mm: implement find_get_pages_range()
[sfrench/cifs-2.6.git] / mm / filemap.c
1 /*
2  *      linux/mm/filemap.c
3  *
4  * Copyright (C) 1994-1999  Linus Torvalds
5  */
6
7 /*
8  * This file handles the generic file mmap semantics used by
9  * most "normal" filesystems (but you don't /have/ to use this:
10  * the NFS filesystem used to do this differently, for example)
11  */
12 #include <linux/export.h>
13 #include <linux/compiler.h>
14 #include <linux/dax.h>
15 #include <linux/fs.h>
16 #include <linux/sched/signal.h>
17 #include <linux/uaccess.h>
18 #include <linux/capability.h>
19 #include <linux/kernel_stat.h>
20 #include <linux/gfp.h>
21 #include <linux/mm.h>
22 #include <linux/swap.h>
23 #include <linux/mman.h>
24 #include <linux/pagemap.h>
25 #include <linux/file.h>
26 #include <linux/uio.h>
27 #include <linux/hash.h>
28 #include <linux/writeback.h>
29 #include <linux/backing-dev.h>
30 #include <linux/pagevec.h>
31 #include <linux/blkdev.h>
32 #include <linux/security.h>
33 #include <linux/cpuset.h>
34 #include <linux/hardirq.h> /* for BUG_ON(!in_atomic()) only */
35 #include <linux/hugetlb.h>
36 #include <linux/memcontrol.h>
37 #include <linux/cleancache.h>
38 #include <linux/rmap.h>
39 #include "internal.h"
40
41 #define CREATE_TRACE_POINTS
42 #include <trace/events/filemap.h>
43
44 /*
45  * FIXME: remove all knowledge of the buffer layer from the core VM
46  */
47 #include <linux/buffer_head.h> /* for try_to_free_buffers */
48
49 #include <asm/mman.h>
50
51 /*
52  * Shared mappings implemented 30.11.1994. It's not fully working yet,
53  * though.
54  *
55  * Shared mappings now work. 15.8.1995  Bruno.
56  *
57  * finished 'unifying' the page and buffer cache and SMP-threaded the
58  * page-cache, 21.05.1999, Ingo Molnar <mingo@redhat.com>
59  *
60  * SMP-threaded pagemap-LRU 1999, Andrea Arcangeli <andrea@suse.de>
61  */
62
63 /*
64  * Lock ordering:
65  *
66  *  ->i_mmap_rwsem              (truncate_pagecache)
67  *    ->private_lock            (__free_pte->__set_page_dirty_buffers)
68  *      ->swap_lock             (exclusive_swap_page, others)
69  *        ->mapping->tree_lock
70  *
71  *  ->i_mutex
72  *    ->i_mmap_rwsem            (truncate->unmap_mapping_range)
73  *
74  *  ->mmap_sem
75  *    ->i_mmap_rwsem
76  *      ->page_table_lock or pte_lock   (various, mainly in memory.c)
77  *        ->mapping->tree_lock  (arch-dependent flush_dcache_mmap_lock)
78  *
79  *  ->mmap_sem
80  *    ->lock_page               (access_process_vm)
81  *
82  *  ->i_mutex                   (generic_perform_write)
83  *    ->mmap_sem                (fault_in_pages_readable->do_page_fault)
84  *
85  *  bdi->wb.list_lock
86  *    sb_lock                   (fs/fs-writeback.c)
87  *    ->mapping->tree_lock      (__sync_single_inode)
88  *
89  *  ->i_mmap_rwsem
90  *    ->anon_vma.lock           (vma_adjust)
91  *
92  *  ->anon_vma.lock
93  *    ->page_table_lock or pte_lock     (anon_vma_prepare and various)
94  *
95  *  ->page_table_lock or pte_lock
96  *    ->swap_lock               (try_to_unmap_one)
97  *    ->private_lock            (try_to_unmap_one)
98  *    ->tree_lock               (try_to_unmap_one)
99  *    ->zone_lru_lock(zone)     (follow_page->mark_page_accessed)
100  *    ->zone_lru_lock(zone)     (check_pte_range->isolate_lru_page)
101  *    ->private_lock            (page_remove_rmap->set_page_dirty)
102  *    ->tree_lock               (page_remove_rmap->set_page_dirty)
103  *    bdi.wb->list_lock         (page_remove_rmap->set_page_dirty)
104  *    ->inode->i_lock           (page_remove_rmap->set_page_dirty)
105  *    ->memcg->move_lock        (page_remove_rmap->lock_page_memcg)
106  *    bdi.wb->list_lock         (zap_pte_range->set_page_dirty)
107  *    ->inode->i_lock           (zap_pte_range->set_page_dirty)
108  *    ->private_lock            (zap_pte_range->__set_page_dirty_buffers)
109  *
110  * ->i_mmap_rwsem
111  *   ->tasklist_lock            (memory_failure, collect_procs_ao)
112  */
113
114 static int page_cache_tree_insert(struct address_space *mapping,
115                                   struct page *page, void **shadowp)
116 {
117         struct radix_tree_node *node;
118         void **slot;
119         int error;
120
121         error = __radix_tree_create(&mapping->page_tree, page->index, 0,
122                                     &node, &slot);
123         if (error)
124                 return error;
125         if (*slot) {
126                 void *p;
127
128                 p = radix_tree_deref_slot_protected(slot, &mapping->tree_lock);
129                 if (!radix_tree_exceptional_entry(p))
130                         return -EEXIST;
131
132                 mapping->nrexceptional--;
133                 if (shadowp)
134                         *shadowp = p;
135         }
136         __radix_tree_replace(&mapping->page_tree, node, slot, page,
137                              workingset_update_node, mapping);
138         mapping->nrpages++;
139         return 0;
140 }
141
142 static void page_cache_tree_delete(struct address_space *mapping,
143                                    struct page *page, void *shadow)
144 {
145         int i, nr;
146
147         /* hugetlb pages are represented by one entry in the radix tree */
148         nr = PageHuge(page) ? 1 : hpage_nr_pages(page);
149
150         VM_BUG_ON_PAGE(!PageLocked(page), page);
151         VM_BUG_ON_PAGE(PageTail(page), page);
152         VM_BUG_ON_PAGE(nr != 1 && shadow, page);
153
154         for (i = 0; i < nr; i++) {
155                 struct radix_tree_node *node;
156                 void **slot;
157
158                 __radix_tree_lookup(&mapping->page_tree, page->index + i,
159                                     &node, &slot);
160
161                 VM_BUG_ON_PAGE(!node && nr != 1, page);
162
163                 radix_tree_clear_tags(&mapping->page_tree, node, slot);
164                 __radix_tree_replace(&mapping->page_tree, node, slot, shadow,
165                                      workingset_update_node, mapping);
166         }
167
168         if (shadow) {
169                 mapping->nrexceptional += nr;
170                 /*
171                  * Make sure the nrexceptional update is committed before
172                  * the nrpages update so that final truncate racing
173                  * with reclaim does not see both counters 0 at the
174                  * same time and miss a shadow entry.
175                  */
176                 smp_wmb();
177         }
178         mapping->nrpages -= nr;
179 }
180
181 /*
182  * Delete a page from the page cache and free it. Caller has to make
183  * sure the page is locked and that nobody else uses it - or that usage
184  * is safe.  The caller must hold the mapping's tree_lock.
185  */
186 void __delete_from_page_cache(struct page *page, void *shadow)
187 {
188         struct address_space *mapping = page->mapping;
189         int nr = hpage_nr_pages(page);
190
191         trace_mm_filemap_delete_from_page_cache(page);
192         /*
193          * if we're uptodate, flush out into the cleancache, otherwise
194          * invalidate any existing cleancache entries.  We can't leave
195          * stale data around in the cleancache once our page is gone
196          */
197         if (PageUptodate(page) && PageMappedToDisk(page))
198                 cleancache_put_page(page);
199         else
200                 cleancache_invalidate_page(mapping, page);
201
202         VM_BUG_ON_PAGE(PageTail(page), page);
203         VM_BUG_ON_PAGE(page_mapped(page), page);
204         if (!IS_ENABLED(CONFIG_DEBUG_VM) && unlikely(page_mapped(page))) {
205                 int mapcount;
206
207                 pr_alert("BUG: Bad page cache in process %s  pfn:%05lx\n",
208                          current->comm, page_to_pfn(page));
209                 dump_page(page, "still mapped when deleted");
210                 dump_stack();
211                 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
212
213                 mapcount = page_mapcount(page);
214                 if (mapping_exiting(mapping) &&
215                     page_count(page) >= mapcount + 2) {
216                         /*
217                          * All vmas have already been torn down, so it's
218                          * a good bet that actually the page is unmapped,
219                          * and we'd prefer not to leak it: if we're wrong,
220                          * some other bad page check should catch it later.
221                          */
222                         page_mapcount_reset(page);
223                         page_ref_sub(page, mapcount);
224                 }
225         }
226
227         page_cache_tree_delete(mapping, page, shadow);
228
229         page->mapping = NULL;
230         /* Leave page->index set: truncation lookup relies upon it */
231
232         /* hugetlb pages do not participate in page cache accounting. */
233         if (PageHuge(page))
234                 return;
235
236         __mod_node_page_state(page_pgdat(page), NR_FILE_PAGES, -nr);
237         if (PageSwapBacked(page)) {
238                 __mod_node_page_state(page_pgdat(page), NR_SHMEM, -nr);
239                 if (PageTransHuge(page))
240                         __dec_node_page_state(page, NR_SHMEM_THPS);
241         } else {
242                 VM_BUG_ON_PAGE(PageTransHuge(page), page);
243         }
244
245         /*
246          * At this point page must be either written or cleaned by truncate.
