dad935769055ff741e06e8caa1270d800278d268
[sfrench/cifs-2.6.git] / mm / filemap.c
1 /*
2  *      linux/mm/filemap.c
3  *
4  * Copyright (C) 1994-1999  Linus Torvalds
5  */
6
7 /*
8  * This file handles the generic file mmap semantics used by
9  * most "normal" filesystems (but you don't /have/ to use this:
10  * the NFS filesystem used to do this differently, for example)
11  */
12 #include <linux/export.h>
13 #include <linux/compiler.h>
14 #include <linux/dax.h>
15 #include <linux/fs.h>
16 #include <linux/sched/signal.h>
17 #include <linux/uaccess.h>
18 #include <linux/capability.h>
19 #include <linux/kernel_stat.h>
20 #include <linux/gfp.h>
21 #include <linux/mm.h>
22 #include <linux/swap.h>
23 #include <linux/mman.h>
24 #include <linux/pagemap.h>
25 #include <linux/file.h>
26 #include <linux/uio.h>
27 #include <linux/hash.h>
28 #include <linux/writeback.h>
29 #include <linux/backing-dev.h>
30 #include <linux/pagevec.h>
31 #include <linux/blkdev.h>
32 #include <linux/security.h>
33 #include <linux/cpuset.h>
34 #include <linux/hardirq.h> /* for BUG_ON(!in_atomic()) only */
35 #include <linux/hugetlb.h>
36 #include <linux/memcontrol.h>
37 #include <linux/cleancache.h>
38 #include <linux/rmap.h>
39 #include "internal.h"
40
41 #define CREATE_TRACE_POINTS
42 #include <trace/events/filemap.h>
43
44 /*
45  * FIXME: remove all knowledge of the buffer layer from the core VM
46  */
47 #include <linux/buffer_head.h> /* for try_to_free_buffers */
48
49 #include <asm/mman.h>
50
51 /*
52  * Shared mappings implemented 30.11.1994. It's not fully working yet,
53  * though.
54  *
55  * Shared mappings now work. 15.8.1995  Bruno.
56  *
57  * finished 'unifying' the page and buffer cache and SMP-threaded the
58  * page-cache, 21.05.1999, Ingo Molnar <mingo@redhat.com>
59  *
60  * SMP-threaded pagemap-LRU 1999, Andrea Arcangeli <andrea@suse.de>
61  */
62
63 /*
64  * Lock ordering:
65  *
66  *  ->i_mmap_rwsem              (truncate_pagecache)
67  *    ->private_lock            (__free_pte->__set_page_dirty_buffers)
68  *      ->swap_lock             (exclusive_swap_page, others)
69  *        ->mapping->tree_lock
70  *
71  *  ->i_mutex
72  *    ->i_mmap_rwsem            (truncate->unmap_mapping_range)
73  *
74  *  ->mmap_sem
75  *    ->i_mmap_rwsem
76  *      ->page_table_lock or pte_lock   (various, mainly in memory.c)
77  *        ->mapping->tree_lock  (arch-dependent flush_dcache_mmap_lock)
78  *
79  *  ->mmap_sem
80  *    ->lock_page               (access_process_vm)
81  *
82  *  ->i_mutex                   (generic_perform_write)
83  *    ->mmap_sem                (fault_in_pages_readable->do_page_fault)
84  *
85  *  bdi->wb.list_lock
86  *    sb_lock                   (fs/fs-writeback.c)
87  *    ->mapping->tree_lock      (__sync_single_inode)
88  *
89  *  ->i_mmap_rwsem
90  *    ->anon_vma.lock           (vma_adjust)
91  *
92  *  ->anon_vma.lock
93  *    ->page_table_lock or pte_lock     (anon_vma_prepare and various)
94  *
95  *  ->page_table_lock or pte_lock
96  *    ->swap_lock               (try_to_unmap_one)
97  *    ->private_lock            (try_to_unmap_one)
98  *    ->tree_lock               (try_to_unmap_one)
99  *    ->zone_lru_lock(zone)     (follow_page->mark_page_accessed)
100  *    ->zone_lru_lock(zone)     (check_pte_range->isolate_lru_page)
101  *    ->private_lock            (page_remove_rmap->set_page_dirty)
102  *    ->tree_lock               (page_remove_rmap->set_page_dirty)
103  *    bdi.wb->list_lock         (page_remove_rmap->set_page_dirty)
104  *    ->inode->i_lock           (page_remove_rmap->set_page_dirty)
105  *    ->memcg->move_lock        (page_remove_rmap->lock_page_memcg)
106  *    bdi.wb->list_lock         (zap_pte_range->set_page_dirty)
107  *    ->inode->i_lock           (zap_pte_range->set_page_dirty)
108  *    ->private_lock            (zap_pte_range->__set_page_dirty_buffers)
109  *
110  * ->i_mmap_rwsem
111  *   ->tasklist_lock            (memory_failure, collect_procs_ao)
112  */
113
114 static int page_cache_tree_insert(struct address_space *mapping,
115                                   struct page *page, void **shadowp)
116 {
117         struct radix_tree_node *node;
118         void **slot;
119         int error;
120
121         error = __radix_tree_create(&mapping->page_tree, page->index, 0,
122                                     &node, &slot);
123         if (error)
124                 return error;
125         if (*slot) {
126                 void *p;
127
128                 p = radix_tree_deref_slot_protected(slot, &mapping->tree_lock);
129                 if (!radix_tree_exceptional_entry(p))
130                         return -EEXIST;
131
132                 mapping->nrexceptional--;
133                 if (shadowp)
134                         *shadowp = p;
135         }
136         __radix_tree_replace(&mapping->page_tree, node, slot, page,
137                              workingset_update_node, mapping);
138         mapping->nrpages++;
139         return 0;
140 }
141
142 static void page_cache_tree_delete(struct address_space *mapping,
143                                    struct page *page, void *shadow)
144 {
145         int i, nr;
146
147         /* hugetlb pages are represented by one entry in the radix tree */
148         nr = PageHuge(page) ? 1 : hpage_nr_pages(page);
149
150         VM_BUG_ON_PAGE(!PageLocked(page), page);
151         VM_BUG_ON_PAGE(PageTail(page), page);
152         VM_BUG_ON_PAGE(nr != 1 && shadow, page);
153
154         for (i = 0; i < nr; i++) {
155                 struct radix_tree_node *node;
156                 void **slot;
157
158                 __radix_tree_lookup(&mapping->page_tree, page->index + i,
159                                     &node, &slot);
160
161                 VM_BUG_ON_PAGE(!node && nr != 1, page);
162
163                 radix_tree_clear_tags(&mapping->page_tree, node, slot);
164                 __radix_tree_replace(&mapping->page_tree, node, slot, shadow,
165                                      workingset_update_node, mapping);
166         }
167
168         if (shadow) {
169                 mapping->nrexceptional += nr;
170                 /*
171                  * Make sure the nrexceptional update is committed before
172                  * the nrpages update so that final truncate racing
173                  * with reclaim does not see both counters 0 at the
174                  * same time and miss a shadow entry.
175                  */
176                 smp_wmb();
177         }
178         mapping->nrpages -= nr;
179 }
180
181 /*
182  * Delete a page from the page cache and free it. Caller has to make
183  * sure the page is locked and that nobody else uses it - or that usage
184  * is safe.  The caller must hold the mapping's tree_lock.
185  */
186 void __delete_from_page_cache(struct page *page, void *shadow)
187 {
188         struct address_space *mapping = page->mapping;
189         int nr = hpage_nr_pages(page);
190
191         trace_mm_filemap_delete_from_page_cache(page);
192         /*
193          * if we're uptodate, flush out into the cleancache, otherwise
194          * invalidate any existing cleancache entries.  We can't leave
195          * stale data around in the cleancache once our page is gone
196          */
197         if (PageUptodate(page) && PageMappedToDisk(page))
198                 cleancache_put_page(page);
199         else
200                 cleancache_invalidate_page(mapping, page);
201
202         VM_BUG_ON_PAGE(PageTail(page), page);
203         VM_BUG_ON_PAGE(page_mapped(page), page);
204         if (!IS_ENABLED(CONFIG_DEBUG_VM) && unlikely(page_mapped(page))) {
205                 int mapcount;
206
207                 pr_alert("BUG: Bad page cache in process %s  pfn:%05lx\n",
208                          current->comm, page_to_pfn(page));
209                 dump_page(page, "still mapped when deleted");
210                 dump_stack();
211                 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
212
213                 mapcount = page_mapcount(page);
214                 if (mapping_exiting(mapping) &&
215                     page_count(page) >= mapcount + 2) {
216                         /*
217                          * All vmas have already been torn down, so it's
218                          * a good bet that actually the page is unmapped,
219                          * and we'd prefer not to leak it: if we're wrong,
220                          * some other bad page check should catch it later.
221                          */
222                         page_mapcount_reset(page);
223                         page_ref_sub(page, mapcount);
224                 }
225         }
226
227         page_cache_tree_delete(mapping, page, shadow);
228
229         page->mapping = NULL;
230         /* Leave page->index set: truncation lookup relies upon it */
231
232         /* hugetlb pages do not participate in page cache accounting. */
233         if (PageHuge(page))
234                 return;
235
236         __mod_node_page_state(page_pgdat(page), NR_FILE_PAGES, -nr);
237         if (PageSwapBacked(page)) {
238                 __mod_node_page_state(page_pgdat(page), NR_SHMEM, -nr);
239                 if (PageTransHuge(page))
240                         __dec_node_page_state(page, NR_SHMEM_THPS);
241         } else {
242                 VM_BUG_ON_PAGE(PageTransHuge(page), page);
243         }
244
245         /*
246          * At this point page must be either written or cleaned by truncate.
247          * Dirty page here signals a bug and loss of unwritten data.
248          *
249          * This fixes dirty accounting after removing the page entirely but
250          * leaves PageDirty set: it has no effect for truncated page and
251          * anyway will be cleared before returning page into buddy allocator.
252          */
253         if (WARN_ON_ONCE(PageDirty(page)))
254                 account_page_cleaned(page, mapping, inode_to_wb(mapping->host));
255 }
256
257 /**
258  * delete_from_page_cache - delete page from page cache
259  * @page: the page which the kernel is trying to remove from page cache
260  *
261  * This must be called only on pages that have been verified to be in the page
262  * cache and locked.  It will never put the page into the free list, the caller
263  * has a reference on the page.
264  */
265 void delete_from_page_cache(struct page *page)
266 {
267         struct address_space *mapping = page_mapping(page);
268         unsigned long flags;
269         void (*freepage)(struct page *);
270
271         BUG_ON(!PageLocked(page));
272
273         freepage = mapping->a_ops->freepage;
274
275         spin_lock_irqsave(&mapping->tree_lock, flags);
276         __delete_from_page_cache(page, NULL);
277         spin_unlock_irqrestore(&mapping->tree_lock, flags);
278
279         if (freepage)
280                 freepage(page);
281
282         if (PageTransHuge(page) && !PageHuge(page)) {
283                 page_ref_sub(page, HPAGE_PMD_NR);
284                 VM_BUG_ON_PAGE(page_count(page) <= 0, page);
285         } else {
286                 put_page(page);
287         }
288 }
289 EXPORT_SYMBOL(delete_from_page_cache);
290
291 int filemap_check_errors(struct address_space *mapping)
292 {
293         int ret = 0;
294         /* Check for outstanding write errors */
295         if (test_bit(AS_ENOSPC, &mapping->flags) &&
296             test_and_clear_bit(AS_ENOSPC, &mapping->flags))
297                 ret = -ENOSPC;
298         if (test_bit(AS_EIO, &mapping->flags) &&
299             test_and_clear_bit(AS_EIO, &mapping->flags))
300                 ret = -EIO;
301         return ret;
302 }
303 EXPORT_SYMBOL(filemap_check_errors);
304
305 static int filemap_check_and_keep_errors(struct address_space *mapping)
306 {
307         /* Check for outstanding write errors */
308         if (test_bit(AS_EIO, &mapping->flags))
309                 return -EIO;
310         if (test_bit(AS_ENOSPC, &mapping->flags))
311                 return -ENOSPC;
312         return 0;
313 }
314
315 /**
316  * __filemap_fdatawrite_range - start writeback on mapping dirty pages in range
317  * @mapping:    address space structure to write
318  * @start:      offset in bytes where the range starts
319  * @end:        offset in bytes where the range ends (inclusive)
320  * @sync_mode:  enable synchronous operation
321  *
322  * Start writeback against all of a mapping's dirty pages that lie
323  * within the byte offsets <start, end> inclusive.
324  *
325  * If sync_mode is WB_SYNC_ALL then this is a "data integrity" operation, as
326  * opposed to a regular memory cleansing writeback.  The difference between
327  * these two operations is that if a dirty page/buffer is encountered, it must
328  * be waited upon, and not just skipped over.
329  */
330 int __filemap_fdatawrite_range(struct address_space *mapping, loff_t start,
331                                 loff_t end, int sync_mode)
332 {
333         int ret;
334         struct writeback_control wbc = {
335                 .sync_mode = sync_mode,
336                 .nr_to_write = LONG_MAX,
337                 .range_start = start,
338                 .range_end = end,
339         };
340
341         if (!mapping_cap_writeback_dirty(mapping))
342                 return 0;
343
344         wbc_attach_fdatawrite_inode(&wbc, mapping->host);
345         ret = do_writepages(mapping, &wbc);
346         wbc_detach_inode(&wbc);
347         return ret;
348 }
349
350 static inline int __filemap_fdatawrite(struct address_space *mapping,
351         int sync_mode)
352 {
353         return __filemap_fdatawrite_range(mapping, 0, LLONG_MAX, sync_mode);
354 }
355
356 int filemap_fdatawrite(struct address_space *mapping)
357 {
358         return __filemap_fdatawrite(mapping, WB_SYNC_ALL);
359 }
360 EXPORT_SYMBOL(filemap_fdatawrite);
361
362 int filemap_fdatawrite_range(struct address_space *mapping, loff_t start,
363                                 loff_t end)
364 {
365         return __filemap_fdatawrite_range(mapping, start, end, WB_SYNC_ALL);
366 }
367 EXPORT_SYMBOL(filemap_fdatawrite_range);
368
369 /**
370  * filemap_flush - mostly a non-blocking flush
371  * @mapping:    target address_space
372  *
373  * This is a mostly non-blocking flush.  Not suitable for data-integrity
374  * purposes - I/O may not be started against all dirty pages.
