9d21afd692b92ae243c1f404ce788048e1c4d768
[sfrench/cifs-2.6.git] / mm / filemap.c
1 /*
2  *      linux/mm/filemap.c
3  *
4  * Copyright (C) 1994-1999  Linus Torvalds
5  */
6
7 /*
8  * This file handles the generic file mmap semantics used by
9  * most "normal" filesystems (but you don't /have/ to use this:
10  * the NFS filesystem used to do this differently, for example)
11  */
12 #include <linux/export.h>
13 #include <linux/compiler.h>
14 #include <linux/dax.h>
15 #include <linux/fs.h>
16 #include <linux/sched/signal.h>
17 #include <linux/uaccess.h>
18 #include <linux/capability.h>
19 #include <linux/kernel_stat.h>
20 #include <linux/gfp.h>
21 #include <linux/mm.h>
22 #include <linux/swap.h>
23 #include <linux/mman.h>
24 #include <linux/pagemap.h>
25 #include <linux/file.h>
26 #include <linux/uio.h>
27 #include <linux/hash.h>
28 #include <linux/writeback.h>
29 #include <linux/backing-dev.h>
30 #include <linux/pagevec.h>
31 #include <linux/blkdev.h>
32 #include <linux/security.h>
33 #include <linux/cpuset.h>
34 #include <linux/hardirq.h> /* for BUG_ON(!in_atomic()) only */
35 #include <linux/hugetlb.h>
36 #include <linux/memcontrol.h>
37 #include <linux/cleancache.h>
38 #include <linux/rmap.h>
39 #include "internal.h"
40
41 #define CREATE_TRACE_POINTS
42 #include <trace/events/filemap.h>
43
44 /*
45  * FIXME: remove all knowledge of the buffer layer from the core VM
46  */
47 #include <linux/buffer_head.h> /* for try_to_free_buffers */
48
49 #include <asm/mman.h>
50
51 /*
52  * Shared mappings implemented 30.11.1994. It's not fully working yet,
53  * though.
54  *
55  * Shared mappings now work. 15.8.1995  Bruno.
56  *
57  * finished 'unifying' the page and buffer cache and SMP-threaded the
58  * page-cache, 21.05.1999, Ingo Molnar <mingo@redhat.com>
59  *
60  * SMP-threaded pagemap-LRU 1999, Andrea Arcangeli <andrea@suse.de>
61  */
62
63 /*
64  * Lock ordering:
65  *
66  *  ->i_mmap_rwsem              (truncate_pagecache)
67  *    ->private_lock            (__free_pte->__set_page_dirty_buffers)
68  *      ->swap_lock             (exclusive_swap_page, others)
69  *        ->mapping->tree_lock
70  *
71  *  ->i_mutex
72  *    ->i_mmap_rwsem            (truncate->unmap_mapping_range)
73  *
74  *  ->mmap_sem
75  *    ->i_mmap_rwsem
76  *      ->page_table_lock or pte_lock   (various, mainly in memory.c)
77  *        ->mapping->tree_lock  (arch-dependent flush_dcache_mmap_lock)
78  *
79  *  ->mmap_sem
80  *    ->lock_page               (access_process_vm)
81  *
82  *  ->i_mutex                   (generic_perform_write)
83  *    ->mmap_sem                (fault_in_pages_readable->do_page_fault)
84  *
85  *  bdi->wb.list_lock
86  *    sb_lock                   (fs/fs-writeback.c)
87  *    ->mapping->tree_lock      (__sync_single_inode)
88  *
89  *  ->i_mmap_rwsem
90  *    ->anon_vma.lock           (vma_adjust)
91  *
92  *  ->anon_vma.lock
93  *    ->page_table_lock or pte_lock     (anon_vma_prepare and various)
94  *
95  *  ->page_table_lock or pte_lock
96  *    ->swap_lock               (try_to_unmap_one)
97  *    ->private_lock            (try_to_unmap_one)
98  *    ->tree_lock               (try_to_unmap_one)
99  *    ->zone_lru_lock(zone)     (follow_page->mark_page_accessed)
100  *    ->zone_lru_lock(zone)     (check_pte_range->isolate_lru_page)
101  *    ->private_lock            (page_remove_rmap->set_page_dirty)
102  *    ->tree_lock               (page_remove_rmap->set_page_dirty)
103  *    bdi.wb->list_lock         (page_remove_rmap->set_page_dirty)
104  *    ->inode->i_lock           (page_remove_rmap->set_page_dirty)
105  *    ->memcg->move_lock        (page_remove_rmap->lock_page_memcg)
106  *    bdi.wb->list_lock         (zap_pte_range->set_page_dirty)
107  *    ->inode->i_lock           (zap_pte_range->set_page_dirty)
108  *    ->private_lock            (zap_pte_range->__set_page_dirty_buffers)
109  *
110  * ->i_mmap_rwsem
111  *   ->tasklist_lock            (memory_failure, collect_procs_ao)
112  */
113
114 static int page_cache_tree_insert(struct address_space *mapping,
115                                   struct page *page, void **shadowp)
116 {
117         struct radix_tree_node *node;
118         void **slot;
119         int error;
120
121         error = __radix_tree_create(&mapping->page_tree, page->index, 0,
122                                     &node, &slot);
123         if (error)
124                 return error;
125         if (*slot) {
126                 void *p;
127
128                 p = radix_tree_deref_slot_protected(slot, &mapping->tree_lock);
129                 if (!radix_tree_exceptional_entry(p))
130                         return -EEXIST;
131
132                 mapping->nrexceptional--;
133                 if (shadowp)
134                         *shadowp = p;
135         }
136         __radix_tree_replace(&mapping->page_tree, node, slot, page,
137                              workingset_update_node, mapping);
138         mapping->nrpages++;
139         return 0;
140 }
141
142 static void page_cache_tree_delete(struct address_space *mapping,
143                                    struct page *page, void *shadow)
144 {
145         int i, nr;
146
147         /* hugetlb pages are represented by one entry in the radix tree */
148         nr = PageHuge(page) ? 1 : hpage_nr_pages(page);
149
150         VM_BUG_ON_PAGE(!PageLocked(page), page);
151         VM_BUG_ON_PAGE(PageTail(page), page);
152         VM_BUG_ON_PAGE(nr != 1 && shadow, page);
153
154         for (i = 0; i < nr; i++) {
155                 struct radix_tree_node *node;
156                 void **slot;
157
158                 __radix_tree_lookup(&mapping->page_tree, page->index + i,
159                                     &node, &slot);
160
161                 VM_BUG_ON_PAGE(!node && nr != 1, page);
162
163                 radix_tree_clear_tags(&mapping->page_tree, node, slot);
164                 __radix_tree_replace(&mapping->page_tree, node, slot, shadow,
165                                      workingset_update_node, mapping);
166         }
167
168         if (shadow) {
169                 mapping->nrexceptional += nr;
170                 /*
171                  * Make sure the nrexceptional update is committed before
172                  * the nrpages update so that final truncate racing
173                  * with reclaim does not see both counters 0 at the
174                  * same time and miss a shadow entry.
175                  */
176                 smp_wmb();
177         }
178         mapping->nrpages -= nr;
179 }
180
181 /*
182  * Delete a page from the page cache and free it. Caller has to make
183  * sure the page is locked and that nobody else uses it - or that usage
184  * is safe.  The caller must hold the mapping's tree_lock.
185  */
186 void __delete_from_page_cache(struct page *page, void *shadow)
187 {
188         struct address_space *mapping = page->mapping;
189         int nr = hpage_nr_pages(page);
190
191         trace_mm_filemap_delete_from_page_cache(page);
192         /*
193          * if we're uptodate, flush out into the cleancache, otherwise
194          * invalidate any existing cleancache entries.  We can't leave
195          * stale data around in the cleancache once our page is gone
196          */
197         if (PageUptodate(page) && PageMappedToDisk(page))
198                 cleancache_put_page(page);
199         else
200                 cleancache_invalidate_page(mapping, page);
201
202         VM_BUG_ON_PAGE(PageTail(page), page);
203         VM_BUG_ON_PAGE(page_mapped(page), page);
204         if (!IS_ENABLED(CONFIG_DEBUG_VM) && unlikely(page_mapped(page))) {
205                 int mapcount;
206
207                 pr_alert("BUG: Bad page cache in process %s  pfn:%05lx\n",
208                          current->comm, page_to_pfn(page));
209                 dump_page(page, "still mapped when deleted");
210                 dump_stack();
211                 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
212
213                 mapcount = page_mapcount(page);
214                 if (mapping_exiting(mapping) &&
215                     page_count(page) >= mapcount + 2) {
216                         /*
217                          * All vmas have already been torn down, so it's
218                          * a good bet that actually the page is unmapped,
219                          * and we'd prefer not to leak it: if we're wrong,
220                          * some other bad page check should catch it later.
221                          */
222                         page_mapcount_reset(page);
223                         page_ref_sub(page, mapcount);
224                 }
225         }
226
227         page_cache_tree_delete(mapping, page, shadow);
228
229         page->mapping = NULL;
230         /* Leave page->index set: truncation lookup relies upon it */
231
232         /* hugetlb pages do not participate in page cache accounting. */
233         if (PageHuge(page))
234                 return;
235
236         __mod_node_page_state(page_pgdat(page), NR_FILE_PAGES, -nr);
237         if (PageSwapBacked(page)) {
238                 __mod_node_page_state(page_pgdat(page), NR_SHMEM, -nr);
239                 if (PageTransHuge(page))
240                         __dec_node_page_state(page, NR_SHMEM_THPS);
241         } else {
242                 VM_BUG_ON_PAGE(PageTransHuge(page), page);
243         }
244
245         /*
246          * At this point page must be either written or cleaned by truncate.
247          * Dirty page here signals a bug and loss of unwritten data.
248          *
249          * This fixes dirty accounting after removing the page entirely but
250          * leaves PageDirty set: it has no effect for truncated page and
251          * anyway will be cleared before returning page into buddy allocator.
252          */
253         if (WARN_ON_ONCE(PageDirty(page)))
254                 account_page_cleaned(page, mapping, inode_to_wb(mapping->host));
255 }
256
257 /**
258  * delete_from_page_cache - delete page from page cache
259  * @page: the page which the kernel is trying to remove from page cache
260  *
261  * This must be called only on pages that have been verified to be in the page
262  * cache and locked.  It will never put the page into the free list, the caller
263  * has a reference on the page.
264  */
265 void delete_from_page_cache(struct page *page)
266 {
267         struct address_space *mapping = page_mapping(page);
268         unsigned long flags;
269         void (*freepage)(struct page *);
270
271         BUG_ON(!PageLocked(page));
272
273         freepage = mapping->a_ops->freepage;
274
275         spin_lock_irqsave(&mapping->tree_lock, flags);
276         __delete_from_page_cache(page, NULL);
277         spin_unlock_irqrestore(&mapping->tree_lock, flags);
278
279         if (freepage)
280                 freepage(page);
281
282         if (PageTransHuge(page) && !PageHuge(page)) {
283                 page_ref_sub(page, HPAGE_PMD_NR);
284                 VM_BUG_ON_PAGE(page_count(page) <= 0, page);
285         } else {
286                 put_page(page);
287         }
288 }
289 EXPORT_SYMBOL(delete_from_page_cache);
290
291 int filemap_check_errors(struct address_space *mapping)
292 {
293         int ret = 0;
294         /* Check for outstanding write errors */
295         if (test_bit(AS_ENOSPC, &mapping->flags) &&
296             test_and_clear_bit(AS_ENOSPC, &mapping->flags))
297                 ret = -ENOSPC;
298         if (test_bit(AS_EIO, &mapping->flags) &&
299             test_and_clear_bit(AS_EIO, &mapping->flags))
300                 ret = -EIO;
301         return ret;
302 }
303 EXPORT_SYMBOL(filemap_check_errors);
304
305 static int filemap_check_and_keep_errors(struct address_space *mapping)
306 {
307         /* Check for outstanding write errors */
308         if (test_bit(AS_EIO, &mapping->flags))
309                 return -EIO;
310         if (test_bit(AS_ENOSPC, &mapping->flags))
311                 return -ENOSPC;
312         return 0;
313 }
314
315 /**
316  * __filemap_fdatawrite_range - start writeback on mapping dirty pages in range
317  * @mapping:    address space structure to write
318  * @start:      offset in bytes where the range starts
319  * @end:        offset in bytes where the range ends (inclusive)
320  * @sync_mode:  enable synchronous operation
321  *
322  * Start writeback against all of a mapping's dirty pages that lie
323  * within the byte offsets <start, end> inclusive.
324  *
325  * If sync_mode is WB_SYNC_ALL then this is a "data integrity" operation, as
326  * opposed to a regular memory cleansing writeback.  The difference between
327  * these two operations is that if a dirty page/buffer is encountered, it must
328  * be waited upon, and not just skipped over.
