ARM64: dts: rk3328: add pdm node
[firefly-linux-kernel-4.4.55.git] / mm / filemap.c
1 /*
2  *      linux/mm/filemap.c
3  *
4  * Copyright (C) 1994-1999  Linus Torvalds
5  */
6
7 /*
8  * This file handles the generic file mmap semantics used by
9  * most "normal" filesystems (but you don't /have/ to use this:
10  * the NFS filesystem used to do this differently, for example)
11  */
12 #include <linux/export.h>
13 #include <linux/compiler.h>
14 #include <linux/fs.h>
15 #include <linux/uaccess.h>
16 #include <linux/capability.h>
17 #include <linux/kernel_stat.h>
18 #include <linux/gfp.h>
19 #include <linux/mm.h>
20 #include <linux/swap.h>
21 #include <linux/mman.h>
22 #include <linux/pagemap.h>
23 #include <linux/file.h>
24 #include <linux/uio.h>
25 #include <linux/hash.h>
26 #include <linux/writeback.h>
27 #include <linux/backing-dev.h>
28 #include <linux/pagevec.h>
29 #include <linux/blkdev.h>
30 #include <linux/security.h>
31 #include <linux/cpuset.h>
32 #include <linux/hardirq.h> /* for BUG_ON(!in_atomic()) only */
33 #include <linux/hugetlb.h>
34 #include <linux/memcontrol.h>
35 #include <linux/cleancache.h>
36 #include <linux/rmap.h>
37 #include "internal.h"
38
39 #define CREATE_TRACE_POINTS
40 #include <trace/events/filemap.h>
41
42 /*
43  * FIXME: remove all knowledge of the buffer layer from the core VM
44  */
45 #include <linux/buffer_head.h> /* for try_to_free_buffers */
46
47 #include <asm/mman.h>
48
49 /*
50  * Shared mappings implemented 30.11.1994. It's not fully working yet,
51  * though.
52  *
53  * Shared mappings now work. 15.8.1995  Bruno.
54  *
55  * finished 'unifying' the page and buffer cache and SMP-threaded the
56  * page-cache, 21.05.1999, Ingo Molnar <mingo@redhat.com>
57  *
58  * SMP-threaded pagemap-LRU 1999, Andrea Arcangeli <andrea@suse.de>
59  */
60
61 /*
62  * Lock ordering:
63  *
64  *  ->i_mmap_rwsem              (truncate_pagecache)
65  *    ->private_lock            (__free_pte->__set_page_dirty_buffers)
66  *      ->swap_lock             (exclusive_swap_page, others)
67  *        ->mapping->tree_lock
68  *
69  *  ->i_mutex
70  *    ->i_mmap_rwsem            (truncate->unmap_mapping_range)
71  *
72  *  ->mmap_sem
73  *    ->i_mmap_rwsem
74  *      ->page_table_lock or pte_lock   (various, mainly in memory.c)
75  *        ->mapping->tree_lock  (arch-dependent flush_dcache_mmap_lock)
76  *
77  *  ->mmap_sem
78  *    ->lock_page               (access_process_vm)
79  *
80  *  ->i_mutex                   (generic_perform_write)
81  *    ->mmap_sem                (fault_in_pages_readable->do_page_fault)
82  *
83  *  bdi->wb.list_lock
84  *    sb_lock                   (fs/fs-writeback.c)
85  *    ->mapping->tree_lock      (__sync_single_inode)
86  *
87  *  ->i_mmap_rwsem
88  *    ->anon_vma.lock           (vma_adjust)
89  *
90  *  ->anon_vma.lock
91  *    ->page_table_lock or pte_lock     (anon_vma_prepare and various)
92  *
93  *  ->page_table_lock or pte_lock
94  *    ->swap_lock               (try_to_unmap_one)
95  *    ->private_lock            (try_to_unmap_one)
96  *    ->tree_lock               (try_to_unmap_one)
97  *    ->zone.lru_lock           (follow_page->mark_page_accessed)
98  *    ->zone.lru_lock           (check_pte_range->isolate_lru_page)
99  *    ->private_lock            (page_remove_rmap->set_page_dirty)
100  *    ->tree_lock               (page_remove_rmap->set_page_dirty)
101  *    bdi.wb->list_lock         (page_remove_rmap->set_page_dirty)
102  *    ->inode->i_lock           (page_remove_rmap->set_page_dirty)
103  *    ->memcg->move_lock        (page_remove_rmap->mem_cgroup_begin_page_stat)
104  *    bdi.wb->list_lock         (zap_pte_range->set_page_dirty)
105  *    ->inode->i_lock           (zap_pte_range->set_page_dirty)
106  *    ->private_lock            (zap_pte_range->__set_page_dirty_buffers)
107  *
108  * ->i_mmap_rwsem
109  *   ->tasklist_lock            (memory_failure, collect_procs_ao)
110  */
111
112 static int page_cache_tree_insert(struct address_space *mapping,
113                                   struct page *page, void **shadowp)
114 {
115         struct radix_tree_node *node;
116         void **slot;
117         int error;
118
119         error = __radix_tree_create(&mapping->page_tree, page->index,
120                                     &node, &slot);
121         if (error)
122                 return error;
123         if (*slot) {
124                 void *p;
125
126                 p = radix_tree_deref_slot_protected(slot, &mapping->tree_lock);
127                 if (!radix_tree_exceptional_entry(p))
128                         return -EEXIST;
129                 if (shadowp)
130                         *shadowp = p;
131                 mapping->nrshadows--;
132                 if (node)
133                         workingset_node_shadows_dec(node);
134         }
135         radix_tree_replace_slot(slot, page);
136         mapping->nrpages++;
137         if (node) {
138                 workingset_node_pages_inc(node);
139                 /*
140                  * Don't track node that contains actual pages.
141                  *
142                  * Avoid acquiring the list_lru lock if already
143                  * untracked.  The list_empty() test is safe as
144                  * node->private_list is protected by
145                  * mapping->tree_lock.
146                  */
147                 if (!list_empty(&node->private_list))
148                         list_lru_del(&workingset_shadow_nodes,
149                                      &node->private_list);
150         }
151         return 0;
152 }
153
154 static void page_cache_tree_delete(struct address_space *mapping,
155                                    struct page *page, void *shadow)
156 {
157         struct radix_tree_node *node;
158         unsigned long index;
159         unsigned int offset;
160         unsigned int tag;
161         void **slot;
162
163         VM_BUG_ON(!PageLocked(page));
164
165         __radix_tree_lookup(&mapping->page_tree, page->index, &node, &slot);
166
167         if (!node) {
168                 /*
169                  * We need a node to properly account shadow
170                  * entries. Don't plant any without. XXX
171                  */
172                 shadow = NULL;
173         }
174
175         if (shadow) {
176                 mapping->nrshadows++;
177                 /*
178                  * Make sure the nrshadows update is committed before
179                  * the nrpages update so that final truncate racing
180                  * with reclaim does not see both counters 0 at the
181                  * same time and miss a shadow entry.
182                  */
183                 smp_wmb();
184         }
185         mapping->nrpages--;
186
187         if (!node) {
188                 /* Clear direct pointer tags in root node */
189                 mapping->page_tree.gfp_mask &= __GFP_BITS_MASK;
190                 radix_tree_replace_slot(slot, shadow);
191                 return;
192         }
193
194         /* Clear tree tags for the removed page */
195         index = page->index;
196         offset = index & RADIX_TREE_MAP_MASK;
197         for (tag = 0; tag < RADIX_TREE_MAX_TAGS; tag++) {
198                 if (test_bit(offset, node->tags[tag]))
199                         radix_tree_tag_clear(&mapping->page_tree, index, tag);
200         }
201
202         /* Delete page, swap shadow entry */
203         radix_tree_replace_slot(slot, shadow);
204         workingset_node_pages_dec(node);
205         if (shadow)
206                 workingset_node_shadows_inc(node);
207         else
208                 if (__radix_tree_delete_node(&mapping->page_tree, node))
209                         return;
210
211         /*
212          * Track node that only contains shadow entries.
213          *
214          * Avoid acquiring the list_lru lock if already tracked.  The
215          * list_empty() test is safe as node->private_list is
216          * protected by mapping->tree_lock.
217          */
218         if (!workingset_node_pages(node) &&
219             list_empty(&node->private_list)) {
220                 node->private_data = mapping;
221                 list_lru_add(&workingset_shadow_nodes, &node->private_list);
222         }
223 }
224
225 /*
226  * Delete a page from the page cache and free it. Caller has to make
227  * sure the page is locked and that nobody else uses it - or that usage
228  * is safe.  The caller must hold the mapping's tree_lock and
229  * mem_cgroup_begin_page_stat().
230  */
231 void __delete_from_page_cache(struct page *page, void *shadow,
232                               struct mem_cgroup *memcg)
233 {
234         struct address_space *mapping = page->mapping;
235
236         trace_mm_filemap_delete_from_page_cache(page);
237         /*
238          * if we're uptodate, flush out into the cleancache, otherwise
239          * invalidate any existing cleancache entries.  We can't leave
240          * stale data around in the cleancache once our page is gone
241          */
242         if (PageUptodate(page) && PageMappedToDisk(page))
243                 cleancache_put_page(page);
244         else
245                 cleancache_invalidate_page(mapping, page);
246
247         page_cache_tree_delete(mapping, page, shadow);
248
249         page->mapping = NULL;
250         /* Leave page->index set: truncation lookup relies upon it */
251
252         /* hugetlb pages do not participate in page cache accounting. */
253         if (!PageHuge(page))
254                 __dec_zone_page_state(page, NR_FILE_PAGES);
255         if (PageSwapBacked(page))
256                 __dec_zone_page_state(page, NR_SHMEM);
257         BUG_ON(page_mapped(page));
258
259         /*
260          * At this point page must be either written or cleaned by truncate.
261          * Dirty page here signals a bug and loss of unwritten data.
262          *
263          * This fixes dirty accounting after removing the page entirely but
264          * leaves PageDirty set: it has no effect for truncated page and
265          * anyway will be cleared before returning page into buddy allocator.
266          */
267         if (WARN_ON_ONCE(PageDirty(page)))
268                 account_page_cleaned(page, mapping, memcg,
269                                      inode_to_wb(mapping->host));
270 }
271
272 /**
273  * delete_from_page_cache - delete page from page cache
274  * @page: the page which the kernel is trying to remove from page cache
275  *
276  * This must be called only on pages that have been verified to be in the page
277  * cache and locked.  It will never put the page into the free list, the caller
278  * has a reference on the page.
279  */
280 void delete_from_page_cache(struct page *page)
281 {
282         struct address_space *mapping = page->mapping;
283         struct mem_cgroup *memcg;
284         unsigned long flags;
285
286         void (*freepage)(struct page *);
287
288         BUG_ON(!PageLocked(page));
289
290         freepage = mapping->a_ops->freepage;
291
292         memcg = mem_cgroup_begin_page_stat(page);
293         spin_lock_irqsave(&mapping->tree_lock, flags);
294         __delete_from_page_cache(page, NULL, memcg);
295         spin_unlock_irqrestore(&mapping->tree_lock, flags);
296         mem_cgroup_end_page_stat(memcg);
297
298         if (freepage)
299                 freepage(page);
300         page_cache_release(page);
301 }
302 EXPORT_SYMBOL(delete_from_page_cache);
303
304 static int filemap_check_errors(struct address_space *mapping)
305 {
306         int ret = 0;
307         /* Check for outstanding write errors */
308         if (test_bit(AS_ENOSPC, &mapping->flags) &&
309             test_and_clear_bit(AS_ENOSPC, &mapping->flags))
310                 ret = -ENOSPC;
311         if (test_bit(AS_EIO, &mapping->flags) &&
312             test_and_clear_bit(AS_EIO, &mapping->flags))
313                 ret = -EIO;
314         return ret;
315 }
316
317 /**
318  * __filemap_fdatawrite_range - start writeback on mapping dirty pages in range
319  * @mapping:    address space structure to write
320  * @start:      offset in bytes where the range starts
321  * @end:        offset in bytes where the range ends (inclusive)
322  * @sync_mode:  enable synchronous operation
323  *
324  * Start writeback against all of a mapping's dirty pages that lie
325  * within the byte offsets <start, end> inclusive.
326  *
327  * If sync_mode is WB_SYNC_ALL then this is a "data integrity" operation, as
328  * opposed to a regular memory cleansing writeback.  The difference between
329  * these two operations is that if a dirty page/buffer is encountered, it must
330  * be waited upon, and not just skipped over.
