mm: implement find_get_pages_range_tag()
[sfrench/cifs-2.6.git] / mm / filemap.c
1 /*
2  *      linux/mm/filemap.c
3  *
4  * Copyright (C) 1994-1999  Linus Torvalds
5  */
6
7 /*
8  * This file handles the generic file mmap semantics used by
9  * most "normal" filesystems (but you don't /have/ to use this:
10  * the NFS filesystem used to do this differently, for example)
11  */
12 #include <linux/export.h>
13 #include <linux/compiler.h>
14 #include <linux/dax.h>
15 #include <linux/fs.h>
16 #include <linux/sched/signal.h>
17 #include <linux/uaccess.h>
18 #include <linux/capability.h>
19 #include <linux/kernel_stat.h>
20 #include <linux/gfp.h>
21 #include <linux/mm.h>
22 #include <linux/swap.h>
23 #include <linux/mman.h>
24 #include <linux/pagemap.h>
25 #include <linux/file.h>
26 #include <linux/uio.h>
27 #include <linux/hash.h>
28 #include <linux/writeback.h>
29 #include <linux/backing-dev.h>
30 #include <linux/pagevec.h>
31 #include <linux/blkdev.h>
32 #include <linux/security.h>
33 #include <linux/cpuset.h>
34 #include <linux/hardirq.h> /* for BUG_ON(!in_atomic()) only */
35 #include <linux/hugetlb.h>
36 #include <linux/memcontrol.h>
37 #include <linux/cleancache.h>
38 #include <linux/rmap.h>
39 #include "internal.h"
40
41 #define CREATE_TRACE_POINTS
42 #include <trace/events/filemap.h>
43
44 /*
45  * FIXME: remove all knowledge of the buffer layer from the core VM
46  */
47 #include <linux/buffer_head.h> /* for try_to_free_buffers */
48
49 #include <asm/mman.h>
50
51 /*
52  * Shared mappings implemented 30.11.1994. It's not fully working yet,
53  * though.
54  *
55  * Shared mappings now work. 15.8.1995  Bruno.
56  *
57  * finished 'unifying' the page and buffer cache and SMP-threaded the
58  * page-cache, 21.05.1999, Ingo Molnar <mingo@redhat.com>
59  *
60  * SMP-threaded pagemap-LRU 1999, Andrea Arcangeli <andrea@suse.de>
61  */
62
63 /*
64  * Lock ordering:
65  *
66  *  ->i_mmap_rwsem              (truncate_pagecache)
67  *    ->private_lock            (__free_pte->__set_page_dirty_buffers)
68  *      ->swap_lock             (exclusive_swap_page, others)
69  *        ->mapping->tree_lock
70  *
71  *  ->i_mutex
72  *    ->i_mmap_rwsem            (truncate->unmap_mapping_range)
73  *
74  *  ->mmap_sem
75  *    ->i_mmap_rwsem
76  *      ->page_table_lock or pte_lock   (various, mainly in memory.c)
77  *        ->mapping->tree_lock  (arch-dependent flush_dcache_mmap_lock)
78  *
79  *  ->mmap_sem
80  *    ->lock_page               (access_process_vm)
81  *
82  *  ->i_mutex                   (generic_perform_write)
83  *    ->mmap_sem                (fault_in_pages_readable->do_page_fault)
84  *
85  *  bdi->wb.list_lock
86  *    sb_lock                   (fs/fs-writeback.c)
87  *    ->mapping->tree_lock      (__sync_single_inode)
88  *
89  *  ->i_mmap_rwsem
90  *    ->anon_vma.lock           (vma_adjust)
91  *
92  *  ->anon_vma.lock
93  *    ->page_table_lock or pte_lock     (anon_vma_prepare and various)
94  *
95  *  ->page_table_lock or pte_lock
96  *    ->swap_lock               (try_to_unmap_one)
97  *    ->private_lock            (try_to_unmap_one)
98  *    ->tree_lock               (try_to_unmap_one)
99  *    ->zone_lru_lock(zone)     (follow_page->mark_page_accessed)
100  *    ->zone_lru_lock(zone)     (check_pte_range->isolate_lru_page)
101  *    ->private_lock            (page_remove_rmap->set_page_dirty)
102  *    ->tree_lock               (page_remove_rmap->set_page_dirty)
103  *    bdi.wb->list_lock         (page_remove_rmap->set_page_dirty)
104  *    ->inode->i_lock           (page_remove_rmap->set_page_dirty)
105  *    ->memcg->move_lock        (page_remove_rmap->lock_page_memcg)
106  *    bdi.wb->list_lock         (zap_pte_range->set_page_dirty)
107  *    ->inode->i_lock           (zap_pte_range->set_page_dirty)
108  *    ->private_lock            (zap_pte_range->__set_page_dirty_buffers)
109  *
110  * ->i_mmap_rwsem
111  *   ->tasklist_lock            (memory_failure, collect_procs_ao)
112  */
113
114 static int page_cache_tree_insert(struct address_space *mapping,
115                                   struct page *page, void **shadowp)
116 {
117         struct radix_tree_node *node;
118         void **slot;
119         int error;
120
121         error = __radix_tree_create(&mapping->page_tree, page->index, 0,
122                                     &node, &slot);
123         if (error)
124                 return error;
125         if (*slot) {
126                 void *p;
127
128                 p = radix_tree_deref_slot_protected(slot, &mapping->tree_lock);
129                 if (!radix_tree_exceptional_entry(p))
130                         return -EEXIST;
131
132                 mapping->nrexceptional--;
133                 if (shadowp)
134                         *shadowp = p;
135         }
136         __radix_tree_replace(&mapping->page_tree, node, slot, page,
137                              workingset_update_node, mapping);
138         mapping->nrpages++;
139         return 0;
140 }
141
142 static void page_cache_tree_delete(struct address_space *mapping,
143                                    struct page *page, void *shadow)
144 {
145         int i, nr;
146
147         /* hugetlb pages are represented by one entry in the radix tree */
148         nr = PageHuge(page) ? 1 : hpage_nr_pages(page);
149
150         VM_BUG_ON_PAGE(!PageLocked(page), page);
151         VM_BUG_ON_PAGE(PageTail(page), page);
152         VM_BUG_ON_PAGE(nr != 1 && shadow, page);
153
154         for (i = 0; i < nr; i++) {
155                 struct radix_tree_node *node;
156                 void **slot;
157
158                 __radix_tree_lookup(&mapping->page_tree, page->index + i,
159                                     &node, &slot);
160
161                 VM_BUG_ON_PAGE(!node && nr != 1, page);
162
163                 radix_tree_clear_tags(&mapping->page_tree, node, slot);
164                 __radix_tree_replace(&mapping->page_tree, node, slot, shadow,
165                                      workingset_update_node, mapping);
166         }
167
168         if (shadow) {
169                 mapping->nrexceptional += nr;
170                 /*
171                  * Make sure the nrexceptional update is committed before
172                  * the nrpages update so that final truncate racing
173                  * with reclaim does not see both counters 0 at the
174                  * same time and miss a shadow entry.
175                  */
176                 smp_wmb();
177         }
178         mapping->nrpages -= nr;
179 }
180
181 /*
182  * Delete a page from the page cache and free it. Caller has to make
183  * sure the page is locked and that nobody else uses it - or that usage
184  * is safe.  The caller must hold the mapping's tree_lock.
185  */
186 void __delete_from_page_cache(struct page *page, void *shadow)
187 {
188         struct address_space *mapping = page->mapping;
189         int nr = hpage_nr_pages(page);
190
191         trace_mm_filemap_delete_from_page_cache(page);
192         /*
193          * if we're uptodate, flush out into the cleancache, otherwise
194          * invalidate any existing cleancache entries.  We can't leave
195          * stale data around in the cleancache once our page is gone
196          */
197         if (PageUptodate(page) && PageMappedToDisk(page))
198                 cleancache_put_page(page);
199         else
200                 cleancache_invalidate_page(mapping, page);
201
202         VM_BUG_ON_PAGE(PageTail(page), page);
203         VM_BUG_ON_PAGE(page_mapped(page), page);
204         if (!IS_ENABLED(CONFIG_DEBUG_VM) && unlikely(page_mapped(page))) {
205                 int mapcount;
206
207                 pr_alert("BUG: Bad page cache in process %s  pfn:%05lx\n",
208                          current->comm, page_to_pfn(page));
209                 dump_page(page, "still mapped when deleted");
210                 dump_stack();
211                 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
212
213                 mapcount = page_mapcount(page);
214                 if (mapping_exiting(mapping) &&
215                     page_count(page) >= mapcount + 2) {
216                         /*
217                          * All vmas have already been torn down, so it's
218                          * a good bet that actually the page is unmapped,
219                          * and we'd prefer not to leak it: if we're wrong,
220                          * some other bad page check should catch it later.
221                          */
222                         page_mapcount_reset(page);
223                         page_ref_sub(page, mapcount);
224                 }
225         }
226
227         page_cache_tree_delete(mapping, page, shadow);
228
229         page->mapping = NULL;
230         /* Leave page->index set: truncation lookup relies upon it */
231
232         /* hugetlb pages do not participate in page cache accounting. */
233         if (PageHuge(page))
234                 return;
235
236         __mod_node_page_state(page_pgdat(page), NR_FILE_PAGES, -nr);
237         if (PageSwapBacked(page)) {
238                 __mod_node_page_state(page_pgdat(page), NR_SHMEM, -nr);
239                 if (PageTransHuge(page))
240                         __dec_node_page_state(page, NR_SHMEM_THPS);
241         } else {
242                 VM_BUG_ON_PAGE(PageTransHuge(page), page);
243         }
244
245         /*
246          * At this point page must be either written or cleaned by truncate.
247          * Dirty page here signals a bug and loss of unwritten data.
248          *
249          * This fixes dirty accounting after removing the page entirely but
250          * leaves PageDirty set: it has no effect for truncated page and
251          * anyway will be cleared before returning page into buddy allocator.
252          */
253         if (WARN_ON_ONCE(PageDirty(page)))
254                 account_page_cleaned(page, mapping, inode_to_wb(mapping->host));
255 }
256
257 /**
258  * delete_from_page_cache - delete page from page cache
259  * @page: the page which the kernel is trying to remove from page cache
260  *
261  * This must be called only on pages that have been verified to be in the page
262  * cache and locked.  It will never put the page into the free list, the caller
263  * has a reference on the page.
264  */
265 void delete_from_page_cache(struct page *page)
266 {
267         struct address_space *mapping = page_mapping(page);
268         unsigned long flags;
269         void (*freepage)(struct page *);
270
271         BUG_ON(!PageLocked(page));
272
273         freepage = mapping->a_ops->freepage;
274
275         spin_lock_irqsave(&mapping->tree_lock, flags);
276         __delete_from_page_cache(page, NULL);
277         spin_unlock_irqrestore(&mapping->tree_lock, flags);
278
279         if (freepage)
280                 freepage(page);
281
282         if (PageTransHuge(page) && !PageHuge(page)) {
283                 page_ref_sub(page, HPAGE_PMD_NR);
284                 VM_BUG_ON_PAGE(page_count(page) <= 0, page);
285         } else {
286                 put_page(page);
287         }
288 }
289 EXPORT_SYMBOL(delete_from_page_cache);
290
291 int filemap_check_errors(struct address_space *mapping)
292 {
293         int ret = 0;
294         /* Check for outstanding write errors */
295         if (test_bit(AS_ENOSPC, &mapping->flags) &&
296             test_and_clear_bit(AS_ENOSPC, &mapping->flags))
297                 ret = -ENOSPC;
298         if (test_bit(AS_EIO, &mapping->flags) &&
299             test_and_clear_bit(AS_EIO, &mapping->flags))
300                 ret = -EIO;
301         return ret;
302 }
303 EXPORT_SYMBOL(filemap_check_errors);
304
305 static int filemap_check_and_keep_errors(struct address_space *mapping)
306 {
307         /* Check for outstanding write errors */
308         if (test_bit(AS_EIO, &mapping->flags))
309                 return -EIO;
310         if (test_bit(AS_ENOSPC, &mapping->flags))
311                 return -ENOSPC;
312         return 0;
313 }
314
315 /**
316  * __filemap_fdatawrite_range - start writeback on mapping dirty pages in range
317  * @mapping:    address space structure to write
318  * @start:      offset in bytes where the range starts
319  * @end:        offset in bytes where the range ends (inclusive)
320  * @sync_mode:  enable synchronous operation
321  *
322  * Start writeback against all of a mapping's dirty pages that lie
323  * within the byte offsets <start, end> inclusive.
324  *
325  * If sync_mode is WB_SYNC_ALL then this is a "data integrity" operation, as
326  * opposed to a regular memory cleansing writeback.  The difference between
327  * these two operations is that if a dirty page/buffer is encountered, it must
328  * be waited upon, and not just skipped over.
329  */
330 int __filemap_fdatawrite_range(struct address_space *mapping, loff_t start,
331                                 loff_t end, int sync_mode)
332 {
333         int ret;
334         struct writeback_control wbc = {
335                 .sync_mode = sync_mode,
336                 .nr_to_write = LONG_MAX,
337                 .range_start = start,
338                 .range_end = end,
339         };
340
341         if (!mapping_cap_writeback_dirty(mapping))
342                 return 0;
343
344         wbc_attach_fdatawrite_inode(&wbc, mapping->host);
345         ret = do_writepages(mapping, &wbc);
346         wbc_detach_inode(&wbc);
347         return ret;
348 }
349
350 static inline int __filemap_fdatawrite(struct address_space *mapping,
351         int sync_mode)
352 {
353         return __filemap_fdatawrite_range(mapping, 0, LLONG_MAX, sync_mode);
354 }
355
356 int filemap_fdatawrite(struct address_space *mapping)
357 {
358         return __filemap_fdatawrite(mapping, WB_SYNC_ALL);
359 }
360 EXPORT_SYMBOL(filemap_fdatawrite);
361
362 int filemap_fdatawrite_range(struct address_space *mapping, loff_t start,
363                                 loff_t end)
364 {
365         return __filemap_fdatawrite_range(mapping, start, end, WB_SYNC_ALL);
366 }
367 EXPORT_SYMBOL(filemap_fdatawrite_range);
368
369 /**
370  * filemap_flush - mostly a non-blocking flush
371  * @mapping:    target address_space
372  *
373  * This is a mostly non-blocking flush.  Not suitable for data-integrity
374  * purposes - I/O may not be started against all dirty pages.
375  */
376 int filemap_flush(struct address_space *mapping)
377 {
378         return __filemap_fdatawrite(mapping, WB_SYNC_NONE);
379 }
380 EXPORT_SYMBOL(filemap_flush);
381
382 /**
383  * filemap_range_has_page - check if a page exists in range.