375  */
376 int filemap_flush(struct address_space *mapping)
377 {
378         return __filemap_fdatawrite(mapping, WB_SYNC_NONE);
379 }
380 EXPORT_SYMBOL(filemap_flush);
381
382 /**
383  * filemap_range_has_page - check if a page exists in range.
384  * @mapping:           address space within which to check
385  * @start_byte:        offset in bytes where the range starts
386  * @end_byte:          offset in bytes where the range ends (inclusive)
387  *
388  * Find at least one page in the range supplied, usually used to check if
389  * direct writing in this range will trigger a writeback.
390  */
391 bool filemap_range_has_page(struct address_space *mapping,
392                            loff_t start_byte, loff_t end_byte)
393 {
394         pgoff_t index = start_byte >> PAGE_SHIFT;
395         pgoff_t end = end_byte >> PAGE_SHIFT;
396         struct pagevec pvec;
397         bool ret;
398
399         if (end_byte < start_byte)
400                 return false;
401
402         if (mapping->nrpages == 0)
403                 return false;
404
405         pagevec_init(&pvec, 0);
406         if (!pagevec_lookup(&pvec, mapping, index, 1))
407                 return false;
408         ret = (pvec.pages[0]->index <= end);
409         pagevec_release(&pvec);
410         return ret;
411 }
412 EXPORT_SYMBOL(filemap_range_has_page);
413
414 static void __filemap_fdatawait_range(struct address_space *mapping,
415                                      loff_t start_byte, loff_t end_byte)
416 {
417         pgoff_t index = start_byte >> PAGE_SHIFT;
418         pgoff_t end = end_byte >> PAGE_SHIFT;
419         struct pagevec pvec;
420         int nr_pages;
421
422         if (end_byte < start_byte)
423                 return;
424
425         pagevec_init(&pvec, 0);
426         while ((index <= end) &&
427                         (nr_pages = pagevec_lookup_tag(&pvec, mapping, &index,
428                         PAGECACHE_TAG_WRITEBACK,
429                         min(end - index, (pgoff_t)PAGEVEC_SIZE-1) + 1)) != 0) {
430                 unsigned i;
431
432                 for (i = 0; i < nr_pages; i++) {
433                         struct page *page = pvec.pages[i];
434
435                         /* until radix tree lookup accepts end_index */
436                         if (page->index > end)
437                                 continue;
438
439                         wait_on_page_writeback(page);
440                         ClearPageError(page);
441                 }
442                 pagevec_release(&pvec);
443                 cond_resched();
444         }
445 }
446
447 /**
448  * filemap_fdatawait_range - wait for writeback to complete
449  * @mapping:            address space structure to wait for
450  * @start_byte:         offset in bytes where the range starts
451  * @end_byte:           offset in bytes where the range ends (inclusive)
452  *
453  * Walk the list of under-writeback pages of the given address space
454  * in the given range and wait for all of them.  Check error status of
455  * the address space and return it.
456  *
457  * Since the error status of the address space is cleared by this function,
458  * callers are responsible for checking the return value and handling and/or
459  * reporting the error.
460  */
461 int filemap_fdatawait_range(struct address_space *mapping, loff_t start_byte,
462                             loff_t end_byte)
463 {
464         __filemap_fdatawait_range(mapping, start_byte, end_byte);
465         return filemap_check_errors(mapping);
466 }
467 EXPORT_SYMBOL(filemap_fdatawait_range);
468
469 /**
470  * filemap_fdatawait_keep_errors - wait for writeback without clearing errors
471  * @mapping: address space structure to wait for
472  *
473  * Walk the list of under-writeback pages of the given address space
474  * and wait for all of them.  Unlike filemap_fdatawait(), this function
475  * does not clear error status of the address space.
476  *
477  * Use this function if callers don't handle errors themselves.  Expected
478  * call sites are system-wide / filesystem-wide data flushers: e.g. sync(2),
479  * fsfreeze(8)
480  */
481 int filemap_fdatawait_keep_errors(struct address_space *mapping)
482 {
483         loff_t i_size = i_size_read(mapping->host);
484
485         if (i_size == 0)
486                 return 0;
487
488         __filemap_fdatawait_range(mapping, 0, i_size - 1);
489         return filemap_check_and_keep_errors(mapping);
490 }
491 EXPORT_SYMBOL(filemap_fdatawait_keep_errors);
492
493 /**
494  * filemap_fdatawait - wait for all under-writeback pages to complete
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.  Check error status of the address space
499  * and return it.
500  *
501  * Since the error status of the address space is cleared by this function,
502  * callers are responsible for checking the return value and handling and/or
503  * reporting the error.
504  */
505 int filemap_fdatawait(struct address_space *mapping)
506 {
507         loff_t i_size = i_size_read(mapping->host);
508
509         if (i_size == 0)
510                 return 0;
511
512         return filemap_fdatawait_range(mapping, 0, i_size - 1);
513 }
514 EXPORT_SYMBOL(filemap_fdatawait);
515
516 int filemap_write_and_wait(struct address_space *mapping)
517 {
518         int err = 0;
519
520         if ((!dax_mapping(mapping) && mapping->nrpages) ||
521             (dax_mapping(mapping) && mapping->nrexceptional)) {
522                 err = filemap_fdatawrite(mapping);
523                 /*
524                  * Even if the above returned error, the pages may be
525                  * written partially (e.g. -ENOSPC), so we wait for it.
526                  * But the -EIO is special case, it may indicate the worst
527                  * thing (e.g. bug) happened, so we avoid waiting for it.
528                  */
529                 if (err != -EIO) {
530                         int err2 = filemap_fdatawait(mapping);
531                         if (!err)
532                                 err = err2;
533                 } else {
534                         /* Clear any previously stored errors */
535                         filemap_check_errors(mapping);
536                 }
537         } else {
538                 err = filemap_check_errors(mapping);
539         }
540         return err;
541 }
542 EXPORT_SYMBOL(filemap_write_and_wait);
543
544 /**
545  * filemap_write_and_wait_range - write out & wait on a file range
546  * @mapping:    the address_space for the pages
547  * @lstart:     offset in bytes where the range starts
548  * @lend:       offset in bytes where the range ends (inclusive)
549  *
550  * Write out and wait upon file offsets lstart->lend, inclusive.
551  *
552  * Note that @lend is inclusive (describes the last byte to be written) so
553  * that this function can be used to write to the very end-of-file (end = -1).
554  */
555 int filemap_write_and_wait_range(struct address_space *mapping,
556                                  loff_t lstart, loff_t lend)
557 {
558         int err = 0;
559
560         if ((!dax_mapping(mapping) && mapping->nrpages) ||
561             (dax_mapping(mapping) && mapping->nrexceptional)) {
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         return err;
628 }
629 EXPORT_SYMBOL(file_check_and_advance_wb_err);
630
631 /**
632  * file_write_and_wait_range - write out & wait on a file range
633  * @file:       file pointing to address_space with pages
634  * @lstart:     offset in bytes where the range starts
635  * @lend:       offset in bytes where the range ends (inclusive)
636  *
637  * Write out and wait upon file offsets lstart->lend, inclusive.
638  *
639  * Note that @lend is inclusive (describes the last byte to be written) so
640  * that this function can be used to write to the very end-of-file (end = -1).
641  *
642  * After writing out and waiting on the data, we check and advance the
643  * f_wb_err cursor to the latest value, and return any errors detected there.
644  */
645 int file_write_and_wait_range(struct file *file, loff_t lstart, loff_t lend)
646 {
647         int err = 0, err2;
648         struct address_space *mapping = file->f_mapping;
649
650         if ((!dax_mapping(mapping) && mapping->nrpages) ||
651             (dax_mapping(mapping) && mapping->nrexceptional)) {
652                 err = __filemap_fdatawrite_range(mapping, lstart, lend,
653                                                  WB_SYNC_ALL);
654                 /* See comment of filemap_write_and_wait() */
655                 if (err != -EIO)
656                         __filemap_fdatawait_range(mapping, lstart, lend);
657         }
658         err2 = file_check_and_advance_wb_err(file);
659         if (!err)
660                 err = err2;
661         return err;
662 }
663 EXPORT_SYMBOL(file_write_and_wait_range);
664
665 /**
666  * replace_page_cache_page - replace a pagecache page with a new one
667  * @old:        page to be replaced
668  * @new:        page to replace with
669  * @gfp_mask:   allocation mode
670  *
671  * This function replaces a page in the pagecache with a new one.  On
672  * success it acquires the pagecache reference for the new page and
673  * drops it for the old page.  Both the old and new pages must be
674  * locked.  This function does not add the new page to the LRU, the
675  * caller must do that.
676  *
677  * The remove + add is atomic.  The only way this function can fail is
678  * memory allocation failure.
679  */
680 int replace_page_cache_page(struct page *old, struct page *new, gfp_t gfp_mask)
681 {
682         int error;
683
684         VM_BUG_ON_PAGE(!PageLocked(old), old);
685         VM_BUG_ON_PAGE(!PageLocked(new), new);
686         VM_BUG_ON_PAGE(new->mapping, new);
687
688         error = radix_tree_preload(gfp_mask & ~__GFP_HIGHMEM);
689         if (!error) {
690                 struct address_space *mapping = old->mapping;
691                 void (*freepage)(struct page *);
692                 unsigned long flags;
693
694                 pgoff_t offset = old->index;
695                 freepage = mapping->a_ops->freepage;
696
697                 get_page(new);
698                 new->mapping = mapping;
699                 new->index = offset;
700
701                 spin_lock_irqsave(&mapping->tree_lock, flags);
702                 __delete_from_page_cache(old, NULL);
703                 error = page_cache_tree_insert(mapping, new, NULL);
704                 BUG_ON(error);
705
706                 /*
707                  * hugetlb pages do not participate in page cache accounting.
708                  */
709                 if (!PageHuge(new))
710                         __inc_node_page_state(new, NR_FILE_PAGES);
711                 if (PageSwapBacked(new))
712                         __inc_node_page_state(new, NR_SHMEM);
713                 spin_unlock_irqrestore(&mapping->tree_lock, flags);
714                 mem_cgroup_migrate(old, new);
715                 radix_tree_preload_end();
716                 if (freepage)
717                         freepage(old);
718                 put_page(old);
719         }
720
721         return error;
722 }
723 EXPORT_SYMBOL_GPL(replace_page_cache_page);
724
725 static int __add_to_page_cache_locked(struct page *page,
726                                       struct address_space *mapping,
727                                       pgoff_t offset, gfp_t gfp_mask,
728                                       void **shadowp)
729 {
730         int huge = PageHuge(page);
731         struct mem_cgroup *memcg;
732         int error;
733
734         VM_BUG_ON_PAGE(!PageLocked(page), page);
735         VM_BUG_ON_PAGE(PageSwapBacked(page), page);
736
737         if (!huge) {
738                 error = mem_cgroup_try_charge(page, current->mm,
739                                               gfp_mask, &memcg, false);
740                 if (error)
741                         return error;
742         }
743
744         error = radix_tree_maybe_preload(gfp_mask & ~__GFP_HIGHMEM);
745         if (error) {
746                 if (!huge)
747                         mem_cgroup_cancel_charge(page, memcg, false);
748                 return error;
749         }
750
751         get_page(page);
752         page->mapping = mapping;
753         page->index = offset;
754
755         spin_lock_irq(&mapping->tree_lock);
756         error = page_cache_tree_insert(mapping, page, shadowp);
757         radix_tree_preload_end();
758         if (unlikely(error))
759                 goto err_insert;
760
761         /* hugetlb pages do not participate in page cache accounting. */
762         if (!huge)
763                 __inc_node_page_state(page, NR_FILE_PAGES);
764         spin_unlock_irq(&mapping->tree_lock);
765         if (!huge)
766                 mem_cgroup_commit_charge(page, memcg, false, false);
767         trace_mm_filemap_add_to_page_cache(page);
768         return 0;
769 err_insert:
770         page->mapping = NULL;
771         /* Leave page->index set: truncation relies upon it */
772         spin_unlock_irq(&mapping->tree_lock);
773         if (!huge)
774                 mem_cgroup_cancel_charge(page, memcg, false);
775         put_page(page);
776         return error;
777 }
778
779 /**
780  * add_to_page_cache_locked - add a locked page to the pagecache
781  * @page:       page to add
782  * @mapping:    the page's address_space
783  * @offset:     page index
784  * @gfp_mask:   page allocation mode
785  *
786  * This function is used to add a page to the pagecache. It must be locked.
787  * This function does not add the page to the LRU.  The caller must do that.
788  */
789 int add_to_page_cache_locked(struct page *page, struct address_space *mapping,
790                 pgoff_t offset, gfp_t gfp_mask)
791 {
792         return __add_to_page_cache_locked(page, mapping, offset,
793                                           gfp_mask, NULL);
794 }
795 EXPORT_SYMBOL(add_to_page_cache_locked);
796
797 int add_to_page_cache_lru(struct page *page, struct address_space *mapping,
798                                 pgoff_t offset, gfp_t gfp_mask)
799 {
800         void *shadow = NULL;
801         int ret;
802
803         __SetPageLocked(page);
804         ret = __add_to_page_cache_locked(page, mapping, offset,
805                                          gfp_mask, &shadow);
806         if (unlikely(ret))
807                 __ClearPageLocked(page);
808         else {
809                 /*
810                  * The page might have been evicted from cache only
811                  * recently, in which case it should be activated like
812                  * any other repeatedly accessed page.
813                  * The exception is pages getting rewritten; evicting other
814                  * data from the working set, only to cache data that will
815                  * get overwritten with something else, is a waste of memory.
