proc: try to remove use of FOLL_FORCE entirely
[muen/linux.git] / mm / filemap.c
1 /*
2  *      linux/mm/filemap.c
3  *
4  * Copyright (C) 1994-1999  Linus Torvalds
5  */
6
7 /*
8  * This file handles the generic file mmap semantics used by
9  * most "normal" filesystems (but you don't /have/ to use this:
10  * the NFS filesystem used to do this differently, for example)
11  */
12 #include <linux/export.h>
13 #include <linux/compiler.h>
14 #include <linux/dax.h>
15 #include <linux/fs.h>
16 #include <linux/sched/signal.h>
17 #include <linux/uaccess.h>
18 #include <linux/capability.h>
19 #include <linux/kernel_stat.h>
20 #include <linux/gfp.h>
21 #include <linux/mm.h>
22 #include <linux/swap.h>
23 #include <linux/mman.h>
24 #include <linux/pagemap.h>
25 #include <linux/file.h>
26 #include <linux/uio.h>
27 #include <linux/hash.h>
28 #include <linux/writeback.h>
29 #include <linux/backing-dev.h>
30 #include <linux/pagevec.h>
31 #include <linux/blkdev.h>
32 #include <linux/security.h>
33 #include <linux/cpuset.h>
34 #include <linux/hardirq.h> /* for BUG_ON(!in_atomic()) only */
35 #include <linux/hugetlb.h>
36 #include <linux/memcontrol.h>
37 #include <linux/cleancache.h>
38 #include <linux/rmap.h>
39 #include "internal.h"
40
41 #define CREATE_TRACE_POINTS
42 #include <trace/events/filemap.h>
43
44 /*
45  * FIXME: remove all knowledge of the buffer layer from the core VM
46  */
47 #include <linux/buffer_head.h> /* for try_to_free_buffers */
48
49 #include <asm/mman.h>
50
51 /*
52  * Shared mappings implemented 30.11.1994. It's not fully working yet,
53  * though.
54  *
55  * Shared mappings now work. 15.8.1995  Bruno.
56  *
57  * finished 'unifying' the page and buffer cache and SMP-threaded the
58  * page-cache, 21.05.1999, Ingo Molnar <mingo@redhat.com>
59  *
60  * SMP-threaded pagemap-LRU 1999, Andrea Arcangeli <andrea@suse.de>
61  */
62
63 /*
64  * Lock ordering:
65  *
66  *  ->i_mmap_rwsem              (truncate_pagecache)
67  *    ->private_lock            (__free_pte->__set_page_dirty_buffers)
68  *      ->swap_lock             (exclusive_swap_page, others)
69  *        ->mapping->tree_lock
70  *
71  *  ->i_mutex
72  *    ->i_mmap_rwsem            (truncate->unmap_mapping_range)
73  *
74  *  ->mmap_sem
75  *    ->i_mmap_rwsem
76  *      ->page_table_lock or pte_lock   (various, mainly in memory.c)
77  *        ->mapping->tree_lock  (arch-dependent flush_dcache_mmap_lock)
78  *
79  *  ->mmap_sem
80  *    ->lock_page               (access_process_vm)
81  *
82  *  ->i_mutex                   (generic_perform_write)
83  *    ->mmap_sem                (fault_in_pages_readable->do_page_fault)
84  *
85  *  bdi->wb.list_lock
86  *    sb_lock                   (fs/fs-writeback.c)
87  *    ->mapping->tree_lock      (__sync_single_inode)
88  *
89  *  ->i_mmap_rwsem
90  *    ->anon_vma.lock           (vma_adjust)
91  *
92  *  ->anon_vma.lock
93  *    ->page_table_lock or pte_lock     (anon_vma_prepare and various)
94  *
95  *  ->page_table_lock or pte_lock
96  *    ->swap_lock               (try_to_unmap_one)
97  *    ->private_lock            (try_to_unmap_one)
98  *    ->tree_lock               (try_to_unmap_one)
99  *    ->zone_lru_lock(zone)     (follow_page->mark_page_accessed)
100  *    ->zone_lru_lock(zone)     (check_pte_range->isolate_lru_page)
101  *    ->private_lock            (page_remove_rmap->set_page_dirty)
102  *    ->tree_lock               (page_remove_rmap->set_page_dirty)
103  *    bdi.wb->list_lock         (page_remove_rmap->set_page_dirty)
104  *    ->inode->i_lock           (page_remove_rmap->set_page_dirty)
105  *    ->memcg->move_lock        (page_remove_rmap->lock_page_memcg)
106  *    bdi.wb->list_lock         (zap_pte_range->set_page_dirty)
107  *    ->inode->i_lock           (zap_pte_range->set_page_dirty)
108  *    ->private_lock            (zap_pte_range->__set_page_dirty_buffers)
109  *
110  * ->i_mmap_rwsem
111  *   ->tasklist_lock            (memory_failure, collect_procs_ao)
112  */
113
114 static int page_cache_tree_insert(struct address_space *mapping,
115                                   struct page *page, void **shadowp)
116 {
117         struct radix_tree_node *node;
118         void **slot;
119         int error;
120
121         error = __radix_tree_create(&mapping->page_tree, page->index, 0,
122                                     &node, &slot);
123         if (error)
124                 return error;
125         if (*slot) {
126                 void *p;
127
128                 p = radix_tree_deref_slot_protected(slot, &mapping->tree_lock);
129                 if (!radix_tree_exceptional_entry(p))
130                         return -EEXIST;
131
132                 mapping->nrexceptional--;
133                 if (!dax_mapping(mapping)) {
134                         if (shadowp)
135                                 *shadowp = p;
136                 } else {
137                         /* DAX can replace empty locked entry with a hole */
138                         WARN_ON_ONCE(p !=
139                                 dax_radix_locked_entry(0, RADIX_DAX_EMPTY));
140                         /* Wakeup waiters for exceptional entry lock */
141                         dax_wake_mapping_entry_waiter(mapping, page->index, p,
142                                                       true);
143                 }
144         }
145         __radix_tree_replace(&mapping->page_tree, node, slot, page,
146                              workingset_update_node, mapping);
147         mapping->nrpages++;
148         return 0;
149 }
150
151 static void page_cache_tree_delete(struct address_space *mapping,
152                                    struct page *page, void *shadow)
153 {
154         int i, nr;
155
156         /* hugetlb pages are represented by one entry in the radix tree */
157         nr = PageHuge(page) ? 1 : hpage_nr_pages(page);
158
159         VM_BUG_ON_PAGE(!PageLocked(page), page);
160         VM_BUG_ON_PAGE(PageTail(page), page);
161         VM_BUG_ON_PAGE(nr != 1 && shadow, page);
162
163         for (i = 0; i < nr; i++) {
164                 struct radix_tree_node *node;
165                 void **slot;
166
167                 __radix_tree_lookup(&mapping->page_tree, page->index + i,
168                                     &node, &slot);
169
170                 VM_BUG_ON_PAGE(!node && nr != 1, page);
171
172                 radix_tree_clear_tags(&mapping->page_tree, node, slot);
173                 __radix_tree_replace(&mapping->page_tree, node, slot, shadow,
174                                      workingset_update_node, mapping);
175         }
176
177         if (shadow) {
178                 mapping->nrexceptional += nr;
179                 /*
180                  * Make sure the nrexceptional update is committed before
181                  * the nrpages update so that final truncate racing
182                  * with reclaim does not see both counters 0 at the
183                  * same time and miss a shadow entry.
184                  */
185                 smp_wmb();
186         }
187         mapping->nrpages -= nr;
188 }
189
190 /*
191  * Delete a page from the page cache and free it. Caller has to make
192  * sure the page is locked and that nobody else uses it - or that usage
193  * is safe.  The caller must hold the mapping's tree_lock.
194  */
195 void __delete_from_page_cache(struct page *page, void *shadow)
196 {
197         struct address_space *mapping = page->mapping;
198         int nr = hpage_nr_pages(page);
199
200         trace_mm_filemap_delete_from_page_cache(page);
201         /*
202          * if we're uptodate, flush out into the cleancache, otherwise
203          * invalidate any existing cleancache entries.  We can't leave
204          * stale data around in the cleancache once our page is gone
205          */
206         if (PageUptodate(page) && PageMappedToDisk(page))
207                 cleancache_put_page(page);
208         else
209                 cleancache_invalidate_page(mapping, page);
210
211         VM_BUG_ON_PAGE(PageTail(page), page);
212         VM_BUG_ON_PAGE(page_mapped(page), page);
213         if (!IS_ENABLED(CONFIG_DEBUG_VM) && unlikely(page_mapped(page))) {
214                 int mapcount;
215
216                 pr_alert("BUG: Bad page cache in process %s  pfn:%05lx\n",
217                          current->comm, page_to_pfn(page));
218                 dump_page(page, "still mapped when deleted");
219                 dump_stack();
220                 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
221
222                 mapcount = page_mapcount(page);
223                 if (mapping_exiting(mapping) &&
224                     page_count(page) >= mapcount + 2) {
225                         /*
226                          * All vmas have already been torn down, so it's
227                          * a good bet that actually the page is unmapped,
228                          * and we'd prefer not to leak it: if we're wrong,
229                          * some other bad page check should catch it later.
230                          */
231                         page_mapcount_reset(page);
232                         page_ref_sub(page, mapcount);
233                 }
234         }
235
236         page_cache_tree_delete(mapping, page, shadow);
237
238         page->mapping = NULL;
239         /* Leave page->index set: truncation lookup relies upon it */
240
241         /* hugetlb pages do not participate in page cache accounting. */
242         if (!PageHuge(page))
243                 __mod_node_page_state(page_pgdat(page), NR_FILE_PAGES, -nr);
244         if (PageSwapBacked(page)) {
245                 __mod_node_page_state(page_pgdat(page), NR_SHMEM, -nr);
246                 if (PageTransHuge(page))
247                         __dec_node_page_state(page, NR_SHMEM_THPS);
248         } else {
249                 VM_BUG_ON_PAGE(PageTransHuge(page) && !PageHuge(page), page);
250         }
251
252         /*
253          * At this point page must be either written or cleaned by truncate.
254          * Dirty page here signals a bug and loss of unwritten data.
255          *
256          * This fixes dirty accounting after removing the page entirely but
257          * leaves PageDirty set: it has no effect for truncated page and
258          * anyway will be cleared before returning page into buddy allocator.
259          */
260         if (WARN_ON_ONCE(PageDirty(page)))
261                 account_page_cleaned(page, mapping, inode_to_wb(mapping->host));
262 }
263
264 /**
265  * delete_from_page_cache - delete page from page cache
266  * @page: the page which the kernel is trying to remove from page cache
267  *
268  * This must be called only on pages that have been verified to be in the page
269  * cache and locked.  It will never put the page into the free list, the caller
270  * has a reference on the page.
271  */
272 void delete_from_page_cache(struct page *page)
273 {
274         struct address_space *mapping = page_mapping(page);
275         unsigned long flags;
276         void (*freepage)(struct page *);
277
278         BUG_ON(!PageLocked(page));
279
280         freepage = mapping->a_ops->freepage;
281
282         spin_lock_irqsave(&mapping->tree_lock, flags);
283         __delete_from_page_cache(page, NULL);
284         spin_unlock_irqrestore(&mapping->tree_lock, flags);
285
286         if (freepage)
287                 freepage(page);
288
289         if (PageTransHuge(page) && !PageHuge(page)) {
290                 page_ref_sub(page, HPAGE_PMD_NR);
291                 VM_BUG_ON_PAGE(page_count(page) <= 0, page);
292         } else {
293                 put_page(page);
294         }
295 }
296 EXPORT_SYMBOL(delete_from_page_cache);
297
298 int filemap_check_errors(struct address_space *mapping)
299 {
300         int ret = 0;
301         /* Check for outstanding write errors */
302         if (test_bit(AS_ENOSPC, &mapping->flags) &&
303             test_and_clear_bit(AS_ENOSPC, &mapping->flags))
304                 ret = -ENOSPC;
305         if (test_bit(AS_EIO, &mapping->flags) &&
306             test_and_clear_bit(AS_EIO, &mapping->flags))
307                 ret = -EIO;
308         return ret;
309 }
310 EXPORT_SYMBOL(filemap_check_errors);
311
312 /**
313  * __filemap_fdatawrite_range - start writeback on mapping dirty pages in range
314  * @mapping:    address space structure to write
315  * @start:      offset in bytes where the range starts
316  * @end:        offset in bytes where the range ends (inclusive)
317  * @sync_mode:  enable synchronous operation
318  *
319  * Start writeback against all of a mapping's dirty pages that lie
320  * within the byte offsets <start, end> inclusive.
321  *
322  * If sync_mode is WB_SYNC_ALL then this is a "data integrity" operation, as
323  * opposed to a regular memory cleansing writeback.  The difference between
324  * these two operations is that if a dirty page/buffer is encountered, it must
325  * be waited upon, and not just skipped over.
