]> Pileus Git - ~andy/linux/blob - fs/fscache/page.c
Merge tag 'staging-3.15-rc6' of git://git.kernel.org/pub/scm/linux/kernel/git/gregkh...
[~andy/linux] / 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_stat(&fscache_n_cop_attr_changed);
168                 ret = object->cache->ops->attr_changed(object);
169                 fscache_stat_d(&fscache_n_cop_attr_changed);
170                 if (ret < 0)
171                         fscache_abort_object(object);
172         }
173
174         fscache_op_complete(op, true);
175         _leave("");
176 }
177
178 /*
179  * notification that the attributes on an object have changed
180  */
181 int __fscache_attr_changed(struct fscache_cookie *cookie)
182 {
183         struct fscache_operation *op;
184         struct fscache_object *object;
185         bool wake_cookie;
186
187         _enter("%p", cookie);
188
189         ASSERTCMP(cookie->def->type, !=, FSCACHE_COOKIE_TYPE_INDEX);
190
191         fscache_stat(&fscache_n_attr_changed);
192
193         op = kzalloc(sizeof(*op), GFP_KERNEL);
194         if (!op) {
195                 fscache_stat(&fscache_n_attr_changed_nomem);
196                 _leave(" = -ENOMEM");
197                 return -ENOMEM;
198         }
199
200         fscache_operation_init(op, fscache_attr_changed_op, NULL);
201         op->flags = FSCACHE_OP_ASYNC |
202                 (1 << FSCACHE_OP_EXCLUSIVE) |
203                 (1 << FSCACHE_OP_UNUSE_COOKIE);
204
205         spin_lock(&cookie->lock);
206
207         if (!fscache_cookie_enabled(cookie) ||
208             hlist_empty(&cookie->backing_objects))
209                 goto nobufs;
210         object = hlist_entry(cookie->backing_objects.first,
211                              struct fscache_object, cookie_link);
212
213         __fscache_use_cookie(cookie);
214         if (fscache_submit_exclusive_op(object, op) < 0)
215                 goto nobufs;
216         spin_unlock(&cookie->lock);
217         fscache_stat(&fscache_n_attr_changed_ok);
218         fscache_put_operation(op);
219         _leave(" = 0");
220         return 0;
221
222 nobufs:
223         wake_cookie = __fscache_unuse_cookie(cookie);
224         spin_unlock(&cookie->lock);
225         kfree(op);
226         if (wake_cookie)
227                 __fscache_wake_unused_cookie(cookie);
228         fscache_stat(&fscache_n_attr_changed_nobufs);
229         _leave(" = %d", -ENOBUFS);
230         return -ENOBUFS;
231 }
232 EXPORT_SYMBOL(__fscache_attr_changed);
233
234 /*
235  * release a retrieval op reference
236  */
237 static void fscache_release_retrieval_op(struct fscache_operation *_op)
238 {
239         struct fscache_retrieval *op =
240                 container_of(_op, struct fscache_retrieval, op);
241
242         _enter("{OP%x}", op->op.debug_id);
243
244         ASSERTCMP(atomic_read(&op->n_pages), ==, 0);
245
246         fscache_hist(fscache_retrieval_histogram, op->start_time);
247         if (op->context)
248                 fscache_put_context(op->op.object->cookie, op->context);
249
250         _leave("");
251 }
252
253 /*
254  * allocate a retrieval op
255  */
256 static struct fscache_retrieval *fscache_alloc_retrieval(
257         struct fscache_cookie *cookie,
258         struct address_space *mapping,
259         fscache_rw_complete_t end_io_func,
260         void *context)
261 {
262         struct fscache_retrieval *op;
263
264         /* allocate a retrieval operation and attempt to submit it */
265         op = kzalloc(sizeof(*op), GFP_NOIO);
266         if (!op) {
267                 fscache_stat(&fscache_n_retrievals_nomem);
268                 return NULL;
269         }
270
271         fscache_operation_init(&op->op, NULL, fscache_release_retrieval_op);
272         op->op.flags    = FSCACHE_OP_MYTHREAD |
273                 (1UL << FSCACHE_OP_WAITING) |
274                 (1UL << FSCACHE_OP_UNUSE_COOKIE);
275         op->mapping     = mapping;
276         op->end_io_func = end_io_func;
277         op->context     = context;
278         op->start_time  = jiffies;
279         INIT_LIST_HEAD(&op->to_do);
280         return op;
281 }
282
283 /*
284  * wait for a deferred lookup to complete
285  */
286 int fscache_wait_for_deferred_lookup(struct fscache_cookie *cookie)
287 {
288         unsigned long jif;
289
290         _enter("");
291
292         if (!test_bit(FSCACHE_COOKIE_LOOKING_UP, &cookie->flags)) {
293                 _leave(" = 0 [imm]");
294                 return 0;
295         }
296
297         fscache_stat(&fscache_n_retrievals_wait);
298
299         jif = jiffies;
300         if (wait_on_bit(&cookie->flags, FSCACHE_COOKIE_LOOKING_UP,