247          * Dirty page here signals a bug and loss of unwritten data.
248          *
249          * This fixes dirty accounting after removing the page entirely but
250          * leaves PageDirty set: it has no effect for truncated page and
251          * anyway will be cleared before returning page into buddy allocator.
252          */
253         if (WARN_ON_ONCE(PageDirty(page)))
254                 account_page_cleaned(page, mapping, inode_to_wb(mapping->host));
255 }
256
257 /**
258  * delete_from_page_cache - delete page from page cache
259  * @page: the page which the kernel is trying to remove from page cache
260  *
261  * This must be called only on pages that have been verified to be in the page
262  * cache and locked.  It will never put the page into the free list, the caller
263  * has a reference on the page.
264  */
265 void delete_from_page_cache(struct page *page)
266 {
267         struct address_space *mapping = page_mapping(page);
268         unsigned long flags;
269         void (*freepage)(struct page *);
270
271         BUG_ON(!PageLocked(page));
272
273         freepage = mapping->a_ops->freepage;
274
275         spin_lock_irqsave(&mapping->tree_lock, flags);
276         __delete_from_page_cache(page, NULL);
277         spin_unlock_irqrestore(&mapping->tree_lock, flags);
278
279         if (freepage)
280                 freepage(page);
281
282         if (PageTransHuge(page) && !PageHuge(page)) {
283                 page_ref_sub(page, HPAGE_PMD_NR);
284                 VM_BUG_ON_PAGE(page_count(page) <= 0, page);
285         } else {
286                 put_page(page);
287         }
288 }
289 EXPORT_SYMBOL(delete_from_page_cache);
290
291 int filemap_check_errors(struct address_space *mapping)
292 {
293         int ret = 0;
294         /* Check for outstanding write errors */
295         if (test_bit(AS_ENOSPC, &mapping->flags) &&
296             test_and_clear_bit(AS_ENOSPC, &mapping->flags))
297                 ret = -ENOSPC;
298         if (test_bit(AS_EIO, &mapping->flags) &&
299             test_and_clear_bit(AS_EIO, &mapping->flags))
300                 ret = -EIO;
301         return ret;
302 }
303 EXPORT_SYMBOL(filemap_check_errors);
304
305 static int filemap_check_and_keep_errors(struct address_space *mapping)
306 {
307         /* Check for outstanding write errors */
308         if (test_bit(AS_EIO, &mapping->flags))
309                 return -EIO;
310         if (test_bit(AS_ENOSPC, &mapping->flags))
311                 return -ENOSPC;
312         return 0;
313 }
314
315 /**
316  * __filemap_fdatawrite_range - start writeback on mapping dirty pages in range
317  * @mapping:    address space structure to write
318  * @start:      offset in bytes where the range starts
319  * @end:        offset in bytes where the range ends (inclusive)
320  * @sync_mode:  enable synchronous operation
321  *
322  * Start writeback against all of a mapping's dirty pages that lie
323  * within the byte offsets <start, end> inclusive.
324  *
325  * If sync_mode is WB_SYNC_ALL then this is a "data integrity" operation, as
326  * opposed to a regular memory cleansing writeback.  The difference between
327  * these two operations is that if a dirty page/buffer is encountered, it must
328  * be waited upon, and not just skipped over.
329  */
330 int __filemap_fdatawrite_range(struct address_space *mapping, loff_t start,
331                                 loff_t end, int sync_mode)
332 {
333         int ret;
334         struct writeback_control wbc = {
335                 .sync_mode = sync_mode,
336                 .nr_to_write = LONG_MAX,
337                 .range_start = start,
338                 .range_end = end,
339         };
340
341         if (!mapping_cap_writeback_dirty(mapping))
342                 return 0;
343
344         wbc_attach_fdatawrite_inode(&wbc, mapping->host);
345         ret = do_writepages(mapping, &wbc);
346         wbc_detach_inode(&wbc);
347         return ret;
348 }
349
350 static inline int __filemap_fdatawrite(struct address_space *mapping,
351         int sync_mode)
352 {
353         return __filemap_fdatawrite_range(mapping, 0, LLONG_MAX, sync_mode);
354 }
355
356 int filemap_fdatawrite(struct address_space *mapping)
357 {
358         return __filemap_fdatawrite(mapping, WB_SYNC_ALL);
359 }
360 EXPORT_SYMBOL(filemap_fdatawrite);
361
362 int filemap_fdatawrite_range(struct address_space *mapping, loff_t start,
363                                 loff_t end)
364 {
365         return __filemap_fdatawrite_range(mapping, start, end, WB_SYNC_ALL);
366 }
367 EXPORT_SYMBOL(filemap_fdatawrite_range);
368
369 /**
370  * filemap_flush - mostly a non-blocking flush
371  * @mapping:    target address_space
372  *
373  * This is a mostly non-blocking flush.  Not suitable for data-integrity
374  * purposes - I/O may not be started against all dirty pages.
375  */
376 int filemap_flush(struct address_space *mapping)
377 {
378         return __filemap_fdatawrite(mapping, WB_SYNC_NONE);
379 }
380 EXPORT_SYMBOL(filemap_flush);
381
382 /**
383  * filemap_range_has_page - check if a page exists in range.
384  * @mapping:           address space within which to check
385  * @start_byte:        offset in bytes where the range starts
386  * @end_byte:          offset in bytes where the range ends (inclusive)
387  *
388  * Find at least one page in the range supplied, usually used to check if
389  * direct writing in this range will trigger a writeback.
390  */
391 bool filemap_range_has_page(struct address_space *mapping,
392                            loff_t start_byte, loff_t end_byte)
393 {
394         pgoff_t index = start_byte >> PAGE_SHIFT;
395         pgoff_t end = end_byte >> PAGE_SHIFT;
396         struct 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_range - gang pagecache lookup
1561  * @mapping:    The address_space to search
1562  * @start:      The starting page index
1563  * @end:        The final page index (inclusive)
1564  * @nr_pages:   The maximum number of pages
1565  * @pages:      Where the resulting pages are placed
1566  *
1567  * find_get_pages_range() will search for and return a group of up to @nr_pages
1568  * pages in the mapping starting at index @start and up to index @end
1569  * (inclusive).  The pages are placed at @pages.  find_get_pages_range() takes
1570  * a reference against the returned pages.
1571  *
1572  * The search returns a group of mapping-contiguous pages with ascending
1573  * indexes.  There may be holes in the indices due to not-present pages.
1574  * We also update @start to index the next page for the traversal.
1575  *
1576  * find_get_pages_range() returns the number of pages which were found. If this
1577  * number is smaller than @nr_pages, the end of specified range has been
1578  * reached.
1579  */
1580 unsigned find_get_pages_range(struct address_space *mapping, pgoff_t *start,
1581                               pgoff_t end, unsigned int nr_pages,
1582                               struct page **pages)
1583 {
1584         struct radix_tree_iter iter;
1585         void **slot;
1586         unsigned ret = 0;
1587
1588         if (unlikely(!nr_pages))
1589                 return 0;
1590
1591         rcu_read_lock();
1592         radix_tree_for_each_slot(slot, &mapping->page_tree, &iter, *start) {
1593                 struct page *head, *page;
1594
1595                 if (iter.index > end)
1596                         break;
1597 repeat:
1598                 page = radix_tree_deref_slot(slot);
1599                 if (unlikely(!page))
1600                         continue;
1601
1602                 if (radix_tree_exception(page)) {
1603                         if (radix_tree_deref_retry(page)) {
1604                                 slot = radix_tree_iter_retry(&iter);
1605                                 continue;
1606                         }
1607                         /*
1608                          * A shadow entry of a recently evicted page,
1609                          * or a swap entry from shmem/tmpfs.  Skip
1610                          * over it.
1611                          */
1612                         continue;
1613                 }
1614
1615                 head = compound_head(page);
1616                 if (!page_cache_get_speculative(head))
1617                         goto repeat;
1618
1619                 /* The page was split under us? */
1620                 if (compound_head(page) != head) {
1621                         put_page(head);
1622                         goto repeat;
1623                 }
1624
1625                 /* Has the page moved? */
1626                 if (unlikely(page != *slot)) {
1627                         put_page(head);
1628                         goto repeat;
1629                 }
1630
1631                 pages[ret] = page;
1632                 if (++ret == nr_pages) {
1633                         *start = pages[ret - 1]->index + 1;
1634                         goto out;
1635                 }
1636         }
1637
1638         /*
1639          * We come here when there is no page beyond @end. We take care to not
1640          * overflow the index @start as it confuses some of the callers. This
1641          * breaks the iteration when there is page at index -1 but that is
1642          * already broken anyway.
