a11b421894369f330f31ba323107593fde8c5ae0
[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         page->mapping = NULL;
169         /* Leave page->index set: truncation lookup relies upon it */
170
171         if (shadow) {
172                 mapping->nrexceptional += nr;
173                 /*
174                  * Make sure the nrexceptional update is committed before
175                  * the nrpages update so that final truncate racing
176                  * with reclaim does not see both counters 0 at the
177                  * same time and miss a shadow entry.
178                  */
179                 smp_wmb();
180         }
181         mapping->nrpages -= nr;
182 }
183
184 static void unaccount_page_cache_page(struct address_space *mapping,
185                                       struct page *page)
186 {
187         int nr;
188
189         /*
190          * if we're uptodate, flush out into the cleancache, otherwise
191          * invalidate any existing cleancache entries.  We can't leave
192          * stale data around in the cleancache once our page is gone
193          */
194         if (PageUptodate(page) && PageMappedToDisk(page))
195                 cleancache_put_page(page);
196         else
197                 cleancache_invalidate_page(mapping, page);
198
199         VM_BUG_ON_PAGE(PageTail(page), page);
200         VM_BUG_ON_PAGE(page_mapped(page), page);
201         if (!IS_ENABLED(CONFIG_DEBUG_VM) && unlikely(page_mapped(page))) {
202                 int mapcount;
203
204                 pr_alert("BUG: Bad page cache in process %s  pfn:%05lx\n",
205                          current->comm, page_to_pfn(page));
206                 dump_page(page, "still mapped when deleted");
207                 dump_stack();
208                 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
209
210                 mapcount = page_mapcount(page);
211                 if (mapping_exiting(mapping) &&
212                     page_count(page) >= mapcount + 2) {
213                         /*
214                          * All vmas have already been torn down, so it's
215                          * a good bet that actually the page is unmapped,
216                          * and we'd prefer not to leak it: if we're wrong,
217                          * some other bad page check should catch it later.
218                          */
219                         page_mapcount_reset(page);
220                         page_ref_sub(page, mapcount);
221                 }
222         }
223
224         /* hugetlb pages do not participate in page cache accounting. */
225         if (PageHuge(page))
226                 return;
227
228         nr = hpage_nr_pages(page);
229
230         __mod_node_page_state(page_pgdat(page), NR_FILE_PAGES, -nr);
231         if (PageSwapBacked(page)) {
232                 __mod_node_page_state(page_pgdat(page), NR_SHMEM, -nr);
233                 if (PageTransHuge(page))
234                         __dec_node_page_state(page, NR_SHMEM_THPS);
235         } else {
236                 VM_BUG_ON_PAGE(PageTransHuge(page), page);
237         }
238
239         /*
240          * At this point page must be either written or cleaned by
241          * truncate.  Dirty page here signals a bug and loss of
242          * unwritten data.
243          *
244          * This fixes dirty accounting after removing the page entirely
245          * but leaves PageDirty set: it has no effect for truncated
246          * page and anyway will be cleared before returning page into
247          * buddy allocator.
248          */
249         if (WARN_ON_ONCE(PageDirty(page)))
250                 account_page_cleaned(page, mapping, inode_to_wb(mapping->host));
251 }
252
253 /*
254  * Delete a page from the page cache and free it. Caller has to make
255  * sure the page is locked and that nobody else uses it - or that usage
256  * is safe.  The caller must hold the mapping's tree_lock.
257  */
258 void __delete_from_page_cache(struct page *page, void *shadow)
259 {
260         struct address_space *mapping = page->mapping;
261
262         trace_mm_filemap_delete_from_page_cache(page);
263
264         unaccount_page_cache_page(mapping, page);
265         page_cache_tree_delete(mapping, page, shadow);
266 }
267
268 static void page_cache_free_page(struct address_space *mapping,
269                                 struct page *page)
270 {
271         void (*freepage)(struct page *);
272
273         freepage = mapping->a_ops->freepage;
274         if (freepage)
275                 freepage(page);
276
277         if (PageTransHuge(page) && !PageHuge(page)) {
278                 page_ref_sub(page, HPAGE_PMD_NR);
279                 VM_BUG_ON_PAGE(page_count(page) <= 0, page);
280         } else {
281                 put_page(page);
282         }
283 }
284
285 /**
286  * delete_from_page_cache - delete page from page cache
287  * @page: the page which the kernel is trying to remove from page cache
288  *
289  * This must be called only on pages that have been verified to be in the page
290  * cache and locked.  It will never put the page into the free list, the caller
291  * has a reference on the page.
292  */
293 void delete_from_page_cache(struct page *page)
294 {
295         struct address_space *mapping = page_mapping(page);
296         unsigned long flags;
297
298         BUG_ON(!PageLocked(page));
299         spin_lock_irqsave(&mapping->tree_lock, flags);
300         __delete_from_page_cache(page, NULL);
301         spin_unlock_irqrestore(&mapping->tree_lock, flags);
302
303         page_cache_free_page(mapping, page);
304 }
305 EXPORT_SYMBOL(delete_from_page_cache);
306
307 int filemap_check_errors(struct address_space *mapping)
308 {
309         int ret = 0;
310         /* Check for outstanding write errors */
311         if (test_bit(AS_ENOSPC, &mapping->flags) &&
312             test_and_clear_bit(AS_ENOSPC, &mapping->flags))
313                 ret = -ENOSPC;
314         if (test_bit(AS_EIO, &mapping->flags) &&
315             test_and_clear_bit(AS_EIO, &mapping->flags))
316                 ret = -EIO;
317         return ret;
318 }
319 EXPORT_SYMBOL(filemap_check_errors);
320
321 static int filemap_check_and_keep_errors(struct address_space *mapping)
322 {
323         /* Check for outstanding write errors */
324         if (test_bit(AS_EIO, &mapping->flags))
325                 return -EIO;
326         if (test_bit(AS_ENOSPC, &mapping->flags))
327                 return -ENOSPC;
328         return 0;
329 }
330
331 /**
332  * __filemap_fdatawrite_range - start writeback on mapping dirty pages in range
333  * @mapping:    address space structure to write
334  * @start:      offset in bytes where the range starts
335  * @end:        offset in bytes where the range ends (inclusive)
336  * @sync_mode:  enable synchronous operation
337  *
338  * Start writeback against all of a mapping's dirty pages that lie
339  * within the byte offsets <start, end> inclusive.
340  *
341  * If sync_mode is WB_SYNC_ALL then this is a "data integrity" operation, as
342  * opposed to a regular memory cleansing writeback.  The difference between
343  * these two operations is that if a dirty page/buffer is encountered, it must
344  * be waited upon, and not just skipped over.
345  */
346 int __filemap_fdatawrite_range(struct address_space *mapping, loff_t start,
347                                 loff_t end, int sync_mode)
348 {
349         int ret;
350         struct writeback_control wbc = {
351                 .sync_mode = sync_mode,
352                 .nr_to_write = LONG_MAX,
353                 .range_start = start,
354                 .range_end = end,
355         };
356
357         if (!mapping_cap_writeback_dirty(mapping))
358                 return 0;
359
360         wbc_attach_fdatawrite_inode(&wbc, mapping->host);
361         ret = do_writepages(mapping, &wbc);
362         wbc_detach_inode(&wbc);
363         return ret;
364 }
365
366 static inline int __filemap_fdatawrite(struct address_space *mapping,
367         int sync_mode)
368 {
369         return __filemap_fdatawrite_range(mapping, 0, LLONG_MAX, sync_mode);
370 }
371
372 int filemap_fdatawrite(struct address_space *mapping)
373 {
374         return __filemap_fdatawrite(mapping, WB_SYNC_ALL);
375 }
376 EXPORT_SYMBOL(filemap_fdatawrite);
377
378 int filemap_fdatawrite_range(struct address_space *mapping, loff_t start,
379                                 loff_t end)
380 {
381         return __filemap_fdatawrite_range(mapping, start, end, WB_SYNC_ALL);
382 }
383 EXPORT_SYMBOL(filemap_fdatawrite_range);
384
385 /**
386  * filemap_flush - mostly a non-blocking flush
387  * @mapping:    target address_space
388  *
389  * This is a mostly non-blocking flush.  Not suitable for data-integrity
390  * purposes - I/O may not be started against all dirty pages.
391  */
392 int filemap_flush(struct address_space *mapping)
393 {
394         return __filemap_fdatawrite(mapping, WB_SYNC_NONE);
395 }
396 EXPORT_SYMBOL(filemap_flush);
397
398 /**
399  * filemap_range_has_page - check if a page exists in range.
400  * @mapping:           address space within which to check
401  * @start_byte:        offset in bytes where the range starts
402  * @end_byte:          offset in bytes where the range ends (inclusive)
403  *
404  * Find at least one page in the range supplied, usually used to check if
405  * direct writing in this range will trigger a writeback.
406  */
407 bool filemap_range_has_page(struct address_space *mapping,
408                            loff_t start_byte, loff_t end_byte)
409 {
410         pgoff_t index = start_byte >> PAGE_SHIFT;
411         pgoff_t end = end_byte >> PAGE_SHIFT;
412         struct page *page;
413
414         if (end_byte < start_byte)
415                 return false;
416
417         if (mapping->nrpages == 0)
418                 return false;
419
420         if (!find_get_pages_range(mapping, &index, end, 1, &page))
421                 return false;
422         put_page(page);
423         return true;
424 }
425 EXPORT_SYMBOL(filemap_range_has_page);
426
427 static void __filemap_fdatawait_range(struct address_space *mapping,
428                                      loff_t start_byte, loff_t end_byte)
429 {
430         pgoff_t index = start_byte >> PAGE_SHIFT;
431         pgoff_t end = end_byte >> PAGE_SHIFT;
432         struct pagevec pvec;
433         int nr_pages;
434
435         if (end_byte < start_byte)
436                 return;
437
438         pagevec_init(&pvec, 0);
439         while (index <= end) {
440                 unsigned i;
441
442                 nr_pages = pagevec_lookup_range_tag(&pvec, mapping, &index,
443                                 end, PAGECACHE_TAG_WRITEBACK);
444                 if (!nr_pages)
445                         break;
446
447                 for (i = 0; i < nr_pages; i++) {
448                         struct page *page = pvec.pages[i];
449
450                         wait_on_page_writeback(page);
451                         ClearPageError(page);
452                 }
453                 pagevec_release(&pvec);
454                 cond_resched();
455         }
456 }
457
458 /**
459  * filemap_fdatawait_range - wait for writeback to complete
460  * @mapping:            address space structure to wait for
461  * @start_byte:         offset in bytes where the range starts
462  * @end_byte:           offset in bytes where the range ends (inclusive)
463  *
464  * Walk the list of under-writeback pages of the given address space
465  * in the given range and wait for all of them.  Check error status of
466  * the address space and return it.
467  *
468  * Since the error status of the address space is cleared by this function,
469  * callers are responsible for checking the return value and handling and/or
470  * reporting the error.
471  */
472 int filemap_fdatawait_range(struct address_space *mapping, loff_t start_byte,
473                             loff_t end_byte)
474 {
475         __filemap_fdatawait_range(mapping, start_byte, end_byte);
476         return filemap_check_errors(mapping);
477 }
478 EXPORT_SYMBOL(filemap_fdatawait_range);
479
480 /**
481  * file_fdatawait_range - wait for writeback to complete
482  * @file:               file pointing to address space structure to wait for
483  * @start_byte:         offset in bytes where the range starts
484  * @end_byte:           offset in bytes where the range ends (inclusive)
485  *
486  * Walk the list of under-writeback pages of the address space that file
487  * refers to, in the given range and wait for all of them.  Check error
488  * status of the address space vs. the file->f_wb_err cursor and return it.
489  *
490  * Since the error status of the file is advanced by this function,
491  * callers are responsible for checking the return value and handling and/or
492  * reporting the error.
493  */
494 int file_fdatawait_range(struct file *file, loff_t start_byte, loff_t end_byte)
495 {
496         struct address_space *mapping = file->f_mapping;
497
498         __filemap_fdatawait_range(mapping, start_byte, end_byte);
499         return file_check_and_advance_wb_err(file);
500 }
501 EXPORT_SYMBOL(file_fdatawait_range);
502
503 /**
504  * filemap_fdatawait_keep_errors - wait for writeback without clearing errors
505  * @mapping: address space structure to wait for
506  *
507  * Walk the list of under-writeback pages of the given address space
508  * and wait for all of them.  Unlike filemap_fdatawait(), this function
509  * does not clear error status of the address space.
510  *
511  * Use this function if callers don't handle errors themselves.  Expected
512  * call sites are system-wide / filesystem-wide data flushers: e.g. sync(2),
513  * fsfreeze(8)
514  */
515 int filemap_fdatawait_keep_errors(struct address_space *mapping)
516 {
517         __filemap_fdatawait_range(mapping, 0, LLONG_MAX);
518         return filemap_check_and_keep_errors(mapping);
519 }
520 EXPORT_SYMBOL(filemap_fdatawait_keep_errors);
521
522 static bool mapping_needs_writeback(struct address_space *mapping)
523 {
524         return (!dax_mapping(mapping) && mapping->nrpages) ||
525             (dax_mapping(mapping) && mapping->nrexceptional);
526 }
527
528 int filemap_write_and_wait(struct address_space *mapping)
529 {
530         int err = 0;
531
532         if (mapping_needs_writeback(mapping)) {
533                 err = filemap_fdatawrite(mapping);
534                 /*
535                  * Even if the above returned error, the pages may be
536                  * written partially (e.g. -ENOSPC), so we wait for it.
537                  * But the -EIO is special case, it may indicate the worst
538                  * thing (e.g. bug) happened, so we avoid waiting for it.
