background storage of a page into the cache
729  */
730 static void fscache_write_op(struct fscache_operation *_op)
731 {
732         struct fscache_storage *op =
733                 container_of(_op, struct fscache_storage, op);
734         struct fscache_object *object = op->op.object;
735         struct fscache_cookie *cookie;
736         struct page *page;
737         unsigned n;
738         void *results[1];
739         int ret;
740
741         _enter("{OP%x,%d}", op->op.debug_id, atomic_read(&op->op.usage));
742
743         spin_lock(&object->lock);
744         cookie = object->cookie;
745
746         if (!fscache_object_is_active(object)) {
747                 /* If we get here, then the on-disk cache object likely longer
748                  * exists, so we should just cancel this write operation.
749                  */
750                 spin_unlock(&object->lock);
751                 fscache_op_complete(&op->op, false);
752                 _leave(" [inactive]");
753                 return;
754         }
755
756         if (!cookie) {
757                 /* If we get here, then the cookie belonging to the object was
758                  * detached, probably by the cookie being withdrawn due to
759                  * memory pressure, which means that the pages we might write
760                  * to the cache from no longer exist - therefore, we can just
761                  * cancel this write operation.
762                  */
763                 spin_unlock(&object->lock);
764                 fscache_op_complete(&op->op, false);
765                 _leave(" [cancel] op{f=%lx s=%u} obj{s=%s f=%lx}",
766                        _op->flags, _op->state, object->state->short_name,
767                        object->flags);
768                 return;
769         }
770
771         spin_lock(&cookie->stores_lock);
772
773         fscache_stat(&fscache_n_store_calls);
774
775         /* find a page to store */
776         page = NULL;
777         n = radix_tree_gang_lookup_tag(&cookie->stores, results, 0, 1,
778                                        FSCACHE_COOKIE_PENDING_TAG);
779         if (n != 1)
780                 goto superseded;
781         page = results[0];
782         _debug("gang %d [%lx]", n, page->index);
783         if (page->index > op->store_limit) {
784                 fscache_stat(&fscache_n_store_pages_over_limit);
785                 goto superseded;
786         }
787
788         radix_tree_tag_set(&cookie->stores, page->index,
789                            FSCACHE_COOKIE_STORING_TAG);
790         radix_tree_tag_clear(&cookie->stores, page->index,
791                              FSCACHE_COOKIE_PENDING_TAG);
792
793         spin_unlock(&cookie->stores_lock);
794         spin_unlock(&object->lock);
795
796         fscache_stat(&fscache_n_store_pages);
797         fscache_stat(&fscache_n_cop_write_page);
798         ret = object->cache->ops->write_page(op, page);
799         fscache_stat_d(&fscache_n_cop_write_page);
800         fscache_end_page_write(object, page);
801         if (ret < 0) {
802                 fscache_abort_object(object);
803                 fscache_op_complete(&op->op, true);
804         } else {
805                 fscache_enqueue_operation(&op->op);
806         }
807
808         _leave("");
809         return;
810
811 superseded:
812         /* this writer is going away and there aren't any more things to
813          * write */
814         _debug("cease");
815         spin_unlock(&cookie->stores_lock);
816         clear_bit(FSCACHE_OBJECT_PENDING_WRITE, &object->flags);
817         spin_unlock(&object->lock);
818         fscache_op_complete(&op->op, true);
819         _leave("");
820 }
821
822 /*
823  * Clear the pages pending writing for invalidation
824  */
825 void fscache_invalidate_writes(struct fscache_cookie *cookie)
826 {
827         struct page *page;
828         void *results[16];
829         int n, i;
830
831         _enter("");
832
833         for (;;) {
834                 spin_lock(&cookie->stores_lock);
835                 n = radix_tree_gang_lookup_tag(&cookie->stores, results, 0,
836                                                ARRAY_SIZE(results),