1643          */
1644         if (end == (pgoff_t)-1)
1645                 *start = (pgoff_t)-1;
1646         else
1647                 *start = end + 1;
1648 out:
1649         rcu_read_unlock();
1650
1651         return ret;
1652 }
1653
1654 /**
1655  * find_get_pages_contig - gang contiguous pagecache lookup
1656  * @mapping:    The address_space to search
1657  * @index:      The starting page index
1658  * @nr_pages:   The maximum number of pages
1659  * @pages:      Where the resulting pages are placed
1660  *
1661  * find_get_pages_contig() works exactly like find_get_pages(), except
1662  * that the returned number of pages are guaranteed to be contiguous.
1663  *
1664  * find_get_pages_contig() returns the number of pages which were found.
1665  */
1666 unsigned find_get_pages_contig(struct address_space *mapping, pgoff_t index,
1667                                unsigned int nr_pages, struct page **pages)
1668 {
1669         struct radix_tree_iter iter;
1670         void **slot;
1671         unsigned int ret = 0;
1672
1673         if (unlikely(!nr_pages))
1674                 return 0;
1675
1676         rcu_read_lock();
1677         radix_tree_for_each_contig(slot, &mapping->page_tree, &iter, index) {
1678                 struct page *head, *page;
1679 repeat:
1680                 page = radix_tree_deref_slot(slot);
1681                 /* The hole, there no reason to continue */
1682                 if (unlikely(!page))
1683                         break;
1684
1685                 if (radix_tree_exception(page)) {
1686                         if (radix_tree_deref_retry(page)) {
1687                                 slot = radix_tree_iter_retry(&iter);
1688                                 continue;
1689                         }
1690                         /*
1691                          * A shadow entry of a recently evicted page,
1692                          * or a swap entry from shmem/tmpfs.  Stop
1693                          * looking for contiguous pages.
1694                          */
1695                         break;
1696                 }
1697
1698                 head = compound_head(page);
1699                 if (!page_cache_get_speculative(head))
1700                         goto repeat;
1701
1702                 /* The page was split under us? */
1703                 if (compound_head(page) != head) {
1704                         put_page(head);
1705                         goto repeat;
1706                 }
1707
1708                 /* Has the page moved? */
1709                 if (unlikely(page != *slot)) {
1710                         put_page(head);
1711                         goto repeat;
1712                 }
1713
1714                 /*
1715                  * must check mapping and index after taking the ref.
1716                  * otherwise we can get both false positives and false
1717                  * negatives, which is just confusing to the caller.
1718                  */
1719                 if (page->mapping == NULL || page_to_pgoff(page) != iter.index) {
1720                         put_page(page);
1721                         break;
1722                 }
1723
1724                 pages[ret] = page;
1725                 if (++ret == nr_pages)
1726                         break;
1727         }
1728         rcu_read_unlock();
1729         return ret;
1730 }
1731 EXPORT_SYMBOL(find_get_pages_contig);
1732
1733 /**
1734  * find_get_pages_tag - find and return pages that match @tag
1735  * @mapping:    the address_space to search
1736  * @index:      the starting page index
1737  * @tag:        the tag index
1738  * @nr_pages:   the maximum number of pages
1739  * @pages:      where the resulting pages are placed
1740  *
1741  * Like find_get_pages, except we only return pages which are tagged with
1742  * @tag.   We update @index to index the next page for the traversal.
1743  */
1744 unsigned find_get_pages_tag(struct address_space *mapping, pgoff_t *index,
1745                         int tag, unsigned int nr_pages, struct page **pages)
1746 {
1747         struct radix_tree_iter iter;
1748         void **slot;
1749         unsigned ret = 0;
1750
1751         if (unlikely(!nr_pages))
1752                 return 0;
1753
1754         rcu_read_lock();
1755         radix_tree_for_each_tagged(slot, &mapping->page_tree,
1756                                    &iter, *index, tag) {
1757                 struct page *head, *page;
1758 repeat:
1759                 page = radix_tree_deref_slot(slot);
1760                 if (unlikely(!page))
1761                         continue;
1762
1763                 if (radix_tree_exception(page)) {
1764                         if (radix_tree_deref_retry(page)) {
1765                                 slot = radix_tree_iter_retry(&iter);
1766                                 continue;
1767                         }
1768                         /*
1769                          * A shadow entry of a recently evicted page.
1770                          *
1771                          * Those entries should never be tagged, but
1772                          * this tree walk is lockless and the tags are
1773                          * looked up in bulk, one radix tree node at a
1774                          * time, so there is a sizable window for page
1775                          * reclaim to evict a page we saw tagged.
1776                          *
1777                          * Skip over it.
1778                          */
1779                         continue;
1780                 }
1781
1782                 head = compound_head(page);
1783                 if (!page_cache_get_speculative(head))
1784                         goto repeat;
1785
1786                 /* The page was split under us? */
1787                 if (compound_head(page) != head) {
1788                         put_page(head);
1789                         goto repeat;
1790                 }
1791
1792                 /* Has the page moved? */
1793                 if (unlikely(page != *slot)) {
1794                         put_page(head);
1795                         goto repeat;
1796                 }
1797
1798                 pages[ret] = page;
1799                 if (++ret == nr_pages)
1800                         break;
1801         }
1802
1803         rcu_read_unlock();
1804
1805         if (ret)
1806                 *index = pages[ret - 1]->index + 1;
1807
1808         return ret;
1809 }
1810 EXPORT_SYMBOL(find_get_pages_tag);
1811
1812 /**
1813  * find_get_entries_tag - find and return entries that match @tag
1814  * @mapping:    the address_space to search
1815  * @start:      the starting page cache index
1816  * @tag:        the tag index
1817  * @nr_entries: the maximum number of entries
1818  * @entries:    where the resulting entries are placed
1819  * @indices:    the cache indices corresponding to the entries in @entries
1820  *
1821  * Like find_get_entries, except we only return entries which are tagged with
1822  * @tag.
1823  */
1824 unsigned find_get_entries_tag(struct address_space *mapping, pgoff_t start,
1825                         int tag, unsigned int nr_entries,
1826                         struct page **entries, pgoff_t *indices)
1827 {
1828         void **slot;
1829         unsigned int ret = 0;
1830         struct radix_tree_iter iter;
1831
1832         if (!nr_entries)
1833                 return 0;
1834
1835         rcu_read_lock();
1836         radix_tree_for_each_tagged(slot, &mapping->page_tree,
1837                                    &iter, start, tag) {
1838                 struct page *head, *page;
1839 repeat:
1840                 page = radix_tree_deref_slot(slot);
1841                 if (unlikely(!page))
1842                         continue;
1843                 if (radix_tree_exception(page)) {
1844                         if (radix_tree_deref_retry(page)) {
1845                                 slot = radix_tree_iter_retry(&iter);
1846                                 continue;
1847                         }
1848
1849                         /*
1850                          * A shadow entry of a recently evicted page, a swap
1851                          * entry from shmem/tmpfs or a DAX entry.  Return it
1852                          * without attempting to raise page count.
1853                          */
1854                         goto export;
1855                 }
1856
1857                 head = compound_head(page);
1858                 if (!page_cache_get_speculative(head))
1859                         goto repeat;
1860
1861                 /* The page was split under us? */
1862                 if (compound_head(page) != head) {
1863                         put_page(head);
1864                         goto repeat;
1865                 }
1866
1867                 /* Has the page moved? */
1868                 if (unlikely(page != *slot)) {
1869                         put_page(head);
1870                         goto repeat;
1871                 }
1872 export:
1873                 indices[ret] = iter.index;
1874                 entries[ret] = page;
1875                 if (++ret == nr_entries)
1876                         break;
1877         }
1878         rcu_read_unlock();
1879         return ret;
1880 }
1881 EXPORT_SYMBOL(find_get_entries_tag);
1882
1883 /*
1884  * CD/DVDs are error prone. When a medium error occurs, the driver may fail
1885  * a _large_ part of the i/o request. Imagine the worst scenario:
1886  *
1887  *      ---R__________________________________________B__________
1888  *         ^ reading here                             ^ bad block(assume 4k)
1889  *
1890  * read(R) => miss => readahead(R...B) => media error => frustrating retries
1891  * => failing the whole request => read(R) => read(R+1) =>
1892  * readahead(R+1...B+1) => bang => read(R+2) => read(R+3) =>
1893  * readahead(R+3...B+2) => bang => read(R+3) => read(R+4) =>
1894  * readahead(R+4...B+3) => bang => read(R+4) => read(R+5) => ......
1895  *
1896  * It is going insane. Fix it by quickly scaling down the readahead size.
1897  */
1898 static void shrink_readahead_size_eio(struct file *filp,
1899                                         struct file_ra_state *ra)
1900 {
1901         ra->ra_pages /= 4;
1902 }
1903
1904 /**
1905  * do_generic_file_read - generic file read routine
1906  * @filp:       the file to read
1907  * @ppos:       current file position
1908  * @iter:       data destination
1909  * @written:    already copied
1910  *
1911  * This is a generic file read routine, and uses the
1912  * mapping->a_ops->readpage() function for the actual low-level stuff.
1913  *
1914  * This is really ugly. But the goto's actually try to clarify some
1915  * of the logic when it comes to error handling etc.
