delayacct: track delays from thrashing cache pages
[muen/linux.git] / mm / filemap.c
1 /*
2  *      linux/mm/filemap.c
3  *
4  * Copyright (C) 1994-1999  Linus Torvalds
5  */
6
7 /*
8  * This file handles the generic file mmap semantics used by
9  * most "normal" filesystems (but you don't /have/ to use this:
10  * the NFS filesystem used to do this differently, for example)
11  */
12 #include <linux/export.h>
13 #include <linux/compiler.h>
14 #include <linux/dax.h>
15 #include <linux/fs.h>
16 #include <linux/sched/signal.h>
17 #include <linux/uaccess.h>
18 #include <linux/capability.h>
19 #include <linux/kernel_stat.h>
20 #include <linux/gfp.h>
21 #include <linux/mm.h>
22 #include <linux/swap.h>
23 #include <linux/mman.h>
24 #include <linux/pagemap.h>
25 #include <linux/file.h>
26 #include <linux/uio.h>
27 #include <linux/hash.h>
28 #include <linux/writeback.h>
29 #include <linux/backing-dev.h>
30 #include <linux/pagevec.h>
31 #include <linux/blkdev.h>
32 #include <linux/security.h>
33 #include <linux/cpuset.h>
34 #include <linux/hugetlb.h>
35 #include <linux/memcontrol.h>
36 #include <linux/cleancache.h>
37 #include <linux/shmem_fs.h>
38 #include <linux/rmap.h>
39 #include <linux/delayacct.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  *        ->i_pages 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  *        ->i_pages 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  *    ->i_pages 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  *    ->i_pages 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  *    ->i_pages 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->i_pages, 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,
130                                                     &mapping->i_pages.xa_lock);
131                 if (!radix_tree_exceptional_entry(p))
132                         return -EEXIST;
133
134                 mapping->nrexceptional--;
135                 if (shadowp)
136                         *shadowp = p;
137         }
138         __radix_tree_replace(&mapping->i_pages, node, slot, page,
139                              workingset_lookup_update(mapping));
140         mapping->nrpages++;
141         return 0;
142 }
143
144 static void page_cache_tree_delete(struct address_space *mapping,
145                                    struct page *page, void *shadow)
146 {
147         int i, nr;
148
149         /* hugetlb pages are represented by one entry in the radix tree */
150         nr = PageHuge(page) ? 1 : hpage_nr_pages(page);
151
152         VM_BUG_ON_PAGE(!PageLocked(page), page);
153         VM_BUG_ON_PAGE(PageTail(page), page);
154         VM_BUG_ON_PAGE(nr != 1 && shadow, page);
155
156         for (i = 0; i < nr; i++) {
157                 struct radix_tree_node *node;
158                 void **slot;
159
160                 __radix_tree_lookup(&mapping->i_pages, page->index + i,
161                                     &node, &slot);
162
163                 VM_BUG_ON_PAGE(!node && nr != 1, page);
164
165                 radix_tree_clear_tags(&mapping->i_pages, node, slot);
166                 __radix_tree_replace(&mapping->i_pages, node, slot, shadow,
167                                 workingset_lookup_update(mapping));
168         }
169
170         page->mapping = NULL;
171         /* Leave page->index set: truncation lookup relies upon it */
172
173         if (shadow) {
174                 mapping->nrexceptional += nr;
175                 /*
176                  * Make sure the nrexceptional update is committed before
177                  * the nrpages update so that final truncate racing
178                  * with reclaim does not see both counters 0 at the
179                  * same time and miss a shadow entry.
180                  */
181                 smp_wmb();
182         }
183         mapping->nrpages -= nr;
184 }
185
186 static void unaccount_page_cache_page(struct address_space *mapping,
187                                       struct page *page)
188 {
189         int nr;
190
191         /*
192          * if we're uptodate, flush out into the cleancache, otherwise
193          * invalidate any existing cleancache entries.  We can't leave
194          * stale data around in the cleancache once our page is gone
195          */
196         if (PageUptodate(page) && PageMappedToDisk(page))
197                 cleancache_put_page(page);
198         else
199                 cleancache_invalidate_page(mapping, page);
200
201         VM_BUG_ON_PAGE(PageTail(page), page);
202         VM_BUG_ON_PAGE(page_mapped(page), page);
203         if (!IS_ENABLED(CONFIG_DEBUG_VM) && unlikely(page_mapped(page))) {
204                 int mapcount;
205
206                 pr_alert("BUG: Bad page cache in process %s  pfn:%05lx\n",
207                          current->comm, page_to_pfn(page));
208                 dump_page(page, "still mapped when deleted");
209                 dump_stack();
210                 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
211
212                 mapcount = page_mapcount(page);
213                 if (mapping_exiting(mapping) &&
214                     page_count(page) >= mapcount + 2) {
215                         /*
216                          * All vmas have already been torn down, so it's
217                          * a good bet that actually the page is unmapped,
218                          * and we'd prefer not to leak it: if we're wrong,
219                          * some other bad page check should catch it later.
220                          */
221                         page_mapcount_reset(page);
222                         page_ref_sub(page, mapcount);
223                 }
224         }
225
226         /* hugetlb pages do not participate in page cache accounting. */
227         if (PageHuge(page))
228                 return;
229
230         nr = hpage_nr_pages(page);
231
232         __mod_node_page_state(page_pgdat(page), NR_FILE_PAGES, -nr);
233         if (PageSwapBacked(page)) {
234                 __mod_node_page_state(page_pgdat(page), NR_SHMEM, -nr);
235                 if (PageTransHuge(page))
236                         __dec_node_page_state(page, NR_SHMEM_THPS);
237         } else {
238                 VM_BUG_ON_PAGE(PageTransHuge(page), page);
239         }
240
241         /*
242          * At this point page must be either written or cleaned by
243          * truncate.  Dirty page here signals a bug and loss of
244          * unwritten data.
245          *
246          * This fixes dirty accounting after removing the page entirely
247          * but leaves PageDirty set: it has no effect for truncated
248          * page and anyway will be cleared before returning page into
249          * buddy allocator.
250          */
251         if (WARN_ON_ONCE(PageDirty(page)))
252                 account_page_cleaned(page, mapping, inode_to_wb(mapping->host));
253 }
254
255 /*
256  * Delete a page from the page cache and free it. Caller has to make
257  * sure the page is locked and that nobody else uses it - or that usage
258  * is safe.  The caller must hold the i_pages lock.
259  */
260 void __delete_from_page_cache(struct page *page, void *shadow)
261 {
262         struct address_space *mapping = page->mapping;
263
264         trace_mm_filemap_delete_from_page_cache(page);
265
266         unaccount_page_cache_page(mapping, page);
267         page_cache_tree_delete(mapping, page, shadow);
268 }
269
270 static void page_cache_free_page(struct address_space *mapping,
271                                 struct page *page)
272 {
273         void (*freepage)(struct page *);
274
275         freepage = mapping->a_ops->freepage;
276         if (freepage)
277                 freepage(page);
278
279         if (PageTransHuge(page) && !PageHuge(page)) {
280                 page_ref_sub(page, HPAGE_PMD_NR);
281                 VM_BUG_ON_PAGE(page_count(page) <= 0, page);
282         } else {
283                 put_page(page);
284         }
285 }
286
287 /**
288  * delete_from_page_cache - delete page from page cache
289  * @page: the page which the kernel is trying to remove from page cache
290  *
291  * This must be called only on pages that have been verified to be in the page
292  * cache and locked.  It will never put the page into the free list, the caller
293  * has a reference on the page.
294  */
295 void delete_from_page_cache(struct page *page)
296 {
297         struct address_space *mapping = page_mapping(page);
298         unsigned long flags;
299
300         BUG_ON(!PageLocked(page));
301         xa_lock_irqsave(&mapping->i_pages, flags);
302         __delete_from_page_cache(page, NULL);
303         xa_unlock_irqrestore(&mapping->i_pages, flags);
304
305         page_cache_free_page(mapping, page);
306 }
307 EXPORT_SYMBOL(delete_from_page_cache);
308
309 /*
310  * page_cache_tree_delete_batch - delete several pages from page cache
311  * @mapping: the mapping to which pages belong
312  * @pvec: pagevec with pages to delete
313  *
314  * The function walks over mapping->i_pages and removes pages passed in @pvec
315  * from the mapping. The function expects @pvec to be sorted by page index.
316  * It tolerates holes in @pvec (mapping entries at those indices are not
317  * modified). The function expects only THP head pages to be present in the
318  * @pvec and takes care to delete all corresponding tail pages from the
319  * mapping as well.
320  *
321  * The function expects the i_pages lock to be held.
322  */
323 static void
324 page_cache_tree_delete_batch(struct address_space *mapping,
325                              struct pagevec *pvec)
326 {
327         struct radix_tree_iter iter;
328         void **slot;
329         int total_pages = 0;
330         int i = 0, tail_pages = 0;
331         struct page *page;
332         pgoff_t start;
333
334         start = pvec->pages[0]->index;
335         radix_tree_for_each_slot(slot, &mapping->i_pages, &iter, start) {
336                 if (i >= pagevec_count(pvec) && !tail_pages)
337                         break;
338                 page = radix_tree_deref_slot_protected(slot,
339                                                        &mapping->i_pages.xa_lock);
340                 if (radix_tree_exceptional_entry(page))
341                         continue;
342                 if (!tail_pages) {
343                         /*
344                          * Some page got inserted in our range? Skip it. We
345                          * have our pages locked so they are protected from
346                          * being removed.
347                          */
348                         if (page != pvec->pages[i])
349                                 continue;
350                         WARN_ON_ONCE(!PageLocked(page));
351                         if (PageTransHuge(page) && !PageHuge(page))
352                                 tail_pages = HPAGE_PMD_NR - 1;
353                         page->mapping = NULL;
354                         /*
355                          * Leave page->index set: truncation lookup relies
356                          * upon it
357                          */
358                         i++;
359                 } else {
360                         tail_pages--;
361                 }
362                 radix_tree_clear_tags(&mapping->i_pages, iter.node, slot);
363                 __radix_tree_replace(&mapping->i_pages, iter.node, slot, NULL,
364                                 workingset_lookup_update(mapping));
365                 total_pages++;
366         }
367         mapping->nrpages -= total_pages;
368 }
369
370 void delete_from_page_cache_batch(struct address_space *mapping,
371                                   struct pagevec *pvec)
372 {
373         int i;
374         unsigned long flags;
375
376         if (!pagevec_count(pvec))
377                 return;
378
379         xa_lock_irqsave(&mapping->i_pages, flags);
380         for (i = 0; i < pagevec_count(pvec); i++) {
381                 trace_mm_filemap_delete_from_page_cache(pvec->pages[i]);
382
383                 unaccount_page_cache_page(mapping, pvec->pages[i]);
384         }
385         page_cache_tree_delete_batch(mapping, pvec);
386         xa_unlock_irqrestore(&mapping->i_pages, flags);
387
388         for (i = 0; i < pagevec_count(pvec); i++)
389                 page_cache_free_page(mapping, pvec->pages[i]);
390 }
391
392 int filemap_check_errors(struct address_space *mapping)
393 {
394         int ret = 0;
395         /* Check for outstanding write errors */
396         if (test_bit(AS_ENOSPC, &mapping->flags) &&
397             test_and_clear_bit(AS_ENOSPC, &mapping->flags))
398                 ret = -ENOSPC;
399         if (test_bit(AS_EIO, &mapping->flags) &&
400             test_and_clear_bit(AS_EIO, &mapping->flags))
401                 ret = -EIO;
402         return ret;
403 }
404 EXPORT_SYMBOL(filemap_check_errors);
405
406 static int filemap_check_and_keep_errors(struct address_space *mapping)
407 {
408         /* Check for outstanding write errors */
409         if (test_bit(AS_EIO, &mapping->flags))
410                 return -EIO;
411         if (test_bit(AS_ENOSPC, &mapping->flags))
412                 return -ENOSPC;
413         return 0;
414 }
415
416 /**
417  * __filemap_fdatawrite_range - start writeback on mapping dirty pages in range
418  * @mapping:    address space structure to write
419  * @start:      offset in bytes where the range starts
420  * @end:        offset in bytes where the range ends (inclusive)
421  * @sync_mode:  enable synchronous operation
422  *
423  * Start writeback against all of a mapping's dirty pages that lie
424  * within the byte offsets <start, end> inclusive.
425  *
426  * If sync_mode is WB_SYNC_ALL then this is a "data integrity" operation, as
427  * opposed to a regular memory cleansing writeback.  The difference between
428  * these two operations is that if a dirty page/buffer is encountered, it must
429  * be waited upon, and not just skipped over.
430  */
431 int __filemap_fdatawrite_range(struct address_space *mapping, loff_t start,
432                                 loff_t end, int sync_mode)
433 {
434         int ret;
435         struct writeback_control wbc = {
436                 .sync_mode = sync_mode,
437                 .nr_to_write = LONG_MAX,
438                 .range_start = start,
439                 .range_end = end,
440         };
441
442         if (!mapping_cap_writeback_dirty(mapping))
443                 return 0;
444
445         wbc_attach_fdatawrite_inode(&wbc, mapping->host);
446         ret = do_writepages(mapping, &wbc);
447         wbc_detach_inode(&wbc);
448         return ret;
449 }
450
451 static inline int __filemap_fdatawrite(struct address_space *mapping,
452         int sync_mode)
453 {
454         return __filemap_fdatawrite_range(mapping, 0, LLONG_MAX, sync_mode);
455 }
456
457 int filemap_fdatawrite(struct address_space *mapping)
458 {
459         return __filemap_fdatawrite(mapping, WB_SYNC_ALL);
460 }
461 EXPORT_SYMBOL(filemap_fdatawrite);
462
463 int filemap_fdatawrite_range(struct address_space *mapping, loff_t start,
464                                 loff_t end)
465 {
466         return __filemap_fdatawrite_range(mapping, start, end, WB_SYNC_ALL);
467 }
468 EXPORT_SYMBOL(filemap_fdatawrite_range);
469
470 /**
471  * filemap_flush - mostly a non-blocking flush
472  * @mapping:    target address_space
473  *
474  * This is a mostly non-blocking flush.  Not suitable for data-integrity
475  * purposes - I/O may not be started against all dirty pages.
476  */
477 int filemap_flush(struct address_space *mapping)
478 {
479         return __filemap_fdatawrite(mapping, WB_SYNC_NONE);
480 }
481 EXPORT_SYMBOL(filemap_flush);
482
483 /**
484  * filemap_range_has_page - check if a page exists in range.
485  * @mapping:           address space within which to check
486  * @start_byte:        offset in bytes where the range starts
487  * @end_byte:          offset in bytes where the range ends (inclusive)
488  *
489  * Find at least one page in the range supplied, usually used to check if
490  * direct writing in this range will trigger a writeback.