384  * @mapping:           address space within which to check
385  * @start_byte:        offset in bytes where the range starts
386  * @end_byte:          offset in bytes where the range ends (inclusive)
387  *
388  * Find at least one page in the range supplied, usually used to check if
389  * direct writing in this range will trigger a writeback.
390  */
391 bool filemap_range_has_page(struct address_space *mapping,
392                            loff_t start_byte, loff_t end_byte)
393 {
394         pgoff_t index = start_byte >> PAGE_SHIFT;
395         pgoff_t end = end_byte >> PAGE_SHIFT;
396         struct page *page;
397
398         if (end_byte < start_byte)
399                 return false;
400
401         if (mapping->nrpages == 0)
402                 return false;
403
404         if (!find_get_pages_range(mapping, &index, end, 1, &page))
405                 return false;
406         put_page(page);
407         return true;
408 }
409 EXPORT_SYMBOL(filemap_range_has_page);
410
411 static void __filemap_fdatawait_range(struct address_space *mapping,
412                                      loff_t start_byte, loff_t end_byte)
413 {
414         pgoff_t index = start_byte >> PAGE_SHIFT;
415         pgoff_t end = end_byte >> PAGE_SHIFT;
416         struct pagevec pvec;
417         int nr_pages;
418
419         if (end_byte < start_byte)
420                 return;
421
422         pagevec_init(&pvec, 0);
423         while ((index <= end) &&
424                         (nr_pages = pagevec_lookup_tag(&pvec, mapping, &index,
425                         PAGECACHE_TAG_WRITEBACK,
426                         min(end - index, (pgoff_t)PAGEVEC_SIZE-1) + 1)) != 0) {
427                 unsigned i;
428
429                 for (i = 0; i < nr_pages; i++) {
430                         struct page *page = pvec.pages[i];
431
432                         /* until radix tree lookup accepts end_index */
433                         if (page->index > end)
434                                 continue;
435
436                         wait_on_page_writeback(page);
437                         ClearPageError(page);
438                 }
439                 pagevec_release(&pvec);
440                 cond_resched();
441         }
442 }
443
444 /**
445  * filemap_fdatawait_range - wait for writeback to complete
446  * @mapping:            address space structure to wait for
447  * @start_byte:         offset in bytes where the range starts
448  * @end_byte:           offset in bytes where the range ends (inclusive)
449  *
450  * Walk the list of under-writeback pages of the given address space
451  * in the given range and wait for all of them.  Check error status of
452  * the address space and return it.
453  *
454  * Since the error status of the address space is cleared by this function,
455  * callers are responsible for checking the return value and handling and/or
456  * reporting the error.
457  */
458 int filemap_fdatawait_range(struct address_space *mapping, loff_t start_byte,
459                             loff_t end_byte)
460 {
461         __filemap_fdatawait_range(mapping, start_byte, end_byte);
462         return filemap_check_errors(mapping);
463 }
464 EXPORT_SYMBOL(filemap_fdatawait_range);
465
466 /**
467  * file_fdatawait_range - wait for writeback to complete
468  * @file:               file pointing to address space structure to wait for
469  * @start_byte:         offset in bytes where the range starts
470  * @end_byte:           offset in bytes where the range ends (inclusive)
471  *
472  * Walk the list of under-writeback pages of the address space that file
473  * refers to, in the given range and wait for all of them.  Check error
474  * status of the address space vs. the file->f_wb_err cursor and return it.
475  *
476  * Since the error status of the file is advanced by this function,
477  * callers are responsible for checking the return value and handling and/or
478  * reporting the error.
479  */
480 int file_fdatawait_range(struct file *file, loff_t start_byte, loff_t end_byte)
481 {
482         struct address_space *mapping = file->f_mapping;
483
484         __filemap_fdatawait_range(mapping, start_byte, end_byte);
485         return file_check_and_advance_wb_err(file);
486 }
487 EXPORT_SYMBOL(file_fdatawait_range);
488
489 /**
490  * filemap_fdatawait_keep_errors - wait for writeback without clearing errors
491  * @mapping: address space structure to wait for
492  *
493  * Walk the list of under-writeback pages of the given address space
494  * and wait for all of them.  Unlike filemap_fdatawait(), this function
495  * does not clear error status of the address space.
496  *
497  * Use this function if callers don't handle errors themselves.  Expected
498  * call sites are system-wide / filesystem-wide data flushers: e.g. sync(2),
499  * fsfreeze(8)
500  */
501 int filemap_fdatawait_keep_errors(struct address_space *mapping)
502 {
503         __filemap_fdatawait_range(mapping, 0, LLONG_MAX);
504         return filemap_check_and_keep_errors(mapping);
505 }
506 EXPORT_SYMBOL(filemap_fdatawait_keep_errors);
507
508 static bool mapping_needs_writeback(struct address_space *mapping)
509 {
510         return (!dax_mapping(mapping) && mapping->nrpages) ||
511             (dax_mapping(mapping) && mapping->nrexceptional);
512 }
513
514 int filemap_write_and_wait(struct address_space *mapping)
515 {
516         int err = 0;
517
518         if (mapping_needs_writeback(mapping)) {
519                 err = filemap_fdatawrite(mapping);
520                 /*
521                  * Even if the above returned error, the pages may be
522                  * written partially (e.g. -ENOSPC), so we wait for it.
523                  * But the -EIO is special case, it may indicate the worst
524                  * thing (e.g. bug) happened, so we avoid waiting for it.
525                  */
526                 if (err != -EIO) {
527                         int err2 = filemap_fdatawait(mapping);
528                         if (!err)
529                                 err = err2;
530                 } else {
531                         /* Clear any previously stored errors */
532                         filemap_check_errors(mapping);
533                 }
534         } else {
535                 err = filemap_check_errors(mapping);
536         }
537         return err;
538 }
539 EXPORT_SYMBOL(filemap_write_and_wait);
540
541 /**
542  * filemap_write_and_wait_range - write out & wait on a file range
543  * @mapping:    the address_space for the pages
544  * @lstart:     offset in bytes where the range starts
545  * @lend:       offset in bytes where the range ends (inclusive)
546  *
547  * Write out and wait upon file offsets lstart->lend, inclusive.
548  *
549  * Note that @lend is inclusive (describes the last byte to be written) so
550  * that this function can be used to write to the very end-of-file (end = -1).
551  */
552 int filemap_write_and_wait_range(struct address_space *mapping,
553                                  loff_t lstart, loff_t lend)
554 {
555         int err = 0;
556
557         if (mapping_needs_writeback(mapping)) {
558                 err = __filemap_fdatawrite_range(mapping, lstart, lend,
559                                                  WB_SYNC_ALL);
560                 /* See comment of filemap_write_and_wait() */
561                 if (err != -EIO) {
562                         int err2 = filemap_fdatawait_range(mapping,
563                                                 lstart, lend);
564                         if (!err)
565                                 err = err2;
566                 } else {
567                         /* Clear any previously stored errors */
568                         filemap_check_errors(mapping);
569                 }
570         } else {
571                 err = filemap_check_errors(mapping);
572         }
573         return err;
574 }
575 EXPORT_SYMBOL(filemap_write_and_wait_range);
576
577 void __filemap_set_wb_err(struct address_space *mapping, int err)
578 {
579         errseq_t eseq = errseq_set(&mapping->wb_err, err);
580
581         trace_filemap_set_wb_err(mapping, eseq);
582 }
583 EXPORT_SYMBOL(__filemap_set_wb_err);
584
585 /**
586  * file_check_and_advance_wb_err - report wb error (if any) that was previously
587  *                                 and advance wb_err to current one
588  * @file: struct file on which the error is being reported
589  *
590  * When userland calls fsync (or something like nfsd does the equivalent), we
591  * want to report any writeback errors that occurred since the last fsync (or
592  * since the file was opened if there haven't been any).
593  *
594  * Grab the wb_err from the mapping. If it matches what we have in the file,
595  * then just quickly return 0. The file is all caught up.
596  *
597  * If it doesn't match, then take the mapping value, set the "seen" flag in
598  * it and try to swap it into place. If it works, or another task beat us
599  * to it with the new value, then update the f_wb_err and return the error
600  * portion. The error at this point must be reported via proper channels
601  * (a'la fsync, or NFS COMMIT operation, etc.).
602  *
603  * While we handle mapping->wb_err with atomic operations, the f_wb_err
604  * value is protected by the f_lock since we must ensure that it reflects
605  * the latest value swapped in for this file descriptor.
606  */
607 int file_check_and_advance_wb_err(struct file *file)
608 {
609         int err = 0;
610         errseq_t old = READ_ONCE(file->f_wb_err);
611         struct address_space *mapping = file->f_mapping;
612
613         /* Locklessly handle the common case where nothing has changed */
614         if (errseq_check(&mapping->wb_err, old)) {
615                 /* Something changed, must use slow path */
616                 spin_lock(&file->f_lock);
617                 old = file->f_wb_err;
618                 err = errseq_check_and_advance(&mapping->wb_err,
619                                                 &file->f_wb_err);
620                 trace_file_check_and_advance_wb_err(file, old);
621                 spin_unlock(&file->f_lock);
622         }
623
624         /*
625          * We're mostly using this function as a drop in replacement for
626          * filemap_check_errors. Clear AS_EIO/AS_ENOSPC to emulate the effect
627          * that the legacy code would have had on these flags.
628          */
629         clear_bit(AS_EIO, &mapping->flags);
630         clear_bit(AS_ENOSPC, &mapping->flags);
631         return err;
632 }
633 EXPORT_SYMBOL(file_check_and_advance_wb_err);
634
635 /**
636  * file_write_and_wait_range - write out & wait on a file range
637  * @file:       file pointing to address_space with pages
638  * @lstart:     offset in bytes where the range starts
639  * @lend:       offset in bytes where the range ends (inclusive)
640  *
641  * Write out and wait upon file offsets lstart->lend, inclusive.
642  *
643  * Note that @lend is inclusive (describes the last byte to be written) so
644  * that this function can be used to write to the very end-of-file (end = -1).
645  *
646  * After writing out and waiting on the data, we check and advance the
647  * f_wb_err cursor to the latest value, and return any errors detected there.
648  */
649 int file_write_and_wait_range(struct file *file, loff_t lstart, loff_t lend)
650 {
651         int err = 0, err2;
652         struct address_space *mapping = file->f_mapping;
653
654         if (mapping_needs_writeback(mapping)) {
655                 err = __filemap_fdatawrite_range(mapping, lstart, lend,
656                                                  WB_SYNC_ALL);
657                 /* See comment of filemap_write_and_wait() */
658                 if (err != -EIO)
659                         __filemap_fdatawait_range(mapping, lstart, lend);
660         }
661         err2 = file_check_and_advance_wb_err(file);
662         if (!err)
663                 err = err2;
664         return err;
665 }
666 EXPORT_SYMBOL(file_write_and_wait_range);
667
668 /**
669  * replace_page_cache_page - replace a pagecache page with a new one
670  * @old:        page to be replaced
671  * @new:        page to replace with
672  * @gfp_mask:   allocation mode
673  *
674  * This function replaces a page in the pagecache with a new one.  On
675  * success it acquires the pagecache reference for the new page and
676  * drops it for the old page.  Both the old and new pages must be
677  * locked.  This function does not add the new page to the LRU, the
678  * caller must do that.
679  *
680  * The remove + add is atomic.  The only way this function can fail is
681  * memory allocation failure.
682  */
683 int replace_page_cache_page(struct page *old, struct page *new, gfp_t gfp_mask)
684 {
685         int error;
686
687         VM_BUG_ON_PAGE(!PageLocked(old), old);
688         VM_BUG_ON_PAGE(!PageLocked(new), new);
689         VM_BUG_ON_PAGE(new->mapping, new);
690
691         error = radix_tree_preload(gfp_mask & ~__GFP_HIGHMEM);
692         if (!error) {
693                 struct address_space *mapping = old->mapping;
694                 void (*freepage)(struct page *);
695                 unsigned long flags;
696
697                 pgoff_t offset = old->index;
698                 freepage = mapping->a_ops->freepage;
699
700                 get_page(new);
701                 new->mapping = mapping;
702                 new->index = offset;
703
704                 spin_lock_irqsave(&mapping->tree_lock, flags);
705                 __delete_from_page_cache(old, NULL);
706                 error = page_cache_tree_insert(mapping, new, NULL);
707                 BUG_ON(error);
708
709                 /*
710                  * hugetlb pages do not participate in page cache accounting.
711                  */
712                 if (!PageHuge(new))
713                         __inc_node_page_state(new, NR_FILE_PAGES);
714                 if (PageSwapBacked(new))
715                         __inc_node_page_state(new, NR_SHMEM);
716                 spin_unlock_irqrestore(&mapping->tree_lock, flags);
717                 mem_cgroup_migrate(old, new);
718                 radix_tree_preload_end();
719                 if (freepage)
720                         freepage(old);
721                 put_page(old);
722         }
723
724         return error;
725 }
726 EXPORT_SYMBOL_GPL(replace_page_cache_page);
727
728 static int __add_to_page_cache_locked(struct page *page,
729                                       struct address_space *mapping,
730                                       pgoff_t offset, gfp_t gfp_mask,
731                                       void **shadowp)
732 {
733         int huge = PageHuge(page);
734         struct mem_cgroup *memcg;
735         int error;
736
737         VM_BUG_ON_PAGE(!PageLocked(page), page);
738         VM_BUG_ON_PAGE(PageSwapBacked(page), page);
739
740         if (!huge) {
741                 error = mem_cgroup_try_charge(page, current->mm,
742                                               gfp_mask, &memcg, false);
743                 if (error)
744                         return error;
745         }
746
747         error = radix_tree_maybe_preload(gfp_mask & ~__GFP_HIGHMEM);
748         if (error) {
749                 if (!huge)
750                         mem_cgroup_cancel_charge(page, memcg, false);
751                 return error;
752         }
753
754         get_page(page);
755         page->mapping = mapping;
756         page->index = offset;
757
758         spin_lock_irq(&mapping->tree_lock);
759         error = page_cache_tree_insert(mapping, page, shadowp);
760         radix_tree_preload_end();
761         if (unlikely(error))
762                 goto err_insert;
763
764         /* hugetlb pages do not participate in page cache accounting. */
765         if (!huge)
766                 __inc_node_page_state(page, NR_FILE_PAGES);
767         spin_unlock_irq(&mapping->tree_lock);
768         if (!huge)
769                 mem_cgroup_commit_charge(page, memcg, false, false);
770         trace_mm_filemap_add_to_page_cache(page);
771         return 0;
772 err_insert:
773         page->mapping = NULL;
774         /* Leave page->index set: truncation relies upon it */
775         spin_unlock_irq(&mapping->tree_lock);
776         if (!huge)
777                 mem_cgroup_cancel_charge(page, memcg, false);
778         put_page(page);
779         return error;
780 }
781
782 /**
783  * add_to_page_cache_locked - add a locked page to the pagecache
784  * @page:       page to add
785  * @mapping:    the page's address_space
786  * @offset:     page index
787  * @gfp_mask:   page allocation mode
788  *
789  * This function is used to add a page to the pagecache. It must be locked.