539                  */
540                 if (err != -EIO) {
541                         int err2 = filemap_fdatawait(mapping);
542                         if (!err)
543                                 err = err2;
544                 } else {
545                         /* Clear any previously stored errors */
546                         filemap_check_errors(mapping);
547                 }
548         } else {
549                 err = filemap_check_errors(mapping);
550         }
551         return err;
552 }
553 EXPORT_SYMBOL(filemap_write_and_wait);
554
555 /**
556  * filemap_write_and_wait_range - write out & wait on a file range
557  * @mapping:    the address_space for the pages
558  * @lstart:     offset in bytes where the range starts
559  * @lend:       offset in bytes where the range ends (inclusive)
560  *
561  * Write out and wait upon file offsets lstart->lend, inclusive.
562  *
563  * Note that @lend is inclusive (describes the last byte to be written) so
564  * that this function can be used to write to the very end-of-file (end = -1).
565  */
566 int filemap_write_and_wait_range(struct address_space *mapping,
567                                  loff_t lstart, loff_t lend)
568 {
569         int err = 0;
570
571         if (mapping_needs_writeback(mapping)) {
572                 err = __filemap_fdatawrite_range(mapping, lstart, lend,
573                                                  WB_SYNC_ALL);
574                 /* See comment of filemap_write_and_wait() */
575                 if (err != -EIO) {
576                         int err2 = filemap_fdatawait_range(mapping,
577                                                 lstart, lend);
578                         if (!err)
579                                 err = err2;
580                 } else {
581                         /* Clear any previously stored errors */
582                         filemap_check_errors(mapping);
583                 }
584         } else {
585                 err = filemap_check_errors(mapping);
586         }
587         return err;
588 }
589 EXPORT_SYMBOL(filemap_write_and_wait_range);
590
591 void __filemap_set_wb_err(struct address_space *mapping, int err)
592 {
593         errseq_t eseq = errseq_set(&mapping->wb_err, err);
594
595         trace_filemap_set_wb_err(mapping, eseq);
596 }
597 EXPORT_SYMBOL(__filemap_set_wb_err);
598
599 /**
600  * file_check_and_advance_wb_err - report wb error (if any) that was previously
601  *                                 and advance wb_err to current one
602  * @file: struct file on which the error is being reported
603  *
604  * When userland calls fsync (or something like nfsd does the equivalent), we
605  * want to report any writeback errors that occurred since the last fsync (or
606  * since the file was opened if there haven't been any).
607  *
608  * Grab the wb_err from the mapping. If it matches what we have in the file,
609  * then just quickly return 0. The file is all caught up.
610  *
611  * If it doesn't match, then take the mapping value, set the "seen" flag in
612  * it and try to swap it into place. If it works, or another task beat us
613  * to it with the new value, then update the f_wb_err and return the error
614  * portion. The error at this point must be reported via proper channels
615  * (a'la fsync, or NFS COMMIT operation, etc.).
616  *
617  * While we handle mapping->wb_err with atomic operations, the f_wb_err
618  * value is protected by the f_lock since we must ensure that it reflects
619  * the latest value swapped in for this file descriptor.
620  */
621 int file_check_and_advance_wb_err(struct file *file)
622 {
623         int err = 0;
624         errseq_t old = READ_ONCE(file->f_wb_err);
625         struct address_space *mapping = file->f_mapping;
626
627         /* Locklessly handle the common case where nothing has changed */
628         if (errseq_check(&mapping->wb_err, old)) {
629                 /* Something changed, must use slow path */
630                 spin_lock(&file->f_lock);
631                 old = file->f_wb_err;
632                 err = errseq_check_and_advance(&mapping->wb_err,
633                                                 &file->f_wb_err);
634                 trace_file_check_and_advance_wb_err(file, old);
635                 spin_unlock(&file->f_lock);
636         }
637
638         /*
639          * We're mostly using this function as a drop in replacement for
640          * filemap_check_errors. Clear AS_EIO/AS_ENOSPC to emulate the effect
641          * that the legacy code would have had on these flags.
642          */
643         clear_bit(AS_EIO, &mapping->flags);
644         clear_bit(AS_ENOSPC, &mapping->flags);
645         return err;
646 }
647 EXPORT_SYMBOL(file_check_and_advance_wb_err);
648
649 /**
650  * file_write_and_wait_range - write out & wait on a file range
651  * @file:       file pointing to address_space with pages
652  * @lstart:     offset in bytes where the range starts
653  * @lend:       offset in bytes where the range ends (inclusive)
654  *
655  * Write out and wait upon file offsets lstart->lend, inclusive.
656  *
657  * Note that @lend is inclusive (describes the last byte to be written) so
658  * that this function can be used to write to the very end-of-file (end = -1).
659  *
660  * After writing out and waiting on the data, we check and advance the
661  * f_wb_err cursor to the latest value, and return any errors detected there.
662  */
663 int file_write_and_wait_range(struct file *file, loff_t lstart, loff_t lend)
664 {
665         int err = 0, err2;
666         struct address_space *mapping = file->f_mapping;
667
668         if (mapping_needs_writeback(mapping)) {
669                 err = __filemap_fdatawrite_range(mapping, lstart, lend,
670                                                  WB_SYNC_ALL);
671                 /* See comment of filemap_write_and_wait() */
672                 if (err != -EIO)
673                         __filemap_fdatawait_range(mapping, lstart, lend);
674         }
675         err2 = file_check_and_advance_wb_err(file);
676         if (!err)
677                 err = err2;
678         return err;
679 }
680 EXPORT_SYMBOL(file_write_and_wait_range);
681
682 /**
683  * replace_page_cache_page - replace a pagecache page with a new one
684  * @old:        page to be replaced
685  * @new:        page to replace with
686  * @gfp_mask:   allocation mode
687  *
688  * This function replaces a page in the pagecache with a new one.  On
689  * success it acquires the pagecache reference for the new page and
690  * drops it for the old page.  Both the old and new pages must be
691  * locked.  This function does not add the new page to the LRU, the
692  * caller must do that.
693  *
694  * The remove + add is atomic.  The only way this function can fail is
695  * memory allocation failure.
696  */
697 int replace_page_cache_page(struct page *old, struct page *new, gfp_t gfp_mask)
698 {
699         int error;
700
701         VM_BUG_ON_PAGE(!PageLocked(old), old);
702         VM_BUG_ON_PAGE(!PageLocked(new), new);
703         VM_BUG_ON_PAGE(new->mapping, new);
704
705         error = radix_tree_preload(gfp_mask & ~__GFP_HIGHMEM);
706         if (!error) {
707                 struct address_space *mapping = old->mapping;
708                 void (*freepage)(struct page *);
709                 unsigned long flags;
710
711                 pgoff_t offset = old->index;
712                 freepage = mapping->a_ops->freepage;
713
714                 get_page(new);
715                 new->mapping = mapping;
716                 new->index = offset;
717
718                 spin_lock_irqsave(&mapping->tree_lock, flags);
719                 __delete_from_page_cache(old, NULL);
720                 error = page_cache_tree_insert(mapping, new, NULL);
721                 BUG_ON(error);
722
723                 /*
724                  * hugetlb pages do not participate in page cache accounting.
725                  */
726                 if (!PageHuge(new))
727                         __inc_node_page_state(new, NR_FILE_PAGES);
728                 if (PageSwapBacked(new))
729                         __inc_node_page_state(new, NR_SHMEM);
730                 spin_unlock_irqrestore(&mapping->tree_lock, flags);
731                 mem_cgroup_migrate(old, new);
732                 radix_tree_preload_end();
733                 if (freepage)
734                         freepage(old);
735                 put_page(old);
736         }
737
738         return error;
739 }
740 EXPORT_SYMBOL_GPL(replace_page_cache_page);
741
742 static int __add_to_page_cache_locked(struct page *page,
743                                       struct address_space *mapping,
744                                       pgoff_t offset, gfp_t gfp_mask,
745                                       void **shadowp)
746 {
747         int huge = PageHuge(page);
748         struct mem_cgroup *memcg;
749         int error;
750
751         VM_BUG_ON_PAGE(!PageLocked(page), page);
752         VM_BUG_ON_PAGE(PageSwapBacked(page), page);
753
754         if (!huge) {
755                 error = mem_cgroup_try_charge(page, current->mm,
756                                               gfp_mask, &memcg, false);
757                 if (error)
758                         return error;
759         }
760
761         error = radix_tree_maybe_preload(gfp_mask & ~__GFP_HIGHMEM);
762         if (error) {
763                 if (!huge)
764                         mem_cgroup_cancel_charge(page, memcg, false);
765                 return error;
766         }
767
768         get_page(page);
769         page->mapping = mapping;
770         page->index = offset;
771
772         spin_lock_irq(&mapping->tree_lock);
773         error = page_cache_tree_insert(mapping, page, shadowp);
774         radix_tree_preload_end();
775         if (unlikely(error))
776                 goto err_insert;
777
778         /* hugetlb pages do not participate in page cache accounting. */
779         if (!huge)
780                 __inc_node_page_state(page, NR_FILE_PAGES);
781         spin_unlock_irq(&mapping->tree_lock);
782         if (!huge)
783                 mem_cgroup_commit_charge(page, memcg, false, false);
784         trace_mm_filemap_add_to_page_cache(page);
785         return 0;
786 err_insert:
787         page->mapping = NULL;
788         /* Leave page->index set: truncation relies upon it */
789         spin_unlock_irq(&mapping->tree_lock);
790         if (!huge)
791                 mem_cgroup_cancel_charge(page, memcg, false);
792         put_page(page);
793         return error;
794 }
795
796 /**
797  * add_to_page_cache_locked - add a locked page to the pagecache
798  * @page:       page to add
799  * @mapping:    the page's address_space
800  * @offset:     page index
801  * @gfp_mask:   page allocation mode
802  *
803  * This function is used to add a page to the pagecache. It must be locked.
804  * This function does not add the page to the LRU.  The caller must do that.
805  */
806 int add_to_page_cache_locked(struct page *page, struct address_space *mapping,
807                 pgoff_t offset, gfp_t gfp_mask)
808 {
809         return __add_to_page_cache_locked(page, mapping, offset,
810                                           gfp_mask, NULL);
811 }
812 EXPORT_SYMBOL(add_to_page_cache_locked);
813
814 int add_to_page_cache_lru(struct page *page, struct address_space *mapping,
815                                 pgoff_t offset, gfp_t gfp_mask)
816 {
817         void *shadow = NULL;
818         int ret;
819
820         __SetPageLocked(page);
821         ret = __add_to_page_cache_locked(page, mapping, offset,
822                                          gfp_mask, &shadow);
823         if (unlikely(ret))
824                 __ClearPageLocked(page);
825         else {
826                 /*
827                  * The page might have been evicted from cache only
828                  * recently, in which case it should be activated like
829                  * any other repeatedly accessed page.
830                  * The exception is pages getting rewritten; evicting other
831                  * data from the working set, only to cache data that will
832                  * get overwritten with something else, is a waste of memory.
833                  */
834                 if (!(gfp_mask & __GFP_WRITE) &&
835                     shadow && workingset_refault(shadow)) {
836                         SetPageActive(page);
837                         workingset_activation(page);
838                 } else
839                         ClearPageActive(page);
840                 lru_cache_add(page);
841         }
842         return ret;
843 }
844 EXPORT_SYMBOL_GPL(add_to_page_cache_lru);
845
846 #ifdef CONFIG_NUMA
847 struct page *__page_cache_alloc(gfp_t gfp)
848 {
849         int n;
850         struct page *page;
851
852         if (cpuset_do_page_mem_spread()) {
853                 unsigned int cpuset_mems_cookie;
854                 do {
855                         cpuset_mems_cookie = read_mems_allowed_begin();
856                         n = cpuset_mem_spread_node();
857                         page = __alloc_pages_node(n, gfp, 0);
858                 } while (!page && read_mems_allowed_retry(cpuset_mems_cookie));
859
860                 return page;
861         }
862         return alloc_pages(gfp, 0);
863 }
864 EXPORT_SYMBOL(__page_cache_alloc);
865 #endif
866
867 /*
868  * In order to wait for pages to become available there must be
869  * waitqueues associated with pages. By using a hash table of
870  * waitqueues where the bucket discipline is to maintain all
871  * waiters on the same queue and wake all when any of the pages
872  * become available, and for the woken contexts to check to be
873  * sure the appropriate page became available, this saves space
874  * at a cost of "thundering herd" phenomena during rare hash
875  * collisions.
876  */
877 #define PAGE_WAIT_TABLE_BITS 8
878 #define PAGE_WAIT_TABLE_SIZE (1 << PAGE_WAIT_TABLE_BITS)
879 static wait_queue_head_t page_wait_table[PAGE_WAIT_TABLE_SIZE] __cacheline_aligned;
880
881 static wait_queue_head_t *page_waitqueue(struct page *page)
882 {
883         return &page_wait_table[hash_ptr(page, PAGE_WAIT_TABLE_BITS)];
884 }
885
886 void __init pagecache_init(void)
887 {
888         int i;
889
890         for (i = 0; i < PAGE_WAIT_TABLE_SIZE; i++)
891                 init_waitqueue_head(&page_wait_table[i]);
892
893         page_writeback_init();
894 }
895
896 /* This has the same layout as wait_bit_key - see fs/cachefiles/rdwr.c */
897 struct wait_page_key {
898         struct page *page;
899         int bit_nr;
900         int page_match;
901 };
902
903 struct wait_page_queue {
904         struct page *page;
905         int bit_nr;
906         wait_queue_entry_t wait;
907 };
908
909 static int wake_page_function(wait_queue_entry_t *wait, unsigned mode, int sync, void *arg)
910 {
911         struct wait_page_key *key = arg;
912         struct wait_page_queue *wait_page
913                 = container_of(wait, struct wait_page_queue, wait);
914
915         if (wait_page->page != key->page)
916                return 0;
917         key->page_match = 1;
918
919         if (wait_page->bit_nr != key->bit_nr)
920                 return 0;
921
922         /* Stop walking if it's locked */
923         if (test_bit(key->bit_nr, &key->page->flags))
924                 return -1;
925
926         return autoremove_wake_function(wait, mode, sync, key);
927 }
928
929 static void wake_up_page_bit(struct page *page, int bit_nr)
930 {
931         wait_queue_head_t *q = page_waitqueue(page);
932         struct wait_page_key key;
933         unsigned long flags;
934         wait_queue_entry_t bookmark;
935
936         key.page = page;
937         key.bit_nr = bit_nr;
938         key.page_match = 0;
939
940         bookmark.flags = 0;
941         bookmark.private = NULL;
942         bookmark.func = NULL;
943         INIT_LIST_HEAD(&bookmark.entry);
944
945         spin_lock_irqsave(&q->lock, flags);
946         __wake_up_locked_key_bookmark(q, TASK_NORMAL, &key, &bookmark);
947
948         while (bookmark.flags & WQ_FLAG_BOOKMARK) {
949                 /*
950                  * Take a breather from holding the lock,
951                  * allow pages that finish wake up asynchronously
952                  * to acquire the lock and remove themselves
953                  * from wait queue
954                  */
955                 spin_unlock_irqrestore(&q->lock, flags);
956                 cpu_relax();
957                 spin_lock_irqsave(&q->lock, flags);
958                 __wake_up_locked_key_bookmark(q, TASK_NORMAL, &key, &bookmark);
959         }
960
961         /*
962          * It is possible for other pages to have collided on the waitqueue
963          * hash, so in that case check for a page match. That prevents a long-
964          * term waiter
965          *
966          * It is still possible to miss a case here, when we woke page waiters
967          * and removed them from the waitqueue, but there are still other
968          * page waiters.