491  */
492 bool filemap_range_has_page(struct address_space *mapping,
493                            loff_t start_byte, loff_t end_byte)
494 {
495         pgoff_t index = start_byte >> PAGE_SHIFT;
496         pgoff_t end = end_byte >> PAGE_SHIFT;
497         struct page *page;
498
499         if (end_byte < start_byte)
500                 return false;
501
502         if (mapping->nrpages == 0)
503                 return false;
504
505         if (!find_get_pages_range(mapping, &index, end, 1, &page))
506                 return false;
507         put_page(page);
508         return true;
509 }
510 EXPORT_SYMBOL(filemap_range_has_page);
511
512 static void __filemap_fdatawait_range(struct address_space *mapping,
513                                      loff_t start_byte, loff_t end_byte)
514 {
515         pgoff_t index = start_byte >> PAGE_SHIFT;
516         pgoff_t end = end_byte >> PAGE_SHIFT;
517         struct pagevec pvec;
518         int nr_pages;
519
520         if (end_byte < start_byte)
521                 return;
522
523         pagevec_init(&pvec);
524         while (index <= end) {
525                 unsigned i;
526
527                 nr_pages = pagevec_lookup_range_tag(&pvec, mapping, &index,
528                                 end, PAGECACHE_TAG_WRITEBACK);
529                 if (!nr_pages)
530                         break;
531
532                 for (i = 0; i < nr_pages; i++) {
533                         struct page *page = pvec.pages[i];
534
535                         wait_on_page_writeback(page);
536                         ClearPageError(page);
537                 }
538                 pagevec_release(&pvec);
539                 cond_resched();
540         }
541 }
542
543 /**
544  * filemap_fdatawait_range - wait for writeback to complete
545  * @mapping:            address space structure to wait for
546  * @start_byte:         offset in bytes where the range starts
547  * @end_byte:           offset in bytes where the range ends (inclusive)
548  *
549  * Walk the list of under-writeback pages of the given address space
550  * in the given range and wait for all of them.  Check error status of
551  * the address space and return it.
552  *
553  * Since the error status of the address space is cleared by this function,
554  * callers are responsible for checking the return value and handling and/or
555  * reporting the error.
556  */
557 int filemap_fdatawait_range(struct address_space *mapping, loff_t start_byte,
558                             loff_t end_byte)
559 {
560         __filemap_fdatawait_range(mapping, start_byte, end_byte);
561         return filemap_check_errors(mapping);
562 }
563 EXPORT_SYMBOL(filemap_fdatawait_range);
564
565 /**
566  * file_fdatawait_range - wait for writeback to complete
567  * @file:               file pointing to address space structure to wait for
568  * @start_byte:         offset in bytes where the range starts
569  * @end_byte:           offset in bytes where the range ends (inclusive)
570  *
571  * Walk the list of under-writeback pages of the address space that file
572  * refers to, in the given range and wait for all of them.  Check error
573  * status of the address space vs. the file->f_wb_err cursor and return it.
574  *
575  * Since the error status of the file is advanced by this function,
576  * callers are responsible for checking the return value and handling and/or
577  * reporting the error.
578  */
579 int file_fdatawait_range(struct file *file, loff_t start_byte, loff_t end_byte)
580 {
581         struct address_space *mapping = file->f_mapping;
582
583         __filemap_fdatawait_range(mapping, start_byte, end_byte);
584         return file_check_and_advance_wb_err(file);
585 }
586 EXPORT_SYMBOL(file_fdatawait_range);
587
588 /**
589  * filemap_fdatawait_keep_errors - wait for writeback without clearing errors
590  * @mapping: address space structure to wait for
591  *
592  * Walk the list of under-writeback pages of the given address space
593  * and wait for all of them.  Unlike filemap_fdatawait(), this function
594  * does not clear error status of the address space.
595  *
596  * Use this function if callers don't handle errors themselves.  Expected
597  * call sites are system-wide / filesystem-wide data flushers: e.g. sync(2),
598  * fsfreeze(8)
599  */
600 int filemap_fdatawait_keep_errors(struct address_space *mapping)
601 {
602         __filemap_fdatawait_range(mapping, 0, LLONG_MAX);
603         return filemap_check_and_keep_errors(mapping);
604 }
605 EXPORT_SYMBOL(filemap_fdatawait_keep_errors);
606
607 static bool mapping_needs_writeback(struct address_space *mapping)
608 {
609         return (!dax_mapping(mapping) && mapping->nrpages) ||
610             (dax_mapping(mapping) && mapping->nrexceptional);
611 }
612
613 int filemap_write_and_wait(struct address_space *mapping)
614 {
615         int err = 0;
616
617         if (mapping_needs_writeback(mapping)) {
618                 err = filemap_fdatawrite(mapping);
619                 /*
620                  * Even if the above returned error, the pages may be
621                  * written partially (e.g. -ENOSPC), so we wait for it.
622                  * But the -EIO is special case, it may indicate the worst
623                  * thing (e.g. bug) happened, so we avoid waiting for it.
624                  */
625                 if (err != -EIO) {
626                         int err2 = filemap_fdatawait(mapping);
627                         if (!err)
628                                 err = err2;
629                 } else {
630                         /* Clear any previously stored errors */
631                         filemap_check_errors(mapping);
632                 }
633         } else {
634                 err = filemap_check_errors(mapping);
635         }
636         return err;
637 }
638 EXPORT_SYMBOL(filemap_write_and_wait);
639
640 /**
641  * filemap_write_and_wait_range - write out & wait on a file range
642  * @mapping:    the address_space for the pages
643  * @lstart:     offset in bytes where the range starts
644  * @lend:       offset in bytes where the range ends (inclusive)
645  *
646  * Write out and wait upon file offsets lstart->lend, inclusive.
647  *
648  * Note that @lend is inclusive (describes the last byte to be written) so
649  * that this function can be used to write to the very end-of-file (end = -1).
650  */
651 int filemap_write_and_wait_range(struct address_space *mapping,
652                                  loff_t lstart, loff_t lend)
653 {
654         int err = 0;
655
656         if (mapping_needs_writeback(mapping)) {
657                 err = __filemap_fdatawrite_range(mapping, lstart, lend,
658                                                  WB_SYNC_ALL);
659                 /* See comment of filemap_write_and_wait() */
660                 if (err != -EIO) {
661                         int err2 = filemap_fdatawait_range(mapping,
662                                                 lstart, lend);
663                         if (!err)
664                                 err = err2;
665                 } else {
666                         /* Clear any previously stored errors */
667                         filemap_check_errors(mapping);
668                 }
669         } else {
670                 err = filemap_check_errors(mapping);
671         }
672         return err;
673 }
674 EXPORT_SYMBOL(filemap_write_and_wait_range);
675
676 void __filemap_set_wb_err(struct address_space *mapping, int err)
677 {
678         errseq_t eseq = errseq_set(&mapping->wb_err, err);
679
680         trace_filemap_set_wb_err(mapping, eseq);
681 }
682 EXPORT_SYMBOL(__filemap_set_wb_err);
683
684 /**
685  * file_check_and_advance_wb_err - report wb error (if any) that was previously
686  *                                 and advance wb_err to current one
687  * @file: struct file on which the error is being reported
688  *
689  * When userland calls fsync (or something like nfsd does the equivalent), we
690  * want to report any writeback errors that occurred since the last fsync (or
691  * since the file was opened if there haven't been any).
692  *
693  * Grab the wb_err from the mapping. If it matches what we have in the file,
694  * then just quickly return 0. The file is all caught up.
695  *
696  * If it doesn't match, then take the mapping value, set the "seen" flag in
697  * it and try to swap it into place. If it works, or another task beat us
698  * to it with the new value, then update the f_wb_err and return the error
699  * portion. The error at this point must be reported via proper channels
700  * (a'la fsync, or NFS COMMIT operation, etc.).
701  *
702  * While we handle mapping->wb_err with atomic operations, the f_wb_err
703  * value is protected by the f_lock since we must ensure that it reflects
704  * the latest value swapped in for this file descriptor.
705  */
706 int file_check_and_advance_wb_err(struct file *file)
707 {
708         int err = 0;
709         errseq_t old = READ_ONCE(file->f_wb_err);
710         struct address_space *mapping = file->f_mapping;
711
712         /* Locklessly handle the common case where nothing has changed */
713         if (errseq_check(&mapping->wb_err, old)) {
714                 /* Something changed, must use slow path */
715                 spin_lock(&file->f_lock);
716                 old = file->f_wb_err;
717                 err = errseq_check_and_advance(&mapping->wb_err,
718                                                 &file->f_wb_err);
719                 trace_file_check_and_advance_wb_err(file, old);
720                 spin_unlock(&file->f_lock);
721         }
722
723         /*
724          * We're mostly using this function as a drop in replacement for
725          * filemap_check_errors. Clear AS_EIO/AS_ENOSPC to emulate the effect
726          * that the legacy code would have had on these flags.
727          */
728         clear_bit(AS_EIO, &mapping->flags);
729         clear_bit(AS_ENOSPC, &mapping->flags);
730         return err;
731 }
732 EXPORT_SYMBOL(file_check_and_advance_wb_err);
733
734 /**
735  * file_write_and_wait_range - write out & wait on a file range
736  * @file:       file pointing to address_space with pages
737  * @lstart:     offset in bytes where the range starts
738  * @lend:       offset in bytes where the range ends (inclusive)
739  *
740  * Write out and wait upon file offsets lstart->lend, inclusive.
741  *
742  * Note that @lend is inclusive (describes the last byte to be written) so
743  * that this function can be used to write to the very end-of-file (end = -1).
744  *
745  * After writing out and waiting on the data, we check and advance the
746  * f_wb_err cursor to the latest value, and return any errors detected there.
747  */
748 int file_write_and_wait_range(struct file *file, loff_t lstart, loff_t lend)
749 {
750         int err = 0, err2;
751         struct address_space *mapping = file->f_mapping;
752
753         if (mapping_needs_writeback(mapping)) {
754                 err = __filemap_fdatawrite_range(mapping, lstart, lend,
755                                                  WB_SYNC_ALL);
756                 /* See comment of filemap_write_and_wait() */
757                 if (err != -EIO)
758                         __filemap_fdatawait_range(mapping, lstart, lend);
759         }
760         err2 = file_check_and_advance_wb_err(file);
761         if (!err)
762                 err = err2;
763         return err;
764 }
765 EXPORT_SYMBOL(file_write_and_wait_range);
766
767 /**
768  * replace_page_cache_page - replace a pagecache page with a new one
769  * @old:        page to be replaced
770  * @new:        page to replace with
771  * @gfp_mask:   allocation mode
772  *
773  * This function replaces a page in the pagecache with a new one.  On
774  * success it acquires the pagecache reference for the new page and
775  * drops it for the old page.  Both the old and new pages must be
776  * locked.  This function does not add the new page to the LRU, the
777  * caller must do that.
778  *
779  * The remove + add is atomic.  The only way this function can fail is
780  * memory allocation failure.
781  */
782 int replace_page_cache_page(struct page *old, struct page *new, gfp_t gfp_mask)
783 {
784         int error;
785
786         VM_BUG_ON_PAGE(!PageLocked(old), old);
787         VM_BUG_ON_PAGE(!PageLocked(new), new);
788         VM_BUG_ON_PAGE(new->mapping, new);
789
790         error = radix_tree_preload(gfp_mask & GFP_RECLAIM_MASK);
791         if (!error) {
792                 struct address_space *mapping = old->mapping;
793                 void (*freepage)(struct page *);
794                 unsigned long flags;
795
796                 pgoff_t offset = old->index;
797                 freepage = mapping->a_ops->freepage;
798
799                 get_page(new);
800                 new->mapping = mapping;
801                 new->index = offset;
802
803                 xa_lock_irqsave(&mapping->i_pages, flags);
804                 __delete_from_page_cache(old, NULL);
805                 error = page_cache_tree_insert(mapping, new, NULL);
806                 BUG_ON(error);
807
808                 /*
809                  * hugetlb pages do not participate in page cache accounting.
810                  */
811                 if (!PageHuge(new))
812                         __inc_node_page_state(new, NR_FILE_PAGES);
813                 if (PageSwapBacked(new))
814                         __inc_node_page_state(new, NR_SHMEM);
815                 xa_unlock_irqrestore(&mapping->i_pages, flags);
816                 mem_cgroup_migrate(old, new);
817                 radix_tree_preload_end();
818                 if (freepage)
819                         freepage(old);
820                 put_page(old);
821         }
822
823         return error;
824 }
825 EXPORT_SYMBOL_GPL(replace_page_cache_page);
826
827 static int __add_to_page_cache_locked(struct page *page,
828                                       struct address_space *mapping,
829                                       pgoff_t offset, gfp_t gfp_mask,
830                                       void **shadowp)
831 {
832         int huge = PageHuge(page);
833         struct mem_cgroup *memcg;
834         int error;
835
836         VM_BUG_ON_PAGE(!PageLocked(page), page);
837         VM_BUG_ON_PAGE(PageSwapBacked(page), page);
838
839         if (!huge) {
840                 error = mem_cgroup_try_charge(page, current->mm,
841                                               gfp_mask, &memcg, false);
842                 if (error)
843                         return error;
844         }
845
846         error = radix_tree_maybe_preload(gfp_mask & GFP_RECLAIM_MASK);
847         if (error) {
848                 if (!huge)
849                         mem_cgroup_cancel_charge(page, memcg, false);
850                 return error;
851         }
852
853         get_page(page);
854         page->mapping = mapping;
855         page->index = offset;
856
857         xa_lock_irq(&mapping->i_pages);
858         error = page_cache_tree_insert(mapping, page, shadowp);
859         radix_tree_preload_end();
860         if (unlikely(error))
861                 goto err_insert;
862
863         /* hugetlb pages do not participate in page cache accounting. */
864         if (!huge)
865                 __inc_node_page_state(page, NR_FILE_PAGES);
866         xa_unlock_irq(&mapping->i_pages);
867         if (!huge)
868                 mem_cgroup_commit_charge(page, memcg, false, false);
869         trace_mm_filemap_add_to_page_cache(page);
870         return 0;
871 err_insert:
872         page->mapping = NULL;
873         /* Leave page->index set: truncation relies upon it */
874         xa_unlock_irq(&mapping->i_pages);
875         if (!huge)
876                 mem_cgroup_cancel_charge(page, memcg, false);
877         put_page(page);
878         return error;
879 }
880
881 /**
882  * add_to_page_cache_locked - add a locked page to the pagecache
883  * @page:       page to add
884  * @mapping:    the page's address_space
885  * @offset:     page index
886  * @gfp_mask:   page allocation mode
887  *
888  * This function is used to add a page to the pagecache. It must be locked.
