mm: use pagevec_lookup_range_tag() in __filemap_fdatawait_range()
[sfrench/cifs-2.6.git] / mm / filemap.c
1 /*
2  *      linux/mm/filemap.c
3  *
4  * Copyright (C) 1994-1999  Linus Torvalds
5  */
6
7 /*
8  * This file handles the generic file mmap semantics used by
9  * most "normal" filesystems (but you don't /have/ to use this:
10  * the NFS filesystem used to do this differently, for example)
11  */
12 #include <linux/export.h>
13 #include <linux/compiler.h>
14 #include <linux/dax.h>
15 #include <linux/fs.h>
16 #include <linux/sched/signal.h>
17 #include <linux/uaccess.h>
18 #include <linux/capability.h>
19 #include <linux/kernel_stat.h>
20 #include <linux/gfp.h>
21 #include <linux/mm.h>
22 #include <linux/swap.h>
23 #include <linux/mman.h>
24 #include <linux/pagemap.h>
25 #include <linux/file.h>
26 #include <linux/uio.h>
27 #include <linux/hash.h>
28 #include <linux/writeback.h>
29 #include <linux/backing-dev.h>
30 #include <linux/pagevec.h>
31 #include <linux/blkdev.h>
32 #include <linux/security.h>
33 #include <linux/cpuset.h>
34 #include <linux/hardirq.h> /* for BUG_ON(!in_atomic()) only */
35 #include <linux/hugetlb.h>
36 #include <linux/memcontrol.h>
37 #include <linux/cleancache.h>
38 #include <linux/rmap.h>
39 #include "internal.h"
40
41 #define CREATE_TRACE_POINTS
42 #include <trace/events/filemap.h>
43
44 /*
45  * FIXME: remove all knowledge of the buffer layer from the core VM
46  */
47 #include <linux/buffer_head.h> /* for try_to_free_buffers */
48
49 #include <asm/mman.h>
50
51 /*
52  * Shared mappings implemented 30.11.1994. It's not fully working yet,
53  * though.
54  *
55  * Shared mappings now work. 15.8.1995  Bruno.
56  *
57  * finished 'unifying' the page and buffer cache and SMP-threaded the
58  * page-cache, 21.05.1999, Ingo Molnar <mingo@redhat.com>
59  *
60  * SMP-threaded pagemap-LRU 1999, Andrea Arcangeli <andrea@suse.de>
61  */
62
63 /*
64  * Lock ordering:
65  *
66  *  ->i_mmap_rwsem              (truncate_pagecache)
67  *    ->private_lock            (__free_pte->__set_page_dirty_buffers)
68  *      ->swap_lock             (exclusive_swap_page, others)
69  *        ->mapping->tree_lock
70  *
71  *  ->i_mutex
72  *    ->i_mmap_rwsem            (truncate->unmap_mapping_range)
73  *
74  *  ->mmap_sem
75  *    ->i_mmap_rwsem
76  *      ->page_table_lock or pte_lock   (various, mainly in memory.c)
77  *        ->mapping->tree_lock  (arch-dependent flush_dcache_mmap_lock)
78  *
79  *  ->mmap_sem
80  *    ->lock_page               (access_process_vm)
81  *
82  *  ->i_mutex                   (generic_perform_write)
83  *    ->mmap_sem                (fault_in_pages_readable->do_page_fault)
84  *
85  *  bdi->wb.list_lock
86  *    sb_lock                   (fs/fs-writeback.c)
87  *    ->mapping->tree_lock      (__sync_single_inode)
88  *
89  *  ->i_mmap_rwsem
90  *    ->anon_vma.lock           (vma_adjust)
91  *
92  *  ->anon_vma.lock
93  *    ->page_table_lock or pte_lock     (anon_vma_prepare and various)
94  *
95  *  ->page_table_lock or pte_lock
96  *    ->swap_lock               (try_to_unmap_one)
97  *    ->private_lock            (try_to_unmap_one)
98  *    ->tree_lock               (try_to_unmap_one)
99  *    ->zone_lru_lock(zone)     (follow_page->mark_page_accessed)
100  *    ->zone_lru_lock(zone)     (check_pte_range->isolate_lru_page)
101  *    ->private_lock            (page_remove_rmap->set_page_dirty)
102  *    ->tree_lock               (page_remove_rmap->set_page_dirty)
103  *    bdi.wb->list_lock         (page_remove_rmap->set_page_dirty)
104  *    ->inode->i_lock           (page_remove_rmap->set_page_dirty)
105  *    ->memcg->move_lock        (page_remove_rmap->lock_page_memcg)
106  *    bdi.wb->list_lock         (zap_pte_range->set_page_dirty)
107  *    ->inode->i_lock           (zap_pte_range->set_page_dirty)
108  *    ->private_lock            (zap_pte_range->__set_page_dirty_buffers)
109  *
110  * ->i_mmap_rwsem
111  *   ->tasklist_lock            (memory_failure, collect_procs_ao)
112  */
113
114 static int page_cache_tree_insert(struct address_space *mapping,
115                                   struct page *page, void **shadowp)
116 {
117         struct radix_tree_node *node;
118         void **slot;
119         int error;
120
121         error = __radix_tree_create(&mapping->page_tree, page->index, 0,
122                                     &node, &slot);
123         if (error)
124                 return error;
125         if (*slot) {
126                 void *p;
127
128                 p = radix_tree_deref_slot_protected(slot, &mapping->tree_lock);
129                 if (!radix_tree_exceptional_entry(p))
130                         return -EEXIST;
131
132                 mapping->nrexceptional--;
133                 if (shadowp)
134                         *shadowp = p;
135         }
136         __radix_tree_replace(&mapping->page_tree, node, slot, page,
137                              workingset_update_node, mapping);
138         mapping->nrpages++;
139         return 0;
140 }
141
142 static void page_cache_tree_delete(struct address_space *mapping,
143                                    struct page *page, void *shadow)
144 {
145         int i, nr;
146
147         /* hugetlb pages are represented by one entry in the radix tree */
148         nr = PageHuge(page) ? 1 : hpage_nr_pages(page);
149
150         VM_BUG_ON_PAGE(!PageLocked(page), page);
151         VM_BUG_ON_PAGE(PageTail(page), page);
152         VM_BUG_ON_PAGE(nr != 1 && shadow, page);
153
154         for (i = 0; i < nr; i++) {
155                 struct radix_tree_node *node;
156                 void **slot;
157
158                 __radix_tree_lookup(&mapping->page_tree, page->index + i,
159                                     &node, &slot);
160
161                 VM_BUG_ON_PAGE(!node && nr != 1, page);
162
163                 radix_tree_clear_tags(&mapping->page_tree, node, slot);
164                 __radix_tree_replace(&mapping->page_tree, node, slot, shadow,
165                                      workingset_update_node, mapping);
166         }
167
168         if (shadow) {
169                 mapping->nrexceptional += nr;
170                 /*
171                  * Make sure the nrexceptional update is committed before
172                  * the nrpages update so that final truncate racing
173                  * with reclaim does not see both counters 0 at the
174                  * same time and miss a shadow entry.
175                  */
176                 smp_wmb();
177         }
178         mapping->nrpages -= nr;
179 }
180
181 /*
182  * Delete a page from the page cache and free it. Caller has to make
183  * sure the page is locked and that nobody else uses it - or that usage
184  * is safe.  The caller must hold the mapping's tree_lock.
185  */
186 void __delete_from_page_cache(struct page *page, void *shadow)
187 {
188         struct address_space *mapping = page->mapping;
189         int nr = hpage_nr_pages(page);
190
191         trace_mm_filemap_delete_from_page_cache(page);
192         /*
193          * if we're uptodate, flush out into the cleancache, otherwise
194          * invalidate any existing cleancache entries.  We can't leave
195          * stale data around in the cleancache once our page is gone
196          */
197         if (PageUptodate(page) && PageMappedToDisk(page))
198                 cleancache_put_page(page);
199         else
200                 cleancache_invalidate_page(mapping, page);
201
202         VM_BUG_ON_PAGE(PageTail(page), page);
203         VM_BUG_ON_PAGE(page_mapped(page), page);
204         if (!IS_ENABLED(CONFIG_DEBUG_VM) && unlikely(page_mapped(page))) {
205                 int mapcount;
206
207                 pr_alert("BUG: Bad page cache in process %s  pfn:%05lx\n",
208                          current->comm, page_to_pfn(page));
209                 dump_page(page, "still mapped when deleted");
210                 dump_stack();
211                 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
212
213                 mapcount = page_mapcount(page);
214                 if (mapping_exiting(mapping) &&
215                     page_count(page) >= mapcount + 2) {
216                         /*
217                          * All vmas have already been torn down, so it's
218                          * a good bet that actually the page is unmapped,
219                          * and we'd prefer not to leak it: if we're wrong,
220                          * some other bad page check should catch it later.
221                          */
222                         page_mapcount_reset(page);
223                         page_ref_sub(page, mapcount);
224                 }
225         }
226
227         page_cache_tree_delete(mapping, page, shadow);
228
229         page->mapping = NULL;
230         /* Leave page->index set: truncation lookup relies upon it */
231
232         /* hugetlb pages do not participate in page cache accounting. */
233         if (PageHuge(page))
234                 return;
235
236         __mod_node_page_state(page_pgdat(page), NR_FILE_PAGES, -nr);
237         if (PageSwapBacked(page)) {
238                 __mod_node_page_state(page_pgdat(page), NR_SHMEM, -nr);
239                 if (PageTransHuge(page))
240                         __dec_node_page_state(page, NR_SHMEM_THPS);
241         } else {
242                 VM_BUG_ON_PAGE(PageTransHuge(page), page);
243         }
244
245         /*
246          * At this point page must be either written or cleaned by truncate.
247          * Dirty page here signals a bug and loss of unwritten data.
248          *
249          * This fixes dirty accounting after removing the page entirely but
250          * leaves PageDirty set: it has no effect for truncated page and
251          * anyway will be cleared before returning page into buddy allocator.
252          */
253         if (WARN_ON_ONCE(PageDirty(page)))
254                 account_page_cleaned(page, mapping, inode_to_wb(mapping->host));
255 }
256
257 /**
258  * delete_from_page_cache - delete page from page cache
259  * @page: the page which the kernel is trying to remove from page cache
260  *
261  * This must be called only on pages that have been verified to be in the page
262  * cache and locked.  It will never put the page into the free list, the caller
263  * has a reference on the page.
264  */
265 void delete_from_page_cache(struct page *page)
266 {
267         struct address_space *mapping = page_mapping(page);
268         unsigned long flags;
269         void (*freepage)(struct page *);
270
271         BUG_ON(!PageLocked(page));
272
273         freepage = mapping->a_ops->freepage;
274
275         spin_lock_irqsave(&mapping->tree_lock, flags);
276         __delete_from_page_cache(page, NULL);
277         spin_unlock_irqrestore(&mapping->tree_lock, flags);
278
279         if (freepage)
280                 freepage(page);
281
282         if (PageTransHuge(page) && !PageHuge(page)) {
283                 page_ref_sub(page, HPAGE_PMD_NR);
284                 VM_BUG_ON_PAGE(page_count(page) <= 0, page);
285         } else {
286                 put_page(page);
287         }
288 }
289 EXPORT_SYMBOL(delete_from_page_cache);
290
291 int filemap_check_errors(struct address_space *mapping)
292 {
293         int ret = 0;
294         /* Check for outstanding write errors */
295         if (test_bit(AS_ENOSPC, &mapping->flags) &&
296             test_and_clear_bit(AS_ENOSPC, &mapping->flags))
297                 ret = -ENOSPC;
298         if (test_bit(AS_EIO, &mapping->flags) &&
299             test_and_clear_bit(AS_EIO, &mapping->flags))
300                 ret = -EIO;
301         return ret;
302 }
303 EXPORT_SYMBOL(filemap_check_errors);
304
305 static int filemap_check_and_keep_errors(struct address_space *mapping)
306 {
307         /* Check for outstanding write errors */
308         if (test_bit(AS_EIO, &mapping->flags))
309                 return -EIO;
310         if (test_bit(AS_ENOSPC, &mapping->flags))
311                 return -ENOSPC;
312         return 0;
313 }
314
315 /**
316  * __filemap_fdatawrite_range - start writeback on mapping dirty pages in range
317  * @mapping:    address space structure to write
318  * @start:      offset in bytes where the range starts
319  * @end:        offset in bytes where the range ends (inclusive)
320  * @sync_mode:  enable synchronous operation
321  *
322  * Start writeback against all of a mapping's dirty pages that lie
323  * within the byte offsets <start, end> inclusive.
324  *
325  * If sync_mode is WB_SYNC_ALL then this is a "data integrity" operation, as
326  * opposed to a regular memory cleansing writeback.  The difference between
327  * these two operations is that if a dirty page/buffer is encountered, it must
328  * be waited upon, and not just skipped over.
329  */
330 int __filemap_fdatawrite_range(struct address_space *mapping, loff_t start,
331                                 loff_t end, int sync_mode)
332 {
333         int ret;
334         struct writeback_control wbc = {
335                 .sync_mode = sync_mode,
336                 .nr_to_write = LONG_MAX,
337                 .range_start = start,
338                 .range_end = end,
339         };
340
341         if (!mapping_cap_writeback_dirty(mapping))
342                 return 0;
343
344         wbc_attach_fdatawrite_inode(&wbc, mapping->host);
345         ret = do_writepages(mapping, &wbc);
346         wbc_detach_inode(&wbc);
347         return ret;
348 }
349
350 static inline int __filemap_fdatawrite(struct address_space *mapping,
351         int sync_mode)
352 {
353         return __filemap_fdatawrite_range(mapping, 0, LLONG_MAX, sync_mode);
354 }
355
356 int filemap_fdatawrite(struct address_space *mapping)
357 {
358         return __filemap_fdatawrite(mapping, WB_SYNC_ALL);
359 }
360 EXPORT_SYMBOL(filemap_fdatawrite);
361
362 int filemap_fdatawrite_range(struct address_space *mapping, loff_t start,
363                                 loff_t end)
364 {
365         return __filemap_fdatawrite_range(mapping, start, end, WB_SYNC_ALL);
366 }
367 EXPORT_SYMBOL(filemap_fdatawrite_range);
368
369 /**
370  * filemap_flush - mostly a non-blocking flush
371  * @mapping:    target address_space
372  *
373  * This is a mostly non-blocking flush.  Not suitable for data-integrity
374  * purposes - I/O may not be started against all dirty pages.
375  */
376 int filemap_flush(struct address_space *mapping)
377 {
378         return __filemap_fdatawrite(mapping, WB_SYNC_NONE);
379 }
380 EXPORT_SYMBOL(filemap_flush);
381
382 /**
383  * filemap_range_has_page - check if a page exists in range.
