7997adce5a29a2a515107c06d1ecadefd2151472
[muen/linux.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/dax.h>
15 #include <linux/fs.h>
16 #include <linux/sched/signal.h>
17 #include <linux/uaccess.h>
18 #include <linux/capability.h>
19 #include <linux/kernel_stat.h>
20 #include <linux/gfp.h>
21 #include <linux/mm.h>
22 #include <linux/swap.h>
23 #include <linux/mman.h>
24 #include <linux/pagemap.h>
25 #include <linux/file.h>
26 #include <linux/uio.h>
27 #include <linux/hash.h>
28 #include <linux/writeback.h>
29 #include <linux/backing-dev.h>
30 #include <linux/pagevec.h>
31 #include <linux/blkdev.h>
32 #include <linux/security.h>
33 #include <linux/cpuset.h>
34 #include <linux/hugetlb.h>
35 #include <linux/memcontrol.h>
36 #include <linux/cleancache.h>
37 #include <linux/shmem_fs.h>
38 #include <linux/rmap.h>
39 #include "internal.h"
40
41 #define CREATE_TRACE_POINTS
42 #include <trace/events/filemap.h>
43
44 /*
45  * FIXME: remove all knowledge of the buffer layer from the core VM
46  */
47 #include <linux/buffer_head.h> /* for try_to_free_buffers */
48
49 #include <asm/mman.h>
50
51 /*
52  * Shared mappings implemented 30.11.1994. It's not fully working yet,
53  * though.
54  *
55  * Shared mappings now work. 15.8.1995  Bruno.
56  *
57  * finished 'unifying' the page and buffer cache and SMP-threaded the
58  * page-cache, 21.05.1999, Ingo Molnar <mingo@redhat.com>
59  *
60  * SMP-threaded pagemap-LRU 1999, Andrea Arcangeli <andrea@suse.de>
61  */
62
63 /*
64  * Lock ordering:
65  *
66  *  ->i_mmap_rwsem              (truncate_pagecache)
67  *    ->private_lock            (__free_pte->__set_page_dirty_buffers)
68  *      ->swap_lock             (exclusive_swap_page, others)
69  *        ->i_pages lock
70  *
71  *  ->i_mutex
72  *    ->i_mmap_rwsem            (truncate->unmap_mapping_range)
73  *
74  *  ->mmap_sem
75  *    ->i_mmap_rwsem
76  *      ->page_table_lock or pte_lock   (various, mainly in memory.c)
77  *        ->i_pages lock        (arch-dependent flush_dcache_mmap_lock)
78  *
79  *  ->mmap_sem
80  *    ->lock_page               (access_process_vm)
81  *
82  *  ->i_mutex                   (generic_perform_write)
83  *    ->mmap_sem                (fault_in_pages_readable->do_page_fault)
84  *
85  *  bdi->wb.list_lock
86  *    sb_lock                   (fs/fs-writeback.c)
87  *    ->i_pages lock            (__sync_single_inode)
88  *
89  *  ->i_mmap_rwsem
90  *    ->anon_vma.lock           (vma_adjust)
91  *
92  *  ->anon_vma.lock
93  *    ->page_table_lock or pte_lock     (anon_vma_prepare and various)
94  *
95  *  ->page_table_lock or pte_lock
96  *    ->swap_lock               (try_to_unmap_one)
97  *    ->private_lock            (try_to_unmap_one)
98  *    ->i_pages lock            (try_to_unmap_one)
99  *    ->zone_lru_lock(zone)     (follow_page->mark_page_accessed)
100  *    ->zone_lru_lock(zone)     (check_pte_range->isolate_lru_page)
101  *    ->private_lock            (page_remove_rmap->set_page_dirty)
102  *    ->i_pages lock            (page_remove_rmap->set_page_dirty)
103  *    bdi.wb->list_lock         (page_remove_rmap->set_page_dirty)
104  *    ->inode->i_lock           (page_remove_rmap->set_page_dirty)
105  *    ->memcg->move_lock        (page_remove_rmap->lock_page_memcg)
106  *    bdi.wb->list_lock         (zap_pte_range->set_page_dirty)
107  *    ->inode->i_lock           (zap_pte_range->set_page_dirty)
108  *    ->private_lock            (zap_pte_range->__set_page_dirty_buffers)
109  *
110  * ->i_mmap_rwsem
111  *   ->tasklist_lock            (memory_failure, collect_procs_ao)
112  */
113
114 static int page_cache_tree_insert(struct address_space *mapping,
115                                   struct page *page, void **shadowp)
116 {
117         struct radix_tree_node *node;
118         void **slot;
119         int error;
120
121         error = __radix_tree_create(&mapping->i_pages, page->index, 0,
122                                     &node, &slot);
123         if (error)
124                 return error;
125         if (*slot) {
126                 void *p;
127
128                 p = radix_tree_deref_slot_protected(slot,
129                                                     &mapping->i_pages.xa_lock);
130                 if (!radix_tree_exceptional_entry(p))
131                         return -EEXIST;
132
133                 mapping->nrexceptional--;
134                 if (shadowp)
135                         *shadowp = p;
136         }
137         __radix_tree_replace(&mapping->i_pages, node, slot, page,
138                              workingset_lookup_update(mapping));
139         mapping->nrpages++;
140         return 0;
141 }
142
143 static void page_cache_tree_delete(struct address_space *mapping,
144                                    struct page *page, void *shadow)
145 {
146         int i, nr;
147
148         /* hugetlb pages are represented by one entry in the radix tree */
149         nr = PageHuge(page) ? 1 : hpage_nr_pages(page);
150
151         VM_BUG_ON_PAGE(!PageLocked(page), page);
152         VM_BUG_ON_PAGE(PageTail(page), page);
153         VM_BUG_ON_PAGE(nr != 1 && shadow, page);
154
155         for (i = 0; i < nr; i++) {
156                 struct radix_tree_node *node;
157                 void **slot;
158
159                 __radix_tree_lookup(&mapping->i_pages, page->index + i,
160                                     &node, &slot);
161
162                 VM_BUG_ON_PAGE(!node && nr != 1, page);
163
164                 radix_tree_clear_tags(&mapping->i_pages, node, slot);
165                 __radix_tree_replace(&mapping->i_pages, node, slot, shadow,
166                                 workingset_lookup_update(mapping));
167         }
168
169         page->mapping = NULL;
170         /* Leave page->index set: truncation lookup relies upon it */
171
172         if (shadow) {
173                 mapping->nrexceptional += nr;
174                 /*
175                  * Make sure the nrexceptional update is committed before
176                  * the nrpages update so that final truncate racing
177                  * with reclaim does not see both counters 0 at the
178                  * same time and miss a shadow entry.
179                  */
180                 smp_wmb();
181         }
182         mapping->nrpages -= nr;
183 }
184
185 static void unaccount_page_cache_page(struct address_space *mapping,
186                                       struct page *page)
187 {
188         int nr;
189
190         /*
191          * if we're uptodate, flush out into the cleancache, otherwise
192          * invalidate any existing cleancache entries.  We can't leave
193          * stale data around in the cleancache once our page is gone
194          */
195         if (PageUptodate(page) && PageMappedToDisk(page))
196                 cleancache_put_page(page);
197         else
198                 cleancache_invalidate_page(mapping, page);
199
200         VM_BUG_ON_PAGE(PageTail(page), page);
201         VM_BUG_ON_PAGE(page_mapped(page), page);
202         if (!IS_ENABLED(CONFIG_DEBUG_VM) && unlikely(page_mapped(page))) {
203                 int mapcount;
204
205                 pr_alert("BUG: Bad page cache in process %s  pfn:%05lx\n",
206                          current->comm, page_to_pfn(page));
207                 dump_page(page, "still mapped when deleted");
208                 dump_stack();
209                 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
210
211                 mapcount = page_mapcount(page);
212                 if (mapping_exiting(mapping) &&
213                     page_count(page) >= mapcount + 2) {
214                         /*
215                          * All vmas have already been torn down, so it's
216                          * a good bet that actually the page is unmapped,
217                          * and we'd prefer not to leak it: if we're wrong,
218                          * some other bad page check should catch it later.
219                          */
220                         page_mapcount_reset(page);
221                         page_ref_sub(page, mapcount);
222                 }
223         }
224
225         /* hugetlb pages do not participate in page cache accounting. */
226         if (PageHuge(page))
227                 return;
228
229         nr = hpage_nr_pages(page);
230
231         __mod_node_page_state(page_pgdat(page), NR_FILE_PAGES, -nr);
232         if (PageSwapBacked(page)) {
233                 __mod_node_page_state(page_pgdat(page), NR_SHMEM, -nr);
234                 if (PageTransHuge(page))
235                         __dec_node_page_state(page, NR_SHMEM_THPS);
236         } else {
237                 VM_BUG_ON_PAGE(PageTransHuge(page), page);
238         }
239
240         /*
241          * At this point page must be either written or cleaned by
242          * truncate.  Dirty page here signals a bug and loss of
243          * unwritten data.
244          *
245          * This fixes dirty accounting after removing the page entirely
246          * but leaves PageDirty set: it has no effect for truncated
247          * page and anyway will be cleared before returning page into
248          * buddy allocator.
249          */
250         if (WARN_ON_ONCE(PageDirty(page)))
251                 account_page_cleaned(page, mapping, inode_to_wb(mapping->host));
252 }
253
254 /*
255  * Delete a page from the page cache and free it. Caller has to make
256  * sure the page is locked and that nobody else uses it - or that usage
257  * is safe.  The caller must hold the i_pages lock.
258  */
259 void __delete_from_page_cache(struct page *page, void *shadow)
260 {
261         struct address_space *mapping = page->mapping;
262
263         trace_mm_filemap_delete_from_page_cache(page);
264
265         unaccount_page_cache_page(mapping, page);
266         page_cache_tree_delete(mapping, page, shadow);
267 }
268
269 static void page_cache_free_page(struct address_space *mapping,
270                                 struct page *page)
271 {
272         void (*freepage)(struct page *);
273
274         freepage = mapping->a_ops->freepage;
275         if (freepage)
276                 freepage(page);
277
278         if (PageTransHuge(page) && !PageHuge(page)) {
279                 page_ref_sub(page, HPAGE_PMD_NR);
280                 VM_BUG_ON_PAGE(page_count(page) <= 0, page);
281         } else {
282                 put_page(page);
283         }
284 }
285
286 /**
287  * delete_from_page_cache - delete page from page cache
288  * @page: the page which the kernel is trying to remove from page cache
289  *
290  * This must be called only on pages that have been verified to be in the page
291  * cache and locked.  It will never put the page into the free list, the caller
292  * has a reference on the page.
293  */
294 void delete_from_page_cache(struct page *page)
295 {
296         struct address_space *mapping = page_mapping(page);
297         unsigned long flags;
298
299         BUG_ON(!PageLocked(page));
300         xa_lock_irqsave(&mapping->i_pages, flags);
301         __delete_from_page_cache(page, NULL);
302         xa_unlock_irqrestore(&mapping->i_pages, flags);
303
304         page_cache_free_page(mapping, page);
305 }
306 EXPORT_SYMBOL(delete_from_page_cache);
307
308 /*
309  * page_cache_tree_delete_batch - delete several pages from page cache
310  * @mapping: the mapping to which pages belong
311  * @pvec: pagevec with pages to delete
312  *
313  * The function walks over mapping->i_pages and removes pages passed in @pvec
314  * from the mapping. The function expects @pvec to be sorted by page index.
315  * It tolerates holes in @pvec (mapping entries at those indices are not
316  * modified). The function expects only THP head pages to be present in the
317  * @pvec and takes care to delete all corresponding tail pages from the
318  * mapping as well.
319  *
320  * The function expects the i_pages lock to be held.
321  */
322 static void
323 page_cache_tree_delete_batch(struct address_space *mapping,
324                              struct pagevec *pvec)
325 {
326         struct radix_tree_iter iter;
327         void **slot;
328         int total_pages = 0;
329         int i = 0, tail_pages = 0;
330         struct page *page;
331         pgoff_t start;
332
333         start = pvec->pages[0]->index;
334         radix_tree_for_each_slot(slot, &mapping->i_pages, &iter, start) {
335                 if (i >= pagevec_count(pvec) && !tail_pages)
336                         break;
337                 page = radix_tree_deref_slot_protected(slot,
338                                                        &mapping->i_pages.xa_lock);
339                 if (radix_tree_exceptional_entry(page))
340                         continue;
341                 if (!tail_pages) {
342                         /*
343                          * Some page got inserted in our range? Skip it. We
344                          * have our pages locked so they are protected from
345                          * being removed.
346                          */
347                         if (page != pvec->pages[i])
348                                 continue;
349                         WARN_ON_ONCE(!PageLocked(page));
350                         if (PageTransHuge(page) && !PageHuge(page))
351                                 tail_pages = HPAGE_PMD_NR - 1;
352                         page->mapping = NULL;
353                         /*
354                          * Leave page->index set: truncation lookup relies
355                          * upon it
356                          */
357                         i++;
358                 } else {
359                         tail_pages--;
360                 }
361                 radix_tree_clear_tags(&mapping->i_pages, iter.node, slot);
362                 __radix_tree_replace(&mapping->i_pages, iter.node, slot, NULL,
363                                 workingset_lookup_update(mapping));
364                 total_pages++;
365         }
366         mapping->nrpages -= total_pages;
367 }
368
369 void delete_from_page_cache_batch(struct address_space *mapping,
370                                   struct pagevec *pvec)
371 {
372         int i;
373         unsigned long flags;
374
375         if (!pagevec_count(pvec))
376                 return;
377
378         xa_lock_irqsave(&mapping->i_pages, flags);
379         for (i = 0; i < pagevec_count(pvec); i++) {
380                 trace_mm_filemap_delete_from_page_cache(pvec->pages[i]);
381
382                 unaccount_page_cache_page(mapping, pvec->pages[i]);
383         }
384         page_cache_tree_delete_batch(mapping, pvec);
385         xa_unlock_irqrestore(&mapping->i_pages, flags);
386
387         for (i = 0; i < pagevec_count(pvec); i++)
388                 page_cache_free_page(mapping, pvec->pages[i]);
389 }
390
391 int filemap_check_errors(struct address_space *mapping)
392 {
393         int ret = 0;
394         /* Check for outstanding write errors */
395         if (test_bit(AS_ENOSPC, &mapping->flags) &&
396             test_and_clear_bit(AS_ENOSPC, &mapping->flags))
397                 ret = -ENOSPC;
398         if (test_bit(AS_EIO, &mapping->flags) &&
399             test_and_clear_bit(AS_EIO, &mapping->flags))
400                 ret = -EIO;
401         return ret;
402 }
403 EXPORT_SYMBOL(filemap_check_errors);
404
405 static int filemap_check_and_keep_errors(struct address_space *mapping)
406 {
407         /* Check for outstanding write errors */
408         if (test_bit(AS_EIO, &mapping->flags))
409                 return -EIO;
410         if (test_bit(AS_ENOSPC, &mapping->flags))
411                 return -ENOSPC;
412         return 0;
413 }
414
415 /**
416  * __filemap_fdatawrite_range - start writeback on mapping dirty pages in range
417  * @mapping:    address space structure to write
418  * @start:      offset in bytes where the range starts
419  * @end:        offset in bytes where the range ends (inclusive)
420  * @sync_mode:  enable synchronous operation
421  *
422  * Start writeback against all of a mapping's dirty pages that lie
423  * within the byte offsets <start, end> inclusive.