326  */
327 int __filemap_fdatawrite_range(struct address_space *mapping, loff_t start,
328                                 loff_t end, int sync_mode)
329 {
330         int ret;
331         struct writeback_control wbc = {
332                 .sync_mode = sync_mode,
333                 .nr_to_write = LONG_MAX,
334                 .range_start = start,
335                 .range_end = end,
336         };
337
338         if (!mapping_cap_writeback_dirty(mapping))
339                 return 0;
340
341         wbc_attach_fdatawrite_inode(&wbc, mapping->host);
342         ret = do_writepages(mapping, &wbc);
343         wbc_detach_inode(&wbc);
344         return ret;
345 }
346
347 static inline int __filemap_fdatawrite(struct address_space *mapping,
348         int sync_mode)
349 {
350         return __filemap_fdatawrite_range(mapping, 0, LLONG_MAX, sync_mode);
351 }
352
353 int filemap_fdatawrite(struct address_space *mapping)
354 {
355         return __filemap_fdatawrite(mapping, WB_SYNC_ALL);
356 }
357 EXPORT_SYMBOL(filemap_fdatawrite);
358
359 int filemap_fdatawrite_range(struct address_space *mapping, loff_t start,
360                                 loff_t end)
361 {
362         return __filemap_fdatawrite_range(mapping, start, end, WB_SYNC_ALL);
363 }
364 EXPORT_SYMBOL(filemap_fdatawrite_range);
365
366 /**
367  * filemap_flush - mostly a non-blocking flush
368  * @mapping:    target address_space
369  *
370  * This is a mostly non-blocking flush.  Not suitable for data-integrity
371  * purposes - I/O may not be started against all dirty pages.
372  */
373 int filemap_flush(struct address_space *mapping)
374 {
375         return __filemap_fdatawrite(mapping, WB_SYNC_NONE);
376 }
377 EXPORT_SYMBOL(filemap_flush);
378
379 static int __filemap_fdatawait_range(struct address_space *mapping,
380                                      loff_t start_byte, loff_t end_byte)
381 {
382         pgoff_t index = start_byte >> PAGE_SHIFT;
383         pgoff_t end = end_byte >> PAGE_SHIFT;
384         struct pagevec pvec;
385         int nr_pages;
386         int ret = 0;
387
388         if (end_byte < start_byte)
389                 goto out;
390
391         pagevec_init(&pvec, 0);
392         while ((index <= end) &&
393                         (nr_pages = pagevec_lookup_tag(&pvec, mapping, &index,
394                         PAGECACHE_TAG_WRITEBACK,
395                         min(end - index, (pgoff_t)PAGEVEC_SIZE-1) + 1)) != 0) {
396                 unsigned i;
397
398                 for (i = 0; i < nr_pages; i++) {
399                         struct page *page = pvec.pages[i];
400
401                         /* until radix tree lookup accepts end_index */
402                         if (page->index > end)
403                                 continue;
404
405                         wait_on_page_writeback(page);
406                         if (TestClearPageError(page))
407                                 ret = -EIO;
408                 }
409                 pagevec_release(&pvec);
410                 cond_resched();
411         }
412 out:
413         return ret;
414 }
415
416 /**
417  * filemap_fdatawait_range - wait for writeback to complete
418  * @mapping:            address space structure to wait for
419  * @start_byte:         offset in bytes where the range starts
420  * @end_byte:           offset in bytes where the range ends (inclusive)
421  *
422  * Walk the list of under-writeback pages of the given address space
423  * in the given range and wait for all of them.  Check error status of
424  * the address space and return it.
425  *
426  * Since the error status of the address space is cleared by this function,
427  * callers are responsible for checking the return value and handling and/or
428  * reporting the error.
429  */
430 int filemap_fdatawait_range(struct address_space *mapping, loff_t start_byte,
431                             loff_t end_byte)
432 {
433         int ret, ret2;
434
435         ret = __filemap_fdatawait_range(mapping, start_byte, end_byte);
436         ret2 = filemap_check_errors(mapping);
437         if (!ret)
438                 ret = ret2;
439
440         return ret;
441 }
442 EXPORT_SYMBOL(filemap_fdatawait_range);
443
444 /**
445  * filemap_fdatawait_keep_errors - wait for writeback without clearing errors
446  * @mapping: address space structure to wait for
447  *
448  * Walk the list of under-writeback pages of the given address space
449  * and wait for all of them.  Unlike filemap_fdatawait(), this function
450  * does not clear error status of the address space.
451  *
452  * Use this function if callers don't handle errors themselves.  Expected
453  * call sites are system-wide / filesystem-wide data flushers: e.g. sync(2),
454  * fsfreeze(8)
455  */
456 void filemap_fdatawait_keep_errors(struct address_space *mapping)
457 {
458         loff_t i_size = i_size_read(mapping->host);
459
460         if (i_size == 0)
461                 return;
462
463         __filemap_fdatawait_range(mapping, 0, i_size - 1);
464 }
465
466 /**
467  * filemap_fdatawait - wait for all under-writeback pages to complete
468  * @mapping: address space structure to wait for
469  *
470  * Walk the list of under-writeback pages of the given address space
471  * and wait for all of them.  Check error status of the address space
472  * and return it.
473  *
474  * Since the error status of the address space is cleared by this function,
475  * callers are responsible for checking the return value and handling and/or
476  * reporting the error.
477  */
478 int filemap_fdatawait(struct address_space *mapping)
479 {
480         loff_t i_size = i_size_read(mapping->host);
481
482         if (i_size == 0)
483                 return 0;
484
485         return filemap_fdatawait_range(mapping, 0, i_size - 1);
486 }
487 EXPORT_SYMBOL(filemap_fdatawait);
488
489 int filemap_write_and_wait(struct address_space *mapping)
490 {
491         int err = 0;
492
493         if ((!dax_mapping(mapping) && mapping->nrpages) ||
494             (dax_mapping(mapping) && mapping->nrexceptional)) {
495                 err = filemap_fdatawrite(mapping);
496                 /*
497                  * Even if the above returned error, the pages may be
498                  * written partially (e.g. -ENOSPC), so we wait for it.
499                  * But the -EIO is special case, it may indicate the worst
500                  * thing (e.g. bug) happened, so we avoid waiting for it.
501                  */
502                 if (err != -EIO) {
503                         int err2 = filemap_fdatawait(mapping);
504                         if (!err)
505                                 err = err2;
506                 }
507         } else {
508                 err = filemap_check_errors(mapping);
509         }
510         return err;
511 }
512 EXPORT_SYMBOL(filemap_write_and_wait);
513
514 /**
515  * filemap_write_and_wait_range - write out & wait on a file range
516  * @mapping:    the address_space for the pages
517  * @lstart:     offset in bytes where the range starts
518  * @lend:       offset in bytes where the range ends (inclusive)
519  *
520  * Write out and wait upon file offsets lstart->lend, inclusive.
521  *
522  * Note that @lend is inclusive (describes the last byte to be written) so
523  * that this function can be used to write to the very end-of-file (end = -1).
524  */
525 int filemap_write_and_wait_range(struct address_space *mapping,
526                                  loff_t lstart, loff_t lend)
527 {
528         int err = 0;
529
530         if ((!dax_mapping(mapping) && mapping->nrpages) ||
531             (dax_mapping(mapping) && mapping->nrexceptional)) {
532                 err = __filemap_fdatawrite_range(mapping, lstart, lend,
533                                                  WB_SYNC_ALL);
534                 /* See comment of filemap_write_and_wait() */
535                 if (err != -EIO) {
536                         int err2 = filemap_fdatawait_range(mapping,
537                                                 lstart, lend);
538                         if (!err)
539                                 err = err2;
540                 }
541         } else {
542                 err = filemap_check_errors(mapping);
543         }
544         return err;
545 }
546 EXPORT_SYMBOL(filemap_write_and_wait_range);
547
548 /**
549  * replace_page_cache_page - replace a pagecache page with a new one
550  * @old:        page to be replaced
551  * @new:        page to replace with
552  * @gfp_mask:   allocation mode
553  *
554  * This function replaces a page in the pagecache with a new one.  On
555  * success it acquires the pagecache reference for the new page and
556  * drops it for the old page.  Both the old and new pages must be
557  * locked.  This function does not add the new page to the LRU, the
558  * caller must do that.
559  *
560  * The remove + add is atomic.  The only way this function can fail is
561  * memory allocation failure.
562  */
563 int replace_page_cache_page(struct page *old, struct page *new, gfp_t gfp_mask)
564 {
565         int error;
566
567         VM_BUG_ON_PAGE(!PageLocked(old), old);
568         VM_BUG_ON_PAGE(!PageLocked(new), new);
569         VM_BUG_ON_PAGE(new->mapping, new);
570
571         error = radix_tree_preload(gfp_mask & ~__GFP_HIGHMEM);
572         if (!error) {
573                 struct address_space *mapping = old->mapping;
574                 void (*freepage)(struct page *);
575                 unsigned long flags;
576
577                 pgoff_t offset = old->index;
578                 freepage = mapping->a_ops->freepage;
579
580                 get_page(new);
581                 new->mapping = mapping;
582                 new->index = offset;
583
584                 spin_lock_irqsave(&mapping->tree_lock, flags);
585                 __delete_from_page_cache(old, NULL);
586                 error = page_cache_tree_insert(mapping, new, NULL);
587                 BUG_ON(error);
588
589                 /*
590                  * hugetlb pages do not participate in page cache accounting.
591                  */
592                 if (!PageHuge(new))
593                         __inc_node_page_state(new, NR_FILE_PAGES);
594                 if (PageSwapBacked(new))
595                         __inc_node_page_state(new, NR_SHMEM);
596                 spin_unlock_irqrestore(&mapping->tree_lock, flags);
597                 mem_cgroup_migrate(old, new);
598                 radix_tree_preload_end();
599                 if (freepage)
600                         freepage(old);
601                 put_page(old);
602         }
603
604         return error;
605 }
606 EXPORT_SYMBOL_GPL(replace_page_cache_page);
607
608 static int __add_to_page_cache_locked(struct page *page,
609                                       struct address_space *mapping,
610                                       pgoff_t offset, gfp_t gfp_mask,
611                                       void **shadowp)
612 {
613         int huge = PageHuge(page);
614         struct mem_cgroup *memcg;
615         int error;
616
617         VM_BUG_ON_PAGE(!PageLocked(page), page);
618         VM_BUG_ON_PAGE(PageSwapBacked(page), page);
619
620         if (!huge) {
621                 error = mem_cgroup_try_charge(page, current->mm,
622                                               gfp_mask, &memcg, false);
623                 if (error)
624                         return error;
625         }
626
627         error = radix_tree_maybe_preload(gfp_mask & ~__GFP_HIGHMEM);
628         if (error) {
629                 if (!huge)
630                         mem_cgroup_cancel_charge(page, memcg, false);
631                 return error;
632         }
633
634         get_page(page);
635         page->mapping = mapping;
636         page->index = offset;
637
638         spin_lock_irq(&mapping->tree_lock);
639         error = page_cache_tree_insert(mapping, page, shadowp);
640         radix_tree_preload_end();
641         if (unlikely(error))
642                 goto err_insert;
643
644         /* hugetlb pages do not participate in page cache accounting. */
645         if (!huge)
646                 __inc_node_page_state(page, NR_FILE_PAGES);
647         spin_unlock_irq(&mapping->tree_lock);
648         if (!huge)
649                 mem_cgroup_commit_charge(page, memcg, false, false);
650         trace_mm_filemap_add_to_page_cache(page);
651         return 0;
652 err_insert:
653         page->mapping = NULL;
654         /* Leave page->index set: truncation relies upon it */
655         spin_unlock_irq(&mapping->tree_lock);
656         if (!huge)
657                 mem_cgroup_cancel_charge(page, memcg, false);
658         put_page(page);
659         return error;
660 }
661
662 /**
663  * add_to_page_cache_locked - add a locked page to the pagecache
664  * @page:       page to add
665  * @mapping:    the page's address_space
666  * @offset:     page index
667  * @gfp_mask:   page allocation mode
668  *
669  * This function is used to add a page to the pagecache. It must be locked.
670  * This function does not add the page to the LRU.  The caller must do that.
671  */
672 int add_to_page_cache_locked(struct page *page, struct address_space *mapping,
673                 pgoff_t offset, gfp_t gfp_mask)
674 {
675         return __add_to_page_cache_locked(page, mapping, offset,
676                                           gfp_mask, NULL);
677 }
678 EXPORT_SYMBOL(add_to_page_cache_locked);
679
680 int add_to_page_cache_lru(struct page *page, struct address_space *mapping,
681                                 pgoff_t offset, gfp_t gfp_mask)
682 {
683         void *shadow = NULL;
684         int ret;
685
686         __SetPageLocked(page);
687         ret = __add_to_page_cache_locked(page, mapping, offset,
688                                          gfp_mask, &shadow);
689         if (unlikely(ret))
690                 __ClearPageLocked(page);
691         else {
692                 /*
693                  * The page might have been evicted from cache only
694                  * recently, in which case it should be activated like
695                  * any other repeatedly accessed page.
696                  * The exception is pages getting rewritten; evicting other
697                  * data from the working set, only to cache data that will
698                  * get overwritten with something else, is a waste of memory.
699                  */
700                 if (!(gfp_mask & __GFP_WRITE) &&
701                     shadow && workingset_refault(shadow)) {
702                         SetPageActive(page);
703                         workingset_activation(page);
704                 } else
705                         ClearPageActive(page);
706                 lru_cache_add(page);
707         }
708         return ret;
709 }
710 EXPORT_SYMBOL_GPL(add_to_page_cache_lru);
711
712 #ifdef CONFIG_NUMA
713 struct page *__page_cache_alloc(gfp_t gfp)
714 {
715         int n;
716         struct page *page;
717
718         if (cpuset_do_page_mem_spread()) {
719                 unsigned int cpuset_mems_cookie;
720                 do {
721                         cpuset_mems_cookie = read_mems_allowed_begin();
722                         n = cpuset_mem_spread_node();
723                         page = __alloc_pages_node(n, gfp, 0);
724                 } while (!page && read_mems_allowed_retry(cpuset_mems_cookie));
725
726                 return page;
727         }
728         return alloc_pages(gfp, 0);
729 }
730 EXPORT_SYMBOL(__page_cache_alloc);
731 #endif
732
733 /*
734  * In order to wait for pages to become available there must be
735  * waitqueues associated with pages. By using a hash table of
736  * waitqueues where the bucket discipline is to maintain all
737  * waiters on the same queue and wake all when any of the pages
738  * become available, and for the woken contexts to check to be
739  * sure the appropriate page became available, this saves space
740  * at a cost of "thundering herd" phenomena during rare hash
741  * collisions.