329  */
330 int __filemap_fdatawrite_range(struct address_space *mapping, loff_t start,
331                                 loff_t end, int sync_mode)
332 {
333         int ret;
334         struct writeback_control wbc = {
335                 .sync_mode = sync_mode,
336                 .nr_to_write = LONG_MAX,
337                 .range_start = start,
338                 .range_end = end,
339         };
340
341         if (!mapping_cap_writeback_dirty(mapping))
342                 return 0;
343
344         wbc_attach_fdatawrite_inode(&wbc, mapping->host);
345         ret = do_writepages(mapping, &wbc);
346         wbc_detach_inode(&wbc);
347         return ret;
348 }
349
350 static inline int __filemap_fdatawrite(struct address_space *mapping,
351         int sync_mode)
352 {
353         return __filemap_fdatawrite_range(mapping, 0, LLONG_MAX, sync_mode);
354 }
355
356 int filemap_fdatawrite(struct address_space *mapping)
357 {
358         return __filemap_fdatawrite(mapping, WB_SYNC_ALL);
359 }
360 EXPORT_SYMBOL(filemap_fdatawrite);
361
362 int filemap_fdatawrite_range(struct address_space *mapping, loff_t start,
363                                 loff_t end)
364 {
365         return __filemap_fdatawrite_range(mapping, start, end, WB_SYNC_ALL);
366 }
367 EXPORT_SYMBOL(filemap_fdatawrite_range);
368
369 /**
370  * filemap_flush - mostly a non-blocking flush
371  * @mapping:    target address_space
372  *
373  * This is a mostly non-blocking flush.  Not suitable for data-integrity
374  * purposes - I/O may not be started against all dirty pages.
375  */
376 int filemap_flush(struct address_space *mapping)
377 {
378         return __filemap_fdatawrite(mapping, WB_SYNC_NONE);
379 }
380 EXPORT_SYMBOL(filemap_flush);
381
382 /**
383  * filemap_range_has_page - check if a page exists in range.
384  * @mapping:           address space within which to check
385  * @start_byte:        offset in bytes where the range starts
386  * @end_byte:          offset in bytes where the range ends (inclusive)
387  *
388  * Find at least one page in the range supplied, usually used to check if
389  * direct writing in this range will trigger a writeback.
390  */
391 bool filemap_range_has_page(struct address_space *mapping,
392                            loff_t start_byte, loff_t end_byte)
393 {
394         pgoff_t index = start_byte >> PAGE_SHIFT;
395         pgoff_t end = end_byte >> PAGE_SHIFT;
396         struct page *page;
397
398         if (end_byte < start_byte)
399                 return false;
400
401         if (mapping->nrpages == 0)
402                 return false;
403
404         if (!find_get_pages_range(mapping, &index, end, 1, &page))
405                 return false;
406         put_page(page);
407         return true;
408 }
409 EXPORT_SYMBOL(filemap_range_has_page);
410
411 static void __filemap_fdatawait_range(struct address_space *mapping,
412                                      loff_t start_byte, loff_t end_byte)
413 {
414         pgoff_t index = start_byte >> PAGE_SHIFT;
415         pgoff_t end = end_byte >> PAGE_SHIFT;
416         struct pagevec pvec;
417         int nr_pages;
418
419         if (end_byte < start_byte)
420                 return;
421
422         pagevec_init(&pvec, 0);
423         while ((index <= end) &&
424                         (nr_pages = pagevec_lookup_tag(&pvec, mapping, &index,
425                         PAGECACHE_TAG_WRITEBACK,
426                         min(end - index, (pgoff_t)PAGEVEC_SIZE-1) + 1)) != 0) {
427                 unsigned i;
428
429                 for (i = 0; i < nr_pages; i++) {
430                         struct page *page = pvec.pages[i];
431
432                         /* until radix tree lookup accepts end_index */
433                         if (page->index > end)
434                                 continue;
435
436                         wait_on_page_writeback(page);
437                         ClearPageError(page);
438                 }
439                 pagevec_release(&pvec);
440                 cond_resched();
441         }
442 }
443
444 /**
445  * filemap_fdatawait_range - wait for writeback to complete
446  * @mapping:            address space structure to wait for
447  * @start_byte:         offset in bytes where the range starts
448  * @end_byte:           offset in bytes where the range ends (inclusive)
449  *
450  * Walk the list of under-writeback pages of the given address space
451  * in the given range and wait for all of them.  Check error status of
452  * the address space and return it.
453  *
454  * Since the error status of the address space is cleared by this function,
455  * callers are responsible for checking the return value and handling and/or
456  * reporting the error.
457  */
458 int filemap_fdatawait_range(struct address_space *mapping, loff_t start_byte,
459                             loff_t end_byte)
460 {
461         __filemap_fdatawait_range(mapping, start_byte, end_byte);
462         return filemap_check_errors(mapping);
463 }
464 EXPORT_SYMBOL(filemap_fdatawait_range);
465
466 /**
467  * file_fdatawait_range - wait for writeback to complete
468  * @file:               file pointing to address space structure to wait for
469  * @start_byte:         offset in bytes where the range starts
470  * @end_byte:           offset in bytes where the range ends (inclusive)
471  *
472  * Walk the list of under-writeback pages of the address space that file
473  * refers to, in the given range and wait for all of them.  Check error
474  * status of the address space vs. the file->f_wb_err cursor and return it.
475  *
476  * Since the error status of the file is advanced by this function,
477  * callers are responsible for checking the return value and handling and/or
478  * reporting the error.
479  */
480 int file_fdatawait_range(struct file *file, loff_t start_byte, loff_t end_byte)
481 {
482         struct address_space *mapping = file->f_mapping;
483
484         __filemap_fdatawait_range(mapping, start_byte, end_byte);
485         return file_check_and_advance_wb_err(file);
486 }
487 EXPORT_SYMBOL(file_fdatawait_range);
488
489 /**
490  * filemap_fdatawait_keep_errors - wait for writeback without clearing errors
491  * @mapping: address space structure to wait for
492  *
493  * Walk the list of under-writeback pages of the given address space
494  * and wait for all of them.  Unlike filemap_fdatawait(), this function
495  * does not clear error status of the address space.
496  *
497  * Use this function if callers don't handle errors themselves.  Expected
498  * call sites are system-wide / filesystem-wide data flushers: e.g. sync(2),
499  * fsfreeze(8)
500  */
501 int filemap_fdatawait_keep_errors(struct address_space *mapping)
502 {
503         __filemap_fdatawait_range(mapping, 0, LLONG_MAX);
504         return filemap_check_and_keep_errors(mapping);
505 }
506 EXPORT_SYMBOL(filemap_fdatawait_keep_errors);
507
508 static bool mapping_needs_writeback(struct address_space *mapping)
509 {
510         return (!dax_mapping(mapping) && mapping->nrpages) ||
511             (dax_mapping(mapping) && mapping->nrexceptional);
512 }
513
514 int filemap_write_and_wait(struct address_space *mapping)
515 {
516         int err = 0;
517
518         if (mapping_needs_writeback(mapping)) {
519                 err = filemap_fdatawrite(mapping);
520                 /*
521                  * Even if the above returned error, the pages may be
522                  * written partially (e.g. -ENOSPC), so we wait for it.
523                  * But the -EIO is special case, it may indicate the worst
524                  * thing (e.g. bug) happened, so we avoid waiting for it.
525                  */
526                 if (err != -EIO) {
527                         int err2 = filemap_fdatawait(mapping);
528                         if (!err)
529                                 err = err2;
530                 } else {
531                         /* Clear any previously stored errors */
532                         filemap_check_errors(mapping);
533                 }
534         } else {
535                 err = filemap_check_errors(mapping);
536         }
537         return err;
538 }
539 EXPORT_SYMBOL(filemap_write_and_wait);
540
541 /**
542  * filemap_write_and_wait_range - write out & wait on a file range
543  * @mapping:    the address_space for the pages
544  * @lstart:     offset in bytes where the range starts
545  * @lend:       offset in bytes where the range ends (inclusive)
546  *
547  * Write out and wait upon file offsets lstart->lend, inclusive.
548  *
549  * Note that @lend is inclusive (describes the last byte to be written) so
550  * that this function can be used to write to the very end-of-file (end = -1).
551  */
552 int filemap_write_and_wait_range(struct address_space *mapping,
553                                  loff_t lstart, loff_t lend)
554 {
555         int err = 0;
556
557         if (mapping_needs_writeback(mapping)) {
558                 err = __filemap_fdatawrite_range(mapping, lstart, lend,
559                                                  WB_SYNC_ALL);
560                 /* See comment of filemap_write_and_wait() */
561                 if (err != -EIO) {
562                         int err2 = filemap_fdatawait_range(mapping,
563                                                 lstart, lend);
564                         if (!err)
565                                 err = err2;
566                 } else {
567                         /* Clear any previously stored errors */
568                         filemap_check_errors(mapping);
569                 }
570         } else {
571                 err = filemap_check_errors(mapping);
572         }
573         return err;
574 }
575 EXPORT_SYMBOL(filemap_write_and_wait_range);
576
577 void __filemap_set_wb_err(struct address_space *mapping, int err)
578 {
579         errseq_t eseq = errseq_set(&mapping->wb_err, err);
580
581         trace_filemap_set_wb_err(mapping, eseq);
582 }
583 EXPORT_SYMBOL(__filemap_set_wb_err);
584
585 /**
586  * file_check_and_advance_wb_err - report wb error (if any) that was previously
587  *                                 and advance wb_err to current one
588  * @file: struct file on which the error is being reported
589  *
590  * When userland calls fsync (or something like nfsd does the equivalent), we
591  * want to report any writeback errors that occurred since the last fsync (or
592  * since the file was opened if there haven't been any).
593  *
594  * Grab the wb_err from the mapping. If it matches what we have in the file,
595  * then just quickly return 0. The file is all caught up.
596  *
597  * If it doesn't match, then take the mapping value, set the "seen" flag in
598  * it and try to swap it into place. If it works, or another task beat us
599  * to it with the new value, then update the f_wb_err and return the error
600  * portion. The error at this point must be reported via proper channels
601  * (a'la fsync, or NFS COMMIT operation, etc.).
602  *
603  * While we handle mapping->wb_err with atomic operations, the f_wb_err
604  * value is protected by the f_lock since we must ensure that it reflects
605  * the latest value swapped in for this file descriptor.
606  */
607 int file_check_and_advance_wb_err(struct file *file)
608 {
609         int err = 0;
610         errseq_t old = READ_ONCE(file->f_wb_err);
611         struct address_space *mapping = file->f_mapping;
612
613         /* Locklessly handle the common case where nothing has changed */
614         if (errseq_check(&mapping->wb_err, old)) {
615                 /* Something changed, must use slow path */
616                 spin_lock(&file->f_lock);
617                 old = file->f_wb_err;
618                 err = errseq_check_and_advance(&mapping->wb_err,
619                                                 &file->f_wb_err);
620                 trace_file_check_and_advance_wb_err(file, old);
621                 spin_unlock(&file->f_lock);
622         }
623         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
913         key.page = page;
914         key.bit_nr = bit_nr;
915         key.page_match = 0;
916
917         spin_lock_irqsave(&q->lock, flags);
918         __wake_up_locked_key(q, TASK_NORMAL, &key);
919         /*
920          * It is possible for other pages to have collided on the waitqueue
921          * hash, so in that case check for a page match. That prevents a long-
922          * term waiter
923          *
924          * It is still possible to miss a case here, when we woke page waiters
925          * and removed them from the waitqueue, but there are still other
926          * page waiters.
927          */
928         if (!waitqueue_active(q) || !key.page_match) {
929                 ClearPageWaiters(page);
930                 /*
931                  * It's possible to miss clearing Waiters here, when we woke
932                  * our page waiters, but the hashed waitqueue has waiters for
933                  * other pages on it.
934                  *
935                  * That's okay, it's a rare case. The next waker will clear it.
936                  */
937         }
938         spin_unlock_irqrestore(&q->lock, flags);
939 }
940
941 static void wake_up_page(struct page *page, int bit)
942 {
943         if (!PageWaiters(page))
944                 return;
945         wake_up_page_bit(page, bit);
946 }
947
948 static inline int wait_on_page_bit_common(wait_queue_head_t *q,
949                 struct page *page, int bit_nr, int state, bool lock)
950 {
951         struct wait_page_queue wait_page;
952         wait_queue_entry_t *wait = &wait_page.wait;
953         int ret = 0;
954
955         init_wait(wait);
956         wait->flags = lock ? WQ_FLAG_EXCLUSIVE : 0;
957         wait->func = wake_page_function;
958         wait_page.page = page;
959         wait_page.bit_nr = bit_nr;
960
961         for (;;) {
962                 spin_lock_irq(&q->lock);
963
964                 if (likely(list_empty(&wait->entry))) {
965                         __add_wait_queue_entry_tail(q, wait);
966                         SetPageWaiters(page);
967                 }
968
969                 set_current_state(state);
970
971                 spin_unlock_irq(&q->lock);
972
973                 if (likely(test_bit(bit_nr, &page->flags))) {
974                         io_schedule();
975                 }
976
977                 if (lock) {
978                         if (!test_and_set_bit_lock(bit_nr, &page->flags))
979                                 break;
980                 } else {
981                         if (!test_bit(bit_nr, &page->flags))
982                                 break;
983                 }
984
985                 if (unlikely(signal_pending_state(state, current))) {
986                         ret = -EINTR;
987                         break;
988                 }
989         }
990
991         finish_wait(q, wait);
992
993         /*
994          * A signal could leave PageWaiters set. Clearing it here if
995          * !waitqueue_active would be possible (by open-coding finish_wait),
996          * but still fail to catch it in the case of wait hash collision. We
997          * already can fail to clear wait hash collision cases, so don't
998          * bother with signals either.
