mm: thp: set the accessed flag for old pages on access fault
[pandora-kernel.git] / mm / memory.c
1 /*
2  *  linux/mm/memory.c
3  *
4  *  Copyright (C) 1991, 1992, 1993, 1994  Linus Torvalds
5  */
6
7 /*
8  * demand-loading started 01.12.91 - seems it is high on the list of
9  * things wanted, and it should be easy to implement. - Linus
10  */
11
12 /*
13  * Ok, demand-loading was easy, shared pages a little bit tricker. Shared
14  * pages started 02.12.91, seems to work. - Linus.
15  *
16  * Tested sharing by executing about 30 /bin/sh: under the old kernel it
17  * would have taken more than the 6M I have free, but it worked well as
18  * far as I could see.
19  *
20  * Also corrected some "invalidate()"s - I wasn't doing enough of them.
21  */
22
23 /*
24  * Real VM (paging to/from disk) started 18.12.91. Much more work and
25  * thought has to go into this. Oh, well..
26  * 19.12.91  -  works, somewhat. Sometimes I get faults, don't know why.
27  *              Found it. Everything seems to work now.
28  * 20.12.91  -  Ok, making the swap-device changeable like the root.
29  */
30
31 /*
32  * 05.04.94  -  Multi-page memory management added for v1.1.
33  *              Idea by Alex Bligh (alex@cconcepts.co.uk)
34  *
35  * 16.07.99  -  Support of BIGMEM added by Gerhard Wichert, Siemens AG
36  *              (Gerhard.Wichert@pdb.siemens.de)
37  *
38  * Aug/Sep 2004 Changed to four level page tables (Andi Kleen)
39  */
40
41 #include <linux/kernel_stat.h>
42 #include <linux/mm.h>
43 #include <linux/hugetlb.h>
44 #include <linux/mman.h>
45 #include <linux/swap.h>
46 #include <linux/highmem.h>
47 #include <linux/pagemap.h>
48 #include <linux/ksm.h>
49 #include <linux/rmap.h>
50 #include <linux/export.h>
51 #include <linux/delayacct.h>
52 #include <linux/init.h>
53 #include <linux/writeback.h>
54 #include <linux/memcontrol.h>
55 #include <linux/mmu_notifier.h>
56 #include <linux/kallsyms.h>
57 #include <linux/swapops.h>
58 #include <linux/elf.h>
59 #include <linux/gfp.h>
60
61 #include <asm/io.h>
62 #include <asm/pgalloc.h>
63 #include <asm/uaccess.h>
64 #include <asm/tlb.h>
65 #include <asm/tlbflush.h>
66 #include <asm/pgtable.h>
67
68 #include "internal.h"
69
70 #ifndef CONFIG_NEED_MULTIPLE_NODES
71 /* use the per-pgdat data instead for discontigmem - mbligh */
72 unsigned long max_mapnr;
73 struct page *mem_map;
74
75 EXPORT_SYMBOL(max_mapnr);
76 EXPORT_SYMBOL(mem_map);
77 #endif
78
79 unsigned long num_physpages;
80 /*
81  * A number of key systems in x86 including ioremap() rely on the assumption
82  * that high_memory defines the upper bound on direct map memory, then end
83  * of ZONE_NORMAL.  Under CONFIG_DISCONTIG this means that max_low_pfn and
84  * highstart_pfn must be the same; there must be no gap between ZONE_NORMAL
85  * and ZONE_HIGHMEM.
86  */
87 void * high_memory;
88
89 EXPORT_SYMBOL(num_physpages);
90 EXPORT_SYMBOL(high_memory);
91
92 /*
93  * Randomize the address space (stacks, mmaps, brk, etc.).