742  */
743 #define PAGE_WAIT_TABLE_BITS 8
744 #define PAGE_WAIT_TABLE_SIZE (1 << PAGE_WAIT_TABLE_BITS)
745 static wait_queue_head_t page_wait_table[PAGE_WAIT_TABLE_SIZE] __cacheline_aligned;
746
747 static wait_queue_head_t *page_waitqueue(struct page *page)
748 {
749         return &page_wait_table[hash_ptr(page, PAGE_WAIT_TABLE_BITS)];
750 }
751
752 void __init pagecache_init(void)
753 {
754         int i;
755
756         for (i = 0; i < PAGE_WAIT_TABLE_SIZE; i++)
757                 init_waitqueue_head(&page_wait_table[i]);
758
759         page_writeback_init();
760 }
761
762 struct wait_page_key {
763         struct page *page;
764         int bit_nr;
765         int page_match;
766 };
767
768 struct wait_page_queue {
769         struct page *page;
770         int bit_nr;
771         wait_queue_t wait;
772 };
773
774 static int wake_page_function(wait_queue_t *wait, unsigned mode, int sync, void *arg)
775 {
776         struct wait_page_key *key = arg;
777         struct wait_page_queue *wait_page
778                 = container_of(wait, struct wait_page_queue, wait);
779
780         if (wait_page->page != key->page)
781                return 0;
782         key->page_match = 1;
783
784         if (wait_page->bit_nr != key->bit_nr)
785                 return 0;
786         if (test_bit(key->bit_nr, &key->page->flags))
787                 return 0;
788
789         return autoremove_wake_function(wait, mode, sync, key);
790 }
791
792 static void wake_up_page_bit(struct page *page, int bit_nr)
793 {
794         wait_queue_head_t *q = page_waitqueue(page);
795         struct wait_page_key key;
796         unsigned long flags;
797
798         key.page = page;
799         key.bit_nr = bit_nr;
800         key.page_match = 0;
801
802         spin_lock_irqsave(&q->lock, flags);
803         __wake_up_locked_key(q, TASK_NORMAL, &key);
804         /*
805          * It is possible for other pages to have collided on the waitqueue
806          * hash, so in that case check for a page match. That prevents a long-
807          * term waiter
808          *
809          * It is still possible to miss a case here, when we woke page waiters
810          * and removed them from the waitqueue, but there are still other
811          * page waiters.
812          */
813         if (!waitqueue_active(q) || !key.page_match) {
814                 ClearPageWaiters(page);
815                 /*
816                  * It's possible to miss clearing Waiters here, when we woke
817                  * our page waiters, but the hashed waitqueue has waiters for
818                  * other pages on it.
819                  *
820                  * That's okay, it's a rare case. The next waker will clear it.
821                  */
822         }
823         spin_unlock_irqrestore(&q->lock, flags);
824 }
825
826 static void wake_up_page(struct page *page, int bit)
827 {
828         if (!PageWaiters(page))
829                 return;
830         wake_up_page_bit(page, bit);
831 }
832
833 static inline int wait_on_page_bit_common(wait_queue_head_t *q,
834                 struct page *page, int bit_nr, int state, bool lock)
835 {
836         struct wait_page_queue wait_page;
837         wait_queue_t *wait = &wait_page.wait;
838         int ret = 0;
839
840         init_wait(wait);
841         wait->func = wake_page_function;
842         wait_page.page = page;
843         wait_page.bit_nr = bit_nr;
844
845         for (;;) {
846                 spin_lock_irq(&q->lock);
847
848                 if (likely(list_empty(&wait->task_list))) {
849                         if (lock)
850                                 __add_wait_queue_tail_exclusive(q, wait);
851                         else
852                                 __add_wait_queue(q, wait);
853                         SetPageWaiters(page);
854                 }
855
856                 set_current_state(state);
857
858                 spin_unlock_irq(&q->lock);
859
860                 if (likely(test_bit(bit_nr, &page->flags))) {
861                         io_schedule();
862                         if (unlikely(signal_pending_state(state, current))) {
863                                 ret = -EINTR;
864                                 break;
865                         }
866                 }
867
868                 if (lock) {
869                         if (!test_and_set_bit_lock(bit_nr, &page->flags))
870                                 break;
871                 } else {
872                         if (!test_bit(bit_nr, &page->flags))
873                                 break;
874                 }
875         }
876
877         finish_wait(q, wait);
878
879         /*
880          * A signal could leave PageWaiters set. Clearing it here if
881          * !waitqueue_active would be possible (by open-coding finish_wait),
882          * but still fail to catch it in the case of wait hash collision. We
883          * already can fail to clear wait hash collision cases, so don't
884          * bother with signals either.
885          */
886
887         return ret;
888 }
889
890 void wait_on_page_bit(struct page *page, int bit_nr)
891 {
892         wait_queue_head_t *q = page_waitqueue(page);
893         wait_on_page_bit_common(q, page, bit_nr, TASK_UNINTERRUPTIBLE, false);
894 }
895 EXPORT_SYMBOL(wait_on_page_bit);
896
897 int wait_on_page_bit_killable(struct page *page, int bit_nr)
898 {
899         wait_queue_head_t *q = page_waitqueue(page);
900         return wait_on_page_bit_common(q, page, bit_nr, TASK_KILLABLE, false);
901 }
902
903 /**
904  * add_page_wait_queue - Add an arbitrary waiter to a page's wait queue
905  * @page: Page defining the wait queue of interest
906  * @waiter: Waiter to add to the queue
907  *
908  * Add an arbitrary @waiter to the wait queue for the nominated @page.
909  */
910 void add_page_wait_queue(struct page *page, wait_queue_t *waiter)
911 {
912         wait_queue_head_t *q = page_waitqueue(page);
913         unsigned long flags;
914
915         spin_lock_irqsave(&q->lock, flags);
916         __add_wait_queue(q, waiter);
917         SetPageWaiters(page);
918         spin_unlock_irqrestore(&q->lock, flags);
919 }
920 EXPORT_SYMBOL_GPL(add_page_wait_queue);
921
922 #ifndef clear_bit_unlock_is_negative_byte
923
924 /*
925  * PG_waiters is the high bit in the same byte as PG_lock.
926  *
927  * On x86 (and on many other architectures), we can clear PG_lock and
928  * test the sign bit at the same time. But if the architecture does
929  * not support that special operation, we just do this all by hand
930  * instead.
931  *
932  * The read of PG_waiters has to be after (or concurrently with) PG_locked
933  * being cleared, but a memory barrier should be unneccssary since it is
934  * in the same byte as PG_locked.
935  */
936 static inline bool clear_bit_unlock_is_negative_byte(long nr, volatile void *mem)
937 {
938         clear_bit_unlock(nr, mem);
939         /* smp_mb__after_atomic(); */
940         return test_bit(PG_waiters, mem);
941 }
942
943 #endif
944
945 /**
946  * unlock_page - unlock a locked page
947  * @page: the page
948  *
949  * Unlocks the page and wakes up sleepers in ___wait_on_page_locked().
950  * Also wakes sleepers in wait_on_page_writeback() because the wakeup
951  * mechanism between PageLocked pages and PageWriteback pages is shared.
952  * But that's OK - sleepers in wait_on_page_writeback() just go back to sleep.
953  *
954  * Note that this depends on PG_waiters being the sign bit in the byte
955  * that contains PG_locked - thus the BUILD_BUG_ON(). That allows us to
956  * clear the PG_locked bit and test PG_waiters at the same time fairly
957  * portably (architectures that do LL/SC can test any bit, while x86 can
958  * test the sign bit).
959  */
960 void unlock_page(struct page *page)
961 {
962         BUILD_BUG_ON(PG_waiters != 7);
963         page = compound_head(page);
964         VM_BUG_ON_PAGE(!PageLocked(page), page);
965         if (clear_bit_unlock_is_negative_byte(PG_locked, &page->flags))
966                 wake_up_page_bit(page, PG_locked);
967 }
968 EXPORT_SYMBOL(unlock_page);
969
970 /**
971  * end_page_writeback - end writeback against a page
972  * @page: the page
973  */
974 void end_page_writeback(struct page *page)
975 {
976         /*
977          * TestClearPageReclaim could be used here but it is an atomic
978          * operation and overkill in this particular case. Failing to
979          * shuffle a page marked for immediate reclaim is too mild to
980          * justify taking an atomic operation penalty at the end of
981          * ever page writeback.
982          */
983         if (PageReclaim(page)) {
984                 ClearPageReclaim(page);
985                 rotate_reclaimable_page(page);
986         }
987
988         if (!test_clear_page_writeback(page))
989                 BUG();
990
991         smp_mb__after_atomic();
992         wake_up_page(page, PG_writeback);
993 }
994 EXPORT_SYMBOL(end_page_writeback);
995
996 /*
997  * After completing I/O on a page, call this routine to update the page
998  * flags appropriately
999  */
1000 void page_endio(struct page *page, bool is_write, int err)
1001 {
1002         if (!is_write) {
1003                 if (!err) {
1004                         SetPageUptodate(page);
1005                 } else {
1006                         ClearPageUptodate(page);
1007                         SetPageError(page);
1008                 }
1009                 unlock_page(page);
1010         } else {
1011                 if (err) {
1012                         struct address_space *mapping;
1013
1014                         SetPageError(page);
1015                         mapping = page_mapping(page);
1016                         if (mapping)
1017                                 mapping_set_error(mapping, err);
1018                 }
1019                 end_page_writeback(page);
1020         }
1021 }
1022 EXPORT_SYMBOL_GPL(page_endio);
1023
1024 /**
1025  * __lock_page - get a lock on the page, assuming we need to sleep to get it
1026  * @__page: the page to lock
1027  */
1028 void __lock_page(struct page *__page)
1029 {
1030         struct page *page = compound_head(__page);
1031         wait_queue_head_t *q = page_waitqueue(page);
1032         wait_on_page_bit_common(q, page, PG_locked, TASK_UNINTERRUPTIBLE, true);
1033 }
1034 EXPORT_SYMBOL(__lock_page);
1035
1036 int __lock_page_killable(struct page *__page)
1037 {
1038         struct page *page = compound_head(__page);
1039         wait_queue_head_t *q = page_waitqueue(page);
1040         return wait_on_page_bit_common(q, page, PG_locked, TASK_KILLABLE, true);
1041 }
1042 EXPORT_SYMBOL_GPL(__lock_page_killable);
1043
1044 /*
1045  * Return values:
1046  * 1 - page is locked; mmap_sem is still held.
1047  * 0 - page is not locked.
1048  *     mmap_sem has been released (up_read()), unless flags had both
1049  *     FAULT_FLAG_ALLOW_RETRY and FAULT_FLAG_RETRY_NOWAIT set, in
1050  *     which case mmap_sem is still held.
1051  *
1052  * If neither ALLOW_RETRY nor KILLABLE are set, will always return 1
1053  * with the page locked and the mmap_sem unperturbed.
1054  */
1055 int __lock_page_or_retry(struct page *page, struct mm_struct *mm,
1056                          unsigned int flags)
1057 {
1058         if (flags & FAULT_FLAG_ALLOW_RETRY) {
1059                 /*
1060                  * CAUTION! In this case, mmap_sem is not released
1061                  * even though return 0.
1062                  */
1063                 if (flags & FAULT_FLAG_RETRY_NOWAIT)
1064                         return 0;
1065
1066                 up_read(&mm->mmap_sem);
1067                 if (flags & FAULT_FLAG_KILLABLE)
1068                         wait_on_page_locked_killable(page);
1069                 else
1070                         wait_on_page_locked(page);
1071                 return 0;
1072         } else {
1073                 if (flags & FAULT_FLAG_KILLABLE) {
1074                         int ret;
1075
1076                         ret = __lock_page_killable(page);
1077                         if (ret) {
1078                                 up_read(&mm->mmap_sem);
1079                                 return 0;
1080                         }
1081                 } else
1082                         __lock_page(page);
1083                 return 1;
1084         }
1085 }
1086
1087 /**
1088  * page_cache_next_hole - find the next hole (not-present entry)
1089  * @mapping: mapping
1090  * @index: index
1091  * @max_scan: maximum range to search
1092  *
1093  * Search the set [index, min(index+max_scan-1, MAX_INDEX)] for the
1094  * lowest indexed hole.
1095  *
1096  * Returns: the index of the hole if found, otherwise returns an index
1097  * outside of the set specified (in which case 'return - index >=
1098  * max_scan' will be true). In rare cases of index wrap-around, 0 will
1099  * be returned.
1100  *
1101  * page_cache_next_hole may be called under rcu_read_lock. However,
1102  * like radix_tree_gang_lookup, this will not atomically search a
1103  * snapshot of the tree at a single point in time. For example, if a
1104  * hole is created at index 5, then subsequently a hole is created at
1105  * index 10, page_cache_next_hole covering both indexes may return 10
1106  * if called under rcu_read_lock.