790  * This function does not add the page to the LRU.  The caller must do that.
791  */
792 int add_to_page_cache_locked(struct page *page, struct address_space *mapping,
793                 pgoff_t offset, gfp_t gfp_mask)
794 {
795         return __add_to_page_cache_locked(page, mapping, offset,
796                                           gfp_mask, NULL);
797 }
798 EXPORT_SYMBOL(add_to_page_cache_locked);
799
800 int add_to_page_cache_lru(struct page *page, struct address_space *mapping,
801                                 pgoff_t offset, gfp_t gfp_mask)
802 {
803         void *shadow = NULL;
804         int ret;
805
806         __SetPageLocked(page);
807         ret = __add_to_page_cache_locked(page, mapping, offset,
808                                          gfp_mask, &shadow);
809         if (unlikely(ret))
810                 __ClearPageLocked(page);
811         else {
812                 /*
813                  * The page might have been evicted from cache only
814                  * recently, in which case it should be activated like
815                  * any other repeatedly accessed page.
816                  * The exception is pages getting rewritten; evicting other
817                  * data from the working set, only to cache data that will
818                  * get overwritten with something else, is a waste of memory.
819                  */
820                 if (!(gfp_mask & __GFP_WRITE) &&
821                     shadow && workingset_refault(shadow)) {
822                         SetPageActive(page);
823                         workingset_activation(page);
824                 } else
825                         ClearPageActive(page);
826                 lru_cache_add(page);
827         }
828         return ret;
829 }
830 EXPORT_SYMBOL_GPL(add_to_page_cache_lru);
831
832 #ifdef CONFIG_NUMA
833 struct page *__page_cache_alloc(gfp_t gfp)
834 {
835         int n;
836         struct page *page;
837
838         if (cpuset_do_page_mem_spread()) {
839                 unsigned int cpuset_mems_cookie;
840                 do {
841                         cpuset_mems_cookie = read_mems_allowed_begin();
842                         n = cpuset_mem_spread_node();
843                         page = __alloc_pages_node(n, gfp, 0);
844                 } while (!page && read_mems_allowed_retry(cpuset_mems_cookie));
845
846                 return page;
847         }
848         return alloc_pages(gfp, 0);
849 }
850 EXPORT_SYMBOL(__page_cache_alloc);
851 #endif
852
853 /*
854  * In order to wait for pages to become available there must be
855  * waitqueues associated with pages. By using a hash table of
856  * waitqueues where the bucket discipline is to maintain all
857  * waiters on the same queue and wake all when any of the pages
858  * become available, and for the woken contexts to check to be
859  * sure the appropriate page became available, this saves space
860  * at a cost of "thundering herd" phenomena during rare hash
861  * collisions.
862  */
863 #define PAGE_WAIT_TABLE_BITS 8
864 #define PAGE_WAIT_TABLE_SIZE (1 << PAGE_WAIT_TABLE_BITS)
865 static wait_queue_head_t page_wait_table[PAGE_WAIT_TABLE_SIZE] __cacheline_aligned;
866
867 static wait_queue_head_t *page_waitqueue(struct page *page)
868 {
869         return &page_wait_table[hash_ptr(page, PAGE_WAIT_TABLE_BITS)];
870 }
871
872 void __init pagecache_init(void)
873 {
874         int i;
875
876         for (i = 0; i < PAGE_WAIT_TABLE_SIZE; i++)
877                 init_waitqueue_head(&page_wait_table[i]);
878
879         page_writeback_init();
880 }
881
882 /* This has the same layout as wait_bit_key - see fs/cachefiles/rdwr.c */
883 struct wait_page_key {
884         struct page *page;
885         int bit_nr;
886         int page_match;
887 };
888
889 struct wait_page_queue {
890         struct page *page;
891         int bit_nr;
892         wait_queue_entry_t wait;
893 };
894
895 static int wake_page_function(wait_queue_entry_t *wait, unsigned mode, int sync, void *arg)
896 {
897         struct wait_page_key *key = arg;
898         struct wait_page_queue *wait_page
899                 = container_of(wait, struct wait_page_queue, wait);
900
901         if (wait_page->page != key->page)
902                return 0;
903         key->page_match = 1;
904
905         if (wait_page->bit_nr != key->bit_nr)
906                 return 0;
907
908         /* Stop walking if it's locked */
909         if (test_bit(key->bit_nr, &key->page->flags))
910                 return -1;
911
912         return autoremove_wake_function(wait, mode, sync, key);
913 }
914
915 static void wake_up_page_bit(struct page *page, int bit_nr)
916 {
917         wait_queue_head_t *q = page_waitqueue(page);
918         struct wait_page_key key;
919         unsigned long flags;
920         wait_queue_entry_t bookmark;
921
922         key.page = page;
923         key.bit_nr = bit_nr;
924         key.page_match = 0;
925
926         bookmark.flags = 0;
927         bookmark.private = NULL;
928         bookmark.func = NULL;
929         INIT_LIST_HEAD(&bookmark.entry);
930
931         spin_lock_irqsave(&q->lock, flags);
932         __wake_up_locked_key_bookmark(q, TASK_NORMAL, &key, &bookmark);
933
934         while (bookmark.flags & WQ_FLAG_BOOKMARK) {
935                 /*
936                  * Take a breather from holding the lock,
937                  * allow pages that finish wake up asynchronously
938                  * to acquire the lock and remove themselves
939                  * from wait queue
940                  */
941                 spin_unlock_irqrestore(&q->lock, flags);
942                 cpu_relax();
943                 spin_lock_irqsave(&q->lock, flags);
944                 __wake_up_locked_key_bookmark(q, TASK_NORMAL, &key, &bookmark);
945         }
946
947         /*
948          * It is possible for other pages to have collided on the waitqueue
949          * hash, so in that case check for a page match. That prevents a long-
950          * term waiter
951          *
952          * It is still possible to miss a case here, when we woke page waiters
953          * and removed them from the waitqueue, but there are still other
954          * page waiters.
955          */
956         if (!waitqueue_active(q) || !key.page_match) {
957                 ClearPageWaiters(page);
958                 /*
959                  * It's possible to miss clearing Waiters here, when we woke
960                  * our page waiters, but the hashed waitqueue has waiters for
961                  * other pages on it.
962                  *
963                  * That's okay, it's a rare case. The next waker will clear it.
964                  */
965         }
966         spin_unlock_irqrestore(&q->lock, flags);
967 }
968
969 static void wake_up_page(struct page *page, int bit)
970 {
971         if (!PageWaiters(page))
972                 return;
973         wake_up_page_bit(page, bit);
974 }
975
976 static inline int wait_on_page_bit_common(wait_queue_head_t *q,
977                 struct page *page, int bit_nr, int state, bool lock)
978 {
979         struct wait_page_queue wait_page;
980         wait_queue_entry_t *wait = &wait_page.wait;
981         int ret = 0;
982
983         init_wait(wait);
984         wait->flags = lock ? WQ_FLAG_EXCLUSIVE : 0;
985         wait->func = wake_page_function;
986         wait_page.page = page;
987         wait_page.bit_nr = bit_nr;
988
989         for (;;) {
990                 spin_lock_irq(&q->lock);
991
992                 if (likely(list_empty(&wait->entry))) {
993                         __add_wait_queue_entry_tail(q, wait);
994                         SetPageWaiters(page);
995                 }
996
997                 set_current_state(state);
998
999                 spin_unlock_irq(&q->lock);
1000
1001                 if (likely(test_bit(bit_nr, &page->flags))) {
1002                         io_schedule();
1003                 }
1004
1005                 if (lock) {
1006                         if (!test_and_set_bit_lock(bit_nr, &page->flags))
1007                                 break;
1008                 } else {
1009                         if (!test_bit(bit_nr, &page->flags))
1010                                 break;
1011                 }
1012
1013                 if (unlikely(signal_pending_state(state, current))) {
1014                         ret = -EINTR;
1015                         break;
1016                 }
1017         }
1018
1019         finish_wait(q, wait);
1020
1021         /*
1022          * A signal could leave PageWaiters set. Clearing it here if
1023          * !waitqueue_active would be possible (by open-coding finish_wait),
1024          * but still fail to catch it in the case of wait hash collision. We
1025          * already can fail to clear wait hash collision cases, so don't
1026          * bother with signals either.
1027          */
1028
1029         return ret;
1030 }
1031
1032 void wait_on_page_bit(struct page *page, int bit_nr)
1033 {
1034         wait_queue_head_t *q = page_waitqueue(page);
1035         wait_on_page_bit_common(q, page, bit_nr, TASK_UNINTERRUPTIBLE, false);
1036 }
1037 EXPORT_SYMBOL(wait_on_page_bit);
1038
1039 int wait_on_page_bit_killable(struct page *page, int bit_nr)
1040 {
1041         wait_queue_head_t *q = page_waitqueue(page);
1042         return wait_on_page_bit_common(q, page, bit_nr, TASK_KILLABLE, false);
1043 }
1044
1045 /**
1046  * add_page_wait_queue - Add an arbitrary waiter to a page's wait queue
1047  * @page: Page defining the wait queue of interest
1048  * @waiter: Waiter to add to the queue
1049  *
1050  * Add an arbitrary @waiter to the wait queue for the nominated @page.
1051  */
1052 void add_page_wait_queue(struct page *page, wait_queue_entry_t *waiter)
1053 {
1054         wait_queue_head_t *q = page_waitqueue(page);
1055         unsigned long flags;
1056
1057         spin_lock_irqsave(&q->lock, flags);
1058         __add_wait_queue_entry_tail(q, waiter);
1059         SetPageWaiters(page);
1060         spin_unlock_irqrestore(&q->lock, flags);
1061 }
1062 EXPORT_SYMBOL_GPL(add_page_wait_queue);
1063
1064 #ifndef clear_bit_unlock_is_negative_byte
1065
1066 /*
1067  * PG_waiters is the high bit in the same byte as PG_lock.
1068  *
1069  * On x86 (and on many other architectures), we can clear PG_lock and
1070  * test the sign bit at the same time. But if the architecture does
1071  * not support that special operation, we just do this all by hand
1072  * instead.
1073  *
1074  * The read of PG_waiters has to be after (or concurrently with) PG_locked
1075  * being cleared, but a memory barrier should be unneccssary since it is
1076  * in the same byte as PG_locked.
1077  */
1078 static inline bool clear_bit_unlock_is_negative_byte(long nr, volatile void *mem)
1079 {
1080         clear_bit_unlock(nr, mem);
1081         /* smp_mb__after_atomic(); */
1082         return test_bit(PG_waiters, mem);
1083 }
1084
1085 #endif
1086
1087 /**
1088  * unlock_page - unlock a locked page
1089  * @page: the page
1090  *
1091  * Unlocks the page and wakes up sleepers in ___wait_on_page_locked().
1092  * Also wakes sleepers in wait_on_page_writeback() because the wakeup
1093  * mechanism between PageLocked pages and PageWriteback pages is shared.
1094  * But that's OK - sleepers in wait_on_page_writeback() just go back to sleep.
1095  *
1096  * Note that this depends on PG_waiters being the sign bit in the byte
1097  * that contains PG_locked - thus the BUILD_BUG_ON(). That allows us to
1098  * clear the PG_locked bit and test PG_waiters at the same time fairly
1099  * portably (architectures that do LL/SC can test any bit, while x86 can
1100  * test the sign bit).
1101  */
1102 void unlock_page(struct page *page)
1103 {
1104         BUILD_BUG_ON(PG_waiters != 7);
1105         page = compound_head(page);
1106         VM_BUG_ON_PAGE(!PageLocked(page), page);
1107         if (clear_bit_unlock_is_negative_byte(PG_locked, &page->flags))
1108                 wake_up_page_bit(page, PG_locked);
1109 }
1110 EXPORT_SYMBOL(unlock_page);
1111
1112 /**
1113  * end_page_writeback - end writeback against a page
1114  * @page: the page
1115  */
1116 void end_page_writeback(struct page *page)
1117 {
1118         /*
1119          * TestClearPageReclaim could be used here but it is an atomic
1120          * operation and overkill in this particular case. Failing to
1121          * shuffle a page marked for immediate reclaim is too mild to
1122          * justify taking an atomic operation penalty at the end of
1123          * ever page writeback.
1124          */
1125         if (PageReclaim(page)) {
1126                 ClearPageReclaim(page);
1127                 rotate_reclaimable_page(page);
1128         }
1129
1130         if (!test_clear_page_writeback(page))
1131                 BUG();
1132
1133         smp_mb__after_atomic();
1134         wake_up_page(page, PG_writeback);
1135 }
1136 EXPORT_SYMBOL(end_page_writeback);
1137
1138 /*
1139  * After completing I/O on a page, call this routine to update the page
1140  * flags appropriately
1141  */
1142 void page_endio(struct page *page, bool is_write, int err)
1143 {
1144         if (!is_write) {
1145                 if (!err) {
1146                         SetPageUptodate(page);
1147                 } else {
1148                         ClearPageUptodate(page);
1149                         SetPageError(page);
1150                 }
1151                 unlock_page(page);
1152         } else {
1153                 if (err) {
1154                         struct address_space *mapping;
1155
1156                         SetPageError(page);
1157                         mapping = page_mapping(page);
1158                         if (mapping)
1159                                 mapping_set_error(mapping, err);
1160                 }
1161                 end_page_writeback(page);
1162         }
1163 }
1164 EXPORT_SYMBOL_GPL(page_endio);
1165
1166 /**
1167  * __lock_page - get a lock on the page, assuming we need to sleep to get it
1168  * @__page: the page to lock
1169  */
1170 void __lock_page(struct page *__page)
1171 {
1172         struct page *page = compound_head(__page);
1173         wait_queue_head_t *q = page_waitqueue(page);
1174         wait_on_page_bit_common(q, page, PG_locked, TASK_UNINTERRUPTIBLE, true);
1175 }
1176 EXPORT_SYMBOL(__lock_page);
1177
1178 int __lock_page_killable(struct page *__page)
1179 {
1180         struct page *page = compound_head(__page);
1181         wait_queue_head_t *q = page_waitqueue(page);
1182         return wait_on_page_bit_common(q, page, PG_locked, TASK_KILLABLE, true);
1183 }
1184 EXPORT_SYMBOL_GPL(__lock_page_killable);
1185
1186 /*
1187  * Return values:
1188  * 1 - page is locked; mmap_sem is still held.