424  *
425  * If sync_mode is WB_SYNC_ALL then this is a "data integrity" operation, as
426  * opposed to a regular memory cleansing writeback.  The difference between
427  * these two operations is that if a dirty page/buffer is encountered, it must
428  * be waited upon, and not just skipped over.
429  */
430 int __filemap_fdatawrite_range(struct address_space *mapping, loff_t start,
431                                 loff_t end, int sync_mode)
432 {
433         int ret;
434         struct writeback_control wbc = {
435                 .sync_mode = sync_mode,
436                 .nr_to_write = LONG_MAX,
437                 .range_start = start,
438                 .range_end = end,
439         };
440
441         if (!mapping_cap_writeback_dirty(mapping))
442                 return 0;
443
444         wbc_attach_fdatawrite_inode(&wbc, mapping->host);
445         ret = do_writepages(mapping, &wbc);
446         wbc_detach_inode(&wbc);
447         return ret;
448 }
449
450 static inline int __filemap_fdatawrite(struct address_space *mapping,
451         int sync_mode)
452 {
453         return __filemap_fdatawrite_range(mapping, 0, LLONG_MAX, sync_mode);
454 }
455
456 int filemap_fdatawrite(struct address_space *mapping)
457 {
458         return __filemap_fdatawrite(mapping, WB_SYNC_ALL);
459 }
460 EXPORT_SYMBOL(filemap_fdatawrite);
461
462 int filemap_fdatawrite_range(struct address_space *mapping, loff_t start,
463                                 loff_t end)
464 {
465         return __filemap_fdatawrite_range(mapping, start, end, WB_SYNC_ALL);
466 }
467 EXPORT_SYMBOL(filemap_fdatawrite_range);
468
469 /**
470  * filemap_flush - mostly a non-blocking flush
471  * @mapping:    target address_space
472  *
473  * This is a mostly non-blocking flush.  Not suitable for data-integrity
474  * purposes - I/O may not be started against all dirty pages.
475  */
476 int filemap_flush(struct address_space *mapping)
477 {
478         return __filemap_fdatawrite(mapping, WB_SYNC_NONE);
479 }
480 EXPORT_SYMBOL(filemap_flush);
481
482 /**
483  * filemap_range_has_page - check if a page exists in range.
484  * @mapping:           address space within which to check
485  * @start_byte:        offset in bytes where the range starts
486  * @end_byte:          offset in bytes where the range ends (inclusive)
487  *
488  * Find at least one page in the range supplied, usually used to check if
489  * direct writing in this range will trigger a writeback.
490  */
491 bool filemap_range_has_page(struct address_space *mapping,
492                            loff_t start_byte, loff_t end_byte)
493 {
494         pgoff_t index = start_byte >> PAGE_SHIFT;
495         pgoff_t end = end_byte >> PAGE_SHIFT;
496         struct page *page;
497
498         if (end_byte < start_byte)
499                 return false;
500
501         if (mapping->nrpages == 0)
502                 return false;
503
504         if (!find_get_pages_range(mapping, &index, end, 1, &page))
505                 return false;
506         put_page(page);
507         return true;
508 }
509 EXPORT_SYMBOL(filemap_range_has_page);
510
511 static void __filemap_fdatawait_range(struct address_space *mapping,
512                                      loff_t start_byte, loff_t end_byte)
513 {
514         pgoff_t index = start_byte >> PAGE_SHIFT;
515         pgoff_t end = end_byte >> PAGE_SHIFT;
516         struct pagevec pvec;
517         int nr_pages;
518
519         if (end_byte < start_byte)
520                 return;
521
522         pagevec_init(&pvec);
523         while (index <= end) {
524                 unsigned i;
525
526                 nr_pages = pagevec_lookup_range_tag(&pvec, mapping, &index,
527                                 end, PAGECACHE_TAG_WRITEBACK);
528                 if (!nr_pages)
529                         break;
530
531                 for (i = 0; i < nr_pages; i++) {
532                         struct page *page = pvec.pages[i];
533
534                         wait_on_page_writeback(page);
535                         ClearPageError(page);
536                 }
537                 pagevec_release(&pvec);
538                 cond_resched();
539         }
540 }
541
542 /**
543  * filemap_fdatawait_range - wait for writeback to complete
544  * @mapping:            address space structure to wait for
545  * @start_byte:         offset in bytes where the range starts
546  * @end_byte:           offset in bytes where the range ends (inclusive)
547  *
548  * Walk the list of under-writeback pages of the given address space
549  * in the given range and wait for all of them.  Check error status of
550  * the address space and return it.
551  *
552  * Since the error status of the address space is cleared by this function,
553  * callers are responsible for checking the return value and handling and/or
554  * reporting the error.
555  */
556 int filemap_fdatawait_range(struct address_space *mapping, loff_t start_byte,
557                             loff_t end_byte)
558 {
559         __filemap_fdatawait_range(mapping, start_byte, end_byte);
560         return filemap_check_errors(mapping);
561 }
562 EXPORT_SYMBOL(filemap_fdatawait_range);
563
564 /**
565  * file_fdatawait_range - wait for writeback to complete
566  * @file:               file pointing to address space structure to wait for
567  * @start_byte:         offset in bytes where the range starts
568  * @end_byte:           offset in bytes where the range ends (inclusive)
569  *
570  * Walk the list of under-writeback pages of the address space that file
571  * refers to, in the given range and wait for all of them.  Check error
572  * status of the address space vs. the file->f_wb_err cursor and return it.
573  *
574  * Since the error status of the file is advanced by this function,
575  * callers are responsible for checking the return value and handling and/or
576  * reporting the error.
577  */
578 int file_fdatawait_range(struct file *file, loff_t start_byte, loff_t end_byte)
579 {
580         struct address_space *mapping = file->f_mapping;
581
582         __filemap_fdatawait_range(mapping, start_byte, end_byte);
583         return file_check_and_advance_wb_err(file);
584 }
585 EXPORT_SYMBOL(file_fdatawait_range);
586
587 /**
588  * filemap_fdatawait_keep_errors - wait for writeback without clearing errors
589  * @mapping: address space structure to wait for
590  *
591  * Walk the list of under-writeback pages of the given address space
592  * and wait for all of them.  Unlike filemap_fdatawait(), this function
593  * does not clear error status of the address space.
594  *
595  * Use this function if callers don't handle errors themselves.  Expected
596  * call sites are system-wide / filesystem-wide data flushers: e.g. sync(2),
597  * fsfreeze(8)
598  */
599 int filemap_fdatawait_keep_errors(struct address_space *mapping)
600 {
601         __filemap_fdatawait_range(mapping, 0, LLONG_MAX);
602         return filemap_check_and_keep_errors(mapping);
603 }
604 EXPORT_SYMBOL(filemap_fdatawait_keep_errors);
605
606 static bool mapping_needs_writeback(struct address_space *mapping)
607 {
608         return (!dax_mapping(mapping) && mapping->nrpages) ||
609             (dax_mapping(mapping) && mapping->nrexceptional);
610 }
611
612 int filemap_write_and_wait(struct address_space *mapping)
613 {
614         int err = 0;
615
616         if (mapping_needs_writeback(mapping)) {
617                 err = filemap_fdatawrite(mapping);
618                 /*
619                  * Even if the above returned error, the pages may be
620                  * written partially (e.g. -ENOSPC), so we wait for it.
621                  * But the -EIO is special case, it may indicate the worst
622                  * thing (e.g. bug) happened, so we avoid waiting for it.
623                  */
624                 if (err != -EIO) {
625                         int err2 = filemap_fdatawait(mapping);
626                         if (!err)
627                                 err = err2;
628                 } else {
629                         /* Clear any previously stored errors */
630                         filemap_check_errors(mapping);
631                 }
632         } else {
633                 err = filemap_check_errors(mapping);
634         }
635         return err;
636 }
637 EXPORT_SYMBOL(filemap_write_and_wait);
638
639 /**
640  * filemap_write_and_wait_range - write out & wait on a file range
641  * @mapping:    the address_space for the pages
642  * @lstart:     offset in bytes where the range starts
643  * @lend:       offset in bytes where the range ends (inclusive)
644  *
645  * Write out and wait upon file offsets lstart->lend, inclusive.
646  *
647  * Note that @lend is inclusive (describes the last byte to be written) so
648  * that this function can be used to write to the very end-of-file (end = -1).
649  */
650 int filemap_write_and_wait_range(struct address_space *mapping,
651                                  loff_t lstart, loff_t lend)
652 {
653         int err = 0;
654
655         if (mapping_needs_writeback(mapping)) {
656                 err = __filemap_fdatawrite_range(mapping, lstart, lend,
657                                                  WB_SYNC_ALL);
658                 /* See comment of filemap_write_and_wait() */
659                 if (err != -EIO) {
660                         int err2 = filemap_fdatawait_range(mapping,
661                                                 lstart, lend);
662                         if (!err)
663                                 err = err2;
664                 } else {
665                         /* Clear any previously stored errors */
666                         filemap_check_errors(mapping);
667                 }
668         } else {
669                 err = filemap_check_errors(mapping);
670         }
671         return err;
672 }
673 EXPORT_SYMBOL(filemap_write_and_wait_range);
674
675 void __filemap_set_wb_err(struct address_space *mapping, int err)
676 {
677         errseq_t eseq = errseq_set(&mapping->wb_err, err);
678
679         trace_filemap_set_wb_err(mapping, eseq);
680 }
681 EXPORT_SYMBOL(__filemap_set_wb_err);
682
683 /**
684  * file_check_and_advance_wb_err - report wb error (if any) that was previously
685  *                                 and advance wb_err to current one
686  * @file: struct file on which the error is being reported
687  *
688  * When userland calls fsync (or something like nfsd does the equivalent), we
689  * want to report any writeback errors that occurred since the last fsync (or
690  * since the file was opened if there haven't been any).
691  *
692  * Grab the wb_err from the mapping. If it matches what we have in the file,
693  * then just quickly return 0. The file is all caught up.
694  *
695  * If it doesn't match, then take the mapping value, set the "seen" flag in
696  * it and try to swap it into place. If it works, or another task beat us
697  * to it with the new value, then update the f_wb_err and return the error
698  * portion. The error at this point must be reported via proper channels
699  * (a'la fsync, or NFS COMMIT operation, etc.).
700  *
701  * While we handle mapping->wb_err with atomic operations, the f_wb_err
702  * value is protected by the f_lock since we must ensure that it reflects
703  * the latest value swapped in for this file descriptor.
704  */
705 int file_check_and_advance_wb_err(struct file *file)
706 {
707         int err = 0;
708         errseq_t old = READ_ONCE(file->f_wb_err);
709         struct address_space *mapping = file->f_mapping;
710
711         /* Locklessly handle the common case where nothing has changed */
712         if (errseq_check(&mapping->wb_err, old)) {
713                 /* Something changed, must use slow path */
714                 spin_lock(&file->f_lock);
715                 old = file->f_wb_err;
716                 err = errseq_check_and_advance(&mapping->wb_err,
717                                                 &file->f_wb_err);
718                 trace_file_check_and_advance_wb_err(file, old);
719                 spin_unlock(&file->f_lock);
720         }
721
722         /*
723          * We're mostly using this function as a drop in replacement for
724          * filemap_check_errors. Clear AS_EIO/AS_ENOSPC to emulate the effect
725          * that the legacy code would have had on these flags.
726          */
727         clear_bit(AS_EIO, &mapping->flags);
728         clear_bit(AS_ENOSPC, &mapping->flags);
729         return err;
730 }
731 EXPORT_SYMBOL(file_check_and_advance_wb_err);
732
733 /**
734  * file_write_and_wait_range - write out & wait on a file range
735  * @file:       file pointing to address_space with pages
736  * @lstart:     offset in bytes where the range starts
737  * @lend:       offset in bytes where the range ends (inclusive)
738  *
739  * Write out and wait upon file offsets lstart->lend, inclusive.
740  *
741  * Note that @lend is inclusive (describes the last byte to be written) so
742  * that this function can be used to write to the very end-of-file (end = -1).
743  *
744  * After writing out and waiting on the data, we check and advance the
745  * f_wb_err cursor to the latest value, and return any errors detected there.
746  */
747 int file_write_and_wait_range(struct file *file, loff_t lstart, loff_t lend)
748 {
749         int err = 0, err2;
750         struct address_space *mapping = file->f_mapping;
751
752         if (mapping_needs_writeback(mapping)) {
753                 err = __filemap_fdatawrite_range(mapping, lstart, lend,
754                                                  WB_SYNC_ALL);
755                 /* See comment of filemap_write_and_wait() */
756                 if (err != -EIO)
757                         __filemap_fdatawait_range(mapping, lstart, lend);
758         }
759         err2 = file_check_and_advance_wb_err(file);
760         if (!err)
761                 err = err2;
762         return err;
763 }
764 EXPORT_SYMBOL(file_write_and_wait_range);
765
766 /**
767  * replace_page_cache_page - replace a pagecache page with a new one
768  * @old:        page to be replaced
769  * @new:        page to replace with
770  * @gfp_mask:   allocation mode
771  *
772  * This function replaces a page in the pagecache with a new one.  On
773  * success it acquires the pagecache reference for the new page and
774  * drops it for the old page.  Both the old and new pages must be
775  * locked.  This function does not add the new page to the LRU, the
776  * caller must do that.
777  *
778  * The remove + add is atomic.  The only way this function can fail is
779  * memory allocation failure.
780  */
781 int replace_page_cache_page(struct page *old, struct page *new, gfp_t gfp_mask)
782 {
783         int error;
784
785         VM_BUG_ON_PAGE(!PageLocked(old), old);
786         VM_BUG_ON_PAGE(!PageLocked(new), new);
787         VM_BUG_ON_PAGE(new->mapping, new);
788
789         error = radix_tree_preload(gfp_mask & GFP_RECLAIM_MASK);
790         if (!error) {
791                 struct address_space *mapping = old->mapping;
792                 void (*freepage)(struct page *);
793                 unsigned long flags;
794
795                 pgoff_t offset = old->index;
796                 freepage = mapping->a_ops->freepage;
797
798                 get_page(new);
799                 new->mapping = mapping;
800                 new->index = offset;
801
802                 xa_lock_irqsave(&mapping->i_pages, flags);
803                 __delete_from_page_cache(old, NULL);
804                 error = page_cache_tree_insert(mapping, new, NULL);
805                 BUG_ON(error);
806
807                 /*
808                  * hugetlb pages do not participate in page cache accounting.