999          */
1000
1001         return ret;
1002 }
1003
1004 void wait_on_page_bit(struct page *page, int bit_nr)
1005 {
1006         wait_queue_head_t *q = page_waitqueue(page);
1007         wait_on_page_bit_common(q, page, bit_nr, TASK_UNINTERRUPTIBLE, false);
1008 }
1009 EXPORT_SYMBOL(wait_on_page_bit);
1010
1011 int wait_on_page_bit_killable(struct page *page, int bit_nr)
1012 {
1013         wait_queue_head_t *q = page_waitqueue(page);
1014         return wait_on_page_bit_common(q, page, bit_nr, TASK_KILLABLE, false);
1015 }
1016
1017 /**
1018  * add_page_wait_queue - Add an arbitrary waiter to a page's wait queue
1019  * @page: Page defining the wait queue of interest
1020  * @waiter: Waiter to add to the queue
1021  *
1022  * Add an arbitrary @waiter to the wait queue for the nominated @page.
1023  */
1024 void add_page_wait_queue(struct page *page, wait_queue_entry_t *waiter)
1025 {
1026         wait_queue_head_t *q = page_waitqueue(page);
1027         unsigned long flags;
1028
1029         spin_lock_irqsave(&q->lock, flags);
1030         __add_wait_queue_entry_tail(q, waiter);
1031         SetPageWaiters(page);
1032         spin_unlock_irqrestore(&q->lock, flags);
1033 }
1034 EXPORT_SYMBOL_GPL(add_page_wait_queue);
1035
1036 #ifndef clear_bit_unlock_is_negative_byte
1037
1038 /*
1039  * PG_waiters is the high bit in the same byte as PG_lock.
1040  *
1041  * On x86 (and on many other architectures), we can clear PG_lock and
1042  * test the sign bit at the same time. But if the architecture does
1043  * not support that special operation, we just do this all by hand
1044  * instead.
1045  *
1046  * The read of PG_waiters has to be after (or concurrently with) PG_locked
1047  * being cleared, but a memory barrier should be unneccssary since it is
1048  * in the same byte as PG_locked.
1049  */
1050 static inline bool clear_bit_unlock_is_negative_byte(long nr, volatile void *mem)
1051 {
1052         clear_bit_unlock(nr, mem);
1053         /* smp_mb__after_atomic(); */
1054         return test_bit(PG_waiters, mem);
1055 }
1056
1057 #endif
1058
1059 /**
1060  * unlock_page - unlock a locked page
1061  * @page: the page
1062  *
1063  * Unlocks the page and wakes up sleepers in ___wait_on_page_locked().
1064  * Also wakes sleepers in wait_on_page_writeback() because the wakeup
1065  * mechanism between PageLocked pages and PageWriteback pages is shared.
1066  * But that's OK - sleepers in wait_on_page_writeback() just go back to sleep.
1067  *
1068  * Note that this depends on PG_waiters being the sign bit in the byte
1069  * that contains PG_locked - thus the BUILD_BUG_ON(). That allows us to
1070  * clear the PG_locked bit and test PG_waiters at the same time fairly
1071  * portably (architectures that do LL/SC can test any bit, while x86 can
1072  * test the sign bit).
1073  */
1074 void unlock_page(struct page *page)
1075 {
1076         BUILD_BUG_ON(PG_waiters != 7);
1077         page = compound_head(page);
1078         VM_BUG_ON_PAGE(!PageLocked(page), page);
1079         if (clear_bit_unlock_is_negative_byte(PG_locked, &page->flags))
1080                 wake_up_page_bit(page, PG_locked);
1081 }
1082 EXPORT_SYMBOL(unlock_page);
1083
1084 /**
1085  * end_page_writeback - end writeback against a page
1086  * @page: the page
1087  */
1088 void end_page_writeback(struct page *page)
1089 {
1090         /*
1091          * TestClearPageReclaim could be used here but it is an atomic
1092          * operation and overkill in this particular case. Failing to
1093          * shuffle a page marked for immediate reclaim is too mild to
1094          * justify taking an atomic operation penalty at the end of
1095          * ever page writeback.
1096          */
1097         if (PageReclaim(page)) {
1098                 ClearPageReclaim(page);
1099                 rotate_reclaimable_page(page);
1100         }
1101
1102         if (!test_clear_page_writeback(page))
1103                 BUG();
1104
1105         smp_mb__after_atomic();
1106         wake_up_page(page, PG_writeback);
1107 }
1108 EXPORT_SYMBOL(end_page_writeback);
1109
1110 /*
1111  * After completing I/O on a page, call this routine to update the page
1112  * flags appropriately
1113  */
1114 void page_endio(struct page *page, bool is_write, int err)
1115 {
1116         if (!is_write) {
1117                 if (!err) {
1118                         SetPageUptodate(page);
1119                 } else {
1120                         ClearPageUptodate(page);
1121                         SetPageError(page);
1122                 }
1123                 unlock_page(page);
1124         } else {
1125                 if (err) {
1126                         struct address_space *mapping;
1127
1128                         SetPageError(page);
1129                         mapping = page_mapping(page);
1130                         if (mapping)
1131                                 mapping_set_error(mapping, err);
1132                 }
1133                 end_page_writeback(page);
1134         }
1135 }
1136 EXPORT_SYMBOL_GPL(page_endio);
1137
1138 /**
1139  * __lock_page - get a lock on the page, assuming we need to sleep to get it
1140  * @__page: the page to lock
1141  */
1142 void __lock_page(struct page *__page)
1143 {
1144         struct page *page = compound_head(__page);
1145         wait_queue_head_t *q = page_waitqueue(page);
1146         wait_on_page_bit_common(q, page, PG_locked, TASK_UNINTERRUPTIBLE, true);
1147 }
1148 EXPORT_SYMBOL(__lock_page);
1149
1150 int __lock_page_killable(struct page *__page)
1151 {
1152         struct page *page = compound_head(__page);
1153         wait_queue_head_t *q = page_waitqueue(page);
1154         return wait_on_page_bit_common(q, page, PG_locked, TASK_KILLABLE, true);
1155 }
1156 EXPORT_SYMBOL_GPL(__lock_page_killable);
1157
1158 /*
1159  * Return values:
1160  * 1 - page is locked; mmap_sem is still held.
1161  * 0 - page is not locked.
1162  *     mmap_sem has been released (up_read()), unless flags had both
1163  *     FAULT_FLAG_ALLOW_RETRY and FAULT_FLAG_RETRY_NOWAIT set, in
1164  *     which case mmap_sem is still held.
1165  *
1166  * If neither ALLOW_RETRY nor KILLABLE are set, will always return 1
1167  * with the page locked and the mmap_sem unperturbed.
1168  */
1169 int __lock_page_or_retry(struct page *page, struct mm_struct *mm,
1170                          unsigned int flags)
1171 {
1172         if (flags & FAULT_FLAG_ALLOW_RETRY) {
1173                 /*
1174                  * CAUTION! In this case, mmap_sem is not released
1175                  * even though return 0.
1176                  */
1177                 if (flags & FAULT_FLAG_RETRY_NOWAIT)
1178                         return 0;
1179
1180                 up_read(&mm->mmap_sem);
1181                 if (flags & FAULT_FLAG_KILLABLE)
1182                         wait_on_page_locked_killable(page);
1183                 else
1184                         wait_on_page_locked(page);
1185                 return 0;
1186         } else {
1187                 if (flags & FAULT_FLAG_KILLABLE) {
1188                         int ret;
1189
1190                         ret = __lock_page_killable(page);
1191                         if (ret) {
1192                                 up_read(&mm->mmap_sem);
1193                                 return 0;
1194                         }
1195                 } else
1196                         __lock_page(page);
1197                 return 1;
1198         }
1199 }
1200
1201 /**
1202  * page_cache_next_hole - find the next hole (not-present entry)
1203  * @mapping: mapping
1204  * @index: index
1205  * @max_scan: maximum range to search
1206  *
1207  * Search the set [index, min(index+max_scan-1, MAX_INDEX)] for the
1208  * lowest indexed hole.
1209  *
1210  * Returns: the index of the hole if found, otherwise returns an index
1211  * outside of the set specified (in which case 'return - index >=
1212  * max_scan' will be true). In rare cases of index wrap-around, 0 will
1213  * be returned.
1214  *
1215  * page_cache_next_hole may be called under rcu_read_lock. However,
1216  * like radix_tree_gang_lookup, this will not atomically search a
1217  * snapshot of the tree at a single point in time. For example, if a
1218  * hole is created at index 5, then subsequently a hole is created at
1219  * index 10, page_cache_next_hole covering both indexes may return 10
1220  * if called under rcu_read_lock.
1221  */
1222 pgoff_t page_cache_next_hole(struct address_space *mapping,
1223                              pgoff_t index, unsigned long max_scan)
1224 {
1225         unsigned long i;
1226
1227         for (i = 0; i < max_scan; i++) {
1228                 struct page *page;
1229
1230                 page = radix_tree_lookup(&mapping->page_tree, index);
1231                 if (!page || radix_tree_exceptional_entry(page))
1232                         break;
1233                 index++;
1234                 if (index == 0)
1235                         break;
1236         }
1237
1238         return index;
1239 }
1240 EXPORT_SYMBOL(page_cache_next_hole);
1241
1242 /**
1243  * page_cache_prev_hole - find the prev hole (not-present entry)
1244  * @mapping: mapping
1245  * @index: index
1246  * @max_scan: maximum range to search
1247  *
1248  * Search backwards in the range [max(index-max_scan+1, 0), index] for
1249  * the first hole.
1250  *
1251  * Returns: the index of the hole if found, otherwise returns an index
1252  * outside of the set specified (in which case 'index - return >=
1253  * max_scan' will be true). In rare cases of wrap-around, ULONG_MAX
1254  * will be returned.
1255  *
1256  * page_cache_prev_hole may be called under rcu_read_lock. However,
1257  * like radix_tree_gang_lookup, this will not atomically search a
1258  * snapshot of the tree at a single point in time. For example, if a
1259  * hole is created at index 10, then subsequently a hole is created at
1260  * index 5, page_cache_prev_hole covering both indexes may return 5 if
1261  * called under rcu_read_lock.
1262  */
1263 pgoff_t page_cache_prev_hole(struct address_space *mapping,
1264                              pgoff_t index, unsigned long max_scan)
1265 {
1266         unsigned long i;
1267
1268         for (i = 0; i < max_scan; i++) {
1269                 struct page *page;
1270
1271                 page = radix_tree_lookup(&mapping->page_tree, index);
1272                 if (!page || radix_tree_exceptional_entry(page))
1273                         break;
1274                 index--;
1275                 if (index == ULONG_MAX)
1276                         break;
1277         }
1278
1279         return index;
1280 }
1281 EXPORT_SYMBOL(page_cache_prev_hole);
1282
1283 /**
1284  * find_get_entry - find and get a page cache entry
1285  * @mapping: the address_space to search
1286  * @offset: the page cache index
1287  *
1288  * Looks up the page cache slot at @mapping & @offset.  If there is a
1289  * page cache page, it is returned with an increased refcount.
1290  *
1291  * If the slot holds a shadow entry of a previously evicted page, or a
1292  * swap entry from shmem/tmpfs, it is returned.
1293  *
1294  * Otherwise, %NULL is returned.
1295  */
1296 struct page *find_get_entry(struct address_space *mapping, pgoff_t offset)
1297 {
1298         void **pagep;
1299         struct page *head, *page;
1300
1301         rcu_read_lock();
1302 repeat:
1303         page = NULL;
1304         pagep = radix_tree_lookup_slot(&mapping->page_tree, offset);
1305         if (pagep) {
1306                 page = radix_tree_deref_slot(pagep);
1307                 if (unlikely(!page))
1308                         goto out;
1309                 if (radix_tree_exception(page)) {
1310                         if (radix_tree_deref_retry(page))
1311                                 goto repeat;
1312                         /*
1313                          * A shadow entry of a recently evicted page,
1314                          * or a swap entry from shmem/tmpfs.  Return
1315                          * it without attempting to raise page count.
1316                          */
1317                         goto out;
1318                 }
1319
1320                 head = compound_head(page);
1321                 if (!page_cache_get_speculative(head))
1322                         goto repeat;
1323
1324                 /* The page was split under us? */
1325                 if (compound_head(page) != head) {
1326                         put_page(head);
1327                         goto repeat;
1328                 }
1329
1330                 /*
1331                  * Has the page moved?
1332                  * This is part of the lockless pagecache protocol. See
1333                  * include/linux/pagemap.h for details.
1334                  */
1335                 if (unlikely(page != *pagep)) {
1336                         put_page(head);
1337                         goto repeat;
1338                 }
1339         }
1340 out:
1341         rcu_read_unlock();
1342
1343         return page;
1344 }
1345 EXPORT_SYMBOL(find_get_entry);
1346
1347 /**
1348  * find_lock_entry - locate, pin and lock a page cache entry
1349  * @mapping: the address_space to search
1350  * @offset: the page cache index
1351  *
1352  * Looks up the page cache slot at @mapping & @offset.  If there is a
1353  * page cache page, it is returned locked and with an increased
1354  * refcount.