837                                                FSCACHE_COOKIE_PENDING_TAG);
838                 if (n == 0) {
839                         spin_unlock(&cookie->stores_lock);
840                         break;
841                 }
842
843                 for (i = n - 1; i >= 0; i--) {
844                         page = results[i];
845                         radix_tree_delete(&cookie->stores, page->index);
846                 }
847
848                 spin_unlock(&cookie->stores_lock);
849
850                 for (i = n - 1; i >= 0; i--)
851                         page_cache_release(results[i]);
852         }
853
854         _leave("");
855 }
856
857 /*
858  * request a page be stored in the cache
859  * - returns:
860  *   -ENOMEM    - out of memory, nothing done
861  *   -ENOBUFS   - no backing object available in which to cache the page
862  *   0          - dispatched a write - it'll call end_io_func() when finished
863  *
864  * if the cookie still has a backing object at this point, that object can be
865  * in one of a few states with respect to storage processing:
866  *
867  *  (1) negative lookup, object not yet created (FSCACHE_COOKIE_CREATING is
868  *      set)
869  *
870  *      (a) no writes yet
871  *
872  *      (b) writes deferred till post-creation (mark page for writing and
873  *          return immediately)
874  *
875  *  (2) negative lookup, object created, initial fill being made from netfs
876  *
877  *      (a) fill point not yet reached this page (mark page for writing and
878  *          return)
879  *
880  *      (b) fill point passed this page (queue op to store this page)
881  *
882  *  (3) object extant (queue op to store this page)
883  *
884  * any other state is invalid
885  */
886 int __fscache_write_page(struct fscache_cookie *cookie,
887                          struct page *page,
888                          gfp_t gfp)
889 {
890         struct fscache_storage *op;
891         struct fscache_object *object;
892         int ret;
893
894         _enter("%p,%x,", cookie, (u32) page->flags);
895
896         ASSERTCMP(cookie->def->type, !=, FSCACHE_COOKIE_TYPE_INDEX);
897         ASSERT(PageFsCache(page));
898
899         fscache_stat(&fscache_n_stores);
900
901         if (test_bit(FSCACHE_COOKIE_INVALIDATING, &cookie->flags)) {
902                 _leave(" = -ENOBUFS [invalidating]");
903                 return -ENOBUFS;
904         }
905
906         op = kzalloc(sizeof(*op), GFP_NOIO | __GFP_NOMEMALLOC | __GFP_NORETRY);
907         if (!op)
908                 goto nomem;
909
910         fscache_operation_init(&op->op, fscache_write_op,
911                                fscache_release_write_op);
912         op->op.flags = FSCACHE_OP_ASYNC |
913                 (1 << FSCACHE_OP_WAITING) |
914                 (1 << FSCACHE_OP_UNUSE_COOKIE);
915
916         ret = radix_tree_preload(gfp & ~__GFP_HIGHMEM);
917         if (ret < 0)
918                 goto nomem_free;
919
920         ret = -ENOBUFS;
921         spin_lock(&cookie->lock);
922
923         if (hlist_empty(&cookie->backing_objects))
924                 goto nobufs;
925         object = hlist_entry(cookie->backing_objects.first,
926                              struct fscache_object, cookie_link);
927         if (test_bit(FSCACHE_IOERROR, &object->cache->flags))
928                 goto nobufs;
929
930         /* add the page to the pending-storage radix tree on the backing
931          * object */
932         spin_lock(&object->lock);
933         spin_lock(&cookie->stores_lock);
934
935         _debug("store limit %llx", (unsigned long long) object->store_limit);
936
937         ret = radix_tree_insert(&cookie->stores, page->index, page);
938         if (ret < 0) {
939                 if (ret == -EEXIST)
940                         goto already_queued;
941                 _debug("insert failed %d", ret);
942                 goto nobufs_unlock_obj;
943         }
944
945         radix_tree_tag_set(&cookie->stores, page->index,
946                            FSCACHE_COOKIE_PENDING_TAG);
947         page_cache_get(page);
948
949         /* we only want one writer at a time, but we do need to queue new
950          * writers after exclusive ops */