384  * @mapping:           address space within which to check
385  * @start_byte:        offset in bytes where the range starts
386  * @end_byte:          offset in bytes where the range ends (inclusive)
387  *
388  * Find at least one page in the range supplied, usually used to check if
389  * direct writing in this range will trigger a writeback.
390  */
391 bool filemap_range_has_page(struct address_space *mapping,
392                            loff_t start_byte, loff_t end_byte)
393 {
394         pgoff_t index = start_byte >> PAGE_SHIFT;
395         pgoff_t end = end_byte >> PAGE_SHIFT;
396         struct page *page;
397
398         if (end_byte < start_byte)
399                 return false;
400
401         if (mapping->nrpages == 0)
402                 return false;
403
404         if (!find_get_pages_range(mapping, &index, end, 1, &page))
405                 return false;
406         put_page(page);
407         return true;
408 }
409 EXPORT_SYMBOL(filemap_range_has_page);
410
411 static void __filemap_fdatawait_range(struct address_space *mapping,
412                                      loff_t start_byte, loff_t end_byte)
413 {
414         pgoff_t index = start_byte >> PAGE_SHIFT;
415         pgoff_t end = end_byte >> PAGE_SHIFT;
416         struct pagevec pvec;
417         int nr_pages;
418
419         if (end_byte < start_byte)
420                 return;
421
422         pagevec_init(&pvec, 0);
423         while (index <= end) {
424                 unsigned i;
425
426                 nr_pages = pagevec_lookup_range_tag(&pvec, mapping, &index,
427                                 end, PAGECACHE_TAG_WRITEBACK, PAGEVEC_SIZE);
428                 if (!nr_pages)
429                         break;
430
431                 for (i = 0; i < nr_pages; i++) {
432                         struct page *page = pvec.pages[i];
433
434                         wait_on_page_writeback(page);
435                         ClearPageError(page);
436                 }
437                 pagevec_release(&pvec);
438                 cond_resched();
439         }
440 }
441
442 /**
443  * filemap_fdatawait_range - wait for writeback to complete
444  * @mapping:            address space structure to wait for
445  * @start_byte:         offset in bytes where the range starts
446  * @end_byte:           offset in bytes where the range ends (inclusive)
447  *
448  * Walk the list of under-writeback pages of the given address space
449  * in the given range and wait for all of them.  Check error status of
450  * the address space and return it.
451  *
452  * Since the error status of the address space is cleared by this function,
453  * callers are responsible for checking the return value and handling and/or
454  * reporting the error.
455  */
456 int filemap_fdatawait_range(struct address_space *mapping, loff_t start_byte,
457                             loff_t end_byte)
458 {
459         __filemap_fdatawait_range(mapping, start_byte, end_byte);
460         return filemap_check_errors(mapping);
461 }
462 EXPORT_SYMBOL(filemap_fdatawait_range);
463
464 /**
465  * file_fdatawait_range - wait for writeback to complete
466  * @file:               file pointing to address space structure to wait for
467  * @start_byte:         offset in bytes where the range starts
468  * @end_byte:           offset in bytes where the range ends (inclusive)
469  *
470  * Walk the list of under-writeback pages of the address space that file
471  * refers to, in the given range and wait for all of them.  Check error
472  * status of the address space vs. the file->f_wb_err cursor and return it.
473  *
474  * Since the error status of the file is advanced by this function,
475  * callers are responsible for checking the return value and handling and/or
476  * reporting the error.
477  */
478 int file_fdatawait_range(struct file *file, loff_t start_byte, loff_t end_byte)
479 {
480         struct address_space *mapping = file->f_mapping;
481
482         __filemap_fdatawait_range(mapping, start_byte, end_byte);
483         return file_check_and_advance_wb_err(file);
484 }
485 EXPORT_SYMBOL(file_fdatawait_range);
486
487 /**
488  * filemap_fdatawait_keep_errors - wait for writeback without clearing errors
489  * @mapping: address space structure to wait for
490  *
491  * Walk the list of under-writeback pages of the given address space
492  * and wait for all of them.  Unlike filemap_fdatawait(), this function
493  * does not clear error status of the address space.
494  *
495  * Use this function if callers don't handle errors themselves.  Expected
496  * call sites are system-wide / filesystem-wide data flushers: e.g. sync(2),
497  * fsfreeze(8)
498  */
499 int filemap_fdatawait_keep_errors(struct address_space *mapping)
500 {
501         __filemap_fdatawait_range(mapping, 0, LLONG_MAX);
502         return filemap_check_and_keep_errors(mapping);
503 }
504 EXPORT_SYMBOL(filemap_fdatawait_keep_errors);
505
506 static bool mapping_needs_writeback(struct address_space *mapping)
507 {
508         return (!dax_mapping(mapping) && mapping->nrpages) ||
509             (dax_mapping(mapping) && mapping->nrexceptional);
510 }
511
512 int filemap_write_and_wait(struct address_space *mapping)
513 {
514         int err = 0;
515
516         if (mapping_needs_writeback(mapping)) {
517                 err = filemap_fdatawrite(mapping);
518                 /*
519                  * Even if the above returned error, the pages may be
520                  * written partially (e.g. -ENOSPC), so we wait for it.
521                  * But the -EIO is special case, it may indicate the worst
522                  * thing (e.g. bug) happened, so we avoid waiting for it.
523                  */
524                 if (err != -EIO) {
525                         int err2 = filemap_fdatawait(mapping);
526                         if (!err)
527                                 err = err2;
528                 } else {
529                         /* Clear any previously stored errors */
530                         filemap_check_errors(mapping);
531                 }
532         } else {
533                 err = filemap_check_errors(mapping);
534         }
535         return err;
536 }
537 EXPORT_SYMBOL(filemap_write_and_wait);
538
539 /**
540  * filemap_write_and_wait_range - write out & wait on a file range
541  * @mapping:    the address_space for the pages
542  * @lstart:     offset in bytes where the range starts
543  * @lend:       offset in bytes where the range ends (inclusive)
544  *
545  * Write out and wait upon file offsets lstart->lend, inclusive.
546  *
547  * Note that @lend is inclusive (describes the last byte to be written) so
548  * that this function can be used to write to the very end-of-file (end = -1).
549  */
550 int filemap_write_and_wait_range(struct address_space *mapping,
551                                  loff_t lstart, loff_t lend)
552 {
553         int err = 0;
554
555         if (mapping_needs_writeback(mapping)) {
556                 err = __filemap_fdatawrite_range(mapping, lstart, lend,
557                                                  WB_SYNC_ALL);
558                 /* See comment of filemap_write_and_wait() */
559                 if (err != -EIO) {
560                         int err2 = filemap_fdatawait_range(mapping,
561                                                 lstart, lend);
562                         if (!err)
563                                 err = err2;
564                 } else {
565                         /* Clear any previously stored errors */
566                         filemap_check_errors(mapping);
567                 }
568         } else {
569                 err = filemap_check_errors(mapping);
570         }
571         return err;
572 }
573 EXPORT_SYMBOL(filemap_write_and_wait_range);
574
575 void __filemap_set_wb_err(struct address_space *mapping, int err)
576 {
577         errseq_t eseq = errseq_set(&mapping->wb_err, err);
578
579         trace_filemap_set_wb_err(mapping, eseq);
580 }
581 EXPORT_SYMBOL(__filemap_set_wb_err);
582
583 /**
584  * file_check_and_advance_wb_err - report wb error (if any) that was previously
585  *                                 and advance wb_err to current one
586  * @file: struct file on which the error is being reported
587  *
588  * When userland calls fsync (or something like nfsd does the equivalent), we
589  * want to report any writeback errors that occurred since the last fsync (or
590  * since the file was opened if there haven't been any).
591  *
592  * Grab the wb_err from the mapping. If it matches what we have in the file,
593  * then just quickly return 0. The file is all caught up.
594  *
595  * If it doesn't match, then take the mapping value, set the "seen" flag in
596  * it and try to swap it into place. If it works, or another task beat us
597  * to it with the new value, then update the f_wb_err and return the error
598  * portion. The error at this point must be reported via proper channels
599  * (a'la fsync, or NFS COMMIT operation, etc.).
600  *
601  * While we handle mapping->wb_err with atomic operations, the f_wb_err
602  * value is protected by the f_lock since we must ensure that it reflects
603  * the latest value swapped in for this file descriptor.
604  */
605 int file_check_and_advance_wb_err(struct file *file)
606 {
607         int err = 0;
608         errseq_t old = READ_ONCE(file->f_wb_err);
609         struct address_space *mapping = file->f_mapping;
610
611         /* Locklessly handle the common case where nothing has changed */
612         if (errseq_check(&mapping->wb_err, old)) {
613                 /* Something changed, must use slow path */
614                 spin_lock(&file->f_lock);
615                 old = file->f_wb_err;
616                 err = errseq_check_and_advance(&mapping->wb_err,
617                                                 &file->f_wb_err);
618                 trace_file_check_and_advance_wb_err(file, old);
619                 spin_unlock(&file->f_lock);
620         }
621
622         /*
623          * We're mostly using this function as a drop in replacement for
624          * filemap_check_errors. Clear AS_EIO/AS_ENOSPC to emulate the effect
625          * that the legacy code would have had on these flags.
626          */
627         clear_bit(AS_EIO, &mapping->flags);
628         clear_bit(AS_ENOSPC, &mapping->flags);
629         return err;
630 }
631 EXPORT_SYMBOL(file_check_and_advance_wb_err);
632
633 /**
634  * file_write_and_wait_range - write out & wait on a file range
635  * @file:       file pointing to address_space with pages
636  * @lstart:     offset in bytes where the range starts
637  * @lend:       offset in bytes where the range ends (inclusive)
638  *
639  * Write out and wait upon file offsets lstart->lend, inclusive.
640  *
641  * Note that @lend is inclusive (describes the last byte to be written) so
642  * that this function can be used to write to the very end-of-file (end = -1).
643  *
644  * After writing out and waiting on the data, we check and advance the
645  * f_wb_err cursor to the latest value, and return any errors detected there.
646  */
647 int file_write_and_wait_range(struct file *file, loff_t lstart, loff_t lend)
648 {
649         int err = 0, err2;
650         struct address_space *mapping = file->f_mapping;
651
652         if (mapping_needs_writeback(mapping)) {
653                 err = __filemap_fdatawrite_range(mapping, lstart, lend,
654                                                  WB_SYNC_ALL);
655                 /* See comment of filemap_write_and_wait() */
656                 if (err != -EIO)
657                         __filemap_fdatawait_range(mapping, lstart, lend);
658         }
659         err2 = file_check_and_advance_wb_err(file);
660         if (!err)
661                 err = err2;
662         return err;
663 }
664 EXPORT_SYMBOL(file_write_and_wait_range);
665
666 /**
667  * replace_page_cache_page - replace a pagecache page with a new one
668  * @old:        page to be replaced
669  * @new:        page to replace with
670  * @gfp_mask:   allocation mode
671  *
672  * This function replaces a page in the pagecache with a new one.  On
673  * success it acquires the pagecache reference for the new page and
674  * drops it for the old page.  Both the old and new pages must be
675  * locked.  This function does not add the new page to the LRU, the
676  * caller must do that.
677  *
678  * The remove + add is atomic.  The only way this function can fail is
679  * memory allocation failure.
680  */
681 int replace_page_cache_page(struct page *old, struct page *new, gfp_t gfp_mask)
682 {
683         int error;
684
685         VM_BUG_ON_PAGE(!PageLocked(old), old);
686         VM_BUG_ON_PAGE(!PageLocked(new), new);
687         VM_BUG_ON_PAGE(new->mapping, new);
688
689         error = radix_tree_preload(gfp_mask & ~__GFP_HIGHMEM);
690         if (!error) {
691                 struct address_space *mapping = old->mapping;
692                 void (*freepage)(struct page *);
693                 unsigned long flags;
694
695                 pgoff_t offset = old->index;
696                 freepage = mapping->a_ops->freepage;
697
698                 get_page(new);
699                 new->mapping = mapping;
700                 new->index = offset;
701
702                 spin_lock_irqsave(&mapping->tree_lock, flags);
703                 __delete_from_page_cache(old, NULL);
704                 error = page_cache_tree_insert(mapping, new, NULL);
705                 BUG_ON(error);
706
707                 /*
708                  * hugetlb pages do not participate in page cache accounting.
709                  */
710                 if (!PageHuge(new))
711                         __inc_node_page_state(new, NR_FILE_PAGES);
712                 if (PageSwapBacked(new))
713                         __inc_node_page_state(new, NR_SHMEM);
714                 spin_unlock_irqrestore(&mapping->tree_lock, flags);
715                 mem_cgroup_migrate(old, new);
716                 radix_tree_preload_end();
717                 if (freepage)
718                         freepage(old);
719                 put_page(old);
720         }
721
722         return error;
723 }
724 EXPORT_SYMBOL_GPL(replace_page_cache_page);
725
726 static int __add_to_page_cache_locked(struct page *page,
727                                       struct address_space *mapping,
728                                       pgoff_t offset, gfp_t gfp_mask,
729                                       void **shadowp)
730 {
731         int huge = PageHuge(page);
732         struct mem_cgroup *memcg;
733         int error;
734
735         VM_BUG_ON_PAGE(!PageLocked(page), page);
736         VM_BUG_ON_PAGE(PageSwapBacked(page), page);
737
738         if (!huge) {
739                 error = mem_cgroup_try_charge(page, current->mm,
740                                               gfp_mask, &memcg, false);
741                 if (error)
742                         return error;
743         }
744
745         error = radix_tree_maybe_preload(gfp_mask & ~__GFP_HIGHMEM);
746         if (error) {
747                 if (!huge)
748                         mem_cgroup_cancel_charge(page, memcg, false);
749                 return error;
750         }
751
752         get_page(page);
753         page->mapping = mapping;
754         page->index = offset;
755
756         spin_lock_irq(&mapping->tree_lock);
757         error = page_cache_tree_insert(mapping, page, shadowp);
758         radix_tree_preload_end();
759         if (unlikely(error))
760                 goto err_insert;
761
762         /* hugetlb pages do not participate in page cache accounting. */
763         if (!huge)
764                 __inc_node_page_state(page, NR_FILE_PAGES);
765         spin_unlock_irq(&mapping->tree_lock);
766         if (!huge)
767                 mem_cgroup_commit_charge(page, memcg, false, false);
768         trace_mm_filemap_add_to_page_cache(page);
769         return 0;
770 err_insert:
771         page->mapping = NULL;
772         /* Leave page->index set: truncation relies upon it */
773         spin_unlock_irq(&mapping->tree_lock);
774         if (!huge)
775                 mem_cgroup_cancel_charge(page, memcg, false);
776         put_page(page);
777         return error;
778 }
779
780 /**
781  * add_to_page_cache_locked - add a locked page to the pagecache
782  * @page:       page to add
783  * @mapping:    the page's address_space
784  * @offset:     page index
785  * @gfp_mask:   page allocation mode
786  *
787  * This function is used to add a page to the pagecache. It must be locked.