94  *
95  * ( When CONFIG_COMPAT_BRK=y we exclude brk from randomization,
96  *   as ancient (libc5 based) binaries can segfault. )
97  */
98 int randomize_va_space __read_mostly =
99 #ifdef CONFIG_COMPAT_BRK
100                                         1;
101 #else
102                                         2;
103 #endif
104
105 static int __init disable_randmaps(char *s)
106 {
107         randomize_va_space = 0;
108         return 1;
109 }
110 __setup("norandmaps", disable_randmaps);
111
112 unsigned long zero_pfn __read_mostly;
113 unsigned long highest_memmap_pfn __read_mostly;
114
115 /*
116  * CONFIG_MMU architectures set up ZERO_PAGE in their paging_init()
117  */
118 static int __init init_zero_pfn(void)
119 {
120         zero_pfn = page_to_pfn(ZERO_PAGE(0));
121         return 0;
122 }
123 core_initcall(init_zero_pfn);
124
125
126 #if defined(SPLIT_RSS_COUNTING)
127
128 static void __sync_task_rss_stat(struct task_struct *task, struct mm_struct *mm)
129 {
130         int i;
131
132         for (i = 0; i < NR_MM_COUNTERS; i++) {
133                 if (task->rss_stat.count[i]) {
134                         add_mm_counter(mm, i, task->rss_stat.count[i]);
135                         task->rss_stat.count[i] = 0;
136                 }
137         }
138         task->rss_stat.events = 0;
139 }
140
141 static void add_mm_counter_fast(struct mm_struct *mm, int member, int val)
142 {
143         struct task_struct *task = current;
144
145         if (likely(task->mm == mm))
146                 task->rss_stat.count[member] += val;
147         else
148                 add_mm_counter(mm, member, val);
149 }
150 #define inc_mm_counter_fast(mm, member) add_mm_counter_fast(mm, member, 1)
151 #define dec_mm_counter_fast(mm, member) add_mm_counter_fast(mm, member, -1)
152
153 /* sync counter once per 64 page faults */
154 #define TASK_RSS_EVENTS_THRESH  (64)
155 static void check_sync_rss_stat(struct task_struct *task)
156 {
157         if (unlikely(task != current))
158                 return;
159         if (unlikely(task->rss_stat.events++ > TASK_RSS_EVENTS_THRESH))
160                 __sync_task_rss_stat(task, task->mm);
161 }
162
163 unsigned long get_mm_counter(struct mm_struct *mm, int member)
164 {
165         long val = 0;
166
167         /*
168          * Don't use task->mm here...for avoiding to use task_get_mm()..
169          * The caller must guarantee task->mm is not invalid.
170          */
171         val = atomic_long_read(&mm->rss_stat.count[member]);
172         /*
173          * counter is updated in asynchronous manner and may go to minus.
174          * But it's never be expected number for users.
175          */
176         if (val < 0)
177                 return 0;
178         return (unsigned long)val;
179 }
180
181 void sync_mm_rss(struct task_struct *task, struct mm_struct *mm)
182 {
183         __sync_task_rss_stat(task, mm);
184 }
185 #else /* SPLIT_RSS_COUNTING */
186
187 #define inc_mm_counter_fast(mm, member) inc_mm_counter(mm, member)
188 #define dec_mm_counter_fast(mm, member) dec_mm_counter(mm, member)
189
190 static void check_sync_rss_stat(struct task_struct *task)
191 {
192 }
193
194 #endif /* SPLIT_RSS_COUNTING */
195
196 #ifdef HAVE_GENERIC_MMU_GATHER
197
198 static int tlb_next_batch(struct mmu_gather *tlb)
199 {
200         struct mmu_gather_batch *batch;
201
202         batch = tlb->active;
203         if (batch->next) {
204                 tlb->active = batch->next;
205                 return 1;
206         }
207
208         if (tlb->batch_count == MAX_GATHER_BATCH_COUNT)
209                 return 0;
210
211         batch = (void *)__get_free_pages(GFP_NOWAIT | __GFP_NOWARN, 0);
212         if (!batch)
213                 return 0;
214
215         tlb->batch_count++;
216         batch->next = NULL;
217         batch->nr   = 0;
218         batch->max  = MAX_GATHER_BATCH;
219
220         tlb->active->next = batch;
221         tlb->active = batch;
222
223         return 1;
224 }
225
226 /* tlb_gather_mmu
227  *      Called to initialize an (on-stack) mmu_gather structure for page-table
228  *      tear-down from @mm. The @fullmm argument is used when @mm is without
229  *      users and we're going to destroy the full address space (exit/execve).
