Merge tag 'pinctrl-v4.15-1' of git://git.kernel.org/pub/scm/linux/kernel/git/linusw...
[sfrench/cifs-2.6.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/hardirq.h> /* for BUG_ON(!in_atomic()) only */
35 #include <linux/hugetlb.h>
36 #include <linux/memcontrol.h>
37 #include <linux/cleancache.h>
38 #include <linux/shmem_fs.h>
39 #include <linux/rmap.h>
40 #include "internal.h"
41
42 #define CREATE_TRACE_POINTS
43 #include <trace/events/filemap.h>
44
45 /*
46  * FIXME: remove all knowledge of the buffer layer from the core VM
47  */
48 #include <linux/buffer_head.h> /* for try_to_free_buffers */
49
50 #include <asm/mman.h>
51
52 /*
53  * Shared mappings implemented 30.11.1994. It's not fully working yet,
54  * though.
55  *
56  * Shared mappings now work. 15.8.1995  Bruno.
57  *
58  * finished 'unifying' the page and buffer cache and SMP-threaded the
59  * page-cache, 21.05.1999, Ingo Molnar <mingo@redhat.com>
60  *
61  * SMP-threaded pagemap-LRU 1999, Andrea Arcangeli <andrea@suse.de>
62  */
63
64 /*
65  * Lock ordering:
66  *
67  *  ->i_mmap_rwsem              (truncate_pagecache)
68  *    ->private_lock            (__free_pte->__set_page_dirty_buffers)
69  *      ->swap_lock             (exclusive_swap_page, others)
70  *        ->mapping->tree_lock
71  *
72  *  ->i_mutex
73  *    ->i_mmap_rwsem            (truncate->unmap_mapping_range)
74  *
75  *  ->mmap_sem
76  *    ->i_mmap_rwsem
77  *      ->page_table_lock or pte_lock   (various, mainly in memory.c)
78  *        ->mapping->tree_lock  (arch-dependent flush_dcache_mmap_lock)
79  *
80  *  ->mmap_sem
81  *    ->lock_page               (access_process_vm)
82  *
83  *  ->i_mutex                   (generic_perform_write)
84  *    ->mmap_sem                (fault_in_pages_readable->do_page_fault)
85  *
86  *  bdi->wb.list_lock
87  *    sb_lock                   (fs/fs-writeback.c)
88  *    ->mapping->tree_lock      (__sync_single_inode)
89  *
90  *  ->i_mmap_rwsem
91  *    ->anon_vma.lock           (vma_adjust)
92  *
93  *  ->anon_vma.lock
94  *    ->page_table_lock or pte_lock     (anon_vma_prepare and various)
95  *
96  *  ->page_table_lock or pte_lock
97  *    ->swap_lock               (try_to_unmap_one)
98  *    ->private_lock            (try_to_unmap_one)
99  *    ->tree_lock               (try_to_unmap_one)
100  *    ->zone_lru_lock(zone)     (follow_page->mark_page_accessed)
101  *    ->zone_lru_lock(zone)     (check_pte_range->isolate_lru_page)
102  *    ->private_lock            (page_remove_rmap->set_page_dirty)
103  *    ->tree_lock               (page_remove_rmap->set_page_dirty)
104  *    bdi.wb->list_lock         (page_remove_rmap->set_page_dirty)
105  *    ->inode->i_lock           (page_remove_rmap->set_page_dirty)
106  *    ->memcg->move_lock        (page_remove_rmap->lock_page_memcg)
107  *    bdi.wb->list_lock         (zap_pte_range->set_page_dirty)
108  *    ->inode->i_lock           (zap_pte_range->set_page_dirty)
109  *    ->private_lock            (zap_pte_range->__set_page_dirty_buffers)
110  *
111  * ->i_mmap_rwsem
112  *   ->tasklist_lock            (memory_failure, collect_procs_ao)
113  */
114
115 static int page_cache_tree_insert(struct address_space *mapping,
116                                   struct page *page, void **shadowp)
117 {
118         struct radix_tree_node *node;
119         void **slot;
120         int error;
121
122         error = __radix_tree_create(&mapping->page_tree, page->index, 0,
123                                     &node, &slot);
124         if (error)
125                 return error;
126         if (*slot) {
127                 void *p;
128
129                 p = radix_tree_deref_slot_protected(slot, &mapping->tree_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->page_tree, 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->page_tree, page->index + i,
160                                     &node, &slot);
161
162                 VM_BUG_ON_PAGE(!node && nr != 1, page);
163
164                 radix_tree_clear_tags(&mapping->page_tree, node, slot);
165                 __radix_tree_replace(&mapping->page_tree, 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 mapping's tree_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         spin_lock_irqsave(&mapping->tree_lock, flags);
301         __delete_from_page_cache(page, NULL);
302         spin_unlock_irqrestore(&mapping->tree_lock, 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->page_tree and removes pages passed in @pvec
314  * from the radix tree. The function expects @pvec to be sorted by page index.
315  * It tolerates holes in @pvec (radix tree 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 radix
318  * tree as well.
319  *
320  * The function expects mapping->tree_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->page_tree, &iter, start) {
335                 if (i >= pagevec_count(pvec) && !tail_pages)
336                         break;
337                 page = radix_tree_deref_slot_protected(slot,
338                                                        &mapping->tree_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->page_tree, iter.node, slot);
362                 __radix_tree_replace(&mapping->page_tree, 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         spin_lock_irqsave(&mapping->tree_lock, 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         spin_unlock_irqrestore(&mapping->tree_lock, 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_HIGHMEM);
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                 spin_lock_irqsave(&mapping->tree_lock, 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                 spin_unlock_irqrestore(&mapping->tree_lock, 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_HIGHMEM);
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         spin_lock_irq(&mapping->tree_lock);
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         spin_unlock_irq(&mapping->tree_lock);
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         spin_unlock_irq(&mapping->tree_lock);
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                 if (!(gfp_mask & __GFP_WRITE) &&
919                     shadow && workingset_refault(shadow)) {
920                         SetPageActive(page);
921                         workingset_activation(page);
922                 } else
923                         ClearPageActive(page);
924                 lru_cache_add(page);
925         }
926         return ret;
927 }
928 EXPORT_SYMBOL_GPL(add_to_page_cache_lru);
929
930 #ifdef CONFIG_NUMA
931 struct page *__page_cache_alloc(gfp_t gfp)
932 {
933         int n;
934         struct page *page;
935
936         if (cpuset_do_page_mem_spread()) {
937                 unsigned int cpuset_mems_cookie;
938                 do {
939                         cpuset_mems_cookie = read_mems_allowed_begin();
940                         n = cpuset_mem_spread_node();
941                         page = __alloc_pages_node(n, gfp, 0);
942                 } while (!page && read_mems_allowed_retry(cpuset_mems_cookie));
943
944                 return page;
945         }
946         return alloc_pages(gfp, 0);
947 }
948 EXPORT_SYMBOL(__page_cache_alloc);
949 #endif
950
951 /*
952  * In order to wait for pages to become available there must be
953  * waitqueues associated with pages. By using a hash table of
954  * waitqueues where the bucket discipline is to maintain all
955  * waiters on the same queue and wake all when any of the pages
956  * become available, and for the woken contexts to check to be
957  * sure the appropriate page became available, this saves space
958  * at a cost of "thundering herd" phenomena during rare hash
959  * collisions.
960  */
961 #define PAGE_WAIT_TABLE_BITS 8
962 #define PAGE_WAIT_TABLE_SIZE (1 << PAGE_WAIT_TABLE_BITS)
963 static wait_queue_head_t page_wait_table[PAGE_WAIT_TABLE_SIZE] __cacheline_aligned;
964
965 static wait_queue_head_t *page_waitqueue(struct page *page)
966 {
967         return &page_wait_table[hash_ptr(page, PAGE_WAIT_TABLE_BITS)];
968 }
969
970 void __init pagecache_init(void)
971 {
972         int i;
973
974         for (i = 0; i < PAGE_WAIT_TABLE_SIZE; i++)
975                 init_waitqueue_head(&page_wait_table[i]);
976
977         page_writeback_init();
978 }
979
980 /* This has the same layout as wait_bit_key - see fs/cachefiles/rdwr.c */
981 struct wait_page_key {
982         struct page *page;
983         int bit_nr;
984         int page_match;
985 };
986
987 struct wait_page_queue {
988         struct page *page;
989         int bit_nr;
990         wait_queue_entry_t wait;
991 };
992
993 static int wake_page_function(wait_queue_entry_t *wait, unsigned mode, int sync, void *arg)
994 {
995         struct wait_page_key *key = arg;
996         struct wait_page_queue *wait_page
997                 = container_of(wait, struct wait_page_queue, wait);
998
999         if (wait_page->page != key->page)
1000                return 0;
1001         key->page_match = 1;
1002
1003         if (wait_page->bit_nr != key->bit_nr)
1004                 return 0;
1005
1006         /* Stop walking if it's locked */
1007         if (test_bit(key->bit_nr, &key->page->flags))
1008                 return -1;
1009
1010         return autoremove_wake_function(wait, mode, sync, key);
1011 }
1012
1013 static void wake_up_page_bit(struct page *page, int bit_nr)
1014 {
1015         wait_queue_head_t *q = page_waitqueue(page);
1016         struct wait_page_key key;
1017         unsigned long flags;
1018         wait_queue_entry_t bookmark;
1019
1020         key.page = page;
1021         key.bit_nr = bit_nr;
1022         key.page_match = 0;
1023
1024         bookmark.flags = 0;
1025         bookmark.private = NULL;
1026         bookmark.func = NULL;
1027         INIT_LIST_HEAD(&bookmark.entry);
1028
1029         spin_lock_irqsave(&q->lock, flags);
1030         __wake_up_locked_key_bookmark(q, TASK_NORMAL, &key, &bookmark);
1031
1032         while (bookmark.flags & WQ_FLAG_BOOKMARK) {
1033                 /*
1034                  * Take a breather from holding the lock,
1035                  * allow pages that finish wake up asynchronously
1036                  * to acquire the lock and remove themselves
1037                  * from wait queue
1038                  */
1039                 spin_unlock_irqrestore(&q->lock, flags);
1040                 cpu_relax();
1041                 spin_lock_irqsave(&q->lock, flags);
1042                 __wake_up_locked_key_bookmark(q, TASK_NORMAL, &key, &bookmark);
1043         }
1044
1045         /*
1046          * It is possible for other pages to have collided on the waitqueue
1047          * hash, so in that case check for a page match. That prevents a long-
1048          * term waiter
1049          *
1050          * It is still possible to miss a case here, when we woke page waiters
1051          * and removed them from the waitqueue, but there are still other
1052          * page waiters.
1053          */
1054         if (!waitqueue_active(q) || !key.page_match) {
1055                 ClearPageWaiters(page);
1056                 /*
1057                  * It's possible to miss clearing Waiters here, when we woke
1058                  * our page waiters, but the hashed waitqueue has waiters for
1059                  * other pages on it.
1060                  *
1061                  * That's okay, it's a rare case. The next waker will clear it.