889  * This function does not add the page to the LRU.  The caller must do that.
890  */
891 int add_to_page_cache_locked(struct page *page, struct address_space *mapping,
892                 pgoff_t offset, gfp_t gfp_mask)
893 {
894         return __add_to_page_cache_locked(page, mapping, offset,
895                                           gfp_mask, NULL);
896 }
897 EXPORT_SYMBOL(add_to_page_cache_locked);
898
899 int add_to_page_cache_lru(struct page *page, struct address_space *mapping,
900                                 pgoff_t offset, gfp_t gfp_mask)
901 {
902         void *shadow = NULL;
903         int ret;
904
905         __SetPageLocked(page);
906         ret = __add_to_page_cache_locked(page, mapping, offset,
907                                          gfp_mask, &shadow);
908         if (unlikely(ret))
909                 __ClearPageLocked(page);
910         else {
911                 /*
912                  * The page might have been evicted from cache only
913                  * recently, in which case it should be activated like
914                  * any other repeatedly accessed page.
915                  * The exception is pages getting rewritten; evicting other
916                  * data from the working set, only to cache data that will
917                  * get overwritten with something else, is a waste of memory.
918                  */
919                 WARN_ON_ONCE(PageActive(page));
920                 if (!(gfp_mask & __GFP_WRITE) && shadow)
921                         workingset_refault(page, shadow);
922                 lru_cache_add(page);
923         }
924         return ret;
925 }
926 EXPORT_SYMBOL_GPL(add_to_page_cache_lru);
927
928 #ifdef CONFIG_NUMA
929 struct page *__page_cache_alloc(gfp_t gfp)
930 {
931         int n;
932         struct page *page;
933
934         if (cpuset_do_page_mem_spread()) {
935                 unsigned int cpuset_mems_cookie;
936                 do {
937                         cpuset_mems_cookie = read_mems_allowed_begin();
938                         n = cpuset_mem_spread_node();
939                         page = __alloc_pages_node(n, gfp, 0);
940                 } while (!page && read_mems_allowed_retry(cpuset_mems_cookie));
941
942                 return page;
943         }
944         return alloc_pages(gfp, 0);
945 }
946 EXPORT_SYMBOL(__page_cache_alloc);
947 #endif
948
949 /*
950  * In order to wait for pages to become available there must be
951  * waitqueues associated with pages. By using a hash table of
952  * waitqueues where the bucket discipline is to maintain all
953  * waiters on the same queue and wake all when any of the pages
954  * become available, and for the woken contexts to check to be
955  * sure the appropriate page became available, this saves space
956  * at a cost of "thundering herd" phenomena during rare hash
957  * collisions.
958  */
959 #define PAGE_WAIT_TABLE_BITS 8
960 #define PAGE_WAIT_TABLE_SIZE (1 << PAGE_WAIT_TABLE_BITS)
961 static wait_queue_head_t page_wait_table[PAGE_WAIT_TABLE_SIZE] __cacheline_aligned;
962
963 static wait_queue_head_t *page_waitqueue(struct page *page)
964 {
965         return &page_wait_table[hash_ptr(page, PAGE_WAIT_TABLE_BITS)];
966 }
967
968 void __init pagecache_init(void)
969 {
970         int i;
971
972         for (i = 0; i < PAGE_WAIT_TABLE_SIZE; i++)
973                 init_waitqueue_head(&page_wait_table[i]);
974
975         page_writeback_init();
976 }
977
978 /* This has the same layout as wait_bit_key - see fs/cachefiles/rdwr.c */
979 struct wait_page_key {
980         struct page *page;
981         int bit_nr;
982         int page_match;
983 };
984
985 struct wait_page_queue {
986         struct page *page;
987         int bit_nr;
988         wait_queue_entry_t wait;
989 };
990
991 static int wake_page_function(wait_queue_entry_t *wait, unsigned mode, int sync, void *arg)
992 {
993         struct wait_page_key *key = arg;
994         struct wait_page_queue *wait_page
995                 = container_of(wait, struct wait_page_queue, wait);
996
997         if (wait_page->page != key->page)
998                return 0;
999         key->page_match = 1;
1000
1001         if (wait_page->bit_nr != key->bit_nr)
1002                 return 0;
1003
1004         /* Stop walking if it's locked */
1005         if (test_bit(key->bit_nr, &key->page->flags))
1006                 return -1;
1007
1008         return autoremove_wake_function(wait, mode, sync, key);
1009 }
1010
1011 static void wake_up_page_bit(struct page *page, int bit_nr)
1012 {
1013         wait_queue_head_t *q = page_waitqueue(page);
1014         struct wait_page_key key;
1015         unsigned long flags;
1016         wait_queue_entry_t bookmark;
1017
1018         key.page = page;
1019         key.bit_nr = bit_nr;
1020         key.page_match = 0;
1021
1022         bookmark.flags = 0;
1023         bookmark.private = NULL;
1024         bookmark.func = NULL;
1025         INIT_LIST_HEAD(&bookmark.entry);
1026
1027         spin_lock_irqsave(&q->lock, flags);
1028         __wake_up_locked_key_bookmark(q, TASK_NORMAL, &key, &bookmark);
1029
1030         while (bookmark.flags & WQ_FLAG_BOOKMARK) {
1031                 /*
1032                  * Take a breather from holding the lock,
1033                  * allow pages that finish wake up asynchronously
1034                  * to acquire the lock and remove themselves
1035                  * from wait queue
1036                  */
1037                 spin_unlock_irqrestore(&q->lock, flags);
1038                 cpu_relax();
1039                 spin_lock_irqsave(&q->lock, flags);
1040                 __wake_up_locked_key_bookmark(q, TASK_NORMAL, &key, &bookmark);
1041         }
1042
1043         /*
1044          * It is possible for other pages to have collided on the waitqueue
1045          * hash, so in that case check for a page match. That prevents a long-
1046          * term waiter
1047          *
1048          * It is still possible to miss a case here, when we woke page waiters
1049          * and removed them from the waitqueue, but there are still other
1050          * page waiters.
1051          */
1052         if (!waitqueue_active(q) || !key.page_match) {
1053                 ClearPageWaiters(page);
1054                 /*
1055                  * It's possible to miss clearing Waiters here, when we woke
1056                  * our page waiters, but the hashed waitqueue has waiters for
1057                  * other pages on it.
1058                  *
1059                  * That's okay, it's a rare case. The next waker will clear it.
1060                  */
1061         }
1062         spin_unlock_irqrestore(&q->lock, flags);
1063 }
1064
1065 static void wake_up_page(struct page *page, int bit)
1066 {
1067         if (!PageWaiters(page))
1068                 return;
1069         wake_up_page_bit(page, bit);
1070 }
1071
1072 static inline int wait_on_page_bit_common(wait_queue_head_t *q,
1073                 struct page *page, int bit_nr, int state, bool lock)
1074 {
1075         struct wait_page_queue wait_page;
1076         wait_queue_entry_t *wait = &wait_page.wait;
1077         bool thrashing = false;
1078         int ret = 0;
1079
1080         if (bit_nr == PG_locked && !PageSwapBacked(page) &&
1081             !PageUptodate(page) && PageWorkingset(page)) {
1082                 delayacct_thrashing_start();
1083                 thrashing = true;
1084         }
1085
1086         init_wait(wait);
1087         wait->flags = lock ? WQ_FLAG_EXCLUSIVE : 0;
1088         wait->func = wake_page_function;
1089         wait_page.page = page;
1090         wait_page.bit_nr = bit_nr;
1091
1092         for (;;) {
1093                 spin_lock_irq(&q->lock);
1094
1095                 if (likely(list_empty(&wait->entry))) {
1096                         __add_wait_queue_entry_tail(q, wait);
1097                         SetPageWaiters(page);
1098                 }
1099
1100                 set_current_state(state);
1101
1102                 spin_unlock_irq(&q->lock);
1103
1104                 if (likely(test_bit(bit_nr, &page->flags))) {
1105                         io_schedule();
1106                 }
1107
1108                 if (lock) {
1109                         if (!test_and_set_bit_lock(bit_nr, &page->flags))
1110                                 break;
1111                 } else {
1112                         if (!test_bit(bit_nr, &page->flags))
1113                                 break;
1114                 }
1115
1116                 if (unlikely(signal_pending_state(state, current))) {
1117                         ret = -EINTR;
1118                         break;
1119                 }
1120         }
1121
1122         finish_wait(q, wait);
1123
1124         if (thrashing)
1125                 delayacct_thrashing_end();
1126
1127         /*
1128          * A signal could leave PageWaiters set. Clearing it here if
1129          * !waitqueue_active would be possible (by open-coding finish_wait),
1130          * but still fail to catch it in the case of wait hash collision. We
1131          * already can fail to clear wait hash collision cases, so don't
1132          * bother with signals either.
1133          */
1134
1135         return ret;
1136 }
1137
1138 void wait_on_page_bit(struct page *page, int bit_nr)
1139 {
1140         wait_queue_head_t *q = page_waitqueue(page);
1141         wait_on_page_bit_common(q, page, bit_nr, TASK_UNINTERRUPTIBLE, false);
1142 }
1143 EXPORT_SYMBOL(wait_on_page_bit);
1144
1145 int wait_on_page_bit_killable(struct page *page, int bit_nr)
1146 {
1147         wait_queue_head_t *q = page_waitqueue(page);
1148         return wait_on_page_bit_common(q, page, bit_nr, TASK_KILLABLE, false);
1149 }
1150 EXPORT_SYMBOL(wait_on_page_bit_killable);
1151
1152 /**
1153  * add_page_wait_queue - Add an arbitrary waiter to a page's wait queue
1154  * @page: Page defining the wait queue of interest
1155  * @waiter: Waiter to add to the queue
1156  *
1157  * Add an arbitrary @waiter to the wait queue for the nominated @page.
1158  */
1159 void add_page_wait_queue(struct page *page, wait_queue_entry_t *waiter)
1160 {
1161         wait_queue_head_t *q = page_waitqueue(page);
1162         unsigned long flags;
1163
1164         spin_lock_irqsave(&q->lock, flags);
1165         __add_wait_queue_entry_tail(q, waiter);
1166         SetPageWaiters(page);
1167         spin_unlock_irqrestore(&q->lock, flags);
1168 }
1169 EXPORT_SYMBOL_GPL(add_page_wait_queue);
1170
1171 #ifndef clear_bit_unlock_is_negative_byte
1172
1173 /*
1174  * PG_waiters is the high bit in the same byte as PG_lock.
1175  *
1176  * On x86 (and on many other architectures), we can clear PG_lock and
1177  * test the sign bit at the same time. But if the architecture does
1178  * not support that special operation, we just do this all by hand
1179  * instead.
1180  *
1181  * The read of PG_waiters has to be after (or concurrently with) PG_locked
1182  * being cleared, but a memory barrier should be unneccssary since it is
1183  * in the same byte as PG_locked.
1184  */
1185 static inline bool clear_bit_unlock_is_negative_byte(long nr, volatile void *mem)
1186 {
1187         clear_bit_unlock(nr, mem);
1188         /* smp_mb__after_atomic(); */
1189         return test_bit(PG_waiters, mem);
1190 }
1191
1192 #endif
1193
1194 /**
1195  * unlock_page - unlock a locked page
1196  * @page: the page
1197  *
1198  * Unlocks the page and wakes up sleepers in ___wait_on_page_locked().
1199  * Also wakes sleepers in wait_on_page_writeback() because the wakeup
1200  * mechanism between PageLocked pages and PageWriteback pages is shared.
1201  * But that's OK - sleepers in wait_on_page_writeback() just go back to sleep.
1202  *
1203  * Note that this depends on PG_waiters being the sign bit in the byte
1204  * that contains PG_locked - thus the BUILD_BUG_ON(). That allows us to
1205  * clear the PG_locked bit and test PG_waiters at the same time fairly
1206  * portably (architectures that do LL/SC can test any bit, while x86 can
1207  * test the sign bit).
1208  */
1209 void unlock_page(struct page *page)
1210 {
1211         BUILD_BUG_ON(PG_waiters != 7);
1212         page = compound_head(page);
1213         VM_BUG_ON_PAGE(!PageLocked(page), page);
1214         if (clear_bit_unlock_is_negative_byte(PG_locked, &page->flags))
1215                 wake_up_page_bit(page, PG_locked);
1216 }
1217 EXPORT_SYMBOL(unlock_page);
1218
1219 /**
1220  * end_page_writeback - end writeback against a page
1221  * @page: the page
1222  */
1223 void end_page_writeback(struct page *page)
1224 {
1225         /*
1226          * TestClearPageReclaim could be used here but it is an atomic
1227          * operation and overkill in this particular case. Failing to
1228          * shuffle a page marked for immediate reclaim is too mild to
1229          * justify taking an atomic operation penalty at the end of
1230          * ever page writeback.
1231          */
1232         if (PageReclaim(page)) {
1233                 ClearPageReclaim(page);
1234                 rotate_reclaimable_page(page);
1235         }
1236
1237         if (!test_clear_page_writeback(page))
1238                 BUG();
1239
1240         smp_mb__after_atomic();
1241         wake_up_page(page, PG_writeback);
1242 }
1243 EXPORT_SYMBOL(end_page_writeback);
1244
1245 /*
1246  * After completing I/O on a page, call this routine to update the page
1247  * flags appropriately
1248  */
1249 void page_endio(struct page *page, bool is_write, int err)
1250 {
1251         if (!is_write) {
1252                 if (!err) {
1253                         SetPageUptodate(page);
1254                 } else {
1255                         ClearPageUptodate(page);
1256                         SetPageError(page);
1257                 }
1258                 unlock_page(page);
1259         } else {
1260                 if (err) {
1261                         struct address_space *mapping;
1262
1263                         SetPageError(page);
1264                         mapping = page_mapping(page);
1265                         if (mapping)
1266                                 mapping_set_error(mapping, err);
1267                 }
1268                 end_page_writeback(page);
1269         }
1270 }
1271 EXPORT_SYMBOL_GPL(page_endio);
1272
1273 /**
1274  * __lock_page - get a lock on the page, assuming we need to sleep to get it
1275  * @__page: the page to lock
1276  */
1277 void __lock_page(struct page *__page)
1278 {
1279         struct page *page = compound_head(__page);
1280         wait_queue_head_t *q = page_waitqueue(page);
1281         wait_on_page_bit_common(q, page, PG_locked, TASK_UNINTERRUPTIBLE, true);
1282 }
1283 EXPORT_SYMBOL(__lock_page);
1284
1285 int __lock_page_killable(struct page *__page)
1286 {
1287         struct page *page = compound_head(__page);
1288         wait_queue_head_t *q = page_waitqueue(page);
1289         return wait_on_page_bit_common(q, page, PG_locked, TASK_KILLABLE, true);
1290 }
1291 EXPORT_SYMBOL_GPL(__lock_page_killable);
1292
1293 /*
1294  * Return values:
1295  * 1 - page is locked; mmap_sem is still held.