331  */
332 int __filemap_fdatawrite_range(struct address_space *mapping, loff_t start,
333                                 loff_t end, int sync_mode)
334 {
335         int ret;
336         struct writeback_control wbc = {
337                 .sync_mode = sync_mode,
338                 .nr_to_write = LONG_MAX,
339                 .range_start = start,
340                 .range_end = end,
341         };
342
343         if (!mapping_cap_writeback_dirty(mapping))
344                 return 0;
345
346         wbc_attach_fdatawrite_inode(&wbc, mapping->host);
347         ret = do_writepages(mapping, &wbc);
348         wbc_detach_inode(&wbc);
349         return ret;
350 }
351
352 static inline int __filemap_fdatawrite(struct address_space *mapping,
353         int sync_mode)
354 {
355         return __filemap_fdatawrite_range(mapping, 0, LLONG_MAX, sync_mode);
356 }
357
358 int filemap_fdatawrite(struct address_space *mapping)
359 {
360         return __filemap_fdatawrite(mapping, WB_SYNC_ALL);
361 }
362 EXPORT_SYMBOL(filemap_fdatawrite);
363
364 int filemap_fdatawrite_range(struct address_space *mapping, loff_t start,
365                                 loff_t end)
366 {
367         return __filemap_fdatawrite_range(mapping, start, end, WB_SYNC_ALL);
368 }
369 EXPORT_SYMBOL(filemap_fdatawrite_range);
370
371 /**
372  * filemap_flush - mostly a non-blocking flush
373  * @mapping:    target address_space
374  *
375  * This is a mostly non-blocking flush.  Not suitable for data-integrity
376  * purposes - I/O may not be started against all dirty pages.
377  */
378 int filemap_flush(struct address_space *mapping)
379 {
380         return __filemap_fdatawrite(mapping, WB_SYNC_NONE);
381 }
382 EXPORT_SYMBOL(filemap_flush);
383
384 static int __filemap_fdatawait_range(struct address_space *mapping,
385                                      loff_t start_byte, loff_t end_byte)
386 {
387         pgoff_t index = start_byte >> PAGE_CACHE_SHIFT;
388         pgoff_t end = end_byte >> PAGE_CACHE_SHIFT;
389         struct pagevec pvec;
390         int nr_pages;
391         int ret = 0;
392
393         if (end_byte < start_byte)
394                 goto out;
395
396         pagevec_init(&pvec, 0);
397         while ((index <= end) &&
398                         (nr_pages = pagevec_lookup_tag(&pvec, mapping, &index,
399                         PAGECACHE_TAG_WRITEBACK,
400                         min(end - index, (pgoff_t)PAGEVEC_SIZE-1) + 1)) != 0) {
401                 unsigned i;
402
403                 for (i = 0; i < nr_pages; i++) {
404                         struct page *page = pvec.pages[i];
405
406                         /* until radix tree lookup accepts end_index */
407                         if (page->index > end)
408                                 continue;
409
410                         wait_on_page_writeback(page);
411                         if (TestClearPageError(page))
412                                 ret = -EIO;
413                 }
414                 pagevec_release(&pvec);
415                 cond_resched();
416         }
417 out:
418         return ret;
419 }
420
421 /**
422  * filemap_fdatawait_range - wait for writeback to complete
423  * @mapping:            address space structure to wait for
424  * @start_byte:         offset in bytes where the range starts
425  * @end_byte:           offset in bytes where the range ends (inclusive)
426  *
427  * Walk the list of under-writeback pages of the given address space
428  * in the given range and wait for all of them.  Check error status of
429  * the address space and return it.
430  *
431  * Since the error status of the address space is cleared by this function,
432  * callers are responsible for checking the return value and handling and/or
433  * reporting the error.
434  */
435 int filemap_fdatawait_range(struct address_space *mapping, loff_t start_byte,
436                             loff_t end_byte)
437 {
438         int ret, ret2;
439
440         ret = __filemap_fdatawait_range(mapping, start_byte, end_byte);
441         ret2 = filemap_check_errors(mapping);
442         if (!ret)
443                 ret = ret2;
444
445         return ret;
446 }
447 EXPORT_SYMBOL(filemap_fdatawait_range);
448
449 /**
450  * filemap_fdatawait_keep_errors - wait for writeback without clearing errors
451  * @mapping: address space structure to wait for
452  *
453  * Walk the list of under-writeback pages of the given address space
454  * and wait for all of them.  Unlike filemap_fdatawait(), this function
455  * does not clear error status of the address space.
456  *
457  * Use this function if callers don't handle errors themselves.  Expected
458  * call sites are system-wide / filesystem-wide data flushers: e.g. sync(2),
459  * fsfreeze(8)
460  */
461 void filemap_fdatawait_keep_errors(struct address_space *mapping)
462 {
463         loff_t i_size = i_size_read(mapping->host);
464
465         if (i_size == 0)
466                 return;
467
468         __filemap_fdatawait_range(mapping, 0, i_size - 1);
469 }
470
471 /**
472  * filemap_fdatawait - wait for all under-writeback pages to complete
473  * @mapping: address space structure to wait for
474  *
475  * Walk the list of under-writeback pages of the given address space
476  * and wait for all of them.  Check error status of the address space
477  * and return it.
478  *
479  * Since the error status of the address space is cleared by this function,
480  * callers are responsible for checking the return value and handling and/or
481  * reporting the error.
482  */
483 int filemap_fdatawait(struct address_space *mapping)
484 {
485         loff_t i_size = i_size_read(mapping->host);
486
487         if (i_size == 0)
488                 return 0;
489
490         return filemap_fdatawait_range(mapping, 0, i_size - 1);
491 }
492 EXPORT_SYMBOL(filemap_fdatawait);
493
494 int filemap_write_and_wait(struct address_space *mapping)
495 {
496         int err = 0;
497
498         if (mapping->nrpages) {
499                 err = filemap_fdatawrite(mapping);
500                 /*
501                  * Even if the above returned error, the pages may be
502                  * written partially (e.g. -ENOSPC), so we wait for it.
503                  * But the -EIO is special case, it may indicate the worst
504                  * thing (e.g. bug) happened, so we avoid waiting for it.
505                  */
506                 if (err != -EIO) {
507                         int err2 = filemap_fdatawait(mapping);
508                         if (!err)
509                                 err = err2;
510                 }
511         } else {
512                 err = filemap_check_errors(mapping);
513         }
514         return err;
515 }
516 EXPORT_SYMBOL(filemap_write_and_wait);
517
518 /**
519  * filemap_write_and_wait_range - write out & wait on a file range
520  * @mapping:    the address_space for the pages
521  * @lstart:     offset in bytes where the range starts
522  * @lend:       offset in bytes where the range ends (inclusive)
523  *
524  * Write out and wait upon file offsets lstart->lend, inclusive.
525  *
526  * Note that `lend' is inclusive (describes the last byte to be written) so
527  * that this function can be used to write to the very end-of-file (end = -1).
528  */
529 int filemap_write_and_wait_range(struct address_space *mapping,
530                                  loff_t lstart, loff_t lend)
531 {
532         int err = 0;
533
534         if (mapping->nrpages) {
535                 err = __filemap_fdatawrite_range(mapping, lstart, lend,
536                                                  WB_SYNC_ALL);
537                 /* See comment of filemap_write_and_wait() */
538                 if (err != -EIO) {
539                         int err2 = filemap_fdatawait_range(mapping,
540                                                 lstart, lend);
541                         if (!err)
542                                 err = err2;
543                 }
544         } else {
545                 err = filemap_check_errors(mapping);
546         }
547         return err;
548 }
549 EXPORT_SYMBOL(filemap_write_and_wait_range);
550
551 /**
552  * replace_page_cache_page - replace a pagecache page with a new one
553  * @old:        page to be replaced
554  * @new:        page to replace with
555  * @gfp_mask:   allocation mode
556  *
557  * This function replaces a page in the pagecache with a new one.  On
558  * success it acquires the pagecache reference for the new page and
559  * drops it for the old page.  Both the old and new pages must be
560  * locked.  This function does not add the new page to the LRU, the
561  * caller must do that.
562  *
563  * The remove + add is atomic.  The only way this function can fail is
564  * memory allocation failure.
565  */
566 int replace_page_cache_page(struct page *old, struct page *new, gfp_t gfp_mask)
567 {
568         int error;
569
570         VM_BUG_ON_PAGE(!PageLocked(old), old);
571         VM_BUG_ON_PAGE(!PageLocked(new), new);
572         VM_BUG_ON_PAGE(new->mapping, new);
573
574         error = radix_tree_preload(gfp_mask & ~__GFP_HIGHMEM);
575         if (!error) {
576                 struct address_space *mapping = old->mapping;
577                 void (*freepage)(struct page *);
578                 struct mem_cgroup *memcg;
579                 unsigned long flags;
580
581                 pgoff_t offset = old->index;
582                 freepage = mapping->a_ops->freepage;
583
584                 page_cache_get(new);
585                 new->mapping = mapping;
586                 new->index = offset;
587
588                 memcg = mem_cgroup_begin_page_stat(old);
589                 spin_lock_irqsave(&mapping->tree_lock, flags);
590                 __delete_from_page_cache(old, NULL, memcg);
591                 error = page_cache_tree_insert(mapping, new, NULL);
592                 BUG_ON(error);
593
594                 /*
595                  * hugetlb pages do not participate in page cache accounting.
596                  */
597                 if (!PageHuge(new))
598                         __inc_zone_page_state(new, NR_FILE_PAGES);
599                 if (PageSwapBacked(new))
600                         __inc_zone_page_state(new, NR_SHMEM);
601                 spin_unlock_irqrestore(&mapping->tree_lock, flags);
602                 mem_cgroup_end_page_stat(memcg);
603                 mem_cgroup_replace_page(old, new);
604                 radix_tree_preload_end();
605                 if (freepage)
606                         freepage(old);
607                 page_cache_release(old);
608         }
609
610         return error;
611 }
612 EXPORT_SYMBOL_GPL(replace_page_cache_page);
613
614 static int __add_to_page_cache_locked(struct page *page,
615                                       struct address_space *mapping,
616                                       pgoff_t offset, gfp_t gfp_mask,
617                                       void **shadowp)
618 {
619         int huge = PageHuge(page);
620         struct mem_cgroup *memcg;
621         int error;
622
623         VM_BUG_ON_PAGE(!PageLocked(page), page);
624         VM_BUG_ON_PAGE(PageSwapBacked(page), page);
625
626         if (!huge) {
627                 error = mem_cgroup_try_charge(page, current->mm,
628                                               gfp_mask, &memcg);
629                 if (error)
630                         return error;
631         }
632
633         error = radix_tree_maybe_preload(gfp_mask & ~__GFP_HIGHMEM);
634         if (error) {
635                 if (!huge)
636                         mem_cgroup_cancel_charge(page, memcg);
637                 return error;
638         }
639
640         page_cache_get(page);
641         page->mapping = mapping;
642         page->index = offset;
643
644         spin_lock_irq(&mapping->tree_lock);
645         error = page_cache_tree_insert(mapping, page, shadowp);
646         radix_tree_preload_end();
647         if (unlikely(error))
648                 goto err_insert;
649
650         /* hugetlb pages do not participate in page cache accounting. */
651         if (!huge)
652                 __inc_zone_page_state(page, NR_FILE_PAGES);
653         spin_unlock_irq(&mapping->tree_lock);
654         if (!huge)
655                 mem_cgroup_commit_charge(page, memcg, false);
656         trace_mm_filemap_add_to_page_cache(page);
657         return 0;
658 err_insert:
659         page->mapping = NULL;
660         /* Leave page->index set: truncation relies upon it */
661         spin_unlock_irq(&mapping->tree_lock);
662         if (!huge)
663                 mem_cgroup_cancel_charge(page, memcg);
664         page_cache_release(page);
665         return error;
666 }
667
668 /**
669  * add_to_page_cache_locked - add a locked page to the pagecache
670  * @page:       page to add
671  * @mapping:    the page's address_space
672  * @offset:     page index
673  * @gfp_mask:   page allocation mode
674  *
675  * This function is used to add a page to the pagecache. It must be locked.
676  * This function does not add the page to the LRU.  The caller must do that.