1062                  */
1063         }
1064         spin_unlock_irqrestore(&q->lock, flags);
1065 }
1066
1067 static void wake_up_page(struct page *page, int bit)
1068 {
1069         if (!PageWaiters(page))
1070                 return;
1071         wake_up_page_bit(page, bit);
1072 }
1073
1074 static inline int wait_on_page_bit_common(wait_queue_head_t *q,
1075                 struct page *page, int bit_nr, int state, bool lock)
1076 {
1077         struct wait_page_queue wait_page;
1078         wait_queue_entry_t *wait = &wait_page.wait;
1079         int ret = 0;
1080
1081         init_wait(wait);
1082         wait->flags = lock ? WQ_FLAG_EXCLUSIVE : 0;
1083         wait->func = wake_page_function;
1084         wait_page.page = page;
1085         wait_page.bit_nr = bit_nr;
1086
1087         for (;;) {
1088                 spin_lock_irq(&q->lock);
1089
1090                 if (likely(list_empty(&wait->entry))) {
1091                         __add_wait_queue_entry_tail(q, wait);
1092                         SetPageWaiters(page);
1093                 }
1094
1095                 set_current_state(state);
1096
1097                 spin_unlock_irq(&q->lock);
1098
1099                 if (likely(test_bit(bit_nr, &page->flags))) {
1100                         io_schedule();
1101                 }
1102
1103                 if (lock) {
1104                         if (!test_and_set_bit_lock(bit_nr, &page->flags))
1105                                 break;
1106                 } else {
1107                         if (!test_bit(bit_nr, &page->flags))
1108                                 break;
1109                 }
1110
1111                 if (unlikely(signal_pending_state(state, current))) {
1112                         ret = -EINTR;
1113                         break;
1114                 }
1115         }
1116
1117         finish_wait(q, wait);
1118
1119         /*
1120          * A signal could leave PageWaiters set. Clearing it here if
1121          * !waitqueue_active would be possible (by open-coding finish_wait),
1122          * but still fail to catch it in the case of wait hash collision. We
1123          * already can fail to clear wait hash collision cases, so don't
1124          * bother with signals either.
1125          */
1126
1127         return ret;
1128 }
1129
1130 void wait_on_page_bit(struct page *page, int bit_nr)
1131 {
1132         wait_queue_head_t *q = page_waitqueue(page);
1133         wait_on_page_bit_common(q, page, bit_nr, TASK_UNINTERRUPTIBLE, false);
1134 }
1135 EXPORT_SYMBOL(wait_on_page_bit);
1136
1137 int wait_on_page_bit_killable(struct page *page, int bit_nr)
1138 {
1139         wait_queue_head_t *q = page_waitqueue(page);
1140         return wait_on_page_bit_common(q, page, bit_nr, TASK_KILLABLE, false);
1141 }
1142
1143 /**
1144  * add_page_wait_queue - Add an arbitrary waiter to a page's wait queue
1145  * @page: Page defining the wait queue of interest
1146  * @waiter: Waiter to add to the queue
1147  *
1148  * Add an arbitrary @waiter to the wait queue for the nominated @page.
1149  */
1150 void add_page_wait_queue(struct page *page, wait_queue_entry_t *waiter)
1151 {
1152         wait_queue_head_t *q = page_waitqueue(page);
1153         unsigned long flags;
1154
1155         spin_lock_irqsave(&q->lock, flags);
1156         __add_wait_queue_entry_tail(q, waiter);
1157         SetPageWaiters(page);
1158         spin_unlock_irqrestore(&q->lock, flags);
1159 }
1160 EXPORT_SYMBOL_GPL(add_page_wait_queue);
1161
1162 #ifndef clear_bit_unlock_is_negative_byte
1163
1164 /*
1165  * PG_waiters is the high bit in the same byte as PG_lock.
1166  *
1167  * On x86 (and on many other architectures), we can clear PG_lock and
1168  * test the sign bit at the same time. But if the architecture does
1169  * not support that special operation, we just do this all by hand
1170  * instead.
1171  *
1172  * The read of PG_waiters has to be after (or concurrently with) PG_locked
1173  * being cleared, but a memory barrier should be unneccssary since it is
1174  * in the same byte as PG_locked.
1175  */
1176 static inline bool clear_bit_unlock_is_negative_byte(long nr, volatile void *mem)
1177 {
1178         clear_bit_unlock(nr, mem);
1179         /* smp_mb__after_atomic(); */
1180         return test_bit(PG_waiters, mem);
1181 }
1182
1183 #endif
1184
1185 /**
1186  * unlock_page - unlock a locked page
1187  * @page: the page
1188  *
1189  * Unlocks the page and wakes up sleepers in ___wait_on_page_locked().
1190  * Also wakes sleepers in wait_on_page_writeback() because the wakeup
1191  * mechanism between PageLocked pages and PageWriteback pages is shared.
1192  * But that's OK - sleepers in wait_on_page_writeback() just go back to sleep.
1193  *
1194  * Note that this depends on PG_waiters being the sign bit in the byte
1195  * that contains PG_locked - thus the BUILD_BUG_ON(). That allows us to
1196  * clear the PG_locked bit and test PG_waiters at the same time fairly
1197  * portably (architectures that do LL/SC can test any bit, while x86 can
1198  * test the sign bit).
1199  */
1200 void unlock_page(struct page *page)
1201 {
1202         BUILD_BUG_ON(PG_waiters != 7);
1203         page = compound_head(page);
1204         VM_BUG_ON_PAGE(!PageLocked(page), page);
1205         if (clear_bit_unlock_is_negative_byte(PG_locked, &page->flags))
1206                 wake_up_page_bit(page, PG_locked);
1207 }
1208 EXPORT_SYMBOL(unlock_page);
1209
1210 /**
1211  * end_page_writeback - end writeback against a page
1212  * @page: the page
1213  */
1214 void end_page_writeback(struct page *page)
1215 {
1216         /*
1217          * TestClearPageReclaim could be used here but it is an atomic
1218          * operation and overkill in this particular case. Failing to
1219          * shuffle a page marked for immediate reclaim is too mild to
1220          * justify taking an atomic operation penalty at the end of
1221          * ever page writeback.
1222          */
1223         if (PageReclaim(page)) {
1224                 ClearPageReclaim(page);
1225                 rotate_reclaimable_page(page);
1226         }
1227
1228         if (!test_clear_page_writeback(page))
1229                 BUG();
1230
1231         smp_mb__after_atomic();
1232         wake_up_page(page, PG_writeback);
1233 }
1234 EXPORT_SYMBOL(end_page_writeback);
1235
1236 /*
1237  * After completing I/O on a page, call this routine to update the page
1238  * flags appropriately
1239  */
1240 void page_endio(struct page *page, bool is_write, int err)
1241 {
1242         if (!is_write) {
1243                 if (!err) {
1244                         SetPageUptodate(page);
1245                 } else {
1246                         ClearPageUptodate(page);
1247                         SetPageError(page);
1248                 }
1249                 unlock_page(page);
1250         } else {
1251                 if (err) {
1252                         struct address_space *mapping;
1253
1254                         SetPageError(page);
1255                         mapping = page_mapping(page);
1256                         if (mapping)
1257                                 mapping_set_error(mapping, err);
1258                 }
1259                 end_page_writeback(page);
1260         }
1261 }
1262 EXPORT_SYMBOL_GPL(page_endio);
1263
1264 /**
1265  * __lock_page - get a lock on the page, assuming we need to sleep to get it
1266  * @__page: the page to lock
1267  */
1268 void __lock_page(struct page *__page)
1269 {
1270         struct page *page = compound_head(__page);
1271         wait_queue_head_t *q = page_waitqueue(page);
1272         wait_on_page_bit_common(q, page, PG_locked, TASK_UNINTERRUPTIBLE, true);
1273 }
1274 EXPORT_SYMBOL(__lock_page);
1275
1276 int __lock_page_killable(struct page *__page)
1277 {
1278         struct page *page = compound_head(__page);
1279         wait_queue_head_t *q = page_waitqueue(page);
1280         return wait_on_page_bit_common(q, page, PG_locked, TASK_KILLABLE, true);
1281 }
1282 EXPORT_SYMBOL_GPL(__lock_page_killable);
1283
1284 /*
1285  * Return values:
1286  * 1 - page is locked; mmap_sem is still held.
1287  * 0 - page is not locked.
1288  *     mmap_sem has been released (up_read()), unless flags had both
1289  *     FAULT_FLAG_ALLOW_RETRY and FAULT_FLAG_RETRY_NOWAIT set, in
1290  *     which case mmap_sem is still held.
1291  *
1292  * If neither ALLOW_RETRY nor KILLABLE are set, will always return 1
1293  * with the page locked and the mmap_sem unperturbed.
1294  */
1295 int __lock_page_or_retry(struct page *page, struct mm_struct *mm,
1296                          unsigned int flags)
1297 {
1298         if (flags & FAULT_FLAG_ALLOW_RETRY) {
1299                 /*
1300                  * CAUTION! In this case, mmap_sem is not released
1301                  * even though return 0.
1302                  */
1303                 if (flags & FAULT_FLAG_RETRY_NOWAIT)
1304                         return 0;
1305
1306                 up_read(&mm->mmap_sem);
1307                 if (flags & FAULT_FLAG_KILLABLE)
1308                         wait_on_page_locked_killable(page);
1309                 else
1310                         wait_on_page_locked(page);
1311                 return 0;
1312         } else {
1313                 if (flags & FAULT_FLAG_KILLABLE) {
1314                         int ret;
1315
1316                         ret = __lock_page_killable(page);
1317                         if (ret) {
1318                                 up_read(&mm->mmap_sem);
1319                                 return 0;
1320                         }
1321                 } else
1322                         __lock_page(page);
1323                 return 1;
1324         }
1325 }
1326
1327 /**
1328  * page_cache_next_hole - find the next hole (not-present entry)
1329  * @mapping: mapping
1330  * @index: index
1331  * @max_scan: maximum range to search
1332  *
1333  * Search the set [index, min(index+max_scan-1, MAX_INDEX)] for the
1334  * lowest indexed hole.
1335  *
1336  * Returns: the index of the hole if found, otherwise returns an index
1337  * outside of the set specified (in which case 'return - index >=
1338  * max_scan' will be true). In rare cases of index wrap-around, 0 will
1339  * be returned.
1340  *
1341  * page_cache_next_hole may be called under rcu_read_lock. However,
1342  * like radix_tree_gang_lookup, this will not atomically search a
1343  * snapshot of the tree at a single point in time. For example, if a
1344  * hole is created at index 5, then subsequently a hole is created at
1345  * index 10, page_cache_next_hole covering both indexes may return 10
1346  * if called under rcu_read_lock.
1347  */
1348 pgoff_t page_cache_next_hole(struct address_space *mapping,
1349                              pgoff_t index, unsigned long max_scan)
1350 {
1351         unsigned long i;
1352
1353         for (i = 0; i < max_scan; i++) {
1354                 struct page *page;
1355
1356                 page = radix_tree_lookup(&mapping->page_tree, index);
1357                 if (!page || radix_tree_exceptional_entry(page))
1358                         break;
1359                 index++;
1360                 if (index == 0)
1361                         break;
1362         }
1363
1364         return index;
1365 }
1366 EXPORT_SYMBOL(page_cache_next_hole);
1367
1368 /**
1369  * page_cache_prev_hole - find the prev hole (not-present entry)
1370  * @mapping: mapping
1371  * @index: index
1372  * @max_scan: maximum range to search
1373  *
1374  * Search backwards in the range [max(index-max_scan+1, 0), index] for
1375  * the first hole.
1376  *
1377  * Returns: the index of the hole if found, otherwise returns an index
1378  * outside of the set specified (in which case 'index - return >=
1379  * max_scan' will be true). In rare cases of wrap-around, ULONG_MAX
1380  * will be returned.