1296  * 0 - page is not locked.
1297  *     mmap_sem has been released (up_read()), unless flags had both
1298  *     FAULT_FLAG_ALLOW_RETRY and FAULT_FLAG_RETRY_NOWAIT set, in
1299  *     which case mmap_sem is still held.
1300  *
1301  * If neither ALLOW_RETRY nor KILLABLE are set, will always return 1
1302  * with the page locked and the mmap_sem unperturbed.
1303  */
1304 int __lock_page_or_retry(struct page *page, struct mm_struct *mm,
1305                          unsigned int flags)
1306 {
1307         if (flags & FAULT_FLAG_ALLOW_RETRY) {
1308                 /*
1309                  * CAUTION! In this case, mmap_sem is not released
1310                  * even though return 0.
1311                  */
1312                 if (flags & FAULT_FLAG_RETRY_NOWAIT)
1313                         return 0;
1314
1315                 up_read(&mm->mmap_sem);
1316                 if (flags & FAULT_FLAG_KILLABLE)
1317                         wait_on_page_locked_killable(page);
1318                 else
1319                         wait_on_page_locked(page);
1320                 return 0;
1321         } else {
1322                 if (flags & FAULT_FLAG_KILLABLE) {
1323                         int ret;
1324
1325                         ret = __lock_page_killable(page);
1326                         if (ret) {
1327                                 up_read(&mm->mmap_sem);
1328                                 return 0;
1329                         }
1330                 } else
1331                         __lock_page(page);
1332                 return 1;
1333         }
1334 }
1335
1336 /**
1337  * page_cache_next_hole - find the next hole (not-present entry)
1338  * @mapping: mapping
1339  * @index: index
1340  * @max_scan: maximum range to search
1341  *
1342  * Search the set [index, min(index+max_scan-1, MAX_INDEX)] for the
1343  * lowest indexed hole.
1344  *
1345  * Returns: the index of the hole if found, otherwise returns an index
1346  * outside of the set specified (in which case 'return - index >=
1347  * max_scan' will be true). In rare cases of index wrap-around, 0 will
1348  * be returned.
1349  *
1350  * page_cache_next_hole may be called under rcu_read_lock. However,
1351  * like radix_tree_gang_lookup, this will not atomically search a
1352  * snapshot of the tree at a single point in time. For example, if a
1353  * hole is created at index 5, then subsequently a hole is created at
1354  * index 10, page_cache_next_hole covering both indexes may return 10
1355  * if called under rcu_read_lock.
1356  */
1357 pgoff_t page_cache_next_hole(struct address_space *mapping,
1358                              pgoff_t index, unsigned long max_scan)
1359 {
1360         unsigned long i;
1361
1362         for (i = 0; i < max_scan; i++) {
1363                 struct page *page;
1364
1365                 page = radix_tree_lookup(&mapping->i_pages, index);
1366                 if (!page || radix_tree_exceptional_entry(page))
1367                         break;
1368                 index++;
1369                 if (index == 0)
1370                         break;
1371         }
1372
1373         return index;
1374 }
1375 EXPORT_SYMBOL(page_cache_next_hole);
1376
1377 /**
1378  * page_cache_prev_hole - find the prev hole (not-present entry)
1379  * @mapping: mapping
1380  * @index: index
1381  * @max_scan: maximum range to search
1382  *
1383  * Search backwards in the range [max(index-max_scan+1, 0), index] for
1384  * the first hole.
1385  *
1386  * Returns: the index of the hole if found, otherwise returns an index
1387  * outside of the set specified (in which case 'index - return >=
1388  * max_scan' will be true). In rare cases of wrap-around, ULONG_MAX
1389  * will be returned.
1390  *
1391  * page_cache_prev_hole may be called under rcu_read_lock. However,
1392  * like radix_tree_gang_lookup, this will not atomically search a
1393  * snapshot of the tree at a single point in time. For example, if a
1394  * hole is created at index 10, then subsequently a hole is created at
1395  * index 5, page_cache_prev_hole covering both indexes may return 5 if
1396  * called under rcu_read_lock.
1397  */
1398 pgoff_t page_cache_prev_hole(struct address_space *mapping,
1399                              pgoff_t index, unsigned long max_scan)
1400 {
1401         unsigned long i;
1402
1403         for (i = 0; i < max_scan; i++) {
1404                 struct page *page;
1405
1406                 page = radix_tree_lookup(&mapping->i_pages, index);
1407                 if (!page || radix_tree_exceptional_entry(page))
1408                         break;
1409                 index--;
1410                 if (index == ULONG_MAX)
1411                         break;
1412         }
1413
1414         return index;
1415 }
1416 EXPORT_SYMBOL(page_cache_prev_hole);
1417
1418 /**
1419  * find_get_entry - find and get a page cache entry
1420  * @mapping: the address_space to search
1421  * @offset: the page cache index
1422  *
1423  * Looks up the page cache slot at @mapping & @offset.  If there is a
1424  * page cache page, it is returned with an increased refcount.
1425  *
1426  * If the slot holds a shadow entry of a previously evicted page, or a
1427  * swap entry from shmem/tmpfs, it is returned.
1428  *
1429  * Otherwise, %NULL is returned.
1430  */
1431 struct page *find_get_entry(struct address_space *mapping, pgoff_t offset)
1432 {
1433         void **pagep;
1434         struct page *head, *page;
1435
1436         rcu_read_lock();
1437 repeat:
1438         page = NULL;
1439         pagep = radix_tree_lookup_slot(&mapping->i_pages, offset);
1440         if (pagep) {
1441                 page = radix_tree_deref_slot(pagep);
1442                 if (unlikely(!page))
1443                         goto out;
1444                 if (radix_tree_exception(page)) {
1445                         if (radix_tree_deref_retry(page))
1446                                 goto repeat;
1447                         /*
1448                          * A shadow entry of a recently evicted page,
1449                          * or a swap entry from shmem/tmpfs.  Return
1450                          * it without attempting to raise page count.
1451                          */
1452                         goto out;
1453                 }
1454
1455                 head = compound_head(page);
1456                 if (!page_cache_get_speculative(head))
1457                         goto repeat;
1458
1459                 /* The page was split under us? */
1460                 if (compound_head(page) != head) {
1461                         put_page(head);
1462                         goto repeat;
1463                 }
1464
1465                 /*
1466                  * Has the page moved?
1467                  * This is part of the lockless pagecache protocol. See
1468                  * include/linux/pagemap.h for details.
1469                  */
1470                 if (unlikely(page != *pagep)) {
1471                         put_page(head);
1472                         goto repeat;
1473                 }
1474         }
1475 out:
1476         rcu_read_unlock();
1477
1478         return page;
1479 }
1480 EXPORT_SYMBOL(find_get_entry);
1481
1482 /**
1483  * find_lock_entry - locate, pin and lock a page cache entry
1484  * @mapping: the address_space to search
1485  * @offset: the page cache index
1486  *
1487  * Looks up the page cache slot at @mapping & @offset.  If there is a
1488  * page cache page, it is returned locked and with an increased
1489  * refcount.
1490  *
1491  * If the slot holds a shadow entry of a previously evicted page, or a
1492  * swap entry from shmem/tmpfs, it is returned.
1493  *
1494  * Otherwise, %NULL is returned.
1495  *
1496  * find_lock_entry() may sleep.
1497  */
1498 struct page *find_lock_entry(struct address_space *mapping, pgoff_t offset)
1499 {
1500         struct page *page;
1501
1502 repeat:
1503         page = find_get_entry(mapping, offset);
1504         if (page && !radix_tree_exception(page)) {
1505                 lock_page(page);
1506                 /* Has the page been truncated? */
1507                 if (unlikely(page_mapping(page) != mapping)) {
1508                         unlock_page(page);
1509                         put_page(page);
1510                         goto repeat;
1511                 }
1512                 VM_BUG_ON_PAGE(page_to_pgoff(page) != offset, page);
1513         }
1514         return page;
1515 }
1516 EXPORT_SYMBOL(find_lock_entry);
1517
1518 /**
1519  * pagecache_get_page - find and get a page reference
1520  * @mapping: the address_space to search
1521  * @offset: the page index
1522  * @fgp_flags: PCG flags
1523  * @gfp_mask: gfp mask to use for the page cache data page allocation
1524  *
1525  * Looks up the page cache slot at @mapping & @offset.
1526  *
1527  * PCG flags modify how the page is returned.
1528  *
1529  * @fgp_flags can be:
1530  *
1531  * - FGP_ACCESSED: the page will be marked accessed
1532  * - FGP_LOCK: Page is return locked
1533  * - FGP_CREAT: If page is not present then a new page is allocated using
1534  *   @gfp_mask and added to the page cache and the VM's LRU
1535  *   list. The page is returned locked and with an increased
1536  *   refcount. Otherwise, NULL is returned.
1537  *
1538  * If FGP_LOCK or FGP_CREAT are specified then the function may sleep even
1539  * if the GFP flags specified for FGP_CREAT are atomic.
1540  *
1541  * If there is a page cache page, it is returned with an increased refcount.
1542  */
1543 struct page *pagecache_get_page(struct address_space *mapping, pgoff_t offset,
1544         int fgp_flags, gfp_t gfp_mask)
1545 {
1546         struct page *page;
1547
1548 repeat:
1549         page = find_get_entry(mapping, offset);
1550         if (radix_tree_exceptional_entry(page))
1551                 page = NULL;
1552         if (!page)
1553                 goto no_page;
1554
1555         if (fgp_flags & FGP_LOCK) {
1556                 if (fgp_flags & FGP_NOWAIT) {
1557                         if (!trylock_page(page)) {
1558                                 put_page(page);
1559                                 return NULL;
1560                         }
1561                 } else {
1562                         lock_page(page);
1563                 }
1564
1565                 /* Has the page been truncated? */
1566                 if (unlikely(page->mapping != mapping)) {
1567                         unlock_page(page);
1568                         put_page(page);
1569                         goto repeat;
1570                 }
1571                 VM_BUG_ON_PAGE(page->index != offset, page);
1572         }
1573
1574         if (page && (fgp_flags & FGP_ACCESSED))
1575                 mark_page_accessed(page);
1576
1577 no_page:
1578         if (!page && (fgp_flags & FGP_CREAT)) {
1579                 int err;
1580                 if ((fgp_flags & FGP_WRITE) && mapping_cap_account_dirty(mapping))
1581                         gfp_mask |= __GFP_WRITE;
1582                 if (fgp_flags & FGP_NOFS)
1583                         gfp_mask &= ~__GFP_FS;
1584
1585                 page = __page_cache_alloc(gfp_mask);
1586                 if (!page)
1587                         return NULL;
1588
1589                 if (WARN_ON_ONCE(!(fgp_flags & FGP_LOCK)))
1590                         fgp_flags |= FGP_LOCK;
1591
1592                 /* Init accessed so avoid atomic mark_page_accessed later */
1593                 if (fgp_flags & FGP_ACCESSED)
1594                         __SetPageReferenced(page);
1595
1596                 err = add_to_page_cache_lru(page, mapping, offset, gfp_mask);
1597                 if (unlikely(err)) {
1598                         put_page(page);
1599                         page = NULL;
1600                         if (err == -EEXIST)
1601                                 goto repeat;
1602                 }
1603         }
1604
1605         return page;
1606 }
1607 EXPORT_SYMBOL(pagecache_get_page);
1608
1609 /**
1610  * find_get_entries - gang pagecache lookup
1611  * @mapping:    The address_space to search
1612  * @start:      The starting page cache index
1613  * @nr_entries: The maximum number of entries
1614  * @entries:    Where the resulting entries are placed
1615  * @indices:    The cache indices corresponding to the entries in @entries
1616  *
1617  * find_get_entries() will search for and return a group of up to
1618  * @nr_entries entries in the mapping.  The entries are placed at
1619  * @entries.  find_get_entries() takes a reference against any actual
1620  * pages it returns.
1621  *
1622  * The search returns a group of mapping-contiguous page cache entries
1623  * with ascending indexes.  There may be holes in the indices due to
1624  * not-present pages.
1625  *
1626  * Any shadow entries of evicted pages, or swap entries from
1627  * shmem/tmpfs, are included in the returned array.
1628  *
1629  * find_get_entries() returns the number of pages and shadow entries
1630  * which were found.
1631  */
1632 unsigned find_get_entries(struct address_space *mapping,
1633                           pgoff_t start, unsigned int nr_entries,
1634                           struct page **entries, pgoff_t *indices)
1635 {
1636         void **slot;
1637         unsigned int ret = 0;
1638         struct radix_tree_iter iter;
1639
1640         if (!nr_entries)
1641                 return 0;
1642
1643         rcu_read_lock();
1644         radix_tree_for_each_slot(slot, &mapping->i_pages, &iter, start) {
1645                 struct page *head, *page;
1646 repeat:
1647                 page = radix_tree_deref_slot(slot);
1648                 if (unlikely(!page))
1649                         continue;
1650                 if (radix_tree_exception(page)) {
1651                         if (radix_tree_deref_retry(page)) {
1652                                 slot = radix_tree_iter_retry(&iter);
1653                                 continue;
1654                         }
1655                         /*
1656                          * A shadow entry of a recently evicted page, a swap
1657                          * entry from shmem/tmpfs or a DAX entry.  Return it
1658                          * without attempting to raise page count.
1659                          */
1660                         goto export;
1661                 }
1662
1663                 head = compound_head(page);
1664                 if (!page_cache_get_speculative(head))
1665                         goto repeat;
1666
1667                 /* The page was split under us? */
1668                 if (compound_head(page) != head) {
1669                         put_page(head);
1670                         goto repeat;
1671                 }
1672
1673                 /* Has the page moved? */
1674                 if (unlikely(page != *slot)) {
1675                         put_page(head);
1676                         goto repeat;
1677                 }
1678 export:
1679                 indices[ret] = iter.index;
1680                 entries[ret] = page;
1681                 if (++ret == nr_entries)
1682                         break;
1683         }
1684         rcu_read_unlock();
1685         return ret;
1686 }
1687
1688 /**
1689  * find_get_pages_range - gang pagecache lookup
1690  * @mapping:    The address_space to search
1691  * @start:      The starting page index
1692  * @end:        The final page index (inclusive)
1693  * @nr_pages:   The maximum number of pages
1694  * @pages:      Where the resulting pages are placed
1695  *
1696  * find_get_pages_range() will search for and return a group of up to @nr_pages
1697  * pages in the mapping starting at index @start and up to index @end
1698  * (inclusive).  The pages are placed at @pages.  find_get_pages_range() takes
1699  * a reference against the returned pages.