301                         fscache_wait_bit_interruptible,
302                         TASK_INTERRUPTIBLE) != 0) {
303                 fscache_stat(&fscache_n_retrievals_intr);
304                 _leave(" = -ERESTARTSYS");
305                 return -ERESTARTSYS;
306         }
307
308         ASSERT(!test_bit(FSCACHE_COOKIE_LOOKING_UP, &cookie->flags));
309
310         smp_rmb();
311         fscache_hist(fscache_retrieval_delay_histogram, jif);
312         _leave(" = 0 [dly]");
313         return 0;
314 }
315
316 /*
317  * Handle cancellation of a pending retrieval op
318  */
319 static void fscache_do_cancel_retrieval(struct fscache_operation *_op)
320 {
321         struct fscache_retrieval *op =
322                 container_of(_op, struct fscache_retrieval, op);
323
324         atomic_set(&op->n_pages, 0);
325 }
326
327 /*
328  * wait for an object to become active (or dead)
329  */
330 int fscache_wait_for_operation_activation(struct fscache_object *object,
331                                           struct fscache_operation *op,
332                                           atomic_t *stat_op_waits,
333                                           atomic_t *stat_object_dead,
334                                           void (*do_cancel)(struct fscache_operation *))
335 {
336         int ret;
337
338         if (!test_bit(FSCACHE_OP_WAITING, &op->flags))
339                 goto check_if_dead;
340
341         _debug(">>> WT");
342         if (stat_op_waits)
343                 fscache_stat(stat_op_waits);
344         if (wait_on_bit(&op->flags, FSCACHE_OP_WAITING,
345                         fscache_wait_bit_interruptible,
346                         TASK_INTERRUPTIBLE) != 0) {
347                 ret = fscache_cancel_op(op, do_cancel);
348                 if (ret == 0)
349                         return -ERESTARTSYS;
350
351                 /* it's been removed from the pending queue by another party,
352                  * so we should get to run shortly */
353                 wait_on_bit(&op->flags, FSCACHE_OP_WAITING,
354                             fscache_wait_bit, TASK_UNINTERRUPTIBLE);
355         }
356         _debug("<<< GO");
357
358 check_if_dead:
359         if (op->state == FSCACHE_OP_ST_CANCELLED) {
360                 if (stat_object_dead)
361                         fscache_stat(stat_object_dead);
362                 _leave(" = -ENOBUFS [cancelled]");
363                 return -ENOBUFS;
364         }
365         if (unlikely(fscache_object_is_dead(object))) {
366                 pr_err("%s() = -ENOBUFS [obj dead %d]\n", __func__, op->state);
367                 fscache_cancel_op(op, do_cancel);
368                 if (stat_object_dead)
369                         fscache_stat(stat_object_dead);
370                 return -ENOBUFS;
371         }
372         return 0;
373 }
374
375 /*
376  * read a page from the cache or allocate a block in which to store it
377  * - we return:
378  *   -ENOMEM    - out of memory, nothing done
379  *   -ERESTARTSYS - interrupted
380  *   -ENOBUFS   - no backing object available in which to cache the block
381  *   -ENODATA   - no data available in the backing object for this block
382  *   0          - dispatched a read - it'll call end_io_func() when finished
383  */
384 int __fscache_read_or_alloc_page(struct fscache_cookie *cookie,
385                                  struct page *page,
386                                  fscache_rw_complete_t end_io_func,
387                                  void *context,
388                                  gfp_t gfp)
389 {
390         struct fscache_retrieval *op;
391         struct fscache_object *object;
392         bool wake_cookie = false;
393         int ret;
394
395         _enter("%p,%p,,,", cookie, page);
396
397         fscache_stat(&fscache_n_retrievals);
398
399         if (hlist_empty(&cookie->backing_objects))
400                 goto nobufs;
401
402         if (test_bit(FSCACHE_COOKIE_INVALIDATING, &cookie->flags)) {
403                 _leave(" = -ENOBUFS [invalidating]");
404                 return -ENOBUFS;
405         }
406
407         ASSERTCMP(cookie->def->type, !=, FSCACHE_COOKIE_TYPE_INDEX);
408         ASSERTCMP(page, !=, NULL);
409
410         if (fscache_wait_for_deferred_lookup(cookie) < 0)
411                 return -ERESTARTSYS;
412
413         op = fscache_alloc_retrieval(cookie, page->mapping,
414                                      end_io_func, context);
415         if (!op) {
416                 _leave(" = -ENOMEM");
417                 return -ENOMEM;
418         }
419         atomic_set(&op->n_pages, 1);
420
421         spin_lock(&cookie->lock);
422
423         if (!fscache_cookie_enabled(cookie) ||