1381  *
1382  * page_cache_prev_hole may be called under rcu_read_lock. However,
1383  * like radix_tree_gang_lookup, this will not atomically search a
1384  * snapshot of the tree at a single point in time. For example, if a
1385  * hole is created at index 10, then subsequently a hole is created at
1386  * index 5, page_cache_prev_hole covering both indexes may return 5 if
1387  * called under rcu_read_lock.
1388  */
1389 pgoff_t page_cache_prev_hole(struct address_space *mapping,
1390                              pgoff_t index, unsigned long max_scan)
1391 {
1392         unsigned long i;
1393
1394         for (i = 0; i < max_scan; i++) {
1395                 struct page *page;
1396
1397                 page = radix_tree_lookup(&mapping->page_tree, index);
1398                 if (!page || radix_tree_exceptional_entry(page))
1399                         break;
1400                 index--;
1401                 if (index == ULONG_MAX)
1402                         break;
1403         }
1404
1405         return index;
1406 }
1407 EXPORT_SYMBOL(page_cache_prev_hole);
1408
1409 /**
1410  * find_get_entry - find and get a page cache entry
1411  * @mapping: the address_space to search
1412  * @offset: the page cache index
1413  *
1414  * Looks up the page cache slot at @mapping & @offset.  If there is a
1415  * page cache page, it is returned with an increased refcount.
1416  *
1417  * If the slot holds a shadow entry of a previously evicted page, or a
1418  * swap entry from shmem/tmpfs, it is returned.
1419  *
1420  * Otherwise, %NULL is returned.
1421  */
1422 struct page *find_get_entry(struct address_space *mapping, pgoff_t offset)
1423 {
1424         void **pagep;
1425         struct page *head, *page;
1426
1427         rcu_read_lock();
1428 repeat:
1429         page = NULL;
1430         pagep = radix_tree_lookup_slot(&mapping->page_tree, offset);
1431         if (pagep) {
1432                 page = radix_tree_deref_slot(pagep);
1433                 if (unlikely(!page))
1434                         goto out;
1435                 if (radix_tree_exception(page)) {
1436                         if (radix_tree_deref_retry(page))
1437                                 goto repeat;
1438                         /*
1439                          * A shadow entry of a recently evicted page,
1440                          * or a swap entry from shmem/tmpfs.  Return
1441                          * it without attempting to raise page count.
1442                          */
1443                         goto out;
1444                 }
1445
1446                 head = compound_head(page);
1447                 if (!page_cache_get_speculative(head))
1448                         goto repeat;
1449
1450                 /* The page was split under us? */
1451                 if (compound_head(page) != head) {
1452                         put_page(head);
1453                         goto repeat;
1454                 }
1455
1456                 /*
1457                  * Has the page moved?
1458                  * This is part of the lockless pagecache protocol. See
1459                  * include/linux/pagemap.h for details.
1460                  */
1461                 if (unlikely(page != *pagep)) {
1462                         put_page(head);
1463                         goto repeat;
1464                 }
1465         }
1466 out:
1467         rcu_read_unlock();
1468
1469         return page;
1470 }
1471 EXPORT_SYMBOL(find_get_entry);
1472
1473 /**
1474  * find_lock_entry - locate, pin and lock a page cache entry
1475  * @mapping: the address_space to search
1476  * @offset: the page cache index
1477  *
1478  * Looks up the page cache slot at @mapping & @offset.  If there is a
1479  * page cache page, it is returned locked and with an increased
1480  * refcount.
1481  *
1482  * If the slot holds a shadow entry of a previously evicted page, or a
1483  * swap entry from shmem/tmpfs, it is returned.
1484  *
1485  * Otherwise, %NULL is returned.
1486  *
1487  * find_lock_entry() may sleep.
1488  */
1489 struct page *find_lock_entry(struct address_space *mapping, pgoff_t offset)
1490 {
1491         struct page *page;
1492
1493 repeat:
1494         page = find_get_entry(mapping, offset);
1495         if (page && !radix_tree_exception(page)) {
1496                 lock_page(page);
1497                 /* Has the page been truncated? */
1498                 if (unlikely(page_mapping(page) != mapping)) {
1499                         unlock_page(page);
1500                         put_page(page);
1501                         goto repeat;
1502                 }
1503                 VM_BUG_ON_PAGE(page_to_pgoff(page) != offset, page);
1504         }
1505         return page;
1506 }
1507 EXPORT_SYMBOL(find_lock_entry);
1508
1509 /**
1510  * pagecache_get_page - find and get a page reference
1511  * @mapping: the address_space to search
1512  * @offset: the page index
1513  * @fgp_flags: PCG flags
1514  * @gfp_mask: gfp mask to use for the page cache data page allocation
1515  *
1516  * Looks up the page cache slot at @mapping & @offset.
1517  *
1518  * PCG flags modify how the page is returned.
1519  *
1520  * @fgp_flags can be:
1521  *
1522  * - FGP_ACCESSED: the page will be marked accessed
1523  * - FGP_LOCK: Page is return locked
1524  * - FGP_CREAT: If page is not present then a new page is allocated using
1525  *   @gfp_mask and added to the page cache and the VM's LRU
1526  *   list. The page is returned locked and with an increased
1527  *   refcount. Otherwise, NULL is returned.
1528  *
1529  * If FGP_LOCK or FGP_CREAT are specified then the function may sleep even
1530  * if the GFP flags specified for FGP_CREAT are atomic.
1531  *
1532  * If there is a page cache page, it is returned with an increased refcount.
1533  */
1534 struct page *pagecache_get_page(struct address_space *mapping, pgoff_t offset,
1535         int fgp_flags, gfp_t gfp_mask)
1536 {
1537         struct page *page;
1538
1539 repeat:
1540         page = find_get_entry(mapping, offset);
1541         if (radix_tree_exceptional_entry(page))
1542                 page = NULL;
1543         if (!page)
1544                 goto no_page;
1545
1546         if (fgp_flags & FGP_LOCK) {
1547                 if (fgp_flags & FGP_NOWAIT) {
1548                         if (!trylock_page(page)) {
1549                                 put_page(page);
1550                                 return NULL;
1551                         }
1552                 } else {
1553                         lock_page(page);
1554                 }
1555
1556                 /* Has the page been truncated? */
1557                 if (unlikely(page->mapping != mapping)) {
1558                         unlock_page(page);
1559                         put_page(page);
1560                         goto repeat;
1561                 }
1562                 VM_BUG_ON_PAGE(page->index != offset, page);
1563         }
1564
1565         if (page && (fgp_flags & FGP_ACCESSED))
1566                 mark_page_accessed(page);
1567
1568 no_page:
1569         if (!page && (fgp_flags & FGP_CREAT)) {
1570                 int err;
1571                 if ((fgp_flags & FGP_WRITE) && mapping_cap_account_dirty(mapping))
1572                         gfp_mask |= __GFP_WRITE;
1573                 if (fgp_flags & FGP_NOFS)
1574                         gfp_mask &= ~__GFP_FS;
1575
1576                 page = __page_cache_alloc(gfp_mask);
1577                 if (!page)
1578                         return NULL;
1579
1580                 if (WARN_ON_ONCE(!(fgp_flags & FGP_LOCK)))
1581                         fgp_flags |= FGP_LOCK;
1582
1583                 /* Init accessed so avoid atomic mark_page_accessed later */
1584                 if (fgp_flags & FGP_ACCESSED)
1585                         __SetPageReferenced(page);
1586
1587                 err = add_to_page_cache_lru(page, mapping, offset,
1588                                 gfp_mask & GFP_RECLAIM_MASK);
1589                 if (unlikely(err)) {
1590                         put_page(page);
1591                         page = NULL;
1592                         if (err == -EEXIST)
1593                                 goto repeat;
1594                 }
1595         }
1596
1597         return page;
1598 }
1599 EXPORT_SYMBOL(pagecache_get_page);
1600
1601 /**
1602  * find_get_entries - gang pagecache lookup
1603  * @mapping:    The address_space to search
1604  * @start:      The starting page cache index
1605  * @nr_entries: The maximum number of entries
1606  * @entries:    Where the resulting entries are placed
1607  * @indices:    The cache indices corresponding to the entries in @entries
1608  *
1609  * find_get_entries() will search for and return a group of up to
1610  * @nr_entries entries in the mapping.  The entries are placed at
1611  * @entries.  find_get_entries() takes a reference against any actual
1612  * pages it returns.
1613  *
1614  * The search returns a group of mapping-contiguous page cache entries
1615  * with ascending indexes.  There may be holes in the indices due to
1616  * not-present pages.
1617  *
1618  * Any shadow entries of evicted pages, or swap entries from
1619  * shmem/tmpfs, are included in the returned array.
1620  *
1621  * find_get_entries() returns the number of pages and shadow entries
1622  * which were found.
1623  */
1624 unsigned find_get_entries(struct address_space *mapping,
1625                           pgoff_t start, unsigned int nr_entries,
1626                           struct page **entries, pgoff_t *indices)
1627 {
1628         void **slot;
1629         unsigned int ret = 0;
1630         struct radix_tree_iter iter;
1631
1632         if (!nr_entries)
1633                 return 0;
1634
1635         rcu_read_lock();
1636         radix_tree_for_each_slot(slot, &mapping->page_tree, &iter, start) {
1637                 struct page *head, *page;
1638 repeat:
1639                 page = radix_tree_deref_slot(slot);
1640                 if (unlikely(!page))
1641                         continue;
1642                 if (radix_tree_exception(page)) {
1643                         if (radix_tree_deref_retry(page)) {
1644                                 slot = radix_tree_iter_retry(&iter);
1645                                 continue;
1646                         }
1647                         /*
1648                          * A shadow entry of a recently evicted page, a swap
1649                          * entry from shmem/tmpfs or a DAX entry.  Return it
1650                          * without attempting to raise page count.
1651                          */
1652                         goto export;
1653                 }
1654
1655                 head = compound_head(page);
1656                 if (!page_cache_get_speculative(head))
1657                         goto repeat;
1658
1659                 /* The page was split under us? */
1660                 if (compound_head(page) != head) {
1661                         put_page(head);
1662                         goto repeat;
1663                 }
1664
1665                 /* Has the page moved? */
1666                 if (unlikely(page != *slot)) {
1667                         put_page(head);
1668                         goto repeat;
1669                 }
1670 export:
1671                 indices[ret] = iter.index;
1672                 entries[ret] = page;
1673                 if (++ret == nr_entries)
1674                         break;
1675         }
1676         rcu_read_unlock();
1677         return ret;
1678 }
1679
1680 /**
1681  * find_get_pages_range - gang pagecache lookup
1682  * @mapping:    The address_space to search
1683  * @start:      The starting page index
1684  * @end:        The final page index (inclusive)
1685  * @nr_pages:   The maximum number of pages
1686  * @pages:      Where the resulting pages are placed
1687  *
1688  * find_get_pages_range() will search for and return a group of up to @nr_pages
1689  * pages in the mapping starting at index @start and up to index @end
1690  * (inclusive).  The pages are placed at @pages.  find_get_pages_range() takes
1691  * a reference against the returned pages.
1692  *
1693  * The search returns a group of mapping-contiguous pages with ascending
1694  * indexes.  There may be holes in the indices due to not-present pages.
1695  * We also update @start to index the next page for the traversal.
1696  *
1697  * find_get_pages_range() returns the number of pages which were found. If this
1698  * number is smaller than @nr_pages, the end of specified range has been
1699  * reached.