230  */
231 void tlb_gather_mmu(struct mmu_gather *tlb, struct mm_struct *mm, bool fullmm)
232 {
233         tlb->mm = mm;
234
235         tlb->fullmm     = fullmm;
236         tlb->need_flush = 0;
237         tlb->fast_mode  = (num_possible_cpus() == 1);
238         tlb->local.next = NULL;
239         tlb->local.nr   = 0;
240         tlb->local.max  = ARRAY_SIZE(tlb->__pages);
241         tlb->active     = &tlb->local;
242         tlb->batch_count = 0;
243
244 #ifdef CONFIG_HAVE_RCU_TABLE_FREE
245         tlb->batch = NULL;
246 #endif
247 }
248
249 void tlb_flush_mmu(struct mmu_gather *tlb)
250 {
251         struct mmu_gather_batch *batch;
252
253         if (!tlb->need_flush)
254                 return;
255         tlb->need_flush = 0;
256         tlb_flush(tlb);
257 #ifdef CONFIG_HAVE_RCU_TABLE_FREE
258         tlb_table_flush(tlb);
259 #endif
260
261         if (tlb_fast_mode(tlb))
262                 return;
263
264         for (batch = &tlb->local; batch; batch = batch->next) {
265                 free_pages_and_swap_cache(batch->pages, batch->nr);
266                 batch->nr = 0;
267         }
268         tlb->active = &tlb->local;
269 }
270
271 /* tlb_finish_mmu
272  *      Called at the end of the shootdown operation to free up any resources
273  *      that were required.
274  */
275 void tlb_finish_mmu(struct mmu_gather *tlb, unsigned long start, unsigned long end)
276 {
277         struct mmu_gather_batch *batch, *next;
278
279         tlb_flush_mmu(tlb);
280
281         /* keep the page table cache within bounds */
282         check_pgt_cache();
283
284         for (batch = tlb->local.next; batch; batch = next) {
285                 next = batch->next;
286                 free_pages((unsigned long)batch, 0);
287         }
288         tlb->local.next = NULL;
289 }
290
291 /* __tlb_remove_page
292  *      Must perform the equivalent to __free_pte(pte_get_and_clear(ptep)), while
293  *      handling the additional races in SMP caused by other CPUs caching valid
294  *      mappings in their TLBs. Returns the number of free page slots left.
295  *      When out of page slots we must call tlb_flush_mmu().
296  */
297 int __tlb_remove_page(struct mmu_gather *tlb, struct page *page)
298 {
299         struct mmu_gather_batch *batch;
300
301         tlb->need_flush = 1;
302
303         if (tlb_fast_mode(tlb)) {
304                 free_page_and_swap_cache(page);
305                 return 1; /* avoid calling tlb_flush_mmu() */
306         }
307
308         batch = tlb->active;
309         batch->pages[batch->nr++] = page;
310         if (batch->nr == batch->max) {
311                 if (!tlb_next_batch(tlb))
312                         return 0;
313                 batch = tlb->active;
314         }
315         VM_BUG_ON(batch->nr > batch->max);
316
317         return batch->max - batch->nr;
318 }
319
320 #endif /* HAVE_GENERIC_MMU_GATHER */
321
322 #ifdef CONFIG_HAVE_RCU_TABLE_FREE
323
324 /*
325  * See the comment near struct mmu_table_batch.
326  */
327
328 static void tlb_remove_table_smp_sync(void *arg)
329 {
330         /* Simply deliver the interrupt */
331 }
332
333 static void tlb_remove_table_one(void *table)
334 {
335         /*
336          * This isn't an RCU grace period and hence the page-tables cannot be
337          * assumed to be actually RCU-freed.
338          *
339          * It is however sufficient for software page-table walkers that rely on
340          * IRQ disabling. See the comment near struct mmu_table_batch.
341          */
342         smp_call_function(tlb_remove_table_smp_sync, NULL, 1);
343         __tlb_remove_table(table);
344 }
345
346 static void tlb_remove_table_rcu(struct rcu_head *head)
347 {
348         struct mmu_table_batch *batch;
349         int i;
350
351         batch = container_of(head, struct mmu_table_batch, rcu);
352
353         for (i = 0; i < batch->nr; i++)
354                 __tlb_remove_table(batch->tables[i]);
355
356         free_page((unsigned long)batch);
357 }
358
359 void tlb_table_flush(struct mmu_gather *tlb)
360 {
361         struct mmu_table_batch **batch = &tlb->batch;
362
363         if (*batch) {
364                 call_rcu_sched(&(*batch)->rcu, tlb_remove_table_rcu);
365                 *batch = NULL;
366         }
367 }
368
369 void tlb_remove_table(struct mmu_gather *tlb, void *table)
370 {
371         struct mmu_table_batch **batch = &tlb->batch;
372
373         tlb->need_flush = 1;
374
375         /*
376          * When there's less then two users of this mm there cannot be a
377          * concurrent page-table walk.
