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