1700  *
1701  * The search returns a group of mapping-contiguous pages with ascending
1702  * indexes.  There may be holes in the indices due to not-present pages.
1703  * We also update @start to index the next page for the traversal.
1704  *
1705  * find_get_pages_range() returns the number of pages which were found. If this
1706  * number is smaller than @nr_pages, the end of specified range has been
1707  * reached.
1708  */
1709 unsigned find_get_pages_range(struct address_space *mapping, pgoff_t *start,
1710                               pgoff_t end, unsigned int nr_pages,
1711                               struct page **pages)
1712 {
1713         struct radix_tree_iter iter;
1714         void **slot;
1715         unsigned ret = 0;
1716
1717         if (unlikely(!nr_pages))
1718                 return 0;
1719
1720         rcu_read_lock();
1721         radix_tree_for_each_slot(slot, &mapping->i_pages, &iter, *start) {
1722                 struct page *head, *page;
1723
1724                 if (iter.index > end)
1725                         break;
1726 repeat:
1727                 page = radix_tree_deref_slot(slot);
1728                 if (unlikely(!page))
1729                         continue;
1730
1731                 if (radix_tree_exception(page)) {
1732                         if (radix_tree_deref_retry(page)) {
1733                                 slot = radix_tree_iter_retry(&iter);
1734                                 continue;
1735                         }
1736                         /*
1737                          * A shadow entry of a recently evicted page,
1738                          * or a swap entry from shmem/tmpfs.  Skip
1739                          * over it.
1740                          */
1741                         continue;
1742                 }
1743
1744                 head = compound_head(page);
1745                 if (!page_cache_get_speculative(head))
1746                         goto repeat;
1747
1748                 /* The page was split under us? */
1749                 if (compound_head(page) != head) {
1750                         put_page(head);
1751                         goto repeat;
1752                 }
1753
1754                 /* Has the page moved? */
1755                 if (unlikely(page != *slot)) {
1756                         put_page(head);
1757                         goto repeat;
1758                 }
1759
1760                 pages[ret] = page;
1761                 if (++ret == nr_pages) {
1762                         *start = pages[ret - 1]->index + 1;
1763                         goto out;
1764                 }
1765         }
1766
1767         /*
1768          * We come here when there is no page beyond @end. We take care to not
1769          * overflow the index @start as it confuses some of the callers. This
1770          * breaks the iteration when there is page at index -1 but that is
1771          * already broken anyway.
1772          */
1773         if (end == (pgoff_t)-1)
1774                 *start = (pgoff_t)-1;
1775         else
1776                 *start = end + 1;
1777 out:
1778         rcu_read_unlock();
1779
1780         return ret;
1781 }
1782
1783 /**
1784  * find_get_pages_contig - gang contiguous pagecache lookup
1785  * @mapping:    The address_space to search
1786  * @index:      The starting page index
1787  * @nr_pages:   The maximum number of pages
1788  * @pages:      Where the resulting pages are placed
1789  *
1790  * find_get_pages_contig() works exactly like find_get_pages(), except
1791  * that the returned number of pages are guaranteed to be contiguous.
1792  *
1793  * find_get_pages_contig() returns the number of pages which were found.
1794  */
1795 unsigned find_get_pages_contig(struct address_space *mapping, pgoff_t index,
1796                                unsigned int nr_pages, struct page **pages)
1797 {
1798         struct radix_tree_iter iter;
1799         void **slot;
1800         unsigned int ret = 0;
1801
1802         if (unlikely(!nr_pages))
1803                 return 0;
1804
1805         rcu_read_lock();
1806         radix_tree_for_each_contig(slot, &mapping->i_pages, &iter, index) {
1807                 struct page *head, *page;
1808 repeat:
1809                 page = radix_tree_deref_slot(slot);
1810                 /* The hole, there no reason to continue */
1811                 if (unlikely(!page))
1812                         break;
1813
1814                 if (radix_tree_exception(page)) {
1815                         if (radix_tree_deref_retry(page)) {
1816                                 slot = radix_tree_iter_retry(&iter);
1817                                 continue;
1818                         }
1819                         /*
1820                          * A shadow entry of a recently evicted page,
1821                          * or a swap entry from shmem/tmpfs.  Stop
1822                          * looking for contiguous pages.
1823                          */
1824                         break;
1825                 }
1826
1827                 head = compound_head(page);
1828                 if (!page_cache_get_speculative(head))
1829                         goto repeat;
1830
1831                 /* The page was split under us? */
1832                 if (compound_head(page) != head) {
1833                         put_page(head);
1834                         goto repeat;
1835                 }
1836
1837                 /* Has the page moved? */
1838                 if (unlikely(page != *slot)) {
1839                         put_page(head);
1840                         goto repeat;
1841                 }
1842
1843                 /*
1844                  * must check mapping and index after taking the ref.
1845                  * otherwise we can get both false positives and false
1846                  * negatives, which is just confusing to the caller.
1847                  */
1848                 if (page->mapping == NULL || page_to_pgoff(page) != iter.index) {
1849                         put_page(page);
1850                         break;
1851                 }
1852
1853                 pages[ret] = page;
1854                 if (++ret == nr_pages)
1855                         break;
1856         }
1857         rcu_read_unlock();
1858         return ret;
1859 }
1860 EXPORT_SYMBOL(find_get_pages_contig);
1861
1862 /**
1863  * find_get_pages_range_tag - find and return pages in given range matching @tag
1864  * @mapping:    the address_space to search
1865  * @index:      the starting page index
1866  * @end:        The final page index (inclusive)
1867  * @tag:        the tag index
1868  * @nr_pages:   the maximum number of pages
1869  * @pages:      where the resulting pages are placed
1870  *
1871  * Like find_get_pages, except we only return pages which are tagged with
1872  * @tag.   We update @index to index the next page for the traversal.
1873  */
1874 unsigned find_get_pages_range_tag(struct address_space *mapping, pgoff_t *index,
1875                         pgoff_t end, int tag, unsigned int nr_pages,
1876                         struct page **pages)
1877 {
1878         struct radix_tree_iter iter;
1879         void **slot;
1880         unsigned ret = 0;
1881
1882         if (unlikely(!nr_pages))
1883                 return 0;
1884
1885         rcu_read_lock();
1886         radix_tree_for_each_tagged(slot, &mapping->i_pages, &iter, *index, tag) {
1887                 struct page *head, *page;
1888
1889                 if (iter.index > end)
1890                         break;
1891 repeat:
1892                 page = radix_tree_deref_slot(slot);
1893                 if (unlikely(!page))
1894                         continue;
1895
1896                 if (radix_tree_exception(page)) {
1897                         if (radix_tree_deref_retry(page)) {
1898                                 slot = radix_tree_iter_retry(&iter);
1899                                 continue;
1900                         }
1901                         /*
1902                          * A shadow entry of a recently evicted page.
1903                          *
1904                          * Those entries should never be tagged, but
1905                          * this tree walk is lockless and the tags are
1906                          * looked up in bulk, one radix tree node at a
1907                          * time, so there is a sizable window for page
1908                          * reclaim to evict a page we saw tagged.
1909                          *
1910                          * Skip over it.
1911                          */
1912                         continue;
1913                 }
1914
1915                 head = compound_head(page);
1916                 if (!page_cache_get_speculative(head))
1917                         goto repeat;
1918
1919                 /* The page was split under us? */
1920                 if (compound_head(page) != head) {
1921                         put_page(head);
1922                         goto repeat;
1923                 }
1924
1925                 /* Has the page moved? */
1926                 if (unlikely(page != *slot)) {
1927                         put_page(head);
1928                         goto repeat;
1929                 }
1930
1931                 pages[ret] = page;
1932                 if (++ret == nr_pages) {
1933                         *index = pages[ret - 1]->index + 1;
1934                         goto out;
1935                 }
1936         }
1937
1938         /*
1939          * We come here when we got at @end. We take care to not overflow the
1940          * index @index as it confuses some of the callers. This breaks the
1941          * iteration when there is page at index -1 but that is already broken
1942          * anyway.
1943          */
1944         if (end == (pgoff_t)-1)
1945                 *index = (pgoff_t)-1;
1946         else
1947                 *index = end + 1;
1948 out:
1949         rcu_read_unlock();
1950
1951         return ret;
1952 }
1953 EXPORT_SYMBOL(find_get_pages_range_tag);
1954
1955 /**
1956  * find_get_entries_tag - find and return entries that match @tag
1957  * @mapping:    the address_space to search
1958  * @start:      the starting page cache index
1959  * @tag:        the tag index
1960  * @nr_entries: the maximum number of entries
1961  * @entries:    where the resulting entries are placed
1962  * @indices:    the cache indices corresponding to the entries in @entries
1963  *
1964  * Like find_get_entries, except we only return entries which are tagged with
1965  * @tag.
1966  */
1967 unsigned find_get_entries_tag(struct address_space *mapping, pgoff_t start,
1968                         int tag, unsigned int nr_entries,
1969                         struct page **entries, pgoff_t *indices)
1970 {
1971         void **slot;
1972         unsigned int ret = 0;
1973         struct radix_tree_iter iter;
1974
1975         if (!nr_entries)
1976                 return 0;
1977
1978         rcu_read_lock();
1979         radix_tree_for_each_tagged(slot, &mapping->i_pages, &iter, start, tag) {
1980                 struct page *head, *page;
1981 repeat:
1982                 page = radix_tree_deref_slot(slot);
1983                 if (unlikely(!page))
1984                         continue;
1985                 if (radix_tree_exception(page)) {
1986                         if (radix_tree_deref_retry(page)) {
1987                                 slot = radix_tree_iter_retry(&iter);
1988                                 continue;
1989                         }
1990
1991                         /*
1992                          * A shadow entry of a recently evicted page, a swap
1993                          * entry from shmem/tmpfs or a DAX entry.  Return it
1994                          * without attempting to raise page count.
1995                          */
1996                         goto export;
1997                 }
1998
1999                 head = compound_head(page);
2000                 if (!page_cache_get_speculative(head))
2001                         goto repeat;
2002
2003                 /* The page was split under us? */
2004                 if (compound_head(page) != head) {
2005                         put_page(head);
2006                         goto repeat;
2007                 }
2008
2009                 /* Has the page moved? */
2010                 if (unlikely(page != *slot)) {
2011                         put_page(head);
2012                         goto repeat;
2013                 }
2014 export:
2015                 indices[ret] = iter.index;
2016                 entries[ret] = page;
2017                 if (++ret == nr_entries)
2018                         break;
2019         }
2020         rcu_read_unlock();
2021         return ret;
2022 }
2023 EXPORT_SYMBOL(find_get_entries_tag);
2024
2025 /*
2026  * CD/DVDs are error prone. When a medium error occurs, the driver may fail
2027  * a _large_ part of the i/o request. Imagine the worst scenario:
2028  *
2029  *      ---R__________________________________________B__________
2030  *         ^ reading here                             ^ bad block(assume 4k)
2031  *
2032  * read(R) => miss => readahead(R...B) => media error => frustrating retries
2033  * => failing the whole request => read(R) => read(R+1) =>
2034  * readahead(R+1...B+1) => bang => read(R+2) => read(R+3) =>
2035  * readahead(R+3...B+2) => bang => read(R+3) => read(R+4) =>
2036  * readahead(R+4...B+3) => bang => read(R+4) => read(R+5) => ......
2037  *
2038  * It is going insane. Fix it by quickly scaling down the readahead size.
2039  */
2040 static void shrink_readahead_size_eio(struct file *filp,
2041                                         struct file_ra_state *ra)
2042 {
2043         ra->ra_pages /= 4;
2044 }
2045
2046 /**
2047  * generic_file_buffered_read - generic file read routine
2048  * @iocb:       the iocb to read
2049  * @iter:       data destination
2050  * @written:    already copied
2051  *
2052  * This is a generic file read routine, and uses the
2053  * mapping->a_ops->readpage() function for the actual low-level stuff.
2054  *
2055  * This is really ugly. But the goto's actually try to clarify some
2056  * of the logic when it comes to error handling etc.
2057  */
2058 static ssize_t generic_file_buffered_read(struct kiocb *iocb,
2059                 struct iov_iter *iter, ssize_t written)
2060 {
2061         struct file *filp = iocb->ki_filp;
2062         struct address_space *mapping = filp->f_mapping;
2063         struct inode *inode = mapping->host;
2064         struct file_ra_state *ra = &filp->f_ra;
2065         loff_t *ppos = &iocb->ki_pos;
2066         pgoff_t index;
2067         pgoff_t last_index;
2068         pgoff_t prev_index;
2069         unsigned long offset;      /* offset into pagecache page */
2070         unsigned int prev_offset;
2071         int error = 0;
2072
2073         if (unlikely(*ppos >= inode->i_sb->s_maxbytes))
2074                 return 0;
2075         iov_iter_truncate(iter, inode->i_sb->s_maxbytes);
2076
2077         index = *ppos >> PAGE_SHIFT;
2078         prev_index = ra->prev_pos >> PAGE_SHIFT;
2079         prev_offset = ra->prev_pos & (PAGE_SIZE-1);
2080         last_index = (*ppos + iter->count + PAGE_SIZE-1) >> PAGE_SHIFT;
2081         offset = *ppos & ~PAGE_MASK;
2082
2083         for (;;) {
2084                 struct page *page;
2085                 pgoff_t end_index;
2086                 loff_t isize;
2087                 unsigned long nr, ret;
2088
2089                 cond_resched();
2090 find_page:
2091                 if (fatal_signal_pending(current)) {
2092                         error = -EINTR;
2093                         goto out;
2094                 }
2095
2096                 page = find_get_page(mapping, index);
2097                 if (!page) {
2098                         if (iocb->ki_flags & IOCB_NOWAIT)
2099                                 goto would_block;
2100                         page_cache_sync_readahead(mapping,
2101                                         ra, filp,
2102                                         index, last_index - index);
2103                         page = find_get_page(mapping, index);
2104                         if (unlikely(page == NULL))
2105                                 goto no_cached_page;
2106                 }
2107                 if (PageReadahead(page)) {
2108                         page_cache_async_readahead(mapping,
2109                                         ra, filp, page,
2110                                         index, last_index - index);
2111                 }
2112                 if (!PageUptodate(page)) {
2113                         if (iocb->ki_flags & IOCB_NOWAIT) {
2114                                 put_page(page);
2115                                 goto would_block;
2116                         }
2117
2118                         /*
2119                          * See comment in do_read_cache_page on why
2120                          * wait_on_page_locked is used to avoid unnecessarily
2121                          * serialisations and why it's safe.