1189  * 0 - page is not locked.
1190  *     mmap_sem has been released (up_read()), unless flags had both
1191  *     FAULT_FLAG_ALLOW_RETRY and FAULT_FLAG_RETRY_NOWAIT set, in
1192  *     which case mmap_sem is still held.
1193  *
1194  * If neither ALLOW_RETRY nor KILLABLE are set, will always return 1
1195  * with the page locked and the mmap_sem unperturbed.
1196  */
1197 int __lock_page_or_retry(struct page *page, struct mm_struct *mm,
1198                          unsigned int flags)
1199 {
1200         if (flags & FAULT_FLAG_ALLOW_RETRY) {
1201                 /*
1202                  * CAUTION! In this case, mmap_sem is not released
1203                  * even though return 0.
1204                  */
1205                 if (flags & FAULT_FLAG_RETRY_NOWAIT)
1206                         return 0;
1207
1208                 up_read(&mm->mmap_sem);
1209                 if (flags & FAULT_FLAG_KILLABLE)
1210                         wait_on_page_locked_killable(page);
1211                 else
1212                         wait_on_page_locked(page);
1213                 return 0;
1214         } else {
1215                 if (flags & FAULT_FLAG_KILLABLE) {
1216                         int ret;
1217
1218                         ret = __lock_page_killable(page);
1219                         if (ret) {
1220                                 up_read(&mm->mmap_sem);
1221                                 return 0;
1222                         }
1223                 } else
1224                         __lock_page(page);
1225                 return 1;
1226         }
1227 }
1228
1229 /**
1230  * page_cache_next_hole - find the next hole (not-present entry)
1231  * @mapping: mapping
1232  * @index: index
1233  * @max_scan: maximum range to search
1234  *
1235  * Search the set [index, min(index+max_scan-1, MAX_INDEX)] for the
1236  * lowest indexed hole.
1237  *
1238  * Returns: the index of the hole if found, otherwise returns an index
1239  * outside of the set specified (in which case 'return - index >=
1240  * max_scan' will be true). In rare cases of index wrap-around, 0 will
1241  * be returned.
1242  *
1243  * page_cache_next_hole may be called under rcu_read_lock. However,
1244  * like radix_tree_gang_lookup, this will not atomically search a
1245  * snapshot of the tree at a single point in time. For example, if a
1246  * hole is created at index 5, then subsequently a hole is created at
1247  * index 10, page_cache_next_hole covering both indexes may return 10
1248  * if called under rcu_read_lock.
1249  */
1250 pgoff_t page_cache_next_hole(struct address_space *mapping,
1251                              pgoff_t index, unsigned long max_scan)
1252 {
1253         unsigned long i;
1254
1255         for (i = 0; i < max_scan; i++) {
1256                 struct page *page;
1257
1258                 page = radix_tree_lookup(&mapping->page_tree, index);
1259                 if (!page || radix_tree_exceptional_entry(page))
1260                         break;
1261                 index++;
1262                 if (index == 0)
1263                         break;
1264         }
1265
1266         return index;
1267 }
1268 EXPORT_SYMBOL(page_cache_next_hole);
1269
1270 /**
1271  * page_cache_prev_hole - find the prev hole (not-present entry)
1272  * @mapping: mapping
1273  * @index: index
1274  * @max_scan: maximum range to search
1275  *
1276  * Search backwards in the range [max(index-max_scan+1, 0), index] for
1277  * the first hole.
1278  *
1279  * Returns: the index of the hole if found, otherwise returns an index
1280  * outside of the set specified (in which case 'index - return >=
1281  * max_scan' will be true). In rare cases of wrap-around, ULONG_MAX
1282  * will be returned.
1283  *
1284  * page_cache_prev_hole may be called under rcu_read_lock. However,
1285  * like radix_tree_gang_lookup, this will not atomically search a
1286  * snapshot of the tree at a single point in time. For example, if a
1287  * hole is created at index 10, then subsequently a hole is created at
1288  * index 5, page_cache_prev_hole covering both indexes may return 5 if
1289  * called under rcu_read_lock.
1290  */
1291 pgoff_t page_cache_prev_hole(struct address_space *mapping,
1292                              pgoff_t index, unsigned long max_scan)
1293 {
1294         unsigned long i;
1295
1296         for (i = 0; i < max_scan; i++) {
1297                 struct page *page;
1298
1299                 page = radix_tree_lookup(&mapping->page_tree, index);
1300                 if (!page || radix_tree_exceptional_entry(page))
1301                         break;
1302                 index--;
1303                 if (index == ULONG_MAX)
1304                         break;
1305         }
1306
1307         return index;
1308 }
1309 EXPORT_SYMBOL(page_cache_prev_hole);
1310
1311 /**
1312  * find_get_entry - find and get a page cache entry
1313  * @mapping: the address_space to search
1314  * @offset: the page cache index
1315  *
1316  * Looks up the page cache slot at @mapping & @offset.  If there is a
1317  * page cache page, it is returned with an increased refcount.
1318  *
1319  * If the slot holds a shadow entry of a previously evicted page, or a
1320  * swap entry from shmem/tmpfs, it is returned.
1321  *
1322  * Otherwise, %NULL is returned.
1323  */
1324 struct page *find_get_entry(struct address_space *mapping, pgoff_t offset)
1325 {
1326         void **pagep;
1327         struct page *head, *page;
1328
1329         rcu_read_lock();
1330 repeat:
1331         page = NULL;
1332         pagep = radix_tree_lookup_slot(&mapping->page_tree, offset);
1333         if (pagep) {
1334                 page = radix_tree_deref_slot(pagep);
1335                 if (unlikely(!page))
1336                         goto out;
1337                 if (radix_tree_exception(page)) {
1338                         if (radix_tree_deref_retry(page))
1339                                 goto repeat;
1340                         /*
1341                          * A shadow entry of a recently evicted page,
1342                          * or a swap entry from shmem/tmpfs.  Return
1343                          * it without attempting to raise page count.
1344                          */
1345                         goto out;
1346                 }
1347
1348                 head = compound_head(page);
1349                 if (!page_cache_get_speculative(head))
1350                         goto repeat;
1351
1352                 /* The page was split under us? */
1353                 if (compound_head(page) != head) {
1354                         put_page(head);
1355                         goto repeat;
1356                 }
1357
1358                 /*
1359                  * Has the page moved?
1360                  * This is part of the lockless pagecache protocol. See
1361                  * include/linux/pagemap.h for details.
1362                  */
1363                 if (unlikely(page != *pagep)) {
1364                         put_page(head);
1365                         goto repeat;
1366                 }
1367         }
1368 out:
1369         rcu_read_unlock();
1370
1371         return page;
1372 }
1373 EXPORT_SYMBOL(find_get_entry);
1374
1375 /**
1376  * find_lock_entry - locate, pin and lock a page cache entry
1377  * @mapping: the address_space to search
1378  * @offset: the page cache index
1379  *
1380  * Looks up the page cache slot at @mapping & @offset.  If there is a
1381  * page cache page, it is returned locked and with an increased
1382  * refcount.
1383  *
1384  * If the slot holds a shadow entry of a previously evicted page, or a
1385  * swap entry from shmem/tmpfs, it is returned.
1386  *
1387  * Otherwise, %NULL is returned.
1388  *
1389  * find_lock_entry() may sleep.
1390  */
1391 struct page *find_lock_entry(struct address_space *mapping, pgoff_t offset)
1392 {
1393         struct page *page;
1394
1395 repeat:
1396         page = find_get_entry(mapping, offset);
1397         if (page && !radix_tree_exception(page)) {
1398                 lock_page(page);
1399                 /* Has the page been truncated? */
1400                 if (unlikely(page_mapping(page) != mapping)) {
1401                         unlock_page(page);
1402                         put_page(page);
1403                         goto repeat;
1404                 }
1405                 VM_BUG_ON_PAGE(page_to_pgoff(page) != offset, page);
1406         }
1407         return page;
1408 }
1409 EXPORT_SYMBOL(find_lock_entry);
1410
1411 /**
1412  * pagecache_get_page - find and get a page reference
1413  * @mapping: the address_space to search
1414  * @offset: the page index
1415  * @fgp_flags: PCG flags
1416  * @gfp_mask: gfp mask to use for the page cache data page allocation
1417  *
1418  * Looks up the page cache slot at @mapping & @offset.
1419  *
1420  * PCG flags modify how the page is returned.
1421  *
1422  * @fgp_flags can be:
1423  *
1424  * - FGP_ACCESSED: the page will be marked accessed
1425  * - FGP_LOCK: Page is return locked
1426  * - FGP_CREAT: If page is not present then a new page is allocated using
1427  *   @gfp_mask and added to the page cache and the VM's LRU
1428  *   list. The page is returned locked and with an increased
1429  *   refcount. Otherwise, NULL is returned.
1430  *
1431  * If FGP_LOCK or FGP_CREAT are specified then the function may sleep even
1432  * if the GFP flags specified for FGP_CREAT are atomic.
1433  *
1434  * If there is a page cache page, it is returned with an increased refcount.
1435  */
1436 struct page *pagecache_get_page(struct address_space *mapping, pgoff_t offset,
1437         int fgp_flags, gfp_t gfp_mask)
1438 {
1439         struct page *page;
1440
1441 repeat:
1442         page = find_get_entry(mapping, offset);
1443         if (radix_tree_exceptional_entry(page))
1444                 page = NULL;
1445         if (!page)
1446                 goto no_page;
1447
1448         if (fgp_flags & FGP_LOCK) {
1449                 if (fgp_flags & FGP_NOWAIT) {
1450                         if (!trylock_page(page)) {
1451                                 put_page(page);
1452                                 return NULL;
1453                         }
1454                 } else {
1455                         lock_page(page);
1456                 }
1457
1458                 /* Has the page been truncated? */
1459                 if (unlikely(page->mapping != mapping)) {
1460                         unlock_page(page);
1461                         put_page(page);
1462                         goto repeat;
1463                 }
1464                 VM_BUG_ON_PAGE(page->index != offset, page);
1465         }
1466
1467         if (page && (fgp_flags & FGP_ACCESSED))
1468                 mark_page_accessed(page);
1469
1470 no_page:
1471         if (!page && (fgp_flags & FGP_CREAT)) {
1472                 int err;
1473                 if ((fgp_flags & FGP_WRITE) && mapping_cap_account_dirty(mapping))
1474                         gfp_mask |= __GFP_WRITE;
1475                 if (fgp_flags & FGP_NOFS)
1476                         gfp_mask &= ~__GFP_FS;
1477
1478                 page = __page_cache_alloc(gfp_mask);
1479                 if (!page)
1480                         return NULL;
1481
1482                 if (WARN_ON_ONCE(!(fgp_flags & FGP_LOCK)))
1483                         fgp_flags |= FGP_LOCK;
1484
1485                 /* Init accessed so avoid atomic mark_page_accessed later */
1486                 if (fgp_flags & FGP_ACCESSED)
1487                         __SetPageReferenced(page);
1488
1489                 err = add_to_page_cache_lru(page, mapping, offset,
1490                                 gfp_mask & GFP_RECLAIM_MASK);
1491                 if (unlikely(err)) {
1492                         put_page(page);
1493                         page = NULL;
1494                         if (err == -EEXIST)
1495                                 goto repeat;
1496                 }
1497         }
1498
1499         return page;
1500 }
1501 EXPORT_SYMBOL(pagecache_get_page);
1502
1503 /**
1504  * find_get_entries - gang pagecache lookup
1505  * @mapping:    The address_space to search
1506  * @start:      The starting page cache index
1507  * @nr_entries: The maximum number of entries
1508  * @entries:    Where the resulting entries are placed
1509  * @indices:    The cache indices corresponding to the entries in @entries
1510  *
1511  * find_get_entries() will search for and return a group of up to
1512  * @nr_entries entries in the mapping.  The entries are placed at
1513  * @entries.  find_get_entries() takes a reference against any actual
1514  * pages it returns.
1515  *
1516  * The search returns a group of mapping-contiguous page cache entries
1517  * with ascending indexes.  There may be holes in the indices due to
1518  * not-present pages.
1519  *
1520  * Any shadow entries of evicted pages, or swap entries from
1521  * shmem/tmpfs, are included in the returned array.
1522  *
1523  * find_get_entries() returns the number of pages and shadow entries
1524  * which were found.
1525  */
1526 unsigned find_get_entries(struct address_space *mapping,
1527                           pgoff_t start, unsigned int nr_entries,
1528                           struct page **entries, pgoff_t *indices)
1529 {
1530         void **slot;
1531         unsigned int ret = 0;
1532         struct radix_tree_iter iter;
1533
1534         if (!nr_entries)
1535                 return 0;
1536
1537         rcu_read_lock();
1538         radix_tree_for_each_slot(slot, &mapping->page_tree, &iter, start) {
1539                 struct page *head, *page;
1540 repeat:
1541                 page = radix_tree_deref_slot(slot);
1542                 if (unlikely(!page))
1543                         continue;
1544                 if (radix_tree_exception(page)) {
1545                         if (radix_tree_deref_retry(page)) {
1546                                 slot = radix_tree_iter_retry(&iter);
1547                                 continue;
1548                         }
1549                         /*
1550                          * A shadow entry of a recently evicted page, a swap
1551                          * entry from shmem/tmpfs or a DAX entry.  Return it
1552                          * without attempting to raise page count.
1553                          */
1554                         goto export;
1555                 }
1556
1557                 head = compound_head(page);
1558                 if (!page_cache_get_speculative(head))
1559                         goto repeat;
1560
1561                 /* The page was split under us? */
1562                 if (compound_head(page) != head) {
1563                         put_page(head);
1564                         goto repeat;
1565                 }
1566
1567                 /* Has the page moved? */
1568                 if (unlikely(page != *slot)) {
1569                         put_page(head);
1570                         goto repeat;
1571                 }
1572 export:
1573                 indices[ret] = iter.index;
1574                 entries[ret] = page;
1575                 if (++ret == nr_entries)
1576                         break;
1577         }
1578         rcu_read_unlock();
1579         return ret;
1580 }
1581
1582 /**
1583  * find_get_pages_range - gang pagecache lookup
1584  * @mapping:    The address_space to search
1585  * @start:      The starting page index
1586  * @end:        The final page index (inclusive)
1587  * @nr_pages:   The maximum number of pages
1588  * @pages:      Where the resulting pages are placed
1589  *
1590  * find_get_pages_range() will search for and return a group of up to @nr_pages
1591  * pages in the mapping starting at index @start and up to index @end
1592  * (inclusive).  The pages are placed at @pages.  find_get_pages_range() takes
1593  * a reference against the returned pages.
1594  *
1595  * The search returns a group of mapping-contiguous pages with ascending
1596  * indexes.  There may be holes in the indices due to not-present pages.