951         if (test_and_set_bit(FSCACHE_OBJECT_PENDING_WRITE, &object->flags))
952                 goto already_pending;
953
954         spin_unlock(&cookie->stores_lock);
955         spin_unlock(&object->lock);
956
957         op->op.debug_id = atomic_inc_return(&fscache_op_debug_id);
958         op->store_limit = object->store_limit;
959
960         atomic_inc(&cookie->n_active);
961         if (fscache_submit_op(object, &op->op) < 0)
962                 goto submit_failed;
963
964         spin_unlock(&cookie->lock);
965         radix_tree_preload_end();
966         fscache_stat(&fscache_n_store_ops);
967         fscache_stat(&fscache_n_stores_ok);
968
969         /* the work queue now carries its own ref on the object */
970         fscache_put_operation(&op->op);
971         _leave(" = 0");
972         return 0;
973
974 already_queued:
975         fscache_stat(&fscache_n_stores_again);
976 already_pending:
977         spin_unlock(&cookie->stores_lock);
978         spin_unlock(&object->lock);
979         spin_unlock(&cookie->lock);
980         radix_tree_preload_end();
981         kfree(op);
982         fscache_stat(&fscache_n_stores_ok);
983         _leave(" = 0");
984         return 0;
985
986 submit_failed:
987         atomic_dec(&cookie->n_active);
988         spin_lock(&cookie->stores_lock);
989         radix_tree_delete(&cookie->stores, page->index);
990         spin_unlock(&cookie->stores_lock);
991         page_cache_release(page);
992         ret = -ENOBUFS;
993         goto nobufs;
994
995 nobufs_unlock_obj:
996         spin_unlock(&cookie->stores_lock);
997         spin_unlock(&object->lock);
998 nobufs:
999         spin_unlock(&cookie->lock);
1000         radix_tree_preload_end();
1001         kfree(op);
1002         fscache_stat(&fscache_n_stores_nobufs);
1003         _leave(" = -ENOBUFS");
1004         return -ENOBUFS;
1005
1006 nomem_free:
1007         kfree(op);
1008 nomem:
1009         fscache_stat(&fscache_n_stores_oom);
1010         _leave(" = -ENOMEM");
1011         return -ENOMEM;
1012 }
1013 EXPORT_SYMBOL(__fscache_write_page);
1014
1015 /*
1016  * remove a page from the cache
1017  */
1018 void __fscache_uncache_page(struct fscache_cookie *cookie, struct page *page)
1019 {
1020         struct fscache_object *object;
1021
1022         _enter(",%p", page);
1023
1024         ASSERTCMP(cookie->def->type, !=, FSCACHE_COOKIE_TYPE_INDEX);
1025         ASSERTCMP(page, !=, NULL);
1026
1027         fscache_stat(&fscache_n_uncaches);
1028
1029         /* cache withdrawal may beat us to it */
1030         if (!PageFsCache(page))
1031                 goto done;
1032
1033         /* get the object */
1034         spin_lock(&cookie->lock);
1035
1036         if (hlist_empty(&cookie->backing_objects)) {
1037                 ClearPageFsCache(page);
1038                 goto done_unlock;
1039         }
1040
1041         object = hlist_entry(cookie->backing_objects.first,
1042                              struct fscache_object, cookie_link);
1043
1044         /* there might now be stuff on disk we could read */
1045         clear_bit(FSCACHE_COOKIE_NO_DATA_YET, &cookie->flags);
1046
1047         /* only invoke the cache backend if we managed to mark the page
1048          * uncached here; this deals with synchronisation vs withdrawal */
1049         if (TestClearPageFsCache(page) &&
1050             object->cache->ops->uncache_page) {
1051                 /* the cache backend releases the cookie lock */
1052                 fscache_stat(&fscache_n_cop_uncache_page);
1053                 object->cache->ops->uncache_page(object, page);
1054                 fscache_stat_d(&fscache_n_cop_uncache_page);
1055                 goto done;
1056         }
1057
1058 done_unlock:
1059         spin_unlock(&cookie->lock);
1060 done:
1061         _leave("");
1062 }
1063 EXPORT_SYMBOL(__fscache_uncache_page);
1064
1065 /**
1066  * fscache_mark_page_cached - Mark a page as being cached
1067  * @op: The retrieval op pages are being marked for
1068  * @page: The page to be marked
1069  *
1070  * Mark a netfs page as being cached.  After this is called, the netfs
1071  * must call fscache_uncache_page() to remove the mark.