809                  */
810                 if (!PageHuge(new))
811                         __inc_node_page_state(new, NR_FILE_PAGES);
812                 if (PageSwapBacked(new))
813                         __inc_node_page_state(new, NR_SHMEM);
814                 xa_unlock_irqrestore(&mapping->i_pages, flags);
815                 mem_cgroup_migrate(old, new);
816                 radix_tree_preload_end();
817                 if (freepage)
818                         freepage(old);
819                 put_page(old);
820         }
821
822         return error;
823 }
824 EXPORT_SYMBOL_GPL(replace_page_cache_page);
825
826 static int __add_to_page_cache_locked(struct page *page,
827                                       struct address_space *mapping,
828                                       pgoff_t offset, gfp_t gfp_mask,
829                                       void **shadowp)
830 {
831         int huge = PageHuge(page);
832         struct mem_cgroup *memcg;
833         int error;
834
835         VM_BUG_ON_PAGE(!PageLocked(page), page);
836         VM_BUG_ON_PAGE(PageSwapBacked(page), page);
837
838         if (!huge) {
839                 error = mem_cgroup_try_charge(page, current->mm,
840                                               gfp_mask, &memcg, false);
841                 if (error)
842                         return error;
843         }
844
845         error = radix_tree_maybe_preload(gfp_mask & GFP_RECLAIM_MASK);
846         if (error) {
847                 if (!huge)
848                         mem_cgroup_cancel_charge(page, memcg, false);
849                 return error;
850         }
851
852         get_page(page);
853         page->mapping = mapping;
854         page->index = offset;
855
856         xa_lock_irq(&mapping->i_pages);
857         error = page_cache_tree_insert(mapping, page, shadowp);
858         radix_tree_preload_end();
859         if (unlikely(error))
860                 goto err_insert;
861
862         /* hugetlb pages do not participate in page cache accounting. */
863         if (!huge)
864                 __inc_node_page_state(page, NR_FILE_PAGES);
865         xa_unlock_irq(&mapping->i_pages);
866         if (!huge)
867                 mem_cgroup_commit_charge(page, memcg, false, false);
868         trace_mm_filemap_add_to_page_cache(page);
869         return 0;
870 err_insert:
871         page->mapping = NULL;
872         /* Leave page->index set: truncation relies upon it */
873         xa_unlock_irq(&mapping->i_pages);
874         if (!huge)
875                 mem_cgroup_cancel_charge(page, memcg, false);
876         put_page(page);
877         return error;
878 }
879
880 /**
881  * add_to_page_cache_locked - add a locked page to the pagecache
882  * @page:       page to add
883  * @mapping:    the page's address_space
884  * @offset:     page index
885  * @gfp_mask:   page allocation mode
886  *
887  * This function is used to add a page to the pagecache. It must be locked.
888  * This function does not add the page to the LRU.  The caller must do that.
889  */
890 int add_to_page_cache_locked(struct page *page, struct address_space *mapping,
891                 pgoff_t offset, gfp_t gfp_mask)
892 {
893         return __add_to_page_cache_locked(page, mapping, offset,
894                                           gfp_mask, NULL);
895 }
896 EXPORT_SYMBOL(add_to_page_cache_locked);
897
898 int add_to_page_cache_lru(struct page *page, struct address_space *mapping,
899                                 pgoff_t offset, gfp_t gfp_mask)
900 {
901         void *shadow = NULL;
902         int ret;
903
904         __SetPageLocked(page);
905         ret = __add_to_page_cache_locked(page, mapping, offset,
906                                          gfp_mask, &shadow);
907         if (unlikely(ret))
908                 __ClearPageLocked(page);
909         else {
910                 /*
911                  * The page might have been evicted from cache only
912                  * recently, in which case it should be activated like
913                  * any other repeatedly accessed page.
914                  * The exception is pages getting rewritten; evicting other
915                  * data from the working set, only to cache data that will
916                  * get overwritten with something else, is a waste of memory.
917                  */
918                 WARN_ON_ONCE(PageActive(page));
919                 if (!(gfp_mask & __GFP_WRITE) && shadow)
920                         workingset_refault(page, shadow);
921                 lru_cache_add(page);
922         }
923         return ret;
924 }
925 EXPORT_SYMBOL_GPL(add_to_page_cache_lru);
926
927 #ifdef CONFIG_NUMA
928 struct page *__page_cache_alloc(gfp_t gfp)
929 {
930         int n;
931         struct page *page;
932
933         if (cpuset_do_page_mem_spread()) {
934                 unsigned int cpuset_mems_cookie;
935                 do {
936                         cpuset_mems_cookie = read_mems_allowed_begin();
937                         n = cpuset_mem_spread_node();
938                         page = __alloc_pages_node(n, gfp, 0);
939                 } while (!page && read_mems_allowed_retry(cpuset_mems_cookie));
940
941                 return page;
942         }
943         return alloc_pages(gfp, 0);
944 }
945 EXPORT_SYMBOL(__page_cache_alloc);
946 #endif
947
948 /*
949  * In order to wait for pages to become available there must be
950  * waitqueues associated with pages. By using a hash table of
951  * waitqueues where the bucket discipline is to maintain all
952  * waiters on the same queue and wake all when any of the pages
953  * become available, and for the woken contexts to check to be
954  * sure the appropriate page became available, this saves space
955  * at a cost of "thundering herd" phenomena during rare hash
956  * collisions.
957  */
958 #define PAGE_WAIT_TABLE_BITS 8
959 #define PAGE_WAIT_TABLE_SIZE (1 << PAGE_WAIT_TABLE_BITS)
960 static wait_queue_head_t page_wait_table[PAGE_WAIT_TABLE_SIZE] __cacheline_aligned;
961
962 static wait_queue_head_t *page_waitqueue(struct page *page)
963 {
964         return &page_wait_table[hash_ptr(page, PAGE_WAIT_TABLE_BITS)];
965 }
966
967 void __init pagecache_init(void)
968 {
969         int i;
970
971         for (i = 0; i < PAGE_WAIT_TABLE_SIZE; i++)
972                 init_waitqueue_head(&page_wait_table[i]);
973
974         page_writeback_init();
975 }
976
977 /* This has the same layout as wait_bit_key - see fs/cachefiles/rdwr.c */
978 struct wait_page_key {
979         struct page *page;
980         int bit_nr;
981         int page_match;
982 };
983
984 struct wait_page_queue {
985         struct page *page;
986         int bit_nr;
987         wait_queue_entry_t wait;
988 };
989
990 static int wake_page_function(wait_queue_entry_t *wait, unsigned mode, int sync, void *arg)
991 {
992         struct wait_page_key *key = arg;
993         struct wait_page_queue *wait_page
994                 = container_of(wait, struct wait_page_queue, wait);
995
996         if (wait_page->page != key->page)
997                return 0;
998         key->page_match = 1;
999
1000         if (wait_page->bit_nr != key->bit_nr)
1001                 return 0;
1002
1003         /* Stop walking if it's locked */
1004         if (test_bit(key->bit_nr, &key->page->flags))
1005                 return -1;
1006
1007         return autoremove_wake_function(wait, mode, sync, key);
1008 }
1009
1010 static void wake_up_page_bit(struct page *page, int bit_nr)
1011 {
1012         wait_queue_head_t *q = page_waitqueue(page);
1013         struct wait_page_key key;
1014         unsigned long flags;
1015         wait_queue_entry_t bookmark;
1016
1017         key.page = page;
1018         key.bit_nr = bit_nr;
1019         key.page_match = 0;
1020
1021         bookmark.flags = 0;
1022         bookmark.private = NULL;
1023         bookmark.func = NULL;
1024         INIT_LIST_HEAD(&bookmark.entry);
1025
1026         spin_lock_irqsave(&q->lock, flags);
1027         __wake_up_locked_key_bookmark(q, TASK_NORMAL, &key, &bookmark);
1028
1029         while (bookmark.flags & WQ_FLAG_BOOKMARK) {
1030                 /*
1031                  * Take a breather from holding the lock,
1032                  * allow pages that finish wake up asynchronously
1033                  * to acquire the lock and remove themselves
1034                  * from wait queue
1035                  */
1036                 spin_unlock_irqrestore(&q->lock, flags);
1037                 cpu_relax();
1038                 spin_lock_irqsave(&q->lock, flags);
1039                 __wake_up_locked_key_bookmark(q, TASK_NORMAL, &key, &bookmark);
1040         }
1041
1042         /*
1043          * It is possible for other pages to have collided on the waitqueue
1044          * hash, so in that case check for a page match. That prevents a long-
1045          * term waiter
1046          *
1047          * It is still possible to miss a case here, when we woke page waiters
1048          * and removed them from the waitqueue, but there are still other
1049          * page waiters.
1050          */
1051         if (!waitqueue_active(q) || !key.page_match) {
1052                 ClearPageWaiters(page);
1053                 /*
1054                  * It's possible to miss clearing Waiters here, when we woke
1055                  * our page waiters, but the hashed waitqueue has waiters for
1056                  * other pages on it.
1057                  *
1058                  * That's okay, it's a rare case. The next waker will clear it.
1059                  */
1060         }
1061         spin_unlock_irqrestore(&q->lock, flags);
1062 }
1063
1064 static void wake_up_page(struct page *page, int bit)
1065 {
1066         if (!PageWaiters(page))
1067                 return;
1068         wake_up_page_bit(page, bit);
1069 }
1070
1071 static inline int wait_on_page_bit_common(wait_queue_head_t *q,
1072                 struct page *page, int bit_nr, int state, bool lock)
1073 {
1074         struct wait_page_queue wait_page;
1075         wait_queue_entry_t *wait = &wait_page.wait;
1076         int ret = 0;
1077
1078         init_wait(wait);
1079         wait->flags = lock ? WQ_FLAG_EXCLUSIVE : 0;
1080         wait->func = wake_page_function;
1081         wait_page.page = page;
1082         wait_page.bit_nr = bit_nr;
1083
1084         for (;;) {
1085                 spin_lock_irq(&q->lock);
1086
1087                 if (likely(list_empty(&wait->entry))) {
1088                         __add_wait_queue_entry_tail(q, wait);
1089                         SetPageWaiters(page);
1090                 }
1091
1092                 set_current_state(state);
1093
1094                 spin_unlock_irq(&q->lock);
1095
1096                 if (likely(test_bit(bit_nr, &page->flags))) {
1097                         io_schedule();
1098                 }
1099
1100                 if (lock) {
1101                         if (!test_and_set_bit_lock(bit_nr, &page->flags))
1102                                 break;
1103                 } else {
1104                         if (!test_bit(bit_nr, &page->flags))
1105                                 break;
1106                 }
1107
1108                 if (unlikely(signal_pending_state(state, current))) {
1109                         ret = -EINTR;
1110                         break;
1111                 }
1112         }
1113
1114         finish_wait(q, wait);
1115
1116         /*
1117          * A signal could leave PageWaiters set. Clearing it here if
1118          * !waitqueue_active would be possible (by open-coding finish_wait),
1119          * but still fail to catch it in the case of wait hash collision. We
1120          * already can fail to clear wait hash collision cases, so don't
1121          * bother with signals either.
1122          */
1123
1124         return ret;
1125 }
1126
1127 void wait_on_page_bit(struct page *page, int bit_nr)
1128 {
1129         wait_queue_head_t *q = page_waitqueue(page);
1130         wait_on_page_bit_common(q, page, bit_nr, TASK_UNINTERRUPTIBLE, false);
1131 }
1132 EXPORT_SYMBOL(wait_on_page_bit);
1133
1134 int wait_on_page_bit_killable(struct page *page, int bit_nr)
1135 {
1136         wait_queue_head_t *q = page_waitqueue(page);
1137         return wait_on_page_bit_common(q, page, bit_nr, TASK_KILLABLE, false);
1138 }
1139 EXPORT_SYMBOL(wait_on_page_bit_killable);
1140
1141 /**
1142  * add_page_wait_queue - Add an arbitrary waiter to a page's wait queue
1143  * @page: Page defining the wait queue of interest
1144  * @waiter: Waiter to add to the queue
1145  *
1146  * Add an arbitrary @waiter to the wait queue for the nominated @page.
1147  */
1148 void add_page_wait_queue(struct page *page, wait_queue_entry_t *waiter)
1149 {
1150         wait_queue_head_t *q = page_waitqueue(page);
1151         unsigned long flags;
1152
1153         spin_lock_irqsave(&q->lock, flags);
1154         __add_wait_queue_entry_tail(q, waiter);
1155         SetPageWaiters(page);
1156         spin_unlock_irqrestore(&q->lock, flags);
1157 }
1158 EXPORT_SYMBOL_GPL(add_page_wait_queue);
1159
1160 #ifndef clear_bit_unlock_is_negative_byte
1161
1162 /*
1163  * PG_waiters is the high bit in the same byte as PG_lock.
1164  *
1165  * On x86 (and on many other architectures), we can clear PG_lock and
1166  * test the sign bit at the same time. But if the architecture does
1167  * not support that special operation, we just do this all by hand
1168  * instead.
1169  *
1170  * The read of PG_waiters has to be after (or concurrently with) PG_locked
1171  * being cleared, but a memory barrier should be unneccssary since it is
1172  * in the same byte as PG_locked.
1173  */
1174 static inline bool clear_bit_unlock_is_negative_byte(long nr, volatile void *mem)
1175 {
1176         clear_bit_unlock(nr, mem);
1177         /* smp_mb__after_atomic(); */
1178         return test_bit(PG_waiters, mem);
1179 }
1180
1181 #endif
1182
1183 /**
1184  * unlock_page - unlock a locked page
1185  * @page: the page
1186  *
1187  * Unlocks the page and wakes up sleepers in ___wait_on_page_locked().
1188  * Also wakes sleepers in wait_on_page_writeback() because the wakeup
1189  * mechanism between PageLocked pages and PageWriteback pages is shared.
1190  * But that's OK - sleepers in wait_on_page_writeback() just go back to sleep.
1191  *
1192  * Note that this depends on PG_waiters being the sign bit in the byte
1193  * that contains PG_locked - thus the BUILD_BUG_ON(). That allows us to
1194  * clear the PG_locked bit and test PG_waiters at the same time fairly
1195  * portably (architectures that do LL/SC can test any bit, while x86 can
1196  * test the sign bit).
1197  */
1198 void unlock_page(struct page *page)
1199 {
1200         BUILD_BUG_ON(PG_waiters != 7);
1201         page = compound_head(page);
1202         VM_BUG_ON_PAGE(!PageLocked(page), page);
1203         if (clear_bit_unlock_is_negative_byte(PG_locked, &page->flags))
1204                 wake_up_page_bit(page, PG_locked);
1205 }
1206 EXPORT_SYMBOL(unlock_page);
1207
1208 /**
1209  * end_page_writeback - end writeback against a page
1210  * @page: the page
1211  */
1212 void end_page_writeback(struct page *page)
1213 {
1214         /*
1215          * TestClearPageReclaim could be used here but it is an atomic
1216          * operation and overkill in this particular case. Failing to
1217          * shuffle a page marked for immediate reclaim is too mild to
1218          * justify taking an atomic operation penalty at the end of
1219          * ever page writeback.