378          */
379         if (atomic_read(&tlb->mm->mm_users) < 2) {
380                 __tlb_remove_table(table);
381                 return;
382         }
383
384         if (*batch == NULL) {
385                 *batch = (struct mmu_table_batch *)__get_free_page(GFP_NOWAIT | __GFP_NOWARN);
386                 if (*batch == NULL) {
387                         tlb_remove_table_one(table);
388                         return;
389                 }
390                 (*batch)->nr = 0;
391         }
392         (*batch)->tables[(*batch)->nr++] = table;
393         if ((*batch)->nr == MAX_TABLE_BATCH)
394                 tlb_table_flush(tlb);
395 }
396
397 #endif /* CONFIG_HAVE_RCU_TABLE_FREE */
398
399 /*
400  * If a p?d_bad entry is found while walking page tables, report
401  * the error, before resetting entry to p?d_none.  Usually (but
402  * very seldom) called out from the p?d_none_or_clear_bad macros.
403  */
404
405 void pgd_clear_bad(pgd_t *pgd)
406 {
407         pgd_ERROR(*pgd);
408         pgd_clear(pgd);
409 }
410
411 void pud_clear_bad(pud_t *pud)
412 {
413         pud_ERROR(*pud);
414         pud_clear(pud);
415 }
416
417 void pmd_clear_bad(pmd_t *pmd)
418 {
419         pmd_ERROR(*pmd);
420         pmd_clear(pmd);
421 }
422
423 /*
424  * Note: this doesn't free the actual pages themselves. That
425  * has been handled earlier when unmapping all the memory regions.
426  */
427 static void free_pte_range(struct mmu_gather *tlb, pmd_t *pmd,
428                            unsigned long addr)
429 {
430         pgtable_t token = pmd_pgtable(*pmd);
431         pmd_clear(pmd);
432         pte_free_tlb(tlb, token, addr);
433         tlb->mm->nr_ptes--;
434 }
435
436 static inline void free_pmd_range(struct mmu_gather *tlb, pud_t *pud,
437                                 unsigned long addr, unsigned long end,
438                                 unsigned long floor, unsigned long ceiling)
439 {
440         pmd_t *pmd;
441         unsigned long next;
442         unsigned long start;
443
444         start = addr;
445         pmd = pmd_offset(pud, addr);
446         do {
447                 next = pmd_addr_end(addr, end);
448                 if (pmd_none_or_clear_bad(pmd))
449                         continue;
450                 free_pte_range(tlb, pmd, addr);
451         } while (pmd++, addr = next, addr != end);
452
453         start &= PUD_MASK;
454         if (start < floor)
455                 return;
456         if (ceiling) {
457                 ceiling &= PUD_MASK;
458                 if (!ceiling)
459                         return;
460         }
461         if (end - 1 > ceiling - 1)
462                 return;
463
464         pmd = pmd_offset(pud, start);
465         pud_clear(pud);
466         pmd_free_tlb(tlb, pmd, start);
467 }
468
469 static inline void free_pud_range(struct mmu_gather *tlb, pgd_t *pgd,
470                                 unsigned long addr, unsigned long end,
471                                 unsigned long floor, unsigned long ceiling)
472 {
473         pud_t *pud;
474         unsigned long next;
475         unsigned long start;
476
477         start = addr;
478         pud = pud_offset(pgd, addr);
479         do {
480                 next = pud_addr_end(addr, end);
481                 if (pud_none_or_clear_bad(pud))
482                         continue;
483                 free_pmd_range(tlb, pud, addr, next, floor, ceiling);
484         } while (pud++, addr = next, addr != end);
485
486         start &= PGDIR_MASK;
487         if (start < floor)
488                 return;
489         if (ceiling) {
490                 ceiling &= PGDIR_MASK;
491                 if (!ceiling)
492                         return;
493         }
494         if (end - 1 > ceiling - 1)
495                 return;
496
497         pud = pud_offset(pgd, start);
498         pgd_clear(pgd);
499         pud_free_tlb(tlb, pud, start);
500 }
501
502 /*
503  * This function frees user-level page tables of a process.
504  *
505  * Must be called with pagetable lock held.