816                  */
817                 if (!(gfp_mask & __GFP_WRITE) &&
818                     shadow && workingset_refault(shadow)) {
819                         SetPageActive(page);
820                         workingset_activation(page);
821                 } else
822                         ClearPageActive(page);
823                 lru_cache_add(page);
824         }
825         return ret;
826 }
827 EXPORT_SYMBOL_GPL(add_to_page_cache_lru);
828
829 #ifdef CONFIG_NUMA
830 struct page *__page_cache_alloc(gfp_t gfp)
831 {
832         int n;
833         struct page *page;
834
835         if (cpuset_do_page_mem_spread()) {
836                 unsigned int cpuset_mems_cookie;
837                 do {
838                         cpuset_mems_cookie = read_mems_allowed_begin();
839                         n = cpuset_mem_spread_node();
840                         page = __alloc_pages_node(n, gfp, 0);
841                 } while (!page && read_mems_allowed_retry(cpuset_mems_cookie));
842
843                 return page;
844         }
845         return alloc_pages(gfp, 0);
846 }
847 EXPORT_SYMBOL(__page_cache_alloc);
848 #endif
849
850 /*
851  * In order to wait for pages to become available there must be
852  * waitqueues associated with pages. By using a hash table of
853  * waitqueues where the bucket discipline is to maintain all
854  * waiters on the same queue and wake all when any of the pages
855  * become available, and for the woken contexts to check to be
856  * sure the appropriate page became available, this saves space
857  * at a cost of "thundering herd" phenomena during rare hash
858  * collisions.
859  */
860 #define PAGE_WAIT_TABLE_BITS 8
861 #define PAGE_WAIT_TABLE_SIZE (1 << PAGE_WAIT_TABLE_BITS)
862 static wait_queue_head_t page_wait_table[PAGE_WAIT_TABLE_SIZE] __cacheline_aligned;
863
864 static wait_queue_head_t *page_waitqueue(struct page *page)
865 {
866         return &page_wait_table[hash_ptr(page, PAGE_WAIT_TABLE_BITS)];
867 }
868
869 void __init pagecache_init(void)
870 {
871         int i;
872
873         for (i = 0; i < PAGE_WAIT_TABLE_SIZE; i++)
874                 init_waitqueue_head(&page_wait_table[i]);
875
876         page_writeback_init();
877 }
878
879 /* This has the same layout as wait_bit_key - see fs/cachefiles/rdwr.c */
880 struct wait_page_key {
881         struct page *page;
882         int bit_nr;
883         int page_match;
884 };
885
886 struct wait_page_queue {
887         struct page *page;
888         int bit_nr;
889         wait_queue_entry_t wait;
890 };
891
892 static int wake_page_function(wait_queue_entry_t *wait, unsigned mode, int sync, void *arg)
893 {
894         struct wait_page_key *key = arg;
895         struct wait_page_queue *wait_page
896                 = container_of(wait, struct wait_page_queue, wait);
897
898         if (wait_page->page != key->page)
899                return 0;
900         key->page_match = 1;
901
902         if (wait_page->bit_nr != key->bit_nr)
903                 return 0;
904
905         /* Stop walking if it's locked */
906         if (test_bit(key->bit_nr, &key->page->flags))
907                 return -1;
908
909         return autoremove_wake_function(wait, mode, sync, key);
910 }
911
912 static void wake_up_page_bit(struct page *page, int bit_nr)
913 {
914         wait_queue_head_t *q = page_waitqueue(page);
915         struct wait_page_key key;
916         unsigned long flags;
917
918         key.page = page;
919         key.bit_nr = bit_nr;
920         key.page_match = 0;
921
922         spin_lock_irqsave(&q->lock, flags);
923         __wake_up_locked_key(q, TASK_NORMAL, &key);
924         /*
925          * It is possible for other pages to have collided on the waitqueue
926          * hash, so in that case check for a page match. That prevents a long-
927          * term waiter
928          *
929          * It is still possible to miss a case here, when we woke page waiters
930          * and removed them from the waitqueue, but there are still other
931          * page waiters.
932          */
933         if (!waitqueue_active(q) || !key.page_match) {
934                 ClearPageWaiters(page);
935                 /*
936                  * It's possible to miss clearing Waiters here, when we woke
937                  * our page waiters, but the hashed waitqueue has waiters for
938                  * other pages on it.
939                  *
940                  * That's okay, it's a rare case. The next waker will clear it.
941                  */
942         }
943         spin_unlock_irqrestore(&q->lock, flags);
944 }
945
946 static void wake_up_page(struct page *page, int bit)
947 {
948         if (!PageWaiters(page))
949                 return;
950         wake_up_page_bit(page, bit);
951 }
952
953 static inline int wait_on_page_bit_common(wait_queue_head_t *q,
954                 struct page *page, int bit_nr, int state, bool lock)
955 {
956         struct wait_page_queue wait_page;
957         wait_queue_entry_t *wait = &wait_page.wait;
958         int ret = 0;
959
960         init_wait(wait);
961         wait->flags = lock ? WQ_FLAG_EXCLUSIVE : 0;
962         wait->func = wake_page_function;
963         wait_page.page = page;
964         wait_page.bit_nr = bit_nr;
965
966         for (;;) {
967                 spin_lock_irq(&q->lock);
968
969                 if (likely(list_empty(&wait->entry))) {
970                         __add_wait_queue_entry_tail(q, wait);
971                         SetPageWaiters(page);
972                 }
973
974                 set_current_state(state);
975
976                 spin_unlock_irq(&q->lock);
977
978                 if (likely(test_bit(bit_nr, &page->flags))) {
979                         io_schedule();
980                 }
981
982                 if (lock) {
983                         if (!test_and_set_bit_lock(bit_nr, &page->flags))
984                                 break;
985                 } else {
986                         if (!test_bit(bit_nr, &page->flags))
987                                 break;
988                 }
989
990                 if (unlikely(signal_pending_state(state, current))) {
991                         ret = -EINTR;
992                         break;
993                 }
994         }
995
996         finish_wait(q, wait);
997
998         /*
999          * A signal could leave PageWaiters set. Clearing it here if
1000          * !waitqueue_active would be possible (by open-coding finish_wait),
1001          * but still fail to catch it in the case of wait hash collision. We
1002          * already can fail to clear wait hash collision cases, so don't
1003          * bother with signals either.
1004          */
1005
1006         return ret;
1007 }
1008
1009 void wait_on_page_bit(struct page *page, int bit_nr)
1010 {
1011         wait_queue_head_t *q = page_waitqueue(page);
1012         wait_on_page_bit_common(q, page, bit_nr, TASK_UNINTERRUPTIBLE, false);
1013 }
1014 EXPORT_SYMBOL(wait_on_page_bit);
1015
1016 int wait_on_page_bit_killable(struct page *page, int bit_nr)
1017 {
1018         wait_queue_head_t *q = page_waitqueue(page);
1019         return wait_on_page_bit_common(q, page, bit_nr, TASK_KILLABLE, false);
1020 }
1021
1022 /**
1023  * add_page_wait_queue - Add an arbitrary waiter to a page's wait queue
1024  * @page: Page defining the wait queue of interest
1025  * @waiter: Waiter to add to the queue
1026  *
1027  * Add an arbitrary @waiter to the wait queue for the nominated @page.
1028  */
1029 void add_page_wait_queue(struct page *page, wait_queue_entry_t *waiter)
1030 {
1031         wait_queue_head_t *q = page_waitqueue(page);
1032         unsigned long flags;
1033
1034         spin_lock_irqsave(&q->lock, flags);
1035         __add_wait_queue_entry_tail(q, waiter);
1036         SetPageWaiters(page);
1037         spin_unlock_irqrestore(&q->lock, flags);
1038 }
1039 EXPORT_SYMBOL_GPL(add_page_wait_queue);
1040
1041 #ifndef clear_bit_unlock_is_negative_byte
1042
1043 /*
1044  * PG_waiters is the high bit in the same byte as PG_lock.
1045  *
1046  * On x86 (and on many other architectures), we can clear PG_lock and
1047  * test the sign bit at the same time. But if the architecture does
1048  * not support that special operation, we just do this all by hand
1049  * instead.
1050  *
1051  * The read of PG_waiters has to be after (or concurrently with) PG_locked
1052  * being cleared, but a memory barrier should be unneccssary since it is
1053  * in the same byte as PG_locked.
1054  */
1055 static inline bool clear_bit_unlock_is_negative_byte(long nr, volatile void *mem)
1056 {
1057         clear_bit_unlock(nr, mem);
1058         /* smp_mb__after_atomic(); */
1059         return test_bit(PG_waiters, mem);
1060 }
1061
1062 #endif
1063
1064 /**
1065  * unlock_page - unlock a locked page
1066  * @page: the page
1067  *
1068  * Unlocks the page and wakes up sleepers in ___wait_on_page_locked().
1069  * Also wakes sleepers in wait_on_page_writeback() because the wakeup
1070  * mechanism between PageLocked pages and PageWriteback pages is shared.
1071  * But that's OK - sleepers in wait_on_page_writeback() just go back to sleep.
1072  *
1073  * Note that this depends on PG_waiters being the sign bit in the byte
1074  * that contains PG_locked - thus the BUILD_BUG_ON(). That allows us to
1075  * clear the PG_locked bit and test PG_waiters at the same time fairly
1076  * portably (architectures that do LL/SC can test any bit, while x86 can
1077  * test the sign bit).
1078  */
1079 void unlock_page(struct page *page)
1080 {
1081         BUILD_BUG_ON(PG_waiters != 7);
1082         page = compound_head(page);
1083         VM_BUG_ON_PAGE(!PageLocked(page), page);
1084         if (clear_bit_unlock_is_negative_byte(PG_locked, &page->flags))
1085                 wake_up_page_bit(page, PG_locked);
1086 }
1087 EXPORT_SYMBOL(unlock_page);
1088
1089 /**
1090  * end_page_writeback - end writeback against a page
1091  * @page: the page
1092  */
1093 void end_page_writeback(struct page *page)
1094 {
1095         /*
1096          * TestClearPageReclaim could be used here but it is an atomic
1097          * operation and overkill in this particular case. Failing to
1098          * shuffle a page marked for immediate reclaim is too mild to
1099          * justify taking an atomic operation penalty at the end of
1100          * ever page writeback.
1101          */
1102         if (PageReclaim(page)) {
1103                 ClearPageReclaim(page);
1104                 rotate_reclaimable_page(page);
1105         }
1106
1107         if (!test_clear_page_writeback(page))
1108                 BUG();
1109
1110         smp_mb__after_atomic();
1111         wake_up_page(page, PG_writeback);
1112 }
1113 EXPORT_SYMBOL(end_page_writeback);
1114
1115 /*
1116  * After completing I/O on a page, call this routine to update the page
1117  * flags appropriately
1118  */
1119 void page_endio(struct page *page, bool is_write, int err)
1120 {
1121         if (!is_write) {
1122                 if (!err) {
1123                         SetPageUptodate(page);
1124                 } else {
1125                         ClearPageUptodate(page);
1126                         SetPageError(page);
1127                 }
1128                 unlock_page(page);
1129         } else {
1130                 if (err) {
1131                         struct address_space *mapping;
1132
1133                         SetPageError(page);
1134                         mapping = page_mapping(page);
1135                         if (mapping)
1136                                 mapping_set_error(mapping, err);
1137                 }
1138                 end_page_writeback(page);
1139         }
1140 }
1141 EXPORT_SYMBOL_GPL(page_endio);
1142
1143 /**
1144  * __lock_page - get a lock on the page, assuming we need to sleep to get it
1145  * @__page: the page to lock
1146  */
1147 void __lock_page(struct page *__page)
1148 {
1149         struct page *page = compound_head(__page);
1150         wait_queue_head_t *q = page_waitqueue(page);
1151         wait_on_page_bit_common(q, page, PG_locked, TASK_UNINTERRUPTIBLE, true);
1152 }
1153 EXPORT_SYMBOL(__lock_page);
1154
1155 int __lock_page_killable(struct page *__page)
1156 {
1157         struct page *page = compound_head(__page);
1158         wait_queue_head_t *q = page_waitqueue(page);
1159         return wait_on_page_bit_common(q, page, PG_locked, TASK_KILLABLE, true);
1160 }
1161 EXPORT_SYMBOL_GPL(__lock_page_killable);
1162
1163 /*
1164  * Return values:
1165  * 1 - page is locked; mmap_sem is still held.
1166  * 0 - page is not locked.
1167  *     mmap_sem has been released (up_read()), unless flags had both
1168  *     FAULT_FLAG_ALLOW_RETRY and FAULT_FLAG_RETRY_NOWAIT set, in
1169  *     which case mmap_sem is still held.
1170  *
1171  * If neither ALLOW_RETRY nor KILLABLE are set, will always return 1
1172  * with the page locked and the mmap_sem unperturbed.
1173  */
1174 int __lock_page_or_retry(struct page *page, struct mm_struct *mm,
1175                          unsigned int flags)
1176 {
1177         if (flags & FAULT_FLAG_ALLOW_RETRY) {
1178                 /*
1179                  * CAUTION! In this case, mmap_sem is not released
1180                  * even though return 0.
1181                  */
1182                 if (flags & FAULT_FLAG_RETRY_NOWAIT)
1183                         return 0;
1184
1185                 up_read(&mm->mmap_sem);
1186                 if (flags & FAULT_FLAG_KILLABLE)
1187                         wait_on_page_locked_killable(page);
1188                 else
1189                         wait_on_page_locked(page);
1190                 return 0;
1191         } else {
1192                 if (flags & FAULT_FLAG_KILLABLE) {
1193                         int ret;
1194
1195                         ret = __lock_page_killable(page);
1196                         if (ret) {
1197                                 up_read(&mm->mmap_sem);
1198                                 return 0;
1199                         }
1200                 } else
1201                         __lock_page(page);
1202                 return 1;
1203         }
1204 }
1205
1206 /**
1207  * page_cache_next_hole - find the next hole (not-present entry)
1208  * @mapping: mapping
1209  * @index: index
1210  * @max_scan: maximum range to search
1211  *
1212  * Search the set [index, min(index+max_scan-1, MAX_INDEX)] for the
1213  * lowest indexed hole.