1220          */
1221         if (PageReclaim(page)) {
1222                 ClearPageReclaim(page);
1223                 rotate_reclaimable_page(page);
1224         }
1225
1226         if (!test_clear_page_writeback(page))
1227                 BUG();
1228
1229         smp_mb__after_atomic();
1230         wake_up_page(page, PG_writeback);
1231 }
1232 EXPORT_SYMBOL(end_page_writeback);
1233
1234 /*
1235  * After completing I/O on a page, call this routine to update the page
1236  * flags appropriately
1237  */
1238 void page_endio(struct page *page, bool is_write, int err)
1239 {
1240         if (!is_write) {
1241                 if (!err) {
1242                         SetPageUptodate(page);
1243                 } else {
1244                         ClearPageUptodate(page);
1245                         SetPageError(page);
1246                 }
1247                 unlock_page(page);
1248         } else {
1249                 if (err) {
1250                         struct address_space *mapping;
1251
1252                         SetPageError(page);
1253                         mapping = page_mapping(page);
1254                         if (mapping)
1255                                 mapping_set_error(mapping, err);
1256                 }
1257                 end_page_writeback(page);
1258         }
1259 }
1260 EXPORT_SYMBOL_GPL(page_endio);
1261
1262 /**
1263  * __lock_page - get a lock on the page, assuming we need to sleep to get it
1264  * @__page: the page to lock
1265  */
1266 void __lock_page(struct page *__page)
1267 {
1268         struct page *page = compound_head(__page);
1269         wait_queue_head_t *q = page_waitqueue(page);
1270         wait_on_page_bit_common(q, page, PG_locked, TASK_UNINTERRUPTIBLE, true);
1271 }
1272 EXPORT_SYMBOL(__lock_page);
1273
1274 int __lock_page_killable(struct page *__page)
1275 {
1276         struct page *page = compound_head(__page);
1277         wait_queue_head_t *q = page_waitqueue(page);
1278         return wait_on_page_bit_common(q, page, PG_locked, TASK_KILLABLE, true);
1279 }
1280 EXPORT_SYMBOL_GPL(__lock_page_killable);
1281
1282 /*
1283  * Return values:
1284  * 1 - page is locked; mmap_sem is still held.
1285  * 0 - page is not locked.
1286  *     mmap_sem has been released (up_read()), unless flags had both
1287  *     FAULT_FLAG_ALLOW_RETRY and FAULT_FLAG_RETRY_NOWAIT set, in
1288  *     which case mmap_sem is still held.
1289  *
1290  * If neither ALLOW_RETRY nor KILLABLE are set, will always return 1
1291  * with the page locked and the mmap_sem unperturbed.
1292  */
1293 int __lock_page_or_retry(struct page *page, struct mm_struct *mm,
1294                          unsigned int flags)
1295 {
1296         if (flags & FAULT_FLAG_ALLOW_RETRY) {
1297                 /*
1298                  * CAUTION! In this case, mmap_sem is not released
1299                  * even though return 0.
1300                  */
1301                 if (flags & FAULT_FLAG_RETRY_NOWAIT)
1302                         return 0;
1303
1304                 up_read(&mm->mmap_sem);
1305                 if (flags & FAULT_FLAG_KILLABLE)
1306                         wait_on_page_locked_killable(page);
1307                 else
1308                         wait_on_page_locked(page);
1309                 return 0;
1310         } else {
1311                 if (flags & FAULT_FLAG_KILLABLE) {
1312                         int ret;
1313
1314                         ret = __lock_page_killable(page);
1315                         if (ret) {
1316                                 up_read(&mm->mmap_sem);
1317                                 return 0;
1318                         }
1319                 } else
1320                         __lock_page(page);
1321                 return 1;
1322         }
1323 }
1324
1325 /**
1326  * page_cache_next_hole - find the next hole (not-present entry)
1327  * @mapping: mapping
1328  * @index: index
1329  * @max_scan: maximum range to search
1330  *
1331  * Search the set [index, min(index+max_scan-1, MAX_INDEX)] for the
1332  * lowest indexed hole.
1333  *
1334  * Returns: the index of the hole if found, otherwise returns an index
1335  * outside of the set specified (in which case 'return - index >=
1336  * max_scan' will be true). In rare cases of index wrap-around, 0 will
1337  * be returned.
1338  *
1339  * page_cache_next_hole may be called under rcu_read_lock. However,
1340  * like radix_tree_gang_lookup, this will not atomically search a
1341  * snapshot of the tree at a single point in time. For example, if a
1342  * hole is created at index 5, then subsequently a hole is created at
1343  * index 10, page_cache_next_hole covering both indexes may return 10
1344  * if called under rcu_read_lock.
1345  */
1346 pgoff_t page_cache_next_hole(struct address_space *mapping,
1347                              pgoff_t index, unsigned long max_scan)
1348 {
1349         unsigned long i;
1350
1351         for (i = 0; i < max_scan; i++) {
1352                 struct page *page;
1353
1354                 page = radix_tree_lookup(&mapping->i_pages, index);
1355                 if (!page || radix_tree_exceptional_entry(page))
1356                         break;
1357                 index++;
1358                 if (index == 0)
1359                         break;
1360         }
1361
1362         return index;
1363 }
1364 EXPORT_SYMBOL(page_cache_next_hole);
1365
1366 /**
1367  * page_cache_prev_hole - find the prev hole (not-present entry)
1368  * @mapping: mapping
1369  * @index: index
1370  * @max_scan: maximum range to search
1371  *
1372  * Search backwards in the range [max(index-max_scan+1, 0), index] for
1373  * the first hole.
1374  *
1375  * Returns: the index of the hole if found, otherwise returns an index
1376  * outside of the set specified (in which case 'index - return >=
1377  * max_scan' will be true). In rare cases of wrap-around, ULONG_MAX
1378  * will be returned.
1379  *
1380  * page_cache_prev_hole may be called under rcu_read_lock. However,
1381  * like radix_tree_gang_lookup, this will not atomically search a
1382  * snapshot of the tree at a single point in time. For example, if a
1383  * hole is created at index 10, then subsequently a hole is created at
1384  * index 5, page_cache_prev_hole covering both indexes may return 5 if
1385  * called under rcu_read_lock.
1386  */
1387 pgoff_t page_cache_prev_hole(struct address_space *mapping,
1388                              pgoff_t index, unsigned long max_scan)
1389 {
1390         unsigned long i;
1391
1392         for (i = 0; i < max_scan; i++) {
1393                 struct page *page;
1394
1395                 page = radix_tree_lookup(&mapping->i_pages, index);
1396                 if (!page || radix_tree_exceptional_entry(page))
1397                         break;
1398                 index--;
1399                 if (index == ULONG_MAX)
1400                         break;
1401         }
1402
1403         return index;
1404 }
1405 EXPORT_SYMBOL(page_cache_prev_hole);
1406
1407 /**
1408  * find_get_entry - find and get a page cache entry
1409  * @mapping: the address_space to search
1410  * @offset: the page cache index
1411  *
1412  * Looks up the page cache slot at @mapping & @offset.  If there is a
1413  * page cache page, it is returned with an increased refcount.
1414  *
1415  * If the slot holds a shadow entry of a previously evicted page, or a
1416  * swap entry from shmem/tmpfs, it is returned.
1417  *
1418  * Otherwise, %NULL is returned.
1419  */
1420 struct page *find_get_entry(struct address_space *mapping, pgoff_t offset)
1421 {
1422         void **pagep;
1423         struct page *head, *page;
1424
1425         rcu_read_lock();
1426 repeat:
1427         page = NULL;
1428         pagep = radix_tree_lookup_slot(&mapping->i_pages, offset);
1429         if (pagep) {
1430                 page = radix_tree_deref_slot(pagep);
1431                 if (unlikely(!page))
1432                         goto out;
1433                 if (radix_tree_exception(page)) {
1434                         if (radix_tree_deref_retry(page))
1435                                 goto repeat;
1436                         /*
1437                          * A shadow entry of a recently evicted page,
1438                          * or a swap entry from shmem/tmpfs.  Return
1439                          * it without attempting to raise page count.
1440                          */
1441                         goto out;
1442                 }
1443
1444                 head = compound_head(page);
1445                 if (!page_cache_get_speculative(head))
1446                         goto repeat;
1447
1448                 /* The page was split under us? */
1449                 if (compound_head(page) != head) {
1450                         put_page(head);
1451                         goto repeat;
1452                 }
1453
1454                 /*
1455                  * Has the page moved?
1456                  * This is part of the lockless pagecache protocol. See
1457                  * include/linux/pagemap.h for details.
1458                  */
1459                 if (unlikely(page != *pagep)) {
1460                         put_page(head);
1461                         goto repeat;
1462                 }
1463         }
1464 out:
1465         rcu_read_unlock();
1466
1467         return page;
1468 }
1469 EXPORT_SYMBOL(find_get_entry);
1470
1471 /**
1472  * find_lock_entry - locate, pin and lock a page cache entry
1473  * @mapping: the address_space to search
1474  * @offset: the page cache index
1475  *
1476  * Looks up the page cache slot at @mapping & @offset.  If there is a
1477  * page cache page, it is returned locked and with an increased
1478  * refcount.
1479  *
1480  * If the slot holds a shadow entry of a previously evicted page, or a
1481  * swap entry from shmem/tmpfs, it is returned.
1482  *
1483  * Otherwise, %NULL is returned.
1484  *
1485  * find_lock_entry() may sleep.
1486  */
1487 struct page *find_lock_entry(struct address_space *mapping, pgoff_t offset)
1488 {
1489         struct page *page;
1490
1491 repeat:
1492         page = find_get_entry(mapping, offset);
1493         if (page && !radix_tree_exception(page)) {
1494                 lock_page(page);
1495                 /* Has the page been truncated? */
1496                 if (unlikely(page_mapping(page) != mapping)) {
1497                         unlock_page(page);
1498                         put_page(page);
1499                         goto repeat;
1500                 }
1501                 VM_BUG_ON_PAGE(page_to_pgoff(page) != offset, page);
1502         }
1503         return page;
1504 }
1505 EXPORT_SYMBOL(find_lock_entry);
1506
1507 /**
1508  * pagecache_get_page - find and get a page reference
1509  * @mapping: the address_space to search
1510  * @offset: the page index
1511  * @fgp_flags: PCG flags
1512  * @gfp_mask: gfp mask to use for the page cache data page allocation
1513  *
1514  * Looks up the page cache slot at @mapping & @offset.
1515  *
1516  * PCG flags modify how the page is returned.
1517  *
1518  * @fgp_flags can be:
1519  *
1520  * - FGP_ACCESSED: the page will be marked accessed
1521  * - FGP_LOCK: Page is return locked
1522  * - FGP_CREAT: If page is not present then a new page is allocated using
1523  *   @gfp_mask and added to the page cache and the VM's LRU
1524  *   list. The page is returned locked and with an increased
1525  *   refcount. Otherwise, NULL is returned.
1526  *
1527  * If FGP_LOCK or FGP_CREAT are specified then the function may sleep even
1528  * if the GFP flags specified for FGP_CREAT are atomic.
1529  *
1530  * If there is a page cache page, it is returned with an increased refcount.
1531  */
1532 struct page *pagecache_get_page(struct address_space *mapping, pgoff_t offset,
1533         int fgp_flags, gfp_t gfp_mask)
1534 {
1535         struct page *page;
1536
1537 repeat:
1538         page = find_get_entry(mapping, offset);
1539         if (radix_tree_exceptional_entry(page))
1540                 page = NULL;
1541         if (!page)
1542                 goto no_page;
1543
1544         if (fgp_flags & FGP_LOCK) {
1545                 if (fgp_flags & FGP_NOWAIT) {
1546                         if (!trylock_page(page)) {
1547                                 put_page(page);
1548                                 return NULL;
1549                         }
1550                 } else {
1551                         lock_page(page);
1552                 }
1553
1554                 /* Has the page been truncated? */
1555                 if (unlikely(page->mapping != mapping)) {
1556                         unlock_page(page);
1557                         put_page(page);
1558                         goto repeat;
1559                 }
1560                 VM_BUG_ON_PAGE(page->index != offset, page);
1561         }
1562
1563         if (page && (fgp_flags & FGP_ACCESSED))
1564                 mark_page_accessed(page);
1565
1566 no_page:
1567         if (!page && (fgp_flags & FGP_CREAT)) {
1568                 int err;
1569                 if ((fgp_flags & FGP_WRITE) && mapping_cap_account_dirty(mapping))
1570                         gfp_mask |= __GFP_WRITE;
1571                 if (fgp_flags & FGP_NOFS)
1572                         gfp_mask &= ~__GFP_FS;
1573
1574                 page = __page_cache_alloc(gfp_mask);
1575                 if (!page)
1576                         return NULL;
1577
1578                 if (WARN_ON_ONCE(!(fgp_flags & FGP_LOCK)))
1579                         fgp_flags |= FGP_LOCK;
1580
1581                 /* Init accessed so avoid atomic mark_page_accessed later */
1582                 if (fgp_flags & FGP_ACCESSED)
1583                         __SetPageReferenced(page);
1584
1585                 err = add_to_page_cache_lru(page, mapping, offset, gfp_mask);
1586                 if (unlikely(err)) {
1587                         put_page(page);
1588                         page = NULL;
1589                         if (err == -EEXIST)
1590                                 goto repeat;
1591                 }
1592         }
1593
1594         return page;
1595 }
1596 EXPORT_SYMBOL(pagecache_get_page);
1597
1598 /**
1599  * find_get_entries - gang pagecache lookup
1600  * @mapping:    The address_space to search
1601  * @start:      The starting page cache index
1602  * @nr_entries: The maximum number of entries
1603  * @entries:    Where the resulting entries are placed
1604  * @indices:    The cache indices corresponding to the entries in @entries
1605  *
1606  * find_get_entries() will search for and return a group of up to
1607  * @nr_entries entries in the mapping.  The entries are placed at
1608  * @entries.  find_get_entries() takes a reference against any actual
1609  * pages it returns.
1610  *
1611  * The search returns a group of mapping-contiguous page cache entries
1612  * with ascending indexes.  There may be holes in the indices due to
1613  * not-present pages.
1614  *
1615  * Any shadow entries of evicted pages, or swap entries from
1616  * shmem/tmpfs, are included in the returned array.
1617  *
1618  * find_get_entries() returns the number of pages and shadow entries
1619  * which were found.
1620  */
1621 unsigned find_get_entries(struct address_space *mapping,
1622                           pgoff_t start, unsigned int nr_entries,
1623                           struct page **entries, pgoff_t *indices)
1624 {
1625         void **slot;
1626         unsigned int ret = 0;
1627         struct radix_tree_iter iter;
1628
1629         if (!nr_entries)
1630                 return 0;
1631
1632         rcu_read_lock();
1633         radix_tree_for_each_slot(slot, &mapping->i_pages, &iter, start) {
1634                 struct page *head, *page;
1635 repeat:
1636                 page = radix_tree_deref_slot(slot);
1637                 if (unlikely(!page))
1638                         continue;
1639                 if (radix_tree_exception(page)) {
1640                         if (radix_tree_deref_retry(page)) {
1641                                 slot = radix_tree_iter_retry(&iter);
1642                                 continue;
1643                         }
1644                         /*
1645                          * A shadow entry of a recently evicted page, a swap
1646                          * entry from shmem/tmpfs or a DAX entry.  Return it
1647                          * without attempting to raise page count.