506  */
507 void free_pgd_range(struct mmu_gather *tlb,
508                         unsigned long addr, unsigned long end,
509                         unsigned long floor, unsigned long ceiling)
510 {
511         pgd_t *pgd;
512         unsigned long next;
513
514         /*
515          * The next few lines have given us lots of grief...
516          *
517          * Why are we testing PMD* at this top level?  Because often
518          * there will be no work to do at all, and we'd prefer not to
519          * go all the way down to the bottom just to discover that.
520          *
521          * Why all these "- 1"s?  Because 0 represents both the bottom
522          * of the address space and the top of it (using -1 for the
523          * top wouldn't help much: the masks would do the wrong thing).
524          * The rule is that addr 0 and floor 0 refer to the bottom of
525          * the address space, but end 0 and ceiling 0 refer to the top
526          * Comparisons need to use "end - 1" and "ceiling - 1" (though
527          * that end 0 case should be mythical).
528          *
529          * Wherever addr is brought up or ceiling brought down, we must
530          * be careful to reject "the opposite 0" before it confuses the
531          * subsequent tests.  But what about where end is brought down
532          * by PMD_SIZE below? no, end can't go down to 0 there.
533          *
534          * Whereas we round start (addr) and ceiling down, by different
535          * masks at different levels, in order to test whether a table
536          * now has no other vmas using it, so can be freed, we don't
537          * bother to round floor or end up - the tests don't need that.
538          */
539
540         addr &= PMD_MASK;
541         if (addr < floor) {
542                 addr += PMD_SIZE;
543                 if (!addr)
544                         return;
545         }
546         if (ceiling) {
547                 ceiling &= PMD_MASK;
548                 if (!ceiling)
549                         return;
550         }
551         if (end - 1 > ceiling - 1)
552                 end -= PMD_SIZE;
553         if (addr > end - 1)
554                 return;
555
556         pgd = pgd_offset(tlb->mm, addr);
557         do {
558                 next = pgd_addr_end(addr, end);
559                 if (pgd_none_or_clear_bad(pgd))
560                         continue;
561                 free_pud_range(tlb, pgd, addr, next, floor, ceiling);
562         } while (pgd++, addr = next, addr != end);
563 }
564
565 void free_pgtables(struct mmu_gather *tlb, struct vm_area_struct *vma,
566                 unsigned long floor, unsigned long ceiling)
567 {
568         while (vma) {
569                 struct vm_area_struct *next = vma->vm_next;
570                 unsigned long addr = vma->vm_start;
571
572                 /*
573                  * Hide vma from rmap and truncate_pagecache before freeing
574                  * pgtables
575                  */
576                 unlink_anon_vmas(vma);
577                 unlink_file_vma(vma);
578
579                 if (is_vm_hugetlb_page(vma)) {
580                         hugetlb_free_pgd_range(tlb, addr, vma->vm_end,
581                                 floor, next? next->vm_start: ceiling);
582                 } else {
583                         /*
584                          * Optimization: gather nearby vmas into one call down
585                          */
586                         while (next && next->vm_start <= vma->vm_end + PMD_SIZE
587                                && !is_vm_hugetlb_page(next)) {
588                                 vma = next;
589                                 next = vma->vm_next;
590                                 unlink_anon_vmas(vma);
591                                 unlink_file_vma(vma);
592                         }
593                         free_pgd_range(tlb, addr, vma->vm_end,
594                                 floor, next? next->vm_start: ceiling);
595                 }
596                 vma = next;
597         }
598 }
599
600 int __pte_alloc(struct mm_struct *mm, struct vm_area_struct *vma,
601                 pmd_t *pmd, unsigned long address)
602 {
603         pgtable_t new = pte_alloc_one(mm, address);
604         int wait_split_huge_page;
605         if (!new)
606                 return -ENOMEM;
607
608         /*
609          * Ensure all pte setup (eg. pte page lock and page clearing) are
610          * visible before the pte is made visible to other CPUs by being
611          * put into page tables.
612          *
613          * The other side of the story is the pointer chasing in the page
614          * table walking code (when walking the page table without locking;
615          * ie. most of the time). Fortunately, these data accesses consist
616          * of a chain of data-dependent loads, meaning most CPUs (alpha
617          * being the notable exception) will already guarantee loads are
618          * seen in-order. See the alpha page table accessors for the
619          * smp_read_barrier_depends() barriers in page table walking code.