1107  */
1108 pgoff_t page_cache_next_hole(struct address_space *mapping,
1109                              pgoff_t index, unsigned long max_scan)
1110 {
1111         unsigned long i;
1112
1113         for (i = 0; i < max_scan; i++) {
1114                 struct page *page;
1115
1116                 page = radix_tree_lookup(&mapping->page_tree, index);
1117                 if (!page || radix_tree_exceptional_entry(page))
1118                         break;
1119                 index++;
1120                 if (index == 0)
1121                         break;
1122         }
1123
1124         return index;
1125 }
1126 EXPORT_SYMBOL(page_cache_next_hole);
1127
1128 /**
1129  * page_cache_prev_hole - find the prev hole (not-present entry)
1130  * @mapping: mapping
1131  * @index: index
1132  * @max_scan: maximum range to search
1133  *
1134  * Search backwards in the range [max(index-max_scan+1, 0), index] for
1135  * the first hole.
1136  *
1137  * Returns: the index of the hole if found, otherwise returns an index
1138  * outside of the set specified (in which case 'index - return >=
1139  * max_scan' will be true). In rare cases of wrap-around, ULONG_MAX
1140  * will be returned.
1141  *
1142  * page_cache_prev_hole may be called under rcu_read_lock. However,
1143  * like radix_tree_gang_lookup, this will not atomically search a
1144  * snapshot of the tree at a single point in time. For example, if a
1145  * hole is created at index 10, then subsequently a hole is created at
1146  * index 5, page_cache_prev_hole covering both indexes may return 5 if
1147  * called under rcu_read_lock.
1148  */
1149 pgoff_t page_cache_prev_hole(struct address_space *mapping,
1150                              pgoff_t index, unsigned long max_scan)
1151 {
1152         unsigned long i;
1153
1154         for (i = 0; i < max_scan; i++) {
1155                 struct page *page;
1156
1157                 page = radix_tree_lookup(&mapping->page_tree, index);
1158                 if (!page || radix_tree_exceptional_entry(page))
1159                         break;
1160                 index--;
1161                 if (index == ULONG_MAX)
1162                         break;
1163         }
1164
1165         return index;
1166 }
1167 EXPORT_SYMBOL(page_cache_prev_hole);
1168
1169 /**
1170  * find_get_entry - find and get a page cache entry
1171  * @mapping: the address_space to search
1172  * @offset: the page cache index
1173  *
1174  * Looks up the page cache slot at @mapping & @offset.  If there is a
1175  * page cache page, it is returned with an increased refcount.
1176  *
1177  * If the slot holds a shadow entry of a previously evicted page, or a
1178  * swap entry from shmem/tmpfs, it is returned.
1179  *
1180  * Otherwise, %NULL is returned.
1181  */
1182 struct page *find_get_entry(struct address_space *mapping, pgoff_t offset)
1183 {
1184         void **pagep;
1185         struct page *head, *page;
1186
1187         rcu_read_lock();
1188 repeat:
1189         page = NULL;
1190         pagep = radix_tree_lookup_slot(&mapping->page_tree, offset);
1191         if (pagep) {
1192                 page = radix_tree_deref_slot(pagep);
1193                 if (unlikely(!page))
1194                         goto out;
1195                 if (radix_tree_exception(page)) {
1196                         if (radix_tree_deref_retry(page))
1197                                 goto repeat;
1198                         /*
1199                          * A shadow entry of a recently evicted page,
1200                          * or a swap entry from shmem/tmpfs.  Return
1201                          * it without attempting to raise page count.
1202                          */
1203                         goto out;
1204                 }
1205
1206                 head = compound_head(page);
1207                 if (!page_cache_get_speculative(head))
1208                         goto repeat;
1209
1210                 /* The page was split under us? */
1211                 if (compound_head(page) != head) {
1212                         put_page(head);
1213                         goto repeat;
1214                 }
1215
1216                 /*
1217                  * Has the page moved?
1218                  * This is part of the lockless pagecache protocol. See
1219                  * include/linux/pagemap.h for details.
1220                  */
1221                 if (unlikely(page != *pagep)) {
1222                         put_page(head);
1223                         goto repeat;
1224                 }
1225         }
1226 out:
1227         rcu_read_unlock();
1228
1229         return page;
1230 }
1231 EXPORT_SYMBOL(find_get_entry);
1232
1233 /**
1234  * find_lock_entry - locate, pin and lock a page cache entry
1235  * @mapping: the address_space to search
1236  * @offset: the page cache index
1237  *
1238  * Looks up the page cache slot at @mapping & @offset.  If there is a
1239  * page cache page, it is returned locked and with an increased
1240  * refcount.
1241  *
1242  * If the slot holds a shadow entry of a previously evicted page, or a
1243  * swap entry from shmem/tmpfs, it is returned.
1244  *
1245  * Otherwise, %NULL is returned.
1246  *
1247  * find_lock_entry() may sleep.
1248  */
1249 struct page *find_lock_entry(struct address_space *mapping, pgoff_t offset)
1250 {
1251         struct page *page;
1252
1253 repeat:
1254         page = find_get_entry(mapping, offset);
1255         if (page && !radix_tree_exception(page)) {
1256                 lock_page(page);
1257                 /* Has the page been truncated? */
1258                 if (unlikely(page_mapping(page) != mapping)) {
1259                         unlock_page(page);
1260                         put_page(page);
1261                         goto repeat;
1262                 }
1263                 VM_BUG_ON_PAGE(page_to_pgoff(page) != offset, page);
1264         }
1265         return page;
1266 }
1267 EXPORT_SYMBOL(find_lock_entry);
1268
1269 /**
1270  * pagecache_get_page - find and get a page reference
1271  * @mapping: the address_space to search
1272  * @offset: the page index
1273  * @fgp_flags: PCG flags
1274  * @gfp_mask: gfp mask to use for the page cache data page allocation
1275  *
1276  * Looks up the page cache slot at @mapping & @offset.
1277  *
1278  * PCG flags modify how the page is returned.
1279  *
1280  * @fgp_flags can be:
1281  *
1282  * - FGP_ACCESSED: the page will be marked accessed
1283  * - FGP_LOCK: Page is return locked
1284  * - FGP_CREAT: If page is not present then a new page is allocated using
1285  *   @gfp_mask and added to the page cache and the VM's LRU
1286  *   list. The page is returned locked and with an increased
1287  *   refcount. Otherwise, NULL is returned.
1288  *
1289  * If FGP_LOCK or FGP_CREAT are specified then the function may sleep even
1290  * if the GFP flags specified for FGP_CREAT are atomic.
1291  *
1292  * If there is a page cache page, it is returned with an increased refcount.
1293  */
1294 struct page *pagecache_get_page(struct address_space *mapping, pgoff_t offset,
1295         int fgp_flags, gfp_t gfp_mask)
1296 {
1297         struct page *page;
1298
1299 repeat:
1300         page = find_get_entry(mapping, offset);
1301         if (radix_tree_exceptional_entry(page))
1302                 page = NULL;
1303         if (!page)
1304                 goto no_page;
1305
1306         if (fgp_flags & FGP_LOCK) {
1307                 if (fgp_flags & FGP_NOWAIT) {
1308                         if (!trylock_page(page)) {
1309                                 put_page(page);
1310                                 return NULL;
1311                         }
1312                 } else {
1313                         lock_page(page);
1314                 }
1315
1316                 /* Has the page been truncated? */
1317                 if (unlikely(page->mapping != mapping)) {
1318                         unlock_page(page);
1319                         put_page(page);
1320                         goto repeat;
1321                 }
1322                 VM_BUG_ON_PAGE(page->index != offset, page);
1323         }
1324
1325         if (page && (fgp_flags & FGP_ACCESSED))
1326                 mark_page_accessed(page);
1327
1328 no_page:
1329         if (!page && (fgp_flags & FGP_CREAT)) {
1330                 int err;
1331                 if ((fgp_flags & FGP_WRITE) && mapping_cap_account_dirty(mapping))
1332                         gfp_mask |= __GFP_WRITE;
1333                 if (fgp_flags & FGP_NOFS)
1334                         gfp_mask &= ~__GFP_FS;
1335
1336                 page = __page_cache_alloc(gfp_mask);
1337                 if (!page)
1338                         return NULL;
1339
1340                 if (WARN_ON_ONCE(!(fgp_flags & FGP_LOCK)))
1341                         fgp_flags |= FGP_LOCK;
1342
1343                 /* Init accessed so avoid atomic mark_page_accessed later */
1344                 if (fgp_flags & FGP_ACCESSED)
1345                         __SetPageReferenced(page);
1346
1347                 err = add_to_page_cache_lru(page, mapping, offset,
1348                                 gfp_mask & GFP_RECLAIM_MASK);
1349                 if (unlikely(err)) {
1350                         put_page(page);
1351                         page = NULL;
1352                         if (err == -EEXIST)
1353                                 goto repeat;
1354                 }
1355         }
1356
1357         return page;
1358 }
1359 EXPORT_SYMBOL(pagecache_get_page);
1360
1361 /**
1362  * find_get_entries - gang pagecache lookup
1363  * @mapping:    The address_space to search
1364  * @start:      The starting page cache index
1365  * @nr_entries: The maximum number of entries
1366  * @entries:    Where the resulting entries are placed
1367  * @indices:    The cache indices corresponding to the entries in @entries
1368  *
1369  * find_get_entries() will search for and return a group of up to
1370  * @nr_entries entries in the mapping.  The entries are placed at
1371  * @entries.  find_get_entries() takes a reference against any actual
1372  * pages it returns.
1373  *
1374  * The search returns a group of mapping-contiguous page cache entries
1375  * with ascending indexes.  There may be holes in the indices due to
1376  * not-present pages.
1377  *
1378  * Any shadow entries of evicted pages, or swap entries from
1379  * shmem/tmpfs, are included in the returned array.
1380  *
1381  * find_get_entries() returns the number of pages and shadow entries
1382  * which were found.
1383  */
1384 unsigned find_get_entries(struct address_space *mapping,
1385                           pgoff_t start, unsigned int nr_entries,
1386                           struct page **entries, pgoff_t *indices)
1387 {
1388         void **slot;
1389         unsigned int ret = 0;
1390         struct radix_tree_iter iter;
1391
1392         if (!nr_entries)
1393                 return 0;
1394
1395         rcu_read_lock();
1396         radix_tree_for_each_slot(slot, &mapping->page_tree, &iter, start) {
1397                 struct page *head, *page;
1398 repeat:
1399                 page = radix_tree_deref_slot(slot);
1400                 if (unlikely(!page))
1401                         continue;
1402                 if (radix_tree_exception(page)) {
1403                         if (radix_tree_deref_retry(page)) {
1404                                 slot = radix_tree_iter_retry(&iter);
1405                                 continue;
1406                         }
1407                         /*
1408                          * A shadow entry of a recently evicted page, a swap
1409                          * entry from shmem/tmpfs or a DAX entry.  Return it
1410                          * without attempting to raise page count.
1411                          */
1412                         goto export;
1413                 }
1414
1415                 head = compound_head(page);
1416                 if (!page_cache_get_speculative(head))
1417                         goto repeat;
1418
1419                 /* The page was split under us? */
1420                 if (compound_head(page) != head) {
1421                         put_page(head);
1422                         goto repeat;
1423                 }
1424
1425                 /* Has the page moved? */
1426                 if (unlikely(page != *slot)) {
1427                         put_page(head);
1428                         goto repeat;
1429                 }
1430 export:
1431                 indices[ret] = iter.index;
1432                 entries[ret] = page;
1433                 if (++ret == nr_entries)
1434                         break;
1435         }
1436         rcu_read_unlock();
1437         return ret;
1438 }
1439
1440 /**
1441  * find_get_pages - gang pagecache lookup
1442  * @mapping:    The address_space to search
1443  * @start:      The starting page index
1444  * @nr_pages:   The maximum number of pages
1445  * @pages:      Where the resulting pages are placed
1446  *
1447  * find_get_pages() will search for and return a group of up to
1448  * @nr_pages pages in the mapping.  The pages are placed at @pages.
1449  * find_get_pages() takes a reference against the returned pages.
1450  *
1451  * The search returns a group of mapping-contiguous pages with ascending
1452  * indexes.  There may be holes in the indices due to not-present pages.
1453  *
1454  * find_get_pages() returns the number of pages which were found.
1455  */
1456 unsigned find_get_pages(struct address_space *mapping, pgoff_t start,
1457                             unsigned int nr_pages, struct page **pages)
1458 {
1459         struct radix_tree_iter iter;
1460         void **slot;
1461         unsigned ret = 0;
1462
1463         if (unlikely(!nr_pages))
1464                 return 0;
1465
1466         rcu_read_lock();
1467         radix_tree_for_each_slot(slot, &mapping->page_tree, &iter, start) {
1468                 struct page *head, *page;
1469 repeat:
1470                 page = radix_tree_deref_slot(slot);
1471                 if (unlikely(!page))
1472                         continue;
1473
1474                 if (radix_tree_exception(page)) {
1475                         if (radix_tree_deref_retry(page)) {
1476                                 slot = radix_tree_iter_retry(&iter);
1477                                 continue;
1478                         }
1479                         /*
1480                          * A shadow entry of a recently evicted page,
1481                          * or a swap entry from shmem/tmpfs.  Skip
1482                          * over it.
1483                          */
1484                         continue;
1485                 }
1486
1487                 head = compound_head(page);
1488                 if (!page_cache_get_speculative(head))
1489                         goto repeat;
1490
1491                 /* The page was split under us? */
1492                 if (compound_head(page) != head) {
1493                         put_page(head);
1494                         goto repeat;
1495                 }
1496
1497                 /* Has the page moved? */
1498                 if (unlikely(page != *slot)) {
1499                         put_page(head);
1500                         goto repeat;
1501                 }
1502
1503                 pages[ret] = page;
1504                 if (++ret == nr_pages)
1505                         break;
1506         }
1507
1508         rcu_read_unlock();
1509         return ret;
1510 }
1511
1512 /**
1513  * find_get_pages_contig - gang contiguous pagecache lookup
1514  * @mapping:    The address_space to search
1515  * @index:      The starting page index
1516  * @nr_pages:   The maximum number of pages
1517  * @pages:      Where the resulting pages are placed
1518  *
1519  * find_get_pages_contig() works exactly like find_get_pages(), except
1520  * that the returned number of pages are guaranteed to be contiguous.