1072  */
1073 void fscache_mark_page_cached(struct fscache_retrieval *op, struct page *page)
1074 {
1075         struct fscache_cookie *cookie = op->op.object->cookie;
1076
1077 #ifdef CONFIG_FSCACHE_STATS
1078         atomic_inc(&fscache_n_marks);
1079 #endif
1080
1081         _debug("- mark %p{%lx}", page, page->index);
1082         if (TestSetPageFsCache(page)) {
1083                 static bool once_only;
1084                 if (!once_only) {
1085                         once_only = true;
1086                         printk(KERN_WARNING "FS-Cache:"
1087                                " Cookie type %s marked page %lx"
1088                                " multiple times\n",
1089                                cookie->def->name, page->index);
1090                 }
1091         }
1092
1093         if (cookie->def->mark_page_cached)
1094                 cookie->def->mark_page_cached(cookie->netfs_data,
1095                                               op->mapping, page);
1096 }
1097 EXPORT_SYMBOL(fscache_mark_page_cached);
1098
1099 /**
1100  * fscache_mark_pages_cached - Mark pages as being cached
1101  * @op: The retrieval op pages are being marked for
1102  * @pagevec: The pages to be marked
1103  *
1104  * Mark a bunch of netfs pages as being cached.  After this is called,
1105  * the netfs must call fscache_uncache_page() to remove the mark.
1106  */
1107 void fscache_mark_pages_cached(struct fscache_retrieval *op,
1108                                struct pagevec *pagevec)
1109 {
1110         unsigned long loop;
1111
1112         for (loop = 0; loop < pagevec->nr; loop++)
1113                 fscache_mark_page_cached(op, pagevec->pages[loop]);
1114
1115         pagevec_reinit(pagevec);
1116 }
1117 EXPORT_SYMBOL(fscache_mark_pages_cached);
1118
1119 /*
1120  * Uncache all the pages in an inode that are marked PG_fscache, assuming them
1121  * to be associated with the given cookie.
1122  */
1123 void __fscache_uncache_all_inode_pages(struct fscache_cookie *cookie,
1124                                        struct inode *inode)
1125 {
1126         struct address_space *mapping = inode->i_mapping;
1127         struct pagevec pvec;
1128         pgoff_t next;
1129         int i;
1130
1131         _enter("%p,%p", cookie, inode);
1132
1133         if (!mapping || mapping->nrpages == 0) {
1134                 _leave(" [no pages]");
1135                 return;
1136         }
1137
1138         pagevec_init(&pvec, 0);
1139         next = 0;
1140         do {
1141                 if (!pagevec_lookup(&pvec, mapping, next, PAGEVEC_SIZE))
1142                         break;
1143                 for (i = 0; i < pagevec_count(&pvec); i++) {
1144                         struct page *page = pvec.pages[i];
1145                         next = page->index;
1146                         if (PageFsCache(page)) {
1147                                 __fscache_wait_on_page_write(cookie, page);
1148                                 __fscache_uncache_page(cookie, page);
1149                         }
1150                 }
1151                 pagevec_release(&pvec);
1152                 cond_resched();
1153         } while (++next);
1154
1155         _leave("");
1156 }
1157 EXPORT_SYMBOL(__fscache_uncache_all_inode_pages);