788  * This function does not add the page to the LRU.  The caller must do that.
789  */
790 int add_to_page_cache_locked(struct page *page, struct address_space *mapping,
791                 pgoff_t offset, gfp_t gfp_mask)
792 {
793         return __add_to_page_cache_locked(page, mapping, offset,
794                                           gfp_mask, NULL);
795 }
796 EXPORT_SYMBOL(add_to_page_cache_locked);
797
798 int add_to_page_cache_lru(struct page *page, struct address_space *mapping,
799                                 pgoff_t offset, gfp_t gfp_mask)
800 {
801         void *shadow = NULL;
802         int ret;
803
804         __SetPageLocked(page);
805         ret = __add_to_page_cache_locked(page, mapping, offset,
806                                          gfp_mask, &shadow);
807         if (unlikely(ret))
808                 __ClearPageLocked(page);
809         else {
810                 /*
811                  * The page might have been evicted from cache only
812                  * recently, in which case it should be activated like
813                  * any other repeatedly accessed page.
814                  * The exception is pages getting rewritten; evicting other
815                  * data from the working set, only to cache data that will
816                  * get overwritten with something else, is a waste of memory.
817                  */
818                 if (!(gfp_mask & __GFP_WRITE) &&
819                     shadow && workingset_refault(shadow)) {
820                         SetPageActive(page);
821                         workingset_activation(page);
822                 } else
823                         ClearPageActive(page);
824                 lru_cache_add(page);
825         }
826         return ret;
827 }
828 EXPORT_SYMBOL_GPL(add_to_page_cache_lru);
829
830 #ifdef CONFIG_NUMA
831 struct page *__page_cache_alloc(gfp_t gfp)
832 {
833         int n;
834         struct page *page;
835
836         if (cpuset_do_page_mem_spread()) {
837                 unsigned int cpuset_mems_cookie;
838                 do {
839                         cpuset_mems_cookie = read_mems_allowed_begin();
840                         n = cpuset_mem_spread_node();
841                         page = __alloc_pages_node(n, gfp, 0);
842                 } while (!page && read_mems_allowed_retry(cpuset_mems_cookie));
843
844                 return page;
845         }
846         return alloc_pages(gfp, 0);
847 }
848 EXPORT_SYMBOL(__page_cache_alloc);
849 #endif
850
851 /*
852  * In order to wait for pages to become available there must be
853  * waitqueues associated with pages. By using a hash table of
854  * waitqueues where the bucket discipline is to maintain all
855  * waiters on the same queue and wake all when any of the pages
856  * become available, and for the woken contexts to check to be
857  * sure the appropriate page became available, this saves space
858  * at a cost of "thundering herd" phenomena during rare hash
859  * collisions.
860  */
861 #define PAGE_WAIT_TABLE_BITS 8
862 #define PAGE_WAIT_TABLE_SIZE (1 << PAGE_WAIT_TABLE_BITS)
863 static wait_queue_head_t page_wait_table[PAGE_WAIT_TABLE_SIZE] __cacheline_aligned;
864
865 static wait_queue_head_t *page_waitqueue(struct page *page)
866 {
867         return &page_wait_table[hash_ptr(page, PAGE_WAIT_TABLE_BITS)];
868 }
869
870 void __init pagecache_init(void)
871 {
872         int i;
873
874         for (i = 0; i < PAGE_WAIT_TABLE_SIZE; i++)
875                 init_waitqueue_head(&page_wait_table[i]);
876
877         page_writeback_init();
878 }
879
880 /* This has the same layout as wait_bit_key - see fs/cachefiles/rdwr.c */
881 struct wait_page_key {
882         struct page *page;
883         int bit_nr;
884         int page_match;
885 };
886
887 struct wait_page_queue {
888         struct page *page;
889         int bit_nr;
890         wait_queue_entry_t wait;
891 };
892
893 static int wake_page_function(wait_queue_entry_t *wait, unsigned mode, int sync, void *arg)
894 {
895         struct wait_page_key *key = arg;
896         struct wait_page_queue *wait_page
897                 = container_of(wait, struct wait_page_queue, wait);
898
899         if (wait_page->page != key->page)
900                return 0;
901         key->page_match = 1;
902
903         if (wait_page->bit_nr != key->bit_nr)
904                 return 0;
905
906         /* Stop walking if it's locked */
907         if (test_bit(key->bit_nr, &key->page->flags))
908                 return -1;
909
910         return autoremove_wake_function(wait, mode, sync, key);
911 }
912
913 static void wake_up_page_bit(struct page *page, int bit_nr)
914 {
915         wait_queue_head_t *q = page_waitqueue(page);
916         struct wait_page_key key;
917         unsigned long flags;
918         wait_queue_entry_t bookmark;
919
920         key.page = page;
921         key.bit_nr = bit_nr;
922         key.page_match = 0;
923
924         bookmark.flags = 0;
925         bookmark.private = NULL;
926         bookmark.func = NULL;
927         INIT_LIST_HEAD(&bookmark.entry);
928
929         spin_lock_irqsave(&q->lock, flags);
930         __wake_up_locked_key_bookmark(q, TASK_NORMAL, &key, &bookmark);
931
932         while (bookmark.flags & WQ_FLAG_BOOKMARK) {
933                 /*
934                  * Take a breather from holding the lock,
935                  * allow pages that finish wake up asynchronously
936                  * to acquire the lock and remove themselves
937                  * from wait queue
938                  */
939                 spin_unlock_irqrestore(&q->lock, flags);
940                 cpu_relax();
941                 spin_lock_irqsave(&q->lock, flags);
942                 __wake_up_locked_key_bookmark(q, TASK_NORMAL, &key, &bookmark);
943         }
944
945         /*
946          * It is possible for other pages to have collided on the waitqueue
947          * hash, so in that case check for a page match. That prevents a long-
948          * term waiter
949          *
950          * It is still possible to miss a case here, when we woke page waiters
951          * and removed them from the waitqueue, but there are still other
952          * page waiters.
953          */
954         if (!waitqueue_active(q) || !key.page_match) {
955                 ClearPageWaiters(page);
956                 /*
957                  * It's possible to miss clearing Waiters here, when we woke
958                  * our page waiters, but the hashed waitqueue has waiters for
959                  * other pages on it.
960                  *
961                  * That's okay, it's a rare case. The next waker will clear it.
962                  */
963         }
964         spin_unlock_irqrestore(&q->lock, flags);
965 }
966
967 static void wake_up_page(struct page *page, int bit)
968 {
969         if (!PageWaiters(page))
970                 return;
971         wake_up_page_bit(page, bit);
972 }
973
974 static inline int wait_on_page_bit_common(wait_queue_head_t *q,
975                 struct page *page, int bit_nr, int state, bool lock)
976 {
977         struct wait_page_queue wait_page;
978         wait_queue_entry_t *wait = &wait_page.wait;
979         int ret = 0;
980
981         init_wait(wait);
982         wait->flags = lock ? WQ_FLAG_EXCLUSIVE : 0;
983         wait->func = wake_page_function;
984         wait_page.page = page;
985         wait_page.bit_nr = bit_nr;
986
987         for (;;) {
988                 spin_lock_irq(&q->lock);
989
990                 if (likely(list_empty(&wait->entry))) {
991                         __add_wait_queue_entry_tail(q, wait);
992                         SetPageWaiters(page);
993                 }
994
995                 set_current_state(state);
996
997                 spin_unlock_irq(&q->lock);
998
999                 if (likely(test_bit(bit_nr, &page->flags))) {
1000                         io_schedule();
1001                 }
1002
1003                 if (lock) {
1004                         if (!test_and_set_bit_lock(bit_nr, &page->flags))
1005                                 break;
1006                 } else {
1007                         if (!test_bit(bit_nr, &page->flags))
1008                                 break;
1009                 }
1010
1011                 if (unlikely(signal_pending_state(state, current))) {
1012                         ret = -EINTR;
1013                         break;
1014                 }
1015         }
1016
1017         finish_wait(q, wait);
1018
1019         /*
1020          * A signal could leave PageWaiters set. Clearing it here if
1021          * !waitqueue_active would be possible (by open-coding finish_wait),
1022          * but still fail to catch it in the case of wait hash collision. We
1023          * already can fail to clear wait hash collision cases, so don't
1024          * bother with signals either.
1025          */
1026
1027         return ret;
1028 }
1029
1030 void wait_on_page_bit(struct page *page, int bit_nr)
1031 {
1032         wait_queue_head_t *q = page_waitqueue(page);
1033         wait_on_page_bit_common(q, page, bit_nr, TASK_UNINTERRUPTIBLE, false);
1034 }
1035 EXPORT_SYMBOL(wait_on_page_bit);
1036
1037 int wait_on_page_bit_killable(struct page *page, int bit_nr)
1038 {
1039         wait_queue_head_t *q = page_waitqueue(page);
1040         return wait_on_page_bit_common(q, page, bit_nr, TASK_KILLABLE, false);
1041 }
1042
1043 /**
1044  * add_page_wait_queue - Add an arbitrary waiter to a page's wait queue
1045  * @page: Page defining the wait queue of interest
1046  * @waiter: Waiter to add to the queue
1047  *
1048  * Add an arbitrary @waiter to the wait queue for the nominated @page.
1049  */
1050 void add_page_wait_queue(struct page *page, wait_queue_entry_t *waiter)
1051 {
1052         wait_queue_head_t *q = page_waitqueue(page);
1053         unsigned long flags;
1054
1055         spin_lock_irqsave(&q->lock, flags);
1056         __add_wait_queue_entry_tail(q, waiter);
1057         SetPageWaiters(page);
1058         spin_unlock_irqrestore(&q->lock, flags);
1059 }
1060 EXPORT_SYMBOL_GPL(add_page_wait_queue);
1061
1062 #ifndef clear_bit_unlock_is_negative_byte
1063
1064 /*
1065  * PG_waiters is the high bit in the same byte as PG_lock.
1066  *
1067  * On x86 (and on many other architectures), we can clear PG_lock and
1068  * test the sign bit at the same time. But if the architecture does
1069  * not support that special operation, we just do this all by hand
1070  * instead.
1071  *
1072  * The read of PG_waiters has to be after (or concurrently with) PG_locked
1073  * being cleared, but a memory barrier should be unneccssary since it is
1074  * in the same byte as PG_locked.
1075  */
1076 static inline bool clear_bit_unlock_is_negative_byte(long nr, volatile void *mem)
1077 {
1078         clear_bit_unlock(nr, mem);
1079         /* smp_mb__after_atomic(); */
1080         return test_bit(PG_waiters, mem);
1081 }
1082
1083 #endif
1084
1085 /**
1086  * unlock_page - unlock a locked page
1087  * @page: the page
1088  *
1089  * Unlocks the page and wakes up sleepers in ___wait_on_page_locked().
1090  * Also wakes sleepers in wait_on_page_writeback() because the wakeup
1091  * mechanism between PageLocked pages and PageWriteback pages is shared.
1092  * But that's OK - sleepers in wait_on_page_writeback() just go back to sleep.
1093  *
1094  * Note that this depends on PG_waiters being the sign bit in the byte
1095  * that contains PG_locked - thus the BUILD_BUG_ON(). That allows us to
1096  * clear the PG_locked bit and test PG_waiters at the same time fairly
1097  * portably (architectures that do LL/SC can test any bit, while x86 can
1098  * test the sign bit).
1099  */
1100 void unlock_page(struct page *page)
1101 {
1102         BUILD_BUG_ON(PG_waiters != 7);
1103         page = compound_head(page);
1104         VM_BUG_ON_PAGE(!PageLocked(page), page);
1105         if (clear_bit_unlock_is_negative_byte(PG_locked, &page->flags))
1106                 wake_up_page_bit(page, PG_locked);
1107 }
1108 EXPORT_SYMBOL(unlock_page);
1109
1110 /**
1111  * end_page_writeback - end writeback against a page
1112  * @page: the page
1113  */
1114 void end_page_writeback(struct page *page)
1115 {
1116         /*
1117          * TestClearPageReclaim could be used here but it is an atomic
1118          * operation and overkill in this particular case. Failing to
1119          * shuffle a page marked for immediate reclaim is too mild to
1120          * justify taking an atomic operation penalty at the end of
1121          * ever page writeback.
1122          */
1123         if (PageReclaim(page)) {
1124                 ClearPageReclaim(page);
1125                 rotate_reclaimable_page(page);
1126         }
1127
1128         if (!test_clear_page_writeback(page))
1129                 BUG();
1130
1131         smp_mb__after_atomic();
1132         wake_up_page(page, PG_writeback);
1133 }
1134 EXPORT_SYMBOL(end_page_writeback);
1135
1136 /*
1137  * After completing I/O on a page, call this routine to update the page
1138  * flags appropriately
1139  */
1140 void page_endio(struct page *page, bool is_write, int err)
1141 {
1142         if (!is_write) {
1143                 if (!err) {
1144                         SetPageUptodate(page);
1145                 } else {
1146                         ClearPageUptodate(page);
1147                         SetPageError(page);
1148                 }
1149                 unlock_page(page);
1150         } else {
1151                 if (err) {
1152                         struct address_space *mapping;
1153
1154                         SetPageError(page);
1155                         mapping = page_mapping(page);
1156                         if (mapping)
1157                                 mapping_set_error(mapping, err);
1158                 }
1159                 end_page_writeback(page);
1160         }
1161 }
1162 EXPORT_SYMBOL_GPL(page_endio);
1163
1164 /**
1165  * __lock_page - get a lock on the page, assuming we need to sleep to get it
1166  * @__page: the page to lock
1167  */
1168 void __lock_page(struct page *__page)
1169 {
1170         struct page *page = compound_head(__page);
1171         wait_queue_head_t *q = page_waitqueue(page);
1172         wait_on_page_bit_common(q, page, PG_locked, TASK_UNINTERRUPTIBLE, true);
1173 }
1174 EXPORT_SYMBOL(__lock_page);
1175
1176 int __lock_page_killable(struct page *__page)
1177 {
1178         struct page *page = compound_head(__page);
1179         wait_queue_head_t *q = page_waitqueue(page);
1180         return wait_on_page_bit_common(q, page, PG_locked, TASK_KILLABLE, true);
1181 }
1182 EXPORT_SYMBOL_GPL(__lock_page_killable);
1183
1184 /*
1185  * Return values:
1186  * 1 - page is locked; mmap_sem is still held.
1187  * 0 - page is not locked.