677  */
678 int add_to_page_cache_locked(struct page *page, struct address_space *mapping,
679                 pgoff_t offset, gfp_t gfp_mask)
680 {
681         return __add_to_page_cache_locked(page, mapping, offset,
682                                           gfp_mask, NULL);
683 }
684 EXPORT_SYMBOL(add_to_page_cache_locked);
685
686 int add_to_page_cache_lru(struct page *page, struct address_space *mapping,
687                                 pgoff_t offset, gfp_t gfp_mask)
688 {
689         void *shadow = NULL;
690         int ret;
691
692         __set_page_locked(page);
693         ret = __add_to_page_cache_locked(page, mapping, offset,
694                                          gfp_mask, &shadow);
695         if (unlikely(ret))
696                 __clear_page_locked(page);
697         else {
698                 /*
699                  * The page might have been evicted from cache only
700                  * recently, in which case it should be activated like
701                  * any other repeatedly accessed page.
702                  */
703                 if (shadow && workingset_refault(shadow)) {
704                         SetPageActive(page);
705                         workingset_activation(page);
706                 } else
707                         ClearPageActive(page);
708                 lru_cache_add(page);
709         }
710         return ret;
711 }
712 EXPORT_SYMBOL_GPL(add_to_page_cache_lru);
713
714 #ifdef CONFIG_NUMA
715 struct page *__page_cache_alloc(gfp_t gfp)
716 {
717         int n;
718         struct page *page;
719
720         if (cpuset_do_page_mem_spread()) {
721                 unsigned int cpuset_mems_cookie;
722                 do {
723                         cpuset_mems_cookie = read_mems_allowed_begin();
724                         n = cpuset_mem_spread_node();
725                         page = __alloc_pages_node(n, gfp, 0);
726                 } while (!page && read_mems_allowed_retry(cpuset_mems_cookie));
727
728                 return page;
729         }
730         return alloc_pages(gfp, 0);
731 }
732 EXPORT_SYMBOL(__page_cache_alloc);
733 #endif
734
735 /*
736  * In order to wait for pages to become available there must be
737  * waitqueues associated with pages. By using a hash table of
738  * waitqueues where the bucket discipline is to maintain all
739  * waiters on the same queue and wake all when any of the pages
740  * become available, and for the woken contexts to check to be
741  * sure the appropriate page became available, this saves space
742  * at a cost of "thundering herd" phenomena during rare hash
743  * collisions.
744  */
745 wait_queue_head_t *page_waitqueue(struct page *page)
746 {
747         const struct zone *zone = page_zone(page);
748
749         return &zone->wait_table[hash_ptr(page, zone->wait_table_bits)];
750 }
751 EXPORT_SYMBOL(page_waitqueue);
752
753 void wait_on_page_bit(struct page *page, int bit_nr)
754 {
755         DEFINE_WAIT_BIT(wait, &page->flags, bit_nr);
756
757         if (test_bit(bit_nr, &page->flags))
758                 __wait_on_bit(page_waitqueue(page), &wait, bit_wait_io,
759                                                         TASK_UNINTERRUPTIBLE);
760 }
761 EXPORT_SYMBOL(wait_on_page_bit);
762
763 int wait_on_page_bit_killable(struct page *page, int bit_nr)
764 {
765         DEFINE_WAIT_BIT(wait, &page->flags, bit_nr);
766
767         if (!test_bit(bit_nr, &page->flags))
768                 return 0;
769
770         return __wait_on_bit(page_waitqueue(page), &wait,
771                              bit_wait_io, TASK_KILLABLE);
772 }
773
774 int wait_on_page_bit_killable_timeout(struct page *page,
775                                        int bit_nr, unsigned long timeout)
776 {
777         DEFINE_WAIT_BIT(wait, &page->flags, bit_nr);
778
779         wait.key.timeout = jiffies + timeout;
780         if (!test_bit(bit_nr, &page->flags))
781                 return 0;
782         return __wait_on_bit(page_waitqueue(page), &wait,
783                              bit_wait_io_timeout, TASK_KILLABLE);
784 }
785 EXPORT_SYMBOL_GPL(wait_on_page_bit_killable_timeout);
786
787 /**
788  * add_page_wait_queue - Add an arbitrary waiter to a page's wait queue
789  * @page: Page defining the wait queue of interest
790  * @waiter: Waiter to add to the queue
791  *
792  * Add an arbitrary @waiter to the wait queue for the nominated @page.
793  */
794 void add_page_wait_queue(struct page *page, wait_queue_t *waiter)
795 {
796         wait_queue_head_t *q = page_waitqueue(page);
797         unsigned long flags;
798
799         spin_lock_irqsave(&q->lock, flags);
800         __add_wait_queue(q, waiter);
801         spin_unlock_irqrestore(&q->lock, flags);
802 }
803 EXPORT_SYMBOL_GPL(add_page_wait_queue);
804
805 /**
806  * unlock_page - unlock a locked page
807  * @page: the page
808  *
809  * Unlocks the page and wakes up sleepers in ___wait_on_page_locked().
810  * Also wakes sleepers in wait_on_page_writeback() because the wakeup
811  * mechanism between PageLocked pages and PageWriteback pages is shared.
812  * But that's OK - sleepers in wait_on_page_writeback() just go back to sleep.
813  *
814  * The mb is necessary to enforce ordering between the clear_bit and the read
815  * of the waitqueue (to avoid SMP races with a parallel wait_on_page_locked()).
816  */
817 void unlock_page(struct page *page)
818 {
819         VM_BUG_ON_PAGE(!PageLocked(page), page);
820         clear_bit_unlock(PG_locked, &page->flags);
821         smp_mb__after_atomic();
822         wake_up_page(page, PG_locked);
823 }
824 EXPORT_SYMBOL(unlock_page);
825
826 /**
827  * end_page_writeback - end writeback against a page
828  * @page: the page
829  */
830 void end_page_writeback(struct page *page)
831 {
832         /*
833          * TestClearPageReclaim could be used here but it is an atomic
834          * operation and overkill in this particular case. Failing to
835          * shuffle a page marked for immediate reclaim is too mild to
836          * justify taking an atomic operation penalty at the end of
837          * ever page writeback.
838          */
839         if (PageReclaim(page)) {
840                 ClearPageReclaim(page);
841                 rotate_reclaimable_page(page);
842         }
843
844         if (!test_clear_page_writeback(page))
845                 BUG();
846
847         smp_mb__after_atomic();
848         wake_up_page(page, PG_writeback);
849 }
850 EXPORT_SYMBOL(end_page_writeback);
851
852 /*
853  * After completing I/O on a page, call this routine to update the page
854  * flags appropriately
855  */
856 void page_endio(struct page *page, int rw, int err)
857 {
858         if (rw == READ) {
859                 if (!err) {
860                         SetPageUptodate(page);
861                 } else {
862                         ClearPageUptodate(page);
863                         SetPageError(page);
864                 }
865                 unlock_page(page);
866         } else { /* rw == WRITE */
867                 if (err) {
868                         SetPageError(page);
869                         if (page->mapping)
870                                 mapping_set_error(page->mapping, err);
871                 }
872                 end_page_writeback(page);
873         }
874 }
875 EXPORT_SYMBOL_GPL(page_endio);
876
877 /**
878  * __lock_page - get a lock on the page, assuming we need to sleep to get it
879  * @page: the page to lock
880  */
881 void __lock_page(struct page *page)
882 {
883         DEFINE_WAIT_BIT(wait, &page->flags, PG_locked);
884
885         __wait_on_bit_lock(page_waitqueue(page), &wait, bit_wait_io,
886                                                         TASK_UNINTERRUPTIBLE);
887 }
888 EXPORT_SYMBOL(__lock_page);
889
890 int __lock_page_killable(struct page *page)
891 {
892         DEFINE_WAIT_BIT(wait, &page->flags, PG_locked);
893
894         return __wait_on_bit_lock(page_waitqueue(page), &wait,
895                                         bit_wait_io, TASK_KILLABLE);
896 }
897 EXPORT_SYMBOL_GPL(__lock_page_killable);
898
899 /*
900  * Return values:
901  * 1 - page is locked; mmap_sem is still held.
902  * 0 - page is not locked.
903  *     mmap_sem has been released (up_read()), unless flags had both
904  *     FAULT_FLAG_ALLOW_RETRY and FAULT_FLAG_RETRY_NOWAIT set, in
905  *     which case mmap_sem is still held.
906  *
907  * If neither ALLOW_RETRY nor KILLABLE are set, will always return 1
908  * with the page locked and the mmap_sem unperturbed.
909  */
910 int __lock_page_or_retry(struct page *page, struct mm_struct *mm,
911                          unsigned int flags)
912 {
913         if (flags & FAULT_FLAG_ALLOW_RETRY) {
914                 /*
915                  * CAUTION! In this case, mmap_sem is not released
916                  * even though return 0.
917                  */
918                 if (flags & FAULT_FLAG_RETRY_NOWAIT)
919                         return 0;
920
921                 up_read(&mm->mmap_sem);
922                 if (flags & FAULT_FLAG_KILLABLE)
923                         wait_on_page_locked_killable(page);
924                 else
925                         wait_on_page_locked(page);
926                 return 0;
927         } else {
928                 if (flags & FAULT_FLAG_KILLABLE) {
929                         int ret;
930
931                         ret = __lock_page_killable(page);
932                         if (ret) {
933                                 up_read(&mm->mmap_sem);
934                                 return 0;
935                         }
936                 } else
937                         __lock_page(page);
938                 return 1;
939         }
940 }
941
942 /**
943  * page_cache_next_hole - find the next hole (not-present entry)
944  * @mapping: mapping
945  * @index: index
946  * @max_scan: maximum range to search
947  *
948  * Search the set [index, min(index+max_scan-1, MAX_INDEX)] for the
949  * lowest indexed hole.
950  *
951  * Returns: the index of the hole if found, otherwise returns an index
952  * outside of the set specified (in which case 'return - index >=
953  * max_scan' will be true). In rare cases of index wrap-around, 0 will
954  * be returned.
955  *
956  * page_cache_next_hole may be called under rcu_read_lock. However,
957  * like radix_tree_gang_lookup, this will not atomically search a
958  * snapshot of the tree at a single point in time. For example, if a
959  * hole is created at index 5, then subsequently a hole is created at
960  * index 10, page_cache_next_hole covering both indexes may return 10
961  * if called under rcu_read_lock.
962  */
963 pgoff_t page_cache_next_hole(struct address_space *mapping,
964                              pgoff_t index, unsigned long max_scan)
965 {
966         unsigned long i;
967
968         for (i = 0; i < max_scan; i++) {
969                 struct page *page;
970
971                 page = radix_tree_lookup(&mapping->page_tree, index);
972                 if (!page || radix_tree_exceptional_entry(page))
973                         break;
974                 index++;
975                 if (index == 0)
976                         break;
977         }
978
979         return index;
980 }
981 EXPORT_SYMBOL(page_cache_next_hole);
982
983 /**
984  * page_cache_prev_hole - find the prev hole (not-present entry)
985  * @mapping: mapping
986  * @index: index
987  * @max_scan: maximum range to search
988  *
989  * Search backwards in the range [max(index-max_scan+1, 0), index] for
990  * the first hole.
991  *
992  * Returns: the index of the hole if found, otherwise returns an index
993  * outside of the set specified (in which case 'index - return >=
994  * max_scan' will be true). In rare cases of wrap-around, ULONG_MAX
995  * will be returned.
996  *
997  * page_cache_prev_hole may be called under rcu_read_lock. However,
998  * like radix_tree_gang_lookup, this will not atomically search a
999  * snapshot of the tree at a single point in time. For example, if a
1000  * hole is created at index 10, then subsequently a hole is created at
1001  * index 5, page_cache_prev_hole covering both indexes may return 5 if
1002  * called under rcu_read_lock.
1003  */
1004 pgoff_t page_cache_prev_hole(struct address_space *mapping,
1005                              pgoff_t index, unsigned long max_scan)
1006 {
1007         unsigned long i;
1008
1009         for (i = 0; i < max_scan; i++) {
1010                 struct page *page;
1011
1012                 page = radix_tree_lookup(&mapping->page_tree, index);
1013                 if (!page || radix_tree_exceptional_entry(page))
1014                         break;
1015                 index--;
1016                 if (index == ULONG_MAX)
1017                         break;
1018         }
1019
1020         return index;
1021 }
1022 EXPORT_SYMBOL(page_cache_prev_hole);
1023
1024 /**
1025  * find_get_entry - find and get a page cache entry
1026  * @mapping: the address_space to search
1027  * @offset: the page cache index
1028  *
1029  * Looks up the page cache slot at @mapping & @offset.  If there is a
1030  * page cache page, it is returned with an increased refcount.
