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