424             hlist_empty(&cookie->backing_objects))
425                 goto nobufs_unlock;
426         object = hlist_entry(cookie->backing_objects.first,
427                              struct fscache_object, cookie_link);
428
429         ASSERT(test_bit(FSCACHE_OBJECT_IS_LOOKED_UP, &object->flags));
430
431         __fscache_use_cookie(cookie);
432         atomic_inc(&object->n_reads);
433         __set_bit(FSCACHE_OP_DEC_READ_CNT, &op->op.flags);
434
435         if (fscache_submit_op(object, &op->op) < 0)
436                 goto nobufs_unlock_dec;
437         spin_unlock(&cookie->lock);
438
439         fscache_stat(&fscache_n_retrieval_ops);
440
441         /* pin the netfs read context in case we need to do the actual netfs
442          * read because we've encountered a cache read failure */
443         fscache_get_context(object->cookie, op->context);
444
445         /* we wait for the operation to become active, and then process it
446          * *here*, in this thread, and not in the thread pool */
447         ret = fscache_wait_for_operation_activation(
448                 object, &op->op,
449                 __fscache_stat(&fscache_n_retrieval_op_waits),
450                 __fscache_stat(&fscache_n_retrievals_object_dead),
451                 fscache_do_cancel_retrieval);
452         if (ret < 0)
453                 goto error;
454
455         /* ask the cache to honour the operation */
456         if (test_bit(FSCACHE_COOKIE_NO_DATA_YET, &object->cookie->flags)) {
457                 fscache_stat(&fscache_n_cop_allocate_page);
458                 ret = object->cache->ops->allocate_page(op, page, gfp);
459                 fscache_stat_d(&fscache_n_cop_allocate_page);
460                 if (ret == 0)
461                         ret = -ENODATA;
462         } else {
463                 fscache_stat(&fscache_n_cop_read_or_alloc_page);
464                 ret = object->cache->ops->read_or_alloc_page(op, page, gfp);
465                 fscache_stat_d(&fscache_n_cop_read_or_alloc_page);
466         }
467
468 error:
469         if (ret == -ENOMEM)
470                 fscache_stat(&fscache_n_retrievals_nomem);
471         else if (ret == -ERESTARTSYS)
472                 fscache_stat(&fscache_n_retrievals_intr);
473         else if (ret == -ENODATA)
474                 fscache_stat(&fscache_n_retrievals_nodata);
475         else if (ret < 0)
476                 fscache_stat(&fscache_n_retrievals_nobufs);
477         else
478                 fscache_stat(&fscache_n_retrievals_ok);
479
480         fscache_put_retrieval(op);
481         _leave(" = %d", ret);
482         return ret;
483
484 nobufs_unlock_dec:
485         atomic_dec(&object->n_reads);
486         wake_cookie = __fscache_unuse_cookie(cookie);
487 nobufs_unlock:
488         spin_unlock(&cookie->lock);
489         if (wake_cookie)
490                 __fscache_wake_unused_cookie(cookie);
491         kfree(op);
492 nobufs:
493         fscache_stat(&fscache_n_retrievals_nobufs);
494         _leave(" = -ENOBUFS");
495         return -ENOBUFS;
496 }
497 EXPORT_SYMBOL(__fscache_read_or_alloc_page);
498
499 /*
500  * read a list of page from the cache or allocate a block in which to store
501  * them
502  * - we return:
503  *   -ENOMEM    - out of memory, some pages may be being read
504  *   -ERESTARTSYS - interrupted, some pages may be being read
505  *   -ENOBUFS   - no backing object or space available in which to cache any
506  *                pages not being read
507  *   -ENODATA   - no data available in the backing object for some or all of
508  *                the pages
509  *   0          - dispatched a read on all pages
510  *
511  * end_io_func() will be called for each page read from the cache as it is
512  * finishes being read
513  *
514  * any pages for which a read is dispatched will be removed from pages and
515  * nr_pages
516  */
517 int __fscache_read_or_alloc_pages(struct fscache_cookie *cookie,
518                                   struct address_space *mapping,
519                                   struct list_head *pages,
520                                   unsigned *nr_pages,
521                                   fscache_rw_complete_t end_io_func,
522                                   void *context,
523                                   gfp_t gfp)
524 {
525         struct fscache_retrieval *op;
526         struct fscache_object *object;
527         bool wake_cookie = false;
528         int ret;
529
530         _enter("%p,,%d,,,", cookie, *nr_pages);
531
532         fscache_stat(&fscache_n_retrievals);
533
534         if (hlist_empty(&cookie->backing_objects))