1648                          */
1649                         goto export;
1650                 }
1651
1652                 head = compound_head(page);
1653                 if (!page_cache_get_speculative(head))
1654                         goto repeat;
1655
1656                 /* The page was split under us? */
1657                 if (compound_head(page) != head) {
1658                         put_page(head);
1659                         goto repeat;
1660                 }
1661
1662                 /* Has the page moved? */
1663                 if (unlikely(page != *slot)) {
1664                         put_page(head);
1665                         goto repeat;
1666                 }
1667 export:
1668                 indices[ret] = iter.index;
1669                 entries[ret] = page;
1670                 if (++ret == nr_entries)
1671                         break;
1672         }
1673         rcu_read_unlock();
1674         return ret;
1675 }
1676
1677 /**
1678  * find_get_pages_range - gang pagecache lookup
1679  * @mapping:    The address_space to search
1680  * @start:      The starting page index
1681  * @end:        The final page index (inclusive)
1682  * @nr_pages:   The maximum number of pages
1683  * @pages:      Where the resulting pages are placed
1684  *
1685  * find_get_pages_range() will search for and return a group of up to @nr_pages
1686  * pages in the mapping starting at index @start and up to index @end
1687  * (inclusive).  The pages are placed at @pages.  find_get_pages_range() takes
1688  * a reference against the returned pages.
1689  *
1690  * The search returns a group of mapping-contiguous pages with ascending
1691  * indexes.  There may be holes in the indices due to not-present pages.
1692  * We also update @start to index the next page for the traversal.
1693  *
1694  * find_get_pages_range() returns the number of pages which were found. If this
1695  * number is smaller than @nr_pages, the end of specified range has been
1696  * reached.
1697  */
1698 unsigned find_get_pages_range(struct address_space *mapping, pgoff_t *start,
1699                               pgoff_t end, unsigned int nr_pages,
1700                               struct page **pages)
1701 {
1702         struct radix_tree_iter iter;
1703         void **slot;
1704         unsigned ret = 0;
1705
1706         if (unlikely(!nr_pages))
1707                 return 0;
1708
1709         rcu_read_lock();
1710         radix_tree_for_each_slot(slot, &mapping->i_pages, &iter, *start) {
1711                 struct page *head, *page;
1712
1713                 if (iter.index > end)
1714                         break;
1715 repeat:
1716                 page = radix_tree_deref_slot(slot);
1717                 if (unlikely(!page))
1718                         continue;
1719
1720                 if (radix_tree_exception(page)) {
1721                         if (radix_tree_deref_retry(page)) {
1722                                 slot = radix_tree_iter_retry(&iter);
1723                                 continue;
1724                         }
1725                         /*
1726                          * A shadow entry of a recently evicted page,
1727                          * or a swap entry from shmem/tmpfs.  Skip
1728                          * over it.
1729                          */
1730                         continue;
1731                 }
1732
1733                 head = compound_head(page);
1734                 if (!page_cache_get_speculative(head))
1735                         goto repeat;
1736
1737                 /* The page was split under us? */
1738                 if (compound_head(page) != head) {
1739                         put_page(head);
1740                         goto repeat;
1741                 }
1742
1743                 /* Has the page moved? */
1744                 if (unlikely(page != *slot)) {
1745                         put_page(head);
1746                         goto repeat;
1747                 }
1748
1749                 pages[ret] = page;
1750                 if (++ret == nr_pages) {
1751                         *start = pages[ret - 1]->index + 1;
1752                         goto out;
1753                 }
1754         }
1755
1756         /*
1757          * We come here when there is no page beyond @end. We take care to not
1758          * overflow the index @start as it confuses some of the callers. This
1759          * breaks the iteration when there is page at index -1 but that is
1760          * already broken anyway.
1761          */
1762         if (end == (pgoff_t)-1)
1763                 *start = (pgoff_t)-1;
1764         else
1765                 *start = end + 1;
1766 out:
1767         rcu_read_unlock();
1768
1769         return ret;
1770 }
1771
1772 /**
1773  * find_get_pages_contig - gang contiguous pagecache lookup
1774  * @mapping:    The address_space to search
1775  * @index:      The starting page index
1776  * @nr_pages:   The maximum number of pages
1777  * @pages:      Where the resulting pages are placed
1778  *
1779  * find_get_pages_contig() works exactly like find_get_pages(), except
1780  * that the returned number of pages are guaranteed to be contiguous.
1781  *
1782  * find_get_pages_contig() returns the number of pages which were found.
1783  */
1784 unsigned find_get_pages_contig(struct address_space *mapping, pgoff_t index,
1785                                unsigned int nr_pages, struct page **pages)
1786 {
1787         struct radix_tree_iter iter;
1788         void **slot;
1789         unsigned int ret = 0;
1790
1791         if (unlikely(!nr_pages))
1792                 return 0;
1793
1794         rcu_read_lock();
1795         radix_tree_for_each_contig(slot, &mapping->i_pages, &iter, index) {
1796                 struct page *head, *page;
1797 repeat:
1798                 page = radix_tree_deref_slot(slot);
1799                 /* The hole, there no reason to continue */
1800                 if (unlikely(!page))
1801                         break;
1802
1803                 if (radix_tree_exception(page)) {
1804                         if (radix_tree_deref_retry(page)) {
1805                                 slot = radix_tree_iter_retry(&iter);
1806                                 continue;
1807                         }
1808                         /*
1809                          * A shadow entry of a recently evicted page,
1810                          * or a swap entry from shmem/tmpfs.  Stop
1811                          * looking for contiguous pages.
1812                          */
1813                         break;
1814                 }
1815
1816                 head = compound_head(page);
1817                 if (!page_cache_get_speculative(head))
1818                         goto repeat;
1819
1820                 /* The page was split under us? */
1821                 if (compound_head(page) != head) {
1822                         put_page(head);
1823                         goto repeat;
1824                 }
1825
1826                 /* Has the page moved? */
1827                 if (unlikely(page != *slot)) {
1828                         put_page(head);
1829                         goto repeat;
1830                 }
1831
1832                 /*
1833                  * must check mapping and index after taking the ref.
1834                  * otherwise we can get both false positives and false
1835                  * negatives, which is just confusing to the caller.
1836                  */
1837                 if (page->mapping == NULL || page_to_pgoff(page) != iter.index) {
1838                         put_page(page);
1839                         break;
1840                 }
1841
1842                 pages[ret] = page;
1843                 if (++ret == nr_pages)
1844                         break;
1845         }
1846         rcu_read_unlock();
1847         return ret;
1848 }
1849 EXPORT_SYMBOL(find_get_pages_contig);
1850
1851 /**
1852  * find_get_pages_range_tag - find and return pages in given range matching @tag
1853  * @mapping:    the address_space to search
1854  * @index:      the starting page index
1855  * @end:        The final page index (inclusive)
1856  * @tag:        the tag index
1857  * @nr_pages:   the maximum number of pages
1858  * @pages:      where the resulting pages are placed
1859  *
1860  * Like find_get_pages, except we only return pages which are tagged with
1861  * @tag.   We update @index to index the next page for the traversal.
1862  */
1863 unsigned find_get_pages_range_tag(struct address_space *mapping, pgoff_t *index,
1864                         pgoff_t end, int tag, unsigned int nr_pages,
1865                         struct page **pages)
1866 {
1867         struct radix_tree_iter iter;
1868         void **slot;
1869         unsigned ret = 0;
1870
1871         if (unlikely(!nr_pages))
1872                 return 0;
1873
1874         rcu_read_lock();
1875         radix_tree_for_each_tagged(slot, &mapping->i_pages, &iter, *index, tag) {
1876                 struct page *head, *page;
1877
1878                 if (iter.index > end)
1879                         break;
1880 repeat:
1881                 page = radix_tree_deref_slot(slot);
1882                 if (unlikely(!page))
1883                         continue;
1884
1885                 if (radix_tree_exception(page)) {
1886                         if (radix_tree_deref_retry(page)) {
1887                                 slot = radix_tree_iter_retry(&iter);
1888                                 continue;
1889                         }
1890                         /*
1891                          * A shadow entry of a recently evicted page.
1892                          *
1893                          * Those entries should never be tagged, but
1894                          * this tree walk is lockless and the tags are
1895                          * looked up in bulk, one radix tree node at a
1896                          * time, so there is a sizable window for page
1897                          * reclaim to evict a page we saw tagged.
1898                          *
1899                          * Skip over it.
1900                          */
1901                         continue;
1902                 }
1903
1904                 head = compound_head(page);
1905                 if (!page_cache_get_speculative(head))
1906                         goto repeat;
1907
1908                 /* The page was split under us? */
1909                 if (compound_head(page) != head) {
1910                         put_page(head);
1911                         goto repeat;
1912                 }
1913
1914                 /* Has the page moved? */
1915                 if (unlikely(page != *slot)) {
1916                         put_page(head);
1917                         goto repeat;
1918                 }
1919
1920                 pages[ret] = page;
1921                 if (++ret == nr_pages) {
1922                         *index = pages[ret - 1]->index + 1;
1923                         goto out;
1924                 }
1925         }
1926
1927         /*
1928          * We come here when we got at @end. We take care to not overflow the
1929          * index @index as it confuses some of the callers. This breaks the
1930          * iteration when there is page at index -1 but that is already broken
1931          * anyway.
1932          */
1933         if (end == (pgoff_t)-1)
1934                 *index = (pgoff_t)-1;
1935         else
1936                 *index = end + 1;
1937 out:
1938         rcu_read_unlock();
1939
1940         return ret;
1941 }
1942 EXPORT_SYMBOL(find_get_pages_range_tag);
1943
1944 /**
1945  * find_get_entries_tag - find and return entries that match @tag
1946  * @mapping:    the address_space to search
1947  * @start:      the starting page cache index
1948  * @tag:        the tag index
1949  * @nr_entries: the maximum number of entries
1950  * @entries:    where the resulting entries are placed
1951  * @indices:    the cache indices corresponding to the entries in @entries
1952  *
1953  * Like find_get_entries, except we only return entries which are tagged with
1954  * @tag.
1955  */
1956 unsigned find_get_entries_tag(struct address_space *mapping, pgoff_t start,
1957                         int tag, unsigned int nr_entries,
1958                         struct page **entries, pgoff_t *indices)
1959 {
1960         void **slot;
1961         unsigned int ret = 0;
1962         struct radix_tree_iter iter;
1963
1964         if (!nr_entries)
1965                 return 0;
1966
1967         rcu_read_lock();
1968         radix_tree_for_each_tagged(slot, &mapping->i_pages, &iter, start, tag) {
1969                 struct page *head, *page;
1970 repeat:
1971                 page = radix_tree_deref_slot(slot);
1972                 if (unlikely(!page))
1973                         continue;
1974                 if (radix_tree_exception(page)) {
1975                         if (radix_tree_deref_retry(page)) {
1976                                 slot = radix_tree_iter_retry(&iter);
1977                                 continue;
1978                         }
1979
1980                         /*
1981                          * A shadow entry of a recently evicted page, a swap
1982                          * entry from shmem/tmpfs or a DAX entry.  Return it
1983                          * without attempting to raise page count.
1984                          */
1985                         goto export;
1986                 }
1987
1988                 head = compound_head(page);
1989                 if (!page_cache_get_speculative(head))
1990                         goto repeat;
1991
1992                 /* The page was split under us? */
1993                 if (compound_head(page) != head) {
1994                         put_page(head);
1995                         goto repeat;
1996                 }
1997
1998                 /* Has the page moved? */
1999                 if (unlikely(page != *slot)) {
2000                         put_page(head);
2001                         goto repeat;
2002                 }
2003 export:
2004                 indices[ret] = iter.index;
2005                 entries[ret] = page;
2006                 if (++ret == nr_entries)
2007                         break;
2008         }
2009         rcu_read_unlock();
2010         return ret;
2011 }
2012 EXPORT_SYMBOL(find_get_entries_tag);
2013
2014 /*
2015  * CD/DVDs are error prone. When a medium error occurs, the driver may fail
2016  * a _large_ part of the i/o request. Imagine the worst scenario:
2017  *
2018  *      ---R__________________________________________B__________
2019  *         ^ reading here                             ^ bad block(assume 4k)
2020  *
2021  * read(R) => miss => readahead(R...B) => media error => frustrating retries
2022  * => failing the whole request => read(R) => read(R+1) =>
2023  * readahead(R+1...B+1) => bang => read(R+2) => read(R+3) =>
2024  * readahead(R+3...B+2) => bang => read(R+3) => read(R+4) =>
2025  * readahead(R+4...B+3) => bang => read(R+4) => read(R+5) => ......
2026  *
2027  * It is going insane. Fix it by quickly scaling down the readahead size.
2028  */
2029 static void shrink_readahead_size_eio(struct file *filp,
2030                                         struct file_ra_state *ra)
2031 {
2032         ra->ra_pages /= 4;
2033 }
2034
2035 /**
2036  * generic_file_buffered_read - generic file read routine
2037  * @iocb:       the iocb to read
2038  * @iter:       data destination
2039  * @written:    already copied
2040  *
2041  * This is a generic file read routine, and uses the
2042  * mapping->a_ops->readpage() function for the actual low-level stuff.
2043  *
2044  * This is really ugly. But the goto's actually try to clarify some
2045  * of the logic when it comes to error handling etc.
2046  */
2047 static ssize_t generic_file_buffered_read(struct kiocb *iocb,
2048                 struct iov_iter *iter, ssize_t written)
2049 {
2050         struct file *filp = iocb->ki_filp;
2051         struct address_space *mapping = filp->f_mapping;
2052         struct inode *inode = mapping->host;
2053         struct file_ra_state *ra = &filp->f_ra;
2054         loff_t *ppos = &iocb->ki_pos;
2055         pgoff_t index;
2056         pgoff_t last_index;
2057         pgoff_t prev_index;
2058         unsigned long offset;      /* offset into pagecache page */
2059         unsigned int prev_offset;
2060         int error = 0;
2061
2062         if (unlikely(*ppos >= inode->i_sb->s_maxbytes))
2063                 return 0;
2064         iov_iter_truncate(iter, inode->i_sb->s_maxbytes);
2065
2066         index = *ppos >> PAGE_SHIFT;
2067         prev_index = ra->prev_pos >> PAGE_SHIFT;
2068         prev_offset = ra->prev_pos & (PAGE_SIZE-1);
2069         last_index = (*ppos + iter->count + PAGE_SIZE-1) >> PAGE_SHIFT;
2070         offset = *ppos & ~PAGE_MASK;
2071
2072         for (;;) {
2073                 struct page *page;
2074                 pgoff_t end_index;
2075                 loff_t isize;
2076                 unsigned long nr, ret;
2077
2078                 cond_resched();
2079 find_page:
2080                 if (fatal_signal_pending(current)) {
2081                         error = -EINTR;
2082                         goto out;
2083                 }
2084
2085                 page = find_get_page(mapping, index);
2086                 if (!page) {
2087                         if (iocb->ki_flags & IOCB_NOWAIT)
2088                                 goto would_block;
2089                         page_cache_sync_readahead(mapping,
2090                                         ra, filp,
2091                                         index, last_index - index);
2092                         page = find_get_page(mapping, index);
2093                         if (unlikely(page == NULL))
2094                                 goto no_cached_page;
2095                 }
2096                 if (PageReadahead(page)) {
2097                         page_cache_async_readahead(mapping,
2098                                         ra, filp, page,
2099                                         index, last_index - index);
2100                 }
2101                 if (!PageUptodate(page)) {
2102                         if (iocb->ki_flags & IOCB_NOWAIT) {
2103                                 put_page(page);
2104                                 goto would_block;
2105                         }
2106
2107                         /*
2108                          * See comment in do_read_cache_page on why
2109                          * wait_on_page_locked is used to avoid unnecessarily
2110                          * serialisations and why it's safe.