969          */
970         if (!waitqueue_active(q) || !key.page_match) {
971                 ClearPageWaiters(page);
972                 /*
973                  * It's possible to miss clearing Waiters here, when we woke
974                  * our page waiters, but the hashed waitqueue has waiters for
975                  * other pages on it.
976                  *
977                  * That's okay, it's a rare case. The next waker will clear it.
978                  */
979         }
980         spin_unlock_irqrestore(&q->lock, flags);
981 }
982
983 static void wake_up_page(struct page *page, int bit)
984 {
985         if (!PageWaiters(page))
986                 return;
987         wake_up_page_bit(page, bit);
988 }
989
990 static inline int wait_on_page_bit_common(wait_queue_head_t *q,
991                 struct page *page, int bit_nr, int state, bool lock)
992 {
993         struct wait_page_queue wait_page;
994         wait_queue_entry_t *wait = &wait_page.wait;
995         int ret = 0;
996
997         init_wait(wait);
998         wait->flags = lock ? WQ_FLAG_EXCLUSIVE : 0;
999         wait->func = wake_page_function;
1000         wait_page.page = page;
1001         wait_page.bit_nr = bit_nr;
1002
1003         for (;;) {
1004                 spin_lock_irq(&q->lock);
1005
1006                 if (likely(list_empty(&wait->entry))) {
1007                         __add_wait_queue_entry_tail(q, wait);
1008                         SetPageWaiters(page);
1009                 }
1010
1011                 set_current_state(state);
1012
1013                 spin_unlock_irq(&q->lock);
1014
1015                 if (likely(test_bit(bit_nr, &page->flags))) {
1016                         io_schedule();
1017                 }
1018
1019                 if (lock) {
1020                         if (!test_and_set_bit_lock(bit_nr, &page->flags))
1021                                 break;
1022                 } else {
1023                         if (!test_bit(bit_nr, &page->flags))
1024                                 break;
1025                 }
1026
1027                 if (unlikely(signal_pending_state(state, current))) {
1028                         ret = -EINTR;
1029                         break;
1030                 }
1031         }
1032
1033         finish_wait(q, wait);
1034
1035         /*
1036          * A signal could leave PageWaiters set. Clearing it here if
1037          * !waitqueue_active would be possible (by open-coding finish_wait),
1038          * but still fail to catch it in the case of wait hash collision. We
1039          * already can fail to clear wait hash collision cases, so don't
1040          * bother with signals either.
1041          */
1042
1043         return ret;
1044 }
1045
1046 void wait_on_page_bit(struct page *page, int bit_nr)
1047 {
1048         wait_queue_head_t *q = page_waitqueue(page);
1049         wait_on_page_bit_common(q, page, bit_nr, TASK_UNINTERRUPTIBLE, false);
1050 }
1051 EXPORT_SYMBOL(wait_on_page_bit);
1052
1053 int wait_on_page_bit_killable(struct page *page, int bit_nr)
1054 {
1055         wait_queue_head_t *q = page_waitqueue(page);
1056         return wait_on_page_bit_common(q, page, bit_nr, TASK_KILLABLE, false);
1057 }
1058
1059 /**
1060  * add_page_wait_queue - Add an arbitrary waiter to a page's wait queue
1061  * @page: Page defining the wait queue of interest
1062  * @waiter: Waiter to add to the queue
1063  *
1064  * Add an arbitrary @waiter to the wait queue for the nominated @page.
1065  */
1066 void add_page_wait_queue(struct page *page, wait_queue_entry_t *waiter)
1067 {
1068         wait_queue_head_t *q = page_waitqueue(page);
1069         unsigned long flags;
1070
1071         spin_lock_irqsave(&q->lock, flags);
1072         __add_wait_queue_entry_tail(q, waiter);
1073         SetPageWaiters(page);
1074         spin_unlock_irqrestore(&q->lock, flags);
1075 }
1076 EXPORT_SYMBOL_GPL(add_page_wait_queue);
1077
1078 #ifndef clear_bit_unlock_is_negative_byte
1079
1080 /*
1081  * PG_waiters is the high bit in the same byte as PG_lock.
1082  *
1083  * On x86 (and on many other architectures), we can clear PG_lock and
1084  * test the sign bit at the same time. But if the architecture does
1085  * not support that special operation, we just do this all by hand
1086  * instead.
1087  *
1088  * The read of PG_waiters has to be after (or concurrently with) PG_locked
1089  * being cleared, but a memory barrier should be unneccssary since it is
1090  * in the same byte as PG_locked.
1091  */
1092 static inline bool clear_bit_unlock_is_negative_byte(long nr, volatile void *mem)
1093 {
1094         clear_bit_unlock(nr, mem);
1095         /* smp_mb__after_atomic(); */
1096         return test_bit(PG_waiters, mem);
1097 }
1098
1099 #endif
1100
1101 /**
1102  * unlock_page - unlock a locked page
1103  * @page: the page
1104  *
1105  * Unlocks the page and wakes up sleepers in ___wait_on_page_locked().
1106  * Also wakes sleepers in wait_on_page_writeback() because the wakeup
1107  * mechanism between PageLocked pages and PageWriteback pages is shared.
1108  * But that's OK - sleepers in wait_on_page_writeback() just go back to sleep.
1109  *
1110  * Note that this depends on PG_waiters being the sign bit in the byte
1111  * that contains PG_locked - thus the BUILD_BUG_ON(). That allows us to
1112  * clear the PG_locked bit and test PG_waiters at the same time fairly
1113  * portably (architectures that do LL/SC can test any bit, while x86 can
1114  * test the sign bit).
1115  */
1116 void unlock_page(struct page *page)
1117 {
1118         BUILD_BUG_ON(PG_waiters != 7);
1119         page = compound_head(page);
1120         VM_BUG_ON_PAGE(!PageLocked(page), page);
1121         if (clear_bit_unlock_is_negative_byte(PG_locked, &page->flags))
1122                 wake_up_page_bit(page, PG_locked);
1123 }
1124 EXPORT_SYMBOL(unlock_page);
1125
1126 /**
1127  * end_page_writeback - end writeback against a page
1128  * @page: the page
1129  */
1130 void end_page_writeback(struct page *page)
1131 {
1132         /*
1133          * TestClearPageReclaim could be used here but it is an atomic
1134          * operation and overkill in this particular case. Failing to
1135          * shuffle a page marked for immediate reclaim is too mild to
1136          * justify taking an atomic operation penalty at the end of
1137          * ever page writeback.
1138          */
1139         if (PageReclaim(page)) {
1140                 ClearPageReclaim(page);
1141                 rotate_reclaimable_page(page);
1142         }
1143
1144         if (!test_clear_page_writeback(page))
1145                 BUG();
1146
1147         smp_mb__after_atomic();
1148         wake_up_page(page, PG_writeback);
1149 }
1150 EXPORT_SYMBOL(end_page_writeback);
1151
1152 /*
1153  * After completing I/O on a page, call this routine to update the page
1154  * flags appropriately
1155  */
1156 void page_endio(struct page *page, bool is_write, int err)
1157 {
1158         if (!is_write) {
1159                 if (!err) {
1160                         SetPageUptodate(page);
1161                 } else {
1162                         ClearPageUptodate(page);
1163                         SetPageError(page);
1164                 }
1165                 unlock_page(page);
1166         } else {
1167                 if (err) {
1168                         struct address_space *mapping;
1169
1170                         SetPageError(page);
1171                         mapping = page_mapping(page);
1172                         if (mapping)
1173                                 mapping_set_error(mapping, err);
1174                 }
1175                 end_page_writeback(page);
1176         }
1177 }
1178 EXPORT_SYMBOL_GPL(page_endio);
1179
1180 /**
1181  * __lock_page - get a lock on the page, assuming we need to sleep to get it
1182  * @__page: the page to lock
1183  */
1184 void __lock_page(struct page *__page)
1185 {
1186         struct page *page = compound_head(__page);
1187         wait_queue_head_t *q = page_waitqueue(page);
1188         wait_on_page_bit_common(q, page, PG_locked, TASK_UNINTERRUPTIBLE, true);
1189 }
1190 EXPORT_SYMBOL(__lock_page);
1191
1192 int __lock_page_killable(struct page *__page)
1193 {
1194         struct page *page = compound_head(__page);
1195         wait_queue_head_t *q = page_waitqueue(page);
1196         return wait_on_page_bit_common(q, page, PG_locked, TASK_KILLABLE, true);
1197 }
1198 EXPORT_SYMBOL_GPL(__lock_page_killable);
1199
1200 /*
1201  * Return values:
1202  * 1 - page is locked; mmap_sem is still held.
1203  * 0 - page is not locked.
1204  *     mmap_sem has been released (up_read()), unless flags had both
1205  *     FAULT_FLAG_ALLOW_RETRY and FAULT_FLAG_RETRY_NOWAIT set, in
1206  *     which case mmap_sem is still held.
1207  *
1208  * If neither ALLOW_RETRY nor KILLABLE are set, will always return 1
1209  * with the page locked and the mmap_sem unperturbed.
1210  */
1211 int __lock_page_or_retry(struct page *page, struct mm_struct *mm,
1212                          unsigned int flags)
1213 {
1214         if (flags & FAULT_FLAG_ALLOW_RETRY) {
1215                 /*
1216                  * CAUTION! In this case, mmap_sem is not released
1217                  * even though return 0.
1218                  */
1219                 if (flags & FAULT_FLAG_RETRY_NOWAIT)
1220                         return 0;
1221
1222                 up_read(&mm->mmap_sem);
1223                 if (flags & FAULT_FLAG_KILLABLE)
1224                         wait_on_page_locked_killable(page);
1225                 else
1226                         wait_on_page_locked(page);
1227                 return 0;
1228         } else {
1229                 if (flags & FAULT_FLAG_KILLABLE) {
1230                         int ret;
1231
1232                         ret = __lock_page_killable(page);
1233                         if (ret) {
1234                                 up_read(&mm->mmap_sem);
1235                                 return 0;
1236                         }
1237                 } else
1238                         __lock_page(page);
1239                 return 1;
1240         }
1241 }
1242
1243 /**
1244  * page_cache_next_hole - find the next hole (not-present entry)
1245  * @mapping: mapping
1246  * @index: index
1247  * @max_scan: maximum range to search
1248  *
1249  * Search the set [index, min(index+max_scan-1, MAX_INDEX)] for the
1250  * lowest indexed hole.
1251  *
1252  * Returns: the index of the hole if found, otherwise returns an index
1253  * outside of the set specified (in which case 'return - index >=
1254  * max_scan' will be true). In rare cases of index wrap-around, 0 will
1255  * be returned.
1256  *
1257  * page_cache_next_hole may be called under rcu_read_lock. However,
1258  * like radix_tree_gang_lookup, this will not atomically search a
1259  * snapshot of the tree at a single point in time. For example, if a
1260  * hole is created at index 5, then subsequently a hole is created at
1261  * index 10, page_cache_next_hole covering both indexes may return 10
1262  * if called under rcu_read_lock.
1263  */
1264 pgoff_t page_cache_next_hole(struct address_space *mapping,
1265                              pgoff_t index, unsigned long max_scan)
1266 {
1267         unsigned long i;
1268
1269         for (i = 0; i < max_scan; i++) {
1270                 struct page *page;
1271
1272                 page = radix_tree_lookup(&mapping->page_tree, index);
1273                 if (!page || radix_tree_exceptional_entry(page))
1274                         break;
1275                 index++;
1276                 if (index == 0)
1277                         break;
1278         }
1279
1280         return index;
1281 }
1282 EXPORT_SYMBOL(page_cache_next_hole);
1283
1284 /**
1285  * page_cache_prev_hole - find the prev hole (not-present entry)
1286  * @mapping: mapping
1287  * @index: index
1288  * @max_scan: maximum range to search
1289  *
1290  * Search backwards in the range [max(index-max_scan+1, 0), index] for
1291  * the first hole.
1292  *
1293  * Returns: the index of the hole if found, otherwise returns an index
1294  * outside of the set specified (in which case 'index - return >=
1295  * max_scan' will be true). In rare cases of wrap-around, ULONG_MAX
1296  * will be returned.
1297  *
1298  * page_cache_prev_hole may be called under rcu_read_lock. However,
1299  * like radix_tree_gang_lookup, this will not atomically search a
1300  * snapshot of the tree at a single point in time. For example, if a
1301  * hole is created at index 10, then subsequently a hole is created at
1302  * index 5, page_cache_prev_hole covering both indexes may return 5 if
1303  * called under rcu_read_lock.