1916  */
1917 static ssize_t do_generic_file_read(struct file *filp, loff_t *ppos,
1918                 struct iov_iter *iter, ssize_t written)
1919 {
1920         struct address_space *mapping = filp->f_mapping;
1921         struct inode *inode = mapping->host;
1922         struct file_ra_state *ra = &filp->f_ra;
1923         pgoff_t index;
1924         pgoff_t last_index;
1925         pgoff_t prev_index;
1926         unsigned long offset;      /* offset into pagecache page */
1927         unsigned int prev_offset;
1928         int error = 0;
1929
1930         if (unlikely(*ppos >= inode->i_sb->s_maxbytes))
1931                 return 0;
1932         iov_iter_truncate(iter, inode->i_sb->s_maxbytes);
1933
1934         index = *ppos >> PAGE_SHIFT;
1935         prev_index = ra->prev_pos >> PAGE_SHIFT;
1936         prev_offset = ra->prev_pos & (PAGE_SIZE-1);
1937         last_index = (*ppos + iter->count + PAGE_SIZE-1) >> PAGE_SHIFT;
1938         offset = *ppos & ~PAGE_MASK;
1939
1940         for (;;) {
1941                 struct page *page;
1942                 pgoff_t end_index;
1943                 loff_t isize;
1944                 unsigned long nr, ret;
1945
1946                 cond_resched();
1947 find_page:
1948                 if (fatal_signal_pending(current)) {
1949                         error = -EINTR;
1950                         goto out;
1951                 }
1952
1953                 page = find_get_page(mapping, index);
1954                 if (!page) {
1955                         page_cache_sync_readahead(mapping,
1956                                         ra, filp,
1957                                         index, last_index - index);
1958                         page = find_get_page(mapping, index);
1959                         if (unlikely(page == NULL))
1960                                 goto no_cached_page;
1961                 }
1962                 if (PageReadahead(page)) {
1963                         page_cache_async_readahead(mapping,
1964                                         ra, filp, page,
1965                                         index, last_index - index);
1966                 }
1967                 if (!PageUptodate(page)) {
1968                         /*
1969                          * See comment in do_read_cache_page on why
1970                          * wait_on_page_locked is used to avoid unnecessarily
1971                          * serialisations and why it's safe.
1972                          */
1973                         error = wait_on_page_locked_killable(page);
1974                         if (unlikely(error))
1975                                 goto readpage_error;
1976                         if (PageUptodate(page))
1977                                 goto page_ok;
1978
1979                         if (inode->i_blkbits == PAGE_SHIFT ||
1980                                         !mapping->a_ops->is_partially_uptodate)
1981                                 goto page_not_up_to_date;
1982                         /* pipes can't handle partially uptodate pages */
1983                         if (unlikely(iter->type & ITER_PIPE))
1984                                 goto page_not_up_to_date;
1985                         if (!trylock_page(page))
1986                                 goto page_not_up_to_date;
1987                         /* Did it get truncated before we got the lock? */
1988                         if (!page->mapping)
1989                                 goto page_not_up_to_date_locked;
1990                         if (!mapping->a_ops->is_partially_uptodate(page,
1991                                                         offset, iter->count))
1992                                 goto page_not_up_to_date_locked;
1993                         unlock_page(page);
1994                 }
1995 page_ok:
1996                 /*
1997                  * i_size must be checked after we know the page is Uptodate.
1998                  *
1999                  * Checking i_size after the check allows us to calculate
2000                  * the correct value for "nr", which means the zero-filled
2001                  * part of the page is not copied back to userspace (unless
2002                  * another truncate extends the file - this is desired though).
2003                  */
2004
2005                 isize = i_size_read(inode);
2006                 end_index = (isize - 1) >> PAGE_SHIFT;
2007                 if (unlikely(!isize || index > end_index)) {
2008                         put_page(page);
2009                         goto out;
2010                 }
2011
2012                 /* nr is the maximum number of bytes to copy from this page */
2013                 nr = PAGE_SIZE;
2014                 if (index == end_index) {
2015                         nr = ((isize - 1) & ~PAGE_MASK) + 1;
2016                         if (nr <= offset) {
2017                                 put_page(page);
2018                                 goto out;
2019                         }
2020                 }
2021                 nr = nr - offset;
2022
2023                 /* If users can be writing to this page using arbitrary
2024                  * virtual addresses, take care about potential aliasing
2025                  * before reading the page on the kernel side.
2026                  */
2027                 if (mapping_writably_mapped(mapping))
2028                         flush_dcache_page(page);
2029
2030                 /*
2031                  * When a sequential read accesses a page several times,
2032                  * only mark it as accessed the first time.
2033                  */
2034                 if (prev_index != index || offset != prev_offset)
2035                         mark_page_accessed(page);
2036                 prev_index = index;
2037
2038                 /*
2039                  * Ok, we have the page, and it's up-to-date, so
2040                  * now we can copy it to user space...
2041                  */
2042
2043                 ret = copy_page_to_iter(page, offset, nr, iter);
2044                 offset += ret;
2045                 index += offset >> PAGE_SHIFT;
2046                 offset &= ~PAGE_MASK;
2047                 prev_offset = offset;
2048
2049                 put_page(page);
2050                 written += ret;
2051                 if (!iov_iter_count(iter))
2052                         goto out;
2053                 if (ret < nr) {
2054                         error = -EFAULT;
2055                         goto out;
2056                 }
2057                 continue;
2058
2059 page_not_up_to_date:
2060                 /* Get exclusive access to the page ... */
2061                 error = lock_page_killable(page);
2062                 if (unlikely(error))
2063                         goto readpage_error;
2064
2065 page_not_up_to_date_locked:
2066                 /* Did it get truncated before we got the lock? */
2067                 if (!page->mapping) {
2068                         unlock_page(page);
2069                         put_page(page);
2070                         continue;
2071                 }
2072
2073                 /* Did somebody else fill it already? */
2074                 if (PageUptodate(page)) {
2075                         unlock_page(page);
2076                         goto page_ok;
2077                 }
2078
2079 readpage:
2080                 /*
2081                  * A previous I/O error may have been due to temporary
2082                  * failures, eg. multipath errors.
2083                  * PG_error will be set again if readpage fails.
2084                  */
2085                 ClearPageError(page);
2086                 /* Start the actual read. The read will unlock the page. */
2087                 error = mapping->a_ops->readpage(filp, page);
2088
2089                 if (unlikely(error)) {
2090                         if (error == AOP_TRUNCATED_PAGE) {
2091                                 put_page(page);
2092                                 error = 0;
2093                                 goto find_page;
2094                         }
2095                         goto readpage_error;
2096                 }
2097
2098                 if (!PageUptodate(page)) {
2099                         error = lock_page_killable(page);
2100                         if (unlikely(error))
2101                                 goto readpage_error;
2102                         if (!PageUptodate(page)) {
2103                                 if (page->mapping == NULL) {
2104                                         /*
2105                                          * invalidate_mapping_pages got it
2106                                          */
2107                                         unlock_page(page);
2108                                         put_page(page);
2109                                         goto find_page;
2110                                 }
2111                                 unlock_page(page);
2112                                 shrink_readahead_size_eio(filp, ra);
2113                                 error = -EIO;
2114                                 goto readpage_error;
2115                         }
2116                         unlock_page(page);
2117                 }
2118
2119                 goto page_ok;
2120
2121 readpage_error:
2122                 /* UHHUH! A synchronous read error occurred. Report it */
2123                 put_page(page);
2124                 goto out;
2125
2126 no_cached_page:
2127                 /*
2128                  * Ok, it wasn't cached, so we need to create a new
2129                  * page..
2130                  */
2131                 page = page_cache_alloc_cold(mapping);
2132                 if (!page) {
2133                         error = -ENOMEM;
2134                         goto out;
2135                 }
2136                 error = add_to_page_cache_lru(page, mapping, index,
2137                                 mapping_gfp_constraint(mapping, GFP_KERNEL));
2138                 if (error) {
2139                         put_page(page);
2140                         if (error == -EEXIST) {
2141                                 error = 0;
2142                                 goto find_page;
2143                         }
2144                         goto out;
2145                 }
2146                 goto readpage;
2147         }
2148
2149 out:
2150         ra->prev_pos = prev_index;
2151         ra->prev_pos <<= PAGE_SHIFT;
2152         ra->prev_pos |= prev_offset;
2153
2154         *ppos = ((loff_t)index << PAGE_SHIFT) + offset;
2155         file_accessed(filp);
2156         return written ? written : error;
2157 }
2158
2159 /**
2160  * generic_file_read_iter - generic filesystem read routine
2161  * @iocb:       kernel I/O control block
2162  * @iter:       destination for the data read
2163  *
2164  * This is the "read_iter()" routine for all filesystems
2165  * that can use the page cache directly.