1214  *
1215  * Returns: the index of the hole if found, otherwise returns an index
1216  * outside of the set specified (in which case 'return - index >=
1217  * max_scan' will be true). In rare cases of index wrap-around, 0 will
1218  * be returned.
1219  *
1220  * page_cache_next_hole may be called under rcu_read_lock. However,
1221  * like radix_tree_gang_lookup, this will not atomically search a
1222  * snapshot of the tree at a single point in time. For example, if a
1223  * hole is created at index 5, then subsequently a hole is created at
1224  * index 10, page_cache_next_hole covering both indexes may return 10
1225  * if called under rcu_read_lock.
1226  */
1227 pgoff_t page_cache_next_hole(struct address_space *mapping,
1228                              pgoff_t index, unsigned long max_scan)
1229 {
1230         unsigned long i;
1231
1232         for (i = 0; i < max_scan; i++) {
1233                 struct page *page;
1234
1235                 page = radix_tree_lookup(&mapping->page_tree, index);
1236                 if (!page || radix_tree_exceptional_entry(page))
1237                         break;
1238                 index++;
1239                 if (index == 0)
1240                         break;
1241         }
1242
1243         return index;
1244 }
1245 EXPORT_SYMBOL(page_cache_next_hole);
1246
1247 /**
1248  * page_cache_prev_hole - find the prev hole (not-present entry)
1249  * @mapping: mapping
1250  * @index: index
1251  * @max_scan: maximum range to search
1252  *
1253  * Search backwards in the range [max(index-max_scan+1, 0), index] for
1254  * the first hole.
1255  *
1256  * Returns: the index of the hole if found, otherwise returns an index
1257  * outside of the set specified (in which case 'index - return >=
1258  * max_scan' will be true). In rare cases of wrap-around, ULONG_MAX
1259  * will be returned.
1260  *
1261  * page_cache_prev_hole may be called under rcu_read_lock. However,
1262  * like radix_tree_gang_lookup, this will not atomically search a
1263  * snapshot of the tree at a single point in time. For example, if a
1264  * hole is created at index 10, then subsequently a hole is created at
1265  * index 5, page_cache_prev_hole covering both indexes may return 5 if
1266  * called under rcu_read_lock.
1267  */
1268 pgoff_t page_cache_prev_hole(struct address_space *mapping,
1269                              pgoff_t index, unsigned long max_scan)
1270 {
1271         unsigned long i;
1272
1273         for (i = 0; i < max_scan; i++) {
1274                 struct page *page;
1275
1276                 page = radix_tree_lookup(&mapping->page_tree, index);
1277                 if (!page || radix_tree_exceptional_entry(page))
1278                         break;
1279                 index--;
1280                 if (index == ULONG_MAX)
1281                         break;
1282         }
1283
1284         return index;
1285 }
1286 EXPORT_SYMBOL(page_cache_prev_hole);
1287
1288 /**
1289  * find_get_entry - find and get a page cache entry
1290  * @mapping: the address_space to search
1291  * @offset: the page cache index
1292  *
1293  * Looks up the page cache slot at @mapping & @offset.  If there is a
1294  * page cache page, it is returned with an increased refcount.
1295  *
1296  * If the slot holds a shadow entry of a previously evicted page, or a
1297  * swap entry from shmem/tmpfs, it is returned.
1298  *
1299  * Otherwise, %NULL is returned.
1300  */
1301 struct page *find_get_entry(struct address_space *mapping, pgoff_t offset)
1302 {
1303         void **pagep;
1304         struct page *head, *page;
1305
1306         rcu_read_lock();
1307 repeat:
1308         page = NULL;
1309         pagep = radix_tree_lookup_slot(&mapping->page_tree, offset);
1310         if (pagep) {
1311                 page = radix_tree_deref_slot(pagep);
1312                 if (unlikely(!page))
1313                         goto out;
1314                 if (radix_tree_exception(page)) {
1315                         if (radix_tree_deref_retry(page))
1316                                 goto repeat;
1317                         /*
1318                          * A shadow entry of a recently evicted page,
1319                          * or a swap entry from shmem/tmpfs.  Return
1320                          * it without attempting to raise page count.
1321                          */
1322                         goto out;
1323                 }
1324
1325                 head = compound_head(page);
1326                 if (!page_cache_get_speculative(head))
1327                         goto repeat;
1328
1329                 /* The page was split under us? */
1330                 if (compound_head(page) != head) {
1331                         put_page(head);
1332                         goto repeat;
1333                 }
1334
1335                 /*
1336                  * Has the page moved?
1337                  * This is part of the lockless pagecache protocol. See
1338                  * include/linux/pagemap.h for details.
1339                  */
1340                 if (unlikely(page != *pagep)) {
1341                         put_page(head);
1342                         goto repeat;
1343                 }
1344         }
1345 out:
1346         rcu_read_unlock();
1347
1348         return page;
1349 }
1350 EXPORT_SYMBOL(find_get_entry);
1351
1352 /**
1353  * find_lock_entry - locate, pin and lock a page cache entry
1354  * @mapping: the address_space to search
1355  * @offset: the page cache index
1356  *
1357  * Looks up the page cache slot at @mapping & @offset.  If there is a
1358  * page cache page, it is returned locked and with an increased
1359  * refcount.
1360  *
1361  * If the slot holds a shadow entry of a previously evicted page, or a
1362  * swap entry from shmem/tmpfs, it is returned.
1363  *
1364  * Otherwise, %NULL is returned.
1365  *
1366  * find_lock_entry() may sleep.
1367  */
1368 struct page *find_lock_entry(struct address_space *mapping, pgoff_t offset)
1369 {
1370         struct page *page;
1371
1372 repeat:
1373         page = find_get_entry(mapping, offset);
1374         if (page && !radix_tree_exception(page)) {
1375                 lock_page(page);
1376                 /* Has the page been truncated? */
1377                 if (unlikely(page_mapping(page) != mapping)) {
1378                         unlock_page(page);
1379                         put_page(page);
1380                         goto repeat;
1381                 }
1382                 VM_BUG_ON_PAGE(page_to_pgoff(page) != offset, page);
1383         }
1384         return page;
1385 }
1386 EXPORT_SYMBOL(find_lock_entry);
1387
1388 /**
1389  * pagecache_get_page - find and get a page reference
1390  * @mapping: the address_space to search
1391  * @offset: the page index
1392  * @fgp_flags: PCG flags
1393  * @gfp_mask: gfp mask to use for the page cache data page allocation
1394  *
1395  * Looks up the page cache slot at @mapping & @offset.
1396  *
1397  * PCG flags modify how the page is returned.
1398  *
1399  * @fgp_flags can be:
1400  *
1401  * - FGP_ACCESSED: the page will be marked accessed
1402  * - FGP_LOCK: Page is return locked
1403  * - FGP_CREAT: If page is not present then a new page is allocated using
1404  *   @gfp_mask and added to the page cache and the VM's LRU
1405  *   list. The page is returned locked and with an increased
1406  *   refcount. Otherwise, NULL is returned.
1407  *
1408  * If FGP_LOCK or FGP_CREAT are specified then the function may sleep even
1409  * if the GFP flags specified for FGP_CREAT are atomic.
1410  *
1411  * If there is a page cache page, it is returned with an increased refcount.
1412  */
1413 struct page *pagecache_get_page(struct address_space *mapping, pgoff_t offset,
1414         int fgp_flags, gfp_t gfp_mask)
1415 {
1416         struct page *page;
1417
1418 repeat:
1419         page = find_get_entry(mapping, offset);
1420         if (radix_tree_exceptional_entry(page))
1421                 page = NULL;
1422         if (!page)
1423                 goto no_page;
1424
1425         if (fgp_flags & FGP_LOCK) {
1426                 if (fgp_flags & FGP_NOWAIT) {
1427                         if (!trylock_page(page)) {
1428                                 put_page(page);
1429                                 return NULL;
1430                         }
1431                 } else {
1432                         lock_page(page);
1433                 }
1434
1435                 /* Has the page been truncated? */
1436                 if (unlikely(page->mapping != mapping)) {
1437                         unlock_page(page);
1438                         put_page(page);
1439                         goto repeat;
1440                 }
1441                 VM_BUG_ON_PAGE(page->index != offset, page);
1442         }
1443
1444         if (page && (fgp_flags & FGP_ACCESSED))
1445                 mark_page_accessed(page);
1446
1447 no_page:
1448         if (!page && (fgp_flags & FGP_CREAT)) {
1449                 int err;
1450                 if ((fgp_flags & FGP_WRITE) && mapping_cap_account_dirty(mapping))
1451                         gfp_mask |= __GFP_WRITE;
1452                 if (fgp_flags & FGP_NOFS)
1453                         gfp_mask &= ~__GFP_FS;
1454
1455                 page = __page_cache_alloc(gfp_mask);
1456                 if (!page)
1457                         return NULL;
1458
1459                 if (WARN_ON_ONCE(!(fgp_flags & FGP_LOCK)))
1460                         fgp_flags |= FGP_LOCK;
1461
1462                 /* Init accessed so avoid atomic mark_page_accessed later */
1463                 if (fgp_flags & FGP_ACCESSED)
1464                         __SetPageReferenced(page);
1465
1466                 err = add_to_page_cache_lru(page, mapping, offset,
1467                                 gfp_mask & GFP_RECLAIM_MASK);
1468                 if (unlikely(err)) {
1469                         put_page(page);
1470                         page = NULL;
1471                         if (err == -EEXIST)
1472                                 goto repeat;
1473                 }
1474         }
1475
1476         return page;
1477 }
1478 EXPORT_SYMBOL(pagecache_get_page);
1479
1480 /**
1481  * find_get_entries - gang pagecache lookup
1482  * @mapping:    The address_space to search
1483  * @start:      The starting page cache index
1484  * @nr_entries: The maximum number of entries
1485  * @entries:    Where the resulting entries are placed
1486  * @indices:    The cache indices corresponding to the entries in @entries
1487  *
1488  * find_get_entries() will search for and return a group of up to
1489  * @nr_entries entries in the mapping.  The entries are placed at
1490  * @entries.  find_get_entries() takes a reference against any actual
1491  * pages it returns.
1492  *
1493  * The search returns a group of mapping-contiguous page cache entries
1494  * with ascending indexes.  There may be holes in the indices due to
1495  * not-present pages.
1496  *
1497  * Any shadow entries of evicted pages, or swap entries from
1498  * shmem/tmpfs, are included in the returned array.
1499  *
1500  * find_get_entries() returns the number of pages and shadow entries
1501  * which were found.
1502  */
1503 unsigned find_get_entries(struct address_space *mapping,
1504                           pgoff_t start, unsigned int nr_entries,
1505                           struct page **entries, pgoff_t *indices)
1506 {
1507         void **slot;
1508         unsigned int ret = 0;
1509         struct radix_tree_iter iter;
1510
1511         if (!nr_entries)
1512                 return 0;
1513
1514         rcu_read_lock();
1515         radix_tree_for_each_slot(slot, &mapping->page_tree, &iter, start) {
1516                 struct page *head, *page;
1517 repeat:
1518                 page = radix_tree_deref_slot(slot);
1519                 if (unlikely(!page))
1520                         continue;
1521                 if (radix_tree_exception(page)) {
1522                         if (radix_tree_deref_retry(page)) {
1523                                 slot = radix_tree_iter_retry(&iter);
1524                                 continue;
1525                         }
1526                         /*
1527                          * A shadow entry of a recently evicted page, a swap
1528                          * entry from shmem/tmpfs or a DAX entry.  Return it
1529                          * without attempting to raise page count.
1530                          */
1531                         goto export;
1532                 }
1533
1534                 head = compound_head(page);
1535                 if (!page_cache_get_speculative(head))
1536                         goto repeat;
1537
1538                 /* The page was split under us? */
1539                 if (compound_head(page) != head) {
1540                         put_page(head);
1541                         goto repeat;
1542                 }
1543
1544                 /* Has the page moved? */
1545                 if (unlikely(page != *slot)) {
1546                         put_page(head);
1547                         goto repeat;
1548                 }
1549 export:
1550                 indices[ret] = iter.index;
1551                 entries[ret] = page;
1552                 if (++ret == nr_entries)
1553                         break;
1554         }
1555         rcu_read_unlock();
1556         return ret;
1557 }
1558
1559 /**
1560  * find_get_pages - gang pagecache lookup
1561  * @mapping:    The address_space to search
1562  * @start:      The starting page index
1563  * @nr_pages:   The maximum number of pages
1564  * @pages:      Where the resulting pages are placed
1565  *
1566  * find_get_pages() will search for and return a group of up to
1567  * @nr_pages pages in the mapping.  The pages are placed at @pages.
1568  * find_get_pages() takes a reference against the returned pages.
1569  *
1570  * The search returns a group of mapping-contiguous pages with ascending
1571  * indexes.  There may be holes in the indices due to not-present pages.
1572  *
1573  * find_get_pages() returns the number of pages which were found.
1574  */
1575 unsigned find_get_pages(struct address_space *mapping, pgoff_t start,
1576                             unsigned int nr_pages, struct page **pages)
1577 {
1578         struct radix_tree_iter iter;
1579         void **slot;
1580         unsigned ret = 0;
1581
1582         if (unlikely(!nr_pages))
1583                 return 0;
1584
1585         rcu_read_lock();
1586         radix_tree_for_each_slot(slot, &mapping->page_tree, &iter, start) {
1587                 struct page *head, *page;
1588 repeat:
1589                 page = radix_tree_deref_slot(slot);
1590                 if (unlikely(!page))
1591                         continue;
1592
1593                 if (radix_tree_exception(page)) {
1594                         if (radix_tree_deref_retry(page)) {
1595                                 slot = radix_tree_iter_retry(&iter);
1596                                 continue;
1597                         }
1598                         /*
1599                          * A shadow entry of a recently evicted page,
1600                          * or a swap entry from shmem/tmpfs.  Skip
1601                          * over it.