1355  *
1356  * If the slot holds a shadow entry of a previously evicted page, or a
1357  * swap entry from shmem/tmpfs, it is returned.
1358  *
1359  * Otherwise, %NULL is returned.
1360  *
1361  * find_lock_entry() may sleep.
1362  */
1363 struct page *find_lock_entry(struct address_space *mapping, pgoff_t offset)
1364 {
1365         struct page *page;
1366
1367 repeat:
1368         page = find_get_entry(mapping, offset);
1369         if (page && !radix_tree_exception(page)) {
1370                 lock_page(page);
1371                 /* Has the page been truncated? */
1372                 if (unlikely(page_mapping(page) != mapping)) {
1373                         unlock_page(page);
1374                         put_page(page);
1375                         goto repeat;
1376                 }
1377                 VM_BUG_ON_PAGE(page_to_pgoff(page) != offset, page);
1378         }
1379         return page;
1380 }
1381 EXPORT_SYMBOL(find_lock_entry);
1382
1383 /**
1384  * pagecache_get_page - find and get a page reference
1385  * @mapping: the address_space to search
1386  * @offset: the page index
1387  * @fgp_flags: PCG flags
1388  * @gfp_mask: gfp mask to use for the page cache data page allocation
1389  *
1390  * Looks up the page cache slot at @mapping & @offset.
1391  *
1392  * PCG flags modify how the page is returned.
1393  *
1394  * @fgp_flags can be:
1395  *
1396  * - FGP_ACCESSED: the page will be marked accessed
1397  * - FGP_LOCK: Page is return locked
1398  * - FGP_CREAT: If page is not present then a new page is allocated using
1399  *   @gfp_mask and added to the page cache and the VM's LRU
1400  *   list. The page is returned locked and with an increased
1401  *   refcount. Otherwise, NULL is returned.
1402  *
1403  * If FGP_LOCK or FGP_CREAT are specified then the function may sleep even
1404  * if the GFP flags specified for FGP_CREAT are atomic.
1405  *
1406  * If there is a page cache page, it is returned with an increased refcount.
1407  */
1408 struct page *pagecache_get_page(struct address_space *mapping, pgoff_t offset,
1409         int fgp_flags, gfp_t gfp_mask)
1410 {
1411         struct page *page;
1412
1413 repeat:
1414         page = find_get_entry(mapping, offset);
1415         if (radix_tree_exceptional_entry(page))
1416                 page = NULL;
1417         if (!page)
1418                 goto no_page;
1419
1420         if (fgp_flags & FGP_LOCK) {
1421                 if (fgp_flags & FGP_NOWAIT) {
1422                         if (!trylock_page(page)) {
1423                                 put_page(page);
1424                                 return NULL;
1425                         }
1426                 } else {
1427                         lock_page(page);
1428                 }
1429
1430                 /* Has the page been truncated? */
1431                 if (unlikely(page->mapping != mapping)) {
1432                         unlock_page(page);
1433                         put_page(page);
1434                         goto repeat;
1435                 }
1436                 VM_BUG_ON_PAGE(page->index != offset, page);
1437         }
1438
1439         if (page && (fgp_flags & FGP_ACCESSED))
1440                 mark_page_accessed(page);
1441
1442 no_page:
1443         if (!page && (fgp_flags & FGP_CREAT)) {
1444                 int err;
1445                 if ((fgp_flags & FGP_WRITE) && mapping_cap_account_dirty(mapping))
1446                         gfp_mask |= __GFP_WRITE;
1447                 if (fgp_flags & FGP_NOFS)
1448                         gfp_mask &= ~__GFP_FS;
1449
1450                 page = __page_cache_alloc(gfp_mask);
1451                 if (!page)
1452                         return NULL;
1453
1454                 if (WARN_ON_ONCE(!(fgp_flags & FGP_LOCK)))
1455                         fgp_flags |= FGP_LOCK;
1456
1457                 /* Init accessed so avoid atomic mark_page_accessed later */
1458                 if (fgp_flags & FGP_ACCESSED)
1459                         __SetPageReferenced(page);
1460
1461                 err = add_to_page_cache_lru(page, mapping, offset,
1462                                 gfp_mask & GFP_RECLAIM_MASK);
1463                 if (unlikely(err)) {
1464                         put_page(page);
1465                         page = NULL;
1466                         if (err == -EEXIST)
1467                                 goto repeat;
1468                 }
1469         }
1470
1471         return page;
1472 }
1473 EXPORT_SYMBOL(pagecache_get_page);
1474
1475 /**
1476  * find_get_entries - gang pagecache lookup
1477  * @mapping:    The address_space to search
1478  * @start:      The starting page cache index
1479  * @nr_entries: The maximum number of entries
1480  * @entries:    Where the resulting entries are placed
1481  * @indices:    The cache indices corresponding to the entries in @entries
1482  *
1483  * find_get_entries() will search for and return a group of up to
1484  * @nr_entries entries in the mapping.  The entries are placed at
1485  * @entries.  find_get_entries() takes a reference against any actual
1486  * pages it returns.
1487  *
1488  * The search returns a group of mapping-contiguous page cache entries
1489  * with ascending indexes.  There may be holes in the indices due to
1490  * not-present pages.
1491  *
1492  * Any shadow entries of evicted pages, or swap entries from
1493  * shmem/tmpfs, are included in the returned array.
1494  *
1495  * find_get_entries() returns the number of pages and shadow entries
1496  * which were found.
1497  */
1498 unsigned find_get_entries(struct address_space *mapping,
1499                           pgoff_t start, unsigned int nr_entries,
1500                           struct page **entries, pgoff_t *indices)
1501 {
1502         void **slot;
1503         unsigned int ret = 0;
1504         struct radix_tree_iter iter;
1505
1506         if (!nr_entries)
1507                 return 0;
1508
1509         rcu_read_lock();
1510         radix_tree_for_each_slot(slot, &mapping->page_tree, &iter, start) {
1511                 struct page *head, *page;
1512 repeat:
1513                 page = radix_tree_deref_slot(slot);
1514                 if (unlikely(!page))
1515                         continue;
1516                 if (radix_tree_exception(page)) {
1517                         if (radix_tree_deref_retry(page)) {
1518                                 slot = radix_tree_iter_retry(&iter);
1519                                 continue;
1520                         }
1521                         /*
1522                          * A shadow entry of a recently evicted page, a swap
1523                          * entry from shmem/tmpfs or a DAX entry.  Return it
1524                          * without attempting to raise page count.
1525                          */
1526                         goto export;
1527                 }
1528
1529                 head = compound_head(page);
1530                 if (!page_cache_get_speculative(head))
1531                         goto repeat;
1532
1533                 /* The page was split under us? */
1534                 if (compound_head(page) != head) {
1535                         put_page(head);
1536                         goto repeat;
1537                 }
1538
1539                 /* Has the page moved? */
1540                 if (unlikely(page != *slot)) {
1541                         put_page(head);
1542                         goto repeat;
1543                 }
1544 export:
1545                 indices[ret] = iter.index;
1546                 entries[ret] = page;
1547                 if (++ret == nr_entries)
1548                         break;
1549         }
1550         rcu_read_unlock();
1551         return ret;
1552 }
1553
1554 /**
1555  * find_get_pages_range - gang pagecache lookup
1556  * @mapping:    The address_space to search
1557  * @start:      The starting page index
1558  * @end:        The final page index (inclusive)
1559  * @nr_pages:   The maximum number of pages
1560  * @pages:      Where the resulting pages are placed
1561  *
1562  * find_get_pages_range() will search for and return a group of up to @nr_pages
1563  * pages in the mapping starting at index @start and up to index @end
1564  * (inclusive).  The pages are placed at @pages.  find_get_pages_range() takes
1565  * a reference against the returned pages.
1566  *
1567  * The search returns a group of mapping-contiguous pages with ascending
1568  * indexes.  There may be holes in the indices due to not-present pages.
1569  * We also update @start to index the next page for the traversal.
1570  *
1571  * find_get_pages_range() returns the number of pages which were found. If this
1572  * number is smaller than @nr_pages, the end of specified range has been
1573  * reached.
1574  */
1575 unsigned find_get_pages_range(struct address_space *mapping, pgoff_t *start,
1576                               pgoff_t end, unsigned int nr_pages,
1577                               struct page **pages)
1578 {
1579         struct radix_tree_iter iter;
1580         void **slot;
1581         unsigned ret = 0;
1582
1583         if (unlikely(!nr_pages))
1584                 return 0;
1585
1586         rcu_read_lock();
1587         radix_tree_for_each_slot(slot, &mapping->page_tree, &iter, *start) {
1588                 struct page *head, *page;
1589
1590                 if (iter.index > end)
1591                         break;
1592 repeat:
1593                 page = radix_tree_deref_slot(slot);
1594                 if (unlikely(!page))
1595                         continue;
1596
1597                 if (radix_tree_exception(page)) {
1598                         if (radix_tree_deref_retry(page)) {
1599                                 slot = radix_tree_iter_retry(&iter);
1600                                 continue;
1601                         }
1602                         /*
1603                          * A shadow entry of a recently evicted page,
1604                          * or a swap entry from shmem/tmpfs.  Skip
1605                          * over it.
1606                          */
1607                         continue;
1608                 }
1609
1610                 head = compound_head(page);
1611                 if (!page_cache_get_speculative(head))
1612                         goto repeat;
1613
1614                 /* The page was split under us? */
1615                 if (compound_head(page) != head) {
1616                         put_page(head);
1617                         goto repeat;
1618                 }
1619
1620                 /* Has the page moved? */
1621                 if (unlikely(page != *slot)) {
1622                         put_page(head);
1623                         goto repeat;
1624                 }
1625
1626                 pages[ret] = page;
1627                 if (++ret == nr_pages) {
1628                         *start = pages[ret - 1]->index + 1;
1629                         goto out;
1630                 }
1631         }
1632
1633         /*
1634          * We come here when there is no page beyond @end. We take care to not
1635          * overflow the index @start as it confuses some of the callers. This
1636          * breaks the iteration when there is page at index -1 but that is
1637          * already broken anyway.
1638          */
1639         if (end == (pgoff_t)-1)
1640                 *start = (pgoff_t)-1;
1641         else
1642                 *start = end + 1;
1643 out:
1644         rcu_read_unlock();
1645
1646         return ret;
1647 }
1648
1649 /**
1650  * find_get_pages_contig - gang contiguous pagecache lookup
1651  * @mapping:    The address_space to search
1652  * @index:      The starting page index
1653  * @nr_pages:   The maximum number of pages
1654  * @pages:      Where the resulting pages are placed
1655  *
1656  * find_get_pages_contig() works exactly like find_get_pages(), except
1657  * that the returned number of pages are guaranteed to be contiguous.
1658  *
1659  * find_get_pages_contig() returns the number of pages which were found.
1660  */
1661 unsigned find_get_pages_contig(struct address_space *mapping, pgoff_t index,
1662                                unsigned int nr_pages, struct page **pages)
1663 {
1664         struct radix_tree_iter iter;
1665         void **slot;
1666         unsigned int ret = 0;
1667
1668         if (unlikely(!nr_pages))
1669                 return 0;
1670
1671         rcu_read_lock();
1672         radix_tree_for_each_contig(slot, &mapping->page_tree, &iter, index) {
1673                 struct page *head, *page;
1674 repeat:
1675                 page = radix_tree_deref_slot(slot);
1676                 /* The hole, there no reason to continue */
1677                 if (unlikely(!page))
1678                         break;
1679
1680                 if (radix_tree_exception(page)) {
1681                         if (radix_tree_deref_retry(page)) {
1682                                 slot = radix_tree_iter_retry(&iter);
1683                                 continue;
1684                         }
1685                         /*
1686                          * A shadow entry of a recently evicted page,
1687                          * or a swap entry from shmem/tmpfs.  Stop
1688                          * looking for contiguous pages.
1689                          */
1690                         break;
1691                 }
1692
1693                 head = compound_head(page);
1694                 if (!page_cache_get_speculative(head))
1695                         goto repeat;
1696
1697                 /* The page was split under us? */
1698                 if (compound_head(page) != head) {
1699                         put_page(head);
1700                         goto repeat;
1701                 }
1702
1703                 /* Has the page moved? */
1704                 if (unlikely(page != *slot)) {
1705                         put_page(head);
1706                         goto repeat;
1707                 }
1708
1709                 /*
1710                  * must check mapping and index after taking the ref.
1711                  * otherwise we can get both false positives and false
1712                  * negatives, which is just confusing to the caller.
1713                  */
1714                 if (page->mapping == NULL || page_to_pgoff(page) != iter.index) {
1715                         put_page(page);
1716                         break;
1717                 }
1718
1719                 pages[ret] = page;
1720                 if (++ret == nr_pages)
1721                         break;
1722         }
1723         rcu_read_unlock();
1724         return ret;
1725 }
1726 EXPORT_SYMBOL(find_get_pages_contig);
1727
1728 /**
1729  * find_get_pages_tag - find and return pages that match @tag
1730  * @mapping:    the address_space to search
1731  * @index:      the starting page index
1732  * @tag:        the tag index
1733  * @nr_pages:   the maximum number of pages
1734  * @pages:      where the resulting pages are placed
1735  *
1736  * Like find_get_pages, except we only return pages which are tagged with
1737  * @tag.   We update @index to index the next page for the traversal.