2166  */
2167 ssize_t
2168 generic_file_read_iter(struct kiocb *iocb, struct iov_iter *iter)
2169 {
2170         struct file *file = iocb->ki_filp;
2171         ssize_t retval = 0;
2172         size_t count = iov_iter_count(iter);
2173
2174         if (!count)
2175                 goto out; /* skip atime */
2176
2177         if (iocb->ki_flags & IOCB_DIRECT) {
2178                 struct address_space *mapping = file->f_mapping;
2179                 struct inode *inode = mapping->host;
2180                 loff_t size;
2181
2182                 size = i_size_read(inode);
2183                 if (iocb->ki_flags & IOCB_NOWAIT) {
2184                         if (filemap_range_has_page(mapping, iocb->ki_pos,
2185                                                    iocb->ki_pos + count - 1))
2186                                 return -EAGAIN;
2187                 } else {
2188                         retval = filemap_write_and_wait_range(mapping,
2189                                                 iocb->ki_pos,
2190                                                 iocb->ki_pos + count - 1);
2191                         if (retval < 0)
2192                                 goto out;
2193                 }
2194
2195                 file_accessed(file);
2196
2197                 retval = mapping->a_ops->direct_IO(iocb, iter);
2198                 if (retval >= 0) {
2199                         iocb->ki_pos += retval;
2200                         count -= retval;
2201                 }
2202                 iov_iter_revert(iter, count - iov_iter_count(iter));
2203
2204                 /*
2205                  * Btrfs can have a short DIO read if we encounter
2206                  * compressed extents, so if there was an error, or if
2207                  * we've already read everything we wanted to, or if
2208                  * there was a short read because we hit EOF, go ahead
2209                  * and return.  Otherwise fallthrough to buffered io for
2210                  * the rest of the read.  Buffered reads will not work for
2211                  * DAX files, so don't bother trying.
2212                  */
2213                 if (retval < 0 || !count || iocb->ki_pos >= size ||
2214                     IS_DAX(inode))
2215                         goto out;
2216         }
2217
2218         retval = do_generic_file_read(file, &iocb->ki_pos, iter, retval);
2219 out:
2220         return retval;
2221 }
2222 EXPORT_SYMBOL(generic_file_read_iter);
2223
2224 #ifdef CONFIG_MMU
2225 /**
2226  * page_cache_read - adds requested page to the page cache if not already there
2227  * @file:       file to read
2228  * @offset:     page index
2229  * @gfp_mask:   memory allocation flags
2230  *
2231  * This adds the requested page to the page cache if it isn't already there,
2232  * and schedules an I/O to read in its contents from disk.
2233  */
2234 static int page_cache_read(struct file *file, pgoff_t offset, gfp_t gfp_mask)
2235 {
2236         struct address_space *mapping = file->f_mapping;
2237         struct page *page;
2238         int ret;
2239
2240         do {
2241                 page = __page_cache_alloc(gfp_mask|__GFP_COLD);
2242                 if (!page)
2243                         return -ENOMEM;
2244
2245                 ret = add_to_page_cache_lru(page, mapping, offset, gfp_mask & GFP_KERNEL);
2246                 if (ret == 0)
2247                         ret = mapping->a_ops->readpage(file, page);
2248                 else if (ret == -EEXIST)
2249                         ret = 0; /* losing race to add is OK */
2250
2251                 put_page(page);
2252
2253         } while (ret == AOP_TRUNCATED_PAGE);
2254
2255         return ret;
2256 }
2257
2258 #define MMAP_LOTSAMISS  (100)
2259
2260 /*
2261  * Synchronous readahead happens when we don't even find
2262  * a page in the page cache at all.
2263  */
2264 static void do_sync_mmap_readahead(struct vm_area_struct *vma,
2265                                    struct file_ra_state *ra,
2266                                    struct file *file,
2267                                    pgoff_t offset)
2268 {
2269         struct address_space *mapping = file->f_mapping;
2270
2271         /* If we don't want any read-ahead, don't bother */
2272         if (vma->vm_flags & VM_RAND_READ)
2273                 return;
2274         if (!ra->ra_pages)
2275                 return;
2276
2277         if (vma->vm_flags & VM_SEQ_READ) {
2278                 page_cache_sync_readahead(mapping, ra, file, offset,
2279                                           ra->ra_pages);
2280                 return;
2281         }
2282
2283         /* Avoid banging the cache line if not needed */
2284         if (ra->mmap_miss < MMAP_LOTSAMISS * 10)
2285                 ra->mmap_miss++;
2286
2287         /*
2288          * Do we miss much more than hit in this file? If so,
2289          * stop bothering with read-ahead. It will only hurt.
2290          */
2291         if (ra->mmap_miss > MMAP_LOTSAMISS)
2292                 return;
2293
2294         /*
2295          * mmap read-around
2296          */
2297         ra->start = max_t(long, 0, offset - ra->ra_pages / 2);
2298         ra->size = ra->ra_pages;
2299         ra->async_size = ra->ra_pages / 4;
2300         ra_submit(ra, mapping, file);
2301 }
2302
2303 /*
2304  * Asynchronous readahead happens when we find the page and PG_readahead,
2305  * so we want to possibly extend the readahead further..
2306  */
2307 static void do_async_mmap_readahead(struct vm_area_struct *vma,
2308                                     struct file_ra_state *ra,
2309                                     struct file *file,
2310                                     struct page *page,
2311                                     pgoff_t offset)
2312 {
2313         struct address_space *mapping = file->f_mapping;
2314
2315         /* If we don't want any read-ahead, don't bother */
2316         if (vma->vm_flags & VM_RAND_READ)
2317                 return;
2318         if (ra->mmap_miss > 0)
2319                 ra->mmap_miss--;
2320         if (PageReadahead(page))
2321                 page_cache_async_readahead(mapping, ra, file,
2322                                            page, offset, ra->ra_pages);
2323 }
2324
2325 /**
2326  * filemap_fault - read in file data for page fault handling
2327  * @vmf:        struct vm_fault containing details of the fault
2328  *
2329  * filemap_fault() is invoked via the vma operations vector for a
2330  * mapped memory region to read in file data during a page fault.
2331  *
2332  * The goto's are kind of ugly, but this streamlines the normal case of having
2333  * it in the page cache, and handles the special cases reasonably without
2334  * having a lot of duplicated code.
2335  *
2336  * vma->vm_mm->mmap_sem must be held on entry.
2337  *
2338  * If our return value has VM_FAULT_RETRY set, it's because
2339  * lock_page_or_retry() returned 0.
2340  * The mmap_sem has usually been released in this case.
2341  * See __lock_page_or_retry() for the exception.
2342  *
2343  * If our return value does not have VM_FAULT_RETRY set, the mmap_sem
2344  * has not been released.
2345  *
2346  * We never return with VM_FAULT_RETRY and a bit from VM_FAULT_ERROR set.
2347  */
2348 int filemap_fault(struct vm_fault *vmf)
2349 {
2350         int error;
2351         struct file *file = vmf->vma->vm_file;
2352         struct address_space *mapping = file->f_mapping;
2353         struct file_ra_state *ra = &file->f_ra;
2354         struct inode *inode = mapping->host;
2355         pgoff_t offset = vmf->pgoff;
2356         pgoff_t max_off;
2357         struct page *page;
2358         int ret = 0;
2359
2360         max_off = DIV_ROUND_UP(i_size_read(inode), PAGE_SIZE);
2361         if (unlikely(offset >= max_off))
2362                 return VM_FAULT_SIGBUS;
2363
2364         /*
2365          * Do we have something in the page cache already?
2366          */
2367         page = find_get_page(mapping, offset);
2368         if (likely(page) && !(vmf->flags & FAULT_FLAG_TRIED)) {
2369                 /*
2370                  * We found the page, so try async readahead before
2371                  * waiting for the lock.
2372                  */
2373                 do_async_mmap_readahead(vmf->vma, ra, file, page, offset);
2374         } else if (!page) {
2375                 /* No page in the page cache at all */
2376                 do_sync_mmap_readahead(vmf->vma, ra, file, offset);
2377                 count_vm_event(PGMAJFAULT);
2378                 count_memcg_event_mm(vmf->vma->vm_mm, PGMAJFAULT);
2379                 ret = VM_FAULT_MAJOR;
2380 retry_find:
2381                 page = find_get_page(mapping, offset);
2382                 if (!page)
2383                         goto no_cached_page;
2384         }
2385
2386         if (!lock_page_or_retry(page, vmf->vma->vm_mm, vmf->flags)) {
2387                 put_page(page);
2388                 return ret | VM_FAULT_RETRY;
2389         }
2390
2391         /* Did it get truncated? */
2392         if (unlikely(page->mapping != mapping)) {
2393                 unlock_page(page);
2394                 put_page(page);
2395                 goto retry_find;
2396         }
2397         VM_BUG_ON_PAGE(page->index != offset, page);
2398
2399         /*
2400          * We have a locked page in the page cache, now we need to check
2401          * that it's up-to-date. If not, it is going to be due to an error.
2402          */
2403         if (unlikely(!PageUptodate(page)))
2404                 goto page_not_uptodate;
2405
2406         /*
2407          * Found the page and have a reference on it.
2408          * We must recheck i_size under page lock.
2409          */
2410         max_off = DIV_ROUND_UP(i_size_read(inode), PAGE_SIZE);
2411         if (unlikely(offset >= max_off)) {
2412                 unlock_page(page);
2413                 put_page(page);
2414                 return VM_FAULT_SIGBUS;
2415         }
2416
2417         vmf->page = page;
2418         return ret | VM_FAULT_LOCKED;
2419
2420 no_cached_page:
2421         /*
2422          * We're only likely to ever get here if MADV_RANDOM is in
2423          * effect.
2424          */
2425         error = page_cache_read(file, offset, vmf->gfp_mask);
2426
2427         /*
2428          * The page we want has now been added to the page cache.
2429          * In the unlikely event that someone removed it in the
2430          * meantime, we'll just come back here and read it again.
2431          */
2432         if (error >= 0)
2433                 goto retry_find;
2434
2435         /*
2436          * An error return from page_cache_read can result if the
2437          * system is low on memory, or a problem occurs while trying
2438          * to schedule I/O.