1602                          */
1603                         continue;
1604                 }
1605
1606                 head = compound_head(page);
1607                 if (!page_cache_get_speculative(head))
1608                         goto repeat;
1609
1610                 /* The page was split under us? */
1611                 if (compound_head(page) != head) {
1612                         put_page(head);
1613                         goto repeat;
1614                 }
1615
1616                 /* Has the page moved? */
1617                 if (unlikely(page != *slot)) {
1618                         put_page(head);
1619                         goto repeat;
1620                 }
1621
1622                 pages[ret] = page;
1623                 if (++ret == nr_pages)
1624                         break;
1625         }
1626
1627         rcu_read_unlock();
1628         return ret;
1629 }
1630
1631 /**
1632  * find_get_pages_contig - gang contiguous pagecache lookup
1633  * @mapping:    The address_space to search
1634  * @index:      The starting page index
1635  * @nr_pages:   The maximum number of pages
1636  * @pages:      Where the resulting pages are placed
1637  *
1638  * find_get_pages_contig() works exactly like find_get_pages(), except
1639  * that the returned number of pages are guaranteed to be contiguous.
1640  *
1641  * find_get_pages_contig() returns the number of pages which were found.
1642  */
1643 unsigned find_get_pages_contig(struct address_space *mapping, pgoff_t index,
1644                                unsigned int nr_pages, struct page **pages)
1645 {
1646         struct radix_tree_iter iter;
1647         void **slot;
1648         unsigned int ret = 0;
1649
1650         if (unlikely(!nr_pages))
1651                 return 0;
1652
1653         rcu_read_lock();
1654         radix_tree_for_each_contig(slot, &mapping->page_tree, &iter, index) {
1655                 struct page *head, *page;
1656 repeat:
1657                 page = radix_tree_deref_slot(slot);
1658                 /* The hole, there no reason to continue */
1659                 if (unlikely(!page))
1660                         break;
1661
1662                 if (radix_tree_exception(page)) {
1663                         if (radix_tree_deref_retry(page)) {
1664                                 slot = radix_tree_iter_retry(&iter);
1665                                 continue;
1666                         }
1667                         /*
1668                          * A shadow entry of a recently evicted page,
1669                          * or a swap entry from shmem/tmpfs.  Stop
1670                          * looking for contiguous pages.
1671                          */
1672                         break;
1673                 }
1674
1675                 head = compound_head(page);
1676                 if (!page_cache_get_speculative(head))
1677                         goto repeat;
1678
1679                 /* The page was split under us? */
1680                 if (compound_head(page) != head) {
1681                         put_page(head);
1682                         goto repeat;
1683                 }
1684
1685                 /* Has the page moved? */
1686                 if (unlikely(page != *slot)) {
1687                         put_page(head);
1688                         goto repeat;
1689                 }
1690
1691                 /*
1692                  * must check mapping and index after taking the ref.
1693                  * otherwise we can get both false positives and false
1694                  * negatives, which is just confusing to the caller.
1695                  */
1696                 if (page->mapping == NULL || page_to_pgoff(page) != iter.index) {
1697                         put_page(page);
1698                         break;
1699                 }
1700
1701                 pages[ret] = page;
1702                 if (++ret == nr_pages)
1703                         break;
1704         }
1705         rcu_read_unlock();
1706         return ret;
1707 }
1708 EXPORT_SYMBOL(find_get_pages_contig);
1709
1710 /**
1711  * find_get_pages_tag - find and return pages that match @tag
1712  * @mapping:    the address_space to search
1713  * @index:      the starting page index
1714  * @tag:        the tag index
1715  * @nr_pages:   the maximum number of pages
1716  * @pages:      where the resulting pages are placed
1717  *
1718  * Like find_get_pages, except we only return pages which are tagged with
1719  * @tag.   We update @index to index the next page for the traversal.
1720  */
1721 unsigned find_get_pages_tag(struct address_space *mapping, pgoff_t *index,
1722                         int tag, unsigned int nr_pages, struct page **pages)
1723 {
1724         struct radix_tree_iter iter;
1725         void **slot;
1726         unsigned ret = 0;
1727
1728         if (unlikely(!nr_pages))
1729                 return 0;
1730
1731         rcu_read_lock();
1732         radix_tree_for_each_tagged(slot, &mapping->page_tree,
1733                                    &iter, *index, tag) {
1734                 struct page *head, *page;
1735 repeat:
1736                 page = radix_tree_deref_slot(slot);
1737                 if (unlikely(!page))
1738                         continue;
1739
1740                 if (radix_tree_exception(page)) {
1741                         if (radix_tree_deref_retry(page)) {
1742                                 slot = radix_tree_iter_retry(&iter);
1743                                 continue;
1744                         }
1745                         /*
1746                          * A shadow entry of a recently evicted page.
1747                          *
1748                          * Those entries should never be tagged, but
1749                          * this tree walk is lockless and the tags are
1750                          * looked up in bulk, one radix tree node at a
1751                          * time, so there is a sizable window for page
1752                          * reclaim to evict a page we saw tagged.
1753                          *
1754                          * Skip over it.
1755                          */
1756                         continue;
1757                 }
1758
1759                 head = compound_head(page);
1760                 if (!page_cache_get_speculative(head))
1761                         goto repeat;
1762
1763                 /* The page was split under us? */
1764                 if (compound_head(page) != head) {
1765                         put_page(head);
1766                         goto repeat;
1767                 }
1768
1769                 /* Has the page moved? */
1770                 if (unlikely(page != *slot)) {
1771                         put_page(head);
1772                         goto repeat;
1773                 }
1774
1775                 pages[ret] = page;
1776                 if (++ret == nr_pages)
1777                         break;
1778         }
1779
1780         rcu_read_unlock();
1781
1782         if (ret)
1783                 *index = pages[ret - 1]->index + 1;
1784
1785         return ret;
1786 }
1787 EXPORT_SYMBOL(find_get_pages_tag);
1788
1789 /**
1790  * find_get_entries_tag - find and return entries that match @tag
1791  * @mapping:    the address_space to search
1792  * @start:      the starting page cache index
1793  * @tag:        the tag index
1794  * @nr_entries: the maximum number of entries
1795  * @entries:    where the resulting entries are placed
1796  * @indices:    the cache indices corresponding to the entries in @entries
1797  *
1798  * Like find_get_entries, except we only return entries which are tagged with
1799  * @tag.
1800  */
1801 unsigned find_get_entries_tag(struct address_space *mapping, pgoff_t start,
1802                         int tag, unsigned int nr_entries,
1803                         struct page **entries, pgoff_t *indices)
1804 {
1805         void **slot;
1806         unsigned int ret = 0;
1807         struct radix_tree_iter iter;
1808
1809         if (!nr_entries)
1810                 return 0;
1811
1812         rcu_read_lock();
1813         radix_tree_for_each_tagged(slot, &mapping->page_tree,
1814                                    &iter, start, tag) {
1815                 struct page *head, *page;
1816 repeat:
1817                 page = radix_tree_deref_slot(slot);
1818                 if (unlikely(!page))
1819                         continue;
1820                 if (radix_tree_exception(page)) {
1821                         if (radix_tree_deref_retry(page)) {
1822                                 slot = radix_tree_iter_retry(&iter);
1823                                 continue;
1824                         }
1825
1826                         /*
1827                          * A shadow entry of a recently evicted page, a swap
1828                          * entry from shmem/tmpfs or a DAX entry.  Return it
1829                          * without attempting to raise page count.
1830                          */
1831                         goto export;
1832                 }
1833
1834                 head = compound_head(page);
1835                 if (!page_cache_get_speculative(head))
1836                         goto repeat;
1837
1838                 /* The page was split under us? */
1839                 if (compound_head(page) != head) {
1840                         put_page(head);
1841                         goto repeat;
1842                 }
1843
1844                 /* Has the page moved? */
1845                 if (unlikely(page != *slot)) {
1846                         put_page(head);
1847                         goto repeat;
1848                 }
1849 export:
1850                 indices[ret] = iter.index;
1851                 entries[ret] = page;
1852                 if (++ret == nr_entries)
1853                         break;
1854         }
1855         rcu_read_unlock();
1856         return ret;
1857 }
1858 EXPORT_SYMBOL(find_get_entries_tag);
1859
1860 /*
1861  * CD/DVDs are error prone. When a medium error occurs, the driver may fail
1862  * a _large_ part of the i/o request. Imagine the worst scenario:
1863  *
1864  *      ---R__________________________________________B__________
1865  *         ^ reading here                             ^ bad block(assume 4k)
1866  *
1867  * read(R) => miss => readahead(R...B) => media error => frustrating retries
1868  * => failing the whole request => read(R) => read(R+1) =>
1869  * readahead(R+1...B+1) => bang => read(R+2) => read(R+3) =>
1870  * readahead(R+3...B+2) => bang => read(R+3) => read(R+4) =>
1871  * readahead(R+4...B+3) => bang => read(R+4) => read(R+5) => ......
1872  *
1873  * It is going insane. Fix it by quickly scaling down the readahead size.
1874  */
1875 static void shrink_readahead_size_eio(struct file *filp,
1876                                         struct file_ra_state *ra)
1877 {
1878         ra->ra_pages /= 4;
1879 }
1880
1881 /**
1882  * do_generic_file_read - generic file read routine
1883  * @filp:       the file to read
1884  * @ppos:       current file position
1885  * @iter:       data destination
1886  * @written:    already copied
1887  *
1888  * This is a generic file read routine, and uses the
1889  * mapping->a_ops->readpage() function for the actual low-level stuff.
1890  *
1891  * This is really ugly. But the goto's actually try to clarify some
1892  * of the logic when it comes to error handling etc.
1893  */
1894 static ssize_t do_generic_file_read(struct file *filp, loff_t *ppos,
1895                 struct iov_iter *iter, ssize_t written)
1896 {
1897         struct address_space *mapping = filp->f_mapping;
1898         struct inode *inode = mapping->host;
1899         struct file_ra_state *ra = &filp->f_ra;
1900         pgoff_t index;
1901         pgoff_t last_index;
1902         pgoff_t prev_index;
1903         unsigned long offset;      /* offset into pagecache page */
1904         unsigned int prev_offset;
1905         int error = 0;
1906
1907         if (unlikely(*ppos >= inode->i_sb->s_maxbytes))
1908                 return 0;
1909         iov_iter_truncate(iter, inode->i_sb->s_maxbytes);
1910
1911         index = *ppos >> PAGE_SHIFT;
1912         prev_index = ra->prev_pos >> PAGE_SHIFT;
1913         prev_offset = ra->prev_pos & (PAGE_SIZE-1);
1914         last_index = (*ppos + iter->count + PAGE_SIZE-1) >> PAGE_SHIFT;
1915         offset = *ppos & ~PAGE_MASK;
1916
1917         for (;;) {
1918                 struct page *page;
1919                 pgoff_t end_index;
1920                 loff_t isize;
1921                 unsigned long nr, ret;
1922
1923                 cond_resched();
1924 find_page:
1925                 if (fatal_signal_pending(current)) {
1926                         error = -EINTR;
1927                         goto out;
1928                 }
1929
1930                 page = find_get_page(mapping, index);
1931                 if (!page) {
1932                         page_cache_sync_readahead(mapping,
1933                                         ra, filp,
1934                                         index, last_index - index);
1935                         page = find_get_page(mapping, index);
1936                         if (unlikely(page == NULL))
1937                                 goto no_cached_page;
1938                 }
1939                 if (PageReadahead(page)) {
1940                         page_cache_async_readahead(mapping,
1941                                         ra, filp, page,
1942                                         index, last_index - index);
1943                 }
1944                 if (!PageUptodate(page)) {
1945                         /*
1946                          * See comment in do_read_cache_page on why
1947                          * wait_on_page_locked is used to avoid unnecessarily
1948                          * serialisations and why it's safe.
1949                          */
1950                         error = wait_on_page_locked_killable(page);
1951                         if (unlikely(error))
1952                                 goto readpage_error;
1953                         if (PageUptodate(page))
1954                                 goto page_ok;
1955
1956                         if (inode->i_blkbits == PAGE_SHIFT ||
1957                                         !mapping->a_ops->is_partially_uptodate)
1958                                 goto page_not_up_to_date;
1959                         /* pipes can't handle partially uptodate pages */
1960                         if (unlikely(iter->type & ITER_PIPE))
1961                                 goto page_not_up_to_date;
1962                         if (!trylock_page(page))
1963                                 goto page_not_up_to_date;
1964                         /* Did it get truncated before we got the lock? */
1965                         if (!page->mapping)
1966                                 goto page_not_up_to_date_locked;
1967                         if (!mapping->a_ops->is_partially_uptodate(page,
1968                                                         offset, iter->count))
1969                                 goto page_not_up_to_date_locked;
1970                         unlock_page(page);
1971                 }
1972 page_ok:
1973                 /*
1974                  * i_size must be checked after we know the page is Uptodate.
1975                  *
1976                  * Checking i_size after the check allows us to calculate
1977                  * the correct value for "nr", which means the zero-filled
1978                  * part of the page is not copied back to userspace (unless
1979                  * another truncate extends the file - this is desired though).
1980                  */
1981
1982                 isize = i_size_read(inode);
1983                 end_index = (isize - 1) >> PAGE_SHIFT;
1984                 if (unlikely(!isize || index > end_index)) {
1985                         put_page(page);
1986                         goto out;
1987                 }
1988
1989                 /* nr is the maximum number of bytes to copy from this page */
1990                 nr = PAGE_SIZE;
1991                 if (index == end_index) {
1992                         nr = ((isize - 1) & ~PAGE_MASK) + 1;
1993                         if (nr <= offset) {
1994                                 put_page(page);
1995                                 goto out;
1996                         }
1997                 }
1998                 nr = nr - offset;
1999
2000                 /* If users can be writing to this page using arbitrary
2001                  * virtual addresses, take care about potential aliasing
2002                  * before reading the page on the kernel side.