1188  *     mmap_sem has been released (up_read()), unless flags had both
1189  *     FAULT_FLAG_ALLOW_RETRY and FAULT_FLAG_RETRY_NOWAIT set, in
1190  *     which case mmap_sem is still held.
1191  *
1192  * If neither ALLOW_RETRY nor KILLABLE are set, will always return 1
1193  * with the page locked and the mmap_sem unperturbed.
1194  */
1195 int __lock_page_or_retry(struct page *page, struct mm_struct *mm,
1196                          unsigned int flags)
1197 {
1198         if (flags & FAULT_FLAG_ALLOW_RETRY) {
1199                 /*
1200                  * CAUTION! In this case, mmap_sem is not released
1201                  * even though return 0.
1202                  */
1203                 if (flags & FAULT_FLAG_RETRY_NOWAIT)
1204                         return 0;
1205
1206                 up_read(&mm->mmap_sem);
1207                 if (flags & FAULT_FLAG_KILLABLE)
1208                         wait_on_page_locked_killable(page);
1209                 else
1210                         wait_on_page_locked(page);
1211                 return 0;
1212         } else {
1213                 if (flags & FAULT_FLAG_KILLABLE) {
1214                         int ret;
1215
1216                         ret = __lock_page_killable(page);
1217                         if (ret) {
1218                                 up_read(&mm->mmap_sem);
1219                                 return 0;
1220                         }
1221                 } else
1222                         __lock_page(page);
1223                 return 1;
1224         }
1225 }
1226
1227 /**
1228  * page_cache_next_hole - find the next hole (not-present entry)
1229  * @mapping: mapping
1230  * @index: index
1231  * @max_scan: maximum range to search
1232  *
1233  * Search the set [index, min(index+max_scan-1, MAX_INDEX)] for the
1234  * lowest indexed hole.
1235  *
1236  * Returns: the index of the hole if found, otherwise returns an index
1237  * outside of the set specified (in which case 'return - index >=
1238  * max_scan' will be true). In rare cases of index wrap-around, 0 will
1239  * be returned.
1240  *
1241  * page_cache_next_hole may be called under rcu_read_lock. However,
1242  * like radix_tree_gang_lookup, this will not atomically search a
1243  * snapshot of the tree at a single point in time. For example, if a
1244  * hole is created at index 5, then subsequently a hole is created at
1245  * index 10, page_cache_next_hole covering both indexes may return 10
1246  * if called under rcu_read_lock.
1247  */
1248 pgoff_t page_cache_next_hole(struct address_space *mapping,
1249                              pgoff_t index, unsigned long max_scan)
1250 {
1251         unsigned long i;
1252
1253         for (i = 0; i < max_scan; i++) {
1254                 struct page *page;
1255
1256                 page = radix_tree_lookup(&mapping->page_tree, index);
1257                 if (!page || radix_tree_exceptional_entry(page))
1258                         break;
1259                 index++;
1260                 if (index == 0)
1261                         break;
1262         }
1263
1264         return index;
1265 }
1266 EXPORT_SYMBOL(page_cache_next_hole);
1267
1268 /**
1269  * page_cache_prev_hole - find the prev hole (not-present entry)
1270  * @mapping: mapping
1271  * @index: index
1272  * @max_scan: maximum range to search
1273  *
1274  * Search backwards in the range [max(index-max_scan+1, 0), index] for
1275  * the first hole.
1276  *
1277  * Returns: the index of the hole if found, otherwise returns an index
1278  * outside of the set specified (in which case 'index - return >=
1279  * max_scan' will be true). In rare cases of wrap-around, ULONG_MAX
1280  * will be returned.
1281  *
1282  * page_cache_prev_hole may be called under rcu_read_lock. However,
1283  * like radix_tree_gang_lookup, this will not atomically search a
1284  * snapshot of the tree at a single point in time. For example, if a
1285  * hole is created at index 10, then subsequently a hole is created at
1286  * index 5, page_cache_prev_hole covering both indexes may return 5 if
1287  * called under rcu_read_lock.
1288  */
1289 pgoff_t page_cache_prev_hole(struct address_space *mapping,
1290                              pgoff_t index, unsigned long max_scan)
1291 {
1292         unsigned long i;
1293
1294         for (i = 0; i < max_scan; i++) {
1295                 struct page *page;
1296
1297                 page = radix_tree_lookup(&mapping->page_tree, index);
1298                 if (!page || radix_tree_exceptional_entry(page))
1299                         break;
1300                 index--;
1301                 if (index == ULONG_MAX)
1302                         break;
1303         }
1304
1305         return index;
1306 }
1307 EXPORT_SYMBOL(page_cache_prev_hole);
1308
1309 /**
1310  * find_get_entry - find and get a page cache entry
1311  * @mapping: the address_space to search
1312  * @offset: the page cache index
1313  *
1314  * Looks up the page cache slot at @mapping & @offset.  If there is a
1315  * page cache page, it is returned with an increased refcount.
1316  *
1317  * If the slot holds a shadow entry of a previously evicted page, or a
1318  * swap entry from shmem/tmpfs, it is returned.
1319  *
1320  * Otherwise, %NULL is returned.
1321  */
1322 struct page *find_get_entry(struct address_space *mapping, pgoff_t offset)
1323 {
1324         void **pagep;
1325         struct page *head, *page;
1326
1327         rcu_read_lock();
1328 repeat:
1329         page = NULL;
1330         pagep = radix_tree_lookup_slot(&mapping->page_tree, offset);
1331         if (pagep) {
1332                 page = radix_tree_deref_slot(pagep);
1333                 if (unlikely(!page))
1334                         goto out;
1335                 if (radix_tree_exception(page)) {
1336                         if (radix_tree_deref_retry(page))
1337                                 goto repeat;
1338                         /*
1339                          * A shadow entry of a recently evicted page,
1340                          * or a swap entry from shmem/tmpfs.  Return
1341                          * it without attempting to raise page count.
1342                          */
1343                         goto out;
1344                 }
1345
1346                 head = compound_head(page);
1347                 if (!page_cache_get_speculative(head))
1348                         goto repeat;
1349
1350                 /* The page was split under us? */
1351                 if (compound_head(page) != head) {
1352                         put_page(head);
1353                         goto repeat;
1354                 }
1355
1356                 /*
1357                  * Has the page moved?
1358                  * This is part of the lockless pagecache protocol. See
1359                  * include/linux/pagemap.h for details.
1360                  */
1361                 if (unlikely(page != *pagep)) {
1362                         put_page(head);
1363                         goto repeat;
1364                 }
1365         }
1366 out:
1367         rcu_read_unlock();
1368
1369         return page;
1370 }
1371 EXPORT_SYMBOL(find_get_entry);
1372
1373 /**
1374  * find_lock_entry - locate, pin and lock a page cache entry
1375  * @mapping: the address_space to search
1376  * @offset: the page cache index
1377  *
1378  * Looks up the page cache slot at @mapping & @offset.  If there is a
1379  * page cache page, it is returned locked and with an increased
1380  * refcount.
1381  *
1382  * If the slot holds a shadow entry of a previously evicted page, or a
1383  * swap entry from shmem/tmpfs, it is returned.
1384  *
1385  * Otherwise, %NULL is returned.
1386  *
1387  * find_lock_entry() may sleep.
1388  */
1389 struct page *find_lock_entry(struct address_space *mapping, pgoff_t offset)
1390 {
1391         struct page *page;
1392
1393 repeat:
1394         page = find_get_entry(mapping, offset);
1395         if (page && !radix_tree_exception(page)) {
1396                 lock_page(page);
1397                 /* Has the page been truncated? */
1398                 if (unlikely(page_mapping(page) != mapping)) {
1399                         unlock_page(page);
1400                         put_page(page);
1401                         goto repeat;
1402                 }
1403                 VM_BUG_ON_PAGE(page_to_pgoff(page) != offset, page);
1404         }
1405         return page;
1406 }
1407 EXPORT_SYMBOL(find_lock_entry);
1408
1409 /**
1410  * pagecache_get_page - find and get a page reference
1411  * @mapping: the address_space to search
1412  * @offset: the page index
1413  * @fgp_flags: PCG flags
1414  * @gfp_mask: gfp mask to use for the page cache data page allocation
1415  *
1416  * Looks up the page cache slot at @mapping & @offset.
1417  *
1418  * PCG flags modify how the page is returned.
1419  *
1420  * @fgp_flags can be:
1421  *
1422  * - FGP_ACCESSED: the page will be marked accessed
1423  * - FGP_LOCK: Page is return locked
1424  * - FGP_CREAT: If page is not present then a new page is allocated using
1425  *   @gfp_mask and added to the page cache and the VM's LRU
1426  *   list. The page is returned locked and with an increased
1427  *   refcount. Otherwise, NULL is returned.
1428  *
1429  * If FGP_LOCK or FGP_CREAT are specified then the function may sleep even
1430  * if the GFP flags specified for FGP_CREAT are atomic.
1431  *
1432  * If there is a page cache page, it is returned with an increased refcount.
1433  */
1434 struct page *pagecache_get_page(struct address_space *mapping, pgoff_t offset,
1435         int fgp_flags, gfp_t gfp_mask)
1436 {
1437         struct page *page;
1438
1439 repeat:
1440         page = find_get_entry(mapping, offset);
1441         if (radix_tree_exceptional_entry(page))
1442                 page = NULL;
1443         if (!page)
1444                 goto no_page;
1445
1446         if (fgp_flags & FGP_LOCK) {
1447                 if (fgp_flags & FGP_NOWAIT) {
1448                         if (!trylock_page(page)) {
1449                                 put_page(page);
1450                                 return NULL;
1451                         }
1452                 } else {
1453                         lock_page(page);
1454                 }
1455
1456                 /* Has the page been truncated? */
1457                 if (unlikely(page->mapping != mapping)) {
1458                         unlock_page(page);
1459                         put_page(page);
1460                         goto repeat;
1461                 }
1462                 VM_BUG_ON_PAGE(page->index != offset, page);
1463         }
1464
1465         if (page && (fgp_flags & FGP_ACCESSED))
1466                 mark_page_accessed(page);
1467
1468 no_page:
1469         if (!page && (fgp_flags & FGP_CREAT)) {
1470                 int err;
1471                 if ((fgp_flags & FGP_WRITE) && mapping_cap_account_dirty(mapping))
1472                         gfp_mask |= __GFP_WRITE;
1473                 if (fgp_flags & FGP_NOFS)
1474                         gfp_mask &= ~__GFP_FS;
1475
1476                 page = __page_cache_alloc(gfp_mask);
1477                 if (!page)
1478                         return NULL;
1479
1480                 if (WARN_ON_ONCE(!(fgp_flags & FGP_LOCK)))
1481                         fgp_flags |= FGP_LOCK;
1482
1483                 /* Init accessed so avoid atomic mark_page_accessed later */
1484                 if (fgp_flags & FGP_ACCESSED)
1485                         __SetPageReferenced(page);
1486
1487                 err = add_to_page_cache_lru(page, mapping, offset,
1488                                 gfp_mask & GFP_RECLAIM_MASK);
1489                 if (unlikely(err)) {
1490                         put_page(page);
1491                         page = NULL;
1492                         if (err == -EEXIST)
1493                                 goto repeat;
1494                 }
1495         }
1496
1497         return page;
1498 }
1499 EXPORT_SYMBOL(pagecache_get_page);
1500
1501 /**
1502  * find_get_entries - gang pagecache lookup
1503  * @mapping:    The address_space to search
1504  * @start:      The starting page cache index
1505  * @nr_entries: The maximum number of entries
1506  * @entries:    Where the resulting entries are placed
1507  * @indices:    The cache indices corresponding to the entries in @entries
1508  *
1509  * find_get_entries() will search for and return a group of up to
1510  * @nr_entries entries in the mapping.  The entries are placed at
1511  * @entries.  find_get_entries() takes a reference against any actual
1512  * pages it returns.
1513  *
1514  * The search returns a group of mapping-contiguous page cache entries
1515  * with ascending indexes.  There may be holes in the indices due to
1516  * not-present pages.
1517  *
1518  * Any shadow entries of evicted pages, or swap entries from
1519  * shmem/tmpfs, are included in the returned array.
1520  *
1521  * find_get_entries() returns the number of pages and shadow entries
1522  * which were found.
1523  */
1524 unsigned find_get_entries(struct address_space *mapping,
1525                           pgoff_t start, unsigned int nr_entries,
1526                           struct page **entries, pgoff_t *indices)
1527 {
1528         void **slot;
1529         unsigned int ret = 0;
1530         struct radix_tree_iter iter;
1531
1532         if (!nr_entries)
1533                 return 0;
1534
1535         rcu_read_lock();
1536         radix_tree_for_each_slot(slot, &mapping->page_tree, &iter, start) {
1537                 struct page *head, *page;
1538 repeat:
1539                 page = radix_tree_deref_slot(slot);
1540                 if (unlikely(!page))
1541                         continue;
1542                 if (radix_tree_exception(page)) {
1543                         if (radix_tree_deref_retry(page)) {
1544                                 slot = radix_tree_iter_retry(&iter);
1545                                 continue;
1546                         }
1547                         /*
1548                          * A shadow entry of a recently evicted page, a swap
1549                          * entry from shmem/tmpfs or a DAX entry.  Return it
1550                          * without attempting to raise page count.
1551                          */
1552                         goto export;
1553                 }
1554
1555                 head = compound_head(page);
1556                 if (!page_cache_get_speculative(head))
1557                         goto repeat;
1558
1559                 /* The page was split under us? */
1560                 if (compound_head(page) != head) {
1561                         put_page(head);
1562                         goto repeat;
1563                 }
1564
1565                 /* Has the page moved? */
1566                 if (unlikely(page != *slot)) {
1567                         put_page(head);
1568                         goto repeat;
1569                 }
1570 export:
1571                 indices[ret] = iter.index;
1572                 entries[ret] = page;
1573                 if (++ret == nr_entries)
1574                         break;
1575         }
1576         rcu_read_unlock();
1577         return ret;
1578 }
1579
1580 /**
1581  * find_get_pages_range - gang pagecache lookup
1582  * @mapping:    The address_space to search
1583  * @start:      The starting page index
1584  * @end:        The final page index (inclusive)
1585  * @nr_pages:   The maximum number of pages
1586  * @pages:      Where the resulting pages are placed
1587  *
1588  * find_get_pages_range() will search for and return a group of up to @nr_pages
1589  * pages in the mapping starting at index @start and up to index @end
1590  * (inclusive).  The pages are placed at @pages.  find_get_pages_range() takes
1591  * a reference against the returned pages.
1592  *
1593  * The search returns a group of mapping-contiguous pages with ascending
1594  * indexes.  There may be holes in the indices due to not-present pages.