1738  */
1739 unsigned find_get_pages_tag(struct address_space *mapping, pgoff_t *index,
1740                         int tag, unsigned int nr_pages, struct page **pages)
1741 {
1742         struct radix_tree_iter iter;
1743         void **slot;
1744         unsigned ret = 0;
1745
1746         if (unlikely(!nr_pages))
1747                 return 0;
1748
1749         rcu_read_lock();
1750         radix_tree_for_each_tagged(slot, &mapping->page_tree,
1751                                    &iter, *index, tag) {
1752                 struct page *head, *page;
1753 repeat:
1754                 page = radix_tree_deref_slot(slot);
1755                 if (unlikely(!page))
1756                         continue;
1757
1758                 if (radix_tree_exception(page)) {
1759                         if (radix_tree_deref_retry(page)) {
1760                                 slot = radix_tree_iter_retry(&iter);
1761                                 continue;
1762                         }
1763                         /*
1764                          * A shadow entry of a recently evicted page.
1765                          *
1766                          * Those entries should never be tagged, but
1767                          * this tree walk is lockless and the tags are
1768                          * looked up in bulk, one radix tree node at a
1769                          * time, so there is a sizable window for page
1770                          * reclaim to evict a page we saw tagged.
1771                          *
1772                          * Skip over it.
1773                          */
1774                         continue;
1775                 }
1776
1777                 head = compound_head(page);
1778                 if (!page_cache_get_speculative(head))
1779                         goto repeat;
1780
1781                 /* The page was split under us? */
1782                 if (compound_head(page) != head) {
1783                         put_page(head);
1784                         goto repeat;
1785                 }
1786
1787                 /* Has the page moved? */
1788                 if (unlikely(page != *slot)) {
1789                         put_page(head);
1790                         goto repeat;
1791                 }
1792
1793                 pages[ret] = page;
1794                 if (++ret == nr_pages)
1795                         break;
1796         }
1797
1798         rcu_read_unlock();
1799
1800         if (ret)
1801                 *index = pages[ret - 1]->index + 1;
1802
1803         return ret;
1804 }
1805 EXPORT_SYMBOL(find_get_pages_tag);
1806
1807 /**
1808  * find_get_entries_tag - find and return entries that match @tag
1809  * @mapping:    the address_space to search
1810  * @start:      the starting page cache index
1811  * @tag:        the tag index
1812  * @nr_entries: the maximum number of entries
1813  * @entries:    where the resulting entries are placed
1814  * @indices:    the cache indices corresponding to the entries in @entries
1815  *
1816  * Like find_get_entries, except we only return entries which are tagged with
1817  * @tag.
1818  */
1819 unsigned find_get_entries_tag(struct address_space *mapping, pgoff_t start,
1820                         int tag, unsigned int nr_entries,
1821                         struct page **entries, pgoff_t *indices)
1822 {
1823         void **slot;
1824         unsigned int ret = 0;
1825         struct radix_tree_iter iter;
1826
1827         if (!nr_entries)
1828                 return 0;
1829
1830         rcu_read_lock();
1831         radix_tree_for_each_tagged(slot, &mapping->page_tree,
1832                                    &iter, start, tag) {
1833                 struct page *head, *page;
1834 repeat:
1835                 page = radix_tree_deref_slot(slot);
1836                 if (unlikely(!page))
1837                         continue;
1838                 if (radix_tree_exception(page)) {
1839                         if (radix_tree_deref_retry(page)) {
1840                                 slot = radix_tree_iter_retry(&iter);
1841                                 continue;
1842                         }
1843
1844                         /*
1845                          * A shadow entry of a recently evicted page, a swap
1846                          * entry from shmem/tmpfs or a DAX entry.  Return it
1847                          * without attempting to raise page count.
1848                          */
1849                         goto export;
1850                 }
1851
1852                 head = compound_head(page);
1853                 if (!page_cache_get_speculative(head))
1854                         goto repeat;
1855
1856                 /* The page was split under us? */
1857                 if (compound_head(page) != head) {
1858                         put_page(head);
1859                         goto repeat;
1860                 }
1861
1862                 /* Has the page moved? */
1863                 if (unlikely(page != *slot)) {
1864                         put_page(head);
1865                         goto repeat;
1866                 }
1867 export:
1868                 indices[ret] = iter.index;
1869                 entries[ret] = page;
1870                 if (++ret == nr_entries)
1871                         break;
1872         }
1873         rcu_read_unlock();
1874         return ret;
1875 }
1876 EXPORT_SYMBOL(find_get_entries_tag);
1877
1878 /*
1879  * CD/DVDs are error prone. When a medium error occurs, the driver may fail
1880  * a _large_ part of the i/o request. Imagine the worst scenario:
1881  *
1882  *      ---R__________________________________________B__________
1883  *         ^ reading here                             ^ bad block(assume 4k)
1884  *
1885  * read(R) => miss => readahead(R...B) => media error => frustrating retries
1886  * => failing the whole request => read(R) => read(R+1) =>
1887  * readahead(R+1...B+1) => bang => read(R+2) => read(R+3) =>
1888  * readahead(R+3...B+2) => bang => read(R+3) => read(R+4) =>
1889  * readahead(R+4...B+3) => bang => read(R+4) => read(R+5) => ......
1890  *
1891  * It is going insane. Fix it by quickly scaling down the readahead size.
1892  */
1893 static void shrink_readahead_size_eio(struct file *filp,
1894                                         struct file_ra_state *ra)
1895 {
1896         ra->ra_pages /= 4;
1897 }
1898
1899 /**
1900  * do_generic_file_read - generic file read routine
1901  * @filp:       the file to read
1902  * @ppos:       current file position
1903  * @iter:       data destination
1904  * @written:    already copied
1905  *
1906  * This is a generic file read routine, and uses the
1907  * mapping->a_ops->readpage() function for the actual low-level stuff.
1908  *
1909  * This is really ugly. But the goto's actually try to clarify some
1910  * of the logic when it comes to error handling etc.
1911  */
1912 static ssize_t do_generic_file_read(struct file *filp, loff_t *ppos,
1913                 struct iov_iter *iter, ssize_t written)
1914 {
1915         struct address_space *mapping = filp->f_mapping;
1916         struct inode *inode = mapping->host;
1917         struct file_ra_state *ra = &filp->f_ra;
1918         pgoff_t index;
1919         pgoff_t last_index;
1920         pgoff_t prev_index;
1921         unsigned long offset;      /* offset into pagecache page */
1922         unsigned int prev_offset;
1923         int error = 0;
1924
1925         if (unlikely(*ppos >= inode->i_sb->s_maxbytes))
1926                 return 0;
1927         iov_iter_truncate(iter, inode->i_sb->s_maxbytes);
1928
1929         index = *ppos >> PAGE_SHIFT;
1930         prev_index = ra->prev_pos >> PAGE_SHIFT;
1931         prev_offset = ra->prev_pos & (PAGE_SIZE-1);
1932         last_index = (*ppos + iter->count + PAGE_SIZE-1) >> PAGE_SHIFT;
1933         offset = *ppos & ~PAGE_MASK;
1934
1935         for (;;) {
1936                 struct page *page;
1937                 pgoff_t end_index;
1938                 loff_t isize;
1939                 unsigned long nr, ret;
1940
1941                 cond_resched();
1942 find_page:
1943                 if (fatal_signal_pending(current)) {
1944                         error = -EINTR;
1945                         goto out;
1946                 }
1947
1948                 page = find_get_page(mapping, index);
1949                 if (!page) {
1950                         page_cache_sync_readahead(mapping,
1951                                         ra, filp,
1952                                         index, last_index - index);
1953                         page = find_get_page(mapping, index);
1954                         if (unlikely(page == NULL))
1955                                 goto no_cached_page;
1956                 }
1957                 if (PageReadahead(page)) {
1958                         page_cache_async_readahead(mapping,
1959                                         ra, filp, page,
1960                                         index, last_index - index);
1961                 }
1962                 if (!PageUptodate(page)) {
1963                         /*
1964                          * See comment in do_read_cache_page on why
1965                          * wait_on_page_locked is used to avoid unnecessarily
1966                          * serialisations and why it's safe.
1967                          */
1968                         error = wait_on_page_locked_killable(page);
1969                         if (unlikely(error))
1970                                 goto readpage_error;
1971                         if (PageUptodate(page))
1972                                 goto page_ok;
1973
1974                         if (inode->i_blkbits == PAGE_SHIFT ||
1975                                         !mapping->a_ops->is_partially_uptodate)
1976                                 goto page_not_up_to_date;
1977                         /* pipes can't handle partially uptodate pages */
1978                         if (unlikely(iter->type & ITER_PIPE))
1979                                 goto page_not_up_to_date;
1980                         if (!trylock_page(page))
1981                                 goto page_not_up_to_date;
1982                         /* Did it get truncated before we got the lock? */
1983                         if (!page->mapping)
1984                                 goto page_not_up_to_date_locked;
1985                         if (!mapping->a_ops->is_partially_uptodate(page,
1986                                                         offset, iter->count))
1987                                 goto page_not_up_to_date_locked;
1988                         unlock_page(page);
1989                 }
1990 page_ok:
1991                 /*
1992                  * i_size must be checked after we know the page is Uptodate.
1993                  *
1994                  * Checking i_size after the check allows us to calculate
1995                  * the correct value for "nr", which means the zero-filled
1996                  * part of the page is not copied back to userspace (unless
1997                  * another truncate extends the file - this is desired though).
1998                  */
1999
2000                 isize = i_size_read(inode);
2001                 end_index = (isize - 1) >> PAGE_SHIFT;
2002                 if (unlikely(!isize || index > end_index)) {
2003                         put_page(page);
2004                         goto out;
2005                 }
2006
2007                 /* nr is the maximum number of bytes to copy from this page */
2008                 nr = PAGE_SIZE;
2009                 if (index == end_index) {
2010                         nr = ((isize - 1) & ~PAGE_MASK) + 1;
2011                         if (nr <= offset) {
2012                                 put_page(page);
2013                                 goto out;
2014                         }
2015                 }
2016                 nr = nr - offset;
2017
2018                 /* If users can be writing to this page using arbitrary
2019                  * virtual addresses, take care about potential aliasing
2020                  * before reading the page on the kernel side.
2021                  */
2022                 if (mapping_writably_mapped(mapping))
2023                         flush_dcache_page(page);
2024
2025                 /*
2026                  * When a sequential read accesses a page several times,
2027                  * only mark it as accessed the first time.
2028                  */
2029                 if (prev_index != index || offset != prev_offset)
2030                         mark_page_accessed(page);
2031                 prev_index = index;
2032
2033                 /*
2034                  * Ok, we have the page, and it's up-to-date, so
2035                  * now we can copy it to user space...
2036                  */
2037
2038                 ret = copy_page_to_iter(page, offset, nr, iter);
2039                 offset += ret;
2040                 index += offset >> PAGE_SHIFT;
2041                 offset &= ~PAGE_MASK;
2042                 prev_offset = offset;
2043
2044                 put_page(page);
2045                 written += ret;
2046                 if (!iov_iter_count(iter))
2047                         goto out;
2048                 if (ret < nr) {
2049                         error = -EFAULT;
2050                         goto out;
2051                 }
2052                 continue;
2053
2054 page_not_up_to_date:
2055                 /* Get exclusive access to the page ... */
2056                 error = lock_page_killable(page);
2057                 if (unlikely(error))
2058                         goto readpage_error;
2059
2060 page_not_up_to_date_locked:
2061                 /* Did it get truncated before we got the lock? */
2062                 if (!page->mapping) {
2063                         unlock_page(page);
2064                         put_page(page);
2065                         continue;
2066                 }
2067
2068                 /* Did somebody else fill it already? */
2069                 if (PageUptodate(page)) {
2070                         unlock_page(page);
2071                         goto page_ok;
2072                 }
2073
2074 readpage:
2075                 /*
2076                  * A previous I/O error may have been due to temporary
2077                  * failures, eg. multipath errors.
2078                  * PG_error will be set again if readpage fails.
2079                  */
2080                 ClearPageError(page);
2081                 /* Start the actual read. The read will unlock the page. */
2082                 error = mapping->a_ops->readpage(filp, page);
2083
2084                 if (unlikely(error)) {
2085                         if (error == AOP_TRUNCATED_PAGE) {
2086                                 put_page(page);
2087                                 error = 0;
2088                                 goto find_page;
2089                         }
2090                         goto readpage_error;
2091                 }
2092
2093                 if (!PageUptodate(page)) {
2094                         error = lock_page_killable(page);
2095                         if (unlikely(error))
2096                                 goto readpage_error;
2097                         if (!PageUptodate(page)) {
2098                                 if (page->mapping == NULL) {
2099                                         /*
2100                                          * invalidate_mapping_pages got it
2101                                          */
2102                                         unlock_page(page);
2103                                         put_page(page);
2104                                         goto find_page;
2105                                 }
2106                                 unlock_page(page);
2107                                 shrink_readahead_size_eio(filp, ra);
2108                                 error = -EIO;
2109                                 goto readpage_error;
2110                         }
2111                         unlock_page(page);
2112                 }
2113
2114                 goto page_ok;
2115
2116 readpage_error:
2117                 /* UHHUH! A synchronous read error occurred. Report it */
2118                 put_page(page);
2119                 goto out;
2120
2121 no_cached_page:
2122                 /*
2123                  * Ok, it wasn't cached, so we need to create a new
2124                  * page..