1521  *
1522  * find_get_pages_contig() returns the number of pages which were found.
1523  */
1524 unsigned find_get_pages_contig(struct address_space *mapping, pgoff_t index,
1525                                unsigned int nr_pages, struct page **pages)
1526 {
1527         struct radix_tree_iter iter;
1528         void **slot;
1529         unsigned int ret = 0;
1530
1531         if (unlikely(!nr_pages))
1532                 return 0;
1533
1534         rcu_read_lock();
1535         radix_tree_for_each_contig(slot, &mapping->page_tree, &iter, index) {
1536                 struct page *head, *page;
1537 repeat:
1538                 page = radix_tree_deref_slot(slot);
1539                 /* The hole, there no reason to continue */
1540                 if (unlikely(!page))
1541                         break;
1542
1543                 if (radix_tree_exception(page)) {
1544                         if (radix_tree_deref_retry(page)) {
1545                                 slot = radix_tree_iter_retry(&iter);
1546                                 continue;
1547                         }
1548                         /*
1549                          * A shadow entry of a recently evicted page,
1550                          * or a swap entry from shmem/tmpfs.  Stop
1551                          * looking for contiguous pages.
1552                          */
1553                         break;
1554                 }
1555
1556                 head = compound_head(page);
1557                 if (!page_cache_get_speculative(head))
1558                         goto repeat;
1559
1560                 /* The page was split under us? */
1561                 if (compound_head(page) != head) {
1562                         put_page(head);
1563                         goto repeat;
1564                 }
1565
1566                 /* Has the page moved? */
1567                 if (unlikely(page != *slot)) {
1568                         put_page(head);
1569                         goto repeat;
1570                 }
1571
1572                 /*
1573                  * must check mapping and index after taking the ref.
1574                  * otherwise we can get both false positives and false
1575                  * negatives, which is just confusing to the caller.
1576                  */
1577                 if (page->mapping == NULL || page_to_pgoff(page) != iter.index) {
1578                         put_page(page);
1579                         break;
1580                 }
1581
1582                 pages[ret] = page;
1583                 if (++ret == nr_pages)
1584                         break;
1585         }
1586         rcu_read_unlock();
1587         return ret;
1588 }
1589 EXPORT_SYMBOL(find_get_pages_contig);
1590
1591 /**
1592  * find_get_pages_tag - find and return pages that match @tag
1593  * @mapping:    the address_space to search
1594  * @index:      the starting page index
1595  * @tag:        the tag index
1596  * @nr_pages:   the maximum number of pages
1597  * @pages:      where the resulting pages are placed
1598  *
1599  * Like find_get_pages, except we only return pages which are tagged with
1600  * @tag.   We update @index to index the next page for the traversal.
1601  */
1602 unsigned find_get_pages_tag(struct address_space *mapping, pgoff_t *index,
1603                         int tag, unsigned int nr_pages, struct page **pages)
1604 {
1605         struct radix_tree_iter iter;
1606         void **slot;
1607         unsigned ret = 0;
1608
1609         if (unlikely(!nr_pages))
1610                 return 0;
1611
1612         rcu_read_lock();
1613         radix_tree_for_each_tagged(slot, &mapping->page_tree,
1614                                    &iter, *index, tag) {
1615                 struct page *head, *page;
1616 repeat:
1617                 page = radix_tree_deref_slot(slot);
1618                 if (unlikely(!page))
1619                         continue;
1620
1621                 if (radix_tree_exception(page)) {
1622                         if (radix_tree_deref_retry(page)) {
1623                                 slot = radix_tree_iter_retry(&iter);
1624                                 continue;
1625                         }
1626                         /*
1627                          * A shadow entry of a recently evicted page.
1628                          *
1629                          * Those entries should never be tagged, but
1630                          * this tree walk is lockless and the tags are
1631                          * looked up in bulk, one radix tree node at a
1632                          * time, so there is a sizable window for page
1633                          * reclaim to evict a page we saw tagged.
1634                          *
1635                          * Skip over it.
1636                          */
1637                         continue;
1638                 }
1639
1640                 head = compound_head(page);
1641                 if (!page_cache_get_speculative(head))
1642                         goto repeat;
1643
1644                 /* The page was split under us? */
1645                 if (compound_head(page) != head) {
1646                         put_page(head);
1647                         goto repeat;
1648                 }
1649
1650                 /* Has the page moved? */
1651                 if (unlikely(page != *slot)) {
1652                         put_page(head);
1653                         goto repeat;
1654                 }
1655
1656                 pages[ret] = page;
1657                 if (++ret == nr_pages)
1658                         break;
1659         }
1660
1661         rcu_read_unlock();
1662
1663         if (ret)
1664                 *index = pages[ret - 1]->index + 1;
1665
1666         return ret;
1667 }
1668 EXPORT_SYMBOL(find_get_pages_tag);
1669
1670 /**
1671  * find_get_entries_tag - find and return entries that match @tag
1672  * @mapping:    the address_space to search
1673  * @start:      the starting page cache index
1674  * @tag:        the tag index
1675  * @nr_entries: the maximum number of entries
1676  * @entries:    where the resulting entries are placed
1677  * @indices:    the cache indices corresponding to the entries in @entries
1678  *
1679  * Like find_get_entries, except we only return entries which are tagged with
1680  * @tag.
1681  */
1682 unsigned find_get_entries_tag(struct address_space *mapping, pgoff_t start,
1683                         int tag, unsigned int nr_entries,
1684                         struct page **entries, pgoff_t *indices)
1685 {
1686         void **slot;
1687         unsigned int ret = 0;
1688         struct radix_tree_iter iter;
1689
1690         if (!nr_entries)
1691                 return 0;
1692
1693         rcu_read_lock();
1694         radix_tree_for_each_tagged(slot, &mapping->page_tree,
1695                                    &iter, start, tag) {
1696                 struct page *head, *page;
1697 repeat:
1698                 page = radix_tree_deref_slot(slot);
1699                 if (unlikely(!page))
1700                         continue;
1701                 if (radix_tree_exception(page)) {
1702                         if (radix_tree_deref_retry(page)) {
1703                                 slot = radix_tree_iter_retry(&iter);
1704                                 continue;
1705                         }
1706
1707                         /*
1708                          * A shadow entry of a recently evicted page, a swap
1709                          * entry from shmem/tmpfs or a DAX entry.  Return it
1710                          * without attempting to raise page count.
1711                          */
1712                         goto export;
1713                 }
1714
1715                 head = compound_head(page);
1716                 if (!page_cache_get_speculative(head))
1717                         goto repeat;
1718
1719                 /* The page was split under us? */
1720                 if (compound_head(page) != head) {
1721                         put_page(head);
1722                         goto repeat;
1723                 }
1724
1725                 /* Has the page moved? */
1726                 if (unlikely(page != *slot)) {
1727                         put_page(head);
1728                         goto repeat;
1729                 }
1730 export:
1731                 indices[ret] = iter.index;
1732                 entries[ret] = page;
1733                 if (++ret == nr_entries)
1734                         break;
1735         }
1736         rcu_read_unlock();
1737         return ret;
1738 }
1739 EXPORT_SYMBOL(find_get_entries_tag);
1740
1741 /*
1742  * CD/DVDs are error prone. When a medium error occurs, the driver may fail
1743  * a _large_ part of the i/o request. Imagine the worst scenario:
1744  *
1745  *      ---R__________________________________________B__________
1746  *         ^ reading here                             ^ bad block(assume 4k)
1747  *
1748  * read(R) => miss => readahead(R...B) => media error => frustrating retries
1749  * => failing the whole request => read(R) => read(R+1) =>
1750  * readahead(R+1...B+1) => bang => read(R+2) => read(R+3) =>
1751  * readahead(R+3...B+2) => bang => read(R+3) => read(R+4) =>
1752  * readahead(R+4...B+3) => bang => read(R+4) => read(R+5) => ......
1753  *
1754  * It is going insane. Fix it by quickly scaling down the readahead size.
1755  */
1756 static void shrink_readahead_size_eio(struct file *filp,
1757                                         struct file_ra_state *ra)
1758 {
1759         ra->ra_pages /= 4;
1760 }
1761
1762 /**
1763  * do_generic_file_read - generic file read routine
1764  * @filp:       the file to read
1765  * @ppos:       current file position
1766  * @iter:       data destination
1767  * @written:    already copied
1768  *
1769  * This is a generic file read routine, and uses the
1770  * mapping->a_ops->readpage() function for the actual low-level stuff.
1771  *
1772  * This is really ugly. But the goto's actually try to clarify some
1773  * of the logic when it comes to error handling etc.
1774  */
1775 static ssize_t do_generic_file_read(struct file *filp, loff_t *ppos,
1776                 struct iov_iter *iter, ssize_t written)
1777 {
1778         struct address_space *mapping = filp->f_mapping;
1779         struct inode *inode = mapping->host;
1780         struct file_ra_state *ra = &filp->f_ra;
1781         pgoff_t index;
1782         pgoff_t last_index;
1783         pgoff_t prev_index;
1784         unsigned long offset;      /* offset into pagecache page */
1785         unsigned int prev_offset;
1786         int error = 0;
1787
1788         if (unlikely(*ppos >= inode->i_sb->s_maxbytes))
1789                 return 0;
1790         iov_iter_truncate(iter, inode->i_sb->s_maxbytes);
1791
1792         index = *ppos >> PAGE_SHIFT;
1793         prev_index = ra->prev_pos >> PAGE_SHIFT;
1794         prev_offset = ra->prev_pos & (PAGE_SIZE-1);
1795         last_index = (*ppos + iter->count + PAGE_SIZE-1) >> PAGE_SHIFT;
1796         offset = *ppos & ~PAGE_MASK;
1797
1798         for (;;) {
1799                 struct page *page;
1800                 pgoff_t end_index;
1801                 loff_t isize;
1802                 unsigned long nr, ret;
1803
1804                 cond_resched();
1805 find_page:
1806                 if (fatal_signal_pending(current)) {
1807                         error = -EINTR;
1808                         goto out;
1809                 }
1810
1811                 page = find_get_page(mapping, index);
1812                 if (!page) {
1813                         page_cache_sync_readahead(mapping,
1814                                         ra, filp,
1815                                         index, last_index - index);
1816                         page = find_get_page(mapping, index);
1817                         if (unlikely(page == NULL))
1818                                 goto no_cached_page;
1819                 }
1820                 if (PageReadahead(page)) {
1821                         page_cache_async_readahead(mapping,
1822                                         ra, filp, page,
1823                                         index, last_index - index);
1824                 }
1825                 if (!PageUptodate(page)) {
1826                         /*
1827                          * See comment in do_read_cache_page on why
1828                          * wait_on_page_locked is used to avoid unnecessarily
1829                          * serialisations and why it's safe.
1830                          */
1831                         error = wait_on_page_locked_killable(page);
1832                         if (unlikely(error))
1833                                 goto readpage_error;
1834                         if (PageUptodate(page))
1835                                 goto page_ok;
1836
1837                         if (inode->i_blkbits == PAGE_SHIFT ||
1838                                         !mapping->a_ops->is_partially_uptodate)
1839                                 goto page_not_up_to_date;
1840                         /* pipes can't handle partially uptodate pages */
1841                         if (unlikely(iter->type & ITER_PIPE))
1842                                 goto page_not_up_to_date;
1843                         if (!trylock_page(page))
1844                                 goto page_not_up_to_date;
1845                         /* Did it get truncated before we got the lock? */
1846                         if (!page->mapping)
1847                                 goto page_not_up_to_date_locked;
1848                         if (!mapping->a_ops->is_partially_uptodate(page,
1849                                                         offset, iter->count))
1850                                 goto page_not_up_to_date_locked;
1851                         unlock_page(page);
1852                 }
1853 page_ok:
1854                 /*
1855                  * i_size must be checked after we know the page is Uptodate.
1856                  *
1857                  * Checking i_size after the check allows us to calculate
1858                  * the correct value for "nr", which means the zero-filled
1859                  * part of the page is not copied back to userspace (unless
1860                  * another truncate extends the file - this is desired though).
1861                  */
1862
1863                 isize = i_size_read(inode);
1864                 end_index = (isize - 1) >> PAGE_SHIFT;
1865                 if (unlikely(!isize || index > end_index)) {
1866                         put_page(page);
1867                         goto out;
1868                 }
1869
1870                 /* nr is the maximum number of bytes to copy from this page */
1871                 nr = PAGE_SIZE;
1872                 if (index == end_index) {
1873                         nr = ((isize - 1) & ~PAGE_MASK) + 1;
1874                         if (nr <= offset) {
1875                                 put_page(page);
1876                                 goto out;
1877                         }
1878                 }
1879                 nr = nr - offset;
1880
1881                 /* If users can be writing to this page using arbitrary
1882                  * virtual addresses, take care about potential aliasing
1883                  * before reading the page on the kernel side.
1884                  */
1885                 if (mapping_writably_mapped(mapping))
1886                         flush_dcache_page(page);
1887
1888                 /*
1889                  * When a sequential read accesses a page several times,
1890                  * only mark it as accessed the first time.
1891                  */
1892                 if (prev_index != index || offset != prev_offset)
1893                         mark_page_accessed(page);
1894                 prev_index = index;
1895
1896                 /*
1897                  * Ok, we have the page, and it's up-to-date, so
1898                  * now we can copy it to user space...