1031  *
1032  * If the slot holds a shadow entry of a previously evicted page, or a
1033  * swap entry from shmem/tmpfs, it is returned.
1034  *
1035  * Otherwise, %NULL is returned.
1036  */
1037 struct page *find_get_entry(struct address_space *mapping, pgoff_t offset)
1038 {
1039         void **pagep;
1040         struct page *page;
1041
1042         rcu_read_lock();
1043 repeat:
1044         page = NULL;
1045         pagep = radix_tree_lookup_slot(&mapping->page_tree, offset);
1046         if (pagep) {
1047                 page = radix_tree_deref_slot(pagep);
1048                 if (unlikely(!page))
1049                         goto out;
1050                 if (radix_tree_exception(page)) {
1051                         if (radix_tree_deref_retry(page))
1052                                 goto repeat;
1053                         /*
1054                          * A shadow entry of a recently evicted page,
1055                          * or a swap entry from shmem/tmpfs.  Return
1056                          * it without attempting to raise page count.
1057                          */
1058                         goto out;
1059                 }
1060                 if (!page_cache_get_speculative(page))
1061                         goto repeat;
1062
1063                 /*
1064                  * Has the page moved?
1065                  * This is part of the lockless pagecache protocol. See
1066                  * include/linux/pagemap.h for details.
1067                  */
1068                 if (unlikely(page != *pagep)) {
1069                         page_cache_release(page);
1070                         goto repeat;
1071                 }
1072         }
1073 out:
1074         rcu_read_unlock();
1075
1076         return page;
1077 }
1078 EXPORT_SYMBOL(find_get_entry);
1079
1080 /**
1081  * find_lock_entry - locate, pin and lock a page cache entry
1082  * @mapping: the address_space to search
1083  * @offset: the page cache index
1084  *
1085  * Looks up the page cache slot at @mapping & @offset.  If there is a
1086  * page cache page, it is returned locked and with an increased
1087  * refcount.
1088  *
1089  * If the slot holds a shadow entry of a previously evicted page, or a
1090  * swap entry from shmem/tmpfs, it is returned.
1091  *
1092  * Otherwise, %NULL is returned.
1093  *
1094  * find_lock_entry() may sleep.
1095  */
1096 struct page *find_lock_entry(struct address_space *mapping, pgoff_t offset)
1097 {
1098         struct page *page;
1099
1100 repeat:
1101         page = find_get_entry(mapping, offset);
1102         if (page && !radix_tree_exception(page)) {
1103                 lock_page(page);
1104                 /* Has the page been truncated? */
1105                 if (unlikely(page->mapping != mapping)) {
1106                         unlock_page(page);
1107                         page_cache_release(page);
1108                         goto repeat;
1109                 }
1110                 VM_BUG_ON_PAGE(page->index != offset, page);
1111         }
1112         return page;
1113 }
1114 EXPORT_SYMBOL(find_lock_entry);
1115
1116 /**
1117  * pagecache_get_page - find and get a page reference
1118  * @mapping: the address_space to search
1119  * @offset: the page index
1120  * @fgp_flags: PCG flags
1121  * @gfp_mask: gfp mask to use for the page cache data page allocation
1122  *
1123  * Looks up the page cache slot at @mapping & @offset.
1124  *
1125  * PCG flags modify how the page is returned.
1126  *
1127  * FGP_ACCESSED: the page will be marked accessed
1128  * FGP_LOCK: Page is return locked
1129  * FGP_CREAT: If page is not present then a new page is allocated using
1130  *              @gfp_mask and added to the page cache and the VM's LRU
1131  *              list. The page is returned locked and with an increased
1132  *              refcount. Otherwise, %NULL is returned.
1133  *
1134  * If FGP_LOCK or FGP_CREAT are specified then the function may sleep even
1135  * if the GFP flags specified for FGP_CREAT are atomic.
1136  *
1137  * If there is a page cache page, it is returned with an increased refcount.
1138  */
1139 struct page *pagecache_get_page(struct address_space *mapping, pgoff_t offset,
1140         int fgp_flags, gfp_t gfp_mask)
1141 {
1142         struct page *page;
1143
1144 repeat:
1145         page = find_get_entry(mapping, offset);
1146         if (radix_tree_exceptional_entry(page))
1147                 page = NULL;
1148         if (!page)
1149                 goto no_page;
1150
1151         if (fgp_flags & FGP_LOCK) {
1152                 if (fgp_flags & FGP_NOWAIT) {
1153                         if (!trylock_page(page)) {
1154                                 page_cache_release(page);
1155                                 return NULL;
1156                         }
1157                 } else {
1158                         lock_page(page);
1159                 }
1160
1161                 /* Has the page been truncated? */
1162                 if (unlikely(page->mapping != mapping)) {
1163                         unlock_page(page);
1164                         page_cache_release(page);
1165                         goto repeat;
1166                 }
1167                 VM_BUG_ON_PAGE(page->index != offset, page);
1168         }
1169
1170         if (page && (fgp_flags & FGP_ACCESSED))
1171                 mark_page_accessed(page);
1172
1173 no_page:
1174         if (!page && (fgp_flags & FGP_CREAT)) {
1175                 int err;
1176                 if ((fgp_flags & FGP_WRITE) && mapping_cap_account_dirty(mapping))
1177                         gfp_mask |= __GFP_WRITE;
1178                 if (fgp_flags & FGP_NOFS)
1179                         gfp_mask &= ~__GFP_FS;
1180
1181                 page = __page_cache_alloc(gfp_mask);
1182                 if (!page)
1183                         return NULL;
1184
1185                 if (WARN_ON_ONCE(!(fgp_flags & FGP_LOCK)))
1186                         fgp_flags |= FGP_LOCK;
1187
1188                 /* Init accessed so avoid atomic mark_page_accessed later */
1189                 if (fgp_flags & FGP_ACCESSED)
1190                         __SetPageReferenced(page);
1191
1192                 err = add_to_page_cache_lru(page, mapping, offset,
1193                                 gfp_mask & GFP_RECLAIM_MASK);
1194                 if (unlikely(err)) {
1195                         page_cache_release(page);
1196                         page = NULL;
1197                         if (err == -EEXIST)
1198                                 goto repeat;
1199                 }
1200         }
1201
1202         return page;
1203 }
1204 EXPORT_SYMBOL(pagecache_get_page);
1205
1206 /**
1207  * find_get_entries - gang pagecache lookup
1208  * @mapping:    The address_space to search
1209  * @start:      The starting page cache index
1210  * @nr_entries: The maximum number of entries
1211  * @entries:    Where the resulting entries are placed
1212  * @indices:    The cache indices corresponding to the entries in @entries
1213  *
1214  * find_get_entries() will search for and return a group of up to
1215  * @nr_entries entries in the mapping.  The entries are placed at
1216  * @entries.  find_get_entries() takes a reference against any actual
1217  * pages it returns.
1218  *
1219  * The search returns a group of mapping-contiguous page cache entries
1220  * with ascending indexes.  There may be holes in the indices due to
1221  * not-present pages.
1222  *
1223  * Any shadow entries of evicted pages, or swap entries from
1224  * shmem/tmpfs, are included in the returned array.
1225  *
1226  * find_get_entries() returns the number of pages and shadow entries
1227  * which were found.
1228  */
1229 unsigned find_get_entries(struct address_space *mapping,
1230                           pgoff_t start, unsigned int nr_entries,
1231                           struct page **entries, pgoff_t *indices)
1232 {
1233         void **slot;
1234         unsigned int ret = 0;
1235         struct radix_tree_iter iter;
1236
1237         if (!nr_entries)
1238                 return 0;
1239
1240         rcu_read_lock();
1241 restart:
1242         radix_tree_for_each_slot(slot, &mapping->page_tree, &iter, start) {
1243                 struct page *page;
1244 repeat:
1245                 page = radix_tree_deref_slot(slot);
1246                 if (unlikely(!page))
1247                         continue;
1248                 if (radix_tree_exception(page)) {
1249                         if (radix_tree_deref_retry(page))
1250                                 goto restart;
1251                         /*
1252                          * A shadow entry of a recently evicted page,
1253                          * or a swap entry from shmem/tmpfs.  Return
1254                          * it without attempting to raise page count.
1255                          */
1256                         goto export;
1257                 }
1258                 if (!page_cache_get_speculative(page))
1259                         goto repeat;
1260
1261                 /* Has the page moved? */
1262                 if (unlikely(page != *slot)) {
1263                         page_cache_release(page);
1264                         goto repeat;
1265                 }
1266 export:
1267                 indices[ret] = iter.index;
1268                 entries[ret] = page;
1269                 if (++ret == nr_entries)
1270                         break;
1271         }
1272         rcu_read_unlock();
1273         return ret;
1274 }
1275
1276 /**
1277  * find_get_pages - gang pagecache lookup
1278  * @mapping:    The address_space to search
1279  * @start:      The starting page index
1280  * @nr_pages:   The maximum number of pages
1281  * @pages:      Where the resulting pages are placed
1282  *
1283  * find_get_pages() will search for and return a group of up to
1284  * @nr_pages pages in the mapping.  The pages are placed at @pages.
1285  * find_get_pages() takes a reference against the returned pages.
1286  *
1287  * The search returns a group of mapping-contiguous pages with ascending
1288  * indexes.  There may be holes in the indices due to not-present pages.
1289  *
1290  * find_get_pages() returns the number of pages which were found.
1291  */
1292 unsigned find_get_pages(struct address_space *mapping, pgoff_t start,
1293                             unsigned int nr_pages, struct page **pages)
1294 {
1295         struct radix_tree_iter iter;
1296         void **slot;
1297         unsigned ret = 0;
1298
1299         if (unlikely(!nr_pages))
1300                 return 0;
1301
1302         rcu_read_lock();
1303 restart:
1304         radix_tree_for_each_slot(slot, &mapping->page_tree, &iter, start) {
1305                 struct page *page;
1306 repeat:
1307                 page = radix_tree_deref_slot(slot);
1308                 if (unlikely(!page))
1309                         continue;
1310
1311                 if (radix_tree_exception(page)) {
1312                         if (radix_tree_deref_retry(page)) {
1313                                 /*
1314                                  * Transient condition which can only trigger
1315                                  * when entry at index 0 moves out of or back
1316                                  * to root: none yet gotten, safe to restart.
1317                                  */
1318                                 WARN_ON(iter.index);
1319                                 goto restart;
1320                         }
1321                         /*
1322                          * A shadow entry of a recently evicted page,
1323                          * or a swap entry from shmem/tmpfs.  Skip
1324                          * over it.
1325                          */
1326                         continue;
1327                 }
1328
1329                 if (!page_cache_get_speculative(page))
1330                         goto repeat;
1331
1332                 /* Has the page moved? */
1333                 if (unlikely(page != *slot)) {
1334                         page_cache_release(page);
1335                         goto repeat;
1336                 }
1337
1338                 pages[ret] = page;
1339                 if (++ret == nr_pages)
1340                         break;
1341         }
1342
1343         rcu_read_unlock();
1344         return ret;
1345 }
1346
1347 /**
1348  * find_get_pages_contig - gang contiguous pagecache lookup
1349  * @mapping:    The address_space to search
1350  * @index:      The starting page index
1351  * @nr_pages:   The maximum number of pages
1352  * @pages:      Where the resulting pages are placed
1353  *
1354  * find_get_pages_contig() works exactly like find_get_pages(), except
1355  * that the returned number of pages are guaranteed to be contiguous.
1356  *
1357  * find_get_pages_contig() returns the number of pages which were found.
1358  */
1359 unsigned find_get_pages_contig(struct address_space *mapping, pgoff_t index,
1360                                unsigned int nr_pages, struct page **pages)
1361 {
1362         struct radix_tree_iter iter;
1363         void **slot;
1364         unsigned int ret = 0;
1365
1366         if (unlikely(!nr_pages))
1367                 return 0;
1368
1369         rcu_read_lock();
1370 restart:
1371         radix_tree_for_each_contig(slot, &mapping->page_tree, &iter, index) {
1372                 struct page *page;
1373 repeat:
1374                 page = radix_tree_deref_slot(slot);
1375                 /* The hole, there no reason to continue */
1376                 if (unlikely(!page))
1377                         break;
1378
1379                 if (radix_tree_exception(page)) {
1380                         if (radix_tree_deref_retry(page)) {
1381                                 /*
1382                                  * Transient condition which can only trigger
1383                                  * when entry at index 0 moves out of or back
1384                                  * to root: none yet gotten, safe to restart.