1597  * We also update @start to index the next page for the traversal.
1598  *
1599  * find_get_pages_range() returns the number of pages which were found. If this
1600  * number is smaller than @nr_pages, the end of specified range has been
1601  * reached.
1602  */
1603 unsigned find_get_pages_range(struct address_space *mapping, pgoff_t *start,
1604                               pgoff_t end, unsigned int nr_pages,
1605                               struct page **pages)
1606 {
1607         struct radix_tree_iter iter;
1608         void **slot;
1609         unsigned ret = 0;
1610
1611         if (unlikely(!nr_pages))
1612                 return 0;
1613
1614         rcu_read_lock();
1615         radix_tree_for_each_slot(slot, &mapping->page_tree, &iter, *start) {
1616                 struct page *head, *page;
1617
1618                 if (iter.index > end)
1619                         break;
1620 repeat:
1621                 page = radix_tree_deref_slot(slot);
1622                 if (unlikely(!page))
1623                         continue;
1624
1625                 if (radix_tree_exception(page)) {
1626                         if (radix_tree_deref_retry(page)) {
1627                                 slot = radix_tree_iter_retry(&iter);
1628                                 continue;
1629                         }
1630                         /*
1631                          * A shadow entry of a recently evicted page,
1632                          * or a swap entry from shmem/tmpfs.  Skip
1633                          * over it.
1634                          */
1635                         continue;
1636                 }
1637
1638                 head = compound_head(page);
1639                 if (!page_cache_get_speculative(head))
1640                         goto repeat;
1641
1642                 /* The page was split under us? */
1643                 if (compound_head(page) != head) {
1644                         put_page(head);
1645                         goto repeat;
1646                 }
1647
1648                 /* Has the page moved? */
1649                 if (unlikely(page != *slot)) {
1650                         put_page(head);
1651                         goto repeat;
1652                 }
1653
1654                 pages[ret] = page;
1655                 if (++ret == nr_pages) {
1656                         *start = pages[ret - 1]->index + 1;
1657                         goto out;
1658                 }
1659         }
1660
1661         /*
1662          * We come here when there is no page beyond @end. We take care to not
1663          * overflow the index @start as it confuses some of the callers. This
1664          * breaks the iteration when there is page at index -1 but that is
1665          * already broken anyway.
1666          */
1667         if (end == (pgoff_t)-1)
1668                 *start = (pgoff_t)-1;
1669         else
1670                 *start = end + 1;
1671 out:
1672         rcu_read_unlock();
1673
1674         return ret;
1675 }
1676
1677 /**
1678  * find_get_pages_contig - gang contiguous pagecache lookup
1679  * @mapping:    The address_space to search
1680  * @index:      The starting page index
1681  * @nr_pages:   The maximum number of pages
1682  * @pages:      Where the resulting pages are placed
1683  *
1684  * find_get_pages_contig() works exactly like find_get_pages(), except
1685  * that the returned number of pages are guaranteed to be contiguous.
1686  *
1687  * find_get_pages_contig() returns the number of pages which were found.
1688  */
1689 unsigned find_get_pages_contig(struct address_space *mapping, pgoff_t index,
1690                                unsigned int nr_pages, struct page **pages)
1691 {
1692         struct radix_tree_iter iter;
1693         void **slot;
1694         unsigned int ret = 0;
1695
1696         if (unlikely(!nr_pages))
1697                 return 0;
1698
1699         rcu_read_lock();
1700         radix_tree_for_each_contig(slot, &mapping->page_tree, &iter, index) {
1701                 struct page *head, *page;
1702 repeat:
1703                 page = radix_tree_deref_slot(slot);
1704                 /* The hole, there no reason to continue */
1705                 if (unlikely(!page))
1706                         break;
1707
1708                 if (radix_tree_exception(page)) {
1709                         if (radix_tree_deref_retry(page)) {
1710                                 slot = radix_tree_iter_retry(&iter);
1711                                 continue;
1712                         }
1713                         /*
1714                          * A shadow entry of a recently evicted page,
1715                          * or a swap entry from shmem/tmpfs.  Stop
1716                          * looking for contiguous pages.
1717                          */
1718                         break;
1719                 }
1720
1721                 head = compound_head(page);
1722                 if (!page_cache_get_speculative(head))
1723                         goto repeat;
1724
1725                 /* The page was split under us? */
1726                 if (compound_head(page) != head) {
1727                         put_page(head);
1728                         goto repeat;
1729                 }
1730
1731                 /* Has the page moved? */
1732                 if (unlikely(page != *slot)) {
1733                         put_page(head);
1734                         goto repeat;
1735                 }
1736
1737                 /*
1738                  * must check mapping and index after taking the ref.
1739                  * otherwise we can get both false positives and false
1740                  * negatives, which is just confusing to the caller.
1741                  */
1742                 if (page->mapping == NULL || page_to_pgoff(page) != iter.index) {
1743                         put_page(page);
1744                         break;
1745                 }
1746
1747                 pages[ret] = page;
1748                 if (++ret == nr_pages)
1749                         break;
1750         }
1751         rcu_read_unlock();
1752         return ret;
1753 }
1754 EXPORT_SYMBOL(find_get_pages_contig);
1755
1756 /**
1757  * find_get_pages_range_tag - find and return pages in given range matching @tag
1758  * @mapping:    the address_space to search
1759  * @index:      the starting page index
1760  * @end:        The final page index (inclusive)
1761  * @tag:        the tag index
1762  * @nr_pages:   the maximum number of pages
1763  * @pages:      where the resulting pages are placed
1764  *
1765  * Like find_get_pages, except we only return pages which are tagged with
1766  * @tag.   We update @index to index the next page for the traversal.
1767  */
1768 unsigned find_get_pages_range_tag(struct address_space *mapping, pgoff_t *index,
1769                         pgoff_t end, int tag, unsigned int nr_pages,
1770                         struct page **pages)
1771 {
1772         struct radix_tree_iter iter;
1773         void **slot;
1774         unsigned ret = 0;
1775
1776         if (unlikely(!nr_pages))
1777                 return 0;
1778
1779         rcu_read_lock();
1780         radix_tree_for_each_tagged(slot, &mapping->page_tree,
1781                                    &iter, *index, tag) {
1782                 struct page *head, *page;
1783
1784                 if (iter.index > end)
1785                         break;
1786 repeat:
1787                 page = radix_tree_deref_slot(slot);
1788                 if (unlikely(!page))
1789                         continue;
1790
1791                 if (radix_tree_exception(page)) {
1792                         if (radix_tree_deref_retry(page)) {
1793                                 slot = radix_tree_iter_retry(&iter);
1794                                 continue;
1795                         }
1796                         /*
1797                          * A shadow entry of a recently evicted page.
1798                          *
1799                          * Those entries should never be tagged, but
1800                          * this tree walk is lockless and the tags are
1801                          * looked up in bulk, one radix tree node at a
1802                          * time, so there is a sizable window for page
1803                          * reclaim to evict a page we saw tagged.
1804                          *
1805                          * Skip over it.
1806                          */
1807                         continue;
1808                 }
1809
1810                 head = compound_head(page);
1811                 if (!page_cache_get_speculative(head))
1812                         goto repeat;
1813
1814                 /* The page was split under us? */
1815                 if (compound_head(page) != head) {
1816                         put_page(head);
1817                         goto repeat;
1818                 }
1819
1820                 /* Has the page moved? */
1821                 if (unlikely(page != *slot)) {
1822                         put_page(head);
1823                         goto repeat;
1824                 }
1825
1826                 pages[ret] = page;
1827                 if (++ret == nr_pages) {
1828                         *index = pages[ret - 1]->index + 1;
1829                         goto out;
1830                 }
1831         }
1832
1833         /*
1834          * We come here when we got at @end. We take care to not overflow the
1835          * index @index as it confuses some of the callers. This breaks the
1836          * iteration when there is page at index -1 but that is already broken
1837          * anyway.
1838          */
1839         if (end == (pgoff_t)-1)
1840                 *index = (pgoff_t)-1;
1841         else
1842                 *index = end + 1;
1843 out:
1844         rcu_read_unlock();
1845
1846         return ret;
1847 }
1848 EXPORT_SYMBOL(find_get_pages_range_tag);
1849
1850 /**
1851  * find_get_entries_tag - find and return entries that match @tag
1852  * @mapping:    the address_space to search
1853  * @start:      the starting page cache index
1854  * @tag:        the tag index
1855  * @nr_entries: the maximum number of entries
1856  * @entries:    where the resulting entries are placed
1857  * @indices:    the cache indices corresponding to the entries in @entries
1858  *
1859  * Like find_get_entries, except we only return entries which are tagged with
1860  * @tag.
1861  */
1862 unsigned find_get_entries_tag(struct address_space *mapping, pgoff_t start,
1863                         int tag, unsigned int nr_entries,
1864                         struct page **entries, pgoff_t *indices)
1865 {
1866         void **slot;
1867         unsigned int ret = 0;
1868         struct radix_tree_iter iter;
1869
1870         if (!nr_entries)
1871                 return 0;
1872
1873         rcu_read_lock();
1874         radix_tree_for_each_tagged(slot, &mapping->page_tree,
1875                                    &iter, start, tag) {
1876                 struct page *head, *page;
1877 repeat:
1878                 page = radix_tree_deref_slot(slot);
1879                 if (unlikely(!page))
1880                         continue;
1881                 if (radix_tree_exception(page)) {
1882                         if (radix_tree_deref_retry(page)) {
1883                                 slot = radix_tree_iter_retry(&iter);
1884                                 continue;
1885                         }
1886
1887                         /*
1888                          * A shadow entry of a recently evicted page, a swap
1889                          * entry from shmem/tmpfs or a DAX entry.  Return it
1890                          * without attempting to raise page count.
1891                          */
1892                         goto export;
1893                 }
1894
1895                 head = compound_head(page);
1896                 if (!page_cache_get_speculative(head))
1897                         goto repeat;
1898
1899                 /* The page was split under us? */
1900                 if (compound_head(page) != head) {
1901                         put_page(head);
1902                         goto repeat;
1903                 }
1904
1905                 /* Has the page moved? */
1906                 if (unlikely(page != *slot)) {
1907                         put_page(head);
1908                         goto repeat;
1909                 }
1910 export:
1911                 indices[ret] = iter.index;
1912                 entries[ret] = page;
1913                 if (++ret == nr_entries)
1914                         break;
1915         }
1916         rcu_read_unlock();
1917         return ret;
1918 }
1919 EXPORT_SYMBOL(find_get_entries_tag);
1920
1921 /*
1922  * CD/DVDs are error prone. When a medium error occurs, the driver may fail
1923  * a _large_ part of the i/o request. Imagine the worst scenario:
1924  *
1925  *      ---R__________________________________________B__________
1926  *         ^ reading here                             ^ bad block(assume 4k)
1927  *
1928  * read(R) => miss => readahead(R...B) => media error => frustrating retries
1929  * => failing the whole request => read(R) => read(R+1) =>
1930  * readahead(R+1...B+1) => bang => read(R+2) => read(R+3) =>
1931  * readahead(R+3...B+2) => bang => read(R+3) => read(R+4) =>
1932  * readahead(R+4...B+3) => bang => read(R+4) => read(R+5) => ......
1933  *
1934  * It is going insane. Fix it by quickly scaling down the readahead size.
1935  */
1936 static void shrink_readahead_size_eio(struct file *filp,
1937                                         struct file_ra_state *ra)
1938 {
1939         ra->ra_pages /= 4;
1940 }
1941
1942 /**
1943  * generic_file_buffered_read - generic file read routine
1944  * @iocb:       the iocb to read
1945  * @iter:       data destination
1946  * @written:    already copied
1947  *
1948  * This is a generic file read routine, and uses the
1949  * mapping->a_ops->readpage() function for the actual low-level stuff.
1950  *
1951  * This is really ugly. But the goto's actually try to clarify some
1952  * of the logic when it comes to error handling etc.
1953  */
1954 static ssize_t generic_file_buffered_read(struct kiocb *iocb,
1955                 struct iov_iter *iter, ssize_t written)
1956 {
1957         struct file *filp = iocb->ki_filp;
1958         struct address_space *mapping = filp->f_mapping;
1959         struct inode *inode = mapping->host;
1960         struct file_ra_state *ra = &filp->f_ra;
1961         loff_t *ppos = &iocb->ki_pos;
1962         pgoff_t index;
1963         pgoff_t last_index;
1964         pgoff_t prev_index;
1965         unsigned long offset;      /* offset into pagecache page */
1966         unsigned int prev_offset;
1967         int error = 0;
1968
1969         if (unlikely(*ppos >= inode->i_sb->s_maxbytes))
1970                 return 0;
1971         iov_iter_truncate(iter, inode->i_sb->s_maxbytes);
1972
1973         index = *ppos >> PAGE_SHIFT;
1974         prev_index = ra->prev_pos >> PAGE_SHIFT;
1975         prev_offset = ra->prev_pos & (PAGE_SIZE-1);
1976         last_index = (*ppos + iter->count + PAGE_SIZE-1) >> PAGE_SHIFT;
1977         offset = *ppos & ~PAGE_MASK;
1978
1979         for (;;) {
1980                 struct page *page;
1981                 pgoff_t end_index;
1982                 loff_t isize;
1983                 unsigned long nr, ret;
1984
1985                 cond_resched();
1986 find_page:
1987                 if (fatal_signal_pending(current)) {
1988                         error = -EINTR;
1989                         goto out;
1990                 }
1991
1992                 page = find_get_page(mapping, index);
1993                 if (!page) {
1994                         if (iocb->ki_flags & IOCB_NOWAIT)
1995                                 goto would_block;
1996                         page_cache_sync_readahead(mapping,
1997                                         ra, filp,
1998                                         index, last_index - index);
1999                         page = find_get_page(mapping, index);
2000                         if (unlikely(page == NULL))
2001                                 goto no_cached_page;
2002                 }
2003                 if (PageReadahead(page)) {
2004                         page_cache_async_readahead(mapping,
2005                                         ra, filp, page,
2006                                         index, last_index - index);
2007                 }
2008                 if (!PageUptodate(page)) {
2009                         if (iocb->ki_flags & IOCB_NOWAIT) {
2010                                 put_page(page);
2011                                 goto would_block;
2012                         }
2013
2014                         /*
2015                          * See comment in do_read_cache_page on why
2016                          * wait_on_page_locked is used to avoid unnecessarily
2017                          * serialisations and why it's safe.