2111                          */
2112                         error = wait_on_page_locked_killable(page);
2113                         if (unlikely(error))
2114                                 goto readpage_error;
2115                         if (PageUptodate(page))
2116                                 goto page_ok;
2117
2118                         if (inode->i_blkbits == PAGE_SHIFT ||
2119                                         !mapping->a_ops->is_partially_uptodate)
2120                                 goto page_not_up_to_date;
2121                         /* pipes can't handle partially uptodate pages */
2122                         if (unlikely(iter->type & ITER_PIPE))
2123                                 goto page_not_up_to_date;
2124                         if (!trylock_page(page))
2125                                 goto page_not_up_to_date;
2126                         /* Did it get truncated before we got the lock? */
2127                         if (!page->mapping)
2128                                 goto page_not_up_to_date_locked;
2129                         if (!mapping->a_ops->is_partially_uptodate(page,
2130                                                         offset, iter->count))
2131                                 goto page_not_up_to_date_locked;
2132                         unlock_page(page);
2133                 }
2134 page_ok:
2135                 /*
2136                  * i_size must be checked after we know the page is Uptodate.
2137                  *
2138                  * Checking i_size after the check allows us to calculate
2139                  * the correct value for "nr", which means the zero-filled
2140                  * part of the page is not copied back to userspace (unless
2141                  * another truncate extends the file - this is desired though).
2142                  */
2143
2144                 isize = i_size_read(inode);
2145                 end_index = (isize - 1) >> PAGE_SHIFT;
2146                 if (unlikely(!isize || index > end_index)) {
2147                         put_page(page);
2148                         goto out;
2149                 }
2150
2151                 /* nr is the maximum number of bytes to copy from this page */
2152                 nr = PAGE_SIZE;
2153                 if (index == end_index) {
2154                         nr = ((isize - 1) & ~PAGE_MASK) + 1;
2155                         if (nr <= offset) {
2156                                 put_page(page);
2157                                 goto out;
2158                         }
2159                 }
2160                 nr = nr - offset;
2161
2162                 /* If users can be writing to this page using arbitrary
2163                  * virtual addresses, take care about potential aliasing
2164                  * before reading the page on the kernel side.
2165                  */
2166                 if (mapping_writably_mapped(mapping))
2167                         flush_dcache_page(page);
2168
2169                 /*
2170                  * When a sequential read accesses a page several times,
2171                  * only mark it as accessed the first time.
2172                  */
2173                 if (prev_index != index || offset != prev_offset)
2174                         mark_page_accessed(page);
2175                 prev_index = index;
2176
2177                 /*
2178                  * Ok, we have the page, and it's up-to-date, so
2179                  * now we can copy it to user space...
2180                  */
2181
2182                 ret = copy_page_to_iter(page, offset, nr, iter);
2183                 offset += ret;
2184                 index += offset >> PAGE_SHIFT;
2185                 offset &= ~PAGE_MASK;
2186                 prev_offset = offset;
2187
2188                 put_page(page);
2189                 written += ret;
2190                 if (!iov_iter_count(iter))
2191                         goto out;
2192                 if (ret < nr) {
2193                         error = -EFAULT;
2194                         goto out;
2195                 }
2196                 continue;
2197
2198 page_not_up_to_date:
2199                 /* Get exclusive access to the page ... */
2200                 error = lock_page_killable(page);
2201                 if (unlikely(error))
2202                         goto readpage_error;
2203
2204 page_not_up_to_date_locked:
2205                 /* Did it get truncated before we got the lock? */
2206                 if (!page->mapping) {
2207                         unlock_page(page);
2208                         put_page(page);
2209                         continue;
2210                 }
2211
2212                 /* Did somebody else fill it already? */
2213                 if (PageUptodate(page)) {
2214                         unlock_page(page);
2215                         goto page_ok;
2216                 }
2217
2218 readpage:
2219                 /*
2220                  * A previous I/O error may have been due to temporary
2221                  * failures, eg. multipath errors.
2222                  * PG_error will be set again if readpage fails.
2223                  */
2224                 ClearPageError(page);
2225                 /* Start the actual read. The read will unlock the page. */
2226                 error = mapping->a_ops->readpage(filp, page);
2227
2228                 if (unlikely(error)) {
2229                         if (error == AOP_TRUNCATED_PAGE) {
2230                                 put_page(page);
2231                                 error = 0;
2232                                 goto find_page;
2233                         }
2234                         goto readpage_error;
2235                 }
2236
2237                 if (!PageUptodate(page)) {
2238                         error = lock_page_killable(page);
2239                         if (unlikely(error))
2240                                 goto readpage_error;
2241                         if (!PageUptodate(page)) {
2242                                 if (page->mapping == NULL) {
2243                                         /*
2244                                          * invalidate_mapping_pages got it
2245                                          */
2246                                         unlock_page(page);
2247                                         put_page(page);
2248                                         goto find_page;
2249                                 }
2250                                 unlock_page(page);
2251                                 shrink_readahead_size_eio(filp, ra);
2252                                 error = -EIO;
2253                                 goto readpage_error;
2254                         }
2255                         unlock_page(page);
2256                 }
2257
2258                 goto page_ok;
2259
2260 readpage_error:
2261                 /* UHHUH! A synchronous read error occurred. Report it */
2262                 put_page(page);
2263                 goto out;
2264
2265 no_cached_page:
2266                 /*
2267                  * Ok, it wasn't cached, so we need to create a new
2268                  * page..
2269                  */
2270                 page = page_cache_alloc(mapping);
2271                 if (!page) {
2272                         error = -ENOMEM;
2273                         goto out;
2274                 }
2275                 error = add_to_page_cache_lru(page, mapping, index,
2276                                 mapping_gfp_constraint(mapping, GFP_KERNEL));
2277                 if (error) {
2278                         put_page(page);
2279                         if (error == -EEXIST) {
2280                                 error = 0;
2281                                 goto find_page;
2282                         }
2283                         goto out;
2284                 }
2285                 goto readpage;
2286         }
2287
2288 would_block:
2289         error = -EAGAIN;
2290 out:
2291         ra->prev_pos = prev_index;
2292         ra->prev_pos <<= PAGE_SHIFT;
2293         ra->prev_pos |= prev_offset;
2294
2295         *ppos = ((loff_t)index << PAGE_SHIFT) + offset;
2296         file_accessed(filp);
2297         return written ? written : error;
2298 }
2299
2300 /**
2301  * generic_file_read_iter - generic filesystem read routine
2302  * @iocb:       kernel I/O control block
2303  * @iter:       destination for the data read
2304  *
2305  * This is the "read_iter()" routine for all filesystems
2306  * that can use the page cache directly.
2307  */
2308 ssize_t
2309 generic_file_read_iter(struct kiocb *iocb, struct iov_iter *iter)
2310 {
2311         size_t count = iov_iter_count(iter);
2312         ssize_t retval = 0;
2313
2314         if (!count)
2315                 goto out; /* skip atime */
2316
2317         if (iocb->ki_flags & IOCB_DIRECT) {
2318                 struct file *file = iocb->ki_filp;
2319                 struct address_space *mapping = file->f_mapping;
2320                 struct inode *inode = mapping->host;
2321                 loff_t size;
2322
2323                 size = i_size_read(inode);
2324                 if (iocb->ki_flags & IOCB_NOWAIT) {
2325                         if (filemap_range_has_page(mapping, iocb->ki_pos,
2326                                                    iocb->ki_pos + count - 1))
2327                                 return -EAGAIN;
2328                 } else {
2329                         retval = filemap_write_and_wait_range(mapping,
2330                                                 iocb->ki_pos,
2331                                                 iocb->ki_pos + count - 1);
2332                         if (retval < 0)
2333                                 goto out;
2334                 }
2335
2336                 file_accessed(file);
2337
2338                 retval = mapping->a_ops->direct_IO(iocb, iter);
2339                 if (retval >= 0) {
2340                         iocb->ki_pos += retval;
2341                         count -= retval;
2342                 }
2343                 iov_iter_revert(iter, count - iov_iter_count(iter));
2344
2345                 /*
2346                  * Btrfs can have a short DIO read if we encounter
2347                  * compressed extents, so if there was an error, or if
2348                  * we've already read everything we wanted to, or if
2349                  * there was a short read because we hit EOF, go ahead
2350                  * and return.  Otherwise fallthrough to buffered io for
2351                  * the rest of the read.  Buffered reads will not work for
2352                  * DAX files, so don't bother trying.
2353                  */
2354                 if (retval < 0 || !count || iocb->ki_pos >= size ||
2355                     IS_DAX(inode))
2356                         goto out;
2357         }
2358
2359         retval = generic_file_buffered_read(iocb, iter, retval);
2360 out:
2361         return retval;
2362 }
2363 EXPORT_SYMBOL(generic_file_read_iter);
2364
2365 #ifdef CONFIG_MMU
2366 /**
2367  * page_cache_read - adds requested page to the page cache if not already there
2368  * @file:       file to read
2369  * @offset:     page index
2370  * @gfp_mask:   memory allocation flags
2371  *
2372  * This adds the requested page to the page cache if it isn't already there,
2373  * and schedules an I/O to read in its contents from disk.
2374  */
2375 static int page_cache_read(struct file *file, pgoff_t offset, gfp_t gfp_mask)
2376 {
2377         struct address_space *mapping = file->f_mapping;
2378         struct page *page;
2379         int ret;
2380
2381         do {
2382                 page = __page_cache_alloc(gfp_mask);
2383                 if (!page)
2384                         return -ENOMEM;
2385
2386                 ret = add_to_page_cache_lru(page, mapping, offset, gfp_mask);
2387                 if (ret == 0)
2388                         ret = mapping->a_ops->readpage(file, page);
2389                 else if (ret == -EEXIST)
2390                         ret = 0; /* losing race to add is OK */
2391
2392                 put_page(page);
2393
2394         } while (ret == AOP_TRUNCATED_PAGE);
2395
2396         return ret;
2397 }
2398
2399 #define MMAP_LOTSAMISS  (100)
2400
2401 /*
2402  * Synchronous readahead happens when we don't even find
2403  * a page in the page cache at all.
2404  */
2405 static void do_sync_mmap_readahead(struct vm_area_struct *vma,
2406                                    struct file_ra_state *ra,
2407                                    struct file *file,
2408                                    pgoff_t offset)
2409 {
2410         struct address_space *mapping = file->f_mapping;
2411
2412         /* If we don't want any read-ahead, don't bother */
2413         if (vma->vm_flags & VM_RAND_READ)
2414                 return;
2415         if (!ra->ra_pages)
2416                 return;
2417
2418         if (vma->vm_flags & VM_SEQ_READ) {
2419                 page_cache_sync_readahead(mapping, ra, file, offset,
2420                                           ra->ra_pages);
2421                 return;
2422         }
2423
2424         /* Avoid banging the cache line if not needed */
2425         if (ra->mmap_miss < MMAP_LOTSAMISS * 10)
2426                 ra->mmap_miss++;
2427
2428         /*
2429          * Do we miss much more than hit in this file? If so,
2430          * stop bothering with read-ahead. It will only hurt.
2431          */
2432         if (ra->mmap_miss > MMAP_LOTSAMISS)
2433                 return;
2434
2435         /*
2436          * mmap read-around
2437          */
2438         ra->start = max_t(long, 0, offset - ra->ra_pages / 2);
2439         ra->size = ra->ra_pages;
2440         ra->async_size = ra->ra_pages / 4;
2441         ra_submit(ra, mapping, file);
2442 }
2443
2444 /*
2445  * Asynchronous readahead happens when we find the page and PG_readahead,
2446  * so we want to possibly extend the readahead further..
2447  */
2448 static void do_async_mmap_readahead(struct vm_area_struct *vma,
2449                                     struct file_ra_state *ra,
2450                                     struct file *file,
2451                                     struct page *page,
2452                                     pgoff_t offset)
2453 {
2454         struct address_space *mapping = file->f_mapping;
2455
2456         /* If we don't want any read-ahead, don't bother */
2457         if (vma->vm_flags & VM_RAND_READ)
2458                 return;
2459         if (ra->mmap_miss > 0)
2460                 ra->mmap_miss--;
2461         if (PageReadahead(page))
2462                 page_cache_async_readahead(mapping, ra, file,
2463                                            page, offset, ra->ra_pages);
2464 }
2465
2466 /**
2467  * filemap_fault - read in file data for page fault handling
2468  * @vmf:        struct vm_fault containing details of the fault
2469  *
2470  * filemap_fault() is invoked via the vma operations vector for a
2471  * mapped memory region to read in file data during a page fault.
2472  *
2473  * The goto's are kind of ugly, but this streamlines the normal case of having
2474  * it in the page cache, and handles the special cases reasonably without
2475  * having a lot of duplicated code.
2476  *
2477  * vma->vm_mm->mmap_sem must be held on entry.
2478  *
2479  * If our return value has VM_FAULT_RETRY set, it's because
2480  * lock_page_or_retry() returned 0.
2481  * The mmap_sem has usually been released in this case.
2482  * See __lock_page_or_retry() for the exception.
2483  *
2484  * If our return value does not have VM_FAULT_RETRY set, the mmap_sem
2485  * has not been released.
2486  *
2487  * We never return with VM_FAULT_RETRY and a bit from VM_FAULT_ERROR set.
2488  */
2489 vm_fault_t filemap_fault(struct vm_fault *vmf)
2490 {
2491         int error;
2492         struct file *file = vmf->vma->vm_file;
2493         struct address_space *mapping = file->f_mapping;
2494         struct file_ra_state *ra = &file->f_ra;
2495         struct inode *inode = mapping->host;
2496         pgoff_t offset = vmf->pgoff;
2497         pgoff_t max_off;
2498         struct page *page;
2499         vm_fault_t ret = 0;
2500
2501         max_off = DIV_ROUND_UP(i_size_read(inode), PAGE_SIZE);
2502         if (unlikely(offset >= max_off))
2503                 return VM_FAULT_SIGBUS;
2504
2505         /*
2506          * Do we have something in the page cache already?
2507          */
2508         page = find_get_page(mapping, offset);
2509         if (likely(page) && !(vmf->flags & FAULT_FLAG_TRIED)) {
2510                 /*
2511                  * We found the page, so try async readahead before
2512                  * waiting for the lock.