2439          */
2440         if (error == -ENOMEM)
2441                 return VM_FAULT_OOM;
2442         return VM_FAULT_SIGBUS;
2443
2444 page_not_uptodate:
2445         /*
2446          * Umm, take care of errors if the page isn't up-to-date.
2447          * Try to re-read it _once_. We do this synchronously,
2448          * because there really aren't any performance issues here
2449          * and we need to check for errors.
2450          */
2451         ClearPageError(page);
2452         error = mapping->a_ops->readpage(file, page);
2453         if (!error) {
2454                 wait_on_page_locked(page);
2455                 if (!PageUptodate(page))
2456                         error = -EIO;
2457         }
2458         put_page(page);
2459
2460         if (!error || error == AOP_TRUNCATED_PAGE)
2461                 goto retry_find;
2462
2463         /* Things didn't work out. Return zero to tell the mm layer so. */
2464         shrink_readahead_size_eio(file, ra);
2465         return VM_FAULT_SIGBUS;
2466 }
2467 EXPORT_SYMBOL(filemap_fault);
2468
2469 void filemap_map_pages(struct vm_fault *vmf,
2470                 pgoff_t start_pgoff, pgoff_t end_pgoff)
2471 {
2472         struct radix_tree_iter iter;
2473         void **slot;
2474         struct file *file = vmf->vma->vm_file;
2475         struct address_space *mapping = file->f_mapping;
2476         pgoff_t last_pgoff = start_pgoff;
2477         unsigned long max_idx;
2478         struct page *head, *page;
2479
2480         rcu_read_lock();
2481         radix_tree_for_each_slot(slot, &mapping->page_tree, &iter,
2482                         start_pgoff) {
2483                 if (iter.index > end_pgoff)
2484                         break;
2485 repeat:
2486                 page = radix_tree_deref_slot(slot);
2487                 if (unlikely(!page))
2488                         goto next;
2489                 if (radix_tree_exception(page)) {
2490                         if (radix_tree_deref_retry(page)) {
2491                                 slot = radix_tree_iter_retry(&iter);
2492                                 continue;
2493                         }
2494                         goto next;
2495                 }
2496
2497                 head = compound_head(page);
2498                 if (!page_cache_get_speculative(head))
2499                         goto repeat;
2500
2501                 /* The page was split under us? */
2502                 if (compound_head(page) != head) {
2503                         put_page(head);
2504                         goto repeat;
2505                 }
2506
2507                 /* Has the page moved? */
2508                 if (unlikely(page != *slot)) {
2509                         put_page(head);
2510                         goto repeat;
2511                 }
2512
2513                 if (!PageUptodate(page) ||
2514                                 PageReadahead(page) ||
2515                                 PageHWPoison(page))
2516                         goto skip;
2517                 if (!trylock_page(page))
2518                         goto skip;
2519
2520                 if (page->mapping != mapping || !PageUptodate(page))
2521                         goto unlock;
2522
2523                 max_idx = DIV_ROUND_UP(i_size_read(mapping->host), PAGE_SIZE);
2524                 if (page->index >= max_idx)
2525                         goto unlock;
2526
2527                 if (file->f_ra.mmap_miss > 0)
2528                         file->f_ra.mmap_miss--;
2529
2530                 vmf->address += (iter.index - last_pgoff) << PAGE_SHIFT;
2531                 if (vmf->pte)
2532                         vmf->pte += iter.index - last_pgoff;
2533                 last_pgoff = iter.index;
2534                 if (alloc_set_pte(vmf, NULL, page))
2535                         goto unlock;
2536                 unlock_page(page);
2537                 goto next;
2538 unlock:
2539                 unlock_page(page);
2540 skip:
2541                 put_page(page);
2542 next:
2543                 /* Huge page is mapped? No need to proceed. */
2544                 if (pmd_trans_huge(*vmf->pmd))
2545                         break;
2546                 if (iter.index == end_pgoff)
2547                         break;
2548         }
2549         rcu_read_unlock();
2550 }
2551 EXPORT_SYMBOL(filemap_map_pages);
2552
2553 int filemap_page_mkwrite(struct vm_fault *vmf)
2554 {
2555         struct page *page = vmf->page;
2556         struct inode *inode = file_inode(vmf->vma->vm_file);
2557         int ret = VM_FAULT_LOCKED;
2558
2559         sb_start_pagefault(inode->i_sb);
2560         file_update_time(vmf->vma->vm_file);
2561         lock_page(page);
2562         if (page->mapping != inode->i_mapping) {
2563                 unlock_page(page);
2564                 ret = VM_FAULT_NOPAGE;
2565                 goto out;
2566         }
2567         /*
2568          * We mark the page dirty already here so that when freeze is in
2569          * progress, we are guaranteed that writeback during freezing will
2570          * see the dirty page and writeprotect it again.
2571          */
2572         set_page_dirty(page);
2573         wait_for_stable_page(page);
2574 out:
2575         sb_end_pagefault(inode->i_sb);
2576         return ret;
2577 }
2578 EXPORT_SYMBOL(filemap_page_mkwrite);
2579
2580 const struct vm_operations_struct generic_file_vm_ops = {
2581         .fault          = filemap_fault,
2582         .map_pages      = filemap_map_pages,
2583         .page_mkwrite   = filemap_page_mkwrite,
2584 };
2585
2586 /* This is used for a general mmap of a disk file */
2587
2588 int generic_file_mmap(struct file * file, struct vm_area_struct * vma)
2589 {
2590         struct address_space *mapping = file->f_mapping;
2591
2592         if (!mapping->a_ops->readpage)
2593                 return -ENOEXEC;
2594         file_accessed(file);
2595         vma->vm_ops = &generic_file_vm_ops;
2596         return 0;
2597 }
2598
2599 /*
2600  * This is for filesystems which do not implement ->writepage.
2601  */
2602 int generic_file_readonly_mmap(struct file *file, struct vm_area_struct *vma)
2603 {
2604         if ((vma->vm_flags & VM_SHARED) && (vma->vm_flags & VM_MAYWRITE))
2605                 return -EINVAL;
2606         return generic_file_mmap(file, vma);
2607 }
2608 #else
2609 int generic_file_mmap(struct file * file, struct vm_area_struct * vma)
2610 {
2611         return -ENOSYS;
2612 }
2613 int generic_file_readonly_mmap(struct file * file, struct vm_area_struct * vma)
2614 {
2615         return -ENOSYS;
2616 }
2617 #endif /* CONFIG_MMU */
2618
2619 EXPORT_SYMBOL(generic_file_mmap);
2620 EXPORT_SYMBOL(generic_file_readonly_mmap);
2621
2622 static struct page *wait_on_page_read(struct page *page)
2623 {
2624         if (!IS_ERR(page)) {
2625                 wait_on_page_locked(page);
2626                 if (!PageUptodate(page)) {
2627                         put_page(page);
2628                         page = ERR_PTR(-EIO);
2629                 }
2630         }
2631         return page;
2632 }
2633
2634 static struct page *do_read_cache_page(struct address_space *mapping,
2635                                 pgoff_t index,
2636                                 int (*filler)(void *, struct page *),
2637                                 void *data,
2638                                 gfp_t gfp)
2639 {
2640         struct page *page;
2641         int err;
2642 repeat:
2643         page = find_get_page(mapping, index);
2644         if (!page) {
2645                 page = __page_cache_alloc(gfp | __GFP_COLD);
2646                 if (!page)
2647                         return ERR_PTR(-ENOMEM);
2648                 err = add_to_page_cache_lru(page, mapping, index, gfp);
2649                 if (unlikely(err)) {
2650                         put_page(page);
2651                         if (err == -EEXIST)
2652                                 goto repeat;
2653                         /* Presumably ENOMEM for radix tree node */
2654                         return ERR_PTR(err);
2655                 }
2656
2657 filler:
2658                 err = filler(data, page);
2659                 if (err < 0) {
2660                         put_page(page);
2661                         return ERR_PTR(err);
2662                 }
2663
2664                 page = wait_on_page_read(page);
2665                 if (IS_ERR(page))
2666                         return page;
2667                 goto out;
2668         }
2669         if (PageUptodate(page))
2670                 goto out;
2671
2672         /*
2673          * Page is not up to date and may be locked due one of the following
2674          * case a: Page is being filled and the page lock is held
2675          * case b: Read/write error clearing the page uptodate status
2676          * case c: Truncation in progress (page locked)
2677          * case d: Reclaim in progress
2678          *
2679          * Case a, the page will be up to date when the page is unlocked.
2680          *    There is no need to serialise on the page lock here as the page
2681          *    is pinned so the lock gives no additional protection. Even if the
2682          *    the page is truncated, the data is still valid if PageUptodate as
2683          *    it's a race vs truncate race.
2684          * Case b, the page will not be up to date
2685          * Case c, the page may be truncated but in itself, the data may still
2686          *    be valid after IO completes as it's a read vs truncate race. The
2687          *    operation must restart if the page is not uptodate on unlock but
2688          *    otherwise serialising on page lock to stabilise the mapping gives
2689          *    no additional guarantees to the caller as the page lock is
2690          *    released before return.