1700  */
1701 unsigned find_get_pages_range(struct address_space *mapping, pgoff_t *start,
1702                               pgoff_t end, unsigned int nr_pages,
1703                               struct page **pages)
1704 {
1705         struct radix_tree_iter iter;
1706         void **slot;
1707         unsigned ret = 0;
1708
1709         if (unlikely(!nr_pages))
1710                 return 0;
1711
1712         rcu_read_lock();
1713         radix_tree_for_each_slot(slot, &mapping->page_tree, &iter, *start) {
1714                 struct page *head, *page;
1715
1716                 if (iter.index > end)
1717                         break;
1718 repeat:
1719                 page = radix_tree_deref_slot(slot);
1720                 if (unlikely(!page))
1721                         continue;
1722
1723                 if (radix_tree_exception(page)) {
1724                         if (radix_tree_deref_retry(page)) {
1725                                 slot = radix_tree_iter_retry(&iter);
1726                                 continue;
1727                         }
1728                         /*
1729                          * A shadow entry of a recently evicted page,
1730                          * or a swap entry from shmem/tmpfs.  Skip
1731                          * over it.
1732                          */
1733                         continue;
1734                 }
1735
1736                 head = compound_head(page);
1737                 if (!page_cache_get_speculative(head))
1738                         goto repeat;
1739
1740                 /* The page was split under us? */
1741                 if (compound_head(page) != head) {
1742                         put_page(head);
1743                         goto repeat;
1744                 }
1745
1746                 /* Has the page moved? */
1747                 if (unlikely(page != *slot)) {
1748                         put_page(head);
1749                         goto repeat;
1750                 }
1751
1752                 pages[ret] = page;
1753                 if (++ret == nr_pages) {
1754                         *start = pages[ret - 1]->index + 1;
1755                         goto out;
1756                 }
1757         }
1758
1759         /*
1760          * We come here when there is no page beyond @end. We take care to not
1761          * overflow the index @start as it confuses some of the callers. This
1762          * breaks the iteration when there is page at index -1 but that is
1763          * already broken anyway.
1764          */
1765         if (end == (pgoff_t)-1)
1766                 *start = (pgoff_t)-1;
1767         else
1768                 *start = end + 1;
1769 out:
1770         rcu_read_unlock();
1771
1772         return ret;
1773 }
1774
1775 /**
1776  * find_get_pages_contig - gang contiguous pagecache lookup
1777  * @mapping:    The address_space to search
1778  * @index:      The starting page index
1779  * @nr_pages:   The maximum number of pages
1780  * @pages:      Where the resulting pages are placed
1781  *
1782  * find_get_pages_contig() works exactly like find_get_pages(), except
1783  * that the returned number of pages are guaranteed to be contiguous.
1784  *
1785  * find_get_pages_contig() returns the number of pages which were found.
1786  */
1787 unsigned find_get_pages_contig(struct address_space *mapping, pgoff_t index,
1788                                unsigned int nr_pages, struct page **pages)
1789 {
1790         struct radix_tree_iter iter;
1791         void **slot;
1792         unsigned int ret = 0;
1793
1794         if (unlikely(!nr_pages))
1795                 return 0;
1796
1797         rcu_read_lock();
1798         radix_tree_for_each_contig(slot, &mapping->page_tree, &iter, index) {
1799                 struct page *head, *page;
1800 repeat:
1801                 page = radix_tree_deref_slot(slot);
1802                 /* The hole, there no reason to continue */
1803                 if (unlikely(!page))
1804                         break;
1805
1806                 if (radix_tree_exception(page)) {
1807                         if (radix_tree_deref_retry(page)) {
1808                                 slot = radix_tree_iter_retry(&iter);
1809                                 continue;
1810                         }
1811                         /*
1812                          * A shadow entry of a recently evicted page,
1813                          * or a swap entry from shmem/tmpfs.  Stop
1814                          * looking for contiguous pages.
1815                          */
1816                         break;
1817                 }
1818
1819                 head = compound_head(page);
1820                 if (!page_cache_get_speculative(head))
1821                         goto repeat;
1822
1823                 /* The page was split under us? */
1824                 if (compound_head(page) != head) {
1825                         put_page(head);
1826                         goto repeat;
1827                 }
1828
1829                 /* Has the page moved? */
1830                 if (unlikely(page != *slot)) {
1831                         put_page(head);
1832                         goto repeat;
1833                 }
1834
1835                 /*
1836                  * must check mapping and index after taking the ref.
1837                  * otherwise we can get both false positives and false
1838                  * negatives, which is just confusing to the caller.
1839                  */
1840                 if (page->mapping == NULL || page_to_pgoff(page) != iter.index) {
1841                         put_page(page);
1842                         break;
1843                 }
1844
1845                 pages[ret] = page;
1846                 if (++ret == nr_pages)
1847                         break;
1848         }
1849         rcu_read_unlock();
1850         return ret;
1851 }
1852 EXPORT_SYMBOL(find_get_pages_contig);
1853
1854 /**
1855  * find_get_pages_range_tag - find and return pages in given range matching @tag
1856  * @mapping:    the address_space to search
1857  * @index:      the starting page index
1858  * @end:        The final page index (inclusive)
1859  * @tag:        the tag index
1860  * @nr_pages:   the maximum number of pages
1861  * @pages:      where the resulting pages are placed
1862  *
1863  * Like find_get_pages, except we only return pages which are tagged with
1864  * @tag.   We update @index to index the next page for the traversal.
1865  */
1866 unsigned find_get_pages_range_tag(struct address_space *mapping, pgoff_t *index,
1867                         pgoff_t end, int tag, unsigned int nr_pages,
1868                         struct page **pages)
1869 {
1870         struct radix_tree_iter iter;
1871         void **slot;
1872         unsigned ret = 0;
1873
1874         if (unlikely(!nr_pages))
1875                 return 0;
1876
1877         rcu_read_lock();
1878         radix_tree_for_each_tagged(slot, &mapping->page_tree,
1879                                    &iter, *index, tag) {
1880                 struct page *head, *page;
1881
1882                 if (iter.index > end)
1883                         break;
1884 repeat:
1885                 page = radix_tree_deref_slot(slot);
1886                 if (unlikely(!page))
1887                         continue;
1888
1889                 if (radix_tree_exception(page)) {
1890                         if (radix_tree_deref_retry(page)) {
1891                                 slot = radix_tree_iter_retry(&iter);
1892                                 continue;
1893                         }
1894                         /*
1895                          * A shadow entry of a recently evicted page.
1896                          *
1897                          * Those entries should never be tagged, but
1898                          * this tree walk is lockless and the tags are
1899                          * looked up in bulk, one radix tree node at a
1900                          * time, so there is a sizable window for page
1901                          * reclaim to evict a page we saw tagged.
1902                          *
1903                          * Skip over it.
1904                          */
1905                         continue;
1906                 }
1907
1908                 head = compound_head(page);
1909                 if (!page_cache_get_speculative(head))
1910                         goto repeat;
1911
1912                 /* The page was split under us? */
1913                 if (compound_head(page) != head) {
1914                         put_page(head);
1915                         goto repeat;
1916                 }
1917
1918                 /* Has the page moved? */
1919                 if (unlikely(page != *slot)) {
1920                         put_page(head);
1921                         goto repeat;
1922                 }
1923
1924                 pages[ret] = page;
1925                 if (++ret == nr_pages) {
1926                         *index = pages[ret - 1]->index + 1;
1927                         goto out;
1928                 }
1929         }
1930
1931         /*
1932          * We come here when we got at @end. We take care to not overflow the
1933          * index @index as it confuses some of the callers. This breaks the
1934          * iteration when there is page at index -1 but that is already broken
1935          * anyway.
1936          */
1937         if (end == (pgoff_t)-1)
1938                 *index = (pgoff_t)-1;
1939         else
1940                 *index = end + 1;
1941 out:
1942         rcu_read_unlock();
1943
1944         return ret;
1945 }
1946 EXPORT_SYMBOL(find_get_pages_range_tag);
1947
1948 /**
1949  * find_get_entries_tag - find and return entries that match @tag
1950  * @mapping:    the address_space to search
1951  * @start:      the starting page cache index
1952  * @tag:        the tag index
1953  * @nr_entries: the maximum number of entries
1954  * @entries:    where the resulting entries are placed
1955  * @indices:    the cache indices corresponding to the entries in @entries
1956  *
1957  * Like find_get_entries, except we only return entries which are tagged with
1958  * @tag.
1959  */
1960 unsigned find_get_entries_tag(struct address_space *mapping, pgoff_t start,
1961                         int tag, unsigned int nr_entries,
1962                         struct page **entries, pgoff_t *indices)
1963 {
1964         void **slot;
1965         unsigned int ret = 0;
1966         struct radix_tree_iter iter;
1967
1968         if (!nr_entries)
1969                 return 0;
1970
1971         rcu_read_lock();
1972         radix_tree_for_each_tagged(slot, &mapping->page_tree,
1973                                    &iter, start, tag) {
1974                 struct page *head, *page;
1975 repeat:
1976                 page = radix_tree_deref_slot(slot);
1977                 if (unlikely(!page))
1978                         continue;
1979                 if (radix_tree_exception(page)) {
1980                         if (radix_tree_deref_retry(page)) {
1981                                 slot = radix_tree_iter_retry(&iter);
1982                                 continue;
1983                         }
1984
1985                         /*
1986                          * A shadow entry of a recently evicted page, a swap
1987                          * entry from shmem/tmpfs or a DAX entry.  Return it
1988                          * without attempting to raise page count.
1989                          */
1990                         goto export;
1991                 }
1992
1993                 head = compound_head(page);
1994                 if (!page_cache_get_speculative(head))
1995                         goto repeat;
1996
1997                 /* The page was split under us? */
1998                 if (compound_head(page) != head) {
1999                         put_page(head);
2000                         goto repeat;
2001                 }
2002
2003                 /* Has the page moved? */
2004                 if (unlikely(page != *slot)) {
2005                         put_page(head);
2006                         goto repeat;
2007                 }
2008 export:
2009                 indices[ret] = iter.index;
2010                 entries[ret] = page;
2011                 if (++ret == nr_entries)
2012                         break;
2013         }
2014         rcu_read_unlock();
2015         return ret;
2016 }
2017 EXPORT_SYMBOL(find_get_entries_tag);
2018
2019 /*
2020  * CD/DVDs are error prone. When a medium error occurs, the driver may fail
2021  * a _large_ part of the i/o request. Imagine the worst scenario:
2022  *
2023  *      ---R__________________________________________B__________
2024  *         ^ reading here                             ^ bad block(assume 4k)
2025  *
2026  * read(R) => miss => readahead(R...B) => media error => frustrating retries
2027  * => failing the whole request => read(R) => read(R+1) =>
2028  * readahead(R+1...B+1) => bang => read(R+2) => read(R+3) =>
2029  * readahead(R+3...B+2) => bang => read(R+3) => read(R+4) =>
2030  * readahead(R+4...B+3) => bang => read(R+4) => read(R+5) => ......
2031  *
2032  * It is going insane. Fix it by quickly scaling down the readahead size.
2033  */
2034 static void shrink_readahead_size_eio(struct file *filp,
2035                                         struct file_ra_state *ra)
2036 {
2037         ra->ra_pages /= 4;
2038 }
2039
2040 /**
2041  * generic_file_buffered_read - generic file read routine
2042  * @iocb:       the iocb to read
2043  * @iter:       data destination
2044  * @written:    already copied
2045  *
2046  * This is a generic file read routine, and uses the
2047  * mapping->a_ops->readpage() function for the actual low-level stuff.
2048  *
2049  * This is really ugly. But the goto's actually try to clarify some
2050  * of the logic when it comes to error handling etc.