2003                  */
2004                 if (mapping_writably_mapped(mapping))
2005                         flush_dcache_page(page);
2006
2007                 /*
2008                  * When a sequential read accesses a page several times,
2009                  * only mark it as accessed the first time.
2010                  */
2011                 if (prev_index != index || offset != prev_offset)
2012                         mark_page_accessed(page);
2013                 prev_index = index;
2014
2015                 /*
2016                  * Ok, we have the page, and it's up-to-date, so
2017                  * now we can copy it to user space...
2018                  */
2019
2020                 ret = copy_page_to_iter(page, offset, nr, iter);
2021                 offset += ret;
2022                 index += offset >> PAGE_SHIFT;
2023                 offset &= ~PAGE_MASK;
2024                 prev_offset = offset;
2025
2026                 put_page(page);
2027                 written += ret;
2028                 if (!iov_iter_count(iter))
2029                         goto out;
2030                 if (ret < nr) {
2031                         error = -EFAULT;
2032                         goto out;
2033                 }
2034                 continue;
2035
2036 page_not_up_to_date:
2037                 /* Get exclusive access to the page ... */
2038                 error = lock_page_killable(page);
2039                 if (unlikely(error))
2040                         goto readpage_error;
2041
2042 page_not_up_to_date_locked:
2043                 /* Did it get truncated before we got the lock? */
2044                 if (!page->mapping) {
2045                         unlock_page(page);
2046                         put_page(page);
2047                         continue;
2048                 }
2049
2050                 /* Did somebody else fill it already? */
2051                 if (PageUptodate(page)) {
2052                         unlock_page(page);
2053                         goto page_ok;
2054                 }
2055
2056 readpage:
2057                 /*
2058                  * A previous I/O error may have been due to temporary
2059                  * failures, eg. multipath errors.
2060                  * PG_error will be set again if readpage fails.
2061                  */
2062                 ClearPageError(page);
2063                 /* Start the actual read. The read will unlock the page. */
2064                 error = mapping->a_ops->readpage(filp, page);
2065
2066                 if (unlikely(error)) {
2067                         if (error == AOP_TRUNCATED_PAGE) {
2068                                 put_page(page);
2069                                 error = 0;
2070                                 goto find_page;
2071                         }
2072                         goto readpage_error;
2073                 }
2074
2075                 if (!PageUptodate(page)) {
2076                         error = lock_page_killable(page);
2077                         if (unlikely(error))
2078                                 goto readpage_error;
2079                         if (!PageUptodate(page)) {
2080                                 if (page->mapping == NULL) {
2081                                         /*
2082                                          * invalidate_mapping_pages got it
2083                                          */
2084                                         unlock_page(page);
2085                                         put_page(page);
2086                                         goto find_page;
2087                                 }
2088                                 unlock_page(page);
2089                                 shrink_readahead_size_eio(filp, ra);
2090                                 error = -EIO;
2091                                 goto readpage_error;
2092                         }
2093                         unlock_page(page);
2094                 }
2095
2096                 goto page_ok;
2097
2098 readpage_error:
2099                 /* UHHUH! A synchronous read error occurred. Report it */
2100                 put_page(page);
2101                 goto out;
2102
2103 no_cached_page:
2104                 /*
2105                  * Ok, it wasn't cached, so we need to create a new
2106                  * page..
2107                  */
2108                 page = page_cache_alloc_cold(mapping);
2109                 if (!page) {
2110                         error = -ENOMEM;
2111                         goto out;
2112                 }
2113                 error = add_to_page_cache_lru(page, mapping, index,
2114                                 mapping_gfp_constraint(mapping, GFP_KERNEL));
2115                 if (error) {
2116                         put_page(page);
2117                         if (error == -EEXIST) {
2118                                 error = 0;
2119                                 goto find_page;
2120                         }
2121                         goto out;
2122                 }
2123                 goto readpage;
2124         }
2125
2126 out:
2127         ra->prev_pos = prev_index;
2128         ra->prev_pos <<= PAGE_SHIFT;
2129         ra->prev_pos |= prev_offset;
2130
2131         *ppos = ((loff_t)index << PAGE_SHIFT) + offset;
2132         file_accessed(filp);
2133         return written ? written : error;
2134 }
2135
2136 /**
2137  * generic_file_read_iter - generic filesystem read routine
2138  * @iocb:       kernel I/O control block
2139  * @iter:       destination for the data read
2140  *
2141  * This is the "read_iter()" routine for all filesystems
2142  * that can use the page cache directly.
2143  */
2144 ssize_t
2145 generic_file_read_iter(struct kiocb *iocb, struct iov_iter *iter)
2146 {
2147         struct file *file = iocb->ki_filp;
2148         ssize_t retval = 0;
2149         size_t count = iov_iter_count(iter);
2150
2151         if (!count)
2152                 goto out; /* skip atime */
2153
2154         if (iocb->ki_flags & IOCB_DIRECT) {
2155                 struct address_space *mapping = file->f_mapping;
2156                 struct inode *inode = mapping->host;
2157                 loff_t size;
2158
2159                 size = i_size_read(inode);
2160                 if (iocb->ki_flags & IOCB_NOWAIT) {
2161                         if (filemap_range_has_page(mapping, iocb->ki_pos,
2162                                                    iocb->ki_pos + count - 1))
2163                                 return -EAGAIN;
2164                 } else {
2165                         retval = filemap_write_and_wait_range(mapping,
2166                                                 iocb->ki_pos,
2167                                                 iocb->ki_pos + count - 1);
2168                         if (retval < 0)
2169                                 goto out;
2170                 }
2171
2172                 file_accessed(file);
2173
2174                 retval = mapping->a_ops->direct_IO(iocb, iter);
2175                 if (retval >= 0) {
2176                         iocb->ki_pos += retval;
2177                         count -= retval;
2178                 }
2179                 iov_iter_revert(iter, count - iov_iter_count(iter));
2180
2181                 /*
2182                  * Btrfs can have a short DIO read if we encounter
2183                  * compressed extents, so if there was an error, or if
2184                  * we've already read everything we wanted to, or if
2185                  * there was a short read because we hit EOF, go ahead
2186                  * and return.  Otherwise fallthrough to buffered io for
2187                  * the rest of the read.  Buffered reads will not work for
2188                  * DAX files, so don't bother trying.
2189                  */
2190                 if (retval < 0 || !count || iocb->ki_pos >= size ||
2191                     IS_DAX(inode))
2192                         goto out;
2193         }
2194
2195         retval = do_generic_file_read(file, &iocb->ki_pos, iter, retval);
2196 out:
2197         return retval;
2198 }
2199 EXPORT_SYMBOL(generic_file_read_iter);
2200
2201 #ifdef CONFIG_MMU
2202 /**
2203  * page_cache_read - adds requested page to the page cache if not already there
2204  * @file:       file to read
2205  * @offset:     page index
2206  * @gfp_mask:   memory allocation flags
2207  *
2208  * This adds the requested page to the page cache if it isn't already there,
2209  * and schedules an I/O to read in its contents from disk.
2210  */
2211 static int page_cache_read(struct file *file, pgoff_t offset, gfp_t gfp_mask)
2212 {
2213         struct address_space *mapping = file->f_mapping;
2214         struct page *page;
2215         int ret;
2216
2217         do {
2218                 page = __page_cache_alloc(gfp_mask|__GFP_COLD);
2219                 if (!page)
2220                         return -ENOMEM;
2221
2222                 ret = add_to_page_cache_lru(page, mapping, offset, gfp_mask & GFP_KERNEL);
2223                 if (ret == 0)
2224                         ret = mapping->a_ops->readpage(file, page);
2225                 else if (ret == -EEXIST)
2226                         ret = 0; /* losing race to add is OK */
2227
2228                 put_page(page);
2229
2230         } while (ret == AOP_TRUNCATED_PAGE);
2231
2232         return ret;
2233 }
2234
2235 #define MMAP_LOTSAMISS  (100)
2236
2237 /*
2238  * Synchronous readahead happens when we don't even find
2239  * a page in the page cache at all.
2240  */
2241 static void do_sync_mmap_readahead(struct vm_area_struct *vma,
2242                                    struct file_ra_state *ra,
2243                                    struct file *file,
2244                                    pgoff_t offset)
2245 {
2246         struct address_space *mapping = file->f_mapping;
2247
2248         /* If we don't want any read-ahead, don't bother */
2249         if (vma->vm_flags & VM_RAND_READ)
2250                 return;
2251         if (!ra->ra_pages)
2252                 return;
2253
2254         if (vma->vm_flags & VM_SEQ_READ) {
2255                 page_cache_sync_readahead(mapping, ra, file, offset,
2256                                           ra->ra_pages);
2257                 return;
2258         }
2259
2260         /* Avoid banging the cache line if not needed */
2261         if (ra->mmap_miss < MMAP_LOTSAMISS * 10)
2262                 ra->mmap_miss++;
2263
2264         /*
2265          * Do we miss much more than hit in this file? If so,
2266          * stop bothering with read-ahead. It will only hurt.
2267          */
2268         if (ra->mmap_miss > MMAP_LOTSAMISS)
2269                 return;
2270
2271         /*
2272          * mmap read-around
2273          */
2274         ra->start = max_t(long, 0, offset - ra->ra_pages / 2);
2275         ra->size = ra->ra_pages;
2276         ra->async_size = ra->ra_pages / 4;
2277         ra_submit(ra, mapping, file);
2278 }
2279
2280 /*
2281  * Asynchronous readahead happens when we find the page and PG_readahead,
2282  * so we want to possibly extend the readahead further..
2283  */
2284 static void do_async_mmap_readahead(struct vm_area_struct *vma,
2285                                     struct file_ra_state *ra,
2286                                     struct file *file,
2287                                     struct page *page,
2288                                     pgoff_t offset)
2289 {
2290         struct address_space *mapping = file->f_mapping;
2291
2292         /* If we don't want any read-ahead, don't bother */
2293         if (vma->vm_flags & VM_RAND_READ)
2294                 return;
2295         if (ra->mmap_miss > 0)
2296                 ra->mmap_miss--;
2297         if (PageReadahead(page))
2298                 page_cache_async_readahead(mapping, ra, file,
2299                                            page, offset, ra->ra_pages);
2300 }
2301
2302 /**
2303  * filemap_fault - read in file data for page fault handling
2304  * @vmf:        struct vm_fault containing details of the fault
2305  *
2306  * filemap_fault() is invoked via the vma operations vector for a
2307  * mapped memory region to read in file data during a page fault.
2308  *
2309  * The goto's are kind of ugly, but this streamlines the normal case of having
2310  * it in the page cache, and handles the special cases reasonably without
2311  * having a lot of duplicated code.
2312  *
2313  * vma->vm_mm->mmap_sem must be held on entry.
2314  *
2315  * If our return value has VM_FAULT_RETRY set, it's because
2316  * lock_page_or_retry() returned 0.
2317  * The mmap_sem has usually been released in this case.
2318  * See __lock_page_or_retry() for the exception.
2319  *
2320  * If our return value does not have VM_FAULT_RETRY set, the mmap_sem
2321  * has not been released.
2322  *
2323  * We never return with VM_FAULT_RETRY and a bit from VM_FAULT_ERROR set.
2324  */
2325 int filemap_fault(struct vm_fault *vmf)
2326 {
2327         int error;
2328         struct file *file = vmf->vma->vm_file;
2329         struct address_space *mapping = file->f_mapping;
2330         struct file_ra_state *ra = &file->f_ra;
2331         struct inode *inode = mapping->host;
2332         pgoff_t offset = vmf->pgoff;
2333         pgoff_t max_off;
2334         struct page *page;
2335         int ret = 0;
2336
2337         max_off = DIV_ROUND_UP(i_size_read(inode), PAGE_SIZE);
2338         if (unlikely(offset >= max_off))
2339                 return VM_FAULT_SIGBUS;
2340
2341         /*
2342          * Do we have something in the page cache already?
2343          */
2344         page = find_get_page(mapping, offset);
2345         if (likely(page) && !(vmf->flags & FAULT_FLAG_TRIED)) {
2346                 /*
2347                  * We found the page, so try async readahead before
2348                  * waiting for the lock.
2349                  */
2350                 do_async_mmap_readahead(vmf->vma, ra, file, page, offset);
2351         } else if (!page) {
2352                 /* No page in the page cache at all */
2353                 do_sync_mmap_readahead(vmf->vma, ra, file, offset);
2354                 count_vm_event(PGMAJFAULT);
2355                 count_memcg_event_mm(vmf->vma->vm_mm, PGMAJFAULT);
2356                 ret = VM_FAULT_MAJOR;
2357 retry_find:
2358                 page = find_get_page(mapping, offset);
2359                 if (!page)
2360                         goto no_cached_page;
2361         }
2362
2363         if (!lock_page_or_retry(page, vmf->vma->vm_mm, vmf->flags)) {
2364                 put_page(page);
2365                 return ret | VM_FAULT_RETRY;
2366         }
2367
2368         /* Did it get truncated? */
2369         if (unlikely(page->mapping != mapping)) {
2370                 unlock_page(page);
2371                 put_page(page);
2372                 goto retry_find;
2373         }
2374         VM_BUG_ON_PAGE(page->index != offset, page);
2375
2376         /*
2377          * We have a locked page in the page cache, now we need to check
2378          * that it's up-to-date. If not, it is going to be due to an error.
2379          */
2380         if (unlikely(!PageUptodate(page)))
2381                 goto page_not_uptodate;
2382
2383         /*
2384          * Found the page and have a reference on it.
2385          * We must recheck i_size under page lock.
2386          */
2387         max_off = DIV_ROUND_UP(i_size_read(inode), PAGE_SIZE);
2388         if (unlikely(offset >= max_off)) {
2389                 unlock_page(page);
2390                 put_page(page);
2391                 return VM_FAULT_SIGBUS;
2392         }
2393
2394         vmf->page = page;
2395         return ret | VM_FAULT_LOCKED;
2396
2397 no_cached_page:
2398         /*
2399          * We're only likely to ever get here if MADV_RANDOM is in
2400          * effect.