2125                  */
2126                 page = page_cache_alloc_cold(mapping);
2127                 if (!page) {
2128                         error = -ENOMEM;
2129                         goto out;
2130                 }
2131                 error = add_to_page_cache_lru(page, mapping, index,
2132                                 mapping_gfp_constraint(mapping, GFP_KERNEL));
2133                 if (error) {
2134                         put_page(page);
2135                         if (error == -EEXIST) {
2136                                 error = 0;
2137                                 goto find_page;
2138                         }
2139                         goto out;
2140                 }
2141                 goto readpage;
2142         }
2143
2144 out:
2145         ra->prev_pos = prev_index;
2146         ra->prev_pos <<= PAGE_SHIFT;
2147         ra->prev_pos |= prev_offset;
2148
2149         *ppos = ((loff_t)index << PAGE_SHIFT) + offset;
2150         file_accessed(filp);
2151         return written ? written : error;
2152 }
2153
2154 /**
2155  * generic_file_read_iter - generic filesystem read routine
2156  * @iocb:       kernel I/O control block
2157  * @iter:       destination for the data read
2158  *
2159  * This is the "read_iter()" routine for all filesystems
2160  * that can use the page cache directly.
2161  */
2162 ssize_t
2163 generic_file_read_iter(struct kiocb *iocb, struct iov_iter *iter)
2164 {
2165         struct file *file = iocb->ki_filp;
2166         ssize_t retval = 0;
2167         size_t count = iov_iter_count(iter);
2168
2169         if (!count)
2170                 goto out; /* skip atime */
2171
2172         if (iocb->ki_flags & IOCB_DIRECT) {
2173                 struct address_space *mapping = file->f_mapping;
2174                 struct inode *inode = mapping->host;
2175                 loff_t size;
2176
2177                 size = i_size_read(inode);
2178                 if (iocb->ki_flags & IOCB_NOWAIT) {
2179                         if (filemap_range_has_page(mapping, iocb->ki_pos,
2180                                                    iocb->ki_pos + count - 1))
2181                                 return -EAGAIN;
2182                 } else {
2183                         retval = filemap_write_and_wait_range(mapping,
2184                                                 iocb->ki_pos,
2185                                                 iocb->ki_pos + count - 1);
2186                         if (retval < 0)
2187                                 goto out;
2188                 }
2189
2190                 file_accessed(file);
2191
2192                 retval = mapping->a_ops->direct_IO(iocb, iter);
2193                 if (retval >= 0) {
2194                         iocb->ki_pos += retval;
2195                         count -= retval;
2196                 }
2197                 iov_iter_revert(iter, count - iov_iter_count(iter));
2198
2199                 /*
2200                  * Btrfs can have a short DIO read if we encounter
2201                  * compressed extents, so if there was an error, or if
2202                  * we've already read everything we wanted to, or if
2203                  * there was a short read because we hit EOF, go ahead
2204                  * and return.  Otherwise fallthrough to buffered io for
2205                  * the rest of the read.  Buffered reads will not work for
2206                  * DAX files, so don't bother trying.
2207                  */
2208                 if (retval < 0 || !count || iocb->ki_pos >= size ||
2209                     IS_DAX(inode))
2210                         goto out;
2211         }
2212
2213         retval = do_generic_file_read(file, &iocb->ki_pos, iter, retval);
2214 out:
2215         return retval;
2216 }
2217 EXPORT_SYMBOL(generic_file_read_iter);
2218
2219 #ifdef CONFIG_MMU
2220 /**
2221  * page_cache_read - adds requested page to the page cache if not already there
2222  * @file:       file to read
2223  * @offset:     page index
2224  * @gfp_mask:   memory allocation flags
2225  *
2226  * This adds the requested page to the page cache if it isn't already there,
2227  * and schedules an I/O to read in its contents from disk.
2228  */
2229 static int page_cache_read(struct file *file, pgoff_t offset, gfp_t gfp_mask)
2230 {
2231         struct address_space *mapping = file->f_mapping;
2232         struct page *page;
2233         int ret;
2234
2235         do {
2236                 page = __page_cache_alloc(gfp_mask|__GFP_COLD);
2237                 if (!page)
2238                         return -ENOMEM;
2239
2240                 ret = add_to_page_cache_lru(page, mapping, offset, gfp_mask & GFP_KERNEL);
2241                 if (ret == 0)
2242                         ret = mapping->a_ops->readpage(file, page);
2243                 else if (ret == -EEXIST)
2244                         ret = 0; /* losing race to add is OK */
2245
2246                 put_page(page);
2247
2248         } while (ret == AOP_TRUNCATED_PAGE);
2249
2250         return ret;
2251 }
2252
2253 #define MMAP_LOTSAMISS  (100)
2254
2255 /*
2256  * Synchronous readahead happens when we don't even find
2257  * a page in the page cache at all.
2258  */
2259 static void do_sync_mmap_readahead(struct vm_area_struct *vma,
2260                                    struct file_ra_state *ra,
2261                                    struct file *file,
2262                                    pgoff_t offset)
2263 {
2264         struct address_space *mapping = file->f_mapping;
2265
2266         /* If we don't want any read-ahead, don't bother */
2267         if (vma->vm_flags & VM_RAND_READ)
2268                 return;
2269         if (!ra->ra_pages)
2270                 return;
2271
2272         if (vma->vm_flags & VM_SEQ_READ) {
2273                 page_cache_sync_readahead(mapping, ra, file, offset,
2274                                           ra->ra_pages);
2275                 return;
2276         }
2277
2278         /* Avoid banging the cache line if not needed */
2279         if (ra->mmap_miss < MMAP_LOTSAMISS * 10)
2280                 ra->mmap_miss++;
2281
2282         /*
2283          * Do we miss much more than hit in this file? If so,
2284          * stop bothering with read-ahead. It will only hurt.
2285          */
2286         if (ra->mmap_miss > MMAP_LOTSAMISS)
2287                 return;
2288
2289         /*
2290          * mmap read-around
2291          */
2292         ra->start = max_t(long, 0, offset - ra->ra_pages / 2);
2293         ra->size = ra->ra_pages;
2294         ra->async_size = ra->ra_pages / 4;
2295         ra_submit(ra, mapping, file);
2296 }
2297
2298 /*
2299  * Asynchronous readahead happens when we find the page and PG_readahead,
2300  * so we want to possibly extend the readahead further..
2301  */
2302 static void do_async_mmap_readahead(struct vm_area_struct *vma,
2303                                     struct file_ra_state *ra,
2304                                     struct file *file,
2305                                     struct page *page,
2306                                     pgoff_t offset)
2307 {
2308         struct address_space *mapping = file->f_mapping;
2309
2310         /* If we don't want any read-ahead, don't bother */
2311         if (vma->vm_flags & VM_RAND_READ)
2312                 return;
2313         if (ra->mmap_miss > 0)
2314                 ra->mmap_miss--;
2315         if (PageReadahead(page))
2316                 page_cache_async_readahead(mapping, ra, file,
2317                                            page, offset, ra->ra_pages);
2318 }
2319
2320 /**
2321  * filemap_fault - read in file data for page fault handling
2322  * @vmf:        struct vm_fault containing details of the fault
2323  *
2324  * filemap_fault() is invoked via the vma operations vector for a
2325  * mapped memory region to read in file data during a page fault.
2326  *
2327  * The goto's are kind of ugly, but this streamlines the normal case of having
2328  * it in the page cache, and handles the special cases reasonably without
2329  * having a lot of duplicated code.
2330  *
2331  * vma->vm_mm->mmap_sem must be held on entry.
2332  *
2333  * If our return value has VM_FAULT_RETRY set, it's because
2334  * lock_page_or_retry() returned 0.
2335  * The mmap_sem has usually been released in this case.
2336  * See __lock_page_or_retry() for the exception.
2337  *
2338  * If our return value does not have VM_FAULT_RETRY set, the mmap_sem
2339  * has not been released.
2340  *
2341  * We never return with VM_FAULT_RETRY and a bit from VM_FAULT_ERROR set.
2342  */
2343 int filemap_fault(struct vm_fault *vmf)
2344 {
2345         int error;
2346         struct file *file = vmf->vma->vm_file;
2347         struct address_space *mapping = file->f_mapping;
2348         struct file_ra_state *ra = &file->f_ra;
2349         struct inode *inode = mapping->host;
2350         pgoff_t offset = vmf->pgoff;
2351         pgoff_t max_off;
2352         struct page *page;
2353         int ret = 0;
2354
2355         max_off = DIV_ROUND_UP(i_size_read(inode), PAGE_SIZE);
2356         if (unlikely(offset >= max_off))
2357                 return VM_FAULT_SIGBUS;
2358
2359         /*
2360          * Do we have something in the page cache already?
2361          */
2362         page = find_get_page(mapping, offset);
2363         if (likely(page) && !(vmf->flags & FAULT_FLAG_TRIED)) {
2364                 /*
2365                  * We found the page, so try async readahead before
2366                  * waiting for the lock.
2367                  */
2368                 do_async_mmap_readahead(vmf->vma, ra, file, page, offset);
2369         } else if (!page) {
2370                 /* No page in the page cache at all */
2371                 do_sync_mmap_readahead(vmf->vma, ra, file, offset);
2372                 count_vm_event(PGMAJFAULT);
2373                 count_memcg_event_mm(vmf->vma->vm_mm, PGMAJFAULT);
2374                 ret = VM_FAULT_MAJOR;
2375 retry_find:
2376                 page = find_get_page(mapping, offset);
2377                 if (!page)
2378                         goto no_cached_page;
2379         }
2380
2381         if (!lock_page_or_retry(page, vmf->vma->vm_mm, vmf->flags)) {
2382                 put_page(page);
2383                 return ret | VM_FAULT_RETRY;
2384         }
2385
2386         /* Did it get truncated? */
2387         if (unlikely(page->mapping != mapping)) {
2388                 unlock_page(page);
2389                 put_page(page);
2390                 goto retry_find;
2391         }
2392         VM_BUG_ON_PAGE(page->index != offset, page);
2393
2394         /*
2395          * We have a locked page in the page cache, now we need to check
2396          * that it's up-to-date. If not, it is going to be due to an error.
2397          */
2398         if (unlikely(!PageUptodate(page)))
2399                 goto page_not_uptodate;
2400
2401         /*
2402          * Found the page and have a reference on it.
2403          * We must recheck i_size under page lock.
2404          */
2405         max_off = DIV_ROUND_UP(i_size_read(inode), PAGE_SIZE);
2406         if (unlikely(offset >= max_off)) {
2407                 unlock_page(page);
2408                 put_page(page);
2409                 return VM_FAULT_SIGBUS;
2410         }
2411
2412         vmf->page = page;
2413         return ret | VM_FAULT_LOCKED;
2414
2415 no_cached_page:
2416         /*
2417          * We're only likely to ever get here if MADV_RANDOM is in
2418          * effect.
2419          */
2420         error = page_cache_read(file, offset, vmf->gfp_mask);
2421
2422         /*
2423          * The page we want has now been added to the page cache.
2424          * In the unlikely event that someone removed it in the
2425          * meantime, we'll just come back here and read it again.
2426          */
2427         if (error >= 0)
2428                 goto retry_find;
2429
2430         /*
2431          * An error return from page_cache_read can result if the
2432          * system is low on memory, or a problem occurs while trying
2433          * to schedule I/O.
2434          */
2435         if (error == -ENOMEM)
2436                 return VM_FAULT_OOM;
2437         return VM_FAULT_SIGBUS;
2438
2439 page_not_uptodate:
2440         /*
2441          * Umm, take care of errors if the page isn't up-to-date.
2442          * Try to re-read it _once_. We do this synchronously,
2443          * because there really aren't any performance issues here
2444          * and we need to check for errors.