1304  */
1305 pgoff_t page_cache_prev_hole(struct address_space *mapping,
1306                              pgoff_t index, unsigned long max_scan)
1307 {
1308         unsigned long i;
1309
1310         for (i = 0; i < max_scan; i++) {
1311                 struct page *page;
1312
1313                 page = radix_tree_lookup(&mapping->page_tree, index);
1314                 if (!page || radix_tree_exceptional_entry(page))
1315                         break;
1316                 index--;
1317                 if (index == ULONG_MAX)
1318                         break;
1319         }
1320
1321         return index;
1322 }
1323 EXPORT_SYMBOL(page_cache_prev_hole);
1324
1325 /**
1326  * find_get_entry - find and get a page cache entry
1327  * @mapping: the address_space to search
1328  * @offset: the page cache index
1329  *
1330  * Looks up the page cache slot at @mapping & @offset.  If there is a
1331  * page cache page, it is returned with an increased refcount.
1332  *
1333  * If the slot holds a shadow entry of a previously evicted page, or a
1334  * swap entry from shmem/tmpfs, it is returned.
1335  *
1336  * Otherwise, %NULL is returned.
1337  */
1338 struct page *find_get_entry(struct address_space *mapping, pgoff_t offset)
1339 {
1340         void **pagep;
1341         struct page *head, *page;
1342
1343         rcu_read_lock();
1344 repeat:
1345         page = NULL;
1346         pagep = radix_tree_lookup_slot(&mapping->page_tree, offset);
1347         if (pagep) {
1348                 page = radix_tree_deref_slot(pagep);
1349                 if (unlikely(!page))
1350                         goto out;
1351                 if (radix_tree_exception(page)) {
1352                         if (radix_tree_deref_retry(page))
1353                                 goto repeat;
1354                         /*
1355                          * A shadow entry of a recently evicted page,
1356                          * or a swap entry from shmem/tmpfs.  Return
1357                          * it without attempting to raise page count.
1358                          */
1359                         goto out;
1360                 }
1361
1362                 head = compound_head(page);
1363                 if (!page_cache_get_speculative(head))
1364                         goto repeat;
1365
1366                 /* The page was split under us? */
1367                 if (compound_head(page) != head) {
1368                         put_page(head);
1369                         goto repeat;
1370                 }
1371
1372                 /*
1373                  * Has the page moved?
1374                  * This is part of the lockless pagecache protocol. See
1375                  * include/linux/pagemap.h for details.
1376                  */
1377                 if (unlikely(page != *pagep)) {
1378                         put_page(head);
1379                         goto repeat;
1380                 }
1381         }
1382 out:
1383         rcu_read_unlock();
1384
1385         return page;
1386 }
1387 EXPORT_SYMBOL(find_get_entry);
1388
1389 /**
1390  * find_lock_entry - locate, pin and lock a page cache entry
1391  * @mapping: the address_space to search
1392  * @offset: the page cache index
1393  *
1394  * Looks up the page cache slot at @mapping & @offset.  If there is a
1395  * page cache page, it is returned locked and with an increased
1396  * refcount.
1397  *
1398  * If the slot holds a shadow entry of a previously evicted page, or a
1399  * swap entry from shmem/tmpfs, it is returned.
1400  *
1401  * Otherwise, %NULL is returned.
1402  *
1403  * find_lock_entry() may sleep.
1404  */
1405 struct page *find_lock_entry(struct address_space *mapping, pgoff_t offset)
1406 {
1407         struct page *page;
1408
1409 repeat:
1410         page = find_get_entry(mapping, offset);
1411         if (page && !radix_tree_exception(page)) {
1412                 lock_page(page);
1413                 /* Has the page been truncated? */
1414                 if (unlikely(page_mapping(page) != mapping)) {
1415                         unlock_page(page);
1416                         put_page(page);
1417                         goto repeat;
1418                 }
1419                 VM_BUG_ON_PAGE(page_to_pgoff(page) != offset, page);
1420         }
1421         return page;
1422 }
1423 EXPORT_SYMBOL(find_lock_entry);
1424
1425 /**
1426  * pagecache_get_page - find and get a page reference
1427  * @mapping: the address_space to search
1428  * @offset: the page index
1429  * @fgp_flags: PCG flags
1430  * @gfp_mask: gfp mask to use for the page cache data page allocation
1431  *
1432  * Looks up the page cache slot at @mapping & @offset.
1433  *
1434  * PCG flags modify how the page is returned.
1435  *
1436  * @fgp_flags can be:
1437  *
1438  * - FGP_ACCESSED: the page will be marked accessed
1439  * - FGP_LOCK: Page is return locked
1440  * - FGP_CREAT: If page is not present then a new page is allocated using
1441  *   @gfp_mask and added to the page cache and the VM's LRU
1442  *   list. The page is returned locked and with an increased
1443  *   refcount. Otherwise, NULL is returned.
1444  *
1445  * If FGP_LOCK or FGP_CREAT are specified then the function may sleep even
1446  * if the GFP flags specified for FGP_CREAT are atomic.
1447  *
1448  * If there is a page cache page, it is returned with an increased refcount.
1449  */
1450 struct page *pagecache_get_page(struct address_space *mapping, pgoff_t offset,
1451         int fgp_flags, gfp_t gfp_mask)
1452 {
1453         struct page *page;
1454
1455 repeat:
1456         page = find_get_entry(mapping, offset);
1457         if (radix_tree_exceptional_entry(page))
1458                 page = NULL;
1459         if (!page)
1460                 goto no_page;
1461
1462         if (fgp_flags & FGP_LOCK) {
1463                 if (fgp_flags & FGP_NOWAIT) {
1464                         if (!trylock_page(page)) {
1465                                 put_page(page);
1466                                 return NULL;
1467                         }
1468                 } else {
1469                         lock_page(page);
1470                 }
1471
1472                 /* Has the page been truncated? */
1473                 if (unlikely(page->mapping != mapping)) {
1474                         unlock_page(page);
1475                         put_page(page);
1476                         goto repeat;
1477                 }
1478                 VM_BUG_ON_PAGE(page->index != offset, page);
1479         }
1480
1481         if (page && (fgp_flags & FGP_ACCESSED))
1482                 mark_page_accessed(page);
1483
1484 no_page:
1485         if (!page && (fgp_flags & FGP_CREAT)) {
1486                 int err;
1487                 if ((fgp_flags & FGP_WRITE) && mapping_cap_account_dirty(mapping))
1488                         gfp_mask |= __GFP_WRITE;
1489                 if (fgp_flags & FGP_NOFS)
1490                         gfp_mask &= ~__GFP_FS;
1491
1492                 page = __page_cache_alloc(gfp_mask);
1493                 if (!page)
1494                         return NULL;
1495
1496                 if (WARN_ON_ONCE(!(fgp_flags & FGP_LOCK)))
1497                         fgp_flags |= FGP_LOCK;
1498
1499                 /* Init accessed so avoid atomic mark_page_accessed later */
1500                 if (fgp_flags & FGP_ACCESSED)
1501                         __SetPageReferenced(page);
1502
1503                 err = add_to_page_cache_lru(page, mapping, offset,
1504                                 gfp_mask & GFP_RECLAIM_MASK);
1505                 if (unlikely(err)) {
1506                         put_page(page);
1507                         page = NULL;
1508                         if (err == -EEXIST)
1509                                 goto repeat;
1510                 }
1511         }
1512
1513         return page;
1514 }
1515 EXPORT_SYMBOL(pagecache_get_page);
1516
1517 /**
1518  * find_get_entries - gang pagecache lookup
1519  * @mapping:    The address_space to search
1520  * @start:      The starting page cache index
1521  * @nr_entries: The maximum number of entries
1522  * @entries:    Where the resulting entries are placed
1523  * @indices:    The cache indices corresponding to the entries in @entries
1524  *
1525  * find_get_entries() will search for and return a group of up to
1526  * @nr_entries entries in the mapping.  The entries are placed at
1527  * @entries.  find_get_entries() takes a reference against any actual
1528  * pages it returns.
1529  *
1530  * The search returns a group of mapping-contiguous page cache entries
1531  * with ascending indexes.  There may be holes in the indices due to
1532  * not-present pages.
1533  *
1534  * Any shadow entries of evicted pages, or swap entries from
1535  * shmem/tmpfs, are included in the returned array.
1536  *
1537  * find_get_entries() returns the number of pages and shadow entries
1538  * which were found.
1539  */
1540 unsigned find_get_entries(struct address_space *mapping,
1541                           pgoff_t start, unsigned int nr_entries,
1542                           struct page **entries, pgoff_t *indices)
1543 {
1544         void **slot;
1545         unsigned int ret = 0;
1546         struct radix_tree_iter iter;
1547
1548         if (!nr_entries)
1549                 return 0;
1550
1551         rcu_read_lock();
1552         radix_tree_for_each_slot(slot, &mapping->page_tree, &iter, start) {
1553                 struct page *head, *page;
1554 repeat:
1555                 page = radix_tree_deref_slot(slot);
1556                 if (unlikely(!page))
1557                         continue;
1558                 if (radix_tree_exception(page)) {
1559                         if (radix_tree_deref_retry(page)) {
1560                                 slot = radix_tree_iter_retry(&iter);
1561                                 continue;
1562                         }
1563                         /*
1564                          * A shadow entry of a recently evicted page, a swap
1565                          * entry from shmem/tmpfs or a DAX entry.  Return it
1566                          * without attempting to raise page count.
1567                          */
1568                         goto export;
1569                 }
1570
1571                 head = compound_head(page);
1572                 if (!page_cache_get_speculative(head))
1573                         goto repeat;
1574
1575                 /* The page was split under us? */
1576                 if (compound_head(page) != head) {
1577                         put_page(head);
1578                         goto repeat;
1579                 }
1580
1581                 /* Has the page moved? */
1582                 if (unlikely(page != *slot)) {
1583                         put_page(head);
1584                         goto repeat;
1585                 }
1586 export:
1587                 indices[ret] = iter.index;
1588                 entries[ret] = page;
1589                 if (++ret == nr_entries)
1590                         break;
1591         }
1592         rcu_read_unlock();
1593         return ret;
1594 }
1595
1596 /**
1597  * find_get_pages_range - gang pagecache lookup
1598  * @mapping:    The address_space to search
1599  * @start:      The starting page index
1600  * @end:        The final page index (inclusive)
1601  * @nr_pages:   The maximum number of pages
1602  * @pages:      Where the resulting pages are placed
1603  *
1604  * find_get_pages_range() will search for and return a group of up to @nr_pages
1605  * pages in the mapping starting at index @start and up to index @end
1606  * (inclusive).  The pages are placed at @pages.  find_get_pages_range() takes
1607  * a reference against the returned pages.
1608  *
1609  * The search returns a group of mapping-contiguous pages with ascending
1610  * indexes.  There may be holes in the indices due to not-present pages.
1611  * We also update @start to index the next page for the traversal.
1612  *
1613  * find_get_pages_range() returns the number of pages which were found. If this
1614  * number is smaller than @nr_pages, the end of specified range has been
1615  * reached.
1616  */
1617 unsigned find_get_pages_range(struct address_space *mapping, pgoff_t *start,
1618                               pgoff_t end, unsigned int nr_pages,
1619                               struct page **pages)
1620 {
1621         struct radix_tree_iter iter;
1622         void **slot;
1623         unsigned ret = 0;
1624
1625         if (unlikely(!nr_pages))
1626                 return 0;
1627
1628         rcu_read_lock();
1629         radix_tree_for_each_slot(slot, &mapping->page_tree, &iter, *start) {
1630                 struct page *head, *page;
1631
1632                 if (iter.index > end)
1633                         break;
1634 repeat:
1635                 page = radix_tree_deref_slot(slot);
1636                 if (unlikely(!page))
1637                         continue;
1638
1639                 if (radix_tree_exception(page)) {
1640                         if (radix_tree_deref_retry(page)) {
1641                                 slot = radix_tree_iter_retry(&iter);
1642                                 continue;
1643                         }
1644                         /*
1645                          * A shadow entry of a recently evicted page,
1646                          * or a swap entry from shmem/tmpfs.  Skip
1647                          * over it.
1648                          */
1649                         continue;
1650                 }
1651
1652                 head = compound_head(page);
1653                 if (!page_cache_get_speculative(head))
1654                         goto repeat;
1655
1656                 /* The page was split under us? */
1657                 if (compound_head(page) != head) {
1658                         put_page(head);
1659                         goto repeat;
1660                 }
1661
1662                 /* Has the page moved? */
1663                 if (unlikely(page != *slot)) {
1664                         put_page(head);
1665                         goto repeat;
1666                 }
1667
1668                 pages[ret] = page;
1669                 if (++ret == nr_pages) {
1670                         *start = pages[ret - 1]->index + 1;
1671                         goto out;
1672                 }
1673         }
1674
1675         /*
1676          * We come here when there is no page beyond @end. We take care to not
1677          * overflow the index @start as it confuses some of the callers. This
1678          * breaks the iteration when there is page at index -1 but that is
1679          * already broken anyway.
1680          */
1681         if (end == (pgoff_t)-1)
1682                 *start = (pgoff_t)-1;
1683         else
1684                 *start = end + 1;
1685 out:
1686         rcu_read_unlock();
1687
1688         return ret;
1689 }
1690
1691 /**
1692  * find_get_pages_contig - gang contiguous pagecache lookup
1693  * @mapping:    The address_space to search
1694  * @index:      The starting page index
1695  * @nr_pages:   The maximum number of pages
1696  * @pages:      Where the resulting pages are placed
1697  *
1698  * find_get_pages_contig() works exactly like find_get_pages(), except
1699  * that the returned number of pages are guaranteed to be contiguous.
1700  *
1701  * find_get_pages_contig() returns the number of pages which were found.