1385                                  */
1386                                 goto restart;
1387                         }
1388                         /*
1389                          * A shadow entry of a recently evicted page,
1390                          * or a swap entry from shmem/tmpfs.  Stop
1391                          * looking for contiguous pages.
1392                          */
1393                         break;
1394                 }
1395
1396                 if (!page_cache_get_speculative(page))
1397                         goto repeat;
1398
1399                 /* Has the page moved? */
1400                 if (unlikely(page != *slot)) {
1401                         page_cache_release(page);
1402                         goto repeat;
1403                 }
1404
1405                 /*
1406                  * must check mapping and index after taking the ref.
1407                  * otherwise we can get both false positives and false
1408                  * negatives, which is just confusing to the caller.
1409                  */
1410                 if (page->mapping == NULL || page->index != iter.index) {
1411                         page_cache_release(page);
1412                         break;
1413                 }
1414
1415                 pages[ret] = page;
1416                 if (++ret == nr_pages)
1417                         break;
1418         }
1419         rcu_read_unlock();
1420         return ret;
1421 }
1422 EXPORT_SYMBOL(find_get_pages_contig);
1423
1424 /**
1425  * find_get_pages_tag - find and return pages that match @tag
1426  * @mapping:    the address_space to search
1427  * @index:      the starting page index
1428  * @tag:        the tag index
1429  * @nr_pages:   the maximum number of pages
1430  * @pages:      where the resulting pages are placed
1431  *
1432  * Like find_get_pages, except we only return pages which are tagged with
1433  * @tag.   We update @index to index the next page for the traversal.
1434  */
1435 unsigned find_get_pages_tag(struct address_space *mapping, pgoff_t *index,
1436                         int tag, unsigned int nr_pages, struct page **pages)
1437 {
1438         struct radix_tree_iter iter;
1439         void **slot;
1440         unsigned ret = 0;
1441
1442         if (unlikely(!nr_pages))
1443                 return 0;
1444
1445         rcu_read_lock();
1446 restart:
1447         radix_tree_for_each_tagged(slot, &mapping->page_tree,
1448                                    &iter, *index, tag) {
1449                 struct page *page;
1450 repeat:
1451                 page = radix_tree_deref_slot(slot);
1452                 if (unlikely(!page))
1453                         continue;
1454
1455                 if (radix_tree_exception(page)) {
1456                         if (radix_tree_deref_retry(page)) {
1457                                 /*
1458                                  * Transient condition which can only trigger
1459                                  * when entry at index 0 moves out of or back
1460                                  * to root: none yet gotten, safe to restart.
1461                                  */
1462                                 goto restart;
1463                         }
1464                         /*
1465                          * A shadow entry of a recently evicted page.
1466                          *
1467                          * Those entries should never be tagged, but
1468                          * this tree walk is lockless and the tags are
1469                          * looked up in bulk, one radix tree node at a
1470                          * time, so there is a sizable window for page
1471                          * reclaim to evict a page we saw tagged.
1472                          *
1473                          * Skip over it.
1474                          */
1475                         continue;
1476                 }
1477
1478                 if (!page_cache_get_speculative(page))
1479                         goto repeat;
1480
1481                 /* Has the page moved? */
1482                 if (unlikely(page != *slot)) {
1483                         page_cache_release(page);
1484                         goto repeat;
1485                 }
1486
1487                 pages[ret] = page;
1488                 if (++ret == nr_pages)
1489                         break;
1490         }
1491
1492         rcu_read_unlock();
1493
1494         if (ret)
1495                 *index = pages[ret - 1]->index + 1;
1496
1497         return ret;
1498 }
1499 EXPORT_SYMBOL(find_get_pages_tag);
1500
1501 /*
1502  * CD/DVDs are error prone. When a medium error occurs, the driver may fail
1503  * a _large_ part of the i/o request. Imagine the worst scenario:
1504  *
1505  *      ---R__________________________________________B__________
1506  *         ^ reading here                             ^ bad block(assume 4k)
1507  *
1508  * read(R) => miss => readahead(R...B) => media error => frustrating retries
1509  * => failing the whole request => read(R) => read(R+1) =>
1510  * readahead(R+1...B+1) => bang => read(R+2) => read(R+3) =>
1511  * readahead(R+3...B+2) => bang => read(R+3) => read(R+4) =>
1512  * readahead(R+4...B+3) => bang => read(R+4) => read(R+5) => ......
1513  *
1514  * It is going insane. Fix it by quickly scaling down the readahead size.
1515  */
1516 static void shrink_readahead_size_eio(struct file *filp,
1517                                         struct file_ra_state *ra)
1518 {
1519         ra->ra_pages /= 4;
1520 }
1521
1522 /**
1523  * do_generic_file_read - generic file read routine
1524  * @filp:       the file to read
1525  * @ppos:       current file position
1526  * @iter:       data destination
1527  * @written:    already copied
1528  *
1529  * This is a generic file read routine, and uses the
1530  * mapping->a_ops->readpage() function for the actual low-level stuff.
1531  *
1532  * This is really ugly. But the goto's actually try to clarify some
1533  * of the logic when it comes to error handling etc.
1534  */
1535 static ssize_t do_generic_file_read(struct file *filp, loff_t *ppos,
1536                 struct iov_iter *iter, ssize_t written)
1537 {
1538         struct address_space *mapping = filp->f_mapping;
1539         struct inode *inode = mapping->host;
1540         struct file_ra_state *ra = &filp->f_ra;
1541         pgoff_t index;
1542         pgoff_t last_index;
1543         pgoff_t prev_index;
1544         unsigned long offset;      /* offset into pagecache page */
1545         unsigned int prev_offset;
1546         int error = 0;
1547
1548         index = *ppos >> PAGE_CACHE_SHIFT;
1549         prev_index = ra->prev_pos >> PAGE_CACHE_SHIFT;
1550         prev_offset = ra->prev_pos & (PAGE_CACHE_SIZE-1);
1551         last_index = (*ppos + iter->count + PAGE_CACHE_SIZE-1) >> PAGE_CACHE_SHIFT;
1552         offset = *ppos & ~PAGE_CACHE_MASK;
1553
1554         for (;;) {
1555                 struct page *page;
1556                 pgoff_t end_index;
1557                 loff_t isize;
1558                 unsigned long nr, ret;
1559
1560                 cond_resched();
1561 find_page:
1562                 if (fatal_signal_pending(current)) {
1563                         error = -EINTR;
1564                         goto out;
1565                 }
1566
1567                 page = find_get_page(mapping, index);
1568                 if (!page) {
1569                         page_cache_sync_readahead(mapping,
1570                                         ra, filp,
1571                                         index, last_index - index);
1572                         page = find_get_page(mapping, index);
1573                         if (unlikely(page == NULL))
1574                                 goto no_cached_page;
1575                 }
1576                 if (PageReadahead(page)) {
1577                         page_cache_async_readahead(mapping,
1578                                         ra, filp, page,
1579                                         index, last_index - index);
1580                 }
1581                 if (!PageUptodate(page)) {
1582                         if (inode->i_blkbits == PAGE_CACHE_SHIFT ||
1583                                         !mapping->a_ops->is_partially_uptodate)
1584                                 goto page_not_up_to_date;
1585                         if (!trylock_page(page))
1586                                 goto page_not_up_to_date;
1587                         /* Did it get truncated before we got the lock? */
1588                         if (!page->mapping)
1589                                 goto page_not_up_to_date_locked;
1590                         if (!mapping->a_ops->is_partially_uptodate(page,
1591                                                         offset, iter->count))
1592                                 goto page_not_up_to_date_locked;
1593                         unlock_page(page);
1594                 }
1595 page_ok:
1596                 /*
1597                  * i_size must be checked after we know the page is Uptodate.
1598                  *
1599                  * Checking i_size after the check allows us to calculate
1600                  * the correct value for "nr", which means the zero-filled
1601                  * part of the page is not copied back to userspace (unless
1602                  * another truncate extends the file - this is desired though).
1603                  */
1604
1605                 isize = i_size_read(inode);
1606                 end_index = (isize - 1) >> PAGE_CACHE_SHIFT;
1607                 if (unlikely(!isize || index > end_index)) {
1608                         page_cache_release(page);
1609                         goto out;
1610                 }
1611
1612                 /* nr is the maximum number of bytes to copy from this page */
1613                 nr = PAGE_CACHE_SIZE;
1614                 if (index == end_index) {
1615                         nr = ((isize - 1) & ~PAGE_CACHE_MASK) + 1;
1616                         if (nr <= offset) {
1617                                 page_cache_release(page);
1618                                 goto out;
1619                         }
1620                 }
1621                 nr = nr - offset;
1622
1623                 /* If users can be writing to this page using arbitrary
1624                  * virtual addresses, take care about potential aliasing
1625                  * before reading the page on the kernel side.
1626                  */
1627                 if (mapping_writably_mapped(mapping))
1628                         flush_dcache_page(page);
1629
1630                 /*
1631                  * When a sequential read accesses a page several times,
1632                  * only mark it as accessed the first time.
1633                  */
1634                 if (prev_index != index || offset != prev_offset)
1635                         mark_page_accessed(page);
1636                 prev_index = index;
1637
1638                 /*
1639                  * Ok, we have the page, and it's up-to-date, so
1640                  * now we can copy it to user space...
1641                  */
1642
1643                 ret = copy_page_to_iter(page, offset, nr, iter);
1644                 offset += ret;
1645                 index += offset >> PAGE_CACHE_SHIFT;
1646                 offset &= ~PAGE_CACHE_MASK;
1647                 prev_offset = offset;
1648
1649                 page_cache_release(page);
1650                 written += ret;
1651                 if (!iov_iter_count(iter))
1652                         goto out;
1653                 if (ret < nr) {
1654                         error = -EFAULT;
1655                         goto out;
1656                 }
1657                 continue;
1658
1659 page_not_up_to_date:
1660                 /* Get exclusive access to the page ... */
1661                 error = lock_page_killable(page);
1662                 if (unlikely(error))
1663                         goto readpage_error;
1664
1665 page_not_up_to_date_locked:
1666                 /* Did it get truncated before we got the lock? */
1667                 if (!page->mapping) {
1668                         unlock_page(page);
1669                         page_cache_release(page);
1670                         continue;
1671                 }
1672
1673                 /* Did somebody else fill it already? */
1674                 if (PageUptodate(page)) {
1675                         unlock_page(page);
1676                         goto page_ok;
1677                 }
1678
1679 readpage:
1680                 /*
1681                  * A previous I/O error may have been due to temporary
1682                  * failures, eg. multipath errors.
1683                  * PG_error will be set again if readpage fails.
1684                  */
1685                 ClearPageError(page);
1686                 /* Start the actual read. The read will unlock the page. */
1687                 error = mapping->a_ops->readpage(filp, page);
1688
1689                 if (unlikely(error)) {
1690                         if (error == AOP_TRUNCATED_PAGE) {
1691                                 page_cache_release(page);
1692                                 error = 0;
1693                                 goto find_page;
1694                         }
1695                         goto readpage_error;
1696                 }
1697
1698                 if (!PageUptodate(page)) {
1699                         error = lock_page_killable(page);
1700                         if (unlikely(error))
1701                                 goto readpage_error;
1702                         if (!PageUptodate(page)) {
1703                                 if (page->mapping == NULL) {
1704                                         /*
1705                                          * invalidate_mapping_pages got it
1706                                          */
1707                                         unlock_page(page);
1708                                         page_cache_release(page);
1709                                         goto find_page;
1710                                 }
1711                                 unlock_page(page);
1712                                 shrink_readahead_size_eio(filp, ra);
1713                                 error = -EIO;
1714                                 goto readpage_error;
1715                         }
1716                         unlock_page(page);
1717                 }
1718
1719                 goto page_ok;
1720
1721 readpage_error:
1722                 /* UHHUH! A synchronous read error occurred. Report it */
1723                 page_cache_release(page);
1724                 goto out;
1725
1726 no_cached_page:
1727                 /*
1728                  * Ok, it wasn't cached, so we need to create a new
1729                  * page..