1899                  */
1900
1901                 ret = copy_page_to_iter(page, offset, nr, iter);
1902                 offset += ret;
1903                 index += offset >> PAGE_SHIFT;
1904                 offset &= ~PAGE_MASK;
1905                 prev_offset = offset;
1906
1907                 put_page(page);
1908                 written += ret;
1909                 if (!iov_iter_count(iter))
1910                         goto out;
1911                 if (ret < nr) {
1912                         error = -EFAULT;
1913                         goto out;
1914                 }
1915                 continue;
1916
1917 page_not_up_to_date:
1918                 /* Get exclusive access to the page ... */
1919                 error = lock_page_killable(page);
1920                 if (unlikely(error))
1921                         goto readpage_error;
1922
1923 page_not_up_to_date_locked:
1924                 /* Did it get truncated before we got the lock? */
1925                 if (!page->mapping) {
1926                         unlock_page(page);
1927                         put_page(page);
1928                         continue;
1929                 }
1930
1931                 /* Did somebody else fill it already? */
1932                 if (PageUptodate(page)) {
1933                         unlock_page(page);
1934                         goto page_ok;
1935                 }
1936
1937 readpage:
1938                 /*
1939                  * A previous I/O error may have been due to temporary
1940                  * failures, eg. multipath errors.
1941                  * PG_error will be set again if readpage fails.
1942                  */
1943                 ClearPageError(page);
1944                 /* Start the actual read. The read will unlock the page. */
1945                 error = mapping->a_ops->readpage(filp, page);
1946
1947                 if (unlikely(error)) {
1948                         if (error == AOP_TRUNCATED_PAGE) {
1949                                 put_page(page);
1950                                 error = 0;
1951                                 goto find_page;
1952                         }
1953                         goto readpage_error;
1954                 }
1955
1956                 if (!PageUptodate(page)) {
1957                         error = lock_page_killable(page);
1958                         if (unlikely(error))
1959                                 goto readpage_error;
1960                         if (!PageUptodate(page)) {
1961                                 if (page->mapping == NULL) {
1962                                         /*
1963                                          * invalidate_mapping_pages got it
1964                                          */
1965                                         unlock_page(page);
1966                                         put_page(page);
1967                                         goto find_page;
1968                                 }
1969                                 unlock_page(page);
1970                                 shrink_readahead_size_eio(filp, ra);
1971                                 error = -EIO;
1972                                 goto readpage_error;
1973                         }
1974                         unlock_page(page);
1975                 }
1976
1977                 goto page_ok;
1978
1979 readpage_error:
1980                 /* UHHUH! A synchronous read error occurred. Report it */
1981                 put_page(page);
1982                 goto out;
1983
1984 no_cached_page:
1985                 /*
1986                  * Ok, it wasn't cached, so we need to create a new
1987                  * page..
1988                  */
1989                 page = page_cache_alloc_cold(mapping);
1990                 if (!page) {
1991                         error = -ENOMEM;
1992                         goto out;
1993                 }
1994                 error = add_to_page_cache_lru(page, mapping, index,
1995                                 mapping_gfp_constraint(mapping, GFP_KERNEL));
1996                 if (error) {
1997                         put_page(page);
1998                         if (error == -EEXIST) {
1999                                 error = 0;
2000                                 goto find_page;
2001                         }
2002                         goto out;
2003                 }
2004                 goto readpage;
2005         }
2006
2007 out:
2008         ra->prev_pos = prev_index;
2009         ra->prev_pos <<= PAGE_SHIFT;
2010         ra->prev_pos |= prev_offset;
2011
2012         *ppos = ((loff_t)index << PAGE_SHIFT) + offset;
2013         file_accessed(filp);
2014         return written ? written : error;
2015 }
2016
2017 /**
2018  * generic_file_read_iter - generic filesystem read routine
2019  * @iocb:       kernel I/O control block
2020  * @iter:       destination for the data read
2021  *
2022  * This is the "read_iter()" routine for all filesystems
2023  * that can use the page cache directly.
2024  */
2025 ssize_t
2026 generic_file_read_iter(struct kiocb *iocb, struct iov_iter *iter)
2027 {
2028         struct file *file = iocb->ki_filp;
2029         ssize_t retval = 0;
2030         size_t count = iov_iter_count(iter);
2031
2032         if (!count)
2033                 goto out; /* skip atime */
2034
2035         if (iocb->ki_flags & IOCB_DIRECT) {
2036                 struct address_space *mapping = file->f_mapping;
2037                 struct inode *inode = mapping->host;
2038                 loff_t size;
2039
2040                 size = i_size_read(inode);
2041                 retval = filemap_write_and_wait_range(mapping, iocb->ki_pos,
2042                                         iocb->ki_pos + count - 1);
2043                 if (retval < 0)
2044                         goto out;
2045
2046                 file_accessed(file);
2047
2048                 retval = mapping->a_ops->direct_IO(iocb, iter);
2049                 if (retval >= 0) {
2050                         iocb->ki_pos += retval;
2051                         count -= retval;
2052                 }
2053                 iov_iter_revert(iter, iov_iter_count(iter) - count);
2054
2055                 /*
2056                  * Btrfs can have a short DIO read if we encounter
2057                  * compressed extents, so if there was an error, or if
2058                  * we've already read everything we wanted to, or if
2059                  * there was a short read because we hit EOF, go ahead
2060                  * and return.  Otherwise fallthrough to buffered io for
2061                  * the rest of the read.  Buffered reads will not work for
2062                  * DAX files, so don't bother trying.
2063                  */
2064                 if (retval < 0 || !count || iocb->ki_pos >= size ||
2065                     IS_DAX(inode))
2066                         goto out;
2067         }
2068
2069         retval = do_generic_file_read(file, &iocb->ki_pos, iter, retval);
2070 out:
2071         return retval;
2072 }
2073 EXPORT_SYMBOL(generic_file_read_iter);
2074
2075 #ifdef CONFIG_MMU
2076 /**
2077  * page_cache_read - adds requested page to the page cache if not already there
2078  * @file:       file to read
2079  * @offset:     page index
2080  * @gfp_mask:   memory allocation flags
2081  *
2082  * This adds the requested page to the page cache if it isn't already there,
2083  * and schedules an I/O to read in its contents from disk.
2084  */
2085 static int page_cache_read(struct file *file, pgoff_t offset, gfp_t gfp_mask)
2086 {
2087         struct address_space *mapping = file->f_mapping;
2088         struct page *page;
2089         int ret;
2090
2091         do {
2092                 page = __page_cache_alloc(gfp_mask|__GFP_COLD);
2093                 if (!page)
2094                         return -ENOMEM;
2095
2096                 ret = add_to_page_cache_lru(page, mapping, offset, gfp_mask & GFP_KERNEL);
2097                 if (ret == 0)
2098                         ret = mapping->a_ops->readpage(file, page);
2099                 else if (ret == -EEXIST)
2100                         ret = 0; /* losing race to add is OK */
2101
2102                 put_page(page);
2103
2104         } while (ret == AOP_TRUNCATED_PAGE);
2105
2106         return ret;
2107 }
2108
2109 #define MMAP_LOTSAMISS  (100)
2110
2111 /*
2112  * Synchronous readahead happens when we don't even find
2113  * a page in the page cache at all.
2114  */
2115 static void do_sync_mmap_readahead(struct vm_area_struct *vma,
2116                                    struct file_ra_state *ra,
2117                                    struct file *file,
2118                                    pgoff_t offset)
2119 {
2120         struct address_space *mapping = file->f_mapping;
2121
2122         /* If we don't want any read-ahead, don't bother */
2123         if (vma->vm_flags & VM_RAND_READ)
2124                 return;
2125         if (!ra->ra_pages)
2126                 return;
2127
2128         if (vma->vm_flags & VM_SEQ_READ) {
2129                 page_cache_sync_readahead(mapping, ra, file, offset,
2130                                           ra->ra_pages);
2131                 return;
2132         }
2133
2134         /* Avoid banging the cache line if not needed */
2135         if (ra->mmap_miss < MMAP_LOTSAMISS * 10)
2136                 ra->mmap_miss++;
2137
2138         /*
2139          * Do we miss much more than hit in this file? If so,
2140          * stop bothering with read-ahead. It will only hurt.
2141          */
2142         if (ra->mmap_miss > MMAP_LOTSAMISS)
2143                 return;
2144
2145         /*
2146          * mmap read-around
2147          */
2148         ra->start = max_t(long, 0, offset - ra->ra_pages / 2);
2149         ra->size = ra->ra_pages;
2150         ra->async_size = ra->ra_pages / 4;
2151         ra_submit(ra, mapping, file);
2152 }
2153
2154 /*
2155  * Asynchronous readahead happens when we find the page and PG_readahead,
2156  * so we want to possibly extend the readahead further..
2157  */
2158 static void do_async_mmap_readahead(struct vm_area_struct *vma,
2159                                     struct file_ra_state *ra,
2160                                     struct file *file,
2161                                     struct page *page,
2162                                     pgoff_t offset)
2163 {
2164         struct address_space *mapping = file->f_mapping;
2165
2166         /* If we don't want any read-ahead, don't bother */
2167         if (vma->vm_flags & VM_RAND_READ)
2168                 return;
2169         if (ra->mmap_miss > 0)
2170                 ra->mmap_miss--;
2171         if (PageReadahead(page))
2172                 page_cache_async_readahead(mapping, ra, file,
2173                                            page, offset, ra->ra_pages);
2174 }
2175
2176 /**
2177  * filemap_fault - read in file data for page fault handling
2178  * @vmf:        struct vm_fault containing details of the fault
2179  *
2180  * filemap_fault() is invoked via the vma operations vector for a
2181  * mapped memory region to read in file data during a page fault.
2182  *
2183  * The goto's are kind of ugly, but this streamlines the normal case of having
2184  * it in the page cache, and handles the special cases reasonably without
2185  * having a lot of duplicated code.
2186  *
2187  * vma->vm_mm->mmap_sem must be held on entry.
2188  *
2189  * If our return value has VM_FAULT_RETRY set, it's because
2190  * lock_page_or_retry() returned 0.
2191  * The mmap_sem has usually been released in this case.
2192  * See __lock_page_or_retry() for the exception.
2193  *
2194  * If our return value does not have VM_FAULT_RETRY set, the mmap_sem
2195  * has not been released.
2196  *
2197  * We never return with VM_FAULT_RETRY and a bit from VM_FAULT_ERROR set.
2198  */
2199 int filemap_fault(struct vm_fault *vmf)
2200 {
2201         int error;
2202         struct file *file = vmf->vma->vm_file;
2203         struct address_space *mapping = file->f_mapping;
2204         struct file_ra_state *ra = &file->f_ra;
2205         struct inode *inode = mapping->host;
2206         pgoff_t offset = vmf->pgoff;
2207         pgoff_t max_off;
2208         struct page *page;
2209         int ret = 0;
2210
2211         max_off = DIV_ROUND_UP(i_size_read(inode), PAGE_SIZE);
2212         if (unlikely(offset >= max_off))
2213                 return VM_FAULT_SIGBUS;
2214
2215         /*
2216          * Do we have something in the page cache already?
2217          */
2218         page = find_get_page(mapping, offset);
2219         if (likely(page) && !(vmf->flags & FAULT_FLAG_TRIED)) {
2220                 /*
2221                  * We found the page, so try async readahead before
2222                  * waiting for the lock.
2223                  */
2224                 do_async_mmap_readahead(vmf->vma, ra, file, page, offset);
2225         } else if (!page) {
2226                 /* No page in the page cache at all */
2227                 do_sync_mmap_readahead(vmf->vma, ra, file, offset);
2228                 count_vm_event(PGMAJFAULT);
2229                 mem_cgroup_count_vm_event(vmf->vma->vm_mm, PGMAJFAULT);
2230                 ret = VM_FAULT_MAJOR;
2231 retry_find:
2232                 page = find_get_page(mapping, offset);
2233                 if (!page)
2234                         goto no_cached_page;
2235         }
2236
2237         if (!lock_page_or_retry(page, vmf->vma->vm_mm, vmf->flags)) {
2238                 put_page(page);
2239                 return ret | VM_FAULT_RETRY;
2240         }
2241
2242         /* Did it get truncated? */
2243         if (unlikely(page->mapping != mapping)) {
2244                 unlock_page(page);
2245                 put_page(page);
2246                 goto retry_find;
2247         }
2248         VM_BUG_ON_PAGE(page->index != offset, page);
2249
2250         /*
2251          * We have a locked page in the page cache, now we need to check
2252          * that it's up-to-date. If not, it is going to be due to an error.
2253          */
2254         if (unlikely(!PageUptodate(page)))
2255                 goto page_not_uptodate;
2256
2257         /*
2258          * Found the page and have a reference on it.
2259          * We must recheck i_size under page lock.
2260          */
2261         max_off = DIV_ROUND_UP(i_size_read(inode), PAGE_SIZE);
2262         if (unlikely(offset >= max_off)) {
2263                 unlock_page(page);
2264                 put_page(page);
2265                 return VM_FAULT_SIGBUS;
2266         }
2267
2268         vmf->page = page;
2269         return ret | VM_FAULT_LOCKED;
2270
2271 no_cached_page:
2272         /*
2273          * We're only likely to ever get here if MADV_RANDOM is in
2274          * effect.
2275          */
2276         error = page_cache_read(file, offset, vmf->gfp_mask);
2277
2278         /*
2279          * The page we want has now been added to the page cache.
2280          * In the unlikely event that someone removed it in the
2281          * meantime, we'll just come back here and read it again.