1702  */
1703 unsigned find_get_pages_contig(struct address_space *mapping, pgoff_t index,
1704                                unsigned int nr_pages, struct page **pages)
1705 {
1706         struct radix_tree_iter iter;
1707         void **slot;
1708         unsigned int ret = 0;
1709
1710         if (unlikely(!nr_pages))
1711                 return 0;
1712
1713         rcu_read_lock();
1714         radix_tree_for_each_contig(slot, &mapping->page_tree, &iter, index) {
1715                 struct page *head, *page;
1716 repeat:
1717                 page = radix_tree_deref_slot(slot);
1718                 /* The hole, there no reason to continue */
1719                 if (unlikely(!page))
1720                         break;
1721
1722                 if (radix_tree_exception(page)) {
1723                         if (radix_tree_deref_retry(page)) {
1724                                 slot = radix_tree_iter_retry(&iter);
1725                                 continue;
1726                         }
1727                         /*
1728                          * A shadow entry of a recently evicted page,
1729                          * or a swap entry from shmem/tmpfs.  Stop
1730                          * looking for contiguous pages.
1731                          */
1732                         break;
1733                 }
1734
1735                 head = compound_head(page);
1736                 if (!page_cache_get_speculative(head))
1737                         goto repeat;
1738
1739                 /* The page was split under us? */
1740                 if (compound_head(page) != head) {
1741                         put_page(head);
1742                         goto repeat;
1743                 }
1744
1745                 /* Has the page moved? */
1746                 if (unlikely(page != *slot)) {
1747                         put_page(head);
1748                         goto repeat;
1749                 }
1750
1751                 /*
1752                  * must check mapping and index after taking the ref.
1753                  * otherwise we can get both false positives and false
1754                  * negatives, which is just confusing to the caller.
1755                  */
1756                 if (page->mapping == NULL || page_to_pgoff(page) != iter.index) {
1757                         put_page(page);
1758                         break;
1759                 }
1760
1761                 pages[ret] = page;
1762                 if (++ret == nr_pages)
1763                         break;
1764         }
1765         rcu_read_unlock();
1766         return ret;
1767 }
1768 EXPORT_SYMBOL(find_get_pages_contig);
1769
1770 /**
1771  * find_get_pages_range_tag - find and return pages in given range matching @tag
1772  * @mapping:    the address_space to search
1773  * @index:      the starting page index
1774  * @end:        The final page index (inclusive)
1775  * @tag:        the tag index
1776  * @nr_pages:   the maximum number of pages
1777  * @pages:      where the resulting pages are placed
1778  *
1779  * Like find_get_pages, except we only return pages which are tagged with
1780  * @tag.   We update @index to index the next page for the traversal.
1781  */
1782 unsigned find_get_pages_range_tag(struct address_space *mapping, pgoff_t *index,
1783                         pgoff_t end, int tag, unsigned int nr_pages,
1784                         struct page **pages)
1785 {
1786         struct radix_tree_iter iter;
1787         void **slot;
1788         unsigned ret = 0;
1789
1790         if (unlikely(!nr_pages))
1791                 return 0;
1792
1793         rcu_read_lock();
1794         radix_tree_for_each_tagged(slot, &mapping->page_tree,
1795                                    &iter, *index, tag) {
1796                 struct page *head, *page;
1797
1798                 if (iter.index > end)
1799                         break;
1800 repeat:
1801                 page = radix_tree_deref_slot(slot);
1802                 if (unlikely(!page))
1803                         continue;
1804
1805                 if (radix_tree_exception(page)) {
1806                         if (radix_tree_deref_retry(page)) {
1807                                 slot = radix_tree_iter_retry(&iter);
1808                                 continue;
1809                         }
1810                         /*
1811                          * A shadow entry of a recently evicted page.
1812                          *
1813                          * Those entries should never be tagged, but
1814                          * this tree walk is lockless and the tags are
1815                          * looked up in bulk, one radix tree node at a
1816                          * time, so there is a sizable window for page
1817                          * reclaim to evict a page we saw tagged.
1818                          *
1819                          * Skip over it.
1820                          */
1821                         continue;
1822                 }
1823
1824                 head = compound_head(page);
1825                 if (!page_cache_get_speculative(head))
1826                         goto repeat;
1827
1828                 /* The page was split under us? */
1829                 if (compound_head(page) != head) {
1830                         put_page(head);
1831                         goto repeat;
1832                 }
1833
1834                 /* Has the page moved? */
1835                 if (unlikely(page != *slot)) {
1836                         put_page(head);
1837                         goto repeat;
1838                 }
1839
1840                 pages[ret] = page;
1841                 if (++ret == nr_pages) {
1842                         *index = pages[ret - 1]->index + 1;
1843                         goto out;
1844                 }
1845         }
1846
1847         /*
1848          * We come here when we got at @end. We take care to not overflow the
1849          * index @index as it confuses some of the callers. This breaks the
1850          * iteration when there is page at index -1 but that is already broken
1851          * anyway.
1852          */
1853         if (end == (pgoff_t)-1)
1854                 *index = (pgoff_t)-1;
1855         else
1856                 *index = end + 1;
1857 out:
1858         rcu_read_unlock();
1859
1860         return ret;
1861 }
1862 EXPORT_SYMBOL(find_get_pages_range_tag);
1863
1864 /**
1865  * find_get_entries_tag - find and return entries that match @tag
1866  * @mapping:    the address_space to search
1867  * @start:      the starting page cache index
1868  * @tag:        the tag index
1869  * @nr_entries: the maximum number of entries
1870  * @entries:    where the resulting entries are placed
1871  * @indices:    the cache indices corresponding to the entries in @entries
1872  *
1873  * Like find_get_entries, except we only return entries which are tagged with
1874  * @tag.
1875  */
1876 unsigned find_get_entries_tag(struct address_space *mapping, pgoff_t start,
1877                         int tag, unsigned int nr_entries,
1878                         struct page **entries, pgoff_t *indices)
1879 {
1880         void **slot;
1881         unsigned int ret = 0;
1882         struct radix_tree_iter iter;
1883
1884         if (!nr_entries)
1885                 return 0;
1886
1887         rcu_read_lock();
1888         radix_tree_for_each_tagged(slot, &mapping->page_tree,
1889                                    &iter, start, tag) {
1890                 struct page *head, *page;
1891 repeat:
1892                 page = radix_tree_deref_slot(slot);
1893                 if (unlikely(!page))
1894                         continue;
1895                 if (radix_tree_exception(page)) {
1896                         if (radix_tree_deref_retry(page)) {
1897                                 slot = radix_tree_iter_retry(&iter);
1898                                 continue;
1899                         }
1900
1901                         /*
1902                          * A shadow entry of a recently evicted page, a swap
1903                          * entry from shmem/tmpfs or a DAX entry.  Return it
1904                          * without attempting to raise page count.
1905                          */
1906                         goto export;
1907                 }
1908
1909                 head = compound_head(page);
1910                 if (!page_cache_get_speculative(head))
1911                         goto repeat;
1912
1913                 /* The page was split under us? */
1914                 if (compound_head(page) != head) {
1915                         put_page(head);
1916                         goto repeat;
1917                 }
1918
1919                 /* Has the page moved? */
1920                 if (unlikely(page != *slot)) {
1921                         put_page(head);
1922                         goto repeat;
1923                 }
1924 export:
1925                 indices[ret] = iter.index;
1926                 entries[ret] = page;
1927                 if (++ret == nr_entries)
1928                         break;
1929         }
1930         rcu_read_unlock();
1931         return ret;
1932 }
1933 EXPORT_SYMBOL(find_get_entries_tag);
1934
1935 /*
1936  * CD/DVDs are error prone. When a medium error occurs, the driver may fail
1937  * a _large_ part of the i/o request. Imagine the worst scenario:
1938  *
1939  *      ---R__________________________________________B__________
1940  *         ^ reading here                             ^ bad block(assume 4k)
1941  *
1942  * read(R) => miss => readahead(R...B) => media error => frustrating retries
1943  * => failing the whole request => read(R) => read(R+1) =>
1944  * readahead(R+1...B+1) => bang => read(R+2) => read(R+3) =>
1945  * readahead(R+3...B+2) => bang => read(R+3) => read(R+4) =>
1946  * readahead(R+4...B+3) => bang => read(R+4) => read(R+5) => ......
1947  *
1948  * It is going insane. Fix it by quickly scaling down the readahead size.
1949  */
1950 static void shrink_readahead_size_eio(struct file *filp,
1951                                         struct file_ra_state *ra)
1952 {
1953         ra->ra_pages /= 4;
1954 }
1955
1956 /**
1957  * generic_file_buffered_read - generic file read routine
1958  * @iocb:       the iocb to read
1959  * @iter:       data destination
1960  * @written:    already copied
1961  *
1962  * This is a generic file read routine, and uses the
1963  * mapping->a_ops->readpage() function for the actual low-level stuff.
1964  *
1965  * This is really ugly. But the goto's actually try to clarify some
1966  * of the logic when it comes to error handling etc.
1967  */
1968 static ssize_t generic_file_buffered_read(struct kiocb *iocb,
1969                 struct iov_iter *iter, ssize_t written)
1970 {
1971         struct file *filp = iocb->ki_filp;
1972         struct address_space *mapping = filp->f_mapping;
1973         struct inode *inode = mapping->host;
1974         struct file_ra_state *ra = &filp->f_ra;
1975         loff_t *ppos = &iocb->ki_pos;
1976         pgoff_t index;
1977         pgoff_t last_index;
1978         pgoff_t prev_index;
1979         unsigned long offset;      /* offset into pagecache page */
1980         unsigned int prev_offset;
1981         int error = 0;
1982
1983         if (unlikely(*ppos >= inode->i_sb->s_maxbytes))
1984                 return 0;
1985         iov_iter_truncate(iter, inode->i_sb->s_maxbytes);
1986
1987         index = *ppos >> PAGE_SHIFT;
1988         prev_index = ra->prev_pos >> PAGE_SHIFT;
1989         prev_offset = ra->prev_pos & (PAGE_SIZE-1);
1990         last_index = (*ppos + iter->count + PAGE_SIZE-1) >> PAGE_SHIFT;
1991         offset = *ppos & ~PAGE_MASK;
1992
1993         for (;;) {
1994                 struct page *page;
1995                 pgoff_t end_index;
1996                 loff_t isize;
1997                 unsigned long nr, ret;
1998
1999                 cond_resched();
2000 find_page:
2001                 if (fatal_signal_pending(current)) {
2002                         error = -EINTR;
2003                         goto out;
2004                 }
2005
2006                 page = find_get_page(mapping, index);
2007                 if (!page) {
2008                         if (iocb->ki_flags & IOCB_NOWAIT)
2009                                 goto would_block;
2010                         page_cache_sync_readahead(mapping,
2011                                         ra, filp,
2012                                         index, last_index - index);
2013                         page = find_get_page(mapping, index);
2014                         if (unlikely(page == NULL))
2015                                 goto no_cached_page;
2016                 }
2017                 if (PageReadahead(page)) {
2018                         page_cache_async_readahead(mapping,
2019                                         ra, filp, page,
2020                                         index, last_index - index);
2021                 }
2022                 if (!PageUptodate(page)) {
2023                         if (iocb->ki_flags & IOCB_NOWAIT) {
2024                                 put_page(page);
2025                                 goto would_block;
2026                         }
2027
2028                         /*
2029                          * See comment in do_read_cache_page on why
2030                          * wait_on_page_locked is used to avoid unnecessarily
2031                          * serialisations and why it's safe.
2032                          */
2033                         error = wait_on_page_locked_killable(page);
2034                         if (unlikely(error))
2035                                 goto readpage_error;
2036                         if (PageUptodate(page))
2037                                 goto page_ok;
2038
2039                         if (inode->i_blkbits == PAGE_SHIFT ||
2040                                         !mapping->a_ops->is_partially_uptodate)
2041                                 goto page_not_up_to_date;
2042                         /* pipes can't handle partially uptodate pages */
2043                         if (unlikely(iter->type & ITER_PIPE))
2044                                 goto page_not_up_to_date;
2045                         if (!trylock_page(page))
2046                                 goto page_not_up_to_date;
2047                         /* Did it get truncated before we got the lock? */
2048                         if (!page->mapping)
2049                                 goto page_not_up_to_date_locked;
2050                         if (!mapping->a_ops->is_partially_uptodate(page,
2051                                                         offset, iter->count))
2052                                 goto page_not_up_to_date_locked;
2053                         unlock_page(page);
2054                 }
2055 page_ok:
2056                 /*
2057                  * i_size must be checked after we know the page is Uptodate.
2058                  *
2059                  * Checking i_size after the check allows us to calculate
2060                  * the correct value for "nr", which means the zero-filled
2061                  * part of the page is not copied back to userspace (unless
2062                  * another truncate extends the file - this is desired though).
2063                  */
2064
2065                 isize = i_size_read(inode);
2066                 end_index = (isize - 1) >> PAGE_SHIFT;
2067                 if (unlikely(!isize || index > end_index)) {
2068                         put_page(page);
2069                         goto out;
2070                 }
2071
2072                 /* nr is the maximum number of bytes to copy from this page */
2073                 nr = PAGE_SIZE;
2074                 if (index == end_index) {
2075                         nr = ((isize - 1) & ~PAGE_MASK) + 1;
2076                         if (nr <= offset) {
2077                                 put_page(page);
2078                                 goto out;
2079                         }
2080                 }
2081                 nr = nr - offset;
2082
2083                 /* If users can be writing to this page using arbitrary
2084                  * virtual addresses, take care about potential aliasing
2085                  * before reading the page on the kernel side.
2086                  */
2087                 if (mapping_writably_mapped(mapping))
2088                         flush_dcache_page(page);
2089
2090                 /*
2091                  * When a sequential read accesses a page several times,
2092                  * only mark it as accessed the first time.
2093                  */
2094                 if (prev_index != index || offset != prev_offset)
2095                         mark_page_accessed(page);
2096                 prev_index = index;
2097
2098                 /*
2099                  * Ok, we have the page, and it's up-to-date, so
2100                  * now we can copy it to user space...