1730                  */
1731                 page = page_cache_alloc_cold(mapping);
1732                 if (!page) {
1733                         error = -ENOMEM;
1734                         goto out;
1735                 }
1736                 error = add_to_page_cache_lru(page, mapping, index,
1737                                 mapping_gfp_constraint(mapping, GFP_KERNEL));
1738                 if (error) {
1739                         page_cache_release(page);
1740                         if (error == -EEXIST) {
1741                                 error = 0;
1742                                 goto find_page;
1743                         }
1744                         goto out;
1745                 }
1746                 goto readpage;
1747         }
1748
1749 out:
1750         ra->prev_pos = prev_index;
1751         ra->prev_pos <<= PAGE_CACHE_SHIFT;
1752         ra->prev_pos |= prev_offset;
1753
1754         *ppos = ((loff_t)index << PAGE_CACHE_SHIFT) + offset;
1755         file_accessed(filp);
1756         return written ? written : error;
1757 }
1758
1759 /**
1760  * generic_file_read_iter - generic filesystem read routine
1761  * @iocb:       kernel I/O control block
1762  * @iter:       destination for the data read
1763  *
1764  * This is the "read_iter()" routine for all filesystems
1765  * that can use the page cache directly.
1766  */
1767 ssize_t
1768 generic_file_read_iter(struct kiocb *iocb, struct iov_iter *iter)
1769 {
1770         struct file *file = iocb->ki_filp;
1771         ssize_t retval = 0;
1772         loff_t *ppos = &iocb->ki_pos;
1773         loff_t pos = *ppos;
1774
1775         if (iocb->ki_flags & IOCB_DIRECT) {
1776                 struct address_space *mapping = file->f_mapping;
1777                 struct inode *inode = mapping->host;
1778                 size_t count = iov_iter_count(iter);
1779                 loff_t size;
1780
1781                 if (!count)
1782                         goto out; /* skip atime */
1783                 size = i_size_read(inode);
1784                 retval = filemap_write_and_wait_range(mapping, pos,
1785                                         pos + count - 1);
1786                 if (!retval) {
1787                         struct iov_iter data = *iter;
1788                         retval = mapping->a_ops->direct_IO(iocb, &data, pos);
1789                 }
1790
1791                 if (retval > 0) {
1792                         *ppos = pos + retval;
1793                         iov_iter_advance(iter, retval);
1794                 }
1795
1796                 /*
1797                  * Btrfs can have a short DIO read if we encounter
1798                  * compressed extents, so if there was an error, or if
1799                  * we've already read everything we wanted to, or if
1800                  * there was a short read because we hit EOF, go ahead
1801                  * and return.  Otherwise fallthrough to buffered io for
1802                  * the rest of the read.  Buffered reads will not work for
1803                  * DAX files, so don't bother trying.
1804                  */
1805                 if (retval < 0 || !iov_iter_count(iter) || *ppos >= size ||
1806                     IS_DAX(inode)) {
1807                         file_accessed(file);
1808                         goto out;
1809                 }
1810         }
1811
1812         retval = do_generic_file_read(file, ppos, iter, retval);
1813 out:
1814         return retval;
1815 }
1816 EXPORT_SYMBOL(generic_file_read_iter);
1817
1818 #ifdef CONFIG_MMU
1819 /**
1820  * page_cache_read - adds requested page to the page cache if not already there
1821  * @file:       file to read
1822  * @offset:     page index
1823  *
1824  * This adds the requested page to the page cache if it isn't already there,
1825  * and schedules an I/O to read in its contents from disk.
1826  */
1827 static int page_cache_read(struct file *file, pgoff_t offset)
1828 {
1829         struct address_space *mapping = file->f_mapping;
1830         struct page *page;
1831         int ret;
1832
1833         do {
1834                 page = page_cache_alloc_cold(mapping);
1835                 if (!page)
1836                         return -ENOMEM;
1837
1838                 ret = add_to_page_cache_lru(page, mapping, offset,
1839                                 mapping_gfp_constraint(mapping, GFP_KERNEL));
1840                 if (ret == 0)
1841                         ret = mapping->a_ops->readpage(file, page);
1842                 else if (ret == -EEXIST)
1843                         ret = 0; /* losing race to add is OK */
1844
1845                 page_cache_release(page);
1846
1847         } while (ret == AOP_TRUNCATED_PAGE);
1848
1849         return ret;
1850 }
1851
1852 #define MMAP_LOTSAMISS  (100)
1853
1854 /*
1855  * Synchronous readahead happens when we don't even find
1856  * a page in the page cache at all.
1857  */
1858 static void do_sync_mmap_readahead(struct vm_area_struct *vma,
1859                                    struct file_ra_state *ra,
1860                                    struct file *file,
1861                                    pgoff_t offset)
1862 {
1863         struct address_space *mapping = file->f_mapping;
1864
1865         /* If we don't want any read-ahead, don't bother */
1866         if (vma->vm_flags & VM_RAND_READ)
1867                 return;
1868         if (!ra->ra_pages)
1869                 return;
1870
1871         if (vma->vm_flags & VM_SEQ_READ) {
1872                 page_cache_sync_readahead(mapping, ra, file, offset,
1873                                           ra->ra_pages);
1874                 return;
1875         }
1876
1877         /* Avoid banging the cache line if not needed */
1878         if (ra->mmap_miss < MMAP_LOTSAMISS * 10)
1879                 ra->mmap_miss++;
1880
1881         /*
1882          * Do we miss much more than hit in this file? If so,
1883          * stop bothering with read-ahead. It will only hurt.
1884          */
1885         if (ra->mmap_miss > MMAP_LOTSAMISS)
1886                 return;
1887
1888         /*
1889          * mmap read-around
1890          */
1891         ra->start = max_t(long, 0, offset - ra->ra_pages / 2);
1892         ra->size = ra->ra_pages;
1893         ra->async_size = ra->ra_pages / 4;
1894         ra_submit(ra, mapping, file);
1895 }
1896
1897 /*
1898  * Asynchronous readahead happens when we find the page and PG_readahead,
1899  * so we want to possibly extend the readahead further..
1900  */
1901 static void do_async_mmap_readahead(struct vm_area_struct *vma,
1902                                     struct file_ra_state *ra,
1903                                     struct file *file,
1904                                     struct page *page,
1905                                     pgoff_t offset)
1906 {
1907         struct address_space *mapping = file->f_mapping;
1908
1909         /* If we don't want any read-ahead, don't bother */
1910         if (vma->vm_flags & VM_RAND_READ)
1911                 return;
1912         if (ra->mmap_miss > 0)
1913                 ra->mmap_miss--;
1914         if (PageReadahead(page))
1915                 page_cache_async_readahead(mapping, ra, file,
1916                                            page, offset, ra->ra_pages);
1917 }
1918
1919 /**
1920  * filemap_fault - read in file data for page fault handling
1921  * @vma:        vma in which the fault was taken
1922  * @vmf:        struct vm_fault containing details of the fault
1923  *
1924  * filemap_fault() is invoked via the vma operations vector for a
1925  * mapped memory region to read in file data during a page fault.
1926  *
1927  * The goto's are kind of ugly, but this streamlines the normal case of having
1928  * it in the page cache, and handles the special cases reasonably without
1929  * having a lot of duplicated code.
1930  *
1931  * vma->vm_mm->mmap_sem must be held on entry.
1932  *
1933  * If our return value has VM_FAULT_RETRY set, it's because
1934  * lock_page_or_retry() returned 0.
1935  * The mmap_sem has usually been released in this case.
1936  * See __lock_page_or_retry() for the exception.
1937  *
1938  * If our return value does not have VM_FAULT_RETRY set, the mmap_sem
1939  * has not been released.
1940  *
1941  * We never return with VM_FAULT_RETRY and a bit from VM_FAULT_ERROR set.
1942  */
1943 int filemap_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
1944 {
1945         int error;
1946         struct file *file = vma->vm_file;
1947         struct address_space *mapping = file->f_mapping;
1948         struct file_ra_state *ra = &file->f_ra;
1949         struct inode *inode = mapping->host;
1950         pgoff_t offset = vmf->pgoff;
1951         struct page *page;
1952         loff_t size;
1953         int ret = 0;
1954
1955         size = round_up(i_size_read(inode), PAGE_CACHE_SIZE);
1956         if (offset >= size >> PAGE_CACHE_SHIFT)
1957                 return VM_FAULT_SIGBUS;
1958
1959         /*
1960          * Do we have something in the page cache already?
1961          */
1962         page = find_get_page(mapping, offset);
1963         if (likely(page) && !(vmf->flags & FAULT_FLAG_TRIED)) {
1964                 /*
1965                  * We found the page, so try async readahead before
1966                  * waiting for the lock.
1967                  */
1968                 do_async_mmap_readahead(vma, ra, file, page, offset);
1969         } else if (!page) {
1970                 /* No page in the page cache at all */
1971                 do_sync_mmap_readahead(vma, ra, file, offset);
1972                 count_vm_event(PGMAJFAULT);
1973                 mem_cgroup_count_vm_event(vma->vm_mm, PGMAJFAULT);
1974                 ret = VM_FAULT_MAJOR;
1975 retry_find:
1976                 page = find_get_page(mapping, offset);
1977                 if (!page)
1978                         goto no_cached_page;
1979         }
1980
1981         if (!lock_page_or_retry(page, vma->vm_mm, vmf->flags)) {
1982                 page_cache_release(page);
1983                 return ret | VM_FAULT_RETRY;
1984         }
1985
1986         /* Did it get truncated? */
1987         if (unlikely(page->mapping != mapping)) {
1988                 unlock_page(page);
1989                 put_page(page);
1990                 goto retry_find;
1991         }
1992         VM_BUG_ON_PAGE(page->index != offset, page);
1993
1994         /*
1995          * We have a locked page in the page cache, now we need to check
1996          * that it's up-to-date. If not, it is going to be due to an error.
1997          */
1998         if (unlikely(!PageUptodate(page)))
1999                 goto page_not_uptodate;
2000
2001         /*
2002          * Found the page and have a reference on it.
2003          * We must recheck i_size under page lock.
2004          */
2005         size = round_up(i_size_read(inode), PAGE_CACHE_SIZE);
2006         if (unlikely(offset >= size >> PAGE_CACHE_SHIFT)) {
2007                 unlock_page(page);
2008                 page_cache_release(page);
2009                 return VM_FAULT_SIGBUS;
2010         }
2011
2012         vmf->page = page;
2013         return ret | VM_FAULT_LOCKED;
2014
2015 no_cached_page:
2016         /*
2017          * We're only likely to ever get here if MADV_RANDOM is in
2018          * effect.
2019          */
2020         error = page_cache_read(file, offset);
2021
2022         /*
2023          * The page we want has now been added to the page cache.
2024          * In the unlikely event that someone removed it in the
2025          * meantime, we'll just come back here and read it again.
2026          */
2027         if (error >= 0)
2028                 goto retry_find;
2029
2030         /*
2031          * An error return from page_cache_read can result if the
2032          * system is low on memory, or a problem occurs while trying
2033          * to schedule I/O.
2034          */
2035         if (error == -ENOMEM)
2036                 return VM_FAULT_OOM;
2037         return VM_FAULT_SIGBUS;
2038
2039 page_not_uptodate:
2040         /*
2041          * Umm, take care of errors if the page isn't up-to-date.
2042          * Try to re-read it _once_. We do this synchronously,
2043          * because there really aren't any performance issues here
2044          * and we need to check for errors.