535                 goto nobufs;
536
537         if (test_bit(FSCACHE_COOKIE_INVALIDATING, &cookie->flags)) {
538                 _leave(" = -ENOBUFS [invalidating]");
539                 return -ENOBUFS;
540         }
541
542         ASSERTCMP(cookie->def->type, !=, FSCACHE_COOKIE_TYPE_INDEX);
543         ASSERTCMP(*nr_pages, >, 0);
544         ASSERT(!list_empty(pages));
545
546         if (fscache_wait_for_deferred_lookup(cookie) < 0)
547                 return -ERESTARTSYS;
548
549         op = fscache_alloc_retrieval(cookie, mapping, end_io_func, context);
550         if (!op)
551                 return -ENOMEM;
552         atomic_set(&op->n_pages, *nr_pages);
553
554         spin_lock(&cookie->lock);
555
556         if (!fscache_cookie_enabled(cookie) ||
557             hlist_empty(&cookie->backing_objects))
558                 goto nobufs_unlock;
559         object = hlist_entry(cookie->backing_objects.first,
560                              struct fscache_object, cookie_link);
561
562         __fscache_use_cookie(cookie);
563         atomic_inc(&object->n_reads);
564         __set_bit(FSCACHE_OP_DEC_READ_CNT, &op->op.flags);
565
566         if (fscache_submit_op(object, &op->op) < 0)
567                 goto nobufs_unlock_dec;
568         spin_unlock(&cookie->lock);
569
570         fscache_stat(&fscache_n_retrieval_ops);
571
572         /* pin the netfs read context in case we need to do the actual netfs
573          * read because we've encountered a cache read failure */
574         fscache_get_context(object->cookie, op->context);
575
576         /* we wait for the operation to become active, and then process it
577          * *here*, in this thread, and not in the thread pool */
578         ret = fscache_wait_for_operation_activation(
579                 object, &op->op,
580                 __fscache_stat(&fscache_n_retrieval_op_waits),
581                 __fscache_stat(&fscache_n_retrievals_object_dead),
582                 fscache_do_cancel_retrieval);
583         if (ret < 0)
584                 goto error;
585
586         /* ask the cache to honour the operation */
587         if (test_bit(FSCACHE_COOKIE_NO_DATA_YET, &object->cookie->flags)) {
588                 fscache_stat(&fscache_n_cop_allocate_pages);
589                 ret = object->cache->ops->allocate_pages(
590                         op, pages, nr_pages, gfp);
591                 fscache_stat_d(&fscache_n_cop_allocate_pages);
592         } else {
593                 fscache_stat(&fscache_n_cop_read_or_alloc_pages);
594                 ret = object->cache->ops->read_or_alloc_pages(
595                         op, pages, nr_pages, gfp);
596                 fscache_stat_d(&fscache_n_cop_read_or_alloc_pages);
597         }
598
599 error:
600         if (ret == -ENOMEM)
601                 fscache_stat(&fscache_n_retrievals_nomem);
602         else if (ret == -ERESTARTSYS)
603                 fscache_stat(&fscache_n_retrievals_intr);
604         else if (ret == -ENODATA)
605                 fscache_stat(&fscache_n_retrievals_nodata);
606         else if (ret < 0)
607                 fscache_stat(&fscache_n_retrievals_nobufs);
608         else
609                 fscache_stat(&fscache_n_retrievals_ok);
610
611         fscache_put_retrieval(op);
612         _leave(" = %d", ret);
613         return ret;
614
615 nobufs_unlock_dec:
616         atomic_dec(&object->n_reads);
617         wake_cookie = __fscache_unuse_cookie(cookie);
618 nobufs_unlock:
619         spin_unlock(&cookie->lock);
620         kfree(op);
621         if (wake_cookie)
622                 __fscache_wake_unused_cookie(cookie);
623 nobufs:
624         fscache_stat(&fscache_n_retrievals_nobufs);
625         _leave(" = -ENOBUFS");
626         return -ENOBUFS;
627 }
628 EXPORT_SYMBOL(__fscache_read_or_alloc_pages);
629
630 /*
631  * allocate a block in the cache on which to store a page
632  * - we return:
633  *   -ENOMEM    - out of memory, nothing done
634  *   -ERESTARTSYS - interrupted
635  *   -ENOBUFS   - no backing object available in which to cache the block
636  *   0          - block allocated
637  */
638 int __fscache_alloc_page(struct fscache_cookie *cookie,
639                          struct page *page,
640                          gfp_t gfp)
641 {
642         struct fscache_retrieval *op;
643         struct fscache_object *object;
644         bool wake_cookie = false;
645         int ret;
646
647         _enter("%p,%p,,,", cookie, page);
648
649         fscache_stat(&fscache_n_allocs);
650
651         if (hlist_empty(&cookie->backing_objects))
652                 goto nobufs;
653
654         ASSERTCMP(cookie->def->type, !=, FSCACHE_COOKIE_TYPE_INDEX);