2122                          */
2123                         error = wait_on_page_locked_killable(page);
2124                         if (unlikely(error))
2125                                 goto readpage_error;
2126                         if (PageUptodate(page))
2127                                 goto page_ok;
2128
2129                         if (inode->i_blkbits == PAGE_SHIFT ||
2130                                         !mapping->a_ops->is_partially_uptodate)
2131                                 goto page_not_up_to_date;
2132                         /* pipes can't handle partially uptodate pages */
2133                         if (unlikely(iter->type & ITER_PIPE))
2134                                 goto page_not_up_to_date;
2135                         if (!trylock_page(page))
2136                                 goto page_not_up_to_date;
2137                         /* Did it get truncated before we got the lock? */
2138                         if (!page->mapping)
2139                                 goto page_not_up_to_date_locked;
2140                         if (!mapping->a_ops->is_partially_uptodate(page,
2141                                                         offset, iter->count))
2142                                 goto page_not_up_to_date_locked;
2143                         unlock_page(page);
2144                 }
2145 page_ok:
2146                 /*
2147                  * i_size must be checked after we know the page is Uptodate.
2148                  *
2149                  * Checking i_size after the check allows us to calculate
2150                  * the correct value for "nr", which means the zero-filled
2151                  * part of the page is not copied back to userspace (unless
2152                  * another truncate extends the file - this is desired though).
2153                  */
2154
2155                 isize = i_size_read(inode);
2156                 end_index = (isize - 1) >> PAGE_SHIFT;
2157                 if (unlikely(!isize || index > end_index)) {
2158                         put_page(page);
2159                         goto out;
2160                 }
2161
2162                 /* nr is the maximum number of bytes to copy from this page */
2163                 nr = PAGE_SIZE;
2164                 if (index == end_index) {
2165                         nr = ((isize - 1) & ~PAGE_MASK) + 1;
2166                         if (nr <= offset) {
2167                                 put_page(page);
2168                                 goto out;
2169                         }
2170                 }
2171                 nr = nr - offset;
2172
2173                 /* If users can be writing to this page using arbitrary
2174                  * virtual addresses, take care about potential aliasing
2175                  * before reading the page on the kernel side.
2176                  */
2177                 if (mapping_writably_mapped(mapping))
2178                         flush_dcache_page(page);
2179
2180                 /*
2181                  * When a sequential read accesses a page several times,
2182                  * only mark it as accessed the first time.
2183                  */
2184                 if (prev_index != index || offset != prev_offset)
2185                         mark_page_accessed(page);
2186                 prev_index = index;
2187
2188                 /*
2189                  * Ok, we have the page, and it's up-to-date, so
2190                  * now we can copy it to user space...
2191                  */
2192
2193                 ret = copy_page_to_iter(page, offset, nr, iter);
2194                 offset += ret;
2195                 index += offset >> PAGE_SHIFT;
2196                 offset &= ~PAGE_MASK;
2197                 prev_offset = offset;
2198
2199                 put_page(page);
2200                 written += ret;
2201                 if (!iov_iter_count(iter))
2202                         goto out;
2203                 if (ret < nr) {
2204                         error = -EFAULT;
2205                         goto out;
2206                 }
2207                 continue;
2208
2209 page_not_up_to_date:
2210                 /* Get exclusive access to the page ... */
2211                 error = lock_page_killable(page);
2212                 if (unlikely(error))
2213                         goto readpage_error;
2214
2215 page_not_up_to_date_locked:
2216                 /* Did it get truncated before we got the lock? */
2217                 if (!page->mapping) {
2218                         unlock_page(page);
2219                         put_page(page);
2220                         continue;
2221                 }
2222
2223                 /* Did somebody else fill it already? */
2224                 if (PageUptodate(page)) {
2225                         unlock_page(page);
2226                         goto page_ok;
2227                 }
2228
2229 readpage:
2230                 /*
2231                  * A previous I/O error may have been due to temporary
2232                  * failures, eg. multipath errors.
2233                  * PG_error will be set again if readpage fails.
2234                  */
2235                 ClearPageError(page);
2236                 /* Start the actual read. The read will unlock the page. */
2237                 error = mapping->a_ops->readpage(filp, page);
2238
2239                 if (unlikely(error)) {
2240                         if (error == AOP_TRUNCATED_PAGE) {
2241                                 put_page(page);
2242                                 error = 0;
2243                                 goto find_page;
2244                         }
2245                         goto readpage_error;
2246                 }
2247
2248                 if (!PageUptodate(page)) {
2249                         error = lock_page_killable(page);
2250                         if (unlikely(error))
2251                                 goto readpage_error;
2252                         if (!PageUptodate(page)) {
2253                                 if (page->mapping == NULL) {
2254                                         /*
2255                                          * invalidate_mapping_pages got it
2256                                          */
2257                                         unlock_page(page);
2258                                         put_page(page);
2259                                         goto find_page;
2260                                 }
2261                                 unlock_page(page);
2262                                 shrink_readahead_size_eio(filp, ra);
2263                                 error = -EIO;
2264                                 goto readpage_error;
2265                         }
2266                         unlock_page(page);
2267                 }
2268
2269                 goto page_ok;
2270
2271 readpage_error:
2272                 /* UHHUH! A synchronous read error occurred. Report it */
2273                 put_page(page);
2274                 goto out;
2275
2276 no_cached_page:
2277                 /*
2278                  * Ok, it wasn't cached, so we need to create a new
2279                  * page..
2280                  */
2281                 page = page_cache_alloc(mapping);
2282                 if (!page) {
2283                         error = -ENOMEM;
2284                         goto out;
2285                 }
2286                 error = add_to_page_cache_lru(page, mapping, index,
2287                                 mapping_gfp_constraint(mapping, GFP_KERNEL));
2288                 if (error) {
2289                         put_page(page);
2290                         if (error == -EEXIST) {
2291                                 error = 0;
2292                                 goto find_page;
2293                         }
2294                         goto out;
2295                 }
2296                 goto readpage;
2297         }
2298
2299 would_block:
2300         error = -EAGAIN;
2301 out:
2302         ra->prev_pos = prev_index;
2303         ra->prev_pos <<= PAGE_SHIFT;
2304         ra->prev_pos |= prev_offset;
2305
2306         *ppos = ((loff_t)index << PAGE_SHIFT) + offset;
2307         file_accessed(filp);
2308         return written ? written : error;
2309 }
2310
2311 /**
2312  * generic_file_read_iter - generic filesystem read routine
2313  * @iocb:       kernel I/O control block
2314  * @iter:       destination for the data read
2315  *
2316  * This is the "read_iter()" routine for all filesystems
2317  * that can use the page cache directly.
2318  */
2319 ssize_t
2320 generic_file_read_iter(struct kiocb *iocb, struct iov_iter *iter)
2321 {
2322         size_t count = iov_iter_count(iter);
2323         ssize_t retval = 0;
2324
2325         if (!count)
2326                 goto out; /* skip atime */
2327
2328         if (iocb->ki_flags & IOCB_DIRECT) {
2329                 struct file *file = iocb->ki_filp;
2330                 struct address_space *mapping = file->f_mapping;
2331                 struct inode *inode = mapping->host;
2332                 loff_t size;
2333
2334                 size = i_size_read(inode);
2335                 if (iocb->ki_flags & IOCB_NOWAIT) {
2336                         if (filemap_range_has_page(mapping, iocb->ki_pos,
2337                                                    iocb->ki_pos + count - 1))
2338                                 return -EAGAIN;
2339                 } else {
2340                         retval = filemap_write_and_wait_range(mapping,
2341                                                 iocb->ki_pos,
2342                                                 iocb->ki_pos + count - 1);
2343                         if (retval < 0)
2344                                 goto out;
2345                 }
2346
2347                 file_accessed(file);
2348
2349                 retval = mapping->a_ops->direct_IO(iocb, iter);
2350                 if (retval >= 0) {
2351                         iocb->ki_pos += retval;
2352                         count -= retval;
2353                 }
2354                 iov_iter_revert(iter, count - iov_iter_count(iter));
2355
2356                 /*
2357                  * Btrfs can have a short DIO read if we encounter
2358                  * compressed extents, so if there was an error, or if
2359                  * we've already read everything we wanted to, or if
2360                  * there was a short read because we hit EOF, go ahead
2361                  * and return.  Otherwise fallthrough to buffered io for
2362                  * the rest of the read.  Buffered reads will not work for
2363                  * DAX files, so don't bother trying.
2364                  */
2365                 if (retval < 0 || !count || iocb->ki_pos >= size ||
2366                     IS_DAX(inode))
2367                         goto out;
2368         }
2369
2370         retval = generic_file_buffered_read(iocb, iter, retval);
2371 out:
2372         return retval;
2373 }
2374 EXPORT_SYMBOL(generic_file_read_iter);
2375
2376 #ifdef CONFIG_MMU
2377 /**
2378  * page_cache_read - adds requested page to the page cache if not already there
2379  * @file:       file to read
2380  * @offset:     page index
2381  * @gfp_mask:   memory allocation flags
2382  *
2383  * This adds the requested page to the page cache if it isn't already there,
2384  * and schedules an I/O to read in its contents from disk.
2385  */
2386 static int page_cache_read(struct file *file, pgoff_t offset, gfp_t gfp_mask)
2387 {
2388         struct address_space *mapping = file->f_mapping;
2389         struct page *page;
2390         int ret;
2391
2392         do {
2393                 page = __page_cache_alloc(gfp_mask);
2394                 if (!page)
2395                         return -ENOMEM;
2396
2397                 ret = add_to_page_cache_lru(page, mapping, offset, gfp_mask);
2398                 if (ret == 0)
2399                         ret = mapping->a_ops->readpage(file, page);
2400                 else if (ret == -EEXIST)
2401                         ret = 0; /* losing race to add is OK */
2402
2403                 put_page(page);
2404
2405         } while (ret == AOP_TRUNCATED_PAGE);
2406
2407         return ret;
2408 }
2409
2410 #define MMAP_LOTSAMISS  (100)
2411
2412 /*
2413  * Synchronous readahead happens when we don't even find
2414  * a page in the page cache at all.
2415  */
2416 static void do_sync_mmap_readahead(struct vm_area_struct *vma,
2417                                    struct file_ra_state *ra,
2418                                    struct file *file,
2419                                    pgoff_t offset)
2420 {
2421         struct address_space *mapping = file->f_mapping;
2422
2423         /* If we don't want any read-ahead, don't bother */
2424         if (vma->vm_flags & VM_RAND_READ)
2425                 return;
2426         if (!ra->ra_pages)
2427                 return;
2428
2429         if (vma->vm_flags & VM_SEQ_READ) {
2430                 page_cache_sync_readahead(mapping, ra, file, offset,
2431                                           ra->ra_pages);
2432                 return;
2433         }
2434
2435         /* Avoid banging the cache line if not needed */
2436         if (ra->mmap_miss < MMAP_LOTSAMISS * 10)
2437                 ra->mmap_miss++;
2438
2439         /*
2440          * Do we miss much more than hit in this file? If so,
2441          * stop bothering with read-ahead. It will only hurt.
2442          */
2443         if (ra->mmap_miss > MMAP_LOTSAMISS)
2444                 return;
2445
2446         /*
2447          * mmap read-around
2448          */
2449         ra->start = max_t(long, 0, offset - ra->ra_pages / 2);
2450         ra->size = ra->ra_pages;
2451         ra->async_size = ra->ra_pages / 4;
2452         ra_submit(ra, mapping, file);
2453 }
2454
2455 /*
2456  * Asynchronous readahead happens when we find the page and PG_readahead,
2457  * so we want to possibly extend the readahead further..
2458  */
2459 static void do_async_mmap_readahead(struct vm_area_struct *vma,
2460                                     struct file_ra_state *ra,
2461                                     struct file *file,
2462                                     struct page *page,
2463                                     pgoff_t offset)
2464 {
2465         struct address_space *mapping = file->f_mapping;
2466
2467         /* If we don't want any read-ahead, don't bother */
2468         if (vma->vm_flags & VM_RAND_READ)
2469                 return;
2470         if (ra->mmap_miss > 0)
2471                 ra->mmap_miss--;
2472         if (PageReadahead(page))
2473                 page_cache_async_readahead(mapping, ra, file,
2474                                            page, offset, ra->ra_pages);
2475 }
2476
2477 /**
2478  * filemap_fault - read in file data for page fault handling
2479  * @vmf:        struct vm_fault containing details of the fault
2480  *
2481  * filemap_fault() is invoked via the vma operations vector for a
2482  * mapped memory region to read in file data during a page fault.
2483  *
2484  * The goto's are kind of ugly, but this streamlines the normal case of having
2485  * it in the page cache, and handles the special cases reasonably without
2486  * having a lot of duplicated code.
2487  *
2488  * vma->vm_mm->mmap_sem must be held on entry.
2489  *
2490  * If our return value has VM_FAULT_RETRY set, it's because
2491  * lock_page_or_retry() returned 0.
2492  * The mmap_sem has usually been released in this case.
2493  * See __lock_page_or_retry() for the exception.
2494  *
2495  * If our return value does not have VM_FAULT_RETRY set, the mmap_sem
2496  * has not been released.
2497  *
2498  * We never return with VM_FAULT_RETRY and a bit from VM_FAULT_ERROR set.
2499  */
2500 vm_fault_t filemap_fault(struct vm_fault *vmf)
2501 {
2502         int error;
2503         struct file *file = vmf->vma->vm_file;
2504         struct address_space *mapping = file->f_mapping;
2505         struct file_ra_state *ra = &file->f_ra;
2506         struct inode *inode = mapping->host;
2507         pgoff_t offset = vmf->pgoff;
2508         pgoff_t max_off;
2509         struct page *page;
2510         vm_fault_t ret = 0;
2511
2512         max_off = DIV_ROUND_UP(i_size_read(inode), PAGE_SIZE);
2513         if (unlikely(offset >= max_off))
2514                 return VM_FAULT_SIGBUS;
2515
2516         /*
2517          * Do we have something in the page cache already?
2518          */
2519         page = find_get_page(mapping, offset);
2520         if (likely(page) && !(vmf->flags & FAULT_FLAG_TRIED)) {
2521                 /*
2522                  * We found the page, so try async readahead before
2523                  * waiting for the lock.
2524                  */
2525                 do_async_mmap_readahead(vmf->vma, ra, file, page, offset);
2526         } else if (!page) {
2527                 /* No page in the page cache at all */
2528                 do_sync_mmap_readahead(vmf->vma, ra, file, offset);
2529                 count_vm_event(PGMAJFAULT);
2530                 count_memcg_event_mm(vmf->vma->vm_mm, PGMAJFAULT);
2531                 ret = VM_FAULT_MAJOR;
2532 retry_find:
2533                 page = find_get_page(mapping, offset);
2534                 if (!page)
2535                         goto no_cached_page;
2536         }
2537
2538         if (!lock_page_or_retry(page, vmf->vma->vm_mm, vmf->flags)) {
2539                 put_page(page);
2540                 return ret | VM_FAULT_RETRY;
2541         }
2542
2543         /* Did it get truncated? */
2544         if (unlikely(page->mapping != mapping)) {
2545                 unlock_page(page);
2546                 put_page(page);
2547                 goto retry_find;
2548         }
2549         VM_BUG_ON_PAGE(page->index != offset, page);
2550
2551         /*
2552          * We have a locked page in the page cache, now we need to check
2553          * that it's up-to-date. If not, it is going to be due to an error.