2691          * Case d, similar to truncation. If reclaim holds the page lock, it
2692          *    will be a race with remove_mapping that determines if the mapping
2693          *    is valid on unlock but otherwise the data is valid and there is
2694          *    no need to serialise with page lock.
2695          *
2696          * As the page lock gives no additional guarantee, we optimistically
2697          * wait on the page to be unlocked and check if it's up to date and
2698          * use the page if it is. Otherwise, the page lock is required to
2699          * distinguish between the different cases. The motivation is that we
2700          * avoid spurious serialisations and wakeups when multiple processes
2701          * wait on the same page for IO to complete.
2702          */
2703         wait_on_page_locked(page);
2704         if (PageUptodate(page))
2705                 goto out;
2706
2707         /* Distinguish between all the cases under the safety of the lock */
2708         lock_page(page);
2709
2710         /* Case c or d, restart the operation */
2711         if (!page->mapping) {
2712                 unlock_page(page);
2713                 put_page(page);
2714                 goto repeat;
2715         }
2716
2717         /* Someone else locked and filled the page in a very small window */
2718         if (PageUptodate(page)) {
2719                 unlock_page(page);
2720                 goto out;
2721         }
2722         goto filler;
2723
2724 out:
2725         mark_page_accessed(page);
2726         return page;
2727 }
2728
2729 /**
2730  * read_cache_page - read into page cache, fill it if needed
2731  * @mapping:    the page's address_space
2732  * @index:      the page index
2733  * @filler:     function to perform the read
2734  * @data:       first arg to filler(data, page) function, often left as NULL
2735  *
2736  * Read into the page cache. If a page already exists, and PageUptodate() is
2737  * not set, try to fill the page and wait for it to become unlocked.
2738  *
2739  * If the page does not get brought uptodate, return -EIO.
2740  */
2741 struct page *read_cache_page(struct address_space *mapping,
2742                                 pgoff_t index,
2743                                 int (*filler)(void *, struct page *),
2744                                 void *data)
2745 {
2746         return do_read_cache_page(mapping, index, filler, data, mapping_gfp_mask(mapping));
2747 }
2748 EXPORT_SYMBOL(read_cache_page);
2749
2750 /**
2751  * read_cache_page_gfp - read into page cache, using specified page allocation flags.
2752  * @mapping:    the page's address_space
2753  * @index:      the page index
2754  * @gfp:        the page allocator flags to use if allocating
2755  *
2756  * This is the same as "read_mapping_page(mapping, index, NULL)", but with
2757  * any new page allocations done using the specified allocation flags.
2758  *
2759  * If the page does not get brought uptodate, return -EIO.
2760  */
2761 struct page *read_cache_page_gfp(struct address_space *mapping,
2762                                 pgoff_t index,
2763                                 gfp_t gfp)
2764 {
2765         filler_t *filler = (filler_t *)mapping->a_ops->readpage;
2766
2767         return do_read_cache_page(mapping, index, filler, NULL, gfp);
2768 }
2769 EXPORT_SYMBOL(read_cache_page_gfp);
2770
2771 /*
2772  * Performs necessary checks before doing a write
2773  *
2774  * Can adjust writing position or amount of bytes to write.
2775  * Returns appropriate error code that caller should return or
2776  * zero in case that write should be allowed.
2777  */
2778 inline ssize_t generic_write_checks(struct kiocb *iocb, struct iov_iter *from)
2779 {
2780         struct file *file = iocb->ki_filp;
2781         struct inode *inode = file->f_mapping->host;
2782         unsigned long limit = rlimit(RLIMIT_FSIZE);
2783         loff_t pos;
2784
2785         if (!iov_iter_count(from))
2786                 return 0;
2787
2788         /* FIXME: this is for backwards compatibility with 2.4 */
2789         if (iocb->ki_flags & IOCB_APPEND)
2790                 iocb->ki_pos = i_size_read(inode);
2791
2792         pos = iocb->ki_pos;
2793
2794         if ((iocb->ki_flags & IOCB_NOWAIT) && !(iocb->ki_flags & IOCB_DIRECT))
2795                 return -EINVAL;
2796
2797         if (limit != RLIM_INFINITY) {
2798                 if (iocb->ki_pos >= limit) {
2799                         send_sig(SIGXFSZ, current, 0);
2800                         return -EFBIG;
2801                 }
2802                 iov_iter_truncate(from, limit - (unsigned long)pos);
2803         }
2804
2805         /*
2806          * LFS rule
2807          */
2808         if (unlikely(pos + iov_iter_count(from) > MAX_NON_LFS &&
2809                                 !(file->f_flags & O_LARGEFILE))) {
2810                 if (pos >= MAX_NON_LFS)
2811                         return -EFBIG;
2812                 iov_iter_truncate(from, MAX_NON_LFS - (unsigned long)pos);
2813         }
2814
2815         /*
2816          * Are we about to exceed the fs block limit ?
2817          *
2818          * If we have written data it becomes a short write.  If we have
2819          * exceeded without writing data we send a signal and return EFBIG.
2820          * Linus frestrict idea will clean these up nicely..
2821          */
2822         if (unlikely(pos >= inode->i_sb->s_maxbytes))
2823                 return -EFBIG;
2824
2825         iov_iter_truncate(from, inode->i_sb->s_maxbytes - pos);
2826         return iov_iter_count(from);
2827 }
2828 EXPORT_SYMBOL(generic_write_checks);
2829
2830 int pagecache_write_begin(struct file *file, struct address_space *mapping,
2831                                 loff_t pos, unsigned len, unsigned flags,
2832                                 struct page **pagep, void **fsdata)
2833 {
2834         const struct address_space_operations *aops = mapping->a_ops;
2835
2836         return aops->write_begin(file, mapping, pos, len, flags,
2837                                                         pagep, fsdata);
2838 }
2839 EXPORT_SYMBOL(pagecache_write_begin);
2840
2841 int pagecache_write_end(struct file *file, struct address_space *mapping,
2842                                 loff_t pos, unsigned len, unsigned copied,
2843                                 struct page *page, void *fsdata)
2844 {
2845         const struct address_space_operations *aops = mapping->a_ops;
2846
2847         return aops->write_end(file, mapping, pos, len, copied, page, fsdata);
2848 }
2849 EXPORT_SYMBOL(pagecache_write_end);
2850
2851 ssize_t
2852 generic_file_direct_write(struct kiocb *iocb, struct iov_iter *from)
2853 {
2854         struct file     *file = iocb->ki_filp;
2855         struct address_space *mapping = file->f_mapping;
2856         struct inode    *inode = mapping->host;
2857         loff_t          pos = iocb->ki_pos;
2858         ssize_t         written;
2859         size_t          write_len;
2860         pgoff_t         end;
2861
2862         write_len = iov_iter_count(from);
2863         end = (pos + write_len - 1) >> PAGE_SHIFT;
2864
2865         if (iocb->ki_flags & IOCB_NOWAIT) {
2866                 /* If there are pages to writeback, return */
2867                 if (filemap_range_has_page(inode->i_mapping, pos,
2868                                            pos + iov_iter_count(from)))
2869                         return -EAGAIN;
2870         } else {
2871                 written = filemap_write_and_wait_range(mapping, pos,
2872                                                         pos + write_len - 1);
2873                 if (written)
2874                         goto out;
2875         }
2876
2877         /*
2878          * After a write we want buffered reads to be sure to go to disk to get
2879          * the new data.  We invalidate clean cached page from the region we're
2880          * about to write.  We do this *before* the write so that we can return
2881          * without clobbering -EIOCBQUEUED from ->direct_IO().
2882          */
2883         written = invalidate_inode_pages2_range(mapping,
2884                                         pos >> PAGE_SHIFT, end);
2885         /*
2886          * If a page can not be invalidated, return 0 to fall back
2887          * to buffered write.
2888          */
2889         if (written) {
2890                 if (written == -EBUSY)
2891                         return 0;
2892                 goto out;
2893         }
2894
2895         written = mapping->a_ops->direct_IO(iocb, from);
2896
2897         /*
2898          * Finally, try again to invalidate clean pages which might have been
2899          * cached by non-direct readahead, or faulted in by get_user_pages()
2900          * if the source of the write was an mmap'ed region of the file
2901          * we're writing.  Either one is a pretty crazy thing to do,
2902          * so we don't support it 100%.  If this invalidation
2903          * fails, tough, the write still worked...
2904          */
2905         invalidate_inode_pages2_range(mapping,
2906                                 pos >> PAGE_SHIFT, end);
2907
2908         if (written > 0) {
2909                 pos += written;
2910                 write_len -= written;
2911                 if (pos > i_size_read(inode) && !S_ISBLK(inode->i_mode)) {
2912                         i_size_write(inode, pos);
2913                         mark_inode_dirty(inode);
2914                 }
2915                 iocb->ki_pos = pos;
2916         }
2917         iov_iter_revert(from, write_len - iov_iter_count(from));
2918 out:
2919         return written;
2920 }
2921 EXPORT_SYMBOL(generic_file_direct_write);
2922
2923 /*
2924  * Find or create a page at the given pagecache position. Return the locked
2925  * page. This function is specifically for buffered writes.