620          */
621         smp_wmb(); /* Could be smp_wmb__xxx(before|after)_spin_lock */
622
623         spin_lock(&mm->page_table_lock);
624         wait_split_huge_page = 0;
625         if (likely(pmd_none(*pmd))) {   /* Has another populated it ? */
626                 mm->nr_ptes++;
627                 pmd_populate(mm, pmd, new);
628                 new = NULL;
629         } else if (unlikely(pmd_trans_splitting(*pmd)))
630                 wait_split_huge_page = 1;
631         spin_unlock(&mm->page_table_lock);
632         if (new)
633                 pte_free(mm, new);
634         if (wait_split_huge_page)
635                 wait_split_huge_page(vma->anon_vma, pmd);
636         return 0;
637 }
638
639 int __pte_alloc_kernel(pmd_t *pmd, unsigned long address)
640 {
641         pte_t *new = pte_alloc_one_kernel(&init_mm, address);
642         if (!new)
643                 return -ENOMEM;
644
645         smp_wmb(); /* See comment in __pte_alloc */
646
647         spin_lock(&init_mm.page_table_lock);
648         if (likely(pmd_none(*pmd))) {   /* Has another populated it ? */
649                 pmd_populate_kernel(&init_mm, pmd, new);
650                 new = NULL;
651         } else
652                 VM_BUG_ON(pmd_trans_splitting(*pmd));
653         spin_unlock(&init_mm.page_table_lock);
654         if (new)
655                 pte_free_kernel(&init_mm, new);
656         return 0;
657 }
658
659 static inline void init_rss_vec(int *rss)
660 {
661         memset(rss, 0, sizeof(int) * NR_MM_COUNTERS);
662 }
663
664 static inline void add_mm_rss_vec(struct mm_struct *mm, int *rss)
665 {
666         int i;
667
668         if (current->mm == mm)
669                 sync_mm_rss(current, mm);
670         for (i = 0; i < NR_MM_COUNTERS; i++)
671                 if (rss[i])
672                         add_mm_counter(mm, i, rss[i]);
673 }
674
675 /*
676  * This function is called to print an error when a bad pte
677  * is found. For example, we might have a PFN-mapped pte in
678  * a region that doesn't allow it.
679  *
680  * The calling function must still handle the error.
681  */
682 static void print_bad_pte(struct vm_area_struct *vma, unsigned long addr,
683                           pte_t pte, struct page *page)
684 {
685         pgd_t *pgd = pgd_offset(vma->vm_mm, addr);
686         pud_t *pud = pud_offset(pgd, addr);
687         pmd_t *pmd = pmd_offset(pud, addr);
688         struct address_space *mapping;
689         pgoff_t index;
690         static unsigned long resume;
691         static unsigned long nr_shown;
692         static unsigned long nr_unshown;
693
694         /*
695          * Allow a burst of 60 reports, then keep quiet for that minute;
696          * or allow a steady drip of one report per second.
697          */
698         if (nr_shown == 60) {
699                 if (time_before(jiffies, resume)) {
700                         nr_unshown++;
701                         return;
702                 }
703                 if (nr_unshown) {
704                         printk(KERN_ALERT
705                                 "BUG: Bad page map: %lu messages suppressed\n",
706                                 nr_unshown);
707                         nr_unshown = 0;
708                 }
709                 nr_shown = 0;
710         }
711         if (nr_shown++ == 0)
712                 resume = jiffies + 60 * HZ;
713
714         mapping = vma->vm_file ? vma->vm_file->f_mapping : NULL;
715         index = linear_page_index(vma, addr);
716
717         printk(KERN_ALERT
718                 "BUG: Bad page map in process %s  pte:%08llx pmd:%08llx\n",
719                 current->comm,
720                 (long long)pte_val(pte), (long long)pmd_val(*pmd));
721         if (page)
722                 dump_page(page);
723         printk(KERN_ALERT
724                 "addr:%p vm_flags:%08lx anon_vma:%p mapping:%p index:%lx\n",
725                 (void *)addr, vma->vm_flags, vma->anon_vma, mapping, index);
726         /*
727          * Choose text because data symbols depend on CONFIG_KALLSYMS_ALL=y
728          */
729         if (vma->vm_ops)
730                 print_symbol(KERN_ALERT "vma->vm_ops->fault: %s\n",
731                                 (unsigned long)vma->vm_ops->fault);
732         if (vma->vm_file && vma->vm_file->f_op)
733                 print_symbol(KERN_ALERT "vma->vm_file->f_op->mmap: %s\n",
734                                 (unsigned long)vma->vm_file->f_op->mmap);
735         dump_stack();
736         add_taint(TAINT_BAD_PAGE);
737 }
738
739 static inline int is_cow_mapping(vm_flags_t flags)
740 {
741         return (flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE;
742 }
743
744 #ifndef is_zero_pfn
745 static inline int is_zero_pfn(unsigned long pfn)
746 {
747         return pfn == zero_pfn;
748 }
749 #endif
750
751 #ifndef my_zero_pfn
752 static inline unsigned long my_zero_pfn(unsigned long addr)
753 {
754         return zero_pfn;
755 }
756 #endif
757
758 /*
759  * vm_normal_page -- This function gets the "struct page" associated with a pte.