2045          */
2046         ClearPageError(page);
2047         error = mapping->a_ops->readpage(file, page);
2048         if (!error) {
2049                 wait_on_page_locked(page);
2050                 if (!PageUptodate(page))
2051                         error = -EIO;
2052         }
2053         page_cache_release(page);
2054
2055         if (!error || error == AOP_TRUNCATED_PAGE)
2056                 goto retry_find;
2057
2058         /* Things didn't work out. Return zero to tell the mm layer so. */
2059         shrink_readahead_size_eio(file, ra);
2060         return VM_FAULT_SIGBUS;
2061 }
2062 EXPORT_SYMBOL(filemap_fault);
2063
2064 void filemap_map_pages(struct vm_area_struct *vma, struct vm_fault *vmf)
2065 {
2066         struct radix_tree_iter iter;
2067         void **slot;
2068         struct file *file = vma->vm_file;
2069         struct address_space *mapping = file->f_mapping;
2070         loff_t size;
2071         struct page *page;
2072         unsigned long address = (unsigned long) vmf->virtual_address;
2073         unsigned long addr;
2074         pte_t *pte;
2075
2076         rcu_read_lock();
2077         radix_tree_for_each_slot(slot, &mapping->page_tree, &iter, vmf->pgoff) {
2078                 if (iter.index > vmf->max_pgoff)
2079                         break;
2080 repeat:
2081                 page = radix_tree_deref_slot(slot);
2082                 if (unlikely(!page))
2083                         goto next;
2084                 if (radix_tree_exception(page)) {
2085                         if (radix_tree_deref_retry(page))
2086                                 break;
2087                         else
2088                                 goto next;
2089                 }
2090
2091                 if (!page_cache_get_speculative(page))
2092                         goto repeat;
2093
2094                 /* Has the page moved? */
2095                 if (unlikely(page != *slot)) {
2096                         page_cache_release(page);
2097                         goto repeat;
2098                 }
2099
2100                 if (!PageUptodate(page) ||
2101                                 PageReadahead(page) ||
2102                                 PageHWPoison(page))
2103                         goto skip;
2104                 if (!trylock_page(page))
2105                         goto skip;
2106
2107                 if (page->mapping != mapping || !PageUptodate(page))
2108                         goto unlock;
2109
2110                 size = round_up(i_size_read(mapping->host), PAGE_CACHE_SIZE);
2111                 if (page->index >= size >> PAGE_CACHE_SHIFT)
2112                         goto unlock;
2113
2114                 pte = vmf->pte + page->index - vmf->pgoff;
2115                 if (!pte_none(*pte))
2116                         goto unlock;
2117
2118                 if (file->f_ra.mmap_miss > 0)
2119                         file->f_ra.mmap_miss--;
2120                 addr = address + (page->index - vmf->pgoff) * PAGE_SIZE;
2121                 do_set_pte(vma, addr, page, pte, false, false);
2122                 unlock_page(page);
2123                 goto next;
2124 unlock:
2125                 unlock_page(page);
2126 skip:
2127                 page_cache_release(page);
2128 next:
2129                 if (iter.index == vmf->max_pgoff)
2130                         break;
2131         }
2132         rcu_read_unlock();
2133 }
2134 EXPORT_SYMBOL(filemap_map_pages);
2135
2136 int filemap_page_mkwrite(struct vm_area_struct *vma, struct vm_fault *vmf)
2137 {
2138         struct page *page = vmf->page;
2139         struct inode *inode = file_inode(vma->vm_file);
2140         int ret = VM_FAULT_LOCKED;
2141
2142         sb_start_pagefault(inode->i_sb);
2143         file_update_time(vma->vm_file);
2144         lock_page(page);
2145         if (page->mapping != inode->i_mapping) {
2146                 unlock_page(page);
2147                 ret = VM_FAULT_NOPAGE;
2148                 goto out;
2149         }
2150         /*
2151          * We mark the page dirty already here so that when freeze is in
2152          * progress, we are guaranteed that writeback during freezing will
2153          * see the dirty page and writeprotect it again.
2154          */
2155         set_page_dirty(page);
2156         wait_for_stable_page(page);
2157 out:
2158         sb_end_pagefault(inode->i_sb);
2159         return ret;
2160 }
2161 EXPORT_SYMBOL(filemap_page_mkwrite);
2162
2163 const struct vm_operations_struct generic_file_vm_ops = {
2164         .fault          = filemap_fault,
2165         .map_pages      = filemap_map_pages,
2166         .page_mkwrite   = filemap_page_mkwrite,
2167 };
2168
2169 /* This is used for a general mmap of a disk file */
2170
2171 int generic_file_mmap(struct file * file, struct vm_area_struct * vma)
2172 {
2173         struct address_space *mapping = file->f_mapping;
2174
2175         if (!mapping->a_ops->readpage)
2176                 return -ENOEXEC;
2177         file_accessed(file);
2178         vma->vm_ops = &generic_file_vm_ops;
2179         return 0;
2180 }
2181
2182 /*
2183  * This is for filesystems which do not implement ->writepage.
2184  */
2185 int generic_file_readonly_mmap(struct file *file, struct vm_area_struct *vma)
2186 {
2187         if ((vma->vm_flags & VM_SHARED) && (vma->vm_flags & VM_MAYWRITE))
2188                 return -EINVAL;
2189         return generic_file_mmap(file, vma);
2190 }
2191 #else
2192 int generic_file_mmap(struct file * file, struct vm_area_struct * vma)
2193 {
2194         return -ENOSYS;
2195 }
2196 int generic_file_readonly_mmap(struct file * file, struct vm_area_struct * vma)
2197 {
2198         return -ENOSYS;
2199 }
2200 #endif /* CONFIG_MMU */
2201
2202 EXPORT_SYMBOL(generic_file_mmap);
2203 EXPORT_SYMBOL(generic_file_readonly_mmap);
2204
2205 static struct page *wait_on_page_read(struct page *page)
2206 {
2207         if (!IS_ERR(page)) {
2208                 wait_on_page_locked(page);
2209                 if (!PageUptodate(page)) {
2210                         page_cache_release(page);
2211                         page = ERR_PTR(-EIO);
2212                 }
2213         }
2214         return page;
2215 }
2216
2217 static struct page *__read_cache_page(struct address_space *mapping,
2218                                 pgoff_t index,
2219                                 int (*filler)(void *, struct page *),
2220                                 void *data,
2221                                 gfp_t gfp)
2222 {
2223         struct page *page;
2224         int err;
2225 repeat:
2226         page = find_get_page(mapping, index);
2227         if (!page) {
2228                 page = __page_cache_alloc(gfp | __GFP_COLD);
2229                 if (!page)
2230                         return ERR_PTR(-ENOMEM);
2231                 err = add_to_page_cache_lru(page, mapping, index, gfp);
2232                 if (unlikely(err)) {
2233                         page_cache_release(page);
2234                         if (err == -EEXIST)
2235                                 goto repeat;
2236                         /* Presumably ENOMEM for radix tree node */
2237                         return ERR_PTR(err);
2238                 }
2239                 err = filler(data, page);
2240                 if (err < 0) {
2241                         page_cache_release(page);
2242                         page = ERR_PTR(err);
2243                 } else {
2244                         page = wait_on_page_read(page);
2245                 }
2246         }
2247         return page;
2248 }
2249
2250 static struct page *do_read_cache_page(struct address_space *mapping,
2251                                 pgoff_t index,
2252                                 int (*filler)(void *, struct page *),
2253                                 void *data,
2254                                 gfp_t gfp)
2255
2256 {
2257         struct page *page;
2258         int err;
2259
2260 retry:
2261         page = __read_cache_page(mapping, index, filler, data, gfp);
2262         if (IS_ERR(page))
2263                 return page;
2264         if (PageUptodate(page))
2265                 goto out;
2266
2267         lock_page(page);
2268         if (!page->mapping) {
2269                 unlock_page(page);
2270                 page_cache_release(page);
2271                 goto retry;
2272         }
2273         if (PageUptodate(page)) {
2274                 unlock_page(page);
2275                 goto out;
2276         }
2277         err = filler(data, page);
2278         if (err < 0) {
2279                 page_cache_release(page);
2280                 return ERR_PTR(err);
2281         } else {
2282                 page = wait_on_page_read(page);
2283                 if (IS_ERR(page))
2284                         return page;
2285         }
2286 out:
2287         mark_page_accessed(page);
2288         return page;
2289 }
2290
2291 /**
2292  * read_cache_page - read into page cache, fill it if needed
2293  * @mapping:    the page's address_space
2294  * @index:      the page index
2295  * @filler:     function to perform the read
2296  * @data:       first arg to filler(data, page) function, often left as NULL
2297  *
2298  * Read into the page cache. If a page already exists, and PageUptodate() is
2299  * not set, try to fill the page and wait for it to become unlocked.
2300  *
2301  * If the page does not get brought uptodate, return -EIO.
2302  */
2303 struct page *read_cache_page(struct address_space *mapping,
2304                                 pgoff_t index,
2305                                 int (*filler)(void *, struct page *),
2306                                 void *data)
2307 {
2308         return do_read_cache_page(mapping, index, filler, data, mapping_gfp_mask(mapping));
2309 }
2310 EXPORT_SYMBOL(read_cache_page);
2311
2312 /**
2313  * read_cache_page_gfp - read into page cache, using specified page allocation flags.
2314  * @mapping:    the page's address_space
2315  * @index:      the page index
2316  * @gfp:        the page allocator flags to use if allocating
2317  *
2318  * This is the same as "read_mapping_page(mapping, index, NULL)", but with
2319  * any new page allocations done using the specified allocation flags.
2320  *
2321  * If the page does not get brought uptodate, return -EIO.
2322  */
2323 struct page *read_cache_page_gfp(struct address_space *mapping,
2324                                 pgoff_t index,
2325                                 gfp_t gfp)
2326 {
2327         filler_t *filler = (filler_t *)mapping->a_ops->readpage;
2328
2329         return do_read_cache_page(mapping, index, filler, NULL, gfp);
2330 }
2331 EXPORT_SYMBOL(read_cache_page_gfp);
2332
2333 /*
2334  * Performs necessary checks before doing a write
2335  *
2336  * Can adjust writing position or amount of bytes to write.
2337  * Returns appropriate error code that caller should return or
2338  * zero in case that write should be allowed.
2339  */
2340 inline ssize_t generic_write_checks(struct kiocb *iocb, struct iov_iter *from)
2341 {
2342         struct file *file = iocb->ki_filp;
2343         struct inode *inode = file->f_mapping->host;
2344         unsigned long limit = rlimit(RLIMIT_FSIZE);
2345         loff_t pos;
2346
2347         if (!iov_iter_count(from))
2348                 return 0;
2349
2350         /* FIXME: this is for backwards compatibility with 2.4 */
2351         if (iocb->ki_flags & IOCB_APPEND)
2352                 iocb->ki_pos = i_size_read(inode);
2353
2354         pos = iocb->ki_pos;
2355
2356         if (limit != RLIM_INFINITY) {
2357                 if (iocb->ki_pos >= limit) {
2358                         send_sig(SIGXFSZ, current, 0);
2359                         return -EFBIG;
2360                 }
2361                 iov_iter_truncate(from, limit - (unsigned long)pos);
2362         }
2363
2364         /*
2365          * LFS rule
2366          */
2367         if (unlikely(pos + iov_iter_count(from) > MAX_NON_LFS &&
2368                                 !(file->f_flags & O_LARGEFILE))) {
2369                 if (pos >= MAX_NON_LFS)
2370                         return -EFBIG;
2371                 iov_iter_truncate(from, MAX_NON_LFS - (unsigned long)pos);
2372         }
2373
2374         /*
2375          * Are we about to exceed the fs block limit ?
2376          *
2377          * If we have written data it becomes a short write.  If we have
2378          * exceeded without writing data we send a signal and return EFBIG.
2379          * Linus frestrict idea will clean these up nicely..
2380          */
2381         if (unlikely(pos >= inode->i_sb->s_maxbytes))
2382                 return -EFBIG;
2383
2384         iov_iter_truncate(from, inode->i_sb->s_maxbytes - pos);
2385         return iov_iter_count(from);
2386 }
2387 EXPORT_SYMBOL(generic_write_checks);
2388
2389 int pagecache_write_begin(struct file *file, struct address_space *mapping,
2390                                 loff_t pos, unsigned len, unsigned flags,
2391                                 struct page **pagep, void **fsdata)
2392 {
2393         const struct address_space_operations *aops = mapping->a_ops;
2394
2395         return aops->write_begin(file, mapping, pos, len, flags,
2396                                                         pagep, fsdata);
2397 }
2398 EXPORT_SYMBOL(pagecache_write_begin);
2399
2400 int pagecache_write_end(struct file *file, struct address_space *mapping,
2401                                 loff_t pos, unsigned len, unsigned copied,
2402                                 struct page *page, void *fsdata)
2403 {
2404         const struct address_space_operations *aops = mapping->a_ops;
2405
2406         return aops->write_end(file, mapping, pos, len, copied, page, fsdata);
2407 }
2408 EXPORT_SYMBOL(pagecache_write_end);
2409
2410 ssize_t
2411 generic_file_direct_write(struct kiocb *iocb, struct iov_iter *from, loff_t pos)
2412 {
2413         struct file     *file = iocb->ki_filp;
2414         struct address_space *mapping = file->f_mapping;
2415         struct inode    *inode = mapping->host;
2416         ssize_t         written;
2417         size_t          write_len;
2418         pgoff_t         end;
2419         struct iov_iter data;
2420
2421         write_len = iov_iter_count(from);
2422         end = (pos + write_len - 1) >> PAGE_CACHE_SHIFT;
2423
2424         written = filemap_write_and_wait_range(mapping, pos, pos + write_len - 1);
2425         if (written)
2426                 goto out;
2427
2428         /*
2429          * After a write we want buffered reads to be sure to go to disk to get
2430          * the new data.  We invalidate clean cached page from the region we're
2431          * about to write.  We do this *before* the write so that we can return
2432          * without clobbering -EIOCBQUEUED from ->direct_IO().