2101                  */
2102
2103                 ret = copy_page_to_iter(page, offset, nr, iter);
2104                 offset += ret;
2105                 index += offset >> PAGE_SHIFT;
2106                 offset &= ~PAGE_MASK;
2107                 prev_offset = offset;
2108
2109                 put_page(page);
2110                 written += ret;
2111                 if (!iov_iter_count(iter))
2112                         goto out;
2113                 if (ret < nr) {
2114                         error = -EFAULT;
2115                         goto out;
2116                 }
2117                 continue;
2118
2119 page_not_up_to_date:
2120                 /* Get exclusive access to the page ... */
2121                 error = lock_page_killable(page);
2122                 if (unlikely(error))
2123                         goto readpage_error;
2124
2125 page_not_up_to_date_locked:
2126                 /* Did it get truncated before we got the lock? */
2127                 if (!page->mapping) {
2128                         unlock_page(page);
2129                         put_page(page);
2130                         continue;
2131                 }
2132
2133                 /* Did somebody else fill it already? */
2134                 if (PageUptodate(page)) {
2135                         unlock_page(page);
2136                         goto page_ok;
2137                 }
2138
2139 readpage:
2140                 /*
2141                  * A previous I/O error may have been due to temporary
2142                  * failures, eg. multipath errors.
2143                  * PG_error will be set again if readpage fails.
2144                  */
2145                 ClearPageError(page);
2146                 /* Start the actual read. The read will unlock the page. */
2147                 error = mapping->a_ops->readpage(filp, page);
2148
2149                 if (unlikely(error)) {
2150                         if (error == AOP_TRUNCATED_PAGE) {
2151                                 put_page(page);
2152                                 error = 0;
2153                                 goto find_page;
2154                         }
2155                         goto readpage_error;
2156                 }
2157
2158                 if (!PageUptodate(page)) {
2159                         error = lock_page_killable(page);
2160                         if (unlikely(error))
2161                                 goto readpage_error;
2162                         if (!PageUptodate(page)) {
2163                                 if (page->mapping == NULL) {
2164                                         /*
2165                                          * invalidate_mapping_pages got it
2166                                          */
2167                                         unlock_page(page);
2168                                         put_page(page);
2169                                         goto find_page;
2170                                 }
2171                                 unlock_page(page);
2172                                 shrink_readahead_size_eio(filp, ra);
2173                                 error = -EIO;
2174                                 goto readpage_error;
2175                         }
2176                         unlock_page(page);
2177                 }
2178
2179                 goto page_ok;
2180
2181 readpage_error:
2182                 /* UHHUH! A synchronous read error occurred. Report it */
2183                 put_page(page);
2184                 goto out;
2185
2186 no_cached_page:
2187                 /*
2188                  * Ok, it wasn't cached, so we need to create a new
2189                  * page..
2190                  */
2191                 page = page_cache_alloc_cold(mapping);
2192                 if (!page) {
2193                         error = -ENOMEM;
2194                         goto out;
2195                 }
2196                 error = add_to_page_cache_lru(page, mapping, index,
2197                                 mapping_gfp_constraint(mapping, GFP_KERNEL));
2198                 if (error) {
2199                         put_page(page);
2200                         if (error == -EEXIST) {
2201                                 error = 0;
2202                                 goto find_page;
2203                         }
2204                         goto out;
2205                 }
2206                 goto readpage;
2207         }
2208
2209 would_block:
2210         error = -EAGAIN;
2211 out:
2212         ra->prev_pos = prev_index;
2213         ra->prev_pos <<= PAGE_SHIFT;
2214         ra->prev_pos |= prev_offset;
2215
2216         *ppos = ((loff_t)index << PAGE_SHIFT) + offset;
2217         file_accessed(filp);
2218         return written ? written : error;
2219 }
2220
2221 /**
2222  * generic_file_read_iter - generic filesystem read routine
2223  * @iocb:       kernel I/O control block
2224  * @iter:       destination for the data read
2225  *
2226  * This is the "read_iter()" routine for all filesystems
2227  * that can use the page cache directly.
2228  */
2229 ssize_t
2230 generic_file_read_iter(struct kiocb *iocb, struct iov_iter *iter)
2231 {
2232         size_t count = iov_iter_count(iter);
2233         ssize_t retval = 0;
2234
2235         if (!count)
2236                 goto out; /* skip atime */
2237
2238         if (iocb->ki_flags & IOCB_DIRECT) {
2239                 struct file *file = iocb->ki_filp;
2240                 struct address_space *mapping = file->f_mapping;
2241                 struct inode *inode = mapping->host;
2242                 loff_t size;
2243
2244                 size = i_size_read(inode);
2245                 if (iocb->ki_flags & IOCB_NOWAIT) {
2246                         if (filemap_range_has_page(mapping, iocb->ki_pos,
2247                                                    iocb->ki_pos + count - 1))
2248                                 return -EAGAIN;
2249                 } else {
2250                         retval = filemap_write_and_wait_range(mapping,
2251                                                 iocb->ki_pos,
2252                                                 iocb->ki_pos + count - 1);
2253                         if (retval < 0)
2254                                 goto out;
2255                 }
2256
2257                 file_accessed(file);
2258
2259                 retval = mapping->a_ops->direct_IO(iocb, iter);
2260                 if (retval >= 0) {
2261                         iocb->ki_pos += retval;
2262                         count -= retval;
2263                 }
2264                 iov_iter_revert(iter, count - iov_iter_count(iter));
2265
2266                 /*
2267                  * Btrfs can have a short DIO read if we encounter
2268                  * compressed extents, so if there was an error, or if
2269                  * we've already read everything we wanted to, or if
2270                  * there was a short read because we hit EOF, go ahead
2271                  * and return.  Otherwise fallthrough to buffered io for
2272                  * the rest of the read.  Buffered reads will not work for
2273                  * DAX files, so don't bother trying.
2274                  */
2275                 if (retval < 0 || !count || iocb->ki_pos >= size ||
2276                     IS_DAX(inode))
2277                         goto out;
2278         }
2279
2280         retval = generic_file_buffered_read(iocb, iter, retval);
2281 out:
2282         return retval;
2283 }
2284 EXPORT_SYMBOL(generic_file_read_iter);
2285
2286 #ifdef CONFIG_MMU
2287 /**
2288  * page_cache_read - adds requested page to the page cache if not already there
2289  * @file:       file to read
2290  * @offset:     page index
2291  * @gfp_mask:   memory allocation flags
2292  *
2293  * This adds the requested page to the page cache if it isn't already there,
2294  * and schedules an I/O to read in its contents from disk.
2295  */
2296 static int page_cache_read(struct file *file, pgoff_t offset, gfp_t gfp_mask)
2297 {
2298         struct address_space *mapping = file->f_mapping;
2299         struct page *page;
2300         int ret;
2301
2302         do {
2303                 page = __page_cache_alloc(gfp_mask|__GFP_COLD);
2304                 if (!page)
2305                         return -ENOMEM;
2306
2307                 ret = add_to_page_cache_lru(page, mapping, offset, gfp_mask & GFP_KERNEL);
2308                 if (ret == 0)
2309                         ret = mapping->a_ops->readpage(file, page);
2310                 else if (ret == -EEXIST)
2311                         ret = 0; /* losing race to add is OK */
2312
2313                 put_page(page);
2314
2315         } while (ret == AOP_TRUNCATED_PAGE);
2316
2317         return ret;
2318 }
2319
2320 #define MMAP_LOTSAMISS  (100)
2321
2322 /*
2323  * Synchronous readahead happens when we don't even find
2324  * a page in the page cache at all.
2325  */
2326 static void do_sync_mmap_readahead(struct vm_area_struct *vma,
2327                                    struct file_ra_state *ra,
2328                                    struct file *file,
2329                                    pgoff_t offset)
2330 {
2331         struct address_space *mapping = file->f_mapping;
2332
2333         /* If we don't want any read-ahead, don't bother */
2334         if (vma->vm_flags & VM_RAND_READ)
2335                 return;
2336         if (!ra->ra_pages)
2337                 return;
2338
2339         if (vma->vm_flags & VM_SEQ_READ) {
2340                 page_cache_sync_readahead(mapping, ra, file, offset,
2341                                           ra->ra_pages);
2342                 return;
2343         }
2344
2345         /* Avoid banging the cache line if not needed */
2346         if (ra->mmap_miss < MMAP_LOTSAMISS * 10)
2347                 ra->mmap_miss++;
2348
2349         /*
2350          * Do we miss much more than hit in this file? If so,
2351          * stop bothering with read-ahead. It will only hurt.
2352          */
2353         if (ra->mmap_miss > MMAP_LOTSAMISS)
2354                 return;
2355
2356         /*
2357          * mmap read-around
2358          */
2359         ra->start = max_t(long, 0, offset - ra->ra_pages / 2);
2360         ra->size = ra->ra_pages;
2361         ra->async_size = ra->ra_pages / 4;
2362         ra_submit(ra, mapping, file);
2363 }
2364
2365 /*
2366  * Asynchronous readahead happens when we find the page and PG_readahead,
2367  * so we want to possibly extend the readahead further..
2368  */
2369 static void do_async_mmap_readahead(struct vm_area_struct *vma,
2370                                     struct file_ra_state *ra,
2371                                     struct file *file,
2372                                     struct page *page,
2373                                     pgoff_t offset)
2374 {
2375         struct address_space *mapping = file->f_mapping;
2376
2377         /* If we don't want any read-ahead, don't bother */
2378         if (vma->vm_flags & VM_RAND_READ)
2379                 return;
2380         if (ra->mmap_miss > 0)
2381                 ra->mmap_miss--;
2382         if (PageReadahead(page))
2383                 page_cache_async_readahead(mapping, ra, file,
2384                                            page, offset, ra->ra_pages);
2385 }
2386
2387 /**
2388  * filemap_fault - read in file data for page fault handling
2389  * @vmf:        struct vm_fault containing details of the fault
2390  *
2391  * filemap_fault() is invoked via the vma operations vector for a
2392  * mapped memory region to read in file data during a page fault.
2393  *
2394  * The goto's are kind of ugly, but this streamlines the normal case of having
2395  * it in the page cache, and handles the special cases reasonably without
2396  * having a lot of duplicated code.
2397  *
2398  * vma->vm_mm->mmap_sem must be held on entry.
2399  *
2400  * If our return value has VM_FAULT_RETRY set, it's because
2401  * lock_page_or_retry() returned 0.
2402  * The mmap_sem has usually been released in this case.
2403  * See __lock_page_or_retry() for the exception.
2404  *
2405  * If our return value does not have VM_FAULT_RETRY set, the mmap_sem
2406  * has not been released.
2407  *
2408  * We never return with VM_FAULT_RETRY and a bit from VM_FAULT_ERROR set.
2409  */
2410 int filemap_fault(struct vm_fault *vmf)
2411 {
2412         int error;
2413         struct file *file = vmf->vma->vm_file;
2414         struct address_space *mapping = file->f_mapping;
2415         struct file_ra_state *ra = &file->f_ra;
2416         struct inode *inode = mapping->host;
2417         pgoff_t offset = vmf->pgoff;
2418         pgoff_t max_off;
2419         struct page *page;
2420         int ret = 0;
2421
2422         max_off = DIV_ROUND_UP(i_size_read(inode), PAGE_SIZE);
2423         if (unlikely(offset >= max_off))
2424                 return VM_FAULT_SIGBUS;
2425
2426         /*
2427          * Do we have something in the page cache already?
2428          */
2429         page = find_get_page(mapping, offset);
2430         if (likely(page) && !(vmf->flags & FAULT_FLAG_TRIED)) {
2431                 /*
2432                  * We found the page, so try async readahead before
2433                  * waiting for the lock.
2434                  */
2435                 do_async_mmap_readahead(vmf->vma, ra, file, page, offset);
2436         } else if (!page) {
2437                 /* No page in the page cache at all */
2438                 do_sync_mmap_readahead(vmf->vma, ra, file, offset);
2439                 count_vm_event(PGMAJFAULT);
2440                 count_memcg_event_mm(vmf->vma->vm_mm, PGMAJFAULT);
2441                 ret = VM_FAULT_MAJOR;
2442 retry_find:
2443                 page = find_get_page(mapping, offset);
2444                 if (!page)
2445                         goto no_cached_page;
2446         }
2447
2448         if (!lock_page_or_retry(page, vmf->vma->vm_mm, vmf->flags)) {
2449                 put_page(page);
2450                 return ret | VM_FAULT_RETRY;
2451         }
2452
2453         /* Did it get truncated? */
2454         if (unlikely(page->mapping != mapping)) {
2455                 unlock_page(page);
2456                 put_page(page);
2457                 goto retry_find;
2458         }
2459         VM_BUG_ON_PAGE(page->index != offset, page);
2460
2461         /*
2462          * We have a locked page in the page cache, now we need to check
2463          * that it's up-to-date. If not, it is going to be due to an error.
2464          */
2465         if (unlikely(!PageUptodate(page)))
2466                 goto page_not_uptodate;
2467
2468         /*
2469          * Found the page and have a reference on it.
2470          * We must recheck i_size under page lock.
2471          */
2472         max_off = DIV_ROUND_UP(i_size_read(inode), PAGE_SIZE);
2473         if (unlikely(offset >= max_off)) {
2474                 unlock_page(page);
2475                 put_page(page);
2476                 return VM_FAULT_SIGBUS;
2477         }
2478
2479         vmf->page = page;
2480         return ret | VM_FAULT_LOCKED;
2481
2482 no_cached_page:
2483         /*
2484          * We're only likely to ever get here if MADV_RANDOM is in
2485          * effect.
2486          */
2487         error = page_cache_read(file, offset, vmf->gfp_mask);
2488
2489         /*
2490          * The page we want has now been added to the page cache.
2491          * In the unlikely event that someone removed it in the
2492          * meantime, we'll just come back here and read it again.
2493          */
2494         if (error >= 0)
2495                 goto retry_find;
2496
2497         /*
2498          * An error return from page_cache_read can result if the
2499          * system is low on memory, or a problem occurs while trying
2500          * to schedule I/O.