2018                          */
2019                         error = wait_on_page_locked_killable(page);
2020                         if (unlikely(error))
2021                                 goto readpage_error;
2022                         if (PageUptodate(page))
2023                                 goto page_ok;
2024
2025                         if (inode->i_blkbits == PAGE_SHIFT ||
2026                                         !mapping->a_ops->is_partially_uptodate)
2027                                 goto page_not_up_to_date;
2028                         /* pipes can't handle partially uptodate pages */
2029                         if (unlikely(iter->type & ITER_PIPE))
2030                                 goto page_not_up_to_date;
2031                         if (!trylock_page(page))
2032                                 goto page_not_up_to_date;
2033                         /* Did it get truncated before we got the lock? */
2034                         if (!page->mapping)
2035                                 goto page_not_up_to_date_locked;
2036                         if (!mapping->a_ops->is_partially_uptodate(page,
2037                                                         offset, iter->count))
2038                                 goto page_not_up_to_date_locked;
2039                         unlock_page(page);
2040                 }
2041 page_ok:
2042                 /*
2043                  * i_size must be checked after we know the page is Uptodate.
2044                  *
2045                  * Checking i_size after the check allows us to calculate
2046                  * the correct value for "nr", which means the zero-filled
2047                  * part of the page is not copied back to userspace (unless
2048                  * another truncate extends the file - this is desired though).
2049                  */
2050
2051                 isize = i_size_read(inode);
2052                 end_index = (isize - 1) >> PAGE_SHIFT;
2053                 if (unlikely(!isize || index > end_index)) {
2054                         put_page(page);
2055                         goto out;
2056                 }
2057
2058                 /* nr is the maximum number of bytes to copy from this page */
2059                 nr = PAGE_SIZE;
2060                 if (index == end_index) {
2061                         nr = ((isize - 1) & ~PAGE_MASK) + 1;
2062                         if (nr <= offset) {
2063                                 put_page(page);
2064                                 goto out;
2065                         }
2066                 }
2067                 nr = nr - offset;
2068
2069                 /* If users can be writing to this page using arbitrary
2070                  * virtual addresses, take care about potential aliasing
2071                  * before reading the page on the kernel side.
2072                  */
2073                 if (mapping_writably_mapped(mapping))
2074                         flush_dcache_page(page);
2075
2076                 /*
2077                  * When a sequential read accesses a page several times,
2078                  * only mark it as accessed the first time.
2079                  */
2080                 if (prev_index != index || offset != prev_offset)
2081                         mark_page_accessed(page);
2082                 prev_index = index;
2083
2084                 /*
2085                  * Ok, we have the page, and it's up-to-date, so
2086                  * now we can copy it to user space...
2087                  */
2088
2089                 ret = copy_page_to_iter(page, offset, nr, iter);
2090                 offset += ret;
2091                 index += offset >> PAGE_SHIFT;
2092                 offset &= ~PAGE_MASK;
2093                 prev_offset = offset;
2094
2095                 put_page(page);
2096                 written += ret;
2097                 if (!iov_iter_count(iter))
2098                         goto out;
2099                 if (ret < nr) {
2100                         error = -EFAULT;
2101                         goto out;
2102                 }
2103                 continue;
2104
2105 page_not_up_to_date:
2106                 /* Get exclusive access to the page ... */
2107                 error = lock_page_killable(page);
2108                 if (unlikely(error))
2109                         goto readpage_error;
2110
2111 page_not_up_to_date_locked:
2112                 /* Did it get truncated before we got the lock? */
2113                 if (!page->mapping) {
2114                         unlock_page(page);
2115                         put_page(page);
2116                         continue;
2117                 }
2118
2119                 /* Did somebody else fill it already? */
2120                 if (PageUptodate(page)) {
2121                         unlock_page(page);
2122                         goto page_ok;
2123                 }
2124
2125 readpage:
2126                 /*
2127                  * A previous I/O error may have been due to temporary
2128                  * failures, eg. multipath errors.
2129                  * PG_error will be set again if readpage fails.
2130                  */
2131                 ClearPageError(page);
2132                 /* Start the actual read. The read will unlock the page. */
2133                 error = mapping->a_ops->readpage(filp, page);
2134
2135                 if (unlikely(error)) {
2136                         if (error == AOP_TRUNCATED_PAGE) {
2137                                 put_page(page);
2138                                 error = 0;
2139                                 goto find_page;
2140                         }
2141                         goto readpage_error;
2142                 }
2143
2144                 if (!PageUptodate(page)) {
2145                         error = lock_page_killable(page);
2146                         if (unlikely(error))
2147                                 goto readpage_error;
2148                         if (!PageUptodate(page)) {
2149                                 if (page->mapping == NULL) {
2150                                         /*
2151                                          * invalidate_mapping_pages got it
2152                                          */
2153                                         unlock_page(page);
2154                                         put_page(page);
2155                                         goto find_page;
2156                                 }
2157                                 unlock_page(page);
2158                                 shrink_readahead_size_eio(filp, ra);
2159                                 error = -EIO;
2160                                 goto readpage_error;
2161                         }
2162                         unlock_page(page);
2163                 }
2164
2165                 goto page_ok;
2166
2167 readpage_error:
2168                 /* UHHUH! A synchronous read error occurred. Report it */
2169                 put_page(page);
2170                 goto out;
2171
2172 no_cached_page:
2173                 /*
2174                  * Ok, it wasn't cached, so we need to create a new
2175                  * page..
2176                  */
2177                 page = page_cache_alloc_cold(mapping);
2178                 if (!page) {
2179                         error = -ENOMEM;
2180                         goto out;
2181                 }
2182                 error = add_to_page_cache_lru(page, mapping, index,
2183                                 mapping_gfp_constraint(mapping, GFP_KERNEL));
2184                 if (error) {
2185                         put_page(page);
2186                         if (error == -EEXIST) {
2187                                 error = 0;
2188                                 goto find_page;
2189                         }
2190                         goto out;
2191                 }
2192                 goto readpage;
2193         }
2194
2195 would_block:
2196         error = -EAGAIN;
2197 out:
2198         ra->prev_pos = prev_index;
2199         ra->prev_pos <<= PAGE_SHIFT;
2200         ra->prev_pos |= prev_offset;
2201
2202         *ppos = ((loff_t)index << PAGE_SHIFT) + offset;
2203         file_accessed(filp);
2204         return written ? written : error;
2205 }
2206
2207 /**
2208  * generic_file_read_iter - generic filesystem read routine
2209  * @iocb:       kernel I/O control block
2210  * @iter:       destination for the data read
2211  *
2212  * This is the "read_iter()" routine for all filesystems
2213  * that can use the page cache directly.
2214  */
2215 ssize_t
2216 generic_file_read_iter(struct kiocb *iocb, struct iov_iter *iter)
2217 {
2218         size_t count = iov_iter_count(iter);
2219         ssize_t retval = 0;
2220
2221         if (!count)
2222                 goto out; /* skip atime */
2223
2224         if (iocb->ki_flags & IOCB_DIRECT) {
2225                 struct file *file = iocb->ki_filp;
2226                 struct address_space *mapping = file->f_mapping;
2227                 struct inode *inode = mapping->host;
2228                 loff_t size;
2229
2230                 size = i_size_read(inode);
2231                 if (iocb->ki_flags & IOCB_NOWAIT) {
2232                         if (filemap_range_has_page(mapping, iocb->ki_pos,
2233                                                    iocb->ki_pos + count - 1))
2234                                 return -EAGAIN;
2235                 } else {
2236                         retval = filemap_write_and_wait_range(mapping,
2237                                                 iocb->ki_pos,
2238                                                 iocb->ki_pos + count - 1);
2239                         if (retval < 0)
2240                                 goto out;
2241                 }
2242
2243                 file_accessed(file);
2244
2245                 retval = mapping->a_ops->direct_IO(iocb, iter);
2246                 if (retval >= 0) {
2247                         iocb->ki_pos += retval;
2248                         count -= retval;
2249                 }
2250                 iov_iter_revert(iter, count - iov_iter_count(iter));
2251
2252                 /*
2253                  * Btrfs can have a short DIO read if we encounter
2254                  * compressed extents, so if there was an error, or if
2255                  * we've already read everything we wanted to, or if
2256                  * there was a short read because we hit EOF, go ahead
2257                  * and return.  Otherwise fallthrough to buffered io for
2258                  * the rest of the read.  Buffered reads will not work for
2259                  * DAX files, so don't bother trying.
2260                  */
2261                 if (retval < 0 || !count || iocb->ki_pos >= size ||
2262                     IS_DAX(inode))
2263                         goto out;
2264         }
2265
2266         retval = generic_file_buffered_read(iocb, iter, retval);
2267 out:
2268         return retval;
2269 }
2270 EXPORT_SYMBOL(generic_file_read_iter);
2271
2272 #ifdef CONFIG_MMU
2273 /**
2274  * page_cache_read - adds requested page to the page cache if not already there
2275  * @file:       file to read
2276  * @offset:     page index
2277  * @gfp_mask:   memory allocation flags
2278  *
2279  * This adds the requested page to the page cache if it isn't already there,
2280  * and schedules an I/O to read in its contents from disk.
2281  */
2282 static int page_cache_read(struct file *file, pgoff_t offset, gfp_t gfp_mask)
2283 {
2284         struct address_space *mapping = file->f_mapping;
2285         struct page *page;
2286         int ret;
2287
2288         do {
2289                 page = __page_cache_alloc(gfp_mask|__GFP_COLD);
2290                 if (!page)
2291                         return -ENOMEM;
2292
2293                 ret = add_to_page_cache_lru(page, mapping, offset, gfp_mask & GFP_KERNEL);
2294                 if (ret == 0)
2295                         ret = mapping->a_ops->readpage(file, page);
2296                 else if (ret == -EEXIST)
2297                         ret = 0; /* losing race to add is OK */
2298
2299                 put_page(page);
2300
2301         } while (ret == AOP_TRUNCATED_PAGE);
2302
2303         return ret;
2304 }
2305
2306 #define MMAP_LOTSAMISS  (100)
2307
2308 /*
2309  * Synchronous readahead happens when we don't even find
2310  * a page in the page cache at all.
2311  */
2312 static void do_sync_mmap_readahead(struct vm_area_struct *vma,
2313                                    struct file_ra_state *ra,
2314                                    struct file *file,
2315                                    pgoff_t offset)
2316 {
2317         struct address_space *mapping = file->f_mapping;
2318
2319         /* If we don't want any read-ahead, don't bother */
2320         if (vma->vm_flags & VM_RAND_READ)
2321                 return;
2322         if (!ra->ra_pages)
2323                 return;
2324
2325         if (vma->vm_flags & VM_SEQ_READ) {
2326                 page_cache_sync_readahead(mapping, ra, file, offset,
2327                                           ra->ra_pages);
2328                 return;
2329         }
2330
2331         /* Avoid banging the cache line if not needed */
2332         if (ra->mmap_miss < MMAP_LOTSAMISS * 10)
2333                 ra->mmap_miss++;
2334
2335         /*
2336          * Do we miss much more than hit in this file? If so,
2337          * stop bothering with read-ahead. It will only hurt.
2338          */
2339         if (ra->mmap_miss > MMAP_LOTSAMISS)
2340                 return;
2341
2342         /*
2343          * mmap read-around
2344          */
2345         ra->start = max_t(long, 0, offset - ra->ra_pages / 2);
2346         ra->size = ra->ra_pages;
2347         ra->async_size = ra->ra_pages / 4;
2348         ra_submit(ra, mapping, file);
2349 }
2350
2351 /*
2352  * Asynchronous readahead happens when we find the page and PG_readahead,
2353  * so we want to possibly extend the readahead further..
2354  */
2355 static void do_async_mmap_readahead(struct vm_area_struct *vma,
2356                                     struct file_ra_state *ra,
2357                                     struct file *file,
2358                                     struct page *page,
2359                                     pgoff_t offset)
2360 {
2361         struct address_space *mapping = file->f_mapping;
2362
2363         /* If we don't want any read-ahead, don't bother */
2364         if (vma->vm_flags & VM_RAND_READ)
2365                 return;
2366         if (ra->mmap_miss > 0)
2367                 ra->mmap_miss--;
2368         if (PageReadahead(page))
2369                 page_cache_async_readahead(mapping, ra, file,
2370                                            page, offset, ra->ra_pages);
2371 }
2372
2373 /**
2374  * filemap_fault - read in file data for page fault handling
2375  * @vmf:        struct vm_fault containing details of the fault
2376  *
2377  * filemap_fault() is invoked via the vma operations vector for a
2378  * mapped memory region to read in file data during a page fault.
2379  *
2380  * The goto's are kind of ugly, but this streamlines the normal case of having
2381  * it in the page cache, and handles the special cases reasonably without
2382  * having a lot of duplicated code.
2383  *
2384  * vma->vm_mm->mmap_sem must be held on entry.
2385  *
2386  * If our return value has VM_FAULT_RETRY set, it's because
2387  * lock_page_or_retry() returned 0.
2388  * The mmap_sem has usually been released in this case.
2389  * See __lock_page_or_retry() for the exception.
2390  *
2391  * If our return value does not have VM_FAULT_RETRY set, the mmap_sem
2392  * has not been released.
2393  *
2394  * We never return with VM_FAULT_RETRY and a bit from VM_FAULT_ERROR set.
2395  */
2396 int filemap_fault(struct vm_fault *vmf)
2397 {
2398         int error;
2399         struct file *file = vmf->vma->vm_file;
2400         struct address_space *mapping = file->f_mapping;
2401         struct file_ra_state *ra = &file->f_ra;
2402         struct inode *inode = mapping->host;
2403         pgoff_t offset = vmf->pgoff;
2404         pgoff_t max_off;
2405         struct page *page;
2406         int ret = 0;
2407
2408         max_off = DIV_ROUND_UP(i_size_read(inode), PAGE_SIZE);
2409         if (unlikely(offset >= max_off))
2410                 return VM_FAULT_SIGBUS;
2411
2412         /*
2413          * Do we have something in the page cache already?
2414          */
2415         page = find_get_page(mapping, offset);
2416         if (likely(page) && !(vmf->flags & FAULT_FLAG_TRIED)) {
2417                 /*
2418                  * We found the page, so try async readahead before
2419                  * waiting for the lock.
2420                  */
2421                 do_async_mmap_readahead(vmf->vma, ra, file, page, offset);
2422         } else if (!page) {
2423                 /* No page in the page cache at all */
2424                 do_sync_mmap_readahead(vmf->vma, ra, file, offset);
2425                 count_vm_event(PGMAJFAULT);
2426                 count_memcg_event_mm(vmf->vma->vm_mm, PGMAJFAULT);
2427                 ret = VM_FAULT_MAJOR;
2428 retry_find:
2429                 page = find_get_page(mapping, offset);
2430                 if (!page)
2431                         goto no_cached_page;
2432         }
2433
2434         if (!lock_page_or_retry(page, vmf->vma->vm_mm, vmf->flags)) {
2435                 put_page(page);
2436                 return ret | VM_FAULT_RETRY;
2437         }
2438
2439         /* Did it get truncated? */
2440         if (unlikely(page->mapping != mapping)) {
2441                 unlock_page(page);
2442                 put_page(page);
2443                 goto retry_find;
2444         }
2445         VM_BUG_ON_PAGE(page->index != offset, page);
2446
2447         /*
2448          * We have a locked page in the page cache, now we need to check
2449          * that it's up-to-date. If not, it is going to be due to an error.