655         ASSERTCMP(page, !=, NULL);
656
657         if (test_bit(FSCACHE_COOKIE_INVALIDATING, &cookie->flags)) {
658                 _leave(" = -ENOBUFS [invalidating]");
659                 return -ENOBUFS;
660         }
661
662         if (fscache_wait_for_deferred_lookup(cookie) < 0)
663                 return -ERESTARTSYS;
664
665         op = fscache_alloc_retrieval(cookie, page->mapping, NULL, NULL);
666         if (!op)
667                 return -ENOMEM;
668         atomic_set(&op->n_pages, 1);
669
670         spin_lock(&cookie->lock);
671
672         if (!fscache_cookie_enabled(cookie) ||
673             hlist_empty(&cookie->backing_objects))
674                 goto nobufs_unlock;
675         object = hlist_entry(cookie->backing_objects.first,
676                              struct fscache_object, cookie_link);
677
678         __fscache_use_cookie(cookie);
679         if (fscache_submit_op(object, &op->op) < 0)
680                 goto nobufs_unlock_dec;
681         spin_unlock(&cookie->lock);
682
683         fscache_stat(&fscache_n_alloc_ops);
684
685         ret = fscache_wait_for_operation_activation(
686                 object, &op->op,
687                 __fscache_stat(&fscache_n_alloc_op_waits),
688                 __fscache_stat(&fscache_n_allocs_object_dead),
689                 fscache_do_cancel_retrieval);
690         if (ret < 0)
691                 goto error;
692
693         /* ask the cache to honour the operation */
694         fscache_stat(&fscache_n_cop_allocate_page);
695         ret = object->cache->ops->allocate_page(op, page, gfp);
696         fscache_stat_d(&fscache_n_cop_allocate_page);
697
698 error:
699         if (ret == -ERESTARTSYS)
700                 fscache_stat(&fscache_n_allocs_intr);
701         else if (ret < 0)
702                 fscache_stat(&fscache_n_allocs_nobufs);
703         else
704                 fscache_stat(&fscache_n_allocs_ok);
705
706         fscache_put_retrieval(op);
707         _leave(" = %d", ret);
708         return ret;
709
710 nobufs_unlock_dec:
711         wake_cookie = __fscache_unuse_cookie(cookie);
712 nobufs_unlock:
713         spin_unlock(&cookie->lock);
714         kfree(op);
715         if (wake_cookie)
716                 __fscache_wake_unused_cookie(cookie);
717 nobufs:
718         fscache_stat(&fscache_n_allocs_nobufs);
719         _leave(" = -ENOBUFS");
720         return -ENOBUFS;
721 }
722 EXPORT_SYMBOL(__fscache_alloc_page);
723
724 /*
725  * Unmark pages allocate in the readahead code path (via:
726  * fscache_readpages_or_alloc) after delegating to the base filesystem
727  */
728 void __fscache_readpages_cancel(struct fscache_cookie *cookie,
729                                 struct list_head *pages)
730 {
731         struct page *page;
732
733         list_for_each_entry(page, pages, lru) {
734                 if (PageFsCache(page))
735                         __fscache_uncache_page(cookie, page);
736         }
737 }
738 EXPORT_SYMBOL(__fscache_readpages_cancel);
739
740 /*
741  * release a write op reference
742  */
743 static void fscache_release_write_op(struct fscache_operation *_op)
744 {
745         _enter("{OP%x}", _op->debug_id);
746 }
747
748 /*
749  * perform the background storage of a page into the cache
750  */
751 static void fscache_write_op(struct fscache_operation *_op)
752 {
753         struct fscache_storage *op =
754                 container_of(_op, struct fscache_storage, op);
755         struct fscache_object *object = op->op.object;
756         struct fscache_cookie *cookie;
757         struct page *page;
758         unsigned n;
759         void *results[1];
760         int ret;
761
762         _enter("{OP%x,%d}", op->op.debug_id, atomic_read(&op->op.usage));
763
764         spin_lock(&object->lock);
765         cookie = object->cookie;
766
767         if (!fscache_object_is_active(object)) {
768                 /* If we get here, then the on-disk cache object likely longer
769                  * exists, so we should just cancel this write operation.
770                  */
771                 spin_unlock(&object->lock);
772                 fscache_op_complete(&op->op, false);
773                 _leave(" [inactive]");
774                 return;
775         }
776
777         if (!cookie) {
778                 /* If we get here, then the cookie belonging to the object was
779                  * detached, probably by the cookie being withdrawn due to
780                  * memory pressure, which means that the pages we might write
781                  * to the cache from no longer exist - therefore, we can just
782                  * cancel this write operation.