2501          */
2502         if (error == -ENOMEM)
2503                 return VM_FAULT_OOM;
2504         return VM_FAULT_SIGBUS;
2505
2506 page_not_uptodate:
2507         /*
2508          * Umm, take care of errors if the page isn't up-to-date.
2509          * Try to re-read it _once_. We do this synchronously,
2510          * because there really aren't any performance issues here
2511          * and we need to check for errors.
2512          */
2513         ClearPageError(page);
2514         error = mapping->a_ops->readpage(file, page);
2515         if (!error) {
2516                 wait_on_page_locked(page);
2517                 if (!PageUptodate(page))
2518                         error = -EIO;
2519         }
2520         put_page(page);
2521
2522         if (!error || error == AOP_TRUNCATED_PAGE)
2523                 goto retry_find;
2524
2525         /* Things didn't work out. Return zero to tell the mm layer so. */
2526         shrink_readahead_size_eio(file, ra);
2527         return VM_FAULT_SIGBUS;
2528 }
2529 EXPORT_SYMBOL(filemap_fault);
2530
2531 void filemap_map_pages(struct vm_fault *vmf,
2532                 pgoff_t start_pgoff, pgoff_t end_pgoff)
2533 {
2534         struct radix_tree_iter iter;
2535         void **slot;
2536         struct file *file = vmf->vma->vm_file;
2537         struct address_space *mapping = file->f_mapping;
2538         pgoff_t last_pgoff = start_pgoff;
2539         unsigned long max_idx;
2540         struct page *head, *page;
2541
2542         rcu_read_lock();
2543         radix_tree_for_each_slot(slot, &mapping->page_tree, &iter,
2544                         start_pgoff) {
2545                 if (iter.index > end_pgoff)
2546                         break;
2547 repeat:
2548                 page = radix_tree_deref_slot(slot);
2549                 if (unlikely(!page))
2550                         goto next;
2551                 if (radix_tree_exception(page)) {
2552                         if (radix_tree_deref_retry(page)) {
2553                                 slot = radix_tree_iter_retry(&iter);
2554                                 continue;
2555                         }
2556                         goto next;
2557                 }
2558
2559                 head = compound_head(page);
2560                 if (!page_cache_get_speculative(head))
2561                         goto repeat;
2562
2563                 /* The page was split under us? */
2564                 if (compound_head(page) != head) {
2565                         put_page(head);
2566                         goto repeat;
2567                 }
2568
2569                 /* Has the page moved? */
2570                 if (unlikely(page != *slot)) {
2571                         put_page(head);
2572                         goto repeat;
2573                 }
2574
2575                 if (!PageUptodate(page) ||
2576                                 PageReadahead(page) ||
2577                                 PageHWPoison(page))
2578                         goto skip;
2579                 if (!trylock_page(page))
2580                         goto skip;
2581
2582                 if (page->mapping != mapping || !PageUptodate(page))
2583                         goto unlock;
2584
2585                 max_idx = DIV_ROUND_UP(i_size_read(mapping->host), PAGE_SIZE);
2586                 if (page->index >= max_idx)
2587                         goto unlock;
2588
2589                 if (file->f_ra.mmap_miss > 0)
2590                         file->f_ra.mmap_miss--;
2591
2592                 vmf->address += (iter.index - last_pgoff) << PAGE_SHIFT;
2593                 if (vmf->pte)
2594                         vmf->pte += iter.index - last_pgoff;
2595                 last_pgoff = iter.index;
2596                 if (alloc_set_pte(vmf, NULL, page))
2597                         goto unlock;
2598                 unlock_page(page);
2599                 goto next;
2600 unlock:
2601                 unlock_page(page);
2602 skip:
2603                 put_page(page);
2604 next:
2605                 /* Huge page is mapped? No need to proceed. */
2606                 if (pmd_trans_huge(*vmf->pmd))
2607                         break;
2608                 if (iter.index == end_pgoff)
2609                         break;
2610         }
2611         rcu_read_unlock();
2612 }
2613 EXPORT_SYMBOL(filemap_map_pages);
2614
2615 int filemap_page_mkwrite(struct vm_fault *vmf)
2616 {
2617         struct page *page = vmf->page;
2618         struct inode *inode = file_inode(vmf->vma->vm_file);
2619         int ret = VM_FAULT_LOCKED;
2620
2621         sb_start_pagefault(inode->i_sb);
2622         file_update_time(vmf->vma->vm_file);
2623         lock_page(page);
2624         if (page->mapping != inode->i_mapping) {
2625                 unlock_page(page);
2626                 ret = VM_FAULT_NOPAGE;
2627                 goto out;
2628         }
2629         /*
2630          * We mark the page dirty already here so that when freeze is in
2631          * progress, we are guaranteed that writeback during freezing will
2632          * see the dirty page and writeprotect it again.
2633          */
2634         set_page_dirty(page);
2635         wait_for_stable_page(page);
2636 out:
2637         sb_end_pagefault(inode->i_sb);
2638         return ret;
2639 }
2640 EXPORT_SYMBOL(filemap_page_mkwrite);
2641
2642 const struct vm_operations_struct generic_file_vm_ops = {
2643         .fault          = filemap_fault,
2644         .map_pages      = filemap_map_pages,
2645         .page_mkwrite   = filemap_page_mkwrite,
2646 };
2647
2648 /* This is used for a general mmap of a disk file */
2649
2650 int generic_file_mmap(struct file * file, struct vm_area_struct * vma)
2651 {
2652         struct address_space *mapping = file->f_mapping;
2653
2654         if (!mapping->a_ops->readpage)
2655                 return -ENOEXEC;
2656         file_accessed(file);
2657         vma->vm_ops = &generic_file_vm_ops;
2658         return 0;
2659 }
2660
2661 /*
2662  * This is for filesystems which do not implement ->writepage.
2663  */
2664 int generic_file_readonly_mmap(struct file *file, struct vm_area_struct *vma)
2665 {
2666         if ((vma->vm_flags & VM_SHARED) && (vma->vm_flags & VM_MAYWRITE))
2667                 return -EINVAL;
2668         return generic_file_mmap(file, vma);
2669 }
2670 #else
2671 int generic_file_mmap(struct file * file, struct vm_area_struct * vma)
2672 {
2673         return -ENOSYS;
2674 }
2675 int generic_file_readonly_mmap(struct file * file, struct vm_area_struct * vma)
2676 {
2677         return -ENOSYS;
2678 }
2679 #endif /* CONFIG_MMU */
2680
2681 EXPORT_SYMBOL(generic_file_mmap);
2682 EXPORT_SYMBOL(generic_file_readonly_mmap);
2683
2684 static struct page *wait_on_page_read(struct page *page)
2685 {
2686         if (!IS_ERR(page)) {
2687                 wait_on_page_locked(page);
2688                 if (!PageUptodate(page)) {
2689                         put_page(page);
2690                         page = ERR_PTR(-EIO);
2691                 }
2692         }
2693         return page;
2694 }
2695
2696 static struct page *do_read_cache_page(struct address_space *mapping,
2697                                 pgoff_t index,
2698                                 int (*filler)(void *, struct page *),
2699                                 void *data,
2700                                 gfp_t gfp)
2701 {
2702         struct page *page;
2703         int err;
2704 repeat:
2705         page = find_get_page(mapping, index);
2706         if (!page) {
2707                 page = __page_cache_alloc(gfp | __GFP_COLD);
2708                 if (!page)
2709                         return ERR_PTR(-ENOMEM);
2710                 err = add_to_page_cache_lru(page, mapping, index, gfp);
2711                 if (unlikely(err)) {
2712                         put_page(page);
2713                         if (err == -EEXIST)
2714                                 goto repeat;
2715                         /* Presumably ENOMEM for radix tree node */
2716                         return ERR_PTR(err);
2717                 }
2718
2719 filler:
2720                 err = filler(data, page);
2721                 if (err < 0) {
2722                         put_page(page);
2723                         return ERR_PTR(err);
2724                 }
2725
2726                 page = wait_on_page_read(page);
2727                 if (IS_ERR(page))
2728                         return page;
2729                 goto out;
2730         }
2731         if (PageUptodate(page))
2732                 goto out;
2733
2734         /*
2735          * Page is not up to date and may be locked due one of the following
2736          * case a: Page is being filled and the page lock is held
2737          * case b: Read/write error clearing the page uptodate status
2738          * case c: Truncation in progress (page locked)
2739          * case d: Reclaim in progress
2740          *
2741          * Case a, the page will be up to date when the page is unlocked.
2742          *    There is no need to serialise on the page lock here as the page
2743          *    is pinned so the lock gives no additional protection. Even if the
2744          *    the page is truncated, the data is still valid if PageUptodate as
2745          *    it's a race vs truncate race.
2746          * Case b, the page will not be up to date
2747          * Case c, the page may be truncated but in itself, the data may still
2748          *    be valid after IO completes as it's a read vs truncate race. The
2749          *    operation must restart if the page is not uptodate on unlock but
2750          *    otherwise serialising on page lock to stabilise the mapping gives
2751          *    no additional guarantees to the caller as the page lock is
2752          *    released before return.
2753          * Case d, similar to truncation. If reclaim holds the page lock, it
2754          *    will be a race with remove_mapping that determines if the mapping
2755          *    is valid on unlock but otherwise the data is valid and there is
2756          *    no need to serialise with page lock.
2757          *
2758          * As the page lock gives no additional guarantee, we optimistically
2759          * wait on the page to be unlocked and check if it's up to date and
2760          * use the page if it is. Otherwise, the page lock is required to
2761          * distinguish between the different cases. The motivation is that we
2762          * avoid spurious serialisations and wakeups when multiple processes
2763          * wait on the same page for IO to complete.
2764          */
2765         wait_on_page_locked(page);
2766         if (PageUptodate(page))
2767                 goto out;
2768
2769         /* Distinguish between all the cases under the safety of the lock */
2770         lock_page(page);
2771
2772         /* Case c or d, restart the operation */
2773         if (!page->mapping) {
2774                 unlock_page(page);
2775                 put_page(page);
2776                 goto repeat;
2777         }
2778
2779         /* Someone else locked and filled the page in a very small window */
2780         if (PageUptodate(page)) {
2781                 unlock_page(page);
2782                 goto out;
2783         }
2784         goto filler;
2785
2786 out:
2787         mark_page_accessed(page);
2788         return page;
2789 }
2790
2791 /**
2792  * read_cache_page - read into page cache, fill it if needed
2793  * @mapping:    the page's address_space
2794  * @index:      the page index
2795  * @filler:     function to perform the read
2796  * @data:       first arg to filler(data, page) function, often left as NULL
2797  *
2798  * Read into the page cache. If a page already exists, and PageUptodate() is
2799  * not set, try to fill the page and wait for it to become unlocked.
2800  *
2801  * If the page does not get brought uptodate, return -EIO.
2802  */
2803 struct page *read_cache_page(struct address_space *mapping,
2804                                 pgoff_t index,
2805                                 int (*filler)(void *, struct page *),
2806                                 void *data)
2807 {
2808         return do_read_cache_page(mapping, index, filler, data, mapping_gfp_mask(mapping));
2809 }
2810 EXPORT_SYMBOL(read_cache_page);
2811
2812 /**
2813  * read_cache_page_gfp - read into page cache, using specified page allocation flags.
2814  * @mapping:    the page's address_space
2815  * @index:      the page index
2816  * @gfp:        the page allocator flags to use if allocating
2817  *
2818  * This is the same as "read_mapping_page(mapping, index, NULL)", but with
2819  * any new page allocations done using the specified allocation flags.
2820  *
2821  * If the page does not get brought uptodate, return -EIO.
2822  */
2823 struct page *read_cache_page_gfp(struct address_space *mapping,
2824                                 pgoff_t index,
2825                                 gfp_t gfp)
2826 {
2827         filler_t *filler = (filler_t *)mapping->a_ops->readpage;
2828
2829         return do_read_cache_page(mapping, index, filler, NULL, gfp);
2830 }
2831 EXPORT_SYMBOL(read_cache_page_gfp);
2832
2833 /*
2834  * Performs necessary checks before doing a write
2835  *
2836  * Can adjust writing position or amount of bytes to write.
2837  * Returns appropriate error code that caller should return or
2838  * zero in case that write should be allowed.
2839  */
2840 inline ssize_t generic_write_checks(struct kiocb *iocb, struct iov_iter *from)
2841 {
2842         struct file *file = iocb->ki_filp;
2843         struct inode *inode = file->f_mapping->host;
2844         unsigned long limit = rlimit(RLIMIT_FSIZE);
2845         loff_t pos;
2846
2847         if (!iov_iter_count(from))
2848                 return 0;
2849
2850         /* FIXME: this is for backwards compatibility with 2.4 */
2851         if (iocb->ki_flags & IOCB_APPEND)
2852                 iocb->ki_pos = i_size_read(inode);
2853
2854         pos = iocb->ki_pos;
2855
2856         if ((iocb->ki_flags & IOCB_NOWAIT) && !(iocb->ki_flags & IOCB_DIRECT))
2857                 return -EINVAL;
2858
2859         if (limit != RLIM_INFINITY) {
2860                 if (iocb->ki_pos >= limit) {
2861                         send_sig(SIGXFSZ, current, 0);
2862                         return -EFBIG;
2863                 }
2864                 iov_iter_truncate(from, limit - (unsigned long)pos);
2865         }
2866
2867         /*
2868          * LFS rule
2869          */
2870         if (unlikely(pos + iov_iter_count(from) > MAX_NON_LFS &&
2871                                 !(file->f_flags & O_LARGEFILE))) {
2872                 if (pos >= MAX_NON_LFS)
2873                         return -EFBIG;
2874                 iov_iter_truncate(from, MAX_NON_LFS - (unsigned long)pos);
2875         }
2876
2877         /*
2878          * Are we about to exceed the fs block limit ?