2450          */
2451         if (unlikely(!PageUptodate(page)))
2452                 goto page_not_uptodate;
2453
2454         /*
2455          * Found the page and have a reference on it.
2456          * We must recheck i_size under page lock.
2457          */
2458         max_off = DIV_ROUND_UP(i_size_read(inode), PAGE_SIZE);
2459         if (unlikely(offset >= max_off)) {
2460                 unlock_page(page);
2461                 put_page(page);
2462                 return VM_FAULT_SIGBUS;
2463         }
2464
2465         vmf->page = page;
2466         return ret | VM_FAULT_LOCKED;
2467
2468 no_cached_page:
2469         /*
2470          * We're only likely to ever get here if MADV_RANDOM is in
2471          * effect.
2472          */
2473         error = page_cache_read(file, offset, vmf->gfp_mask);
2474
2475         /*
2476          * The page we want has now been added to the page cache.
2477          * In the unlikely event that someone removed it in the
2478          * meantime, we'll just come back here and read it again.
2479          */
2480         if (error >= 0)
2481                 goto retry_find;
2482
2483         /*
2484          * An error return from page_cache_read can result if the
2485          * system is low on memory, or a problem occurs while trying
2486          * to schedule I/O.
2487          */
2488         if (error == -ENOMEM)
2489                 return VM_FAULT_OOM;
2490         return VM_FAULT_SIGBUS;
2491
2492 page_not_uptodate:
2493         /*
2494          * Umm, take care of errors if the page isn't up-to-date.
2495          * Try to re-read it _once_. We do this synchronously,
2496          * because there really aren't any performance issues here
2497          * and we need to check for errors.
2498          */
2499         ClearPageError(page);
2500         error = mapping->a_ops->readpage(file, page);
2501         if (!error) {
2502                 wait_on_page_locked(page);
2503                 if (!PageUptodate(page))
2504                         error = -EIO;
2505         }
2506         put_page(page);
2507
2508         if (!error || error == AOP_TRUNCATED_PAGE)
2509                 goto retry_find;
2510
2511         /* Things didn't work out. Return zero to tell the mm layer so. */
2512         shrink_readahead_size_eio(file, ra);
2513         return VM_FAULT_SIGBUS;
2514 }
2515 EXPORT_SYMBOL(filemap_fault);
2516
2517 void filemap_map_pages(struct vm_fault *vmf,
2518                 pgoff_t start_pgoff, pgoff_t end_pgoff)
2519 {
2520         struct radix_tree_iter iter;
2521         void **slot;
2522         struct file *file = vmf->vma->vm_file;
2523         struct address_space *mapping = file->f_mapping;
2524         pgoff_t last_pgoff = start_pgoff;
2525         unsigned long max_idx;
2526         struct page *head, *page;
2527
2528         rcu_read_lock();
2529         radix_tree_for_each_slot(slot, &mapping->page_tree, &iter,
2530                         start_pgoff) {
2531                 if (iter.index > end_pgoff)
2532                         break;
2533 repeat:
2534                 page = radix_tree_deref_slot(slot);
2535                 if (unlikely(!page))
2536                         goto next;
2537                 if (radix_tree_exception(page)) {
2538                         if (radix_tree_deref_retry(page)) {
2539                                 slot = radix_tree_iter_retry(&iter);
2540                                 continue;
2541                         }
2542                         goto next;
2543                 }
2544
2545                 head = compound_head(page);
2546                 if (!page_cache_get_speculative(head))
2547                         goto repeat;
2548
2549                 /* The page was split under us? */
2550                 if (compound_head(page) != head) {
2551                         put_page(head);
2552                         goto repeat;
2553                 }
2554
2555                 /* Has the page moved? */
2556                 if (unlikely(page != *slot)) {
2557                         put_page(head);
2558                         goto repeat;
2559                 }
2560
2561                 if (!PageUptodate(page) ||
2562                                 PageReadahead(page) ||
2563                                 PageHWPoison(page))
2564                         goto skip;
2565                 if (!trylock_page(page))
2566                         goto skip;
2567
2568                 if (page->mapping != mapping || !PageUptodate(page))
2569                         goto unlock;
2570
2571                 max_idx = DIV_ROUND_UP(i_size_read(mapping->host), PAGE_SIZE);
2572                 if (page->index >= max_idx)
2573                         goto unlock;
2574
2575                 if (file->f_ra.mmap_miss > 0)
2576                         file->f_ra.mmap_miss--;
2577
2578                 vmf->address += (iter.index - last_pgoff) << PAGE_SHIFT;
2579                 if (vmf->pte)
2580                         vmf->pte += iter.index - last_pgoff;
2581                 last_pgoff = iter.index;
2582                 if (alloc_set_pte(vmf, NULL, page))
2583                         goto unlock;
2584                 unlock_page(page);
2585                 goto next;
2586 unlock:
2587                 unlock_page(page);
2588 skip:
2589                 put_page(page);
2590 next:
2591                 /* Huge page is mapped? No need to proceed. */
2592                 if (pmd_trans_huge(*vmf->pmd))
2593                         break;
2594                 if (iter.index == end_pgoff)
2595                         break;
2596         }
2597         rcu_read_unlock();
2598 }
2599 EXPORT_SYMBOL(filemap_map_pages);
2600
2601 int filemap_page_mkwrite(struct vm_fault *vmf)
2602 {
2603         struct page *page = vmf->page;
2604         struct inode *inode = file_inode(vmf->vma->vm_file);
2605         int ret = VM_FAULT_LOCKED;
2606
2607         sb_start_pagefault(inode->i_sb);
2608         file_update_time(vmf->vma->vm_file);
2609         lock_page(page);
2610         if (page->mapping != inode->i_mapping) {
2611                 unlock_page(page);
2612                 ret = VM_FAULT_NOPAGE;
2613                 goto out;
2614         }
2615         /*
2616          * We mark the page dirty already here so that when freeze is in
2617          * progress, we are guaranteed that writeback during freezing will
2618          * see the dirty page and writeprotect it again.
2619          */
2620         set_page_dirty(page);
2621         wait_for_stable_page(page);
2622 out:
2623         sb_end_pagefault(inode->i_sb);
2624         return ret;
2625 }
2626 EXPORT_SYMBOL(filemap_page_mkwrite);
2627
2628 const struct vm_operations_struct generic_file_vm_ops = {
2629         .fault          = filemap_fault,
2630         .map_pages      = filemap_map_pages,
2631         .page_mkwrite   = filemap_page_mkwrite,
2632 };
2633
2634 /* This is used for a general mmap of a disk file */
2635
2636 int generic_file_mmap(struct file * file, struct vm_area_struct * vma)
2637 {
2638         struct address_space *mapping = file->f_mapping;
2639
2640         if (!mapping->a_ops->readpage)
2641                 return -ENOEXEC;
2642         file_accessed(file);
2643         vma->vm_ops = &generic_file_vm_ops;
2644         return 0;
2645 }
2646
2647 /*
2648  * This is for filesystems which do not implement ->writepage.
2649  */
2650 int generic_file_readonly_mmap(struct file *file, struct vm_area_struct *vma)
2651 {
2652         if ((vma->vm_flags & VM_SHARED) && (vma->vm_flags & VM_MAYWRITE))
2653                 return -EINVAL;
2654         return generic_file_mmap(file, vma);
2655 }
2656 #else
2657 int generic_file_mmap(struct file * file, struct vm_area_struct * vma)
2658 {
2659         return -ENOSYS;
2660 }
2661 int generic_file_readonly_mmap(struct file * file, struct vm_area_struct * vma)
2662 {
2663         return -ENOSYS;
2664 }
2665 #endif /* CONFIG_MMU */
2666
2667 EXPORT_SYMBOL(generic_file_mmap);
2668 EXPORT_SYMBOL(generic_file_readonly_mmap);
2669
2670 static struct page *wait_on_page_read(struct page *page)
2671 {
2672         if (!IS_ERR(page)) {
2673                 wait_on_page_locked(page);
2674                 if (!PageUptodate(page)) {
2675                         put_page(page);
2676                         page = ERR_PTR(-EIO);
2677                 }
2678         }
2679         return page;
2680 }
2681
2682 static struct page *do_read_cache_page(struct address_space *mapping,
2683                                 pgoff_t index,
2684                                 int (*filler)(void *, struct page *),
2685                                 void *data,
2686                                 gfp_t gfp)
2687 {
2688         struct page *page;
2689         int err;
2690 repeat:
2691         page = find_get_page(mapping, index);
2692         if (!page) {
2693                 page = __page_cache_alloc(gfp | __GFP_COLD);
2694                 if (!page)
2695                         return ERR_PTR(-ENOMEM);
2696                 err = add_to_page_cache_lru(page, mapping, index, gfp);
2697                 if (unlikely(err)) {
2698                         put_page(page);
2699                         if (err == -EEXIST)
2700                                 goto repeat;
2701                         /* Presumably ENOMEM for radix tree node */
2702                         return ERR_PTR(err);
2703                 }
2704
2705 filler:
2706                 err = filler(data, page);
2707                 if (err < 0) {
2708                         put_page(page);
2709                         return ERR_PTR(err);
2710                 }
2711
2712                 page = wait_on_page_read(page);
2713                 if (IS_ERR(page))
2714                         return page;
2715                 goto out;
2716         }
2717         if (PageUptodate(page))
2718                 goto out;
2719
2720         /*
2721          * Page is not up to date and may be locked due one of the following
2722          * case a: Page is being filled and the page lock is held
2723          * case b: Read/write error clearing the page uptodate status
2724          * case c: Truncation in progress (page locked)
2725          * case d: Reclaim in progress
2726          *
2727          * Case a, the page will be up to date when the page is unlocked.
2728          *    There is no need to serialise on the page lock here as the page
2729          *    is pinned so the lock gives no additional protection. Even if the
2730          *    the page is truncated, the data is still valid if PageUptodate as
2731          *    it's a race vs truncate race.
2732          * Case b, the page will not be up to date
2733          * Case c, the page may be truncated but in itself, the data may still
2734          *    be valid after IO completes as it's a read vs truncate race. The
2735          *    operation must restart if the page is not uptodate on unlock but
2736          *    otherwise serialising on page lock to stabilise the mapping gives
2737          *    no additional guarantees to the caller as the page lock is
2738          *    released before return.
2739          * Case d, similar to truncation. If reclaim holds the page lock, it
2740          *    will be a race with remove_mapping that determines if the mapping
2741          *    is valid on unlock but otherwise the data is valid and there is
2742          *    no need to serialise with page lock.
2743          *
2744          * As the page lock gives no additional guarantee, we optimistically
2745          * wait on the page to be unlocked and check if it's up to date and
2746          * use the page if it is. Otherwise, the page lock is required to
2747          * distinguish between the different cases. The motivation is that we
2748          * avoid spurious serialisations and wakeups when multiple processes
2749          * wait on the same page for IO to complete.
2750          */
2751         wait_on_page_locked(page);
2752         if (PageUptodate(page))
2753                 goto out;
2754
2755         /* Distinguish between all the cases under the safety of the lock */
2756         lock_page(page);
2757
2758         /* Case c or d, restart the operation */
2759         if (!page->mapping) {
2760                 unlock_page(page);
2761                 put_page(page);
2762                 goto repeat;
2763         }
2764
2765         /* Someone else locked and filled the page in a very small window */
2766         if (PageUptodate(page)) {
2767                 unlock_page(page);
2768                 goto out;
2769         }
2770         goto filler;
2771
2772 out:
2773         mark_page_accessed(page);
2774         return page;
2775 }
2776
2777 /**
2778  * read_cache_page - read into page cache, fill it if needed
2779  * @mapping:    the page's address_space
2780  * @index:      the page index
2781  * @filler:     function to perform the read
2782  * @data:       first arg to filler(data, page) function, often left as NULL
2783  *
2784  * Read into the page cache. If a page already exists, and PageUptodate() is
2785  * not set, try to fill the page and wait for it to become unlocked.
2786  *
2787  * If the page does not get brought uptodate, return -EIO.
2788  */
2789 struct page *read_cache_page(struct address_space *mapping,
2790                                 pgoff_t index,
2791                                 int (*filler)(void *, struct page *),
2792                                 void *data)
2793 {
2794         return do_read_cache_page(mapping, index, filler, data, mapping_gfp_mask(mapping));
2795 }
2796 EXPORT_SYMBOL(read_cache_page);
2797
2798 /**
2799  * read_cache_page_gfp - read into page cache, using specified page allocation flags.
2800  * @mapping:    the page's address_space
2801  * @index:      the page index
2802  * @gfp:        the page allocator flags to use if allocating
2803  *
2804  * This is the same as "read_mapping_page(mapping, index, NULL)", but with
2805  * any new page allocations done using the specified allocation flags.
2806  *
2807  * If the page does not get brought uptodate, return -EIO.
2808  */
2809 struct page *read_cache_page_gfp(struct address_space *mapping,
2810                                 pgoff_t index,
2811                                 gfp_t gfp)
2812 {
2813         filler_t *filler = (filler_t *)mapping->a_ops->readpage;
2814
2815         return do_read_cache_page(mapping, index, filler, NULL, gfp);
2816 }
2817 EXPORT_SYMBOL(read_cache_page_gfp);
2818
2819 /*
2820  * Performs necessary checks before doing a write
2821  *
2822  * Can adjust writing position or amount of bytes to write.
2823  * Returns appropriate error code that caller should return or
2824  * zero in case that write should be allowed.
2825  */
2826 inline ssize_t generic_write_checks(struct kiocb *iocb, struct iov_iter *from)
2827 {
2828         struct file *file = iocb->ki_filp;
2829         struct inode *inode = file->f_mapping->host;
2830         unsigned long limit = rlimit(RLIMIT_FSIZE);
2831         loff_t pos;
2832
2833         if (!iov_iter_count(from))
2834                 return 0;
2835
2836         /* FIXME: this is for backwards compatibility with 2.4 */
2837         if (iocb->ki_flags & IOCB_APPEND)
2838                 iocb->ki_pos = i_size_read(inode);
2839
2840         pos = iocb->ki_pos;
2841
2842         if ((iocb->ki_flags & IOCB_NOWAIT) && !(iocb->ki_flags & IOCB_DIRECT))
2843                 return -EINVAL;
2844
2845         if (limit != RLIM_INFINITY) {
2846                 if (iocb->ki_pos >= limit) {
2847                         send_sig(SIGXFSZ, current, 0);
2848                         return -EFBIG;
2849                 }
2850                 iov_iter_truncate(from, limit - (unsigned long)pos);
2851         }
2852
2853         /*
2854          * LFS rule
2855          */
2856         if (unlikely(pos + iov_iter_count(from) > MAX_NON_LFS &&
2857                                 !(file->f_flags & O_LARGEFILE))) {
2858                 if (pos >= MAX_NON_LFS)
2859                         return -EFBIG;
2860                 iov_iter_truncate(from, MAX_NON_LFS - (unsigned long)pos);
2861         }
2862
2863         /*
2864          * Are we about to exceed the fs block limit ?