1595  * We also update @start to index the next page for the traversal.
1596  *
1597  * find_get_pages_range() returns the number of pages which were found. If this
1598  * number is smaller than @nr_pages, the end of specified range has been
1599  * reached.
1600  */
1601 unsigned find_get_pages_range(struct address_space *mapping, pgoff_t *start,
1602                               pgoff_t end, unsigned int nr_pages,
1603                               struct page **pages)
1604 {
1605         struct radix_tree_iter iter;
1606         void **slot;
1607         unsigned ret = 0;
1608
1609         if (unlikely(!nr_pages))
1610                 return 0;
1611
1612         rcu_read_lock();
1613         radix_tree_for_each_slot(slot, &mapping->page_tree, &iter, *start) {
1614                 struct page *head, *page;
1615
1616                 if (iter.index > end)
1617                         break;
1618 repeat:
1619                 page = radix_tree_deref_slot(slot);
1620                 if (unlikely(!page))
1621                         continue;
1622
1623                 if (radix_tree_exception(page)) {
1624                         if (radix_tree_deref_retry(page)) {
1625                                 slot = radix_tree_iter_retry(&iter);
1626                                 continue;
1627                         }
1628                         /*
1629                          * A shadow entry of a recently evicted page,
1630                          * or a swap entry from shmem/tmpfs.  Skip
1631                          * over it.
1632                          */
1633                         continue;
1634                 }
1635
1636                 head = compound_head(page);
1637                 if (!page_cache_get_speculative(head))
1638                         goto repeat;
1639
1640                 /* The page was split under us? */
1641                 if (compound_head(page) != head) {
1642                         put_page(head);
1643                         goto repeat;
1644                 }
1645
1646                 /* Has the page moved? */
1647                 if (unlikely(page != *slot)) {
1648                         put_page(head);
1649                         goto repeat;
1650                 }
1651
1652                 pages[ret] = page;
1653                 if (++ret == nr_pages) {
1654                         *start = pages[ret - 1]->index + 1;
1655                         goto out;
1656                 }
1657         }
1658
1659         /*
1660          * We come here when there is no page beyond @end. We take care to not
1661          * overflow the index @start as it confuses some of the callers. This
1662          * breaks the iteration when there is page at index -1 but that is
1663          * already broken anyway.
1664          */
1665         if (end == (pgoff_t)-1)
1666                 *start = (pgoff_t)-1;
1667         else
1668                 *start = end + 1;
1669 out:
1670         rcu_read_unlock();
1671
1672         return ret;
1673 }
1674
1675 /**
1676  * find_get_pages_contig - gang contiguous pagecache lookup
1677  * @mapping:    The address_space to search
1678  * @index:      The starting page index
1679  * @nr_pages:   The maximum number of pages
1680  * @pages:      Where the resulting pages are placed
1681  *
1682  * find_get_pages_contig() works exactly like find_get_pages(), except
1683  * that the returned number of pages are guaranteed to be contiguous.
1684  *
1685  * find_get_pages_contig() returns the number of pages which were found.
1686  */
1687 unsigned find_get_pages_contig(struct address_space *mapping, pgoff_t index,
1688                                unsigned int nr_pages, struct page **pages)
1689 {
1690         struct radix_tree_iter iter;
1691         void **slot;
1692         unsigned int ret = 0;
1693
1694         if (unlikely(!nr_pages))
1695                 return 0;
1696
1697         rcu_read_lock();
1698         radix_tree_for_each_contig(slot, &mapping->page_tree, &iter, index) {
1699                 struct page *head, *page;
1700 repeat:
1701                 page = radix_tree_deref_slot(slot);
1702                 /* The hole, there no reason to continue */
1703                 if (unlikely(!page))
1704                         break;
1705
1706                 if (radix_tree_exception(page)) {
1707                         if (radix_tree_deref_retry(page)) {
1708                                 slot = radix_tree_iter_retry(&iter);
1709                                 continue;
1710                         }
1711                         /*
1712                          * A shadow entry of a recently evicted page,
1713                          * or a swap entry from shmem/tmpfs.  Stop
1714                          * looking for contiguous pages.
1715                          */
1716                         break;
1717                 }
1718
1719                 head = compound_head(page);
1720                 if (!page_cache_get_speculative(head))
1721                         goto repeat;
1722
1723                 /* The page was split under us? */
1724                 if (compound_head(page) != head) {
1725                         put_page(head);
1726                         goto repeat;
1727                 }
1728
1729                 /* Has the page moved? */
1730                 if (unlikely(page != *slot)) {
1731                         put_page(head);
1732                         goto repeat;
1733                 }
1734
1735                 /*
1736                  * must check mapping and index after taking the ref.
1737                  * otherwise we can get both false positives and false
1738                  * negatives, which is just confusing to the caller.
1739                  */
1740                 if (page->mapping == NULL || page_to_pgoff(page) != iter.index) {
1741                         put_page(page);
1742                         break;
1743                 }
1744
1745                 pages[ret] = page;
1746                 if (++ret == nr_pages)
1747                         break;
1748         }
1749         rcu_read_unlock();
1750         return ret;
1751 }
1752 EXPORT_SYMBOL(find_get_pages_contig);
1753
1754 /**
1755  * find_get_pages_range_tag - find and return pages in given range matching @tag
1756  * @mapping:    the address_space to search
1757  * @index:      the starting page index
1758  * @end:        The final page index (inclusive)
1759  * @tag:        the tag index
1760  * @nr_pages:   the maximum number of pages
1761  * @pages:      where the resulting pages are placed
1762  *
1763  * Like find_get_pages, except we only return pages which are tagged with
1764  * @tag.   We update @index to index the next page for the traversal.
1765  */
1766 unsigned find_get_pages_range_tag(struct address_space *mapping, pgoff_t *index,
1767                         pgoff_t end, int tag, unsigned int nr_pages,
1768                         struct page **pages)
1769 {
1770         struct radix_tree_iter iter;
1771         void **slot;
1772         unsigned ret = 0;
1773
1774         if (unlikely(!nr_pages))
1775                 return 0;
1776
1777         rcu_read_lock();
1778         radix_tree_for_each_tagged(slot, &mapping->page_tree,
1779                                    &iter, *index, tag) {
1780                 struct page *head, *page;
1781
1782                 if (iter.index > end)
1783                         break;
1784 repeat:
1785                 page = radix_tree_deref_slot(slot);
1786                 if (unlikely(!page))
1787                         continue;
1788
1789                 if (radix_tree_exception(page)) {
1790                         if (radix_tree_deref_retry(page)) {
1791                                 slot = radix_tree_iter_retry(&iter);
1792                                 continue;
1793                         }
1794                         /*
1795                          * A shadow entry of a recently evicted page.
1796                          *
1797                          * Those entries should never be tagged, but
1798                          * this tree walk is lockless and the tags are
1799                          * looked up in bulk, one radix tree node at a
1800                          * time, so there is a sizable window for page
1801                          * reclaim to evict a page we saw tagged.
1802                          *
1803                          * Skip over it.
1804                          */
1805                         continue;
1806                 }
1807
1808                 head = compound_head(page);
1809                 if (!page_cache_get_speculative(head))
1810                         goto repeat;
1811
1812                 /* The page was split under us? */
1813                 if (compound_head(page) != head) {
1814                         put_page(head);
1815                         goto repeat;
1816                 }
1817
1818                 /* Has the page moved? */
1819                 if (unlikely(page != *slot)) {
1820                         put_page(head);
1821                         goto repeat;
1822                 }
1823
1824                 pages[ret] = page;
1825                 if (++ret == nr_pages) {
1826                         *index = pages[ret - 1]->index + 1;
1827                         goto out;
1828                 }
1829         }
1830
1831         /*
1832          * We come here when we got at @end. We take care to not overflow the
1833          * index @index as it confuses some of the callers. This breaks the
1834          * iteration when there is page at index -1 but that is already broken
1835          * anyway.
1836          */
1837         if (end == (pgoff_t)-1)
1838                 *index = (pgoff_t)-1;
1839         else
1840                 *index = end + 1;
1841 out:
1842         rcu_read_unlock();
1843
1844         return ret;
1845 }
1846 EXPORT_SYMBOL(find_get_pages_range_tag);
1847
1848 /**
1849  * find_get_entries_tag - find and return entries that match @tag
1850  * @mapping:    the address_space to search
1851  * @start:      the starting page cache index
1852  * @tag:        the tag index
1853  * @nr_entries: the maximum number of entries
1854  * @entries:    where the resulting entries are placed
1855  * @indices:    the cache indices corresponding to the entries in @entries
1856  *
1857  * Like find_get_entries, except we only return entries which are tagged with
1858  * @tag.
1859  */
1860 unsigned find_get_entries_tag(struct address_space *mapping, pgoff_t start,
1861                         int tag, unsigned int nr_entries,
1862                         struct page **entries, pgoff_t *indices)
1863 {
1864         void **slot;
1865         unsigned int ret = 0;
1866         struct radix_tree_iter iter;
1867
1868         if (!nr_entries)
1869                 return 0;
1870
1871         rcu_read_lock();
1872         radix_tree_for_each_tagged(slot, &mapping->page_tree,
1873                                    &iter, start, tag) {
1874                 struct page *head, *page;
1875 repeat:
1876                 page = radix_tree_deref_slot(slot);
1877                 if (unlikely(!page))
1878                         continue;
1879                 if (radix_tree_exception(page)) {
1880                         if (radix_tree_deref_retry(page)) {
1881                                 slot = radix_tree_iter_retry(&iter);
1882                                 continue;
1883                         }
1884
1885                         /*
1886                          * A shadow entry of a recently evicted page, a swap
1887                          * entry from shmem/tmpfs or a DAX entry.  Return it
1888                          * without attempting to raise page count.
1889                          */
1890                         goto export;
1891                 }
1892
1893                 head = compound_head(page);
1894                 if (!page_cache_get_speculative(head))
1895                         goto repeat;
1896
1897                 /* The page was split under us? */
1898                 if (compound_head(page) != head) {
1899                         put_page(head);
1900                         goto repeat;
1901                 }
1902
1903                 /* Has the page moved? */
1904                 if (unlikely(page != *slot)) {
1905                         put_page(head);
1906                         goto repeat;
1907                 }
1908 export:
1909                 indices[ret] = iter.index;
1910                 entries[ret] = page;
1911                 if (++ret == nr_entries)
1912                         break;
1913         }
1914         rcu_read_unlock();
1915         return ret;
1916 }
1917 EXPORT_SYMBOL(find_get_entries_tag);
1918
1919 /*
1920  * CD/DVDs are error prone. When a medium error occurs, the driver may fail
1921  * a _large_ part of the i/o request. Imagine the worst scenario:
1922  *
1923  *      ---R__________________________________________B__________
1924  *         ^ reading here                             ^ bad block(assume 4k)
1925  *
1926  * read(R) => miss => readahead(R...B) => media error => frustrating retries
1927  * => failing the whole request => read(R) => read(R+1) =>
1928  * readahead(R+1...B+1) => bang => read(R+2) => read(R+3) =>
1929  * readahead(R+3...B+2) => bang => read(R+3) => read(R+4) =>
1930  * readahead(R+4...B+3) => bang => read(R+4) => read(R+5) => ......
1931  *
1932  * It is going insane. Fix it by quickly scaling down the readahead size.
1933  */
1934 static void shrink_readahead_size_eio(struct file *filp,
1935                                         struct file_ra_state *ra)
1936 {
1937         ra->ra_pages /= 4;
1938 }
1939
1940 /**
1941  * generic_file_buffered_read - generic file read routine
1942  * @iocb:       the iocb to read
1943  * @iter:       data destination
1944  * @written:    already copied
1945  *
1946  * This is a generic file read routine, and uses the
1947  * mapping->a_ops->readpage() function for the actual low-level stuff.
1948  *
1949  * This is really ugly. But the goto's actually try to clarify some
1950  * of the logic when it comes to error handling etc.
1951  */
1952 static ssize_t generic_file_buffered_read(struct kiocb *iocb,
1953                 struct iov_iter *iter, ssize_t written)
1954 {
1955         struct file *filp = iocb->ki_filp;
1956         struct address_space *mapping = filp->f_mapping;
1957         struct inode *inode = mapping->host;
1958         struct file_ra_state *ra = &filp->f_ra;
1959         loff_t *ppos = &iocb->ki_pos;
1960         pgoff_t index;
1961         pgoff_t last_index;
1962         pgoff_t prev_index;
1963         unsigned long offset;      /* offset into pagecache page */
1964         unsigned int prev_offset;
1965         int error = 0;
1966
1967         if (unlikely(*ppos >= inode->i_sb->s_maxbytes))
1968                 return 0;
1969         iov_iter_truncate(iter, inode->i_sb->s_maxbytes);
1970
1971         index = *ppos >> PAGE_SHIFT;
1972         prev_index = ra->prev_pos >> PAGE_SHIFT;
1973         prev_offset = ra->prev_pos & (PAGE_SIZE-1);
1974         last_index = (*ppos + iter->count + PAGE_SIZE-1) >> PAGE_SHIFT;
1975         offset = *ppos & ~PAGE_MASK;
1976
1977         for (;;) {
1978                 struct page *page;
1979                 pgoff_t end_index;
1980                 loff_t isize;
1981                 unsigned long nr, ret;
1982
1983                 cond_resched();
1984 find_page:
1985                 if (fatal_signal_pending(current)) {
1986                         error = -EINTR;
1987                         goto out;
1988                 }
1989
1990                 page = find_get_page(mapping, index);
1991                 if (!page) {
1992                         if (iocb->ki_flags & IOCB_NOWAIT)
1993                                 goto would_block;
1994                         page_cache_sync_readahead(mapping,
1995                                         ra, filp,
1996                                         index, last_index - index);
1997                         page = find_get_page(mapping, index);
1998                         if (unlikely(page == NULL))
1999                                 goto no_cached_page;
2000                 }
2001                 if (PageReadahead(page)) {
2002                         page_cache_async_readahead(mapping,
2003                                         ra, filp, page,
2004                                         index, last_index - index);
2005                 }
2006                 if (!PageUptodate(page)) {
2007                         if (iocb->ki_flags & IOCB_NOWAIT) {
2008                                 put_page(page);
2009                                 goto would_block;
2010                         }
2011
2012                         /*
2013                          * See comment in do_read_cache_page on why
2014                          * wait_on_page_locked is used to avoid unnecessarily
2015                          * serialisations and why it's safe.