2401          */
2402         error = page_cache_read(file, offset, vmf->gfp_mask);
2403
2404         /*
2405          * The page we want has now been added to the page cache.
2406          * In the unlikely event that someone removed it in the
2407          * meantime, we'll just come back here and read it again.
2408          */
2409         if (error >= 0)
2410                 goto retry_find;
2411
2412         /*
2413          * An error return from page_cache_read can result if the
2414          * system is low on memory, or a problem occurs while trying
2415          * to schedule I/O.
2416          */
2417         if (error == -ENOMEM)
2418                 return VM_FAULT_OOM;
2419         return VM_FAULT_SIGBUS;
2420
2421 page_not_uptodate:
2422         /*
2423          * Umm, take care of errors if the page isn't up-to-date.
2424          * Try to re-read it _once_. We do this synchronously,
2425          * because there really aren't any performance issues here
2426          * and we need to check for errors.
2427          */
2428         ClearPageError(page);
2429         error = mapping->a_ops->readpage(file, page);
2430         if (!error) {
2431                 wait_on_page_locked(page);
2432                 if (!PageUptodate(page))
2433                         error = -EIO;
2434         }
2435         put_page(page);
2436
2437         if (!error || error == AOP_TRUNCATED_PAGE)
2438                 goto retry_find;
2439
2440         /* Things didn't work out. Return zero to tell the mm layer so. */
2441         shrink_readahead_size_eio(file, ra);
2442         return VM_FAULT_SIGBUS;
2443 }
2444 EXPORT_SYMBOL(filemap_fault);
2445
2446 void filemap_map_pages(struct vm_fault *vmf,
2447                 pgoff_t start_pgoff, pgoff_t end_pgoff)
2448 {
2449         struct radix_tree_iter iter;
2450         void **slot;
2451         struct file *file = vmf->vma->vm_file;
2452         struct address_space *mapping = file->f_mapping;
2453         pgoff_t last_pgoff = start_pgoff;
2454         unsigned long max_idx;
2455         struct page *head, *page;
2456
2457         rcu_read_lock();
2458         radix_tree_for_each_slot(slot, &mapping->page_tree, &iter,
2459                         start_pgoff) {
2460                 if (iter.index > end_pgoff)
2461                         break;
2462 repeat:
2463                 page = radix_tree_deref_slot(slot);
2464                 if (unlikely(!page))
2465                         goto next;
2466                 if (radix_tree_exception(page)) {
2467                         if (radix_tree_deref_retry(page)) {
2468                                 slot = radix_tree_iter_retry(&iter);
2469                                 continue;
2470                         }
2471                         goto next;
2472                 }
2473
2474                 head = compound_head(page);
2475                 if (!page_cache_get_speculative(head))
2476                         goto repeat;
2477
2478                 /* The page was split under us? */
2479                 if (compound_head(page) != head) {
2480                         put_page(head);
2481                         goto repeat;
2482                 }
2483
2484                 /* Has the page moved? */
2485                 if (unlikely(page != *slot)) {
2486                         put_page(head);
2487                         goto repeat;
2488                 }
2489
2490                 if (!PageUptodate(page) ||
2491                                 PageReadahead(page) ||
2492                                 PageHWPoison(page))
2493                         goto skip;
2494                 if (!trylock_page(page))
2495                         goto skip;
2496
2497                 if (page->mapping != mapping || !PageUptodate(page))
2498                         goto unlock;
2499
2500                 max_idx = DIV_ROUND_UP(i_size_read(mapping->host), PAGE_SIZE);
2501                 if (page->index >= max_idx)
2502                         goto unlock;
2503
2504                 if (file->f_ra.mmap_miss > 0)
2505                         file->f_ra.mmap_miss--;
2506
2507                 vmf->address += (iter.index - last_pgoff) << PAGE_SHIFT;
2508                 if (vmf->pte)
2509                         vmf->pte += iter.index - last_pgoff;
2510                 last_pgoff = iter.index;
2511                 if (alloc_set_pte(vmf, NULL, page))
2512                         goto unlock;
2513                 unlock_page(page);
2514                 goto next;
2515 unlock:
2516                 unlock_page(page);
2517 skip:
2518                 put_page(page);
2519 next:
2520                 /* Huge page is mapped? No need to proceed. */
2521                 if (pmd_trans_huge(*vmf->pmd))
2522                         break;
2523                 if (iter.index == end_pgoff)
2524                         break;
2525         }
2526         rcu_read_unlock();
2527 }
2528 EXPORT_SYMBOL(filemap_map_pages);
2529
2530 int filemap_page_mkwrite(struct vm_fault *vmf)
2531 {
2532         struct page *page = vmf->page;
2533         struct inode *inode = file_inode(vmf->vma->vm_file);
2534         int ret = VM_FAULT_LOCKED;
2535
2536         sb_start_pagefault(inode->i_sb);
2537         file_update_time(vmf->vma->vm_file);
2538         lock_page(page);
2539         if (page->mapping != inode->i_mapping) {
2540                 unlock_page(page);
2541                 ret = VM_FAULT_NOPAGE;
2542                 goto out;
2543         }
2544         /*
2545          * We mark the page dirty already here so that when freeze is in
2546          * progress, we are guaranteed that writeback during freezing will
2547          * see the dirty page and writeprotect it again.
2548          */
2549         set_page_dirty(page);
2550         wait_for_stable_page(page);
2551 out:
2552         sb_end_pagefault(inode->i_sb);
2553         return ret;
2554 }
2555 EXPORT_SYMBOL(filemap_page_mkwrite);
2556
2557 const struct vm_operations_struct generic_file_vm_ops = {
2558         .fault          = filemap_fault,
2559         .map_pages      = filemap_map_pages,
2560         .page_mkwrite   = filemap_page_mkwrite,
2561 };
2562
2563 /* This is used for a general mmap of a disk file */
2564
2565 int generic_file_mmap(struct file * file, struct vm_area_struct * vma)
2566 {
2567         struct address_space *mapping = file->f_mapping;
2568
2569         if (!mapping->a_ops->readpage)
2570                 return -ENOEXEC;
2571         file_accessed(file);
2572         vma->vm_ops = &generic_file_vm_ops;
2573         return 0;
2574 }
2575
2576 /*
2577  * This is for filesystems which do not implement ->writepage.
2578  */
2579 int generic_file_readonly_mmap(struct file *file, struct vm_area_struct *vma)
2580 {
2581         if ((vma->vm_flags & VM_SHARED) && (vma->vm_flags & VM_MAYWRITE))
2582                 return -EINVAL;
2583         return generic_file_mmap(file, vma);
2584 }
2585 #else
2586 int generic_file_mmap(struct file * file, struct vm_area_struct * vma)
2587 {
2588         return -ENOSYS;
2589 }
2590 int generic_file_readonly_mmap(struct file * file, struct vm_area_struct * vma)
2591 {
2592         return -ENOSYS;
2593 }
2594 #endif /* CONFIG_MMU */
2595
2596 EXPORT_SYMBOL(generic_file_mmap);
2597 EXPORT_SYMBOL(generic_file_readonly_mmap);
2598
2599 static struct page *wait_on_page_read(struct page *page)
2600 {
2601         if (!IS_ERR(page)) {
2602                 wait_on_page_locked(page);
2603                 if (!PageUptodate(page)) {
2604                         put_page(page);
2605                         page = ERR_PTR(-EIO);
2606                 }
2607         }
2608         return page;
2609 }
2610
2611 static struct page *do_read_cache_page(struct address_space *mapping,
2612                                 pgoff_t index,
2613                                 int (*filler)(void *, struct page *),
2614                                 void *data,
2615                                 gfp_t gfp)
2616 {
2617         struct page *page;
2618         int err;
2619 repeat:
2620         page = find_get_page(mapping, index);
2621         if (!page) {
2622                 page = __page_cache_alloc(gfp | __GFP_COLD);
2623                 if (!page)
2624                         return ERR_PTR(-ENOMEM);
2625                 err = add_to_page_cache_lru(page, mapping, index, gfp);
2626                 if (unlikely(err)) {
2627                         put_page(page);
2628                         if (err == -EEXIST)
2629                                 goto repeat;
2630                         /* Presumably ENOMEM for radix tree node */
2631                         return ERR_PTR(err);
2632                 }
2633
2634 filler:
2635                 err = filler(data, page);
2636                 if (err < 0) {
2637                         put_page(page);
2638                         return ERR_PTR(err);
2639                 }
2640
2641                 page = wait_on_page_read(page);
2642                 if (IS_ERR(page))
2643                         return page;
2644                 goto out;
2645         }
2646         if (PageUptodate(page))
2647                 goto out;
2648
2649         /*
2650          * Page is not up to date and may be locked due one of the following
2651          * case a: Page is being filled and the page lock is held
2652          * case b: Read/write error clearing the page uptodate status
2653          * case c: Truncation in progress (page locked)
2654          * case d: Reclaim in progress
2655          *
2656          * Case a, the page will be up to date when the page is unlocked.
2657          *    There is no need to serialise on the page lock here as the page
2658          *    is pinned so the lock gives no additional protection. Even if the
2659          *    the page is truncated, the data is still valid if PageUptodate as
2660          *    it's a race vs truncate race.
2661          * Case b, the page will not be up to date
2662          * Case c, the page may be truncated but in itself, the data may still
2663          *    be valid after IO completes as it's a read vs truncate race. The
2664          *    operation must restart if the page is not uptodate on unlock but
2665          *    otherwise serialising on page lock to stabilise the mapping gives
2666          *    no additional guarantees to the caller as the page lock is
2667          *    released before return.
2668          * Case d, similar to truncation. If reclaim holds the page lock, it
2669          *    will be a race with remove_mapping that determines if the mapping
2670          *    is valid on unlock but otherwise the data is valid and there is
2671          *    no need to serialise with page lock.
2672          *
2673          * As the page lock gives no additional guarantee, we optimistically
2674          * wait on the page to be unlocked and check if it's up to date and
2675          * use the page if it is. Otherwise, the page lock is required to
2676          * distinguish between the different cases. The motivation is that we
2677          * avoid spurious serialisations and wakeups when multiple processes
2678          * wait on the same page for IO to complete.
2679          */
2680         wait_on_page_locked(page);
2681         if (PageUptodate(page))
2682                 goto out;
2683
2684         /* Distinguish between all the cases under the safety of the lock */
2685         lock_page(page);
2686
2687         /* Case c or d, restart the operation */
2688         if (!page->mapping) {
2689                 unlock_page(page);
2690                 put_page(page);
2691                 goto repeat;
2692         }
2693
2694         /* Someone else locked and filled the page in a very small window */
2695         if (PageUptodate(page)) {
2696                 unlock_page(page);
2697                 goto out;
2698         }
2699         goto filler;
2700
2701 out:
2702         mark_page_accessed(page);
2703         return page;
2704 }
2705
2706 /**
2707  * read_cache_page - read into page cache, fill it if needed
2708  * @mapping:    the page's address_space
2709  * @index:      the page index
2710  * @filler:     function to perform the read
2711  * @data:       first arg to filler(data, page) function, often left as NULL
2712  *
2713  * Read into the page cache. If a page already exists, and PageUptodate() is
2714  * not set, try to fill the page and wait for it to become unlocked.
2715  *
2716  * If the page does not get brought uptodate, return -EIO.
2717  */
2718 struct page *read_cache_page(struct address_space *mapping,
2719                                 pgoff_t index,
2720                                 int (*filler)(void *, struct page *),
2721                                 void *data)
2722 {
2723         return do_read_cache_page(mapping, index, filler, data, mapping_gfp_mask(mapping));
2724 }
2725 EXPORT_SYMBOL(read_cache_page);
2726
2727 /**
2728  * read_cache_page_gfp - read into page cache, using specified page allocation flags.
2729  * @mapping:    the page's address_space
2730  * @index:      the page index
2731  * @gfp:        the page allocator flags to use if allocating
2732  *
2733  * This is the same as "read_mapping_page(mapping, index, NULL)", but with
2734  * any new page allocations done using the specified allocation flags.
2735  *
2736  * If the page does not get brought uptodate, return -EIO.
2737  */
2738 struct page *read_cache_page_gfp(struct address_space *mapping,
2739                                 pgoff_t index,
2740                                 gfp_t gfp)
2741 {
2742         filler_t *filler = (filler_t *)mapping->a_ops->readpage;
2743
2744         return do_read_cache_page(mapping, index, filler, NULL, gfp);
2745 }
2746 EXPORT_SYMBOL(read_cache_page_gfp);
2747
2748 /*
2749  * Performs necessary checks before doing a write
2750  *
2751  * Can adjust writing position or amount of bytes to write.
2752  * Returns appropriate error code that caller should return or
2753  * zero in case that write should be allowed.
2754  */
2755 inline ssize_t generic_write_checks(struct kiocb *iocb, struct iov_iter *from)
2756 {
2757         struct file *file = iocb->ki_filp;
2758         struct inode *inode = file->f_mapping->host;
2759         unsigned long limit = rlimit(RLIMIT_FSIZE);
2760         loff_t pos;
2761
2762         if (!iov_iter_count(from))
2763                 return 0;
2764
2765         /* FIXME: this is for backwards compatibility with 2.4 */
2766         if (iocb->ki_flags & IOCB_APPEND)
2767                 iocb->ki_pos = i_size_read(inode);
2768
2769         pos = iocb->ki_pos;
2770
2771         if ((iocb->ki_flags & IOCB_NOWAIT) && !(iocb->ki_flags & IOCB_DIRECT))
2772                 return -EINVAL;
2773
2774         if (limit != RLIM_INFINITY) {
2775                 if (iocb->ki_pos >= limit) {
2776                         send_sig(SIGXFSZ, current, 0);
2777                         return -EFBIG;
2778                 }
2779                 iov_iter_truncate(from, limit - (unsigned long)pos);
2780         }
2781
2782         /*
2783          * LFS rule
2784          */
2785         if (unlikely(pos + iov_iter_count(from) > MAX_NON_LFS &&
2786                                 !(file->f_flags & O_LARGEFILE))) {
2787                 if (pos >= MAX_NON_LFS)
2788                         return -EFBIG;
2789                 iov_iter_truncate(from, MAX_NON_LFS - (unsigned long)pos);
2790         }
2791
2792         /*
2793          * Are we about to exceed the fs block limit ?