2513                  */
2514                 do_async_mmap_readahead(vmf->vma, ra, file, page, offset);
2515         } else if (!page) {
2516                 /* No page in the page cache at all */
2517                 do_sync_mmap_readahead(vmf->vma, ra, file, offset);
2518                 count_vm_event(PGMAJFAULT);
2519                 count_memcg_event_mm(vmf->vma->vm_mm, PGMAJFAULT);
2520                 ret = VM_FAULT_MAJOR;
2521 retry_find:
2522                 page = find_get_page(mapping, offset);
2523                 if (!page)
2524                         goto no_cached_page;
2525         }
2526
2527         if (!lock_page_or_retry(page, vmf->vma->vm_mm, vmf->flags)) {
2528                 put_page(page);
2529                 return ret | VM_FAULT_RETRY;
2530         }
2531
2532         /* Did it get truncated? */
2533         if (unlikely(page->mapping != mapping)) {
2534                 unlock_page(page);
2535                 put_page(page);
2536                 goto retry_find;
2537         }
2538         VM_BUG_ON_PAGE(page->index != offset, page);
2539
2540         /*
2541          * We have a locked page in the page cache, now we need to check
2542          * that it's up-to-date. If not, it is going to be due to an error.
2543          */
2544         if (unlikely(!PageUptodate(page)))
2545                 goto page_not_uptodate;
2546
2547         /*
2548          * Found the page and have a reference on it.
2549          * We must recheck i_size under page lock.
2550          */
2551         max_off = DIV_ROUND_UP(i_size_read(inode), PAGE_SIZE);
2552         if (unlikely(offset >= max_off)) {
2553                 unlock_page(page);
2554                 put_page(page);
2555                 return VM_FAULT_SIGBUS;
2556         }
2557
2558         vmf->page = page;
2559         return ret | VM_FAULT_LOCKED;
2560
2561 no_cached_page:
2562         /*
2563          * We're only likely to ever get here if MADV_RANDOM is in
2564          * effect.
2565          */
2566         error = page_cache_read(file, offset, vmf->gfp_mask);
2567
2568         /*
2569          * The page we want has now been added to the page cache.
2570          * In the unlikely event that someone removed it in the
2571          * meantime, we'll just come back here and read it again.
2572          */
2573         if (error >= 0)
2574                 goto retry_find;
2575
2576         /*
2577          * An error return from page_cache_read can result if the
2578          * system is low on memory, or a problem occurs while trying
2579          * to schedule I/O.
2580          */
2581         if (error == -ENOMEM)
2582                 return VM_FAULT_OOM;
2583         return VM_FAULT_SIGBUS;
2584
2585 page_not_uptodate:
2586         /*
2587          * Umm, take care of errors if the page isn't up-to-date.
2588          * Try to re-read it _once_. We do this synchronously,
2589          * because there really aren't any performance issues here
2590          * and we need to check for errors.
2591          */
2592         ClearPageError(page);
2593         error = mapping->a_ops->readpage(file, page);
2594         if (!error) {
2595                 wait_on_page_locked(page);
2596                 if (!PageUptodate(page))
2597                         error = -EIO;
2598         }
2599         put_page(page);
2600
2601         if (!error || error == AOP_TRUNCATED_PAGE)
2602                 goto retry_find;
2603
2604         /* Things didn't work out. Return zero to tell the mm layer so. */
2605         shrink_readahead_size_eio(file, ra);
2606         return VM_FAULT_SIGBUS;
2607 }
2608 EXPORT_SYMBOL(filemap_fault);
2609
2610 void filemap_map_pages(struct vm_fault *vmf,
2611                 pgoff_t start_pgoff, pgoff_t end_pgoff)
2612 {
2613         struct radix_tree_iter iter;
2614         void **slot;
2615         struct file *file = vmf->vma->vm_file;
2616         struct address_space *mapping = file->f_mapping;
2617         pgoff_t last_pgoff = start_pgoff;
2618         unsigned long max_idx;
2619         struct page *head, *page;
2620
2621         rcu_read_lock();
2622         radix_tree_for_each_slot(slot, &mapping->i_pages, &iter, start_pgoff) {
2623                 if (iter.index > end_pgoff)
2624                         break;
2625 repeat:
2626                 page = radix_tree_deref_slot(slot);
2627                 if (unlikely(!page))
2628                         goto next;
2629                 if (radix_tree_exception(page)) {
2630                         if (radix_tree_deref_retry(page)) {
2631                                 slot = radix_tree_iter_retry(&iter);
2632                                 continue;
2633                         }
2634                         goto next;
2635                 }
2636
2637                 head = compound_head(page);
2638                 if (!page_cache_get_speculative(head))
2639                         goto repeat;
2640
2641                 /* The page was split under us? */
2642                 if (compound_head(page) != head) {
2643                         put_page(head);
2644                         goto repeat;
2645                 }
2646
2647                 /* Has the page moved? */
2648                 if (unlikely(page != *slot)) {
2649                         put_page(head);
2650                         goto repeat;
2651                 }
2652
2653                 if (!PageUptodate(page) ||
2654                                 PageReadahead(page) ||
2655                                 PageHWPoison(page))
2656                         goto skip;
2657                 if (!trylock_page(page))
2658                         goto skip;
2659
2660                 if (page->mapping != mapping || !PageUptodate(page))
2661                         goto unlock;
2662
2663                 max_idx = DIV_ROUND_UP(i_size_read(mapping->host), PAGE_SIZE);
2664                 if (page->index >= max_idx)
2665                         goto unlock;
2666
2667                 if (file->f_ra.mmap_miss > 0)
2668                         file->f_ra.mmap_miss--;
2669
2670                 vmf->address += (iter.index - last_pgoff) << PAGE_SHIFT;
2671                 if (vmf->pte)
2672                         vmf->pte += iter.index - last_pgoff;
2673                 last_pgoff = iter.index;
2674                 if (alloc_set_pte(vmf, NULL, page))
2675                         goto unlock;
2676                 unlock_page(page);
2677                 goto next;
2678 unlock:
2679                 unlock_page(page);
2680 skip:
2681                 put_page(page);
2682 next:
2683                 /* Huge page is mapped? No need to proceed. */
2684                 if (pmd_trans_huge(*vmf->pmd))
2685                         break;
2686                 if (iter.index == end_pgoff)
2687                         break;
2688         }
2689         rcu_read_unlock();
2690 }
2691 EXPORT_SYMBOL(filemap_map_pages);
2692
2693 vm_fault_t filemap_page_mkwrite(struct vm_fault *vmf)
2694 {
2695         struct page *page = vmf->page;
2696         struct inode *inode = file_inode(vmf->vma->vm_file);
2697         vm_fault_t ret = VM_FAULT_LOCKED;
2698
2699         sb_start_pagefault(inode->i_sb);
2700         file_update_time(vmf->vma->vm_file);
2701         lock_page(page);
2702         if (page->mapping != inode->i_mapping) {
2703                 unlock_page(page);
2704                 ret = VM_FAULT_NOPAGE;
2705                 goto out;
2706         }
2707         /*
2708          * We mark the page dirty already here so that when freeze is in
2709          * progress, we are guaranteed that writeback during freezing will
2710          * see the dirty page and writeprotect it again.
2711          */
2712         set_page_dirty(page);
2713         wait_for_stable_page(page);
2714 out:
2715         sb_end_pagefault(inode->i_sb);
2716         return ret;
2717 }
2718
2719 const struct vm_operations_struct generic_file_vm_ops = {
2720         .fault          = filemap_fault,
2721         .map_pages      = filemap_map_pages,
2722         .page_mkwrite   = filemap_page_mkwrite,
2723 };
2724
2725 /* This is used for a general mmap of a disk file */
2726
2727 int generic_file_mmap(struct file * file, struct vm_area_struct * vma)
2728 {
2729         struct address_space *mapping = file->f_mapping;
2730
2731         if (!mapping->a_ops->readpage)
2732                 return -ENOEXEC;
2733         file_accessed(file);
2734         vma->vm_ops = &generic_file_vm_ops;
2735         return 0;
2736 }
2737
2738 /*
2739  * This is for filesystems which do not implement ->writepage.
2740  */
2741 int generic_file_readonly_mmap(struct file *file, struct vm_area_struct *vma)
2742 {
2743         if ((vma->vm_flags & VM_SHARED) && (vma->vm_flags & VM_MAYWRITE))
2744                 return -EINVAL;
2745         return generic_file_mmap(file, vma);
2746 }
2747 #else
2748 vm_fault_t filemap_page_mkwrite(struct vm_fault *vmf)
2749 {
2750         return VM_FAULT_SIGBUS;
2751 }
2752 int generic_file_mmap(struct file * file, struct vm_area_struct * vma)
2753 {
2754         return -ENOSYS;
2755 }
2756 int generic_file_readonly_mmap(struct file * file, struct vm_area_struct * vma)
2757 {
2758         return -ENOSYS;
2759 }
2760 #endif /* CONFIG_MMU */
2761
2762 EXPORT_SYMBOL(filemap_page_mkwrite);
2763 EXPORT_SYMBOL(generic_file_mmap);
2764 EXPORT_SYMBOL(generic_file_readonly_mmap);
2765
2766 static struct page *wait_on_page_read(struct page *page)
2767 {
2768         if (!IS_ERR(page)) {
2769                 wait_on_page_locked(page);
2770                 if (!PageUptodate(page)) {
2771                         put_page(page);
2772                         page = ERR_PTR(-EIO);
2773                 }
2774         }
2775         return page;
2776 }
2777
2778 static struct page *do_read_cache_page(struct address_space *mapping,
2779                                 pgoff_t index,
2780                                 int (*filler)(void *, struct page *),
2781                                 void *data,
2782                                 gfp_t gfp)
2783 {
2784         struct page *page;
2785         int err;
2786 repeat:
2787         page = find_get_page(mapping, index);
2788         if (!page) {
2789                 page = __page_cache_alloc(gfp);
2790                 if (!page)
2791                         return ERR_PTR(-ENOMEM);
2792                 err = add_to_page_cache_lru(page, mapping, index, gfp);
2793                 if (unlikely(err)) {
2794                         put_page(page);
2795                         if (err == -EEXIST)
2796                                 goto repeat;
2797                         /* Presumably ENOMEM for radix tree node */
2798                         return ERR_PTR(err);
2799                 }
2800
2801 filler:
2802                 err = filler(data, page);
2803                 if (err < 0) {
2804                         put_page(page);
2805                         return ERR_PTR(err);
2806                 }
2807
2808                 page = wait_on_page_read(page);
2809                 if (IS_ERR(page))
2810                         return page;
2811                 goto out;
2812         }
2813         if (PageUptodate(page))
2814                 goto out;
2815
2816         /*
2817          * Page is not up to date and may be locked due one of the following
2818          * case a: Page is being filled and the page lock is held
2819          * case b: Read/write error clearing the page uptodate status
2820          * case c: Truncation in progress (page locked)
2821          * case d: Reclaim in progress
2822          *
2823          * Case a, the page will be up to date when the page is unlocked.
2824          *    There is no need to serialise on the page lock here as the page
2825          *    is pinned so the lock gives no additional protection. Even if the
2826          *    the page is truncated, the data is still valid if PageUptodate as
2827          *    it's a race vs truncate race.
2828          * Case b, the page will not be up to date
2829          * Case c, the page may be truncated but in itself, the data may still
2830          *    be valid after IO completes as it's a read vs truncate race. The
2831          *    operation must restart if the page is not uptodate on unlock but
2832          *    otherwise serialising on page lock to stabilise the mapping gives
2833          *    no additional guarantees to the caller as the page lock is
2834          *    released before return.
2835          * Case d, similar to truncation. If reclaim holds the page lock, it
2836          *    will be a race with remove_mapping that determines if the mapping
2837          *    is valid on unlock but otherwise the data is valid and there is
2838          *    no need to serialise with page lock.
2839          *
2840          * As the page lock gives no additional guarantee, we optimistically
2841          * wait on the page to be unlocked and check if it's up to date and
2842          * use the page if it is. Otherwise, the page lock is required to
2843          * distinguish between the different cases. The motivation is that we
2844          * avoid spurious serialisations and wakeups when multiple processes
2845          * wait on the same page for IO to complete.
2846          */
2847         wait_on_page_locked(page);
2848         if (PageUptodate(page))
2849                 goto out;
2850
2851         /* Distinguish between all the cases under the safety of the lock */
2852         lock_page(page);
2853
2854         /* Case c or d, restart the operation */
2855         if (!page->mapping) {
2856                 unlock_page(page);
2857                 put_page(page);
2858                 goto repeat;
2859         }
2860
2861         /* Someone else locked and filled the page in a very small window */
2862         if (PageUptodate(page)) {
2863                 unlock_page(page);
2864                 goto out;
2865         }
2866         goto filler;
2867
2868 out:
2869         mark_page_accessed(page);
2870         return page;
2871 }
2872
2873 /**
2874  * read_cache_page - read into page cache, fill it if needed
2875  * @mapping:    the page's address_space
2876  * @index:      the page index
2877  * @filler:     function to perform the read
2878  * @data:       first arg to filler(data, page) function, often left as NULL
2879  *
2880  * Read into the page cache. If a page already exists, and PageUptodate() is
2881  * not set, try to fill the page and wait for it to become unlocked.
2882  *
2883  * If the page does not get brought uptodate, return -EIO.
2884  */
2885 struct page *read_cache_page(struct address_space *mapping,
2886                                 pgoff_t index,
2887                                 int (*filler)(void *, struct page *),
2888                                 void *data)
2889 {
2890         return do_read_cache_page(mapping, index, filler, data, mapping_gfp_mask(mapping));
2891 }
2892 EXPORT_SYMBOL(read_cache_page);
2893
2894 /**
2895  * read_cache_page_gfp - read into page cache, using specified page allocation flags.
2896  * @mapping:    the page's address_space
2897  * @index:      the page index
2898  * @gfp:        the page allocator flags to use if allocating
2899  *
2900  * This is the same as "read_mapping_page(mapping, index, NULL)", but with
2901  * any new page allocations done using the specified allocation flags.
2902  *
2903  * If the page does not get brought uptodate, return -EIO.
2904  */
2905 struct page *read_cache_page_gfp(struct address_space *mapping,
2906                                 pgoff_t index,
2907                                 gfp_t gfp)
2908 {
2909         filler_t *filler = (filler_t *)mapping->a_ops->readpage;
2910
2911         return do_read_cache_page(mapping, index, filler, NULL, gfp);
2912 }
2913 EXPORT_SYMBOL(read_cache_page_gfp);
2914
2915 /*
2916  * Performs necessary checks before doing a write
2917  *
2918  * Can adjust writing position or amount of bytes to write.
2919  * Returns appropriate error code that caller should return or
2920  * zero in case that write should be allowed.