760  *
761  * "Special" mappings do not wish to be associated with a "struct page" (either
762  * it doesn't exist, or it exists but they don't want to touch it). In this
763  * case, NULL is returned here. "Normal" mappings do have a struct page.
764  *
765  * There are 2 broad cases. Firstly, an architecture may define a pte_special()
766  * pte bit, in which case this function is trivial. Secondly, an architecture
767  * may not have a spare pte bit, which requires a more complicated scheme,
768  * described below.
769  *
770  * A raw VM_PFNMAP mapping (ie. one that is not COWed) is always considered a
771  * special mapping (even if there are underlying and valid "struct pages").
772  * COWed pages of a VM_PFNMAP are always normal.
773  *
774  * The way we recognize COWed pages within VM_PFNMAP mappings is through the
775  * rules set up by "remap_pfn_range()": the vma will have the VM_PFNMAP bit
776  * set, and the vm_pgoff will point to the first PFN mapped: thus every special
777  * mapping will always honor the rule
778  *
779  *      pfn_of_page == vma->vm_pgoff + ((addr - vma->vm_start) >> PAGE_SHIFT)
780  *
781  * And for normal mappings this is false.
782  *
783  * This restricts such mappings to be a linear translation from virtual address
784  * to pfn. To get around this restriction, we allow arbitrary mappings so long
785  * as the vma is not a COW mapping; in that case, we know that all ptes are
786  * special (because none can have been COWed).
787  *
788  *
789  * In order to support COW of arbitrary special mappings, we have VM_MIXEDMAP.
790  *
791  * VM_MIXEDMAP mappings can likewise contain memory with or without "struct
792  * page" backing, however the difference is that _all_ pages with a struct
793  * page (that is, those where pfn_valid is true) are refcounted and considered
794  * normal pages by the VM. The disadvantage is that pages are refcounted
795  * (which can be slower and simply not an option for some PFNMAP users). The
796  * advantage is that we don't have to follow the strict linearity rule of
797  * PFNMAP mappings in order to support COWable mappings.
798  *
799  */
800 #ifdef __HAVE_ARCH_PTE_SPECIAL
801 # define HAVE_PTE_SPECIAL 1
802 #else
803 # define HAVE_PTE_SPECIAL 0
804 #endif
805 struct page *vm_normal_page(struct vm_area_struct *vma, unsigned long addr,
806                                 pte_t pte)
807 {
808         unsigned long pfn = pte_pfn(pte);
809
810         if (HAVE_PTE_SPECIAL) {
811                 if (likely(!pte_special(pte)))
812                         goto check_pfn;
813                 if (vma->vm_flags & (VM_PFNMAP | VM_MIXEDMAP))
814                         return NULL;
815                 if (!is_zero_pfn(pfn))
816                         print_bad_pte(vma, addr, pte, NULL);
817                 return NULL;
818         }
819
820         /* !HAVE_PTE_SPECIAL case follows: */
821
822         if (unlikely(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP))) {
823                 if (vma->vm_flags & VM_MIXEDMAP) {
824                         if (!pfn_valid(pfn))
825                                 return NULL;
826                         goto out;
827                 } else {
828                         unsigned long off;
829                         off = (addr - vma->vm_start) >> PAGE_SHIFT;
830                         if (pfn == vma->vm_pgoff + off)
831                                 return NULL;
832                         if (!is_cow_mapping(vma->vm_flags))
833                                 return NULL;
834                 }
835         }
836
837         if (is_zero_pfn(pfn))
838                 return NULL;
839 check_pfn:
840         if (unlikely(pfn > highest_memmap_pfn)) {
841                 print_bad_pte(vma, addr, pte, NULL);
842                 return NULL;
843         }
844
845         /*
846          * NOTE! We still have PageReserved() pages in the page tables.