2051  */
2052 static ssize_t generic_file_buffered_read(struct kiocb *iocb,
2053                 struct iov_iter *iter, ssize_t written)
2054 {
2055         struct file *filp = iocb->ki_filp;
2056         struct address_space *mapping = filp->f_mapping;
2057         struct inode *inode = mapping->host;
2058         struct file_ra_state *ra = &filp->f_ra;
2059         loff_t *ppos = &iocb->ki_pos;
2060         pgoff_t index;
2061         pgoff_t last_index;
2062         pgoff_t prev_index;
2063         unsigned long offset;      /* offset into pagecache page */
2064         unsigned int prev_offset;
2065         int error = 0;
2066
2067         if (unlikely(*ppos >= inode->i_sb->s_maxbytes))
2068                 return 0;
2069         iov_iter_truncate(iter, inode->i_sb->s_maxbytes);
2070
2071         index = *ppos >> PAGE_SHIFT;
2072         prev_index = ra->prev_pos >> PAGE_SHIFT;
2073         prev_offset = ra->prev_pos & (PAGE_SIZE-1);
2074         last_index = (*ppos + iter->count + PAGE_SIZE-1) >> PAGE_SHIFT;
2075         offset = *ppos & ~PAGE_MASK;
2076
2077         for (;;) {
2078                 struct page *page;
2079                 pgoff_t end_index;
2080                 loff_t isize;
2081                 unsigned long nr, ret;
2082
2083                 cond_resched();
2084 find_page:
2085                 if (fatal_signal_pending(current)) {
2086                         error = -EINTR;
2087                         goto out;
2088                 }
2089
2090                 page = find_get_page(mapping, index);
2091                 if (!page) {
2092                         if (iocb->ki_flags & IOCB_NOWAIT)
2093                                 goto would_block;
2094                         page_cache_sync_readahead(mapping,
2095                                         ra, filp,
2096                                         index, last_index - index);
2097                         page = find_get_page(mapping, index);
2098                         if (unlikely(page == NULL))
2099                                 goto no_cached_page;
2100                 }
2101                 if (PageReadahead(page)) {
2102                         page_cache_async_readahead(mapping,
2103                                         ra, filp, page,
2104                                         index, last_index - index);
2105                 }
2106                 if (!PageUptodate(page)) {
2107                         if (iocb->ki_flags & IOCB_NOWAIT) {
2108                                 put_page(page);
2109                                 goto would_block;
2110                         }
2111
2112                         /*
2113                          * See comment in do_read_cache_page on why
2114                          * wait_on_page_locked is used to avoid unnecessarily
2115                          * serialisations and why it's safe.
2116                          */
2117                         error = wait_on_page_locked_killable(page);
2118                         if (unlikely(error))
2119                                 goto readpage_error;
2120                         if (PageUptodate(page))
2121                                 goto page_ok;
2122
2123                         if (inode->i_blkbits == PAGE_SHIFT ||
2124                                         !mapping->a_ops->is_partially_uptodate)
2125                                 goto page_not_up_to_date;
2126                         /* pipes can't handle partially uptodate pages */
2127                         if (unlikely(iter->type & ITER_PIPE))
2128                                 goto page_not_up_to_date;
2129                         if (!trylock_page(page))
2130                                 goto page_not_up_to_date;
2131                         /* Did it get truncated before we got the lock? */
2132                         if (!page->mapping)
2133                                 goto page_not_up_to_date_locked;
2134                         if (!mapping->a_ops->is_partially_uptodate(page,
2135                                                         offset, iter->count))
2136                                 goto page_not_up_to_date_locked;
2137                         unlock_page(page);
2138                 }
2139 page_ok:
2140                 /*
2141                  * i_size must be checked after we know the page is Uptodate.
2142                  *
2143                  * Checking i_size after the check allows us to calculate
2144                  * the correct value for "nr", which means the zero-filled
2145                  * part of the page is not copied back to userspace (unless
2146                  * another truncate extends the file - this is desired though).
2147                  */
2148
2149                 isize = i_size_read(inode);
2150                 end_index = (isize - 1) >> PAGE_SHIFT;
2151                 if (unlikely(!isize || index > end_index)) {
2152                         put_page(page);
2153                         goto out;
2154                 }
2155
2156                 /* nr is the maximum number of bytes to copy from this page */
2157                 nr = PAGE_SIZE;
2158                 if (index == end_index) {
2159                         nr = ((isize - 1) & ~PAGE_MASK) + 1;
2160                         if (nr <= offset) {
2161                                 put_page(page);
2162                                 goto out;
2163                         }
2164                 }
2165                 nr = nr - offset;
2166
2167                 /* If users can be writing to this page using arbitrary
2168                  * virtual addresses, take care about potential aliasing
2169                  * before reading the page on the kernel side.
2170                  */
2171                 if (mapping_writably_mapped(mapping))
2172                         flush_dcache_page(page);
2173
2174                 /*
2175                  * When a sequential read accesses a page several times,
2176                  * only mark it as accessed the first time.
2177                  */
2178                 if (prev_index != index || offset != prev_offset)
2179                         mark_page_accessed(page);
2180                 prev_index = index;
2181
2182                 /*
2183                  * Ok, we have the page, and it's up-to-date, so
2184                  * now we can copy it to user space...
2185                  */
2186
2187                 ret = copy_page_to_iter(page, offset, nr, iter);
2188                 offset += ret;
2189                 index += offset >> PAGE_SHIFT;
2190                 offset &= ~PAGE_MASK;
2191                 prev_offset = offset;
2192
2193                 put_page(page);
2194                 written += ret;
2195                 if (!iov_iter_count(iter))
2196                         goto out;
2197                 if (ret < nr) {
2198                         error = -EFAULT;
2199                         goto out;
2200                 }
2201                 continue;
2202
2203 page_not_up_to_date:
2204                 /* Get exclusive access to the page ... */
2205                 error = lock_page_killable(page);
2206                 if (unlikely(error))
2207                         goto readpage_error;
2208
2209 page_not_up_to_date_locked:
2210                 /* Did it get truncated before we got the lock? */
2211                 if (!page->mapping) {
2212                         unlock_page(page);
2213                         put_page(page);
2214                         continue;
2215                 }
2216
2217                 /* Did somebody else fill it already? */
2218                 if (PageUptodate(page)) {
2219                         unlock_page(page);
2220                         goto page_ok;
2221                 }
2222
2223 readpage:
2224                 /*
2225                  * A previous I/O error may have been due to temporary
2226                  * failures, eg. multipath errors.
2227                  * PG_error will be set again if readpage fails.
2228                  */
2229                 ClearPageError(page);
2230                 /* Start the actual read. The read will unlock the page. */
2231                 error = mapping->a_ops->readpage(filp, page);
2232
2233                 if (unlikely(error)) {
2234                         if (error == AOP_TRUNCATED_PAGE) {
2235                                 put_page(page);
2236                                 error = 0;
2237                                 goto find_page;
2238                         }
2239                         goto readpage_error;
2240                 }
2241
2242                 if (!PageUptodate(page)) {
2243                         error = lock_page_killable(page);
2244                         if (unlikely(error))
2245                                 goto readpage_error;
2246                         if (!PageUptodate(page)) {
2247                                 if (page->mapping == NULL) {
2248                                         /*
2249                                          * invalidate_mapping_pages got it
2250                                          */
2251                                         unlock_page(page);
2252                                         put_page(page);
2253                                         goto find_page;
2254                                 }
2255                                 unlock_page(page);
2256                                 shrink_readahead_size_eio(filp, ra);
2257                                 error = -EIO;
2258                                 goto readpage_error;
2259                         }
2260                         unlock_page(page);
2261                 }
2262
2263                 goto page_ok;
2264
2265 readpage_error:
2266                 /* UHHUH! A synchronous read error occurred. Report it */
2267                 put_page(page);
2268                 goto out;
2269
2270 no_cached_page:
2271                 /*
2272                  * Ok, it wasn't cached, so we need to create a new
2273                  * page..
2274                  */
2275                 page = page_cache_alloc(mapping);
2276                 if (!page) {
2277                         error = -ENOMEM;
2278                         goto out;
2279                 }
2280                 error = add_to_page_cache_lru(page, mapping, index,
2281                                 mapping_gfp_constraint(mapping, GFP_KERNEL));
2282                 if (error) {
2283                         put_page(page);
2284                         if (error == -EEXIST) {
2285                                 error = 0;
2286                                 goto find_page;
2287                         }
2288                         goto out;
2289                 }
2290                 goto readpage;
2291         }
2292
2293 would_block:
2294         error = -EAGAIN;
2295 out:
2296         ra->prev_pos = prev_index;
2297         ra->prev_pos <<= PAGE_SHIFT;
2298         ra->prev_pos |= prev_offset;
2299
2300         *ppos = ((loff_t)index << PAGE_SHIFT) + offset;
2301         file_accessed(filp);
2302         return written ? written : error;
2303 }
2304
2305 /**
2306  * generic_file_read_iter - generic filesystem read routine
2307  * @iocb:       kernel I/O control block
2308  * @iter:       destination for the data read
2309  *
2310  * This is the "read_iter()" routine for all filesystems
2311  * that can use the page cache directly.
2312  */
2313 ssize_t
2314 generic_file_read_iter(struct kiocb *iocb, struct iov_iter *iter)
2315 {
2316         size_t count = iov_iter_count(iter);
2317         ssize_t retval = 0;
2318
2319         if (!count)
2320                 goto out; /* skip atime */
2321
2322         if (iocb->ki_flags & IOCB_DIRECT) {
2323                 struct file *file = iocb->ki_filp;
2324                 struct address_space *mapping = file->f_mapping;
2325                 struct inode *inode = mapping->host;
2326                 loff_t size;
2327
2328                 size = i_size_read(inode);
2329                 if (iocb->ki_flags & IOCB_NOWAIT) {
2330                         if (filemap_range_has_page(mapping, iocb->ki_pos,
2331                                                    iocb->ki_pos + count - 1))
2332                                 return -EAGAIN;
2333                 } else {
2334                         retval = filemap_write_and_wait_range(mapping,
2335                                                 iocb->ki_pos,
2336                                                 iocb->ki_pos + count - 1);
2337                         if (retval < 0)
2338                                 goto out;
2339                 }
2340
2341                 file_accessed(file);
2342
2343                 retval = mapping->a_ops->direct_IO(iocb, iter);
2344                 if (retval >= 0) {
2345                         iocb->ki_pos += retval;
2346                         count -= retval;
2347                 }
2348                 iov_iter_revert(iter, count - iov_iter_count(iter));
2349
2350                 /*
2351                  * Btrfs can have a short DIO read if we encounter
2352                  * compressed extents, so if there was an error, or if
2353                  * we've already read everything we wanted to, or if
2354                  * there was a short read because we hit EOF, go ahead
2355                  * and return.  Otherwise fallthrough to buffered io for
2356                  * the rest of the read.  Buffered reads will not work for
2357                  * DAX files, so don't bother trying.
2358                  */
2359                 if (retval < 0 || !count || iocb->ki_pos >= size ||
2360                     IS_DAX(inode))
2361                         goto out;
2362         }
2363
2364         retval = generic_file_buffered_read(iocb, iter, retval);
2365 out:
2366         return retval;
2367 }
2368 EXPORT_SYMBOL(generic_file_read_iter);
2369
2370 #ifdef CONFIG_MMU
2371 /**
2372  * page_cache_read - adds requested page to the page cache if not already there
2373  * @file:       file to read
2374  * @offset:     page index
2375  * @gfp_mask:   memory allocation flags
2376  *
2377  * This adds the requested page to the page cache if it isn't already there,
2378  * and schedules an I/O to read in its contents from disk.