2282          */
2283         if (error >= 0)
2284                 goto retry_find;
2285
2286         /*
2287          * An error return from page_cache_read can result if the
2288          * system is low on memory, or a problem occurs while trying
2289          * to schedule I/O.
2290          */
2291         if (error == -ENOMEM)
2292                 return VM_FAULT_OOM;
2293         return VM_FAULT_SIGBUS;
2294
2295 page_not_uptodate:
2296         /*
2297          * Umm, take care of errors if the page isn't up-to-date.
2298          * Try to re-read it _once_. We do this synchronously,
2299          * because there really aren't any performance issues here
2300          * and we need to check for errors.
2301          */
2302         ClearPageError(page);
2303         error = mapping->a_ops->readpage(file, page);
2304         if (!error) {
2305                 wait_on_page_locked(page);
2306                 if (!PageUptodate(page))
2307                         error = -EIO;
2308         }
2309         put_page(page);
2310
2311         if (!error || error == AOP_TRUNCATED_PAGE)
2312                 goto retry_find;
2313
2314         /* Things didn't work out. Return zero to tell the mm layer so. */
2315         shrink_readahead_size_eio(file, ra);
2316         return VM_FAULT_SIGBUS;
2317 }
2318 EXPORT_SYMBOL(filemap_fault);
2319
2320 void filemap_map_pages(struct vm_fault *vmf,
2321                 pgoff_t start_pgoff, pgoff_t end_pgoff)
2322 {
2323         struct radix_tree_iter iter;
2324         void **slot;
2325         struct file *file = vmf->vma->vm_file;
2326         struct address_space *mapping = file->f_mapping;
2327         pgoff_t last_pgoff = start_pgoff;
2328         unsigned long max_idx;
2329         struct page *head, *page;
2330
2331         rcu_read_lock();
2332         radix_tree_for_each_slot(slot, &mapping->page_tree, &iter,
2333                         start_pgoff) {
2334                 if (iter.index > end_pgoff)
2335                         break;
2336 repeat:
2337                 page = radix_tree_deref_slot(slot);
2338                 if (unlikely(!page))
2339                         goto next;
2340                 if (radix_tree_exception(page)) {
2341                         if (radix_tree_deref_retry(page)) {
2342                                 slot = radix_tree_iter_retry(&iter);
2343                                 continue;
2344                         }
2345                         goto next;
2346                 }
2347
2348                 head = compound_head(page);
2349                 if (!page_cache_get_speculative(head))
2350                         goto repeat;
2351
2352                 /* The page was split under us? */
2353                 if (compound_head(page) != head) {
2354                         put_page(head);
2355                         goto repeat;
2356                 }
2357
2358                 /* Has the page moved? */
2359                 if (unlikely(page != *slot)) {
2360                         put_page(head);
2361                         goto repeat;
2362                 }
2363
2364                 if (!PageUptodate(page) ||
2365                                 PageReadahead(page) ||
2366                                 PageHWPoison(page))
2367                         goto skip;
2368                 if (!trylock_page(page))
2369                         goto skip;
2370
2371                 if (page->mapping != mapping || !PageUptodate(page))
2372                         goto unlock;
2373
2374                 max_idx = DIV_ROUND_UP(i_size_read(mapping->host), PAGE_SIZE);
2375                 if (page->index >= max_idx)
2376                         goto unlock;
2377
2378                 if (file->f_ra.mmap_miss > 0)
2379                         file->f_ra.mmap_miss--;
2380
2381                 vmf->address += (iter.index - last_pgoff) << PAGE_SHIFT;
2382                 if (vmf->pte)
2383                         vmf->pte += iter.index - last_pgoff;
2384                 last_pgoff = iter.index;
2385                 if (alloc_set_pte(vmf, NULL, page))
2386                         goto unlock;
2387                 unlock_page(page);
2388                 goto next;
2389 unlock:
2390                 unlock_page(page);
2391 skip:
2392                 put_page(page);
2393 next:
2394                 /* Huge page is mapped? No need to proceed. */
2395                 if (pmd_trans_huge(*vmf->pmd))
2396                         break;
2397                 if (iter.index == end_pgoff)
2398                         break;
2399         }
2400         rcu_read_unlock();
2401 }
2402 EXPORT_SYMBOL(filemap_map_pages);
2403
2404 int filemap_page_mkwrite(struct vm_fault *vmf)
2405 {
2406         struct page *page = vmf->page;
2407         struct inode *inode = file_inode(vmf->vma->vm_file);
2408         int ret = VM_FAULT_LOCKED;
2409
2410         sb_start_pagefault(inode->i_sb);
2411         file_update_time(vmf->vma->vm_file);
2412         lock_page(page);
2413         if (page->mapping != inode->i_mapping) {
2414                 unlock_page(page);
2415                 ret = VM_FAULT_NOPAGE;
2416                 goto out;
2417         }
2418         /*
2419          * We mark the page dirty already here so that when freeze is in
2420          * progress, we are guaranteed that writeback during freezing will
2421          * see the dirty page and writeprotect it again.
2422          */
2423         set_page_dirty(page);
2424         wait_for_stable_page(page);
2425 out:
2426         sb_end_pagefault(inode->i_sb);
2427         return ret;
2428 }
2429 EXPORT_SYMBOL(filemap_page_mkwrite);
2430
2431 const struct vm_operations_struct generic_file_vm_ops = {
2432         .fault          = filemap_fault,
2433         .map_pages      = filemap_map_pages,
2434         .page_mkwrite   = filemap_page_mkwrite,
2435 };
2436
2437 /* This is used for a general mmap of a disk file */
2438
2439 int generic_file_mmap(struct file * file, struct vm_area_struct * vma)
2440 {
2441         struct address_space *mapping = file->f_mapping;
2442
2443         if (!mapping->a_ops->readpage)
2444                 return -ENOEXEC;
2445         file_accessed(file);
2446         vma->vm_ops = &generic_file_vm_ops;
2447         return 0;
2448 }
2449
2450 /*
2451  * This is for filesystems which do not implement ->writepage.
2452  */
2453 int generic_file_readonly_mmap(struct file *file, struct vm_area_struct *vma)
2454 {
2455         if ((vma->vm_flags & VM_SHARED) && (vma->vm_flags & VM_MAYWRITE))
2456                 return -EINVAL;
2457         return generic_file_mmap(file, vma);
2458 }
2459 #else
2460 int generic_file_mmap(struct file * file, struct vm_area_struct * vma)
2461 {
2462         return -ENOSYS;
2463 }
2464 int generic_file_readonly_mmap(struct file * file, struct vm_area_struct * vma)
2465 {
2466         return -ENOSYS;
2467 }
2468 #endif /* CONFIG_MMU */
2469
2470 EXPORT_SYMBOL(generic_file_mmap);
2471 EXPORT_SYMBOL(generic_file_readonly_mmap);
2472
2473 static struct page *wait_on_page_read(struct page *page)
2474 {
2475         if (!IS_ERR(page)) {
2476                 wait_on_page_locked(page);
2477                 if (!PageUptodate(page)) {
2478                         put_page(page);
2479                         page = ERR_PTR(-EIO);
2480                 }
2481         }
2482         return page;
2483 }
2484
2485 static struct page *do_read_cache_page(struct address_space *mapping,
2486                                 pgoff_t index,
2487                                 int (*filler)(void *, struct page *),
2488                                 void *data,
2489                                 gfp_t gfp)
2490 {
2491         struct page *page;
2492         int err;
2493 repeat:
2494         page = find_get_page(mapping, index);
2495         if (!page) {
2496                 page = __page_cache_alloc(gfp | __GFP_COLD);
2497                 if (!page)
2498                         return ERR_PTR(-ENOMEM);
2499                 err = add_to_page_cache_lru(page, mapping, index, gfp);
2500                 if (unlikely(err)) {
2501                         put_page(page);
2502                         if (err == -EEXIST)
2503                                 goto repeat;
2504                         /* Presumably ENOMEM for radix tree node */
2505                         return ERR_PTR(err);
2506                 }
2507
2508 filler:
2509                 err = filler(data, page);
2510                 if (err < 0) {
2511                         put_page(page);
2512                         return ERR_PTR(err);
2513                 }
2514
2515                 page = wait_on_page_read(page);
2516                 if (IS_ERR(page))
2517                         return page;
2518                 goto out;
2519         }
2520         if (PageUptodate(page))
2521                 goto out;
2522
2523         /*
2524          * Page is not up to date and may be locked due one of the following
2525          * case a: Page is being filled and the page lock is held
2526          * case b: Read/write error clearing the page uptodate status
2527          * case c: Truncation in progress (page locked)
2528          * case d: Reclaim in progress
2529          *
2530          * Case a, the page will be up to date when the page is unlocked.
2531          *    There is no need to serialise on the page lock here as the page
2532          *    is pinned so the lock gives no additional protection. Even if the
2533          *    the page is truncated, the data is still valid if PageUptodate as
2534          *    it's a race vs truncate race.
2535          * Case b, the page will not be up to date
2536          * Case c, the page may be truncated but in itself, the data may still
2537          *    be valid after IO completes as it's a read vs truncate race. The
2538          *    operation must restart if the page is not uptodate on unlock but
2539          *    otherwise serialising on page lock to stabilise the mapping gives
2540          *    no additional guarantees to the caller as the page lock is
2541          *    released before return.
2542          * Case d, similar to truncation. If reclaim holds the page lock, it
2543          *    will be a race with remove_mapping that determines if the mapping
2544          *    is valid on unlock but otherwise the data is valid and there is
2545          *    no need to serialise with page lock.
2546          *
2547          * As the page lock gives no additional guarantee, we optimistically
2548          * wait on the page to be unlocked and check if it's up to date and
2549          * use the page if it is. Otherwise, the page lock is required to
2550          * distinguish between the different cases. The motivation is that we
2551          * avoid spurious serialisations and wakeups when multiple processes
2552          * wait on the same page for IO to complete.
2553          */
2554         wait_on_page_locked(page);
2555         if (PageUptodate(page))
2556                 goto out;
2557
2558         /* Distinguish between all the cases under the safety of the lock */
2559         lock_page(page);
2560
2561         /* Case c or d, restart the operation */
2562         if (!page->mapping) {
2563                 unlock_page(page);
2564                 put_page(page);
2565                 goto repeat;
2566         }
2567
2568         /* Someone else locked and filled the page in a very small window */
2569         if (PageUptodate(page)) {
2570                 unlock_page(page);
2571                 goto out;
2572         }
2573         goto filler;
2574
2575 out:
2576         mark_page_accessed(page);
2577         return page;
2578 }
2579
2580 /**
2581  * read_cache_page - read into page cache, fill it if needed
2582  * @mapping:    the page's address_space
2583  * @index:      the page index
2584  * @filler:     function to perform the read
2585  * @data:       first arg to filler(data, page) function, often left as NULL
2586  *
2587  * Read into the page cache. If a page already exists, and PageUptodate() is
2588  * not set, try to fill the page and wait for it to become unlocked.
2589  *
2590  * If the page does not get brought uptodate, return -EIO.
2591  */
2592 struct page *read_cache_page(struct address_space *mapping,
2593                                 pgoff_t index,
2594                                 int (*filler)(void *, struct page *),
2595                                 void *data)
2596 {
2597         return do_read_cache_page(mapping, index, filler, data, mapping_gfp_mask(mapping));
2598 }
2599 EXPORT_SYMBOL(read_cache_page);
2600
2601 /**
2602  * read_cache_page_gfp - read into page cache, using specified page allocation flags.
2603  * @mapping:    the page's address_space
2604  * @index:      the page index
2605  * @gfp:        the page allocator flags to use if allocating
2606  *
2607  * This is the same as "read_mapping_page(mapping, index, NULL)", but with
2608  * any new page allocations done using the specified allocation flags.
2609  *
2610  * If the page does not get brought uptodate, return -EIO.
2611  */
2612 struct page *read_cache_page_gfp(struct address_space *mapping,
2613                                 pgoff_t index,
2614                                 gfp_t gfp)
2615 {
2616         filler_t *filler = (filler_t *)mapping->a_ops->readpage;
2617
2618         return do_read_cache_page(mapping, index, filler, NULL, gfp);
2619 }
2620 EXPORT_SYMBOL(read_cache_page_gfp);
2621
2622 /*
2623  * Performs necessary checks before doing a write
2624  *
2625  * Can adjust writing position or amount of bytes to write.
2626  * Returns appropriate error code that caller should return or
2627  * zero in case that write should be allowed.
2628  */
2629 inline ssize_t generic_write_checks(struct kiocb *iocb, struct iov_iter *from)
2630 {
2631         struct file *file = iocb->ki_filp;
2632         struct inode *inode = file->f_mapping->host;
2633         unsigned long limit = rlimit(RLIMIT_FSIZE);
2634         loff_t pos;
2635
2636         if (!iov_iter_count(from))
2637                 return 0;
2638
2639         /* FIXME: this is for backwards compatibility with 2.4 */
2640         if (iocb->ki_flags & IOCB_APPEND)
2641                 iocb->ki_pos = i_size_read(inode);
2642
2643         pos = iocb->ki_pos;
2644
2645         if (limit != RLIM_INFINITY) {
2646                 if (iocb->ki_pos >= limit) {
2647                         send_sig(SIGXFSZ, current, 0);
2648                         return -EFBIG;
2649                 }
2650                 iov_iter_truncate(from, limit - (unsigned long)pos);
2651         }
2652
2653         /*
2654          * LFS rule
2655          */
2656         if (unlikely(pos + iov_iter_count(from) > MAX_NON_LFS &&
2657                                 !(file->f_flags & O_LARGEFILE))) {
2658                 if (pos >= MAX_NON_LFS)
2659                         return -EFBIG;
2660                 iov_iter_truncate(from, MAX_NON_LFS - (unsigned long)pos);
2661         }
2662
2663         /*
2664          * Are we about to exceed the fs block limit ?