783                  */
784                 spin_unlock(&object->lock);
785                 fscache_op_complete(&op->op, false);
786                 _leave(" [cancel] op{f=%lx s=%u} obj{s=%s f=%lx}",
787                        _op->flags, _op->state, object->state->short_name,
788                        object->flags);
789                 return;
790         }
791
792         spin_lock(&cookie->stores_lock);
793
794         fscache_stat(&fscache_n_store_calls);
795
796         /* find a page to store */
797         page = NULL;
798         n = radix_tree_gang_lookup_tag(&cookie->stores, results, 0, 1,
799                                        FSCACHE_COOKIE_PENDING_TAG);
800         if (n != 1)
801                 goto superseded;
802         page = results[0];
803         _debug("gang %d [%lx]", n, page->index);
804         if (page->index > op->store_limit) {
805                 fscache_stat(&fscache_n_store_pages_over_limit);
806                 goto superseded;
807         }
808
809         radix_tree_tag_set(&cookie->stores, page->index,
810                            FSCACHE_COOKIE_STORING_TAG);
811         radix_tree_tag_clear(&cookie->stores, page->index,
812                              FSCACHE_COOKIE_PENDING_TAG);
813
814         spin_unlock(&cookie->stores_lock);
815         spin_unlock(&object->lock);
816
817         fscache_stat(&fscache_n_store_pages);
818         fscache_stat(&fscache_n_cop_write_page);
819         ret = object->cache->ops->write_page(op, page);
820         fscache_stat_d(&fscache_n_cop_write_page);
821         fscache_end_page_write(object, page);
822         if (ret < 0) {
823                 fscache_abort_object(object);
824                 fscache_op_complete(&op->op, true);
825         } else {
826                 fscache_enqueue_operation(&op->op);
827         }
828
829         _leave("");
830         return;
831
832 superseded:
833         /* this writer is going away and there aren't any more things to
834          * write */
835         _debug("cease");
836         spin_unlock(&cookie->stores_lock);
837         clear_bit(FSCACHE_OBJECT_PENDING_WRITE, &object->flags);
838         spin_unlock(&object->lock);
839         fscache_op_complete(&op->op, true);
840         _leave("");
841 }
842
843 /*
844  * Clear the pages pending writing for invalidation
845  */
846 void fscache_invalidate_writes(struct fscache_cookie *cookie)
847 {
848         struct page *page;
849         void *results[16];
850         int n, i;
851
852         _enter("");
853
854         for (;;) {
855                 spin_lock(&cookie->stores_lock);
856                 n = radix_tree_gang_lookup_tag(&cookie->stores, results, 0,
857                                                ARRAY_SIZE(results),
858                                                FSCACHE_COOKIE_PENDING_TAG);
859                 if (n == 0) {
860                         spin_unlock(&cookie->stores_lock);
861                         break;
862                 }
863
864                 for (i = n - 1; i >= 0; i--) {
865                         page = results[i];
866                         radix_tree_delete(&cookie->stores, page->index);
867                 }
868
869                 spin_unlock(&cookie->stores_lock);
870
871                 for (i = n - 1; i >= 0; i--)
872                         page_cache_release(results[i]);
873         }
874
875         _leave("");
876 }
877
878 /*
879  * request a page be stored in the cache
880  * - returns:
881  *   -ENOMEM    - out of memory, nothing done
882  *   -ENOBUFS   - no backing object available in which to cache the page
883  *   0          - dispatched a write - it'll call end_io_func() when finished
884  *
885  * if the cookie still has a backing object at this point, that object can be
886  * in one of a few states with respect to storage processing:
887  *
888  *  (1) negative lookup, object not yet created (FSCACHE_COOKIE_CREATING is
889  *      set)
890  *
891  *      (a) no writes yet
892  *
893  *      (b) writes deferred till post-creation (mark page for writing and
894  *          return immediately)
895  *
896  *  (2) negative lookup, object created, initial fill being made from netfs
897  *
898  *      (a) fill point not yet reached this page (mark page for writing and
899  *          return)
900  *
901  *      (b) fill point passed this page (queue op to store this page)
902  *
903  *  (3) object extant (queue op to store this page)
904  *
905  * any other state is invalid
906  */
907 int __fscache_write_page(struct fscache_cookie *cookie,
908                          struct page *page,
909                          gfp_t gfp)
910 {
911         struct fscache_storage *op;
912         struct fscache_object *object;
913         bool wake_cookie = false;
914         int ret;
915
916         _enter("%p,%x,", cookie, (u32) page->flags);
917