2433          */
2434         if (mapping->nrpages) {
2435                 written = invalidate_inode_pages2_range(mapping,
2436                                         pos >> PAGE_CACHE_SHIFT, end);
2437                 /*
2438                  * If a page can not be invalidated, return 0 to fall back
2439                  * to buffered write.
2440                  */
2441                 if (written) {
2442                         if (written == -EBUSY)
2443                                 return 0;
2444                         goto out;
2445                 }
2446         }
2447
2448         data = *from;
2449         written = mapping->a_ops->direct_IO(iocb, &data, pos);
2450
2451         /*
2452          * Finally, try again to invalidate clean pages which might have been
2453          * cached by non-direct readahead, or faulted in by get_user_pages()
2454          * if the source of the write was an mmap'ed region of the file
2455          * we're writing.  Either one is a pretty crazy thing to do,
2456          * so we don't support it 100%.  If this invalidation
2457          * fails, tough, the write still worked...
2458          */
2459         if (mapping->nrpages) {
2460                 invalidate_inode_pages2_range(mapping,
2461                                               pos >> PAGE_CACHE_SHIFT, end);
2462         }
2463
2464         if (written > 0) {
2465                 pos += written;
2466                 iov_iter_advance(from, written);
2467                 if (pos > i_size_read(inode) && !S_ISBLK(inode->i_mode)) {
2468                         i_size_write(inode, pos);
2469                         mark_inode_dirty(inode);
2470                 }
2471                 iocb->ki_pos = pos;
2472         }
2473 out:
2474         return written;
2475 }
2476 EXPORT_SYMBOL(generic_file_direct_write);
2477
2478 /*
2479  * Find or create a page at the given pagecache position. Return the locked
2480  * page. This function is specifically for buffered writes.
2481  */
2482 struct page *grab_cache_page_write_begin(struct address_space *mapping,
2483                                         pgoff_t index, unsigned flags)
2484 {
2485         struct page *page;
2486         int fgp_flags = FGP_LOCK|FGP_ACCESSED|FGP_WRITE|FGP_CREAT;
2487
2488         if (flags & AOP_FLAG_NOFS)
2489                 fgp_flags |= FGP_NOFS;
2490
2491         page = pagecache_get_page(mapping, index, fgp_flags,
2492                         mapping_gfp_mask(mapping));
2493         if (page)
2494                 wait_for_stable_page(page);
2495
2496         return page;
2497 }
2498 EXPORT_SYMBOL(grab_cache_page_write_begin);
2499
2500 ssize_t generic_perform_write(struct file *file,
2501                                 struct iov_iter *i, loff_t pos)
2502 {
2503         struct address_space *mapping = file->f_mapping;
2504         const struct address_space_operations *a_ops = mapping->a_ops;
2505         long status = 0;
2506         ssize_t written = 0;
2507         unsigned int flags = 0;
2508
2509         /*
2510          * Copies from kernel address space cannot fail (NFSD is a big user).
2511          */
2512         if (!iter_is_iovec(i))
2513                 flags |= AOP_FLAG_UNINTERRUPTIBLE;
2514
2515         do {
2516                 struct page *page;
2517                 unsigned long offset;   /* Offset into pagecache page */
2518                 unsigned long bytes;    /* Bytes to write to page */
2519                 size_t copied;          /* Bytes copied from user */
2520                 void *fsdata;
2521
2522                 offset = (pos & (PAGE_CACHE_SIZE - 1));
2523                 bytes = min_t(unsigned long, PAGE_CACHE_SIZE - offset,
2524                                                 iov_iter_count(i));
2525
2526 again:
2527                 /*
2528                  * Bring in the user page that we will copy from _first_.
2529                  * Otherwise there's a nasty deadlock on copying from the
2530                  * same page as we're writing to, without it being marked
2531                  * up-to-date.
2532                  *
2533                  * Not only is this an optimisation, but it is also required
2534                  * to check that the address is actually valid, when atomic
2535                  * usercopies are used, below.
2536                  */
2537                 if (unlikely(iov_iter_fault_in_readable(i, bytes))) {
2538                         status = -EFAULT;
2539                         break;
2540                 }
2541
2542                 if (fatal_signal_pending(current)) {
2543                         status = -EINTR;
2544                         break;
2545                 }
2546
2547                 status = a_ops->write_begin(file, mapping, pos, bytes, flags,
2548                                                 &page, &fsdata);
2549                 if (unlikely(status < 0))
2550                         break;
2551
2552                 if (mapping_writably_mapped(mapping))
2553                         flush_dcache_page(page);
2554
2555                 copied = iov_iter_copy_from_user_atomic(page, i, offset, bytes);
2556                 flush_dcache_page(page);
2557
2558                 status = a_ops->write_end(file, mapping, pos, bytes, copied,
2559                                                 page, fsdata);
2560                 if (unlikely(status < 0))
2561                         break;
2562                 copied = status;
2563
2564                 cond_resched();
2565
2566                 iov_iter_advance(i, copied);
2567                 if (unlikely(copied == 0)) {
2568                         /*
2569                          * If we were unable to copy any data at all, we must
2570                          * fall back to a single segment length write.
2571                          *
2572                          * If we didn't fallback here, we could livelock
2573                          * because not all segments in the iov can be copied at
2574                          * once without a pagefault.
2575                          */
2576                         bytes = min_t(unsigned long, PAGE_CACHE_SIZE - offset,
2577                                                 iov_iter_single_seg_count(i));
2578                         goto again;
2579                 }
2580                 pos += copied;
2581                 written += copied;
2582
2583                 balance_dirty_pages_ratelimited(mapping);
2584         } while (iov_iter_count(i));
2585
2586         return written ? written : status;
2587 }
2588 EXPORT_SYMBOL(generic_perform_write);
2589
2590 /**
2591  * __generic_file_write_iter - write data to a file
2592  * @iocb:       IO state structure (file, offset, etc.)
2593  * @from:       iov_iter with data to write
2594  *
2595  * This function does all the work needed for actually writing data to a
2596  * file. It does all basic checks, removes SUID from the file, updates
2597  * modification times and calls proper subroutines depending on whether we
2598  * do direct IO or a standard buffered write.
2599  *
2600  * It expects i_mutex to be grabbed unless we work on a block device or similar
2601  * object which does not need locking at all.
2602  *
2603  * This function does *not* take care of syncing data in case of O_SYNC write.
2604  * A caller has to handle it. This is mainly due to the fact that we want to
2605  * avoid syncing under i_mutex.
2606  */
2607 ssize_t __generic_file_write_iter(struct kiocb *iocb, struct iov_iter *from)
2608 {
2609         struct file *file = iocb->ki_filp;
2610         struct address_space * mapping = file->f_mapping;
2611         struct inode    *inode = mapping->host;
2612         ssize_t         written = 0;
2613         ssize_t         err;
2614         ssize_t         status;
2615
2616         /* We can write back this queue in page reclaim */
2617         current->backing_dev_info = inode_to_bdi(inode);
2618         err = file_remove_privs(file);
2619         if (err)
2620                 goto out;
2621
2622         err = file_update_time(file);
2623         if (err)
2624                 goto out;
2625
2626         if (iocb->ki_flags & IOCB_DIRECT) {
2627                 loff_t pos, endbyte;
2628
2629                 written = generic_file_direct_write(iocb, from, iocb->ki_pos);
2630                 /*
2631                  * If the write stopped short of completing, fall back to
2632                  * buffered writes.  Some filesystems do this for writes to
2633                  * holes, for example.  For DAX files, a buffered write will
2634                  * not succeed (even if it did, DAX does not handle dirty
2635                  * page-cache pages correctly).
2636                  */
2637                 if (written < 0 || !iov_iter_count(from) || IS_DAX(inode))
2638                         goto out;
2639
2640                 status = generic_perform_write(file, from, pos = iocb->ki_pos);
2641                 /*
2642                  * If generic_perform_write() returned a synchronous error
2643                  * then we want to return the number of bytes which were
2644                  * direct-written, or the error code if that was zero.  Note
2645                  * that this differs from normal direct-io semantics, which
2646                  * will return -EFOO even if some bytes were written.
2647                  */
2648                 if (unlikely(status < 0)) {
2649                         err = status;
2650                         goto out;
2651                 }
2652                 /*
2653                  * We need to ensure that the page cache pages are written to
2654                  * disk and invalidated to preserve the expected O_DIRECT
2655                  * semantics.
2656                  */
2657                 endbyte = pos + status - 1;
2658                 err = filemap_write_and_wait_range(mapping, pos, endbyte);
2659                 if (err == 0) {
2660                         iocb->ki_pos = endbyte + 1;
2661                         written += status;
2662                         invalidate_mapping_pages(mapping,
2663                                                  pos >> PAGE_CACHE_SHIFT,
2664                                                  endbyte >> PAGE_CACHE_SHIFT);
2665                 } else {
2666                         /*
2667                          * We don't know how much we wrote, so just return
2668                          * the number of bytes which were direct-written
2669                          */
2670                 }
2671         } else {
2672                 written = generic_perform_write(file, from, iocb->ki_pos);
2673                 if (likely(written > 0))
2674                         iocb->ki_pos += written;
2675         }
2676 out:
2677         current->backing_dev_info = NULL;
2678         return written ? written : err;
2679 }
2680 EXPORT_SYMBOL(__generic_file_write_iter);
2681
2682 /**
2683  * generic_file_write_iter - write data to a file
2684  * @iocb:       IO state structure
2685  * @from:       iov_iter with data to write
2686  *
2687  * This is a wrapper around __generic_file_write_iter() to be used by most
2688  * filesystems. It takes care of syncing the file in case of O_SYNC file
2689  * and acquires i_mutex as needed.
2690  */
2691 ssize_t generic_file_write_iter(struct kiocb *iocb, struct iov_iter *from)
2692 {
2693         struct file *file = iocb->ki_filp;
2694         struct inode *inode = file->f_mapping->host;
2695         ssize_t ret;
2696
2697         mutex_lock(&inode->i_mutex);
2698         ret = generic_write_checks(iocb, from);
2699         if (ret > 0)
2700                 ret = __generic_file_write_iter(iocb, from);
2701         mutex_unlock(&inode->i_mutex);
2702
2703         if (ret > 0) {
2704                 ssize_t err;
2705
2706                 err = generic_write_sync(file, iocb->ki_pos - ret, ret);
2707                 if (err < 0)
2708                         ret = err;
2709         }
2710         return ret;
2711 }
2712 EXPORT_SYMBOL(generic_file_write_iter);
2713
2714 /**
2715  * try_to_release_page() - release old fs-specific metadata on a page
2716  *
2717  * @page: the page which the kernel is trying to free
2718  * @gfp_mask: memory allocation flags (and I/O mode)
2719  *
2720  * The address_space is to try to release any data against the page
2721  * (presumably at page->private).  If the release was successful, return `1'.
2722  * Otherwise return zero.
2723  *
2724  * This may also be called if PG_fscache is set on a page, indicating that the
2725  * page is known to the local caching routines.
2726  *
2727  * The @gfp_mask argument specifies whether I/O may be performed to release
2728  * this page (__GFP_IO), and whether the call may block (__GFP_RECLAIM & __GFP_FS).
2729  *
2730  */
2731 int try_to_release_page(struct page *page, gfp_t gfp_mask)
2732 {
2733         struct address_space * const mapping = page->mapping;
2734
2735         BUG_ON(!PageLocked(page));
2736         if (PageWriteback(page))
2737                 return 0;
2738
2739         if (mapping && mapping->a_ops->releasepage)
2740                 return mapping->a_ops->releasepage(page, gfp_mask);
2741         return try_to_free_buffers(page);
2742 }
2743
2744 EXPORT_SYMBOL(try_to_release_page);