2445          */
2446         ClearPageError(page);
2447         error = mapping->a_ops->readpage(file, page);
2448         if (!error) {
2449                 wait_on_page_locked(page);
2450                 if (!PageUptodate(page))
2451                         error = -EIO;
2452         }
2453         put_page(page);
2454
2455         if (!error || error == AOP_TRUNCATED_PAGE)
2456                 goto retry_find;
2457
2458         /* Things didn't work out. Return zero to tell the mm layer so. */
2459         shrink_readahead_size_eio(file, ra);
2460         return VM_FAULT_SIGBUS;
2461 }
2462 EXPORT_SYMBOL(filemap_fault);
2463
2464 void filemap_map_pages(struct vm_fault *vmf,
2465                 pgoff_t start_pgoff, pgoff_t end_pgoff)
2466 {
2467         struct radix_tree_iter iter;
2468         void **slot;
2469         struct file *file = vmf->vma->vm_file;
2470         struct address_space *mapping = file->f_mapping;
2471         pgoff_t last_pgoff = start_pgoff;
2472         unsigned long max_idx;
2473         struct page *head, *page;
2474
2475         rcu_read_lock();
2476         radix_tree_for_each_slot(slot, &mapping->page_tree, &iter,
2477                         start_pgoff) {
2478                 if (iter.index > end_pgoff)
2479                         break;
2480 repeat:
2481                 page = radix_tree_deref_slot(slot);
2482                 if (unlikely(!page))
2483                         goto next;
2484                 if (radix_tree_exception(page)) {
2485                         if (radix_tree_deref_retry(page)) {
2486                                 slot = radix_tree_iter_retry(&iter);
2487                                 continue;
2488                         }
2489                         goto next;
2490                 }
2491
2492                 head = compound_head(page);
2493                 if (!page_cache_get_speculative(head))
2494                         goto repeat;
2495
2496                 /* The page was split under us? */
2497                 if (compound_head(page) != head) {
2498                         put_page(head);
2499                         goto repeat;
2500                 }
2501
2502                 /* Has the page moved? */
2503                 if (unlikely(page != *slot)) {
2504                         put_page(head);
2505                         goto repeat;
2506                 }
2507
2508                 if (!PageUptodate(page) ||
2509                                 PageReadahead(page) ||
2510                                 PageHWPoison(page))
2511                         goto skip;
2512                 if (!trylock_page(page))
2513                         goto skip;
2514
2515                 if (page->mapping != mapping || !PageUptodate(page))
2516                         goto unlock;
2517
2518                 max_idx = DIV_ROUND_UP(i_size_read(mapping->host), PAGE_SIZE);
2519                 if (page->index >= max_idx)
2520                         goto unlock;
2521
2522                 if (file->f_ra.mmap_miss > 0)
2523                         file->f_ra.mmap_miss--;
2524
2525                 vmf->address += (iter.index - last_pgoff) << PAGE_SHIFT;
2526                 if (vmf->pte)
2527                         vmf->pte += iter.index - last_pgoff;
2528                 last_pgoff = iter.index;
2529                 if (alloc_set_pte(vmf, NULL, page))
2530                         goto unlock;
2531                 unlock_page(page);
2532                 goto next;
2533 unlock:
2534                 unlock_page(page);
2535 skip:
2536                 put_page(page);
2537 next:
2538                 /* Huge page is mapped? No need to proceed. */
2539                 if (pmd_trans_huge(*vmf->pmd))
2540                         break;
2541                 if (iter.index == end_pgoff)
2542                         break;
2543         }
2544         rcu_read_unlock();
2545 }
2546 EXPORT_SYMBOL(filemap_map_pages);
2547
2548 int filemap_page_mkwrite(struct vm_fault *vmf)
2549 {
2550         struct page *page = vmf->page;
2551         struct inode *inode = file_inode(vmf->vma->vm_file);
2552         int ret = VM_FAULT_LOCKED;
2553
2554         sb_start_pagefault(inode->i_sb);
2555         file_update_time(vmf->vma->vm_file);
2556         lock_page(page);
2557         if (page->mapping != inode->i_mapping) {
2558                 unlock_page(page);
2559                 ret = VM_FAULT_NOPAGE;
2560                 goto out;
2561         }
2562         /*
2563          * We mark the page dirty already here so that when freeze is in
2564          * progress, we are guaranteed that writeback during freezing will
2565          * see the dirty page and writeprotect it again.
2566          */
2567         set_page_dirty(page);
2568         wait_for_stable_page(page);
2569 out:
2570         sb_end_pagefault(inode->i_sb);
2571         return ret;
2572 }
2573 EXPORT_SYMBOL(filemap_page_mkwrite);
2574
2575 const struct vm_operations_struct generic_file_vm_ops = {
2576         .fault          = filemap_fault,
2577         .map_pages      = filemap_map_pages,
2578         .page_mkwrite   = filemap_page_mkwrite,
2579 };
2580
2581 /* This is used for a general mmap of a disk file */
2582
2583 int generic_file_mmap(struct file * file, struct vm_area_struct * vma)
2584 {
2585         struct address_space *mapping = file->f_mapping;
2586
2587         if (!mapping->a_ops->readpage)
2588                 return -ENOEXEC;
2589         file_accessed(file);
2590         vma->vm_ops = &generic_file_vm_ops;
2591         return 0;
2592 }
2593
2594 /*
2595  * This is for filesystems which do not implement ->writepage.
2596  */
2597 int generic_file_readonly_mmap(struct file *file, struct vm_area_struct *vma)
2598 {
2599         if ((vma->vm_flags & VM_SHARED) && (vma->vm_flags & VM_MAYWRITE))
2600                 return -EINVAL;
2601         return generic_file_mmap(file, vma);
2602 }
2603 #else
2604 int generic_file_mmap(struct file * file, struct vm_area_struct * vma)
2605 {
2606         return -ENOSYS;
2607 }
2608 int generic_file_readonly_mmap(struct file * file, struct vm_area_struct * vma)
2609 {
2610         return -ENOSYS;
2611 }
2612 #endif /* CONFIG_MMU */
2613
2614 EXPORT_SYMBOL(generic_file_mmap);
2615 EXPORT_SYMBOL(generic_file_readonly_mmap);
2616
2617 static struct page *wait_on_page_read(struct page *page)
2618 {
2619         if (!IS_ERR(page)) {
2620                 wait_on_page_locked(page);
2621                 if (!PageUptodate(page)) {
2622                         put_page(page);
2623                         page = ERR_PTR(-EIO);
2624                 }
2625         }
2626         return page;
2627 }
2628
2629 static struct page *do_read_cache_page(struct address_space *mapping,
2630                                 pgoff_t index,
2631                                 int (*filler)(void *, struct page *),
2632                                 void *data,
2633                                 gfp_t gfp)
2634 {
2635         struct page *page;
2636         int err;
2637 repeat:
2638         page = find_get_page(mapping, index);
2639         if (!page) {
2640                 page = __page_cache_alloc(gfp | __GFP_COLD);
2641                 if (!page)
2642                         return ERR_PTR(-ENOMEM);
2643                 err = add_to_page_cache_lru(page, mapping, index, gfp);
2644                 if (unlikely(err)) {
2645                         put_page(page);
2646                         if (err == -EEXIST)
2647                                 goto repeat;
2648                         /* Presumably ENOMEM for radix tree node */
2649                         return ERR_PTR(err);
2650                 }
2651
2652 filler:
2653                 err = filler(data, page);
2654                 if (err < 0) {
2655                         put_page(page);
2656                         return ERR_PTR(err);
2657                 }
2658
2659                 page = wait_on_page_read(page);
2660                 if (IS_ERR(page))
2661                         return page;
2662                 goto out;
2663         }
2664         if (PageUptodate(page))
2665                 goto out;
2666
2667         /*
2668          * Page is not up to date and may be locked due one of the following
2669          * case a: Page is being filled and the page lock is held
2670          * case b: Read/write error clearing the page uptodate status
2671          * case c: Truncation in progress (page locked)
2672          * case d: Reclaim in progress
2673          *
2674          * Case a, the page will be up to date when the page is unlocked.
2675          *    There is no need to serialise on the page lock here as the page
2676          *    is pinned so the lock gives no additional protection. Even if the
2677          *    the page is truncated, the data is still valid if PageUptodate as
2678          *    it's a race vs truncate race.
2679          * Case b, the page will not be up to date
2680          * Case c, the page may be truncated but in itself, the data may still
2681          *    be valid after IO completes as it's a read vs truncate race. The
2682          *    operation must restart if the page is not uptodate on unlock but
2683          *    otherwise serialising on page lock to stabilise the mapping gives
2684          *    no additional guarantees to the caller as the page lock is
2685          *    released before return.
2686          * Case d, similar to truncation. If reclaim holds the page lock, it
2687          *    will be a race with remove_mapping that determines if the mapping
2688          *    is valid on unlock but otherwise the data is valid and there is
2689          *    no need to serialise with page lock.
2690          *
2691          * As the page lock gives no additional guarantee, we optimistically
2692          * wait on the page to be unlocked and check if it's up to date and
2693          * use the page if it is. Otherwise, the page lock is required to
2694          * distinguish between the different cases. The motivation is that we
2695          * avoid spurious serialisations and wakeups when multiple processes
2696          * wait on the same page for IO to complete.
2697          */
2698         wait_on_page_locked(page);
2699         if (PageUptodate(page))
2700                 goto out;
2701
2702         /* Distinguish between all the cases under the safety of the lock */
2703         lock_page(page);
2704
2705         /* Case c or d, restart the operation */
2706         if (!page->mapping) {
2707                 unlock_page(page);
2708                 put_page(page);
2709                 goto repeat;
2710         }
2711
2712         /* Someone else locked and filled the page in a very small window */
2713         if (PageUptodate(page)) {
2714                 unlock_page(page);
2715                 goto out;
2716         }
2717         goto filler;
2718
2719 out:
2720         mark_page_accessed(page);
2721         return page;
2722 }
2723
2724 /**
2725  * read_cache_page - read into page cache, fill it if needed
2726  * @mapping:    the page's address_space
2727  * @index:      the page index
2728  * @filler:     function to perform the read
2729  * @data:       first arg to filler(data, page) function, often left as NULL
2730  *
2731  * Read into the page cache. If a page already exists, and PageUptodate() is
2732  * not set, try to fill the page and wait for it to become unlocked.
2733  *
2734  * If the page does not get brought uptodate, return -EIO.
2735  */
2736 struct page *read_cache_page(struct address_space *mapping,
2737                                 pgoff_t index,
2738                                 int (*filler)(void *, struct page *),
2739                                 void *data)
2740 {
2741         return do_read_cache_page(mapping, index, filler, data, mapping_gfp_mask(mapping));
2742 }
2743 EXPORT_SYMBOL(read_cache_page);
2744
2745 /**
2746  * read_cache_page_gfp - read into page cache, using specified page allocation flags.
2747  * @mapping:    the page's address_space
2748  * @index:      the page index
2749  * @gfp:        the page allocator flags to use if allocating
2750  *
2751  * This is the same as "read_mapping_page(mapping, index, NULL)", but with
2752  * any new page allocations done using the specified allocation flags.
2753  *
2754  * If the page does not get brought uptodate, return -EIO.
2755  */
2756 struct page *read_cache_page_gfp(struct address_space *mapping,
2757                                 pgoff_t index,
2758                                 gfp_t gfp)
2759 {
2760         filler_t *filler = (filler_t *)mapping->a_ops->readpage;
2761
2762         return do_read_cache_page(mapping, index, filler, NULL, gfp);
2763 }
2764 EXPORT_SYMBOL(read_cache_page_gfp);
2765
2766 /*
2767  * Performs necessary checks before doing a write
2768  *
2769  * Can adjust writing position or amount of bytes to write.
2770  * Returns appropriate error code that caller should return or
2771  * zero in case that write should be allowed.
2772  */
2773 inline ssize_t generic_write_checks(struct kiocb *iocb, struct iov_iter *from)
2774 {
2775         struct file *file = iocb->ki_filp;
2776         struct inode *inode = file->f_mapping->host;
2777         unsigned long limit = rlimit(RLIMIT_FSIZE);
2778         loff_t pos;
2779
2780         if (!iov_iter_count(from))
2781                 return 0;
2782
2783         /* FIXME: this is for backwards compatibility with 2.4 */
2784         if (iocb->ki_flags & IOCB_APPEND)
2785                 iocb->ki_pos = i_size_read(inode);
2786
2787         pos = iocb->ki_pos;
2788
2789         if ((iocb->ki_flags & IOCB_NOWAIT) && !(iocb->ki_flags & IOCB_DIRECT))
2790                 return -EINVAL;
2791
2792         if (limit != RLIM_INFINITY) {
2793                 if (iocb->ki_pos >= limit) {
2794                         send_sig(SIGXFSZ, current, 0);
2795                         return -EFBIG;
2796                 }
2797                 iov_iter_truncate(from, limit - (unsigned long)pos);
2798         }
2799
2800         /*
2801          * LFS rule
2802          */
2803         if (unlikely(pos + iov_iter_count(from) > MAX_NON_LFS &&
2804                                 !(file->f_flags & O_LARGEFILE))) {
2805                 if (pos >= MAX_NON_LFS)
2806                         return -EFBIG;
2807                 iov_iter_truncate(from, MAX_NON_LFS - (unsigned long)pos);
2808         }
2809
2810         /*
2811          * Are we about to exceed the fs block limit ?
2812          *
2813          * If we have written data it becomes a short write.  If we have
2814          * exceeded without writing data we send a signal and return EFBIG.
2815          * Linus frestrict idea will clean these up nicely..