918         ASSERTCMP(cookie->def->type, !=, FSCACHE_COOKIE_TYPE_INDEX);
919         ASSERT(PageFsCache(page));
920
921         fscache_stat(&fscache_n_stores);
922
923         if (test_bit(FSCACHE_COOKIE_INVALIDATING, &cookie->flags)) {
924                 _leave(" = -ENOBUFS [invalidating]");
925                 return -ENOBUFS;
926         }
927
928         op = kzalloc(sizeof(*op), GFP_NOIO | __GFP_NOMEMALLOC | __GFP_NORETRY);
929         if (!op)
930                 goto nomem;
931
932         fscache_operation_init(&op->op, fscache_write_op,
933                                fscache_release_write_op);
934         op->op.flags = FSCACHE_OP_ASYNC |
935                 (1 << FSCACHE_OP_WAITING) |
936                 (1 << FSCACHE_OP_UNUSE_COOKIE);
937
938         ret = radix_tree_maybe_preload(gfp & ~__GFP_HIGHMEM);
939         if (ret < 0)
940                 goto nomem_free;
941
942         ret = -ENOBUFS;
943         spin_lock(&cookie->lock);
944
945         if (!fscache_cookie_enabled(cookie) ||
946             hlist_empty(&cookie->backing_objects))
947                 goto nobufs;
948         object = hlist_entry(cookie->backing_objects.first,
949                              struct fscache_object, cookie_link);
950         if (test_bit(FSCACHE_IOERROR, &object->cache->flags))
951                 goto nobufs;
952
953         /* add the page to the pending-storage radix tree on the backing
954          * object */
955         spin_lock(&object->lock);
956         spin_lock(&cookie->stores_lock);
957
958         _debug("store limit %llx", (unsigned long long) object->store_limit);
959
960         ret = radix_tree_insert(&cookie->stores, page->index, page);
961         if (ret < 0) {
962                 if (ret == -EEXIST)
963                         goto already_queued;
964                 _debug("insert failed %d", ret);
965                 goto nobufs_unlock_obj;
966         }
967
968         radix_tree_tag_set(&cookie->stores, page->index,
969                            FSCACHE_COOKIE_PENDING_TAG);
970         page_cache_get(page);
971
972         /* we only want one writer at a time, but we do need to queue new
973          * writers after exclusive ops */
974         if (test_and_set_bit(FSCACHE_OBJECT_PENDING_WRITE, &object->flags))
975                 goto already_pending;
976
977         spin_unlock(&cookie->stores_lock);
978         spin_unlock(&object->lock);
979
980         op->op.debug_id = atomic_inc_return(&fscache_op_debug_id);
981         op->store_limit = object->store_limit;
982
983         __fscache_use_cookie(cookie);
984         if (fscache_submit_op(object, &op->op) < 0)
985                 goto submit_failed;
986
987         spin_unlock(&cookie->lock);
988         radix_tree_preload_end();
989         fscache_stat(&fscache_n_store_ops);
990         fscache_stat(&fscache_n_stores_ok);
991
992         /* the work queue now carries its own ref on the object */
993         fscache_put_operation(&op->op);
994         _leave(" = 0");
995         return 0;
996
997 already_queued:
998         fscache_stat(&fscache_n_stores_again);
999 already_pending:
1000         spin_unlock(&cookie->stores_lock);
1001         spin_unlock(&object->lock);
1002         spin_unlock(&cookie->lock);
1003         radix_tree_preload_end();
1004         kfree(op);
1005         fscache_stat(&fscache_n_stores_ok);
1006         _leave(" = 0");
1007         return 0;
1008
1009 submit_failed:
1010         spin_lock(&cookie->stores_lock);
1011         radix_tree_delete(&cookie->stores, page->index);
1012         spin_unlock(&cookie->stores_lock);
1013         wake_cookie = __fscache_unuse_cookie(cookie);
1014         page_cache_release(page);
1015         ret = -ENOBUFS;
1016         goto nobufs;
1017
1018 nobufs_unlock_obj:
1019         spin_unlock(&cookie->stores_lock);
1020         spin_unlock(&object->lock);
1021 nobufs:
1022         spin_unlock(&cookie->lock);
1023         radix_tree_preload_end();
1024         kfree(op);
1025         if (wake_cookie)
1026                 __fscache_wake_unused_cookie(cookie);
1027         fscache_stat(&fscache_n_stores_nobufs);
1028         _leave(" = -ENOBUFS");
1029         return -ENOBUFS;
1030
1031 nomem_free:
1032         kfree(op);
1033 nomem:
1034         fscache_stat(&fscache_n_stores_oom);
1035         _leave(" = -ENOMEM");
1036         return -ENOMEM;
1037 }
1038 EXPORT_SYMBOL(__fscache_write_page);
1039
1040 /*
1041  * remove a page from the cache
1042  */
1043 void __fscache_uncache_page(struct fscache_cookie *cookie, struct page *page)
1044 {
1045         struct fscache_object *object;
1046
1047         _enter(",%p", page);
1048
1049         ASSERTCMP(cookie->def->type, !=, FSCACHE_COOKIE_TYPE_INDEX);
1050         ASSERTCMP(page, !=, NULL);
1051
1052         fscache_stat(&fscache_n_uncaches);
1053
1054         /* cache withdrawal may beat us to it */