2794          *
2795          * If we have written data it becomes a short write.  If we have
2796          * exceeded without writing data we send a signal and return EFBIG.
2797          * Linus frestrict idea will clean these up nicely..
2798          */
2799         if (unlikely(pos >= inode->i_sb->s_maxbytes))
2800                 return -EFBIG;
2801
2802         iov_iter_truncate(from, inode->i_sb->s_maxbytes - pos);
2803         return iov_iter_count(from);
2804 }
2805 EXPORT_SYMBOL(generic_write_checks);
2806
2807 int pagecache_write_begin(struct file *file, struct address_space *mapping,
2808                                 loff_t pos, unsigned len, unsigned flags,
2809                                 struct page **pagep, void **fsdata)
2810 {
2811         const struct address_space_operations *aops = mapping->a_ops;
2812
2813         return aops->write_begin(file, mapping, pos, len, flags,
2814                                                         pagep, fsdata);
2815 }
2816 EXPORT_SYMBOL(pagecache_write_begin);
2817
2818 int pagecache_write_end(struct file *file, struct address_space *mapping,
2819                                 loff_t pos, unsigned len, unsigned copied,
2820                                 struct page *page, void *fsdata)
2821 {
2822         const struct address_space_operations *aops = mapping->a_ops;
2823
2824         return aops->write_end(file, mapping, pos, len, copied, page, fsdata);
2825 }
2826 EXPORT_SYMBOL(pagecache_write_end);
2827
2828 ssize_t
2829 generic_file_direct_write(struct kiocb *iocb, struct iov_iter *from)
2830 {
2831         struct file     *file = iocb->ki_filp;
2832         struct address_space *mapping = file->f_mapping;
2833         struct inode    *inode = mapping->host;
2834         loff_t          pos = iocb->ki_pos;
2835         ssize_t         written;
2836         size_t          write_len;
2837         pgoff_t         end;
2838
2839         write_len = iov_iter_count(from);
2840         end = (pos + write_len - 1) >> PAGE_SHIFT;
2841
2842         if (iocb->ki_flags & IOCB_NOWAIT) {
2843                 /* If there are pages to writeback, return */
2844                 if (filemap_range_has_page(inode->i_mapping, pos,
2845                                            pos + iov_iter_count(from)))
2846                         return -EAGAIN;
2847         } else {
2848                 written = filemap_write_and_wait_range(mapping, pos,
2849                                                         pos + write_len - 1);
2850                 if (written)
2851                         goto out;
2852         }
2853
2854         /*
2855          * After a write we want buffered reads to be sure to go to disk to get
2856          * the new data.  We invalidate clean cached page from the region we're
2857          * about to write.  We do this *before* the write so that we can return
2858          * without clobbering -EIOCBQUEUED from ->direct_IO().
2859          */
2860         written = invalidate_inode_pages2_range(mapping,
2861                                         pos >> PAGE_SHIFT, end);
2862         /*
2863          * If a page can not be invalidated, return 0 to fall back
2864          * to buffered write.
2865          */
2866         if (written) {
2867                 if (written == -EBUSY)
2868                         return 0;
2869                 goto out;
2870         }
2871
2872         written = mapping->a_ops->direct_IO(iocb, from);
2873
2874         /*
2875          * Finally, try again to invalidate clean pages which might have been
2876          * cached by non-direct readahead, or faulted in by get_user_pages()
2877          * if the source of the write was an mmap'ed region of the file
2878          * we're writing.  Either one is a pretty crazy thing to do,
2879          * so we don't support it 100%.  If this invalidation
2880          * fails, tough, the write still worked...
2881          */
2882         invalidate_inode_pages2_range(mapping,
2883                                 pos >> PAGE_SHIFT, end);
2884
2885         if (written > 0) {
2886                 pos += written;
2887                 write_len -= written;
2888                 if (pos > i_size_read(inode) && !S_ISBLK(inode->i_mode)) {
2889                         i_size_write(inode, pos);
2890                         mark_inode_dirty(inode);
2891                 }
2892                 iocb->ki_pos = pos;
2893         }
2894         iov_iter_revert(from, write_len - iov_iter_count(from));
2895 out:
2896         return written;
2897 }
2898 EXPORT_SYMBOL(generic_file_direct_write);
2899
2900 /*
2901  * Find or create a page at the given pagecache position. Return the locked
2902  * page. This function is specifically for buffered writes.
2903  */
2904 struct page *grab_cache_page_write_begin(struct address_space *mapping,
2905                                         pgoff_t index, unsigned flags)
2906 {
2907         struct page *page;
2908         int fgp_flags = FGP_LOCK|FGP_WRITE|FGP_CREAT;
2909
2910         if (flags & AOP_FLAG_NOFS)
2911                 fgp_flags |= FGP_NOFS;
2912
2913         page = pagecache_get_page(mapping, index, fgp_flags,
2914                         mapping_gfp_mask(mapping));
2915         if (page)
2916                 wait_for_stable_page(page);
2917
2918         return page;
2919 }
2920 EXPORT_SYMBOL(grab_cache_page_write_begin);
2921
2922 ssize_t generic_perform_write(struct file *file,
2923                                 struct iov_iter *i, loff_t pos)
2924 {
2925         struct address_space *mapping = file->f_mapping;
2926         const struct address_space_operations *a_ops = mapping->a_ops;
2927         long status = 0;
2928         ssize_t written = 0;
2929         unsigned int flags = 0;
2930
2931         do {
2932                 struct page *page;
2933                 unsigned long offset;   /* Offset into pagecache page */
2934                 unsigned long bytes;    /* Bytes to write to page */
2935                 size_t copied;          /* Bytes copied from user */
2936                 void *fsdata;
2937
2938                 offset = (pos & (PAGE_SIZE - 1));
2939                 bytes = min_t(unsigned long, PAGE_SIZE - offset,
2940                                                 iov_iter_count(i));
2941
2942 again:
2943                 /*
2944                  * Bring in the user page that we will copy from _first_.
2945                  * Otherwise there's a nasty deadlock on copying from the
2946                  * same page as we're writing to, without it being marked
2947                  * up-to-date.
2948                  *
2949                  * Not only is this an optimisation, but it is also required
2950                  * to check that the address is actually valid, when atomic
2951                  * usercopies are used, below.
2952                  */
2953                 if (unlikely(iov_iter_fault_in_readable(i, bytes))) {
2954                         status = -EFAULT;
2955                         break;
2956                 }
2957
2958                 if (fatal_signal_pending(current)) {
2959                         status = -EINTR;
2960                         break;
2961                 }
2962
2963                 status = a_ops->write_begin(file, mapping, pos, bytes, flags,
2964                                                 &page, &fsdata);
2965                 if (unlikely(status < 0))
2966                         break;
2967
2968                 if (mapping_writably_mapped(mapping))
2969                         flush_dcache_page(page);
2970
2971                 copied = iov_iter_copy_from_user_atomic(page, i, offset, bytes);
2972                 flush_dcache_page(page);
2973
2974                 status = a_ops->write_end(file, mapping, pos, bytes, copied,
2975                                                 page, fsdata);
2976                 if (unlikely(status < 0))
2977                         break;
2978                 copied = status;
2979
2980                 cond_resched();
2981
2982                 iov_iter_advance(i, copied);
2983                 if (unlikely(copied == 0)) {
2984                         /*
2985                          * If we were unable to copy any data at all, we must
2986                          * fall back to a single segment length write.
2987                          *
2988                          * If we didn't fallback here, we could livelock
2989                          * because not all segments in the iov can be copied at
2990                          * once without a pagefault.
2991                          */
2992                         bytes = min_t(unsigned long, PAGE_SIZE - offset,
2993                                                 iov_iter_single_seg_count(i));
2994                         goto again;
2995                 }
2996                 pos += copied;
2997                 written += copied;
2998
2999                 balance_dirty_pages_ratelimited(mapping);
3000         } while (iov_iter_count(i));
3001
3002         return written ? written : status;
3003 }
3004 EXPORT_SYMBOL(generic_perform_write);
3005
3006 /**
3007  * __generic_file_write_iter - write data to a file
3008  * @iocb:       IO state structure (file, offset, etc.)
3009  * @from:       iov_iter with data to write
3010  *
3011  * This function does all the work needed for actually writing data to a
3012  * file. It does all basic checks, removes SUID from the file, updates
3013  * modification times and calls proper subroutines depending on whether we
3014  * do direct IO or a standard buffered write.
3015  *
3016  * It expects i_mutex to be grabbed unless we work on a block device or similar
3017  * object which does not need locking at all.
3018  *
3019  * This function does *not* take care of syncing data in case of O_SYNC write.
3020  * A caller has to handle it. This is mainly due to the fact that we want to
3021  * avoid syncing under i_mutex.
3022  */
3023 ssize_t __generic_file_write_iter(struct kiocb *iocb, struct iov_iter *from)
3024 {
3025         struct file *file = iocb->ki_filp;
3026         struct address_space * mapping = file->f_mapping;
3027         struct inode    *inode = mapping->host;
3028         ssize_t         written = 0;
3029         ssize_t         err;
3030         ssize_t         status;
3031
3032         /* We can write back this queue in page reclaim */
3033         current->backing_dev_info = inode_to_bdi(inode);
3034         err = file_remove_privs(file);
3035         if (err)
3036                 goto out;
3037
3038         err = file_update_time(file);
3039         if (err)
3040                 goto out;
3041
3042         if (iocb->ki_flags & IOCB_DIRECT) {
3043                 loff_t pos, endbyte;
3044
3045                 written = generic_file_direct_write(iocb, from);
3046                 /*
3047                  * If the write stopped short of completing, fall back to
3048                  * buffered writes.  Some filesystems do this for writes to
3049                  * holes, for example.  For DAX files, a buffered write will
3050                  * not succeed (even if it did, DAX does not handle dirty
3051                  * page-cache pages correctly).
3052                  */
3053                 if (written < 0 || !iov_iter_count(from) || IS_DAX(inode))
3054                         goto out;
3055
3056                 status = generic_perform_write(file, from, pos = iocb->ki_pos);
3057                 /*
3058                  * If generic_perform_write() returned a synchronous error
3059                  * then we want to return the number of bytes which were
3060                  * direct-written, or the error code if that was zero.  Note
3061                  * that this differs from normal direct-io semantics, which
3062                  * will return -EFOO even if some bytes were written.
3063                  */
3064                 if (unlikely(status < 0)) {
3065                         err = status;
3066                         goto out;
3067                 }
3068                 /*
3069                  * We need to ensure that the page cache pages are written to
3070                  * disk and invalidated to preserve the expected O_DIRECT
3071                  * semantics.
3072                  */
3073                 endbyte = pos + status - 1;
3074                 err = filemap_write_and_wait_range(mapping, pos, endbyte);
3075                 if (err == 0) {
3076                         iocb->ki_pos = endbyte + 1;
3077                         written += status;
3078                         invalidate_mapping_pages(mapping,
3079                                                  pos >> PAGE_SHIFT,
3080                                                  endbyte >> PAGE_SHIFT);
3081                 } else {
3082                         /*
3083                          * We don't know how much we wrote, so just return
3084                          * the number of bytes which were direct-written
3085                          */
3086                 }
3087         } else {
3088                 written = generic_perform_write(file, from, iocb->ki_pos);
3089                 if (likely(written > 0))
3090                         iocb->ki_pos += written;
3091         }
3092 out:
3093         current->backing_dev_info = NULL;
3094         return written ? written : err;
3095 }
3096 EXPORT_SYMBOL(__generic_file_write_iter);
3097
3098 /**
3099  * generic_file_write_iter - write data to a file
3100  * @iocb:       IO state structure
3101  * @from:       iov_iter with data to write
3102  *
3103  * This is a wrapper around __generic_file_write_iter() to be used by most
3104  * filesystems. It takes care of syncing the file in case of O_SYNC file
3105  * and acquires i_mutex as needed.
3106  */
3107 ssize_t generic_file_write_iter(struct kiocb *iocb, struct iov_iter *from)
3108 {
3109         struct file *file = iocb->ki_filp;
3110         struct inode *inode = file->f_mapping->host;
3111         ssize_t ret;
3112
3113         inode_lock(inode);
3114         ret = generic_write_checks(iocb, from);
3115         if (ret > 0)
3116                 ret = __generic_file_write_iter(iocb, from);
3117         inode_unlock(inode);
3118
3119         if (ret > 0)
3120                 ret = generic_write_sync(iocb, ret);
3121         return ret;
3122 }
3123 EXPORT_SYMBOL(generic_file_write_iter);
3124
3125 /**
3126  * try_to_release_page() - release old fs-specific metadata on a page
3127  *
3128  * @page: the page which the kernel is trying to free
3129  * @gfp_mask: memory allocation flags (and I/O mode)
3130  *
3131  * The address_space is to try to release any data against the page
3132  * (presumably at page->private).  If the release was successful, return '1'.
3133  * Otherwise return zero.
3134  *
3135  * This may also be called if PG_fscache is set on a page, indicating that the
3136  * page is known to the local caching routines.
3137  *
3138  * The @gfp_mask argument specifies whether I/O may be performed to release
3139  * this page (__GFP_IO), and whether the call may block (__GFP_RECLAIM & __GFP_FS).
3140  *
3141  */
3142 int try_to_release_page(struct page *page, gfp_t gfp_mask)
3143 {
3144         struct address_space * const mapping = page->mapping;
3145
3146         BUG_ON(!PageLocked(page));
3147         if (PageWriteback(page))
3148                 return 0;
3149
3150         if (mapping && mapping->a_ops->releasepage)
3151                 return mapping->a_ops->releasepage(page, gfp_mask);
3152         return try_to_free_buffers(page);
3153 }
3154
3155 EXPORT_SYMBOL(try_to_release_page);