2926  */
2927 struct page *grab_cache_page_write_begin(struct address_space *mapping,
2928                                         pgoff_t index, unsigned flags)
2929 {
2930         struct page *page;
2931         int fgp_flags = FGP_LOCK|FGP_WRITE|FGP_CREAT;
2932
2933         if (flags & AOP_FLAG_NOFS)
2934                 fgp_flags |= FGP_NOFS;
2935
2936         page = pagecache_get_page(mapping, index, fgp_flags,
2937                         mapping_gfp_mask(mapping));
2938         if (page)
2939                 wait_for_stable_page(page);
2940
2941         return page;
2942 }
2943 EXPORT_SYMBOL(grab_cache_page_write_begin);
2944
2945 ssize_t generic_perform_write(struct file *file,
2946                                 struct iov_iter *i, loff_t pos)
2947 {
2948         struct address_space *mapping = file->f_mapping;
2949         const struct address_space_operations *a_ops = mapping->a_ops;
2950         long status = 0;
2951         ssize_t written = 0;
2952         unsigned int flags = 0;
2953
2954         do {
2955                 struct page *page;
2956                 unsigned long offset;   /* Offset into pagecache page */
2957                 unsigned long bytes;    /* Bytes to write to page */
2958                 size_t copied;          /* Bytes copied from user */
2959                 void *fsdata;
2960
2961                 offset = (pos & (PAGE_SIZE - 1));
2962                 bytes = min_t(unsigned long, PAGE_SIZE - offset,
2963                                                 iov_iter_count(i));
2964
2965 again:
2966                 /*
2967                  * Bring in the user page that we will copy from _first_.
2968                  * Otherwise there's a nasty deadlock on copying from the
2969                  * same page as we're writing to, without it being marked
2970                  * up-to-date.
2971                  *
2972                  * Not only is this an optimisation, but it is also required
2973                  * to check that the address is actually valid, when atomic
2974                  * usercopies are used, below.
2975                  */
2976                 if (unlikely(iov_iter_fault_in_readable(i, bytes))) {
2977                         status = -EFAULT;
2978                         break;
2979                 }
2980
2981                 if (fatal_signal_pending(current)) {
2982                         status = -EINTR;
2983                         break;
2984                 }
2985
2986                 status = a_ops->write_begin(file, mapping, pos, bytes, flags,
2987                                                 &page, &fsdata);
2988                 if (unlikely(status < 0))
2989                         break;
2990
2991                 if (mapping_writably_mapped(mapping))
2992                         flush_dcache_page(page);
2993
2994                 copied = iov_iter_copy_from_user_atomic(page, i, offset, bytes);
2995                 flush_dcache_page(page);
2996
2997                 status = a_ops->write_end(file, mapping, pos, bytes, copied,
2998                                                 page, fsdata);
2999                 if (unlikely(status < 0))
3000                         break;
3001                 copied = status;
3002
3003                 cond_resched();
3004
3005                 iov_iter_advance(i, copied);
3006                 if (unlikely(copied == 0)) {
3007                         /*
3008                          * If we were unable to copy any data at all, we must
3009                          * fall back to a single segment length write.
3010                          *
3011                          * If we didn't fallback here, we could livelock
3012                          * because not all segments in the iov can be copied at
3013                          * once without a pagefault.
3014                          */
3015                         bytes = min_t(unsigned long, PAGE_SIZE - offset,
3016                                                 iov_iter_single_seg_count(i));
3017                         goto again;
3018                 }
3019                 pos += copied;
3020                 written += copied;
3021
3022                 balance_dirty_pages_ratelimited(mapping);
3023         } while (iov_iter_count(i));
3024
3025         return written ? written : status;
3026 }
3027 EXPORT_SYMBOL(generic_perform_write);
3028
3029 /**
3030  * __generic_file_write_iter - write data to a file
3031  * @iocb:       IO state structure (file, offset, etc.)
3032  * @from:       iov_iter with data to write
3033  *
3034  * This function does all the work needed for actually writing data to a
3035  * file. It does all basic checks, removes SUID from the file, updates
3036  * modification times and calls proper subroutines depending on whether we
3037  * do direct IO or a standard buffered write.
3038  *
3039  * It expects i_mutex to be grabbed unless we work on a block device or similar
3040  * object which does not need locking at all.
3041  *
3042  * This function does *not* take care of syncing data in case of O_SYNC write.
3043  * A caller has to handle it. This is mainly due to the fact that we want to
3044  * avoid syncing under i_mutex.
3045  */
3046 ssize_t __generic_file_write_iter(struct kiocb *iocb, struct iov_iter *from)
3047 {
3048         struct file *file = iocb->ki_filp;
3049         struct address_space * mapping = file->f_mapping;
3050         struct inode    *inode = mapping->host;
3051         ssize_t         written = 0;
3052         ssize_t         err;
3053         ssize_t         status;
3054
3055         /* We can write back this queue in page reclaim */
3056         current->backing_dev_info = inode_to_bdi(inode);
3057         err = file_remove_privs(file);
3058         if (err)
3059                 goto out;
3060
3061         err = file_update_time(file);
3062         if (err)
3063                 goto out;
3064
3065         if (iocb->ki_flags & IOCB_DIRECT) {
3066                 loff_t pos, endbyte;
3067
3068                 written = generic_file_direct_write(iocb, from);
3069                 /*
3070                  * If the write stopped short of completing, fall back to
3071                  * buffered writes.  Some filesystems do this for writes to
3072                  * holes, for example.  For DAX files, a buffered write will
3073                  * not succeed (even if it did, DAX does not handle dirty
3074                  * page-cache pages correctly).
3075                  */
3076                 if (written < 0 || !iov_iter_count(from) || IS_DAX(inode))
3077                         goto out;
3078
3079                 status = generic_perform_write(file, from, pos = iocb->ki_pos);
3080                 /*
3081                  * If generic_perform_write() returned a synchronous error
3082                  * then we want to return the number of bytes which were
3083                  * direct-written, or the error code if that was zero.  Note
3084                  * that this differs from normal direct-io semantics, which
3085                  * will return -EFOO even if some bytes were written.
3086                  */
3087                 if (unlikely(status < 0)) {
3088                         err = status;
3089                         goto out;
3090                 }
3091                 /*
3092                  * We need to ensure that the page cache pages are written to
3093                  * disk and invalidated to preserve the expected O_DIRECT
3094                  * semantics.
3095                  */
3096                 endbyte = pos + status - 1;
3097                 err = filemap_write_and_wait_range(mapping, pos, endbyte);
3098                 if (err == 0) {
3099                         iocb->ki_pos = endbyte + 1;
3100                         written += status;
3101                         invalidate_mapping_pages(mapping,
3102                                                  pos >> PAGE_SHIFT,
3103                                                  endbyte >> PAGE_SHIFT);
3104                 } else {
3105                         /*
3106                          * We don't know how much we wrote, so just return
3107                          * the number of bytes which were direct-written
3108                          */
3109                 }
3110         } else {
3111                 written = generic_perform_write(file, from, iocb->ki_pos);
3112                 if (likely(written > 0))
3113                         iocb->ki_pos += written;
3114         }
3115 out:
3116         current->backing_dev_info = NULL;
3117         return written ? written : err;
3118 }
3119 EXPORT_SYMBOL(__generic_file_write_iter);
3120
3121 /**
3122  * generic_file_write_iter - write data to a file
3123  * @iocb:       IO state structure
3124  * @from:       iov_iter with data to write
3125  *
3126  * This is a wrapper around __generic_file_write_iter() to be used by most
3127  * filesystems. It takes care of syncing the file in case of O_SYNC file
3128  * and acquires i_mutex as needed.
3129  */
3130 ssize_t generic_file_write_iter(struct kiocb *iocb, struct iov_iter *from)
3131 {
3132         struct file *file = iocb->ki_filp;
3133         struct inode *inode = file->f_mapping->host;
3134         ssize_t ret;
3135
3136         inode_lock(inode);
3137         ret = generic_write_checks(iocb, from);
3138         if (ret > 0)
3139                 ret = __generic_file_write_iter(iocb, from);
3140         inode_unlock(inode);
3141
3142         if (ret > 0)
3143                 ret = generic_write_sync(iocb, ret);
3144         return ret;
3145 }
3146 EXPORT_SYMBOL(generic_file_write_iter);
3147
3148 /**
3149  * try_to_release_page() - release old fs-specific metadata on a page
3150  *
3151  * @page: the page which the kernel is trying to free
3152  * @gfp_mask: memory allocation flags (and I/O mode)
3153  *
3154  * The address_space is to try to release any data against the page
3155  * (presumably at page->private).  If the release was successful, return '1'.
3156  * Otherwise return zero.
3157  *
3158  * This may also be called if PG_fscache is set on a page, indicating that the
3159  * page is known to the local caching routines.
3160  *
3161  * The @gfp_mask argument specifies whether I/O may be performed to release
3162  * this page (__GFP_IO), and whether the call may block (__GFP_RECLAIM & __GFP_FS).
3163  *
3164  */
3165 int try_to_release_page(struct page *page, gfp_t gfp_mask)
3166 {
3167         struct address_space * const mapping = page->mapping;
3168
3169         BUG_ON(!PageLocked(page));
3170         if (PageWriteback(page))
3171                 return 0;
3172
3173         if (mapping && mapping->a_ops->releasepage)
3174                 return mapping->a_ops->releasepage(page, gfp_mask);
3175         return try_to_free_buffers(page);
3176 }
3177
3178 EXPORT_SYMBOL(try_to_release_page);