2016                          */
2017                         error = wait_on_page_locked_killable(page);
2018                         if (unlikely(error))
2019                                 goto readpage_error;
2020                         if (PageUptodate(page))
2021                                 goto page_ok;
2022
2023                         if (inode->i_blkbits == PAGE_SHIFT ||
2024                                         !mapping->a_ops->is_partially_uptodate)
2025                                 goto page_not_up_to_date;
2026                         /* pipes can't handle partially uptodate pages */
2027                         if (unlikely(iter->type & ITER_PIPE))
2028                                 goto page_not_up_to_date;
2029                         if (!trylock_page(page))
2030                                 goto page_not_up_to_date;
2031                         /* Did it get truncated before we got the lock? */
2032                         if (!page->mapping)
2033                                 goto page_not_up_to_date_locked;
2034                         if (!mapping->a_ops->is_partially_uptodate(page,
2035                                                         offset, iter->count))
2036                                 goto page_not_up_to_date_locked;
2037                         unlock_page(page);
2038                 }
2039 page_ok:
2040                 /*
2041                  * i_size must be checked after we know the page is Uptodate.
2042                  *
2043                  * Checking i_size after the check allows us to calculate
2044                  * the correct value for "nr", which means the zero-filled
2045                  * part of the page is not copied back to userspace (unless
2046                  * another truncate extends the file - this is desired though).
2047                  */
2048
2049                 isize = i_size_read(inode);
2050                 end_index = (isize - 1) >> PAGE_SHIFT;
2051                 if (unlikely(!isize || index > end_index)) {
2052                         put_page(page);
2053                         goto out;
2054                 }
2055
2056                 /* nr is the maximum number of bytes to copy from this page */
2057                 nr = PAGE_SIZE;
2058                 if (index == end_index) {
2059                         nr = ((isize - 1) & ~PAGE_MASK) + 1;
2060                         if (nr <= offset) {
2061                                 put_page(page);
2062                                 goto out;
2063                         }
2064                 }
2065                 nr = nr - offset;
2066
2067                 /* If users can be writing to this page using arbitrary
2068                  * virtual addresses, take care about potential aliasing
2069                  * before reading the page on the kernel side.
2070                  */
2071                 if (mapping_writably_mapped(mapping))
2072                         flush_dcache_page(page);
2073
2074                 /*
2075                  * When a sequential read accesses a page several times,
2076                  * only mark it as accessed the first time.
2077                  */
2078                 if (prev_index != index || offset != prev_offset)
2079                         mark_page_accessed(page);
2080                 prev_index = index;
2081
2082                 /*
2083                  * Ok, we have the page, and it's up-to-date, so
2084                  * now we can copy it to user space...
2085                  */
2086
2087                 ret = copy_page_to_iter(page, offset, nr, iter);
2088                 offset += ret;
2089                 index += offset >> PAGE_SHIFT;
2090                 offset &= ~PAGE_MASK;
2091                 prev_offset = offset;
2092
2093                 put_page(page);
2094                 written += ret;
2095                 if (!iov_iter_count(iter))
2096                         goto out;
2097                 if (ret < nr) {
2098                         error = -EFAULT;
2099                         goto out;
2100                 }
2101                 continue;
2102
2103 page_not_up_to_date:
2104                 /* Get exclusive access to the page ... */
2105                 error = lock_page_killable(page);
2106                 if (unlikely(error))
2107                         goto readpage_error;
2108
2109 page_not_up_to_date_locked:
2110                 /* Did it get truncated before we got the lock? */
2111                 if (!page->mapping) {
2112                         unlock_page(page);
2113                         put_page(page);
2114                         continue;
2115                 }
2116
2117                 /* Did somebody else fill it already? */
2118                 if (PageUptodate(page)) {
2119                         unlock_page(page);
2120                         goto page_ok;
2121                 }
2122
2123 readpage:
2124                 /*
2125                  * A previous I/O error may have been due to temporary
2126                  * failures, eg. multipath errors.
2127                  * PG_error will be set again if readpage fails.
2128                  */
2129                 ClearPageError(page);
2130                 /* Start the actual read. The read will unlock the page. */
2131                 error = mapping->a_ops->readpage(filp, page);
2132
2133                 if (unlikely(error)) {
2134                         if (error == AOP_TRUNCATED_PAGE) {
2135                                 put_page(page);
2136                                 error = 0;
2137                                 goto find_page;
2138                         }
2139                         goto readpage_error;
2140                 }
2141
2142                 if (!PageUptodate(page)) {
2143                         error = lock_page_killable(page);
2144                         if (unlikely(error))
2145                                 goto readpage_error;
2146                         if (!PageUptodate(page)) {
2147                                 if (page->mapping == NULL) {
2148                                         /*
2149                                          * invalidate_mapping_pages got it
2150                                          */
2151                                         unlock_page(page);
2152                                         put_page(page);
2153                                         goto find_page;
2154                                 }
2155                                 unlock_page(page);
2156                                 shrink_readahead_size_eio(filp, ra);
2157                                 error = -EIO;
2158                                 goto readpage_error;
2159                         }
2160                         unlock_page(page);
2161                 }
2162
2163                 goto page_ok;
2164
2165 readpage_error:
2166                 /* UHHUH! A synchronous read error occurred. Report it */
2167                 put_page(page);
2168                 goto out;
2169
2170 no_cached_page:
2171                 /*
2172                  * Ok, it wasn't cached, so we need to create a new
2173                  * page..
2174                  */
2175                 page = page_cache_alloc_cold(mapping);
2176                 if (!page) {
2177                         error = -ENOMEM;
2178                         goto out;
2179                 }
2180                 error = add_to_page_cache_lru(page, mapping, index,
2181                                 mapping_gfp_constraint(mapping, GFP_KERNEL));
2182                 if (error) {
2183                         put_page(page);
2184                         if (error == -EEXIST) {
2185                                 error = 0;
2186                                 goto find_page;
2187                         }
2188                         goto out;
2189                 }
2190                 goto readpage;
2191         }
2192
2193 would_block:
2194         error = -EAGAIN;
2195 out:
2196         ra->prev_pos = prev_index;
2197         ra->prev_pos <<= PAGE_SHIFT;
2198         ra->prev_pos |= prev_offset;
2199
2200         *ppos = ((loff_t)index << PAGE_SHIFT) + offset;
2201         file_accessed(filp);
2202         return written ? written : error;
2203 }
2204
2205 /**
2206  * generic_file_read_iter - generic filesystem read routine
2207  * @iocb:       kernel I/O control block
2208  * @iter:       destination for the data read
2209  *
2210  * This is the "read_iter()" routine for all filesystems
2211  * that can use the page cache directly.
2212  */
2213 ssize_t
2214 generic_file_read_iter(struct kiocb *iocb, struct iov_iter *iter)
2215 {
2216         size_t count = iov_iter_count(iter);
2217         ssize_t retval = 0;
2218
2219         if (!count)
2220                 goto out; /* skip atime */
2221
2222         if (iocb->ki_flags & IOCB_DIRECT) {
2223                 struct file *file = iocb->ki_filp;
2224                 struct address_space *mapping = file->f_mapping;
2225                 struct inode *inode = mapping->host;
2226                 loff_t size;
2227
2228                 size = i_size_read(inode);
2229                 if (iocb->ki_flags & IOCB_NOWAIT) {
2230                         if (filemap_range_has_page(mapping, iocb->ki_pos,
2231                                                    iocb->ki_pos + count - 1))
2232                                 return -EAGAIN;
2233                 } else {
2234                         retval = filemap_write_and_wait_range(mapping,
2235                                                 iocb->ki_pos,
2236                                                 iocb->ki_pos + count - 1);
2237                         if (retval < 0)
2238                                 goto out;
2239                 }
2240
2241                 file_accessed(file);
2242
2243                 retval = mapping->a_ops->direct_IO(iocb, iter);
2244                 if (retval >= 0) {
2245                         iocb->ki_pos += retval;
2246                         count -= retval;
2247                 }
2248                 iov_iter_revert(iter, count - iov_iter_count(iter));
2249
2250                 /*
2251                  * Btrfs can have a short DIO read if we encounter
2252                  * compressed extents, so if there was an error, or if
2253                  * we've already read everything we wanted to, or if
2254                  * there was a short read because we hit EOF, go ahead
2255                  * and return.  Otherwise fallthrough to buffered io for
2256                  * the rest of the read.  Buffered reads will not work for
2257                  * DAX files, so don't bother trying.
2258                  */
2259                 if (retval < 0 || !count || iocb->ki_pos >= size ||
2260                     IS_DAX(inode))
2261                         goto out;
2262         }
2263
2264         retval = generic_file_buffered_read(iocb, iter, retval);
2265 out:
2266         return retval;
2267 }
2268 EXPORT_SYMBOL(generic_file_read_iter);
2269
2270 #ifdef CONFIG_MMU
2271 /**
2272  * page_cache_read - adds requested page to the page cache if not already there
2273  * @file:       file to read
2274  * @offset:     page index
2275  * @gfp_mask:   memory allocation flags
2276  *
2277  * This adds the requested page to the page cache if it isn't already there,
2278  * and schedules an I/O to read in its contents from disk.
2279  */
2280 static int page_cache_read(struct file *file, pgoff_t offset, gfp_t gfp_mask)
2281 {
2282         struct address_space *mapping = file->f_mapping;
2283         struct page *page;
2284         int ret;
2285
2286         do {
2287                 page = __page_cache_alloc(gfp_mask|__GFP_COLD);
2288                 if (!page)
2289                         return -ENOMEM;
2290
2291                 ret = add_to_page_cache_lru(page, mapping, offset, gfp_mask & GFP_KERNEL);
2292                 if (ret == 0)
2293                         ret = mapping->a_ops->readpage(file, page);
2294                 else if (ret == -EEXIST)
2295                         ret = 0; /* losing race to add is OK */
2296
2297                 put_page(page);
2298
2299         } while (ret == AOP_TRUNCATED_PAGE);
2300
2301         return ret;
2302 }
2303
2304 #define MMAP_LOTSAMISS  (100)
2305
2306 /*
2307  * Synchronous readahead happens when we don't even find
2308  * a page in the page cache at all.
2309  */
2310 static void do_sync_mmap_readahead(struct vm_area_struct *vma,
2311                                    struct file_ra_state *ra,
2312                                    struct file *file,
2313                                    pgoff_t offset)
2314 {
2315         struct address_space *mapping = file->f_mapping;
2316
2317         /* If we don't want any read-ahead, don't bother */
2318         if (vma->vm_flags & VM_RAND_READ)
2319                 return;
2320         if (!ra->ra_pages)
2321                 return;
2322
2323         if (vma->vm_flags & VM_SEQ_READ) {
2324                 page_cache_sync_readahead(mapping, ra, file, offset,
2325                                           ra->ra_pages);
2326                 return;
2327         }
2328
2329         /* Avoid banging the cache line if not needed */
2330         if (ra->mmap_miss < MMAP_LOTSAMISS * 10)
2331                 ra->mmap_miss++;
2332
2333         /*
2334          * Do we miss much more than hit in this file? If so,
2335          * stop bothering with read-ahead. It will only hurt.
2336          */
2337         if (ra->mmap_miss > MMAP_LOTSAMISS)
2338                 return;
2339
2340         /*
2341          * mmap read-around
2342          */
2343         ra->start = max_t(long, 0, offset - ra->ra_pages / 2);
2344         ra->size = ra->ra_pages;
2345         ra->async_size = ra->ra_pages / 4;
2346         ra_submit(ra, mapping, file);
2347 }
2348
2349 /*
2350  * Asynchronous readahead happens when we find the page and PG_readahead,
2351  * so we want to possibly extend the readahead further..
2352  */
2353 static void do_async_mmap_readahead(struct vm_area_struct *vma,
2354                                     struct file_ra_state *ra,
2355                                     struct file *file,
2356                                     struct page *page,
2357                                     pgoff_t offset)
2358 {
2359         struct address_space *mapping = file->f_mapping;
2360
2361         /* If we don't want any read-ahead, don't bother */
2362         if (vma->vm_flags & VM_RAND_READ)
2363                 return;
2364         if (ra->mmap_miss > 0)
2365                 ra->mmap_miss--;
2366         if (PageReadahead(page))
2367                 page_cache_async_readahead(mapping, ra, file,
2368                                            page, offset, ra->ra_pages);
2369 }
2370
2371 /**
2372  * filemap_fault - read in file data for page fault handling
2373  * @vmf:        struct vm_fault containing details of the fault
2374  *
2375  * filemap_fault() is invoked via the vma operations vector for a
2376  * mapped memory region to read in file data during a page fault.
2377  *
2378  * The goto's are kind of ugly, but this streamlines the normal case of having
2379  * it in the page cache, and handles the special cases reasonably without
2380  * having a lot of duplicated code.
2381  *
2382  * vma->vm_mm->mmap_sem must be held on entry.
2383  *
2384  * If our return value has VM_FAULT_RETRY set, it's because
2385  * lock_page_or_retry() returned 0.
2386  * The mmap_sem has usually been released in this case.
2387  * See __lock_page_or_retry() for the exception.
2388  *
2389  * If our return value does not have VM_FAULT_RETRY set, the mmap_sem
2390  * has not been released.
2391  *
2392  * We never return with VM_FAULT_RETRY and a bit from VM_FAULT_ERROR set.
2393  */
2394 int filemap_fault(struct vm_fault *vmf)
2395 {
2396         int error;
2397         struct file *file = vmf->vma->vm_file;
2398         struct address_space *mapping = file->f_mapping;
2399         struct file_ra_state *ra = &file->f_ra;
2400         struct inode *inode = mapping->host;
2401         pgoff_t offset = vmf->pgoff;
2402         pgoff_t max_off;
2403         struct page *page;
2404         int ret = 0;
2405
2406         max_off = DIV_ROUND_UP(i_size_read(inode), PAGE_SIZE);
2407         if (unlikely(offset >= max_off))
2408                 return VM_FAULT_SIGBUS;
2409
2410         /*
2411          * Do we have something in the page cache already?
2412          */
2413         page = find_get_page(mapping, offset);
2414         if (likely(page) && !(vmf->flags & FAULT_FLAG_TRIED)) {
2415                 /*
2416                  * We found the page, so try async readahead before
2417                  * waiting for the lock.
2418                  */
2419                 do_async_mmap_readahead(vmf->vma, ra, file, page, offset);
2420         } else if (!page) {
2421                 /* No page in the page cache at all */
2422                 do_sync_mmap_readahead(vmf->vma, ra, file, offset);
2423                 count_vm_event(PGMAJFAULT);
2424                 count_memcg_event_mm(vmf->vma->vm_mm, PGMAJFAULT);
2425                 ret = VM_FAULT_MAJOR;
2426 retry_find:
2427                 page = find_get_page(mapping, offset);
2428                 if (!page)
2429                         goto no_cached_page;
2430         }
2431
2432         if (!lock_page_or_retry(page, vmf->vma->vm_mm, vmf->flags)) {
2433                 put_page(page);
2434                 return ret | VM_FAULT_RETRY;
2435         }
2436
2437         /* Did it get truncated? */
2438         if (unlikely(page->mapping != mapping)) {
2439                 unlock_page(page);
2440                 put_page(page);
2441                 goto retry_find;
2442         }
2443         VM_BUG_ON_PAGE(page->index != offset, page);
2444
2445         /*
2446          * We have a locked page in the page cache, now we need to check
2447          * that it's up-to-date. If not, it is going to be due to an error.