2554          */
2555         if (unlikely(!PageUptodate(page)))
2556                 goto page_not_uptodate;
2557
2558         /*
2559          * Found the page and have a reference on it.
2560          * We must recheck i_size under page lock.
2561          */
2562         max_off = DIV_ROUND_UP(i_size_read(inode), PAGE_SIZE);
2563         if (unlikely(offset >= max_off)) {
2564                 unlock_page(page);
2565                 put_page(page);
2566                 return VM_FAULT_SIGBUS;
2567         }
2568
2569         vmf->page = page;
2570         return ret | VM_FAULT_LOCKED;
2571
2572 no_cached_page:
2573         /*
2574          * We're only likely to ever get here if MADV_RANDOM is in
2575          * effect.
2576          */
2577         error = page_cache_read(file, offset, vmf->gfp_mask);
2578
2579         /*
2580          * The page we want has now been added to the page cache.
2581          * In the unlikely event that someone removed it in the
2582          * meantime, we'll just come back here and read it again.
2583          */
2584         if (error >= 0)
2585                 goto retry_find;
2586
2587         /*
2588          * An error return from page_cache_read can result if the
2589          * system is low on memory, or a problem occurs while trying
2590          * to schedule I/O.
2591          */
2592         if (error == -ENOMEM)
2593                 return VM_FAULT_OOM;
2594         return VM_FAULT_SIGBUS;
2595
2596 page_not_uptodate:
2597         /*
2598          * Umm, take care of errors if the page isn't up-to-date.
2599          * Try to re-read it _once_. We do this synchronously,
2600          * because there really aren't any performance issues here
2601          * and we need to check for errors.
2602          */
2603         ClearPageError(page);
2604         error = mapping->a_ops->readpage(file, page);
2605         if (!error) {
2606                 wait_on_page_locked(page);
2607                 if (!PageUptodate(page))
2608                         error = -EIO;
2609         }
2610         put_page(page);
2611
2612         if (!error || error == AOP_TRUNCATED_PAGE)
2613                 goto retry_find;
2614
2615         /* Things didn't work out. Return zero to tell the mm layer so. */
2616         shrink_readahead_size_eio(file, ra);
2617         return VM_FAULT_SIGBUS;
2618 }
2619 EXPORT_SYMBOL(filemap_fault);
2620
2621 void filemap_map_pages(struct vm_fault *vmf,
2622                 pgoff_t start_pgoff, pgoff_t end_pgoff)
2623 {
2624         struct radix_tree_iter iter;
2625         void **slot;
2626         struct file *file = vmf->vma->vm_file;
2627         struct address_space *mapping = file->f_mapping;
2628         pgoff_t last_pgoff = start_pgoff;
2629         unsigned long max_idx;
2630         struct page *head, *page;
2631
2632         rcu_read_lock();
2633         radix_tree_for_each_slot(slot, &mapping->i_pages, &iter, start_pgoff) {
2634                 if (iter.index > end_pgoff)
2635                         break;
2636 repeat:
2637                 page = radix_tree_deref_slot(slot);
2638                 if (unlikely(!page))
2639                         goto next;
2640                 if (radix_tree_exception(page)) {
2641                         if (radix_tree_deref_retry(page)) {
2642                                 slot = radix_tree_iter_retry(&iter);
2643                                 continue;
2644                         }
2645                         goto next;
2646                 }
2647
2648                 head = compound_head(page);
2649                 if (!page_cache_get_speculative(head))
2650                         goto repeat;
2651
2652                 /* The page was split under us? */
2653                 if (compound_head(page) != head) {
2654                         put_page(head);
2655                         goto repeat;
2656                 }
2657
2658                 /* Has the page moved? */
2659                 if (unlikely(page != *slot)) {
2660                         put_page(head);
2661                         goto repeat;
2662                 }
2663
2664                 if (!PageUptodate(page) ||
2665                                 PageReadahead(page) ||
2666                                 PageHWPoison(page))
2667                         goto skip;
2668                 if (!trylock_page(page))
2669                         goto skip;
2670
2671                 if (page->mapping != mapping || !PageUptodate(page))
2672                         goto unlock;
2673
2674                 max_idx = DIV_ROUND_UP(i_size_read(mapping->host), PAGE_SIZE);
2675                 if (page->index >= max_idx)
2676                         goto unlock;
2677
2678                 if (file->f_ra.mmap_miss > 0)
2679                         file->f_ra.mmap_miss--;
2680
2681                 vmf->address += (iter.index - last_pgoff) << PAGE_SHIFT;
2682                 if (vmf->pte)
2683                         vmf->pte += iter.index - last_pgoff;
2684                 last_pgoff = iter.index;
2685                 if (alloc_set_pte(vmf, NULL, page))
2686                         goto unlock;
2687                 unlock_page(page);
2688                 goto next;
2689 unlock:
2690                 unlock_page(page);
2691 skip:
2692                 put_page(page);
2693 next:
2694                 /* Huge page is mapped? No need to proceed. */
2695                 if (pmd_trans_huge(*vmf->pmd))
2696                         break;
2697                 if (iter.index == end_pgoff)
2698                         break;
2699         }
2700         rcu_read_unlock();
2701 }
2702 EXPORT_SYMBOL(filemap_map_pages);
2703
2704 vm_fault_t filemap_page_mkwrite(struct vm_fault *vmf)
2705 {
2706         struct page *page = vmf->page;
2707         struct inode *inode = file_inode(vmf->vma->vm_file);
2708         vm_fault_t ret = VM_FAULT_LOCKED;
2709
2710         sb_start_pagefault(inode->i_sb);
2711         file_update_time(vmf->vma->vm_file);
2712         lock_page(page);
2713         if (page->mapping != inode->i_mapping) {
2714                 unlock_page(page);
2715                 ret = VM_FAULT_NOPAGE;
2716                 goto out;
2717         }
2718         /*
2719          * We mark the page dirty already here so that when freeze is in
2720          * progress, we are guaranteed that writeback during freezing will
2721          * see the dirty page and writeprotect it again.
2722          */
2723         set_page_dirty(page);
2724         wait_for_stable_page(page);
2725 out:
2726         sb_end_pagefault(inode->i_sb);
2727         return ret;
2728 }
2729
2730 const struct vm_operations_struct generic_file_vm_ops = {
2731         .fault          = filemap_fault,
2732         .map_pages      = filemap_map_pages,
2733         .page_mkwrite   = filemap_page_mkwrite,
2734 };
2735
2736 /* This is used for a general mmap of a disk file */
2737
2738 int generic_file_mmap(struct file * file, struct vm_area_struct * vma)
2739 {
2740         struct address_space *mapping = file->f_mapping;
2741
2742         if (!mapping->a_ops->readpage)
2743                 return -ENOEXEC;
2744         file_accessed(file);
2745         vma->vm_ops = &generic_file_vm_ops;
2746         return 0;
2747 }
2748
2749 /*
2750  * This is for filesystems which do not implement ->writepage.
2751  */
2752 int generic_file_readonly_mmap(struct file *file, struct vm_area_struct *vma)
2753 {
2754         if ((vma->vm_flags & VM_SHARED) && (vma->vm_flags & VM_MAYWRITE))
2755                 return -EINVAL;
2756         return generic_file_mmap(file, vma);
2757 }
2758 #else
2759 vm_fault_t filemap_page_mkwrite(struct vm_fault *vmf)
2760 {
2761         return VM_FAULT_SIGBUS;
2762 }
2763 int generic_file_mmap(struct file * file, struct vm_area_struct * vma)
2764 {
2765         return -ENOSYS;
2766 }
2767 int generic_file_readonly_mmap(struct file * file, struct vm_area_struct * vma)
2768 {
2769         return -ENOSYS;
2770 }
2771 #endif /* CONFIG_MMU */
2772
2773 EXPORT_SYMBOL(filemap_page_mkwrite);
2774 EXPORT_SYMBOL(generic_file_mmap);
2775 EXPORT_SYMBOL(generic_file_readonly_mmap);
2776
2777 static struct page *wait_on_page_read(struct page *page)
2778 {
2779         if (!IS_ERR(page)) {
2780                 wait_on_page_locked(page);
2781                 if (!PageUptodate(page)) {
2782                         put_page(page);
2783                         page = ERR_PTR(-EIO);
2784                 }
2785         }
2786         return page;
2787 }
2788
2789 static struct page *do_read_cache_page(struct address_space *mapping,
2790                                 pgoff_t index,
2791                                 int (*filler)(void *, struct page *),
2792                                 void *data,
2793                                 gfp_t gfp)
2794 {
2795         struct page *page;
2796         int err;
2797 repeat:
2798         page = find_get_page(mapping, index);
2799         if (!page) {
2800                 page = __page_cache_alloc(gfp);
2801                 if (!page)
2802                         return ERR_PTR(-ENOMEM);
2803                 err = add_to_page_cache_lru(page, mapping, index, gfp);
2804                 if (unlikely(err)) {
2805                         put_page(page);
2806                         if (err == -EEXIST)
2807                                 goto repeat;
2808                         /* Presumably ENOMEM for radix tree node */
2809                         return ERR_PTR(err);
2810                 }
2811
2812 filler:
2813                 err = filler(data, page);
2814                 if (err < 0) {
2815                         put_page(page);
2816                         return ERR_PTR(err);
2817                 }
2818
2819                 page = wait_on_page_read(page);
2820                 if (IS_ERR(page))
2821                         return page;
2822                 goto out;
2823         }
2824         if (PageUptodate(page))
2825                 goto out;
2826
2827         /*
2828          * Page is not up to date and may be locked due one of the following
2829          * case a: Page is being filled and the page lock is held
2830          * case b: Read/write error clearing the page uptodate status
2831          * case c: Truncation in progress (page locked)
2832          * case d: Reclaim in progress
2833          *
2834          * Case a, the page will be up to date when the page is unlocked.
2835          *    There is no need to serialise on the page lock here as the page
2836          *    is pinned so the lock gives no additional protection. Even if the
2837          *    the page is truncated, the data is still valid if PageUptodate as
2838          *    it's a race vs truncate race.
2839          * Case b, the page will not be up to date
2840          * Case c, the page may be truncated but in itself, the data may still
2841          *    be valid after IO completes as it's a read vs truncate race. The
2842          *    operation must restart if the page is not uptodate on unlock but
2843          *    otherwise serialising on page lock to stabilise the mapping gives
2844          *    no additional guarantees to the caller as the page lock is
2845          *    released before return.
2846          * Case d, similar to truncation. If reclaim holds the page lock, it
2847          *    will be a race with remove_mapping that determines if the mapping
2848          *    is valid on unlock but otherwise the data is valid and there is
2849          *    no need to serialise with page lock.
2850          *
2851          * As the page lock gives no additional guarantee, we optimistically
2852          * wait on the page to be unlocked and check if it's up to date and
2853          * use the page if it is. Otherwise, the page lock is required to
2854          * distinguish between the different cases. The motivation is that we
2855          * avoid spurious serialisations and wakeups when multiple processes
2856          * wait on the same page for IO to complete.
2857          */
2858         wait_on_page_locked(page);
2859         if (PageUptodate(page))
2860                 goto out;
2861
2862         /* Distinguish between all the cases under the safety of the lock */
2863         lock_page(page);
2864
2865         /* Case c or d, restart the operation */
2866         if (!page->mapping) {
2867                 unlock_page(page);
2868                 put_page(page);
2869                 goto repeat;
2870         }
2871
2872         /* Someone else locked and filled the page in a very small window */
2873         if (PageUptodate(page)) {
2874                 unlock_page(page);
2875                 goto out;
2876         }
2877         goto filler;
2878
2879 out:
2880         mark_page_accessed(page);
2881         return page;
2882 }
2883
2884 /**
2885  * read_cache_page - read into page cache, fill it if needed
2886  * @mapping:    the page's address_space
2887  * @index:      the page index
2888  * @filler:     function to perform the read
2889  * @data:       first arg to filler(data, page) function, often left as NULL
2890  *
2891  * Read into the page cache. If a page already exists, and PageUptodate() is
2892  * not set, try to fill the page and wait for it to become unlocked.
2893  *
2894  * If the page does not get brought uptodate, return -EIO.
2895  */
2896 struct page *read_cache_page(struct address_space *mapping,
2897                                 pgoff_t index,
2898                                 int (*filler)(void *, struct page *),
2899                                 void *data)
2900 {
2901         return do_read_cache_page(mapping, index, filler, data, mapping_gfp_mask(mapping));
2902 }
2903 EXPORT_SYMBOL(read_cache_page);
2904
2905 /**
2906  * read_cache_page_gfp - read into page cache, using specified page allocation flags.
2907  * @mapping:    the page's address_space
2908  * @index:      the page index
2909  * @gfp:        the page allocator flags to use if allocating
2910  *
2911  * This is the same as "read_mapping_page(mapping, index, NULL)", but with
2912  * any new page allocations done using the specified allocation flags.
2913  *
2914  * If the page does not get brought uptodate, return -EIO.
2915  */
2916 struct page *read_cache_page_gfp(struct address_space *mapping,
2917                                 pgoff_t index,
2918                                 gfp_t gfp)
2919 {
2920         filler_t *filler = (filler_t *)mapping->a_ops->readpage;
2921
2922         return do_read_cache_page(mapping, index, filler, NULL, gfp);
2923 }
2924 EXPORT_SYMBOL(read_cache_page_gfp);
2925
2926 /*
2927  * Performs necessary checks before doing a write
2928  *
2929  * Can adjust writing position or amount of bytes to write.
2930  * Returns appropriate error code that caller should return or
2931  * zero in case that write should be allowed.
2932  */
2933 inline ssize_t generic_write_checks(struct kiocb *iocb, struct iov_iter *from)
2934 {
2935         struct file *file = iocb->ki_filp;
2936         struct inode *inode = file->f_mapping->host;
2937         unsigned long limit = rlimit(RLIMIT_FSIZE);
2938         loff_t pos;
2939
2940         if (!iov_iter_count(from))
2941                 return 0;
2942
2943         /* FIXME: this is for backwards compatibility with 2.4 */
2944         if (iocb->ki_flags & IOCB_APPEND)
2945                 iocb->ki_pos = i_size_read(inode);
2946
2947         pos = iocb->ki_pos;
2948
2949         if ((iocb->ki_flags & IOCB_NOWAIT) && !(iocb->ki_flags & IOCB_DIRECT))
2950                 return -EINVAL;
2951
2952         if (limit != RLIM_INFINITY) {
2953                 if (iocb->ki_pos >= limit) {
2954                         send_sig(SIGXFSZ, current, 0);
2955                         return -EFBIG;
2956                 }
2957                 iov_iter_truncate(from, limit - (unsigned long)pos);
2958         }
2959
2960         /*
2961          * LFS rule
2962          */
2963         if (unlikely(pos + iov_iter_count(from) > MAX_NON_LFS &&
2964                                 !(file->f_flags & O_LARGEFILE))) {
2965                 if (pos >= MAX_NON_LFS)
2966                         return -EFBIG;
2967                 iov_iter_truncate(from, MAX_NON_LFS - (unsigned long)pos);
2968         }
2969
2970         /*
2971          * Are we about to exceed the fs block limit ?