2865          *
2866          * If we have written data it becomes a short write.  If we have
2867          * exceeded without writing data we send a signal and return EFBIG.
2868          * Linus frestrict idea will clean these up nicely..
2869          */
2870         if (unlikely(pos >= inode->i_sb->s_maxbytes))
2871                 return -EFBIG;
2872
2873         iov_iter_truncate(from, inode->i_sb->s_maxbytes - pos);
2874         return iov_iter_count(from);
2875 }
2876 EXPORT_SYMBOL(generic_write_checks);
2877
2878 int pagecache_write_begin(struct file *file, struct address_space *mapping,
2879                                 loff_t pos, unsigned len, unsigned flags,
2880                                 struct page **pagep, void **fsdata)
2881 {
2882         const struct address_space_operations *aops = mapping->a_ops;
2883
2884         return aops->write_begin(file, mapping, pos, len, flags,
2885                                                         pagep, fsdata);
2886 }
2887 EXPORT_SYMBOL(pagecache_write_begin);
2888
2889 int pagecache_write_end(struct file *file, struct address_space *mapping,
2890                                 loff_t pos, unsigned len, unsigned copied,
2891                                 struct page *page, void *fsdata)
2892 {
2893         const struct address_space_operations *aops = mapping->a_ops;
2894
2895         return aops->write_end(file, mapping, pos, len, copied, page, fsdata);
2896 }
2897 EXPORT_SYMBOL(pagecache_write_end);
2898
2899 ssize_t
2900 generic_file_direct_write(struct kiocb *iocb, struct iov_iter *from)
2901 {
2902         struct file     *file = iocb->ki_filp;
2903         struct address_space *mapping = file->f_mapping;
2904         struct inode    *inode = mapping->host;
2905         loff_t          pos = iocb->ki_pos;
2906         ssize_t         written;
2907         size_t          write_len;
2908         pgoff_t         end;
2909
2910         write_len = iov_iter_count(from);
2911         end = (pos + write_len - 1) >> PAGE_SHIFT;
2912
2913         if (iocb->ki_flags & IOCB_NOWAIT) {
2914                 /* If there are pages to writeback, return */
2915                 if (filemap_range_has_page(inode->i_mapping, pos,
2916                                            pos + iov_iter_count(from)))
2917                         return -EAGAIN;
2918         } else {
2919                 written = filemap_write_and_wait_range(mapping, pos,
2920                                                         pos + write_len - 1);
2921                 if (written)
2922                         goto out;
2923         }
2924
2925         /*
2926          * After a write we want buffered reads to be sure to go to disk to get
2927          * the new data.  We invalidate clean cached page from the region we're
2928          * about to write.  We do this *before* the write so that we can return
2929          * without clobbering -EIOCBQUEUED from ->direct_IO().
2930          */
2931         written = invalidate_inode_pages2_range(mapping,
2932                                         pos >> PAGE_SHIFT, end);
2933         /*
2934          * If a page can not be invalidated, return 0 to fall back
2935          * to buffered write.
2936          */
2937         if (written) {
2938                 if (written == -EBUSY)
2939                         return 0;
2940                 goto out;
2941         }
2942
2943         written = mapping->a_ops->direct_IO(iocb, from);
2944
2945         /*
2946          * Finally, try again to invalidate clean pages which might have been
2947          * cached by non-direct readahead, or faulted in by get_user_pages()
2948          * if the source of the write was an mmap'ed region of the file
2949          * we're writing.  Either one is a pretty crazy thing to do,
2950          * so we don't support it 100%.  If this invalidation
2951          * fails, tough, the write still worked...
2952          *
2953          * Most of the time we do not need this since dio_complete() will do
2954          * the invalidation for us. However there are some file systems that
2955          * do not end up with dio_complete() being called, so let's not break
2956          * them by removing it completely
2957          */
2958         if (mapping->nrpages)
2959                 invalidate_inode_pages2_range(mapping,
2960                                         pos >> PAGE_SHIFT, end);
2961
2962         if (written > 0) {
2963                 pos += written;
2964                 write_len -= written;
2965                 if (pos > i_size_read(inode) && !S_ISBLK(inode->i_mode)) {
2966                         i_size_write(inode, pos);
2967                         mark_inode_dirty(inode);
2968                 }
2969                 iocb->ki_pos = pos;
2970         }
2971         iov_iter_revert(from, write_len - iov_iter_count(from));
2972 out:
2973         return written;
2974 }
2975 EXPORT_SYMBOL(generic_file_direct_write);
2976
2977 /*
2978  * Find or create a page at the given pagecache position. Return the locked
2979  * page. This function is specifically for buffered writes.
2980  */
2981 struct page *grab_cache_page_write_begin(struct address_space *mapping,
2982                                         pgoff_t index, unsigned flags)
2983 {
2984         struct page *page;
2985         int fgp_flags = FGP_LOCK|FGP_WRITE|FGP_CREAT;
2986
2987         if (flags & AOP_FLAG_NOFS)
2988                 fgp_flags |= FGP_NOFS;
2989
2990         page = pagecache_get_page(mapping, index, fgp_flags,
2991                         mapping_gfp_mask(mapping));
2992         if (page)
2993                 wait_for_stable_page(page);
2994
2995         return page;
2996 }
2997 EXPORT_SYMBOL(grab_cache_page_write_begin);
2998
2999 ssize_t generic_perform_write(struct file *file,
3000                                 struct iov_iter *i, loff_t pos)
3001 {
3002         struct address_space *mapping = file->f_mapping;
3003         const struct address_space_operations *a_ops = mapping->a_ops;
3004         long status = 0;
3005         ssize_t written = 0;
3006         unsigned int flags = 0;
3007
3008         do {
3009                 struct page *page;
3010                 unsigned long offset;   /* Offset into pagecache page */
3011                 unsigned long bytes;    /* Bytes to write to page */
3012                 size_t copied;          /* Bytes copied from user */
3013                 void *fsdata;
3014
3015                 offset = (pos & (PAGE_SIZE - 1));
3016                 bytes = min_t(unsigned long, PAGE_SIZE - offset,
3017                                                 iov_iter_count(i));
3018
3019 again:
3020                 /*
3021                  * Bring in the user page that we will copy from _first_.
3022                  * Otherwise there's a nasty deadlock on copying from the
3023                  * same page as we're writing to, without it being marked
3024                  * up-to-date.
3025                  *
3026                  * Not only is this an optimisation, but it is also required
3027                  * to check that the address is actually valid, when atomic
3028                  * usercopies are used, below.
3029                  */
3030                 if (unlikely(iov_iter_fault_in_readable(i, bytes))) {
3031                         status = -EFAULT;
3032                         break;
3033                 }
3034
3035                 if (fatal_signal_pending(current)) {
3036                         status = -EINTR;
3037                         break;
3038                 }
3039
3040                 status = a_ops->write_begin(file, mapping, pos, bytes, flags,
3041                                                 &page, &fsdata);
3042                 if (unlikely(status < 0))
3043                         break;
3044
3045                 if (mapping_writably_mapped(mapping))
3046                         flush_dcache_page(page);
3047
3048                 copied = iov_iter_copy_from_user_atomic(page, i, offset, bytes);
3049                 flush_dcache_page(page);
3050
3051                 status = a_ops->write_end(file, mapping, pos, bytes, copied,
3052                                                 page, fsdata);
3053                 if (unlikely(status < 0))
3054                         break;
3055                 copied = status;
3056
3057                 cond_resched();
3058
3059                 iov_iter_advance(i, copied);
3060                 if (unlikely(copied == 0)) {
3061                         /*
3062                          * If we were unable to copy any data at all, we must
3063                          * fall back to a single segment length write.
3064                          *
3065                          * If we didn't fallback here, we could livelock
3066                          * because not all segments in the iov can be copied at
3067                          * once without a pagefault.
3068                          */
3069                         bytes = min_t(unsigned long, PAGE_SIZE - offset,
3070                                                 iov_iter_single_seg_count(i));
3071                         goto again;
3072                 }
3073                 pos += copied;
3074                 written += copied;
3075
3076                 balance_dirty_pages_ratelimited(mapping);
3077         } while (iov_iter_count(i));
3078
3079         return written ? written : status;
3080 }
3081 EXPORT_SYMBOL(generic_perform_write);
3082
3083 /**
3084  * __generic_file_write_iter - write data to a file
3085  * @iocb:       IO state structure (file, offset, etc.)
3086  * @from:       iov_iter with data to write
3087  *
3088  * This function does all the work needed for actually writing data to a
3089  * file. It does all basic checks, removes SUID from the file, updates
3090  * modification times and calls proper subroutines depending on whether we
3091  * do direct IO or a standard buffered write.
3092  *
3093  * It expects i_mutex to be grabbed unless we work on a block device or similar
3094  * object which does not need locking at all.
3095  *
3096  * This function does *not* take care of syncing data in case of O_SYNC write.
3097  * A caller has to handle it. This is mainly due to the fact that we want to
3098  * avoid syncing under i_mutex.
3099  */
3100 ssize_t __generic_file_write_iter(struct kiocb *iocb, struct iov_iter *from)
3101 {
3102         struct file *file = iocb->ki_filp;
3103         struct address_space * mapping = file->f_mapping;
3104         struct inode    *inode = mapping->host;
3105         ssize_t         written = 0;
3106         ssize_t         err;
3107         ssize_t         status;
3108
3109         /* We can write back this queue in page reclaim */
3110         current->backing_dev_info = inode_to_bdi(inode);
3111         err = file_remove_privs(file);
3112         if (err)
3113                 goto out;
3114
3115         err = file_update_time(file);
3116         if (err)
3117                 goto out;
3118
3119         if (iocb->ki_flags & IOCB_DIRECT) {
3120                 loff_t pos, endbyte;
3121
3122                 written = generic_file_direct_write(iocb, from);
3123                 /*
3124                  * If the write stopped short of completing, fall back to
3125                  * buffered writes.  Some filesystems do this for writes to
3126                  * holes, for example.  For DAX files, a buffered write will
3127                  * not succeed (even if it did, DAX does not handle dirty
3128                  * page-cache pages correctly).
3129                  */
3130                 if (written < 0 || !iov_iter_count(from) || IS_DAX(inode))
3131                         goto out;
3132
3133                 status = generic_perform_write(file, from, pos = iocb->ki_pos);
3134                 /*
3135                  * If generic_perform_write() returned a synchronous error
3136                  * then we want to return the number of bytes which were
3137                  * direct-written, or the error code if that was zero.  Note
3138                  * that this differs from normal direct-io semantics, which
3139                  * will return -EFOO even if some bytes were written.
3140                  */
3141                 if (unlikely(status < 0)) {
3142                         err = status;
3143                         goto out;
3144                 }
3145                 /*
3146                  * We need to ensure that the page cache pages are written to
3147                  * disk and invalidated to preserve the expected O_DIRECT
3148                  * semantics.
3149                  */
3150                 endbyte = pos + status - 1;
3151                 err = filemap_write_and_wait_range(mapping, pos, endbyte);
3152                 if (err == 0) {
3153                         iocb->ki_pos = endbyte + 1;
3154                         written += status;
3155                         invalidate_mapping_pages(mapping,
3156                                                  pos >> PAGE_SHIFT,
3157                                                  endbyte >> PAGE_SHIFT);
3158                 } else {
3159                         /*
3160                          * We don't know how much we wrote, so just return
3161                          * the number of bytes which were direct-written
3162                          */
3163                 }
3164         } else {
3165                 written = generic_perform_write(file, from, iocb->ki_pos);
3166                 if (likely(written > 0))
3167                         iocb->ki_pos += written;
3168         }
3169 out:
3170         current->backing_dev_info = NULL;
3171         return written ? written : err;
3172 }
3173 EXPORT_SYMBOL(__generic_file_write_iter);
3174
3175 /**
3176  * generic_file_write_iter - write data to a file
3177  * @iocb:       IO state structure
3178  * @from:       iov_iter with data to write
3179  *
3180  * This is a wrapper around __generic_file_write_iter() to be used by most
3181  * filesystems. It takes care of syncing the file in case of O_SYNC file
3182  * and acquires i_mutex as needed.
3183  */
3184 ssize_t generic_file_write_iter(struct kiocb *iocb, struct iov_iter *from)
3185 {
3186         struct file *file = iocb->ki_filp;
3187         struct inode *inode = file->f_mapping->host;
3188         ssize_t ret;
3189
3190         inode_lock(inode);
3191         ret = generic_write_checks(iocb, from);
3192         if (ret > 0)
3193                 ret = __generic_file_write_iter(iocb, from);
3194         inode_unlock(inode);
3195
3196         if (ret > 0)
3197                 ret = generic_write_sync(iocb, ret);
3198         return ret;
3199 }
3200 EXPORT_SYMBOL(generic_file_write_iter);
3201
3202 /**
3203  * try_to_release_page() - release old fs-specific metadata on a page
3204  *
3205  * @page: the page which the kernel is trying to free
3206  * @gfp_mask: memory allocation flags (and I/O mode)
3207  *
3208  * The address_space is to try to release any data against the page
3209  * (presumably at page->private).  If the release was successful, return '1'.
3210  * Otherwise return zero.
3211  *
3212  * This may also be called if PG_fscache is set on a page, indicating that the
3213  * page is known to the local caching routines.
3214  *
3215  * The @gfp_mask argument specifies whether I/O may be performed to release
3216  * this page (__GFP_IO), and whether the call may block (__GFP_RECLAIM & __GFP_FS).
3217  *
3218  */
3219 int try_to_release_page(struct page *page, gfp_t gfp_mask)
3220 {
3221         struct address_space * const mapping = page->mapping;
3222
3223         BUG_ON(!PageLocked(page));
3224         if (PageWriteback(page))
3225                 return 0;
3226
3227         if (mapping && mapping->a_ops->releasepage)
3228                 return mapping->a_ops->releasepage(page, gfp_mask);
3229         return try_to_free_buffers(page);
3230 }
3231
3232 EXPORT_SYMBOL(try_to_release_page);