2448          */
2449         if (unlikely(!PageUptodate(page)))
2450                 goto page_not_uptodate;
2451
2452         /*
2453          * Found the page and have a reference on it.
2454          * We must recheck i_size under page lock.
2455          */
2456         max_off = DIV_ROUND_UP(i_size_read(inode), PAGE_SIZE);
2457         if (unlikely(offset >= max_off)) {
2458                 unlock_page(page);
2459                 put_page(page);
2460                 return VM_FAULT_SIGBUS;
2461         }
2462
2463         vmf->page = page;
2464         return ret | VM_FAULT_LOCKED;
2465
2466 no_cached_page:
2467         /*
2468          * We're only likely to ever get here if MADV_RANDOM is in
2469          * effect.
2470          */
2471         error = page_cache_read(file, offset, vmf->gfp_mask);
2472
2473         /*
2474          * The page we want has now been added to the page cache.
2475          * In the unlikely event that someone removed it in the
2476          * meantime, we'll just come back here and read it again.
2477          */
2478         if (error >= 0)
2479                 goto retry_find;
2480
2481         /*
2482          * An error return from page_cache_read can result if the
2483          * system is low on memory, or a problem occurs while trying
2484          * to schedule I/O.
2485          */
2486         if (error == -ENOMEM)
2487                 return VM_FAULT_OOM;
2488         return VM_FAULT_SIGBUS;
2489
2490 page_not_uptodate:
2491         /*
2492          * Umm, take care of errors if the page isn't up-to-date.
2493          * Try to re-read it _once_. We do this synchronously,
2494          * because there really aren't any performance issues here
2495          * and we need to check for errors.
2496          */
2497         ClearPageError(page);
2498         error = mapping->a_ops->readpage(file, page);
2499         if (!error) {
2500                 wait_on_page_locked(page);
2501                 if (!PageUptodate(page))
2502                         error = -EIO;
2503         }
2504         put_page(page);
2505
2506         if (!error || error == AOP_TRUNCATED_PAGE)
2507                 goto retry_find;
2508
2509         /* Things didn't work out. Return zero to tell the mm layer so. */
2510         shrink_readahead_size_eio(file, ra);
2511         return VM_FAULT_SIGBUS;
2512 }
2513 EXPORT_SYMBOL(filemap_fault);
2514
2515 void filemap_map_pages(struct vm_fault *vmf,
2516                 pgoff_t start_pgoff, pgoff_t end_pgoff)
2517 {
2518         struct radix_tree_iter iter;
2519         void **slot;
2520         struct file *file = vmf->vma->vm_file;
2521         struct address_space *mapping = file->f_mapping;
2522         pgoff_t last_pgoff = start_pgoff;
2523         unsigned long max_idx;
2524         struct page *head, *page;
2525
2526         rcu_read_lock();
2527         radix_tree_for_each_slot(slot, &mapping->page_tree, &iter,
2528                         start_pgoff) {
2529                 if (iter.index > end_pgoff)
2530                         break;
2531 repeat:
2532                 page = radix_tree_deref_slot(slot);
2533                 if (unlikely(!page))
2534                         goto next;
2535                 if (radix_tree_exception(page)) {
2536                         if (radix_tree_deref_retry(page)) {
2537                                 slot = radix_tree_iter_retry(&iter);
2538                                 continue;
2539                         }
2540                         goto next;
2541                 }
2542
2543                 head = compound_head(page);
2544                 if (!page_cache_get_speculative(head))
2545                         goto repeat;
2546
2547                 /* The page was split under us? */
2548                 if (compound_head(page) != head) {
2549                         put_page(head);
2550                         goto repeat;
2551                 }
2552
2553                 /* Has the page moved? */
2554                 if (unlikely(page != *slot)) {
2555                         put_page(head);
2556                         goto repeat;
2557                 }
2558
2559                 if (!PageUptodate(page) ||
2560                                 PageReadahead(page) ||
2561                                 PageHWPoison(page))
2562                         goto skip;
2563                 if (!trylock_page(page))
2564                         goto skip;
2565
2566                 if (page->mapping != mapping || !PageUptodate(page))
2567                         goto unlock;
2568
2569                 max_idx = DIV_ROUND_UP(i_size_read(mapping->host), PAGE_SIZE);
2570                 if (page->index >= max_idx)
2571                         goto unlock;
2572
2573                 if (file->f_ra.mmap_miss > 0)
2574                         file->f_ra.mmap_miss--;
2575
2576                 vmf->address += (iter.index - last_pgoff) << PAGE_SHIFT;
2577                 if (vmf->pte)
2578                         vmf->pte += iter.index - last_pgoff;
2579                 last_pgoff = iter.index;
2580                 if (alloc_set_pte(vmf, NULL, page))
2581                         goto unlock;
2582                 unlock_page(page);
2583                 goto next;
2584 unlock:
2585                 unlock_page(page);
2586 skip:
2587                 put_page(page);
2588 next:
2589                 /* Huge page is mapped? No need to proceed. */
2590                 if (pmd_trans_huge(*vmf->pmd))
2591                         break;
2592                 if (iter.index == end_pgoff)
2593                         break;
2594         }
2595         rcu_read_unlock();
2596 }
2597 EXPORT_SYMBOL(filemap_map_pages);
2598
2599 int filemap_page_mkwrite(struct vm_fault *vmf)
2600 {
2601         struct page *page = vmf->page;
2602         struct inode *inode = file_inode(vmf->vma->vm_file);
2603         int ret = VM_FAULT_LOCKED;
2604
2605         sb_start_pagefault(inode->i_sb);
2606         file_update_time(vmf->vma->vm_file);
2607         lock_page(page);
2608         if (page->mapping != inode->i_mapping) {
2609                 unlock_page(page);
2610                 ret = VM_FAULT_NOPAGE;
2611                 goto out;
2612         }
2613         /*
2614          * We mark the page dirty already here so that when freeze is in
2615          * progress, we are guaranteed that writeback during freezing will
2616          * see the dirty page and writeprotect it again.
2617          */
2618         set_page_dirty(page);
2619         wait_for_stable_page(page);
2620 out:
2621         sb_end_pagefault(inode->i_sb);
2622         return ret;
2623 }
2624 EXPORT_SYMBOL(filemap_page_mkwrite);
2625
2626 const struct vm_operations_struct generic_file_vm_ops = {
2627         .fault          = filemap_fault,
2628         .map_pages      = filemap_map_pages,
2629         .page_mkwrite   = filemap_page_mkwrite,
2630 };
2631
2632 /* This is used for a general mmap of a disk file */
2633
2634 int generic_file_mmap(struct file * file, struct vm_area_struct * vma)
2635 {
2636         struct address_space *mapping = file->f_mapping;
2637
2638         if (!mapping->a_ops->readpage)
2639                 return -ENOEXEC;
2640         file_accessed(file);
2641         vma->vm_ops = &generic_file_vm_ops;
2642         return 0;
2643 }
2644
2645 /*
2646  * This is for filesystems which do not implement ->writepage.
2647  */
2648 int generic_file_readonly_mmap(struct file *file, struct vm_area_struct *vma)
2649 {
2650         if ((vma->vm_flags & VM_SHARED) && (vma->vm_flags & VM_MAYWRITE))
2651                 return -EINVAL;
2652         return generic_file_mmap(file, vma);
2653 }
2654 #else
2655 int generic_file_mmap(struct file * file, struct vm_area_struct * vma)
2656 {
2657         return -ENOSYS;
2658 }
2659 int generic_file_readonly_mmap(struct file * file, struct vm_area_struct * vma)
2660 {
2661         return -ENOSYS;
2662 }
2663 #endif /* CONFIG_MMU */
2664
2665 EXPORT_SYMBOL(generic_file_mmap);
2666 EXPORT_SYMBOL(generic_file_readonly_mmap);
2667
2668 static struct page *wait_on_page_read(struct page *page)
2669 {
2670         if (!IS_ERR(page)) {
2671                 wait_on_page_locked(page);
2672                 if (!PageUptodate(page)) {
2673                         put_page(page);
2674                         page = ERR_PTR(-EIO);
2675                 }
2676         }
2677         return page;
2678 }
2679
2680 static struct page *do_read_cache_page(struct address_space *mapping,
2681                                 pgoff_t index,
2682                                 int (*filler)(void *, struct page *),
2683                                 void *data,
2684                                 gfp_t gfp)
2685 {
2686         struct page *page;
2687         int err;
2688 repeat:
2689         page = find_get_page(mapping, index);
2690         if (!page) {
2691                 page = __page_cache_alloc(gfp | __GFP_COLD);
2692                 if (!page)
2693                         return ERR_PTR(-ENOMEM);
2694                 err = add_to_page_cache_lru(page, mapping, index, gfp);
2695                 if (unlikely(err)) {
2696                         put_page(page);
2697                         if (err == -EEXIST)
2698                                 goto repeat;
2699                         /* Presumably ENOMEM for radix tree node */
2700                         return ERR_PTR(err);
2701                 }
2702
2703 filler:
2704                 err = filler(data, page);
2705                 if (err < 0) {
2706                         put_page(page);
2707                         return ERR_PTR(err);
2708                 }
2709
2710                 page = wait_on_page_read(page);
2711                 if (IS_ERR(page))
2712                         return page;
2713                 goto out;
2714         }
2715         if (PageUptodate(page))
2716                 goto out;
2717
2718         /*
2719          * Page is not up to date and may be locked due one of the following
2720          * case a: Page is being filled and the page lock is held
2721          * case b: Read/write error clearing the page uptodate status
2722          * case c: Truncation in progress (page locked)
2723          * case d: Reclaim in progress
2724          *
2725          * Case a, the page will be up to date when the page is unlocked.
2726          *    There is no need to serialise on the page lock here as the page
2727          *    is pinned so the lock gives no additional protection. Even if the
2728          *    the page is truncated, the data is still valid if PageUptodate as
2729          *    it's a race vs truncate race.
2730          * Case b, the page will not be up to date
2731          * Case c, the page may be truncated but in itself, the data may still
2732          *    be valid after IO completes as it's a read vs truncate race. The
2733          *    operation must restart if the page is not uptodate on unlock but
2734          *    otherwise serialising on page lock to stabilise the mapping gives
2735          *    no additional guarantees to the caller as the page lock is
2736          *    released before return.
2737          * Case d, similar to truncation. If reclaim holds the page lock, it
2738          *    will be a race with remove_mapping that determines if the mapping
2739          *    is valid on unlock but otherwise the data is valid and there is
2740          *    no need to serialise with page lock.
2741          *
2742          * As the page lock gives no additional guarantee, we optimistically
2743          * wait on the page to be unlocked and check if it's up to date and
2744          * use the page if it is. Otherwise, the page lock is required to
2745          * distinguish between the different cases. The motivation is that we
2746          * avoid spurious serialisations and wakeups when multiple processes
2747          * wait on the same page for IO to complete.
2748          */
2749         wait_on_page_locked(page);
2750         if (PageUptodate(page))
2751                 goto out;
2752
2753         /* Distinguish between all the cases under the safety of the lock */
2754         lock_page(page);
2755
2756         /* Case c or d, restart the operation */
2757         if (!page->mapping) {
2758                 unlock_page(page);
2759                 put_page(page);
2760                 goto repeat;
2761         }
2762
2763         /* Someone else locked and filled the page in a very small window */
2764         if (PageUptodate(page)) {
2765                 unlock_page(page);
2766                 goto out;
2767         }
2768         goto filler;
2769
2770 out:
2771         mark_page_accessed(page);
2772         return page;
2773 }
2774
2775 /**
2776  * read_cache_page - read into page cache, fill it if needed
2777  * @mapping:    the page's address_space
2778  * @index:      the page index
2779  * @filler:     function to perform the read
2780  * @data:       first arg to filler(data, page) function, often left as NULL
2781  *
2782  * Read into the page cache. If a page already exists, and PageUptodate() is
2783  * not set, try to fill the page and wait for it to become unlocked.
2784  *
2785  * If the page does not get brought uptodate, return -EIO.
2786  */
2787 struct page *read_cache_page(struct address_space *mapping,
2788                                 pgoff_t index,
2789                                 int (*filler)(void *, struct page *),
2790                                 void *data)
2791 {
2792         return do_read_cache_page(mapping, index, filler, data, mapping_gfp_mask(mapping));
2793 }
2794 EXPORT_SYMBOL(read_cache_page);
2795
2796 /**
2797  * read_cache_page_gfp - read into page cache, using specified page allocation flags.
2798  * @mapping:    the page's address_space
2799  * @index:      the page index
2800  * @gfp:        the page allocator flags to use if allocating
2801  *
2802  * This is the same as "read_mapping_page(mapping, index, NULL)", but with
2803  * any new page allocations done using the specified allocation flags.
2804  *
2805  * If the page does not get brought uptodate, return -EIO.
2806  */
2807 struct page *read_cache_page_gfp(struct address_space *mapping,
2808                                 pgoff_t index,
2809                                 gfp_t gfp)
2810 {
2811         filler_t *filler = (filler_t *)mapping->a_ops->readpage;
2812
2813         return do_read_cache_page(mapping, index, filler, NULL, gfp);
2814 }
2815 EXPORT_SYMBOL(read_cache_page_gfp);
2816
2817 /*
2818  * Performs necessary checks before doing a write
2819  *
2820  * Can adjust writing position or amount of bytes to write.
2821  * Returns appropriate error code that caller should return or
2822  * zero in case that write should be allowed.
2823  */
2824 inline ssize_t generic_write_checks(struct kiocb *iocb, struct iov_iter *from)
2825 {
2826         struct file *file = iocb->ki_filp;
2827         struct inode *inode = file->f_mapping->host;
2828         unsigned long limit = rlimit(RLIMIT_FSIZE);
2829         loff_t pos;
2830
2831         if (!iov_iter_count(from))
2832                 return 0;
2833
2834         /* FIXME: this is for backwards compatibility with 2.4 */
2835         if (iocb->ki_flags & IOCB_APPEND)
2836                 iocb->ki_pos = i_size_read(inode);
2837
2838         pos = iocb->ki_pos;
2839
2840         if ((iocb->ki_flags & IOCB_NOWAIT) && !(iocb->ki_flags & IOCB_DIRECT))
2841                 return -EINVAL;
2842
2843         if (limit != RLIM_INFINITY) {
2844                 if (iocb->ki_pos >= limit) {
2845                         send_sig(SIGXFSZ, current, 0);
2846                         return -EFBIG;
2847                 }
2848                 iov_iter_truncate(from, limit - (unsigned long)pos);
2849         }
2850
2851         /*
2852          * LFS rule
2853          */
2854         if (unlikely(pos + iov_iter_count(from) > MAX_NON_LFS &&
2855                                 !(file->f_flags & O_LARGEFILE))) {
2856                 if (pos >= MAX_NON_LFS)
2857                         return -EFBIG;
2858                 iov_iter_truncate(from, MAX_NON_LFS - (unsigned long)pos);
2859         }
2860
2861         /*
2862          * Are we about to exceed the fs block limit ?