2665          *
2666          * If we have written data it becomes a short write.  If we have
2667          * exceeded without writing data we send a signal and return EFBIG.
2668          * Linus frestrict idea will clean these up nicely..
2669          */
2670         if (unlikely(pos >= inode->i_sb->s_maxbytes))
2671                 return -EFBIG;
2672
2673         iov_iter_truncate(from, inode->i_sb->s_maxbytes - pos);
2674         return iov_iter_count(from);
2675 }
2676 EXPORT_SYMBOL(generic_write_checks);
2677
2678 int pagecache_write_begin(struct file *file, struct address_space *mapping,
2679                                 loff_t pos, unsigned len, unsigned flags,
2680                                 struct page **pagep, void **fsdata)
2681 {
2682         const struct address_space_operations *aops = mapping->a_ops;
2683
2684         return aops->write_begin(file, mapping, pos, len, flags,
2685                                                         pagep, fsdata);
2686 }
2687 EXPORT_SYMBOL(pagecache_write_begin);
2688
2689 int pagecache_write_end(struct file *file, struct address_space *mapping,
2690                                 loff_t pos, unsigned len, unsigned copied,
2691                                 struct page *page, void *fsdata)
2692 {
2693         const struct address_space_operations *aops = mapping->a_ops;
2694
2695         return aops->write_end(file, mapping, pos, len, copied, page, fsdata);
2696 }
2697 EXPORT_SYMBOL(pagecache_write_end);
2698
2699 ssize_t
2700 generic_file_direct_write(struct kiocb *iocb, struct iov_iter *from)
2701 {
2702         struct file     *file = iocb->ki_filp;
2703         struct address_space *mapping = file->f_mapping;
2704         struct inode    *inode = mapping->host;
2705         loff_t          pos = iocb->ki_pos;
2706         ssize_t         written;
2707         size_t          write_len;
2708         pgoff_t         end;
2709
2710         write_len = iov_iter_count(from);
2711         end = (pos + write_len - 1) >> PAGE_SHIFT;
2712
2713         written = filemap_write_and_wait_range(mapping, pos, pos + write_len - 1);
2714         if (written)
2715                 goto out;
2716
2717         /*
2718          * After a write we want buffered reads to be sure to go to disk to get
2719          * the new data.  We invalidate clean cached page from the region we're
2720          * about to write.  We do this *before* the write so that we can return
2721          * without clobbering -EIOCBQUEUED from ->direct_IO().
2722          */
2723         written = invalidate_inode_pages2_range(mapping,
2724                                         pos >> PAGE_SHIFT, end);
2725         /*
2726          * If a page can not be invalidated, return 0 to fall back
2727          * to buffered write.
2728          */
2729         if (written) {
2730                 if (written == -EBUSY)
2731                         return 0;
2732                 goto out;
2733         }
2734
2735         written = mapping->a_ops->direct_IO(iocb, from);
2736
2737         /*
2738          * Finally, try again to invalidate clean pages which might have been
2739          * cached by non-direct readahead, or faulted in by get_user_pages()
2740          * if the source of the write was an mmap'ed region of the file
2741          * we're writing.  Either one is a pretty crazy thing to do,
2742          * so we don't support it 100%.  If this invalidation
2743          * fails, tough, the write still worked...
2744          */
2745         invalidate_inode_pages2_range(mapping,
2746                                 pos >> PAGE_SHIFT, end);
2747
2748         if (written > 0) {
2749                 pos += written;
2750                 write_len -= written;
2751                 if (pos > i_size_read(inode) && !S_ISBLK(inode->i_mode)) {
2752                         i_size_write(inode, pos);
2753                         mark_inode_dirty(inode);
2754                 }
2755                 iocb->ki_pos = pos;
2756         }
2757         iov_iter_revert(from, write_len - iov_iter_count(from));
2758 out:
2759         return written;
2760 }
2761 EXPORT_SYMBOL(generic_file_direct_write);
2762
2763 /*
2764  * Find or create a page at the given pagecache position. Return the locked
2765  * page. This function is specifically for buffered writes.
2766  */
2767 struct page *grab_cache_page_write_begin(struct address_space *mapping,
2768                                         pgoff_t index, unsigned flags)
2769 {
2770         struct page *page;
2771         int fgp_flags = FGP_LOCK|FGP_WRITE|FGP_CREAT;
2772
2773         if (flags & AOP_FLAG_NOFS)
2774                 fgp_flags |= FGP_NOFS;
2775
2776         page = pagecache_get_page(mapping, index, fgp_flags,
2777                         mapping_gfp_mask(mapping));
2778         if (page)
2779                 wait_for_stable_page(page);
2780
2781         return page;
2782 }
2783 EXPORT_SYMBOL(grab_cache_page_write_begin);
2784
2785 ssize_t generic_perform_write(struct file *file,
2786                                 struct iov_iter *i, loff_t pos)
2787 {
2788         struct address_space *mapping = file->f_mapping;
2789         const struct address_space_operations *a_ops = mapping->a_ops;
2790         long status = 0;
2791         ssize_t written = 0;
2792         unsigned int flags = 0;
2793
2794         do {
2795                 struct page *page;
2796                 unsigned long offset;   /* Offset into pagecache page */
2797                 unsigned long bytes;    /* Bytes to write to page */
2798                 size_t copied;          /* Bytes copied from user */
2799                 void *fsdata;
2800
2801                 offset = (pos & (PAGE_SIZE - 1));
2802                 bytes = min_t(unsigned long, PAGE_SIZE - offset,
2803                                                 iov_iter_count(i));
2804
2805 again:
2806                 /*
2807                  * Bring in the user page that we will copy from _first_.
2808                  * Otherwise there's a nasty deadlock on copying from the
2809                  * same page as we're writing to, without it being marked
2810                  * up-to-date.
2811                  *
2812                  * Not only is this an optimisation, but it is also required
2813                  * to check that the address is actually valid, when atomic
2814                  * usercopies are used, below.
2815                  */
2816                 if (unlikely(iov_iter_fault_in_readable(i, bytes))) {
2817                         status = -EFAULT;
2818                         break;
2819                 }
2820
2821                 if (fatal_signal_pending(current)) {
2822                         status = -EINTR;
2823                         break;
2824                 }
2825
2826                 status = a_ops->write_begin(file, mapping, pos, bytes, flags,
2827                                                 &page, &fsdata);
2828                 if (unlikely(status < 0))
2829                         break;
2830
2831                 if (mapping_writably_mapped(mapping))
2832                         flush_dcache_page(page);
2833
2834                 copied = iov_iter_copy_from_user_atomic(page, i, offset, bytes);
2835                 flush_dcache_page(page);
2836
2837                 status = a_ops->write_end(file, mapping, pos, bytes, copied,
2838                                                 page, fsdata);
2839                 if (unlikely(status < 0))
2840                         break;
2841                 copied = status;
2842
2843                 cond_resched();
2844
2845                 iov_iter_advance(i, copied);
2846                 if (unlikely(copied == 0)) {
2847                         /*
2848                          * If we were unable to copy any data at all, we must
2849                          * fall back to a single segment length write.
2850                          *
2851                          * If we didn't fallback here, we could livelock
2852                          * because not all segments in the iov can be copied at
2853                          * once without a pagefault.
2854                          */
2855                         bytes = min_t(unsigned long, PAGE_SIZE - offset,
2856                                                 iov_iter_single_seg_count(i));
2857                         goto again;
2858                 }
2859                 pos += copied;
2860                 written += copied;
2861
2862                 balance_dirty_pages_ratelimited(mapping);
2863         } while (iov_iter_count(i));
2864
2865         return written ? written : status;
2866 }
2867 EXPORT_SYMBOL(generic_perform_write);
2868
2869 /**
2870  * __generic_file_write_iter - write data to a file
2871  * @iocb:       IO state structure (file, offset, etc.)
2872  * @from:       iov_iter with data to write
2873  *
2874  * This function does all the work needed for actually writing data to a
2875  * file. It does all basic checks, removes SUID from the file, updates
2876  * modification times and calls proper subroutines depending on whether we
2877  * do direct IO or a standard buffered write.
2878  *
2879  * It expects i_mutex to be grabbed unless we work on a block device or similar
2880  * object which does not need locking at all.
2881  *
2882  * This function does *not* take care of syncing data in case of O_SYNC write.
2883  * A caller has to handle it. This is mainly due to the fact that we want to
2884  * avoid syncing under i_mutex.
2885  */
2886 ssize_t __generic_file_write_iter(struct kiocb *iocb, struct iov_iter *from)
2887 {
2888         struct file *file = iocb->ki_filp;
2889         struct address_space * mapping = file->f_mapping;
2890         struct inode    *inode = mapping->host;
2891         ssize_t         written = 0;
2892         ssize_t         err;
2893         ssize_t         status;
2894
2895         /* We can write back this queue in page reclaim */
2896         current->backing_dev_info = inode_to_bdi(inode);
2897         err = file_remove_privs(file);
2898         if (err)
2899                 goto out;
2900
2901         err = file_update_time(file);
2902         if (err)
2903                 goto out;
2904
2905         if (iocb->ki_flags & IOCB_DIRECT) {
2906                 loff_t pos, endbyte;
2907
2908                 written = generic_file_direct_write(iocb, from);
2909                 /*
2910                  * If the write stopped short of completing, fall back to
2911                  * buffered writes.  Some filesystems do this for writes to
2912                  * holes, for example.  For DAX files, a buffered write will
2913                  * not succeed (even if it did, DAX does not handle dirty
2914                  * page-cache pages correctly).
2915                  */
2916                 if (written < 0 || !iov_iter_count(from) || IS_DAX(inode))
2917                         goto out;
2918
2919                 status = generic_perform_write(file, from, pos = iocb->ki_pos);
2920                 /*
2921                  * If generic_perform_write() returned a synchronous error
2922                  * then we want to return the number of bytes which were
2923                  * direct-written, or the error code if that was zero.  Note
2924                  * that this differs from normal direct-io semantics, which
2925                  * will return -EFOO even if some bytes were written.
2926                  */
2927                 if (unlikely(status < 0)) {
2928                         err = status;
2929                         goto out;
2930                 }
2931                 /*
2932                  * We need to ensure that the page cache pages are written to
2933                  * disk and invalidated to preserve the expected O_DIRECT
2934                  * semantics.
2935                  */
2936                 endbyte = pos + status - 1;
2937                 err = filemap_write_and_wait_range(mapping, pos, endbyte);
2938                 if (err == 0) {
2939                         iocb->ki_pos = endbyte + 1;
2940                         written += status;
2941                         invalidate_mapping_pages(mapping,
2942                                                  pos >> PAGE_SHIFT,
2943                                                  endbyte >> PAGE_SHIFT);
2944                 } else {
2945                         /*
2946                          * We don't know how much we wrote, so just return
2947                          * the number of bytes which were direct-written
2948                          */
2949                 }
2950         } else {
2951                 written = generic_perform_write(file, from, iocb->ki_pos);
2952                 if (likely(written > 0))
2953                         iocb->ki_pos += written;
2954         }
2955 out:
2956         current->backing_dev_info = NULL;
2957         return written ? written : err;
2958 }
2959 EXPORT_SYMBOL(__generic_file_write_iter);
2960
2961 /**
2962  * generic_file_write_iter - write data to a file
2963  * @iocb:       IO state structure
2964  * @from:       iov_iter with data to write
2965  *
2966  * This is a wrapper around __generic_file_write_iter() to be used by most
2967  * filesystems. It takes care of syncing the file in case of O_SYNC file
2968  * and acquires i_mutex as needed.
2969  */
2970 ssize_t generic_file_write_iter(struct kiocb *iocb, struct iov_iter *from)
2971 {
2972         struct file *file = iocb->ki_filp;
2973         struct inode *inode = file->f_mapping->host;
2974         ssize_t ret;
2975
2976         inode_lock(inode);
2977         ret = generic_write_checks(iocb, from);
2978         if (ret > 0)
2979                 ret = __generic_file_write_iter(iocb, from);
2980         inode_unlock(inode);
2981
2982         if (ret > 0)
2983                 ret = generic_write_sync(iocb, ret);
2984         return ret;
2985 }
2986 EXPORT_SYMBOL(generic_file_write_iter);
2987
2988 /**
2989  * try_to_release_page() - release old fs-specific metadata on a page
2990  *
2991  * @page: the page which the kernel is trying to free
2992  * @gfp_mask: memory allocation flags (and I/O mode)
2993  *
2994  * The address_space is to try to release any data against the page
2995  * (presumably at page->private).  If the release was successful, return '1'.
2996  * Otherwise return zero.
2997  *
2998  * This may also be called if PG_fscache is set on a page, indicating that the
2999  * page is known to the local caching routines.
3000  *
3001  * The @gfp_mask argument specifies whether I/O may be performed to release
3002  * this page (__GFP_IO), and whether the call may block (__GFP_RECLAIM & __GFP_FS).
3003  *
3004  */
3005 int try_to_release_page(struct page *page, gfp_t gfp_mask)
3006 {
3007         struct address_space * const mapping = page->mapping;
3008
3009         BUG_ON(!PageLocked(page));
3010         if (PageWriteback(page))
3011                 return 0;
3012
3013         if (mapping && mapping->a_ops->releasepage)
3014                 return mapping->a_ops->releasepage(page, gfp_mask);
3015         return try_to_free_buffers(page);
3016 }
3017
3018 EXPORT_SYMBOL(try_to_release_page);