2379  */
2380 static int page_cache_read(struct file *file, pgoff_t offset, gfp_t gfp_mask)
2381 {
2382         struct address_space *mapping = file->f_mapping;
2383         struct page *page;
2384         int ret;
2385
2386         do {
2387                 page = __page_cache_alloc(gfp_mask);
2388                 if (!page)
2389                         return -ENOMEM;
2390
2391                 ret = add_to_page_cache_lru(page, mapping, offset, gfp_mask & GFP_KERNEL);
2392                 if (ret == 0)
2393                         ret = mapping->a_ops->readpage(file, page);
2394                 else if (ret == -EEXIST)
2395                         ret = 0; /* losing race to add is OK */
2396
2397                 put_page(page);
2398
2399         } while (ret == AOP_TRUNCATED_PAGE);
2400
2401         return ret;
2402 }
2403
2404 #define MMAP_LOTSAMISS  (100)
2405
2406 /*
2407  * Synchronous readahead happens when we don't even find
2408  * a page in the page cache at all.
2409  */
2410 static void do_sync_mmap_readahead(struct vm_area_struct *vma,
2411                                    struct file_ra_state *ra,
2412                                    struct file *file,
2413                                    pgoff_t offset)
2414 {
2415         struct address_space *mapping = file->f_mapping;
2416
2417         /* If we don't want any read-ahead, don't bother */
2418         if (vma->vm_flags & VM_RAND_READ)
2419                 return;
2420         if (!ra->ra_pages)
2421                 return;
2422
2423         if (vma->vm_flags & VM_SEQ_READ) {
2424                 page_cache_sync_readahead(mapping, ra, file, offset,
2425                                           ra->ra_pages);
2426                 return;
2427         }
2428
2429         /* Avoid banging the cache line if not needed */
2430         if (ra->mmap_miss < MMAP_LOTSAMISS * 10)
2431                 ra->mmap_miss++;
2432
2433         /*
2434          * Do we miss much more than hit in this file? If so,
2435          * stop bothering with read-ahead. It will only hurt.
2436          */
2437         if (ra->mmap_miss > MMAP_LOTSAMISS)
2438                 return;
2439
2440         /*
2441          * mmap read-around
2442          */
2443         ra->start = max_t(long, 0, offset - ra->ra_pages / 2);
2444         ra->size = ra->ra_pages;
2445         ra->async_size = ra->ra_pages / 4;
2446         ra_submit(ra, mapping, file);
2447 }
2448
2449 /*
2450  * Asynchronous readahead happens when we find the page and PG_readahead,
2451  * so we want to possibly extend the readahead further..
2452  */
2453 static void do_async_mmap_readahead(struct vm_area_struct *vma,
2454                                     struct file_ra_state *ra,
2455                                     struct file *file,
2456                                     struct page *page,
2457                                     pgoff_t offset)
2458 {
2459         struct address_space *mapping = file->f_mapping;
2460
2461         /* If we don't want any read-ahead, don't bother */
2462         if (vma->vm_flags & VM_RAND_READ)
2463                 return;
2464         if (ra->mmap_miss > 0)
2465                 ra->mmap_miss--;
2466         if (PageReadahead(page))
2467                 page_cache_async_readahead(mapping, ra, file,
2468                                            page, offset, ra->ra_pages);
2469 }
2470
2471 /**
2472  * filemap_fault - read in file data for page fault handling
2473  * @vmf:        struct vm_fault containing details of the fault
2474  *
2475  * filemap_fault() is invoked via the vma operations vector for a
2476  * mapped memory region to read in file data during a page fault.
2477  *
2478  * The goto's are kind of ugly, but this streamlines the normal case of having
2479  * it in the page cache, and handles the special cases reasonably without
2480  * having a lot of duplicated code.
2481  *
2482  * vma->vm_mm->mmap_sem must be held on entry.
2483  *
2484  * If our return value has VM_FAULT_RETRY set, it's because
2485  * lock_page_or_retry() returned 0.
2486  * The mmap_sem has usually been released in this case.
2487  * See __lock_page_or_retry() for the exception.
2488  *
2489  * If our return value does not have VM_FAULT_RETRY set, the mmap_sem
2490  * has not been released.
2491  *
2492  * We never return with VM_FAULT_RETRY and a bit from VM_FAULT_ERROR set.
2493  */
2494 int filemap_fault(struct vm_fault *vmf)
2495 {
2496         int error;
2497         struct file *file = vmf->vma->vm_file;
2498         struct address_space *mapping = file->f_mapping;
2499         struct file_ra_state *ra = &file->f_ra;
2500         struct inode *inode = mapping->host;
2501         pgoff_t offset = vmf->pgoff;
2502         pgoff_t max_off;
2503         struct page *page;
2504         int ret = 0;
2505
2506         max_off = DIV_ROUND_UP(i_size_read(inode), PAGE_SIZE);
2507         if (unlikely(offset >= max_off))
2508                 return VM_FAULT_SIGBUS;
2509
2510         /*
2511          * Do we have something in the page cache already?
2512          */
2513         page = find_get_page(mapping, offset);
2514         if (likely(page) && !(vmf->flags & FAULT_FLAG_TRIED)) {
2515                 /*
2516                  * We found the page, so try async readahead before
2517                  * waiting for the lock.
2518                  */
2519                 do_async_mmap_readahead(vmf->vma, ra, file, page, offset);
2520         } else if (!page) {
2521                 /* No page in the page cache at all */
2522                 do_sync_mmap_readahead(vmf->vma, ra, file, offset);
2523                 count_vm_event(PGMAJFAULT);
2524                 count_memcg_event_mm(vmf->vma->vm_mm, PGMAJFAULT);
2525                 ret = VM_FAULT_MAJOR;
2526 retry_find:
2527                 page = find_get_page(mapping, offset);
2528                 if (!page)
2529                         goto no_cached_page;
2530         }
2531
2532         if (!lock_page_or_retry(page, vmf->vma->vm_mm, vmf->flags)) {
2533                 put_page(page);
2534                 return ret | VM_FAULT_RETRY;
2535         }
2536
2537         /* Did it get truncated? */
2538         if (unlikely(page->mapping != mapping)) {
2539                 unlock_page(page);
2540                 put_page(page);
2541                 goto retry_find;
2542         }
2543         VM_BUG_ON_PAGE(page->index != offset, page);
2544
2545         /*
2546          * We have a locked page in the page cache, now we need to check
2547          * that it's up-to-date. If not, it is going to be due to an error.
2548          */
2549         if (unlikely(!PageUptodate(page)))
2550                 goto page_not_uptodate;
2551
2552         /*
2553          * Found the page and have a reference on it.
2554          * We must recheck i_size under page lock.
2555          */
2556         max_off = DIV_ROUND_UP(i_size_read(inode), PAGE_SIZE);
2557         if (unlikely(offset >= max_off)) {
2558                 unlock_page(page);
2559                 put_page(page);
2560                 return VM_FAULT_SIGBUS;
2561         }
2562
2563         vmf->page = page;
2564         return ret | VM_FAULT_LOCKED;
2565
2566 no_cached_page:
2567         /*
2568          * We're only likely to ever get here if MADV_RANDOM is in
2569          * effect.
2570          */
2571         error = page_cache_read(file, offset, vmf->gfp_mask);
2572
2573         /*
2574          * The page we want has now been added to the page cache.
2575          * In the unlikely event that someone removed it in the
2576          * meantime, we'll just come back here and read it again.
2577          */
2578         if (error >= 0)
2579                 goto retry_find;
2580
2581         /*
2582          * An error return from page_cache_read can result if the
2583          * system is low on memory, or a problem occurs while trying
2584          * to schedule I/O.
2585          */
2586         if (error == -ENOMEM)
2587                 return VM_FAULT_OOM;
2588         return VM_FAULT_SIGBUS;
2589
2590 page_not_uptodate:
2591         /*
2592          * Umm, take care of errors if the page isn't up-to-date.
2593          * Try to re-read it _once_. We do this synchronously,
2594          * because there really aren't any performance issues here
2595          * and we need to check for errors.
2596          */
2597         ClearPageError(page);
2598         error = mapping->a_ops->readpage(file, page);
2599         if (!error) {
2600                 wait_on_page_locked(page);
2601                 if (!PageUptodate(page))
2602                         error = -EIO;
2603         }
2604         put_page(page);
2605
2606         if (!error || error == AOP_TRUNCATED_PAGE)
2607                 goto retry_find;
2608
2609         /* Things didn't work out. Return zero to tell the mm layer so. */
2610         shrink_readahead_size_eio(file, ra);
2611         return VM_FAULT_SIGBUS;
2612 }
2613 EXPORT_SYMBOL(filemap_fault);
2614
2615 void filemap_map_pages(struct vm_fault *vmf,
2616                 pgoff_t start_pgoff, pgoff_t end_pgoff)
2617 {
2618         struct radix_tree_iter iter;
2619         void **slot;
2620         struct file *file = vmf->vma->vm_file;
2621         struct address_space *mapping = file->f_mapping;
2622         pgoff_t last_pgoff = start_pgoff;
2623         unsigned long max_idx;
2624         struct page *head, *page;
2625
2626         rcu_read_lock();
2627         radix_tree_for_each_slot(slot, &mapping->page_tree, &iter,
2628                         start_pgoff) {
2629                 if (iter.index > end_pgoff)
2630                         break;
2631 repeat:
2632                 page = radix_tree_deref_slot(slot);
2633                 if (unlikely(!page))
2634                         goto next;
2635                 if (radix_tree_exception(page)) {
2636                         if (radix_tree_deref_retry(page)) {
2637                                 slot = radix_tree_iter_retry(&iter);
2638                                 continue;
2639                         }
2640                         goto next;
2641                 }
2642
2643                 head = compound_head(page);
2644                 if (!page_cache_get_speculative(head))
2645                         goto repeat;
2646
2647                 /* The page was split under us? */
2648                 if (compound_head(page) != head) {
2649                         put_page(head);
2650                         goto repeat;
2651                 }
2652
2653                 /* Has the page moved? */
2654                 if (unlikely(page != *slot)) {
2655                         put_page(head);
2656                         goto repeat;
2657                 }
2658
2659                 if (!PageUptodate(page) ||
2660                                 PageReadahead(page) ||
2661                                 PageHWPoison(page))
2662                         goto skip;
2663                 if (!trylock_page(page))
2664                         goto skip;
2665
2666                 if (page->mapping != mapping || !PageUptodate(page))
2667                         goto unlock;
2668
2669                 max_idx = DIV_ROUND_UP(i_size_read(mapping->host), PAGE_SIZE);
2670                 if (page->index >= max_idx)
2671                         goto unlock;
2672
2673                 if (file->f_ra.mmap_miss > 0)
2674                         file->f_ra.mmap_miss--;
2675
2676                 vmf->address += (iter.index - last_pgoff) << PAGE_SHIFT;
2677                 if (vmf->pte)
2678                         vmf->pte += iter.index - last_pgoff;
2679                 last_pgoff = iter.index;
2680                 if (alloc_set_pte(vmf, NULL, page))
2681                         goto unlock;
2682                 unlock_page(page);
2683                 goto next;
2684 unlock:
2685                 unlock_page(page);
2686 skip:
2687                 put_page(page);
2688 next:
2689                 /* Huge page is mapped? No need to proceed. */
2690                 if (pmd_trans_huge(*vmf->pmd))
2691                         break;
2692                 if (iter.index == end_pgoff)
2693                         break;
2694         }
2695         rcu_read_unlock();
2696 }
2697 EXPORT_SYMBOL(filemap_map_pages);
2698
2699 int filemap_page_mkwrite(struct vm_fault *vmf)
2700 {
2701         struct page *page = vmf->page;
2702         struct inode *inode = file_inode(vmf->vma->vm_file);
2703         int ret = VM_FAULT_LOCKED;
2704
2705         sb_start_pagefault(inode->i_sb);
2706         file_update_time(vmf->vma->vm_file);
2707         lock_page(page);
2708         if (page->mapping != inode->i_mapping) {
2709                 unlock_page(page);
2710                 ret = VM_FAULT_NOPAGE;
2711                 goto out;
2712         }
2713         /*
2714          * We mark the page dirty already here so that when freeze is in
2715          * progress, we are guaranteed that writeback during freezing will
2716          * see the dirty page and writeprotect it again.