2972          *
2973          * If we have written data it becomes a short write.  If we have
2974          * exceeded without writing data we send a signal and return EFBIG.
2975          * Linus frestrict idea will clean these up nicely..
2976          */
2977         if (unlikely(pos >= inode->i_sb->s_maxbytes))
2978                 return -EFBIG;
2979
2980         iov_iter_truncate(from, inode->i_sb->s_maxbytes - pos);
2981         return iov_iter_count(from);
2982 }
2983 EXPORT_SYMBOL(generic_write_checks);
2984
2985 int pagecache_write_begin(struct file *file, struct address_space *mapping,
2986                                 loff_t pos, unsigned len, unsigned flags,
2987                                 struct page **pagep, void **fsdata)
2988 {
2989         const struct address_space_operations *aops = mapping->a_ops;
2990
2991         return aops->write_begin(file, mapping, pos, len, flags,
2992                                                         pagep, fsdata);
2993 }
2994 EXPORT_SYMBOL(pagecache_write_begin);
2995
2996 int pagecache_write_end(struct file *file, struct address_space *mapping,
2997                                 loff_t pos, unsigned len, unsigned copied,
2998                                 struct page *page, void *fsdata)
2999 {
3000         const struct address_space_operations *aops = mapping->a_ops;
3001
3002         return aops->write_end(file, mapping, pos, len, copied, page, fsdata);
3003 }
3004 EXPORT_SYMBOL(pagecache_write_end);
3005
3006 ssize_t
3007 generic_file_direct_write(struct kiocb *iocb, struct iov_iter *from)
3008 {
3009         struct file     *file = iocb->ki_filp;
3010         struct address_space *mapping = file->f_mapping;
3011         struct inode    *inode = mapping->host;
3012         loff_t          pos = iocb->ki_pos;
3013         ssize_t         written;
3014         size_t          write_len;
3015         pgoff_t         end;
3016
3017         write_len = iov_iter_count(from);
3018         end = (pos + write_len - 1) >> PAGE_SHIFT;
3019
3020         if (iocb->ki_flags & IOCB_NOWAIT) {
3021                 /* If there are pages to writeback, return */
3022                 if (filemap_range_has_page(inode->i_mapping, pos,
3023                                            pos + iov_iter_count(from)))
3024                         return -EAGAIN;
3025         } else {
3026                 written = filemap_write_and_wait_range(mapping, pos,
3027                                                         pos + write_len - 1);
3028                 if (written)
3029                         goto out;
3030         }
3031
3032         /*
3033          * After a write we want buffered reads to be sure to go to disk to get
3034          * the new data.  We invalidate clean cached page from the region we're
3035          * about to write.  We do this *before* the write so that we can return
3036          * without clobbering -EIOCBQUEUED from ->direct_IO().
3037          */
3038         written = invalidate_inode_pages2_range(mapping,
3039                                         pos >> PAGE_SHIFT, end);
3040         /*
3041          * If a page can not be invalidated, return 0 to fall back
3042          * to buffered write.
3043          */
3044         if (written) {
3045                 if (written == -EBUSY)
3046                         return 0;
3047                 goto out;
3048         }
3049
3050         written = mapping->a_ops->direct_IO(iocb, from);
3051
3052         /*
3053          * Finally, try again to invalidate clean pages which might have been
3054          * cached by non-direct readahead, or faulted in by get_user_pages()
3055          * if the source of the write was an mmap'ed region of the file
3056          * we're writing.  Either one is a pretty crazy thing to do,
3057          * so we don't support it 100%.  If this invalidation
3058          * fails, tough, the write still worked...
3059          *
3060          * Most of the time we do not need this since dio_complete() will do
3061          * the invalidation for us. However there are some file systems that
3062          * do not end up with dio_complete() being called, so let's not break
3063          * them by removing it completely
3064          */
3065         if (mapping->nrpages)
3066                 invalidate_inode_pages2_range(mapping,
3067                                         pos >> PAGE_SHIFT, end);
3068
3069         if (written > 0) {
3070                 pos += written;
3071                 write_len -= written;
3072                 if (pos > i_size_read(inode) && !S_ISBLK(inode->i_mode)) {
3073                         i_size_write(inode, pos);
3074                         mark_inode_dirty(inode);
3075                 }
3076                 iocb->ki_pos = pos;
3077         }
3078         iov_iter_revert(from, write_len - iov_iter_count(from));
3079 out:
3080         return written;
3081 }
3082 EXPORT_SYMBOL(generic_file_direct_write);
3083
3084 /*
3085  * Find or create a page at the given pagecache position. Return the locked
3086  * page. This function is specifically for buffered writes.
3087  */
3088 struct page *grab_cache_page_write_begin(struct address_space *mapping,
3089                                         pgoff_t index, unsigned flags)
3090 {
3091         struct page *page;
3092         int fgp_flags = FGP_LOCK|FGP_WRITE|FGP_CREAT;
3093
3094         if (flags & AOP_FLAG_NOFS)
3095                 fgp_flags |= FGP_NOFS;
3096
3097         page = pagecache_get_page(mapping, index, fgp_flags,
3098                         mapping_gfp_mask(mapping));
3099         if (page)
3100                 wait_for_stable_page(page);
3101
3102         return page;
3103 }
3104 EXPORT_SYMBOL(grab_cache_page_write_begin);
3105
3106 ssize_t generic_perform_write(struct file *file,
3107                                 struct iov_iter *i, loff_t pos)
3108 {
3109         struct address_space *mapping = file->f_mapping;
3110         const struct address_space_operations *a_ops = mapping->a_ops;
3111         long status = 0;
3112         ssize_t written = 0;
3113         unsigned int flags = 0;
3114
3115         do {
3116                 struct page *page;
3117                 unsigned long offset;   /* Offset into pagecache page */
3118                 unsigned long bytes;    /* Bytes to write to page */
3119                 size_t copied;          /* Bytes copied from user */
3120                 void *fsdata;
3121
3122                 offset = (pos & (PAGE_SIZE - 1));
3123                 bytes = min_t(unsigned long, PAGE_SIZE - offset,
3124                                                 iov_iter_count(i));
3125
3126 again:
3127                 /*
3128                  * Bring in the user page that we will copy from _first_.
3129                  * Otherwise there's a nasty deadlock on copying from the
3130                  * same page as we're writing to, without it being marked
3131                  * up-to-date.
3132                  *
3133                  * Not only is this an optimisation, but it is also required
3134                  * to check that the address is actually valid, when atomic
3135                  * usercopies are used, below.
3136                  */
3137                 if (unlikely(iov_iter_fault_in_readable(i, bytes))) {
3138                         status = -EFAULT;
3139                         break;
3140                 }
3141
3142                 if (fatal_signal_pending(current)) {
3143                         status = -EINTR;
3144                         break;
3145                 }
3146
3147                 status = a_ops->write_begin(file, mapping, pos, bytes, flags,
3148                                                 &page, &fsdata);
3149                 if (unlikely(status < 0))
3150                         break;
3151
3152                 if (mapping_writably_mapped(mapping))
3153                         flush_dcache_page(page);
3154
3155                 copied = iov_iter_copy_from_user_atomic(page, i, offset, bytes);
3156                 flush_dcache_page(page);
3157
3158                 status = a_ops->write_end(file, mapping, pos, bytes, copied,
3159                                                 page, fsdata);
3160                 if (unlikely(status < 0))
3161                         break;
3162                 copied = status;
3163
3164                 cond_resched();
3165
3166                 iov_iter_advance(i, copied);
3167                 if (unlikely(copied == 0)) {
3168                         /*
3169                          * If we were unable to copy any data at all, we must
3170                          * fall back to a single segment length write.
3171                          *
3172                          * If we didn't fallback here, we could livelock
3173                          * because not all segments in the iov can be copied at
3174                          * once without a pagefault.
3175                          */
3176                         bytes = min_t(unsigned long, PAGE_SIZE - offset,
3177                                                 iov_iter_single_seg_count(i));
3178                         goto again;
3179                 }
3180                 pos += copied;
3181                 written += copied;
3182
3183                 balance_dirty_pages_ratelimited(mapping);
3184         } while (iov_iter_count(i));
3185
3186         return written ? written : status;
3187 }
3188 EXPORT_SYMBOL(generic_perform_write);
3189
3190 /**
3191  * __generic_file_write_iter - write data to a file
3192  * @iocb:       IO state structure (file, offset, etc.)
3193  * @from:       iov_iter with data to write
3194  *
3195  * This function does all the work needed for actually writing data to a
3196  * file. It does all basic checks, removes SUID from the file, updates
3197  * modification times and calls proper subroutines depending on whether we
3198  * do direct IO or a standard buffered write.
3199  *
3200  * It expects i_mutex to be grabbed unless we work on a block device or similar
3201  * object which does not need locking at all.
3202  *
3203  * This function does *not* take care of syncing data in case of O_SYNC write.
3204  * A caller has to handle it. This is mainly due to the fact that we want to
3205  * avoid syncing under i_mutex.
3206  */
3207 ssize_t __generic_file_write_iter(struct kiocb *iocb, struct iov_iter *from)
3208 {
3209         struct file *file = iocb->ki_filp;
3210         struct address_space * mapping = file->f_mapping;
3211         struct inode    *inode = mapping->host;
3212         ssize_t         written = 0;
3213         ssize_t         err;
3214         ssize_t         status;
3215
3216         /* We can write back this queue in page reclaim */
3217         current->backing_dev_info = inode_to_bdi(inode);
3218         err = file_remove_privs(file);
3219         if (err)
3220                 goto out;
3221
3222         err = file_update_time(file);
3223         if (err)
3224                 goto out;
3225
3226         if (iocb->ki_flags & IOCB_DIRECT) {
3227                 loff_t pos, endbyte;
3228
3229                 written = generic_file_direct_write(iocb, from);
3230                 /*
3231                  * If the write stopped short of completing, fall back to
3232                  * buffered writes.  Some filesystems do this for writes to
3233                  * holes, for example.  For DAX files, a buffered write will
3234                  * not succeed (even if it did, DAX does not handle dirty
3235                  * page-cache pages correctly).
3236                  */
3237                 if (written < 0 || !iov_iter_count(from) || IS_DAX(inode))
3238                         goto out;
3239
3240                 status = generic_perform_write(file, from, pos = iocb->ki_pos);
3241                 /*
3242                  * If generic_perform_write() returned a synchronous error
3243                  * then we want to return the number of bytes which were
3244                  * direct-written, or the error code if that was zero.  Note
3245                  * that this differs from normal direct-io semantics, which
3246                  * will return -EFOO even if some bytes were written.
3247                  */
3248                 if (unlikely(status < 0)) {
3249                         err = status;
3250                         goto out;
3251                 }
3252                 /*
3253                  * We need to ensure that the page cache pages are written to
3254                  * disk and invalidated to preserve the expected O_DIRECT
3255                  * semantics.
3256                  */
3257                 endbyte = pos + status - 1;
3258                 err = filemap_write_and_wait_range(mapping, pos, endbyte);
3259                 if (err == 0) {
3260                         iocb->ki_pos = endbyte + 1;
3261                         written += status;
3262                         invalidate_mapping_pages(mapping,
3263                                                  pos >> PAGE_SHIFT,
3264                                                  endbyte >> PAGE_SHIFT);
3265                 } else {
3266                         /*
3267                          * We don't know how much we wrote, so just return
3268                          * the number of bytes which were direct-written
3269                          */
3270                 }
3271         } else {
3272                 written = generic_perform_write(file, from, iocb->ki_pos);
3273                 if (likely(written > 0))
3274                         iocb->ki_pos += written;
3275         }
3276 out:
3277         current->backing_dev_info = NULL;
3278         return written ? written : err;
3279 }
3280 EXPORT_SYMBOL(__generic_file_write_iter);
3281
3282 /**
3283  * generic_file_write_iter - write data to a file
3284  * @iocb:       IO state structure
3285  * @from:       iov_iter with data to write
3286  *
3287  * This is a wrapper around __generic_file_write_iter() to be used by most
3288  * filesystems. It takes care of syncing the file in case of O_SYNC file
3289  * and acquires i_mutex as needed.
3290  */
3291 ssize_t generic_file_write_iter(struct kiocb *iocb, struct iov_iter *from)
3292 {
3293         struct file *file = iocb->ki_filp;
3294         struct inode *inode = file->f_mapping->host;
3295         ssize_t ret;
3296
3297         inode_lock(inode);
3298         ret = generic_write_checks(iocb, from);
3299         if (ret > 0)
3300                 ret = __generic_file_write_iter(iocb, from);
3301         inode_unlock(inode);
3302
3303         if (ret > 0)
3304                 ret = generic_write_sync(iocb, ret);
3305         return ret;
3306 }
3307 EXPORT_SYMBOL(generic_file_write_iter);
3308
3309 /**
3310  * try_to_release_page() - release old fs-specific metadata on a page
3311  *
3312  * @page: the page which the kernel is trying to free
3313  * @gfp_mask: memory allocation flags (and I/O mode)
3314  *
3315  * The address_space is to try to release any data against the page
3316  * (presumably at page->private).  If the release was successful, return '1'.
3317  * Otherwise return zero.
3318  *
3319  * This may also be called if PG_fscache is set on a page, indicating that the
3320  * page is known to the local caching routines.
3321  *
3322  * The @gfp_mask argument specifies whether I/O may be performed to release
3323  * this page (__GFP_IO), and whether the call may block (__GFP_RECLAIM & __GFP_FS).
3324  *
3325  */
3326 int try_to_release_page(struct page *page, gfp_t gfp_mask)
3327 {
3328         struct address_space * const mapping = page->mapping;
3329
3330         BUG_ON(!PageLocked(page));
3331         if (PageWriteback(page))
3332                 return 0;
3333
3334         if (mapping && mapping->a_ops->releasepage)
3335                 return mapping->a_ops->releasepage(page, gfp_mask);
3336         return try_to_free_buffers(page);
3337 }
3338
3339 EXPORT_SYMBOL(try_to_release_page);