2863          *
2864          * If we have written data it becomes a short write.  If we have
2865          * exceeded without writing data we send a signal and return EFBIG.
2866          * Linus frestrict idea will clean these up nicely..
2867          */
2868         if (unlikely(pos >= inode->i_sb->s_maxbytes))
2869                 return -EFBIG;
2870
2871         iov_iter_truncate(from, inode->i_sb->s_maxbytes - pos);
2872         return iov_iter_count(from);
2873 }
2874 EXPORT_SYMBOL(generic_write_checks);
2875
2876 int pagecache_write_begin(struct file *file, struct address_space *mapping,
2877                                 loff_t pos, unsigned len, unsigned flags,
2878                                 struct page **pagep, void **fsdata)
2879 {
2880         const struct address_space_operations *aops = mapping->a_ops;
2881
2882         return aops->write_begin(file, mapping, pos, len, flags,
2883                                                         pagep, fsdata);
2884 }
2885 EXPORT_SYMBOL(pagecache_write_begin);
2886
2887 int pagecache_write_end(struct file *file, struct address_space *mapping,
2888                                 loff_t pos, unsigned len, unsigned copied,
2889                                 struct page *page, void *fsdata)
2890 {
2891         const struct address_space_operations *aops = mapping->a_ops;
2892
2893         return aops->write_end(file, mapping, pos, len, copied, page, fsdata);
2894 }
2895 EXPORT_SYMBOL(pagecache_write_end);
2896
2897 ssize_t
2898 generic_file_direct_write(struct kiocb *iocb, struct iov_iter *from)
2899 {
2900         struct file     *file = iocb->ki_filp;
2901         struct address_space *mapping = file->f_mapping;
2902         struct inode    *inode = mapping->host;
2903         loff_t          pos = iocb->ki_pos;
2904         ssize_t         written;
2905         size_t          write_len;
2906         pgoff_t         end;
2907
2908         write_len = iov_iter_count(from);
2909         end = (pos + write_len - 1) >> PAGE_SHIFT;
2910
2911         if (iocb->ki_flags & IOCB_NOWAIT) {
2912                 /* If there are pages to writeback, return */
2913                 if (filemap_range_has_page(inode->i_mapping, pos,
2914                                            pos + iov_iter_count(from)))
2915                         return -EAGAIN;
2916         } else {
2917                 written = filemap_write_and_wait_range(mapping, pos,
2918                                                         pos + write_len - 1);
2919                 if (written)
2920                         goto out;
2921         }
2922
2923         /*
2924          * After a write we want buffered reads to be sure to go to disk to get
2925          * the new data.  We invalidate clean cached page from the region we're
2926          * about to write.  We do this *before* the write so that we can return
2927          * without clobbering -EIOCBQUEUED from ->direct_IO().
2928          */
2929         written = invalidate_inode_pages2_range(mapping,
2930                                         pos >> PAGE_SHIFT, end);
2931         /*
2932          * If a page can not be invalidated, return 0 to fall back
2933          * to buffered write.
2934          */
2935         if (written) {
2936                 if (written == -EBUSY)
2937                         return 0;
2938                 goto out;
2939         }
2940
2941         written = mapping->a_ops->direct_IO(iocb, from);
2942
2943         /*
2944          * Finally, try again to invalidate clean pages which might have been
2945          * cached by non-direct readahead, or faulted in by get_user_pages()
2946          * if the source of the write was an mmap'ed region of the file
2947          * we're writing.  Either one is a pretty crazy thing to do,
2948          * so we don't support it 100%.  If this invalidation
2949          * fails, tough, the write still worked...
2950          *
2951          * Most of the time we do not need this since dio_complete() will do
2952          * the invalidation for us. However there are some file systems that
2953          * do not end up with dio_complete() being called, so let's not break
2954          * them by removing it completely
2955          */
2956         if (mapping->nrpages)
2957                 invalidate_inode_pages2_range(mapping,
2958                                         pos >> PAGE_SHIFT, end);
2959
2960         if (written > 0) {
2961                 pos += written;
2962                 write_len -= written;
2963                 if (pos > i_size_read(inode) && !S_ISBLK(inode->i_mode)) {
2964                         i_size_write(inode, pos);
2965                         mark_inode_dirty(inode);
2966                 }
2967                 iocb->ki_pos = pos;
2968         }
2969         iov_iter_revert(from, write_len - iov_iter_count(from));
2970 out:
2971         return written;
2972 }
2973 EXPORT_SYMBOL(generic_file_direct_write);
2974
2975 /*
2976  * Find or create a page at the given pagecache position. Return the locked
2977  * page. This function is specifically for buffered writes.
2978  */
2979 struct page *grab_cache_page_write_begin(struct address_space *mapping,
2980                                         pgoff_t index, unsigned flags)
2981 {
2982         struct page *page;
2983         int fgp_flags = FGP_LOCK|FGP_WRITE|FGP_CREAT;
2984
2985         if (flags & AOP_FLAG_NOFS)
2986                 fgp_flags |= FGP_NOFS;
2987
2988         page = pagecache_get_page(mapping, index, fgp_flags,
2989                         mapping_gfp_mask(mapping));
2990         if (page)
2991                 wait_for_stable_page(page);
2992
2993         return page;
2994 }
2995 EXPORT_SYMBOL(grab_cache_page_write_begin);
2996
2997 ssize_t generic_perform_write(struct file *file,
2998                                 struct iov_iter *i, loff_t pos)
2999 {
3000         struct address_space *mapping = file->f_mapping;
3001         const struct address_space_operations *a_ops = mapping->a_ops;
3002         long status = 0;
3003         ssize_t written = 0;
3004         unsigned int flags = 0;
3005
3006         do {
3007                 struct page *page;
3008                 unsigned long offset;   /* Offset into pagecache page */
3009                 unsigned long bytes;    /* Bytes to write to page */
3010                 size_t copied;          /* Bytes copied from user */
3011                 void *fsdata;
3012
3013                 offset = (pos & (PAGE_SIZE - 1));
3014                 bytes = min_t(unsigned long, PAGE_SIZE - offset,
3015                                                 iov_iter_count(i));
3016
3017 again:
3018                 /*
3019                  * Bring in the user page that we will copy from _first_.
3020                  * Otherwise there's a nasty deadlock on copying from the
3021                  * same page as we're writing to, without it being marked
3022                  * up-to-date.
3023                  *
3024                  * Not only is this an optimisation, but it is also required
3025                  * to check that the address is actually valid, when atomic
3026                  * usercopies are used, below.
3027                  */
3028                 if (unlikely(iov_iter_fault_in_readable(i, bytes))) {
3029                         status = -EFAULT;
3030                         break;
3031                 }
3032
3033                 if (fatal_signal_pending(current)) {
3034                         status = -EINTR;
3035                         break;
3036                 }
3037
3038                 status = a_ops->write_begin(file, mapping, pos, bytes, flags,
3039                                                 &page, &fsdata);
3040                 if (unlikely(status < 0))
3041                         break;
3042
3043                 if (mapping_writably_mapped(mapping))
3044                         flush_dcache_page(page);
3045
3046                 copied = iov_iter_copy_from_user_atomic(page, i, offset, bytes);
3047                 flush_dcache_page(page);
3048
3049                 status = a_ops->write_end(file, mapping, pos, bytes, copied,
3050                                                 page, fsdata);
3051                 if (unlikely(status < 0))
3052                         break;
3053                 copied = status;
3054
3055                 cond_resched();
3056
3057                 iov_iter_advance(i, copied);
3058                 if (unlikely(copied == 0)) {
3059                         /*
3060                          * If we were unable to copy any data at all, we must
3061                          * fall back to a single segment length write.
3062                          *
3063                          * If we didn't fallback here, we could livelock
3064                          * because not all segments in the iov can be copied at
3065                          * once without a pagefault.
3066                          */
3067                         bytes = min_t(unsigned long, PAGE_SIZE - offset,
3068                                                 iov_iter_single_seg_count(i));
3069                         goto again;
3070                 }
3071                 pos += copied;
3072                 written += copied;
3073
3074                 balance_dirty_pages_ratelimited(mapping);
3075         } while (iov_iter_count(i));
3076
3077         return written ? written : status;
3078 }
3079 EXPORT_SYMBOL(generic_perform_write);
3080
3081 /**
3082  * __generic_file_write_iter - write data to a file
3083  * @iocb:       IO state structure (file, offset, etc.)
3084  * @from:       iov_iter with data to write
3085  *
3086  * This function does all the work needed for actually writing data to a
3087  * file. It does all basic checks, removes SUID from the file, updates
3088  * modification times and calls proper subroutines depending on whether we
3089  * do direct IO or a standard buffered write.
3090  *
3091  * It expects i_mutex to be grabbed unless we work on a block device or similar
3092  * object which does not need locking at all.
3093  *
3094  * This function does *not* take care of syncing data in case of O_SYNC write.
3095  * A caller has to handle it. This is mainly due to the fact that we want to
3096  * avoid syncing under i_mutex.
3097  */
3098 ssize_t __generic_file_write_iter(struct kiocb *iocb, struct iov_iter *from)
3099 {
3100         struct file *file = iocb->ki_filp;
3101         struct address_space * mapping = file->f_mapping;
3102         struct inode    *inode = mapping->host;
3103         ssize_t         written = 0;
3104         ssize_t         err;
3105         ssize_t         status;
3106
3107         /* We can write back this queue in page reclaim */
3108         current->backing_dev_info = inode_to_bdi(inode);
3109         err = file_remove_privs(file);
3110         if (err)
3111                 goto out;
3112
3113         err = file_update_time(file);
3114         if (err)
3115                 goto out;
3116
3117         if (iocb->ki_flags & IOCB_DIRECT) {
3118                 loff_t pos, endbyte;
3119
3120                 written = generic_file_direct_write(iocb, from);
3121                 /*
3122                  * If the write stopped short of completing, fall back to
3123                  * buffered writes.  Some filesystems do this for writes to
3124                  * holes, for example.  For DAX files, a buffered write will
3125                  * not succeed (even if it did, DAX does not handle dirty
3126                  * page-cache pages correctly).
3127                  */
3128                 if (written < 0 || !iov_iter_count(from) || IS_DAX(inode))
3129                         goto out;
3130
3131                 status = generic_perform_write(file, from, pos = iocb->ki_pos);
3132                 /*
3133                  * If generic_perform_write() returned a synchronous error
3134                  * then we want to return the number of bytes which were
3135                  * direct-written, or the error code if that was zero.  Note
3136                  * that this differs from normal direct-io semantics, which
3137                  * will return -EFOO even if some bytes were written.
3138                  */
3139                 if (unlikely(status < 0)) {
3140                         err = status;
3141                         goto out;
3142                 }
3143                 /*
3144                  * We need to ensure that the page cache pages are written to
3145                  * disk and invalidated to preserve the expected O_DIRECT
3146                  * semantics.
3147                  */
3148                 endbyte = pos + status - 1;
3149                 err = filemap_write_and_wait_range(mapping, pos, endbyte);
3150                 if (err == 0) {
3151                         iocb->ki_pos = endbyte + 1;
3152                         written += status;
3153                         invalidate_mapping_pages(mapping,
3154                                                  pos >> PAGE_SHIFT,
3155                                                  endbyte >> PAGE_SHIFT);
3156                 } else {
3157                         /*
3158                          * We don't know how much we wrote, so just return
3159                          * the number of bytes which were direct-written
3160                          */
3161                 }
3162         } else {
3163                 written = generic_perform_write(file, from, iocb->ki_pos);
3164                 if (likely(written > 0))
3165                         iocb->ki_pos += written;
3166         }
3167 out:
3168         current->backing_dev_info = NULL;
3169         return written ? written : err;
3170 }
3171 EXPORT_SYMBOL(__generic_file_write_iter);
3172
3173 /**
3174  * generic_file_write_iter - write data to a file
3175  * @iocb:       IO state structure
3176  * @from:       iov_iter with data to write
3177  *
3178  * This is a wrapper around __generic_file_write_iter() to be used by most
3179  * filesystems. It takes care of syncing the file in case of O_SYNC file
3180  * and acquires i_mutex as needed.
3181  */
3182 ssize_t generic_file_write_iter(struct kiocb *iocb, struct iov_iter *from)
3183 {
3184         struct file *file = iocb->ki_filp;
3185         struct inode *inode = file->f_mapping->host;
3186         ssize_t ret;
3187
3188         inode_lock(inode);
3189         ret = generic_write_checks(iocb, from);
3190         if (ret > 0)
3191                 ret = __generic_file_write_iter(iocb, from);
3192         inode_unlock(inode);
3193
3194         if (ret > 0)
3195                 ret = generic_write_sync(iocb, ret);
3196         return ret;
3197 }
3198 EXPORT_SYMBOL(generic_file_write_iter);
3199
3200 /**
3201  * try_to_release_page() - release old fs-specific metadata on a page
3202  *
3203  * @page: the page which the kernel is trying to free
3204  * @gfp_mask: memory allocation flags (and I/O mode)
3205  *
3206  * The address_space is to try to release any data against the page
3207  * (presumably at page->private).  If the release was successful, return '1'.
3208  * Otherwise return zero.
3209  *
3210  * This may also be called if PG_fscache is set on a page, indicating that the
3211  * page is known to the local caching routines.
3212  *
3213  * The @gfp_mask argument specifies whether I/O may be performed to release
3214  * this page (__GFP_IO), and whether the call may block (__GFP_RECLAIM & __GFP_FS).
3215  *
3216  */
3217 int try_to_release_page(struct page *page, gfp_t gfp_mask)
3218 {
3219         struct address_space * const mapping = page->mapping;
3220
3221         BUG_ON(!PageLocked(page));
3222         if (PageWriteback(page))
3223                 return 0;
3224
3225         if (mapping && mapping->a_ops->releasepage)
3226                 return mapping->a_ops->releasepage(page, gfp_mask);
3227         return try_to_free_buffers(page);
3228 }
3229
3230 EXPORT_SYMBOL(try_to_release_page);