2717          */
2718         set_page_dirty(page);
2719         wait_for_stable_page(page);
2720 out:
2721         sb_end_pagefault(inode->i_sb);
2722         return ret;
2723 }
2724 EXPORT_SYMBOL(filemap_page_mkwrite);
2725
2726 const struct vm_operations_struct generic_file_vm_ops = {
2727         .fault          = filemap_fault,
2728         .map_pages      = filemap_map_pages,
2729         .page_mkwrite   = filemap_page_mkwrite,
2730 };
2731
2732 /* This is used for a general mmap of a disk file */
2733
2734 int generic_file_mmap(struct file * file, struct vm_area_struct * vma)
2735 {
2736         struct address_space *mapping = file->f_mapping;
2737
2738         if (!mapping->a_ops->readpage)
2739                 return -ENOEXEC;
2740         file_accessed(file);
2741         vma->vm_ops = &generic_file_vm_ops;
2742         return 0;
2743 }
2744
2745 /*
2746  * This is for filesystems which do not implement ->writepage.
2747  */
2748 int generic_file_readonly_mmap(struct file *file, struct vm_area_struct *vma)
2749 {
2750         if ((vma->vm_flags & VM_SHARED) && (vma->vm_flags & VM_MAYWRITE))
2751                 return -EINVAL;
2752         return generic_file_mmap(file, vma);
2753 }
2754 #else
2755 int generic_file_mmap(struct file * file, struct vm_area_struct * vma)
2756 {
2757         return -ENOSYS;
2758 }
2759 int generic_file_readonly_mmap(struct file * file, struct vm_area_struct * vma)
2760 {
2761         return -ENOSYS;
2762 }
2763 #endif /* CONFIG_MMU */
2764
2765 EXPORT_SYMBOL(generic_file_mmap);
2766 EXPORT_SYMBOL(generic_file_readonly_mmap);
2767
2768 static struct page *wait_on_page_read(struct page *page)
2769 {
2770         if (!IS_ERR(page)) {
2771                 wait_on_page_locked(page);
2772                 if (!PageUptodate(page)) {
2773                         put_page(page);
2774                         page = ERR_PTR(-EIO);
2775                 }
2776         }
2777         return page;
2778 }
2779
2780 static struct page *do_read_cache_page(struct address_space *mapping,
2781                                 pgoff_t index,
2782                                 int (*filler)(void *, struct page *),
2783                                 void *data,
2784                                 gfp_t gfp)
2785 {
2786         struct page *page;
2787         int err;
2788 repeat:
2789         page = find_get_page(mapping, index);
2790         if (!page) {
2791                 page = __page_cache_alloc(gfp);
2792                 if (!page)
2793                         return ERR_PTR(-ENOMEM);
2794                 err = add_to_page_cache_lru(page, mapping, index, gfp);
2795                 if (unlikely(err)) {
2796                         put_page(page);
2797                         if (err == -EEXIST)
2798                                 goto repeat;
2799                         /* Presumably ENOMEM for radix tree node */
2800                         return ERR_PTR(err);
2801                 }
2802
2803 filler:
2804                 err = filler(data, page);
2805                 if (err < 0) {
2806                         put_page(page);
2807                         return ERR_PTR(err);
2808                 }
2809
2810                 page = wait_on_page_read(page);
2811                 if (IS_ERR(page))
2812                         return page;
2813                 goto out;
2814         }
2815         if (PageUptodate(page))
2816                 goto out;
2817
2818         /*
2819          * Page is not up to date and may be locked due one of the following
2820          * case a: Page is being filled and the page lock is held
2821          * case b: Read/write error clearing the page uptodate status
2822          * case c: Truncation in progress (page locked)
2823          * case d: Reclaim in progress
2824          *
2825          * Case a, the page will be up to date when the page is unlocked.
2826          *    There is no need to serialise on the page lock here as the page
2827          *    is pinned so the lock gives no additional protection. Even if the
2828          *    the page is truncated, the data is still valid if PageUptodate as
2829          *    it's a race vs truncate race.
2830          * Case b, the page will not be up to date
2831          * Case c, the page may be truncated but in itself, the data may still
2832          *    be valid after IO completes as it's a read vs truncate race. The
2833          *    operation must restart if the page is not uptodate on unlock but
2834          *    otherwise serialising on page lock to stabilise the mapping gives
2835          *    no additional guarantees to the caller as the page lock is
2836          *    released before return.
2837          * Case d, similar to truncation. If reclaim holds the page lock, it
2838          *    will be a race with remove_mapping that determines if the mapping
2839          *    is valid on unlock but otherwise the data is valid and there is
2840          *    no need to serialise with page lock.
2841          *
2842          * As the page lock gives no additional guarantee, we optimistically
2843          * wait on the page to be unlocked and check if it's up to date and
2844          * use the page if it is. Otherwise, the page lock is required to
2845          * distinguish between the different cases. The motivation is that we
2846          * avoid spurious serialisations and wakeups when multiple processes
2847          * wait on the same page for IO to complete.
2848          */
2849         wait_on_page_locked(page);
2850         if (PageUptodate(page))
2851                 goto out;
2852
2853         /* Distinguish between all the cases under the safety of the lock */
2854         lock_page(page);
2855
2856         /* Case c or d, restart the operation */
2857         if (!page->mapping) {
2858                 unlock_page(page);
2859                 put_page(page);
2860                 goto repeat;
2861         }
2862
2863         /* Someone else locked and filled the page in a very small window */
2864         if (PageUptodate(page)) {
2865                 unlock_page(page);
2866                 goto out;
2867         }
2868         goto filler;
2869
2870 out:
2871         mark_page_accessed(page);
2872         return page;
2873 }
2874
2875 /**
2876  * read_cache_page - read into page cache, fill it if needed
2877  * @mapping:    the page's address_space
2878  * @index:      the page index
2879  * @filler:     function to perform the read
2880  * @data:       first arg to filler(data, page) function, often left as NULL
2881  *
2882  * Read into the page cache. If a page already exists, and PageUptodate() is
2883  * not set, try to fill the page and wait for it to become unlocked.
2884  *
2885  * If the page does not get brought uptodate, return -EIO.
2886  */
2887 struct page *read_cache_page(struct address_space *mapping,
2888                                 pgoff_t index,
2889                                 int (*filler)(void *, struct page *),
2890                                 void *data)
2891 {
2892         return do_read_cache_page(mapping, index, filler, data, mapping_gfp_mask(mapping));
2893 }
2894 EXPORT_SYMBOL(read_cache_page);
2895
2896 /**
2897  * read_cache_page_gfp - read into page cache, using specified page allocation flags.
2898  * @mapping:    the page's address_space
2899  * @index:      the page index
2900  * @gfp:        the page allocator flags to use if allocating
2901  *
2902  * This is the same as "read_mapping_page(mapping, index, NULL)", but with
2903  * any new page allocations done using the specified allocation flags.
2904  *
2905  * If the page does not get brought uptodate, return -EIO.
2906  */
2907 struct page *read_cache_page_gfp(struct address_space *mapping,
2908                                 pgoff_t index,
2909                                 gfp_t gfp)
2910 {
2911         filler_t *filler = (filler_t *)mapping->a_ops->readpage;
2912
2913         return do_read_cache_page(mapping, index, filler, NULL, gfp);
2914 }
2915 EXPORT_SYMBOL(read_cache_page_gfp);
2916
2917 /*
2918  * Performs necessary checks before doing a write
2919  *
2920  * Can adjust writing position or amount of bytes to write.
2921  * Returns appropriate error code that caller should return or
2922  * zero in case that write should be allowed.
2923  */
2924 inline ssize_t generic_write_checks(struct kiocb *iocb, struct iov_iter *from)
2925 {
2926         struct file *file = iocb->ki_filp;
2927         struct inode *inode = file->f_mapping->host;
2928         unsigned long limit = rlimit(RLIMIT_FSIZE);
2929         loff_t pos;
2930
2931         if (!iov_iter_count(from))
2932                 return 0;
2933
2934         /* FIXME: this is for backwards compatibility with 2.4 */
2935         if (iocb->ki_flags & IOCB_APPEND)
2936                 iocb->ki_pos = i_size_read(inode);
2937
2938         pos = iocb->ki_pos;
2939
2940         if ((iocb->ki_flags & IOCB_NOWAIT) && !(iocb->ki_flags & IOCB_DIRECT))
2941                 return -EINVAL;
2942
2943         if (limit != RLIM_INFINITY) {
2944                 if (iocb->ki_pos >= limit) {
2945                         send_sig(SIGXFSZ, current, 0);
2946                         return -EFBIG;
2947                 }
2948                 iov_iter_truncate(from, limit - (unsigned long)pos);
2949         }
2950
2951         /*
2952          * LFS rule
2953          */
2954         if (unlikely(pos + iov_iter_count(from) > MAX_NON_LFS &&
2955                                 !(file->f_flags & O_LARGEFILE))) {
2956                 if (pos >= MAX_NON_LFS)
2957                         return -EFBIG;
2958                 iov_iter_truncate(from, MAX_NON_LFS - (unsigned long)pos);
2959         }
2960
2961         /*
2962          * Are we about to exceed the fs block limit ?
2963          *
2964          * If we have written data it becomes a short write.  If we have
2965          * exceeded without writing data we send a signal and return EFBIG.
2966          * Linus frestrict idea will clean these up nicely..
2967          */
2968         if (unlikely(pos >= inode->i_sb->s_maxbytes))
2969                 return -EFBIG;
2970
2971         iov_iter_truncate(from, inode->i_sb->s_maxbytes - pos);
2972         return iov_iter_count(from);
2973 }
2974 EXPORT_SYMBOL(generic_write_checks);
2975
2976 int pagecache_write_begin(struct file *file, struct address_space *mapping,
2977                                 loff_t pos, unsigned len, unsigned flags,
2978                                 struct page **pagep, void **fsdata)
2979 {
2980         const struct address_space_operations *aops = mapping->a_ops;
2981
2982         return aops->write_begin(file, mapping, pos, len, flags,
2983                                                         pagep, fsdata);
2984 }
2985 EXPORT_SYMBOL(pagecache_write_begin);
2986
2987 int pagecache_write_end(struct file *file, struct address_space *mapping,
2988                                 loff_t pos, unsigned len, unsigned copied,
2989                                 struct page *page, void *fsdata)
2990 {
2991         const struct address_space_operations *aops = mapping->a_ops;
2992
2993         return aops->write_end(file, mapping, pos, len, copied, page, fsdata);
2994 }
2995 EXPORT_SYMBOL(pagecache_write_end);
2996
2997 ssize_t
2998 generic_file_direct_write(struct kiocb *iocb, struct iov_iter *from)
2999 {
3000         struct file     *file = iocb->ki_filp;
3001         struct address_space *mapping = file->f_mapping;
3002         struct inode    *inode = mapping->host;
3003         loff_t          pos = iocb->ki_pos;
3004         ssize_t         written;
3005         size_t          write_len;
3006         pgoff_t         end;
3007
3008         write_len = iov_iter_count(from);
3009         end = (pos + write_len - 1) >> PAGE_SHIFT;
3010
3011         if (iocb->ki_flags & IOCB_NOWAIT) {
3012                 /* If there are pages to writeback, return */
3013                 if (filemap_range_has_page(inode->i_mapping, pos,
3014                                            pos + iov_iter_count(from)))
3015                         return -EAGAIN;
3016         } else {
3017                 written = filemap_write_and_wait_range(mapping, pos,
3018                                                         pos + write_len - 1);
3019                 if (written)
3020                         goto out;
3021         }
3022
3023         /*
3024          * After a write we want buffered reads to be sure to go to disk to get
3025          * the new data.  We invalidate clean cached page from the region we're
3026          * about to write.  We do this *before* the write so that we can return
3027          * without clobbering -EIOCBQUEUED from ->direct_IO().