2879          *
2880          * If we have written data it becomes a short write.  If we have
2881          * exceeded without writing data we send a signal and return EFBIG.
2882          * Linus frestrict idea will clean these up nicely..
2883          */
2884         if (unlikely(pos >= inode->i_sb->s_maxbytes))
2885                 return -EFBIG;
2886
2887         iov_iter_truncate(from, inode->i_sb->s_maxbytes - pos);
2888         return iov_iter_count(from);
2889 }
2890 EXPORT_SYMBOL(generic_write_checks);
2891
2892 int pagecache_write_begin(struct file *file, struct address_space *mapping,
2893                                 loff_t pos, unsigned len, unsigned flags,
2894                                 struct page **pagep, void **fsdata)
2895 {
2896         const struct address_space_operations *aops = mapping->a_ops;
2897
2898         return aops->write_begin(file, mapping, pos, len, flags,
2899                                                         pagep, fsdata);
2900 }
2901 EXPORT_SYMBOL(pagecache_write_begin);
2902
2903 int pagecache_write_end(struct file *file, struct address_space *mapping,
2904                                 loff_t pos, unsigned len, unsigned copied,
2905                                 struct page *page, void *fsdata)
2906 {
2907         const struct address_space_operations *aops = mapping->a_ops;
2908
2909         return aops->write_end(file, mapping, pos, len, copied, page, fsdata);
2910 }
2911 EXPORT_SYMBOL(pagecache_write_end);
2912
2913 ssize_t
2914 generic_file_direct_write(struct kiocb *iocb, struct iov_iter *from)
2915 {
2916         struct file     *file = iocb->ki_filp;
2917         struct address_space *mapping = file->f_mapping;
2918         struct inode    *inode = mapping->host;
2919         loff_t          pos = iocb->ki_pos;
2920         ssize_t         written;
2921         size_t          write_len;
2922         pgoff_t         end;
2923
2924         write_len = iov_iter_count(from);
2925         end = (pos + write_len - 1) >> PAGE_SHIFT;
2926
2927         if (iocb->ki_flags & IOCB_NOWAIT) {
2928                 /* If there are pages to writeback, return */
2929                 if (filemap_range_has_page(inode->i_mapping, pos,
2930                                            pos + iov_iter_count(from)))
2931                         return -EAGAIN;
2932         } else {
2933                 written = filemap_write_and_wait_range(mapping, pos,
2934                                                         pos + write_len - 1);
2935                 if (written)
2936                         goto out;
2937         }
2938
2939         /*
2940          * After a write we want buffered reads to be sure to go to disk to get
2941          * the new data.  We invalidate clean cached page from the region we're
2942          * about to write.  We do this *before* the write so that we can return
2943          * without clobbering -EIOCBQUEUED from ->direct_IO().
2944          */
2945         written = invalidate_inode_pages2_range(mapping,
2946                                         pos >> PAGE_SHIFT, end);
2947         /*
2948          * If a page can not be invalidated, return 0 to fall back
2949          * to buffered write.
2950          */
2951         if (written) {
2952                 if (written == -EBUSY)
2953                         return 0;
2954                 goto out;
2955         }
2956
2957         written = mapping->a_ops->direct_IO(iocb, from);
2958
2959         /*
2960          * Finally, try again to invalidate clean pages which might have been
2961          * cached by non-direct readahead, or faulted in by get_user_pages()
2962          * if the source of the write was an mmap'ed region of the file
2963          * we're writing.  Either one is a pretty crazy thing to do,
2964          * so we don't support it 100%.  If this invalidation
2965          * fails, tough, the write still worked...
2966          *
2967          * Most of the time we do not need this since dio_complete() will do
2968          * the invalidation for us. However there are some file systems that
2969          * do not end up with dio_complete() being called, so let's not break
2970          * them by removing it completely
2971          */
2972         if (mapping->nrpages)
2973                 invalidate_inode_pages2_range(mapping,
2974                                         pos >> PAGE_SHIFT, end);
2975
2976         if (written > 0) {
2977                 pos += written;
2978                 write_len -= written;
2979                 if (pos > i_size_read(inode) && !S_ISBLK(inode->i_mode)) {
2980                         i_size_write(inode, pos);
2981                         mark_inode_dirty(inode);
2982                 }
2983                 iocb->ki_pos = pos;
2984         }
2985         iov_iter_revert(from, write_len - iov_iter_count(from));
2986 out:
2987         return written;
2988 }
2989 EXPORT_SYMBOL(generic_file_direct_write);
2990
2991 /*
2992  * Find or create a page at the given pagecache position. Return the locked
2993  * page. This function is specifically for buffered writes.
2994  */
2995 struct page *grab_cache_page_write_begin(struct address_space *mapping,
2996                                         pgoff_t index, unsigned flags)
2997 {
2998         struct page *page;
2999         int fgp_flags = FGP_LOCK|FGP_WRITE|FGP_CREAT;
3000
3001         if (flags & AOP_FLAG_NOFS)
3002                 fgp_flags |= FGP_NOFS;
3003
3004         page = pagecache_get_page(mapping, index, fgp_flags,
3005                         mapping_gfp_mask(mapping));
3006         if (page)
3007                 wait_for_stable_page(page);
3008
3009         return page;
3010 }
3011 EXPORT_SYMBOL(grab_cache_page_write_begin);
3012
3013 ssize_t generic_perform_write(struct file *file,
3014                                 struct iov_iter *i, loff_t pos)
3015 {
3016         struct address_space *mapping = file->f_mapping;
3017         const struct address_space_operations *a_ops = mapping->a_ops;
3018         long status = 0;
3019         ssize_t written = 0;
3020         unsigned int flags = 0;
3021
3022         do {
3023                 struct page *page;
3024                 unsigned long offset;   /* Offset into pagecache page */
3025                 unsigned long bytes;    /* Bytes to write to page */
3026                 size_t copied;          /* Bytes copied from user */
3027                 void *fsdata;
3028
3029                 offset = (pos & (PAGE_SIZE - 1));
3030                 bytes = min_t(unsigned long, PAGE_SIZE - offset,
3031                                                 iov_iter_count(i));
3032
3033 again:
3034                 /*
3035                  * Bring in the user page that we will copy from _first_.
3036                  * Otherwise there's a nasty deadlock on copying from the
3037                  * same page as we're writing to, without it being marked
3038                  * up-to-date.
3039                  *
3040                  * Not only is this an optimisation, but it is also required
3041                  * to check that the address is actually valid, when atomic
3042                  * usercopies are used, below.
3043                  */
3044                 if (unlikely(iov_iter_fault_in_readable(i, bytes))) {
3045                         status = -EFAULT;
3046                         break;
3047                 }
3048
3049                 if (fatal_signal_pending(current)) {
3050                         status = -EINTR;
3051                         break;
3052                 }
3053
3054                 status = a_ops->write_begin(file, mapping, pos, bytes, flags,
3055                                                 &page, &fsdata);
3056                 if (unlikely(status < 0))
3057                         break;
3058
3059                 if (mapping_writably_mapped(mapping))
3060                         flush_dcache_page(page);
3061
3062                 copied = iov_iter_copy_from_user_atomic(page, i, offset, bytes);
3063                 flush_dcache_page(page);
3064
3065                 status = a_ops->write_end(file, mapping, pos, bytes, copied,
3066                                                 page, fsdata);
3067                 if (unlikely(status < 0))
3068                         break;
3069                 copied = status;
3070
3071                 cond_resched();
3072
3073                 iov_iter_advance(i, copied);
3074                 if (unlikely(copied == 0)) {
3075                         /*
3076                          * If we were unable to copy any data at all, we must
3077                          * fall back to a single segment length write.
3078                          *
3079                          * If we didn't fallback here, we could livelock
3080                          * because not all segments in the iov can be copied at
3081                          * once without a pagefault.
3082                          */
3083                         bytes = min_t(unsigned long, PAGE_SIZE - offset,
3084                                                 iov_iter_single_seg_count(i));
3085                         goto again;
3086                 }
3087                 pos += copied;
3088                 written += copied;
3089
3090                 balance_dirty_pages_ratelimited(mapping);
3091         } while (iov_iter_count(i));
3092
3093         return written ? written : status;
3094 }
3095 EXPORT_SYMBOL(generic_perform_write);
3096
3097 /**
3098  * __generic_file_write_iter - write data to a file
3099  * @iocb:       IO state structure (file, offset, etc.)
3100  * @from:       iov_iter with data to write
3101  *
3102  * This function does all the work needed for actually writing data to a
3103  * file. It does all basic checks, removes SUID from the file, updates
3104  * modification times and calls proper subroutines depending on whether we
3105  * do direct IO or a standard buffered write.
3106  *
3107  * It expects i_mutex to be grabbed unless we work on a block device or similar
3108  * object which does not need locking at all.
3109  *
3110  * This function does *not* take care of syncing data in case of O_SYNC write.
3111  * A caller has to handle it. This is mainly due to the fact that we want to
3112  * avoid syncing under i_mutex.
3113  */
3114 ssize_t __generic_file_write_iter(struct kiocb *iocb, struct iov_iter *from)
3115 {
3116         struct file *file = iocb->ki_filp;
3117         struct address_space * mapping = file->f_mapping;
3118         struct inode    *inode = mapping->host;
3119         ssize_t         written = 0;
3120         ssize_t         err;
3121         ssize_t         status;
3122
3123         /* We can write back this queue in page reclaim */
3124         current->backing_dev_info = inode_to_bdi(inode);
3125         err = file_remove_privs(file);
3126         if (err)
3127                 goto out;
3128
3129         err = file_update_time(file);
3130         if (err)
3131                 goto out;
3132
3133         if (iocb->ki_flags & IOCB_DIRECT) {
3134                 loff_t pos, endbyte;
3135
3136                 written = generic_file_direct_write(iocb, from);
3137                 /*
3138                  * If the write stopped short of completing, fall back to
3139                  * buffered writes.  Some filesystems do this for writes to
3140                  * holes, for example.  For DAX files, a buffered write will
3141                  * not succeed (even if it did, DAX does not handle dirty
3142                  * page-cache pages correctly).
3143                  */
3144                 if (written < 0 || !iov_iter_count(from) || IS_DAX(inode))
3145                         goto out;
3146
3147                 status = generic_perform_write(file, from, pos = iocb->ki_pos);
3148                 /*
3149                  * If generic_perform_write() returned a synchronous error
3150                  * then we want to return the number of bytes which were
3151                  * direct-written, or the error code if that was zero.  Note
3152                  * that this differs from normal direct-io semantics, which
3153                  * will return -EFOO even if some bytes were written.
3154                  */
3155                 if (unlikely(status < 0)) {
3156                         err = status;
3157                         goto out;
3158                 }
3159                 /*
3160                  * We need to ensure that the page cache pages are written to
3161                  * disk and invalidated to preserve the expected O_DIRECT
3162                  * semantics.
3163                  */
3164                 endbyte = pos + status - 1;
3165                 err = filemap_write_and_wait_range(mapping, pos, endbyte);
3166                 if (err == 0) {
3167                         iocb->ki_pos = endbyte + 1;
3168                         written += status;
3169                         invalidate_mapping_pages(mapping,
3170                                                  pos >> PAGE_SHIFT,
3171                                                  endbyte >> PAGE_SHIFT);
3172                 } else {
3173                         /*
3174                          * We don't know how much we wrote, so just return
3175                          * the number of bytes which were direct-written
3176                          */
3177                 }
3178         } else {
3179                 written = generic_perform_write(file, from, iocb->ki_pos);
3180                 if (likely(written > 0))
3181                         iocb->ki_pos += written;
3182         }
3183 out:
3184         current->backing_dev_info = NULL;
3185         return written ? written : err;
3186 }
3187 EXPORT_SYMBOL(__generic_file_write_iter);
3188
3189 /**
3190  * generic_file_write_iter - write data to a file
3191  * @iocb:       IO state structure
3192  * @from:       iov_iter with data to write
3193  *
3194  * This is a wrapper around __generic_file_write_iter() to be used by most
3195  * filesystems. It takes care of syncing the file in case of O_SYNC file
3196  * and acquires i_mutex as needed.
3197  */
3198 ssize_t generic_file_write_iter(struct kiocb *iocb, struct iov_iter *from)
3199 {
3200         struct file *file = iocb->ki_filp;
3201         struct inode *inode = file->f_mapping->host;
3202         ssize_t ret;
3203
3204         inode_lock(inode);
3205         ret = generic_write_checks(iocb, from);
3206         if (ret > 0)
3207                 ret = __generic_file_write_iter(iocb, from);
3208         inode_unlock(inode);
3209
3210         if (ret > 0)
3211                 ret = generic_write_sync(iocb, ret);
3212         return ret;
3213 }
3214 EXPORT_SYMBOL(generic_file_write_iter);
3215
3216 /**
3217  * try_to_release_page() - release old fs-specific metadata on a page
3218  *
3219  * @page: the page which the kernel is trying to free
3220  * @gfp_mask: memory allocation flags (and I/O mode)
3221  *
3222  * The address_space is to try to release any data against the page
3223  * (presumably at page->private).  If the release was successful, return '1'.
3224  * Otherwise return zero.
3225  *
3226  * This may also be called if PG_fscache is set on a page, indicating that the
3227  * page is known to the local caching routines.
3228  *
3229  * The @gfp_mask argument specifies whether I/O may be performed to release
3230  * this page (__GFP_IO), and whether the call may block (__GFP_RECLAIM & __GFP_FS).
3231  *
3232  */
3233 int try_to_release_page(struct page *page, gfp_t gfp_mask)
3234 {
3235         struct address_space * const mapping = page->mapping;
3236
3237         BUG_ON(!PageLocked(page));
3238         if (PageWriteback(page))
3239                 return 0;
3240
3241         if (mapping && mapping->a_ops->releasepage)
3242                 return mapping->a_ops->releasepage(page, gfp_mask);
3243         return try_to_free_buffers(page);
3244 }
3245
3246 EXPORT_SYMBOL(try_to_release_page);