2816          */
2817         if (unlikely(pos >= inode->i_sb->s_maxbytes))
2818                 return -EFBIG;
2819
2820         iov_iter_truncate(from, inode->i_sb->s_maxbytes - pos);
2821         return iov_iter_count(from);
2822 }
2823 EXPORT_SYMBOL(generic_write_checks);
2824
2825 int pagecache_write_begin(struct file *file, struct address_space *mapping,
2826                                 loff_t pos, unsigned len, unsigned flags,
2827                                 struct page **pagep, void **fsdata)
2828 {
2829         const struct address_space_operations *aops = mapping->a_ops;
2830
2831         return aops->write_begin(file, mapping, pos, len, flags,
2832                                                         pagep, fsdata);
2833 }
2834 EXPORT_SYMBOL(pagecache_write_begin);
2835
2836 int pagecache_write_end(struct file *file, struct address_space *mapping,
2837                                 loff_t pos, unsigned len, unsigned copied,
2838                                 struct page *page, void *fsdata)
2839 {
2840         const struct address_space_operations *aops = mapping->a_ops;
2841
2842         return aops->write_end(file, mapping, pos, len, copied, page, fsdata);
2843 }
2844 EXPORT_SYMBOL(pagecache_write_end);
2845
2846 ssize_t
2847 generic_file_direct_write(struct kiocb *iocb, struct iov_iter *from)
2848 {
2849         struct file     *file = iocb->ki_filp;
2850         struct address_space *mapping = file->f_mapping;
2851         struct inode    *inode = mapping->host;
2852         loff_t          pos = iocb->ki_pos;
2853         ssize_t         written;
2854         size_t          write_len;
2855         pgoff_t         end;
2856
2857         write_len = iov_iter_count(from);
2858         end = (pos + write_len - 1) >> PAGE_SHIFT;
2859
2860         if (iocb->ki_flags & IOCB_NOWAIT) {
2861                 /* If there are pages to writeback, return */
2862                 if (filemap_range_has_page(inode->i_mapping, pos,
2863                                            pos + iov_iter_count(from)))
2864                         return -EAGAIN;
2865         } else {
2866                 written = filemap_write_and_wait_range(mapping, pos,
2867                                                         pos + write_len - 1);
2868                 if (written)
2869                         goto out;
2870         }
2871
2872         /*
2873          * After a write we want buffered reads to be sure to go to disk to get
2874          * the new data.  We invalidate clean cached page from the region we're
2875          * about to write.  We do this *before* the write so that we can return
2876          * without clobbering -EIOCBQUEUED from ->direct_IO().
2877          */
2878         written = invalidate_inode_pages2_range(mapping,
2879                                         pos >> PAGE_SHIFT, end);
2880         /*
2881          * If a page can not be invalidated, return 0 to fall back
2882          * to buffered write.
2883          */
2884         if (written) {
2885                 if (written == -EBUSY)
2886                         return 0;
2887                 goto out;
2888         }
2889
2890         written = mapping->a_ops->direct_IO(iocb, from);
2891
2892         /*
2893          * Finally, try again to invalidate clean pages which might have been
2894          * cached by non-direct readahead, or faulted in by get_user_pages()
2895          * if the source of the write was an mmap'ed region of the file
2896          * we're writing.  Either one is a pretty crazy thing to do,
2897          * so we don't support it 100%.  If this invalidation
2898          * fails, tough, the write still worked...
2899          */
2900         invalidate_inode_pages2_range(mapping,
2901                                 pos >> PAGE_SHIFT, end);
2902
2903         if (written > 0) {
2904                 pos += written;
2905                 write_len -= written;
2906                 if (pos > i_size_read(inode) && !S_ISBLK(inode->i_mode)) {
2907                         i_size_write(inode, pos);
2908                         mark_inode_dirty(inode);
2909                 }
2910                 iocb->ki_pos = pos;
2911         }
2912         iov_iter_revert(from, write_len - iov_iter_count(from));
2913 out:
2914         return written;
2915 }
2916 EXPORT_SYMBOL(generic_file_direct_write);
2917
2918 /*
2919  * Find or create a page at the given pagecache position. Return the locked
2920  * page. This function is specifically for buffered writes.
2921  */
2922 struct page *grab_cache_page_write_begin(struct address_space *mapping,
2923                                         pgoff_t index, unsigned flags)
2924 {
2925         struct page *page;
2926         int fgp_flags = FGP_LOCK|FGP_WRITE|FGP_CREAT;
2927
2928         if (flags & AOP_FLAG_NOFS)
2929                 fgp_flags |= FGP_NOFS;
2930
2931         page = pagecache_get_page(mapping, index, fgp_flags,
2932                         mapping_gfp_mask(mapping));
2933         if (page)
2934                 wait_for_stable_page(page);
2935
2936         return page;
2937 }
2938 EXPORT_SYMBOL(grab_cache_page_write_begin);
2939
2940 ssize_t generic_perform_write(struct file *file,
2941                                 struct iov_iter *i, loff_t pos)
2942 {
2943         struct address_space *mapping = file->f_mapping;
2944         const struct address_space_operations *a_ops = mapping->a_ops;
2945         long status = 0;
2946         ssize_t written = 0;
2947         unsigned int flags = 0;
2948
2949         do {
2950                 struct page *page;
2951                 unsigned long offset;   /* Offset into pagecache page */
2952                 unsigned long bytes;    /* Bytes to write to page */
2953                 size_t copied;          /* Bytes copied from user */
2954                 void *fsdata;
2955
2956                 offset = (pos & (PAGE_SIZE - 1));
2957                 bytes = min_t(unsigned long, PAGE_SIZE - offset,
2958                                                 iov_iter_count(i));
2959
2960 again:
2961                 /*
2962                  * Bring in the user page that we will copy from _first_.
2963                  * Otherwise there's a nasty deadlock on copying from the
2964                  * same page as we're writing to, without it being marked
2965                  * up-to-date.
2966                  *
2967                  * Not only is this an optimisation, but it is also required
2968                  * to check that the address is actually valid, when atomic
2969                  * usercopies are used, below.
2970                  */
2971                 if (unlikely(iov_iter_fault_in_readable(i, bytes))) {
2972                         status = -EFAULT;
2973                         break;
2974                 }
2975
2976                 if (fatal_signal_pending(current)) {
2977                         status = -EINTR;
2978                         break;
2979                 }
2980
2981                 status = a_ops->write_begin(file, mapping, pos, bytes, flags,
2982                                                 &page, &fsdata);
2983                 if (unlikely(status < 0))
2984                         break;
2985
2986                 if (mapping_writably_mapped(mapping))
2987                         flush_dcache_page(page);
2988
2989                 copied = iov_iter_copy_from_user_atomic(page, i, offset, bytes);
2990                 flush_dcache_page(page);
2991
2992                 status = a_ops->write_end(file, mapping, pos, bytes, copied,
2993                                                 page, fsdata);
2994                 if (unlikely(status < 0))
2995                         break;
2996                 copied = status;
2997
2998                 cond_resched();
2999
3000                 iov_iter_advance(i, copied);
3001                 if (unlikely(copied == 0)) {
3002                         /*
3003                          * If we were unable to copy any data at all, we must
3004                          * fall back to a single segment length write.
3005                          *
3006                          * If we didn't fallback here, we could livelock
3007                          * because not all segments in the iov can be copied at
3008                          * once without a pagefault.
3009                          */
3010                         bytes = min_t(unsigned long, PAGE_SIZE - offset,
3011                                                 iov_iter_single_seg_count(i));
3012                         goto again;
3013                 }
3014                 pos += copied;
3015                 written += copied;
3016
3017                 balance_dirty_pages_ratelimited(mapping);
3018         } while (iov_iter_count(i));
3019
3020         return written ? written : status;
3021 }
3022 EXPORT_SYMBOL(generic_perform_write);
3023
3024 /**
3025  * __generic_file_write_iter - write data to a file
3026  * @iocb:       IO state structure (file, offset, etc.)
3027  * @from:       iov_iter with data to write
3028  *
3029  * This function does all the work needed for actually writing data to a
3030  * file. It does all basic checks, removes SUID from the file, updates
3031  * modification times and calls proper subroutines depending on whether we
3032  * do direct IO or a standard buffered write.
3033  *
3034  * It expects i_mutex to be grabbed unless we work on a block device or similar
3035  * object which does not need locking at all.
3036  *
3037  * This function does *not* take care of syncing data in case of O_SYNC write.
3038  * A caller has to handle it. This is mainly due to the fact that we want to
3039  * avoid syncing under i_mutex.
3040  */
3041 ssize_t __generic_file_write_iter(struct kiocb *iocb, struct iov_iter *from)
3042 {
3043         struct file *file = iocb->ki_filp;
3044         struct address_space * mapping = file->f_mapping;
3045         struct inode    *inode = mapping->host;
3046         ssize_t         written = 0;
3047         ssize_t         err;
3048         ssize_t         status;
3049
3050         /* We can write back this queue in page reclaim */
3051         current->backing_dev_info = inode_to_bdi(inode);
3052         err = file_remove_privs(file);
3053         if (err)
3054                 goto out;
3055
3056         err = file_update_time(file);
3057         if (err)
3058                 goto out;
3059
3060         if (iocb->ki_flags & IOCB_DIRECT) {
3061                 loff_t pos, endbyte;
3062
3063                 written = generic_file_direct_write(iocb, from);
3064                 /*
3065                  * If the write stopped short of completing, fall back to
3066                  * buffered writes.  Some filesystems do this for writes to
3067                  * holes, for example.  For DAX files, a buffered write will
3068                  * not succeed (even if it did, DAX does not handle dirty
3069                  * page-cache pages correctly).
3070                  */
3071                 if (written < 0 || !iov_iter_count(from) || IS_DAX(inode))
3072                         goto out;
3073
3074                 status = generic_perform_write(file, from, pos = iocb->ki_pos);
3075                 /*
3076                  * If generic_perform_write() returned a synchronous error
3077                  * then we want to return the number of bytes which were
3078                  * direct-written, or the error code if that was zero.  Note
3079                  * that this differs from normal direct-io semantics, which
3080                  * will return -EFOO even if some bytes were written.
3081                  */
3082                 if (unlikely(status < 0)) {
3083                         err = status;
3084                         goto out;
3085                 }
3086                 /*
3087                  * We need to ensure that the page cache pages are written to
3088                  * disk and invalidated to preserve the expected O_DIRECT
3089                  * semantics.
3090                  */
3091                 endbyte = pos + status - 1;
3092                 err = filemap_write_and_wait_range(mapping, pos, endbyte);
3093                 if (err == 0) {
3094                         iocb->ki_pos = endbyte + 1;
3095                         written += status;
3096                         invalidate_mapping_pages(mapping,
3097                                                  pos >> PAGE_SHIFT,
3098                                                  endbyte >> PAGE_SHIFT);
3099                 } else {
3100                         /*
3101                          * We don't know how much we wrote, so just return
3102                          * the number of bytes which were direct-written
3103                          */
3104                 }
3105         } else {
3106                 written = generic_perform_write(file, from, iocb->ki_pos);
3107                 if (likely(written > 0))
3108                         iocb->ki_pos += written;
3109         }
3110 out:
3111         current->backing_dev_info = NULL;
3112         return written ? written : err;
3113 }
3114 EXPORT_SYMBOL(__generic_file_write_iter);
3115
3116 /**
3117  * generic_file_write_iter - write data to a file
3118  * @iocb:       IO state structure
3119  * @from:       iov_iter with data to write
3120  *
3121  * This is a wrapper around __generic_file_write_iter() to be used by most
3122  * filesystems. It takes care of syncing the file in case of O_SYNC file
3123  * and acquires i_mutex as needed.
3124  */
3125 ssize_t generic_file_write_iter(struct kiocb *iocb, struct iov_iter *from)
3126 {
3127         struct file *file = iocb->ki_filp;
3128         struct inode *inode = file->f_mapping->host;
3129         ssize_t ret;
3130
3131         inode_lock(inode);
3132         ret = generic_write_checks(iocb, from);
3133         if (ret > 0)
3134                 ret = __generic_file_write_iter(iocb, from);
3135         inode_unlock(inode);
3136
3137         if (ret > 0)
3138                 ret = generic_write_sync(iocb, ret);
3139         return ret;
3140 }
3141 EXPORT_SYMBOL(generic_file_write_iter);
3142
3143 /**
3144  * try_to_release_page() - release old fs-specific metadata on a page
3145  *
3146  * @page: the page which the kernel is trying to free
3147  * @gfp_mask: memory allocation flags (and I/O mode)
3148  *
3149  * The address_space is to try to release any data against the page
3150  * (presumably at page->private).  If the release was successful, return '1'.
3151  * Otherwise return zero.
3152  *
3153  * This may also be called if PG_fscache is set on a page, indicating that the
3154  * page is known to the local caching routines.
3155  *
3156  * The @gfp_mask argument specifies whether I/O may be performed to release
3157  * this page (__GFP_IO), and whether the call may block (__GFP_RECLAIM & __GFP_FS).
3158  *
3159  */
3160 int try_to_release_page(struct page *page, gfp_t gfp_mask)
3161 {
3162         struct address_space * const mapping = page->mapping;
3163
3164         BUG_ON(!PageLocked(page));
3165         if (PageWriteback(page))
3166                 return 0;
3167
3168         if (mapping && mapping->a_ops->releasepage)
3169                 return mapping->a_ops->releasepage(page, gfp_mask);
3170         return try_to_free_buffers(page);
3171 }
3172
3173 EXPORT_SYMBOL(try_to_release_page);