2921  */
2922 inline ssize_t generic_write_checks(struct kiocb *iocb, struct iov_iter *from)
2923 {
2924         struct file *file = iocb->ki_filp;
2925         struct inode *inode = file->f_mapping->host;
2926         unsigned long limit = rlimit(RLIMIT_FSIZE);
2927         loff_t pos;
2928
2929         if (!iov_iter_count(from))
2930                 return 0;
2931
2932         /* FIXME: this is for backwards compatibility with 2.4 */
2933         if (iocb->ki_flags & IOCB_APPEND)
2934                 iocb->ki_pos = i_size_read(inode);
2935
2936         pos = iocb->ki_pos;
2937
2938         if ((iocb->ki_flags & IOCB_NOWAIT) && !(iocb->ki_flags & IOCB_DIRECT))
2939                 return -EINVAL;
2940
2941         if (limit != RLIM_INFINITY) {
2942                 if (iocb->ki_pos >= limit) {
2943                         send_sig(SIGXFSZ, current, 0);
2944                         return -EFBIG;
2945                 }
2946                 iov_iter_truncate(from, limit - (unsigned long)pos);
2947         }
2948
2949         /*
2950          * LFS rule
2951          */
2952         if (unlikely(pos + iov_iter_count(from) > MAX_NON_LFS &&
2953                                 !(file->f_flags & O_LARGEFILE))) {
2954                 if (pos >= MAX_NON_LFS)
2955                         return -EFBIG;
2956                 iov_iter_truncate(from, MAX_NON_LFS - (unsigned long)pos);
2957         }
2958
2959         /*
2960          * Are we about to exceed the fs block limit ?
2961          *
2962          * If we have written data it becomes a short write.  If we have
2963          * exceeded without writing data we send a signal and return EFBIG.
2964          * Linus frestrict idea will clean these up nicely..
2965          */
2966         if (unlikely(pos >= inode->i_sb->s_maxbytes))
2967                 return -EFBIG;
2968
2969         iov_iter_truncate(from, inode->i_sb->s_maxbytes - pos);
2970         return iov_iter_count(from);
2971 }
2972 EXPORT_SYMBOL(generic_write_checks);
2973
2974 int pagecache_write_begin(struct file *file, struct address_space *mapping,
2975                                 loff_t pos, unsigned len, unsigned flags,
2976                                 struct page **pagep, void **fsdata)
2977 {
2978         const struct address_space_operations *aops = mapping->a_ops;
2979
2980         return aops->write_begin(file, mapping, pos, len, flags,
2981                                                         pagep, fsdata);
2982 }
2983 EXPORT_SYMBOL(pagecache_write_begin);
2984
2985 int pagecache_write_end(struct file *file, struct address_space *mapping,
2986                                 loff_t pos, unsigned len, unsigned copied,
2987                                 struct page *page, void *fsdata)
2988 {
2989         const struct address_space_operations *aops = mapping->a_ops;
2990
2991         return aops->write_end(file, mapping, pos, len, copied, page, fsdata);
2992 }
2993 EXPORT_SYMBOL(pagecache_write_end);
2994
2995 ssize_t
2996 generic_file_direct_write(struct kiocb *iocb, struct iov_iter *from)
2997 {
2998         struct file     *file = iocb->ki_filp;
2999         struct address_space *mapping = file->f_mapping;
3000         struct inode    *inode = mapping->host;
3001         loff_t          pos = iocb->ki_pos;
3002         ssize_t         written;
3003         size_t          write_len;
3004         pgoff_t         end;
3005
3006         write_len = iov_iter_count(from);
3007         end = (pos + write_len - 1) >> PAGE_SHIFT;
3008
3009         if (iocb->ki_flags & IOCB_NOWAIT) {
3010                 /* If there are pages to writeback, return */
3011                 if (filemap_range_has_page(inode->i_mapping, pos,
3012                                            pos + iov_iter_count(from)))
3013                         return -EAGAIN;
3014         } else {
3015                 written = filemap_write_and_wait_range(mapping, pos,
3016                                                         pos + write_len - 1);
3017                 if (written)
3018                         goto out;
3019         }
3020
3021         /*
3022          * After a write we want buffered reads to be sure to go to disk to get
3023          * the new data.  We invalidate clean cached page from the region we're
3024          * about to write.  We do this *before* the write so that we can return
3025          * without clobbering -EIOCBQUEUED from ->direct_IO().
3026          */
3027         written = invalidate_inode_pages2_range(mapping,
3028                                         pos >> PAGE_SHIFT, end);
3029         /*
3030          * If a page can not be invalidated, return 0 to fall back
3031          * to buffered write.
3032          */
3033         if (written) {
3034                 if (written == -EBUSY)
3035                         return 0;
3036                 goto out;
3037         }
3038
3039         written = mapping->a_ops->direct_IO(iocb, from);
3040
3041         /*
3042          * Finally, try again to invalidate clean pages which might have been
3043          * cached by non-direct readahead, or faulted in by get_user_pages()
3044          * if the source of the write was an mmap'ed region of the file
3045          * we're writing.  Either one is a pretty crazy thing to do,
3046          * so we don't support it 100%.  If this invalidation
3047          * fails, tough, the write still worked...
3048          *
3049          * Most of the time we do not need this since dio_complete() will do
3050          * the invalidation for us. However there are some file systems that
3051          * do not end up with dio_complete() being called, so let's not break
3052          * them by removing it completely
3053          */
3054         if (mapping->nrpages)
3055                 invalidate_inode_pages2_range(mapping,
3056                                         pos >> PAGE_SHIFT, end);
3057
3058         if (written > 0) {
3059                 pos += written;
3060                 write_len -= written;
3061                 if (pos > i_size_read(inode) && !S_ISBLK(inode->i_mode)) {
3062                         i_size_write(inode, pos);
3063                         mark_inode_dirty(inode);
3064                 }
3065                 iocb->ki_pos = pos;
3066         }
3067         iov_iter_revert(from, write_len - iov_iter_count(from));
3068 out:
3069         return written;
3070 }
3071 EXPORT_SYMBOL(generic_file_direct_write);
3072
3073 /*
3074  * Find or create a page at the given pagecache position. Return the locked
3075  * page. This function is specifically for buffered writes.
3076  */
3077 struct page *grab_cache_page_write_begin(struct address_space *mapping,
3078                                         pgoff_t index, unsigned flags)
3079 {
3080         struct page *page;
3081         int fgp_flags = FGP_LOCK|FGP_WRITE|FGP_CREAT;
3082
3083         if (flags & AOP_FLAG_NOFS)
3084                 fgp_flags |= FGP_NOFS;
3085
3086         page = pagecache_get_page(mapping, index, fgp_flags,
3087                         mapping_gfp_mask(mapping));
3088         if (page)
3089                 wait_for_stable_page(page);
3090
3091         return page;
3092 }
3093 EXPORT_SYMBOL(grab_cache_page_write_begin);
3094
3095 ssize_t generic_perform_write(struct file *file,
3096                                 struct iov_iter *i, loff_t pos)
3097 {
3098         struct address_space *mapping = file->f_mapping;
3099         const struct address_space_operations *a_ops = mapping->a_ops;
3100         long status = 0;
3101         ssize_t written = 0;
3102         unsigned int flags = 0;
3103
3104         do {
3105                 struct page *page;
3106                 unsigned long offset;   /* Offset into pagecache page */
3107                 unsigned long bytes;    /* Bytes to write to page */
3108                 size_t copied;          /* Bytes copied from user */
3109                 void *fsdata;
3110
3111                 offset = (pos & (PAGE_SIZE - 1));
3112                 bytes = min_t(unsigned long, PAGE_SIZE - offset,
3113                                                 iov_iter_count(i));
3114
3115 again:
3116                 /*
3117                  * Bring in the user page that we will copy from _first_.
3118                  * Otherwise there's a nasty deadlock on copying from the
3119                  * same page as we're writing to, without it being marked
3120                  * up-to-date.
3121                  *
3122                  * Not only is this an optimisation, but it is also required
3123                  * to check that the address is actually valid, when atomic
3124                  * usercopies are used, below.
3125                  */
3126                 if (unlikely(iov_iter_fault_in_readable(i, bytes))) {
3127                         status = -EFAULT;
3128                         break;
3129                 }
3130
3131                 if (fatal_signal_pending(current)) {
3132                         status = -EINTR;
3133                         break;
3134                 }
3135
3136                 status = a_ops->write_begin(file, mapping, pos, bytes, flags,
3137                                                 &page, &fsdata);
3138                 if (unlikely(status < 0))
3139                         break;
3140
3141                 if (mapping_writably_mapped(mapping))
3142                         flush_dcache_page(page);
3143
3144                 copied = iov_iter_copy_from_user_atomic(page, i, offset, bytes);
3145                 flush_dcache_page(page);
3146
3147                 status = a_ops->write_end(file, mapping, pos, bytes, copied,
3148                                                 page, fsdata);
3149                 if (unlikely(status < 0))
3150                         break;
3151                 copied = status;
3152
3153                 cond_resched();
3154
3155                 iov_iter_advance(i, copied);
3156                 if (unlikely(copied == 0)) {
3157                         /*
3158                          * If we were unable to copy any data at all, we must
3159                          * fall back to a single segment length write.
3160                          *
3161                          * If we didn't fallback here, we could livelock
3162                          * because not all segments in the iov can be copied at
3163                          * once without a pagefault.
3164                          */
3165                         bytes = min_t(unsigned long, PAGE_SIZE - offset,
3166                                                 iov_iter_single_seg_count(i));
3167                         goto again;
3168                 }
3169                 pos += copied;
3170                 written += copied;
3171
3172                 balance_dirty_pages_ratelimited(mapping);
3173         } while (iov_iter_count(i));
3174
3175         return written ? written : status;
3176 }
3177 EXPORT_SYMBOL(generic_perform_write);
3178
3179 /**
3180  * __generic_file_write_iter - write data to a file
3181  * @iocb:       IO state structure (file, offset, etc.)
3182  * @from:       iov_iter with data to write
3183  *
3184  * This function does all the work needed for actually writing data to a
3185  * file. It does all basic checks, removes SUID from the file, updates
3186  * modification times and calls proper subroutines depending on whether we
3187  * do direct IO or a standard buffered write.
3188  *
3189  * It expects i_mutex to be grabbed unless we work on a block device or similar
3190  * object which does not need locking at all.
3191  *
3192  * This function does *not* take care of syncing data in case of O_SYNC write.
3193  * A caller has to handle it. This is mainly due to the fact that we want to
3194  * avoid syncing under i_mutex.
3195  */
3196 ssize_t __generic_file_write_iter(struct kiocb *iocb, struct iov_iter *from)
3197 {
3198         struct file *file = iocb->ki_filp;
3199         struct address_space * mapping = file->f_mapping;
3200         struct inode    *inode = mapping->host;
3201         ssize_t         written = 0;
3202         ssize_t         err;
3203         ssize_t         status;
3204
3205         /* We can write back this queue in page reclaim */
3206         current->backing_dev_info = inode_to_bdi(inode);
3207         err = file_remove_privs(file);
3208         if (err)
3209                 goto out;
3210
3211         err = file_update_time(file);
3212         if (err)
3213                 goto out;
3214
3215         if (iocb->ki_flags & IOCB_DIRECT) {
3216                 loff_t pos, endbyte;
3217
3218                 written = generic_file_direct_write(iocb, from);
3219                 /*
3220                  * If the write stopped short of completing, fall back to
3221                  * buffered writes.  Some filesystems do this for writes to
3222                  * holes, for example.  For DAX files, a buffered write will
3223                  * not succeed (even if it did, DAX does not handle dirty
3224                  * page-cache pages correctly).
3225                  */
3226                 if (written < 0 || !iov_iter_count(from) || IS_DAX(inode))
3227                         goto out;
3228
3229                 status = generic_perform_write(file, from, pos = iocb->ki_pos);
3230                 /*
3231                  * If generic_perform_write() returned a synchronous error
3232                  * then we want to return the number of bytes which were
3233                  * direct-written, or the error code if that was zero.  Note
3234                  * that this differs from normal direct-io semantics, which
3235                  * will return -EFOO even if some bytes were written.
3236                  */
3237                 if (unlikely(status < 0)) {
3238                         err = status;
3239                         goto out;
3240                 }
3241                 /*
3242                  * We need to ensure that the page cache pages are written to
3243                  * disk and invalidated to preserve the expected O_DIRECT
3244                  * semantics.
3245                  */
3246                 endbyte = pos + status - 1;
3247                 err = filemap_write_and_wait_range(mapping, pos, endbyte);
3248                 if (err == 0) {
3249                         iocb->ki_pos = endbyte + 1;
3250                         written += status;
3251                         invalidate_mapping_pages(mapping,
3252                                                  pos >> PAGE_SHIFT,
3253                                                  endbyte >> PAGE_SHIFT);
3254                 } else {
3255                         /*
3256                          * We don't know how much we wrote, so just return
3257                          * the number of bytes which were direct-written
3258                          */
3259                 }
3260         } else {
3261                 written = generic_perform_write(file, from, iocb->ki_pos);
3262                 if (likely(written > 0))
3263                         iocb->ki_pos += written;
3264         }
3265 out:
3266         current->backing_dev_info = NULL;
3267         return written ? written : err;
3268 }
3269 EXPORT_SYMBOL(__generic_file_write_iter);
3270
3271 /**
3272  * generic_file_write_iter - write data to a file
3273  * @iocb:       IO state structure
3274  * @from:       iov_iter with data to write
3275  *
3276  * This is a wrapper around __generic_file_write_iter() to be used by most
3277  * filesystems. It takes care of syncing the file in case of O_SYNC file
3278  * and acquires i_mutex as needed.
3279  */
3280 ssize_t generic_file_write_iter(struct kiocb *iocb, struct iov_iter *from)
3281 {
3282         struct file *file = iocb->ki_filp;
3283         struct inode *inode = file->f_mapping->host;
3284         ssize_t ret;
3285
3286         inode_lock(inode);
3287         ret = generic_write_checks(iocb, from);
3288         if (ret > 0)
3289                 ret = __generic_file_write_iter(iocb, from);
3290         inode_unlock(inode);
3291
3292         if (ret > 0)
3293                 ret = generic_write_sync(iocb, ret);
3294         return ret;
3295 }
3296 EXPORT_SYMBOL(generic_file_write_iter);
3297
3298 /**
3299  * try_to_release_page() - release old fs-specific metadata on a page
3300  *
3301  * @page: the page which the kernel is trying to free
3302  * @gfp_mask: memory allocation flags (and I/O mode)
3303  *
3304  * The address_space is to try to release any data against the page
3305  * (presumably at page->private).  If the release was successful, return '1'.
3306  * Otherwise return zero.
3307  *
3308  * This may also be called if PG_fscache is set on a page, indicating that the
3309  * page is known to the local caching routines.
3310  *
3311  * The @gfp_mask argument specifies whether I/O may be performed to release
3312  * this page (__GFP_IO), and whether the call may block (__GFP_RECLAIM & __GFP_FS).
3313  *
3314  */
3315 int try_to_release_page(struct page *page, gfp_t gfp_mask)
3316 {
3317         struct address_space * const mapping = page->mapping;
3318
3319         BUG_ON(!PageLocked(page));
3320         if (PageWriteback(page))
3321                 return 0;
3322
3323         if (mapping && mapping->a_ops->releasepage)
3324                 return mapping->a_ops->releasepage(page, gfp_mask);
3325         return try_to_free_buffers(page);
3326 }
3327
3328 EXPORT_SYMBOL(try_to_release_page);