1055         if (!PageFsCache(page))
1056                 goto done;
1057
1058         /* get the object */
1059         spin_lock(&cookie->lock);
1060
1061         if (hlist_empty(&cookie->backing_objects)) {
1062                 ClearPageFsCache(page);
1063                 goto done_unlock;
1064         }
1065
1066         object = hlist_entry(cookie->backing_objects.first,
1067                              struct fscache_object, cookie_link);
1068
1069         /* there might now be stuff on disk we could read */
1070         clear_bit(FSCACHE_COOKIE_NO_DATA_YET, &cookie->flags);
1071
1072         /* only invoke the cache backend if we managed to mark the page
1073          * uncached here; this deals with synchronisation vs withdrawal */
1074         if (TestClearPageFsCache(page) &&
1075             object->cache->ops->uncache_page) {
1076                 /* the cache backend releases the cookie lock */
1077                 fscache_stat(&fscache_n_cop_uncache_page);
1078                 object->cache->ops->uncache_page(object, page);
1079                 fscache_stat_d(&fscache_n_cop_uncache_page);
1080                 goto done;
1081         }
1082
1083 done_unlock:
1084         spin_unlock(&cookie->lock);
1085 done:
1086         _leave("");
1087 }
1088 EXPORT_SYMBOL(__fscache_uncache_page);
1089
1090 /**
1091  * fscache_mark_page_cached - Mark a page as being cached
1092  * @op: The retrieval op pages are being marked for
1093  * @page: The page to be marked
1094  *
1095  * Mark a netfs page as being cached.  After this is called, the netfs
1096  * must call fscache_uncache_page() to remove the mark.
1097  */
1098 void fscache_mark_page_cached(struct fscache_retrieval *op, struct page *page)
1099 {
1100         struct fscache_cookie *cookie = op->op.object->cookie;
1101
1102 #ifdef CONFIG_FSCACHE_STATS
1103         atomic_inc(&fscache_n_marks);
1104 #endif
1105
1106         _debug("- mark %p{%lx}", page, page->index);
1107         if (TestSetPageFsCache(page)) {
1108                 static bool once_only;
1109                 if (!once_only) {
1110                         once_only = true;
1111                         printk(KERN_WARNING "FS-Cache:"
1112                                " Cookie type %s marked page %lx"
1113                                " multiple times\n",
1114                                cookie->def->name, page->index);
1115                 }
1116         }
1117
1118         if (cookie->def->mark_page_cached)
1119                 cookie->def->mark_page_cached(cookie->netfs_data,
1120                                               op->mapping, page);
1121 }
1122 EXPORT_SYMBOL(fscache_mark_page_cached);
1123
1124 /**
1125  * fscache_mark_pages_cached - Mark pages as being cached
1126  * @op: The retrieval op pages are being marked for
1127  * @pagevec: The pages to be marked
1128  *
1129  * Mark a bunch of netfs pages as being cached.  After this is called,
1130  * the netfs must call fscache_uncache_page() to remove the mark.
1131  */
1132 void fscache_mark_pages_cached(struct fscache_retrieval *op,
1133                                struct pagevec *pagevec)
1134 {
1135         unsigned long loop;
1136
1137         for (loop = 0; loop < pagevec->nr; loop++)
1138                 fscache_mark_page_cached(op, pagevec->pages[loop]);
1139
1140         pagevec_reinit(pagevec);
1141 }
1142 EXPORT_SYMBOL(fscache_mark_pages_cached);
1143
1144 /*
1145  * Uncache all the pages in an inode that are marked PG_fscache, assuming them
1146  * to be associated with the given cookie.
1147  */
1148 void __fscache_uncache_all_inode_pages(struct fscache_cookie *cookie,
1149                                        struct inode *inode)
1150 {
1151         struct address_space *mapping = inode->i_mapping;
1152         struct pagevec pvec;
1153         pgoff_t next;
1154         int i;
1155
1156         _enter("%p,%p", cookie, inode);
1157
1158         if (!mapping || mapping->nrpages == 0) {
1159                 _leave(" [no pages]");
1160                 return;
1161         }
1162
1163         pagevec_init(&pvec, 0);
1164         next = 0;
1165         do {
1166                 if (!pagevec_lookup(&pvec, mapping, next, PAGEVEC_SIZE))
1167                         break;
1168                 for (i = 0; i < pagevec_count(&pvec); i++) {
1169                         struct page *page = pvec.pages[i];
1170                         next = page->index;
1171                         if (PageFsCache(page)) {
1172                                 __fscache_wait_on_page_write(cookie, page);
1173                                 __fscache_uncache_page(cookie, page);
1174                         }
1175                 }
1176                 pagevec_release(&pvec);
1177                 cond_resched();
1178         } while (++next);
1179
1180         _leave("");
1181 }
1182 EXPORT_SYMBOL(__fscache_uncache_all_inode_pages);