3028          */
3029         written = invalidate_inode_pages2_range(mapping,
3030                                         pos >> PAGE_SHIFT, end);
3031         /*
3032          * If a page can not be invalidated, return 0 to fall back
3033          * to buffered write.
3034          */
3035         if (written) {
3036                 if (written == -EBUSY)
3037                         return 0;
3038                 goto out;
3039         }
3040
3041         written = mapping->a_ops->direct_IO(iocb, from);
3042
3043         /*
3044          * Finally, try again to invalidate clean pages which might have been
3045          * cached by non-direct readahead, or faulted in by get_user_pages()
3046          * if the source of the write was an mmap'ed region of the file
3047          * we're writing.  Either one is a pretty crazy thing to do,
3048          * so we don't support it 100%.  If this invalidation
3049          * fails, tough, the write still worked...
3050          *
3051          * Most of the time we do not need this since dio_complete() will do
3052          * the invalidation for us. However there are some file systems that
3053          * do not end up with dio_complete() being called, so let's not break
3054          * them by removing it completely
3055          */
3056         if (mapping->nrpages)
3057                 invalidate_inode_pages2_range(mapping,
3058                                         pos >> PAGE_SHIFT, end);
3059
3060         if (written > 0) {
3061                 pos += written;
3062                 write_len -= written;
3063                 if (pos > i_size_read(inode) && !S_ISBLK(inode->i_mode)) {
3064                         i_size_write(inode, pos);
3065                         mark_inode_dirty(inode);
3066                 }
3067                 iocb->ki_pos = pos;
3068         }
3069         iov_iter_revert(from, write_len - iov_iter_count(from));
3070 out:
3071         return written;
3072 }
3073 EXPORT_SYMBOL(generic_file_direct_write);
3074
3075 /*
3076  * Find or create a page at the given pagecache position. Return the locked
3077  * page. This function is specifically for buffered writes.
3078  */
3079 struct page *grab_cache_page_write_begin(struct address_space *mapping,
3080                                         pgoff_t index, unsigned flags)
3081 {
3082         struct page *page;
3083         int fgp_flags = FGP_LOCK|FGP_WRITE|FGP_CREAT;
3084
3085         if (flags & AOP_FLAG_NOFS)
3086                 fgp_flags |= FGP_NOFS;
3087
3088         page = pagecache_get_page(mapping, index, fgp_flags,
3089                         mapping_gfp_mask(mapping));
3090         if (page)
3091                 wait_for_stable_page(page);
3092
3093         return page;
3094 }
3095 EXPORT_SYMBOL(grab_cache_page_write_begin);
3096
3097 ssize_t generic_perform_write(struct file *file,
3098                                 struct iov_iter *i, loff_t pos)
3099 {
3100         struct address_space *mapping = file->f_mapping;
3101         const struct address_space_operations *a_ops = mapping->a_ops;
3102         long status = 0;
3103         ssize_t written = 0;
3104         unsigned int flags = 0;
3105
3106         do {
3107                 struct page *page;
3108                 unsigned long offset;   /* Offset into pagecache page */
3109                 unsigned long bytes;    /* Bytes to write to page */
3110                 size_t copied;          /* Bytes copied from user */
3111                 void *fsdata;
3112
3113                 offset = (pos & (PAGE_SIZE - 1));
3114                 bytes = min_t(unsigned long, PAGE_SIZE - offset,
3115                                                 iov_iter_count(i));
3116
3117 again:
3118                 /*
3119                  * Bring in the user page that we will copy from _first_.
3120                  * Otherwise there's a nasty deadlock on copying from the
3121                  * same page as we're writing to, without it being marked
3122                  * up-to-date.
3123                  *
3124                  * Not only is this an optimisation, but it is also required
3125                  * to check that the address is actually valid, when atomic
3126                  * usercopies are used, below.
3127                  */
3128                 if (unlikely(iov_iter_fault_in_readable(i, bytes))) {
3129                         status = -EFAULT;
3130                         break;
3131                 }
3132
3133                 if (fatal_signal_pending(current)) {
3134                         status = -EINTR;
3135                         break;
3136                 }
3137
3138                 status = a_ops->write_begin(file, mapping, pos, bytes, flags,
3139                                                 &page, &fsdata);
3140                 if (unlikely(status < 0))
3141                         break;
3142
3143                 if (mapping_writably_mapped(mapping))
3144                         flush_dcache_page(page);
3145
3146                 copied = iov_iter_copy_from_user_atomic(page, i, offset, bytes);
3147                 flush_dcache_page(page);
3148
3149                 status = a_ops->write_end(file, mapping, pos, bytes, copied,
3150                                                 page, fsdata);
3151                 if (unlikely(status < 0))
3152                         break;
3153                 copied = status;
3154
3155                 cond_resched();
3156
3157                 iov_iter_advance(i, copied);
3158                 if (unlikely(copied == 0)) {
3159                         /*
3160                          * If we were unable to copy any data at all, we must
3161                          * fall back to a single segment length write.
3162                          *
3163                          * If we didn't fallback here, we could livelock
3164                          * because not all segments in the iov can be copied at
3165                          * once without a pagefault.
3166                          */
3167                         bytes = min_t(unsigned long, PAGE_SIZE - offset,
3168                                                 iov_iter_single_seg_count(i));
3169                         goto again;
3170                 }
3171                 pos += copied;
3172                 written += copied;
3173
3174                 balance_dirty_pages_ratelimited(mapping);
3175         } while (iov_iter_count(i));
3176
3177         return written ? written : status;
3178 }
3179 EXPORT_SYMBOL(generic_perform_write);
3180
3181 /**
3182  * __generic_file_write_iter - write data to a file
3183  * @iocb:       IO state structure (file, offset, etc.)
3184  * @from:       iov_iter with data to write
3185  *
3186  * This function does all the work needed for actually writing data to a
3187  * file. It does all basic checks, removes SUID from the file, updates
3188  * modification times and calls proper subroutines depending on whether we
3189  * do direct IO or a standard buffered write.
3190  *
3191  * It expects i_mutex to be grabbed unless we work on a block device or similar
3192  * object which does not need locking at all.
3193  *
3194  * This function does *not* take care of syncing data in case of O_SYNC write.
3195  * A caller has to handle it. This is mainly due to the fact that we want to
3196  * avoid syncing under i_mutex.
3197  */
3198 ssize_t __generic_file_write_iter(struct kiocb *iocb, struct iov_iter *from)
3199 {
3200         struct file *file = iocb->ki_filp;
3201         struct address_space * mapping = file->f_mapping;
3202         struct inode    *inode = mapping->host;
3203         ssize_t         written = 0;
3204         ssize_t         err;
3205         ssize_t         status;
3206
3207         /* We can write back this queue in page reclaim */
3208         current->backing_dev_info = inode_to_bdi(inode);
3209         err = file_remove_privs(file);
3210         if (err)
3211                 goto out;
3212
3213         err = file_update_time(file);
3214         if (err)
3215                 goto out;
3216
3217         if (iocb->ki_flags & IOCB_DIRECT) {
3218                 loff_t pos, endbyte;
3219
3220                 written = generic_file_direct_write(iocb, from);
3221                 /*
3222                  * If the write stopped short of completing, fall back to
3223                  * buffered writes.  Some filesystems do this for writes to
3224                  * holes, for example.  For DAX files, a buffered write will
3225                  * not succeed (even if it did, DAX does not handle dirty
3226                  * page-cache pages correctly).
3227                  */
3228                 if (written < 0 || !iov_iter_count(from) || IS_DAX(inode))
3229                         goto out;
3230
3231                 status = generic_perform_write(file, from, pos = iocb->ki_pos);
3232                 /*
3233                  * If generic_perform_write() returned a synchronous error
3234                  * then we want to return the number of bytes which were
3235                  * direct-written, or the error code if that was zero.  Note
3236                  * that this differs from normal direct-io semantics, which
3237                  * will return -EFOO even if some bytes were written.
3238                  */
3239                 if (unlikely(status < 0)) {
3240                         err = status;
3241                         goto out;
3242                 }
3243                 /*
3244                  * We need to ensure that the page cache pages are written to
3245                  * disk and invalidated to preserve the expected O_DIRECT
3246                  * semantics.
3247                  */
3248                 endbyte = pos + status - 1;
3249                 err = filemap_write_and_wait_range(mapping, pos, endbyte);
3250                 if (err == 0) {
3251                         iocb->ki_pos = endbyte + 1;
3252                         written += status;
3253                         invalidate_mapping_pages(mapping,
3254                                                  pos >> PAGE_SHIFT,
3255                                                  endbyte >> PAGE_SHIFT);
3256                 } else {
3257                         /*
3258                          * We don't know how much we wrote, so just return
3259                          * the number of bytes which were direct-written
3260                          */
3261                 }
3262         } else {
3263                 written = generic_perform_write(file, from, iocb->ki_pos);
3264                 if (likely(written > 0))
3265                         iocb->ki_pos += written;
3266         }
3267 out:
3268         current->backing_dev_info = NULL;
3269         return written ? written : err;
3270 }
3271 EXPORT_SYMBOL(__generic_file_write_iter);
3272
3273 /**
3274  * generic_file_write_iter - write data to a file
3275  * @iocb:       IO state structure
3276  * @from:       iov_iter with data to write
3277  *
3278  * This is a wrapper around __generic_file_write_iter() to be used by most
3279  * filesystems. It takes care of syncing the file in case of O_SYNC file
3280  * and acquires i_mutex as needed.
3281  */
3282 ssize_t generic_file_write_iter(struct kiocb *iocb, struct iov_iter *from)
3283 {
3284         struct file *file = iocb->ki_filp;
3285         struct inode *inode = file->f_mapping->host;
3286         ssize_t ret;
3287
3288         inode_lock(inode);
3289         ret = generic_write_checks(iocb, from);
3290         if (ret > 0)
3291                 ret = __generic_file_write_iter(iocb, from);
3292         inode_unlock(inode);
3293
3294         if (ret > 0)
3295                 ret = generic_write_sync(iocb, ret);
3296         return ret;
3297 }
3298 EXPORT_SYMBOL(generic_file_write_iter);
3299
3300 /**
3301  * try_to_release_page() - release old fs-specific metadata on a page
3302  *
3303  * @page: the page which the kernel is trying to free
3304  * @gfp_mask: memory allocation&