847          * eg. VDSO mappings can cause them to exist.
848          */
849 out:
850         return pfn_to_page(pfn);
851 }
852
853 /*
854  * copy one vm_area from one task to the other. Assumes the page tables
855  * already present in the new task to be cleared in the whole range
856  * covered by this vma.
857  */
858
859 static inline unsigned long
860 copy_one_pte(struct mm_struct *dst_mm, struct mm_struct *src_mm,
861                 pte_t *dst_pte, pte_t *src_pte, struct vm_area_struct *vma,
862                 unsigned long addr, int *rss)
863 {
864         unsigned long vm_flags = vma->vm_flags;
865         pte_t pte = *src_pte;
866         struct page *page;
867
868         /* pte contains position in swap or file, so copy. */
869         if (unlikely(!pte_present(pte))) {
870                 if (!pte_file(pte)) {
871                         swp_entry_t entry = pte_to_swp_entry(pte);
872
873                         if (swap_duplicate(entry) < 0)
874                                 return entry.val;
875
876                         /* make sure dst_mm is on swapoff's mmlist. */
877                         if (unlikely(list_empty(&dst_mm->mmlist))) {
878                                 spin_lock(&mmlist_lock);
879                                 if (list_empty(&dst_mm->mmlist))
880                                         list_add(&dst_mm->mmlist,
881                                                  &src_mm->mmlist);
882                                 spin_unlock(&mmlist_lock);
883                         }
884                         if (likely(!non_swap_entry(entry)))
885                                 rss[MM_SWAPENTS]++;
886                         else if (is_write_migration_entry(entry) &&
887                                         is_cow_mapping(vm_flags)) {
888                                 /*
889                                  * COW mappings require pages in both parent
890                                  * and child to be set to read.
891                                  */
892                                 make_migration_entry_read(&entry);
893                                 pte = swp_entry_to_pte(entry);
894                                 set_pte_at(src_mm, addr, src_pte, pte);
895                         }
896                 }
897                 goto out_set_pte;
898         }
899
900         /*
901          * If it's a COW mapping, write protect it both
902          * in the parent and the child
903          */
904         if (is_cow_mapping(vm_flags)) {
905                 ptep_set_wrprotect(src_mm, addr, src_pte);
906                 pte = pte_wrprotect(pte);
907         }
908
909         /*
910          * If it's a shared mapping, mark it clean in
911          * the child
912          */
913         if (vm_flags & VM_SHARED)
914                 pte = pte_mkclean(pte);
915         pte = pte_mkold(pte);
916
917         page = vm_normal_page(vma, addr, pte);
918         if (page) {
919                 get_page(page);
920                 page_dup_rmap(page);
921                 if (PageAnon(page))
922                         rss[MM_ANONPAGES]++;
923                 else
924                         rss[MM_FILEPAGES]++;
925         }
926
927 out_set_pte:
928         set_pte_at(dst_mm, addr, dst_pte, pte);
929         return 0;
930 }
931
932 int copy_pte_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
933                    pmd_t *dst_pmd, pmd_t *src_pmd, struct vm_area_struct *vma,
934                    unsigned long addr, unsigned long end)
935 {
936         pte_t *orig_src_pte, *orig_dst_pte;
937         pte_t *src_pte, *dst_pte;
938         spinlock_t *src_ptl, *dst_ptl;
939         int progress = 0;
940         int rss[NR_MM_COUNTERS];
941         swp_entry_t entry = (swp_entry_t){0};
942
943 again:
944         init_rss_vec(rss);
945
946         dst_pte = pte_alloc_map_lock(dst_mm, dst_pmd, addr, &dst_ptl);
947         if (!dst_pte)
948                 return -ENOMEM;
949         src_pte = pte_offset_map(src_pmd, addr);
950         src_ptl = pte_lockptr(src_mm, src_pmd);
951         spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
952         orig_src_pte = src_pte;
953         orig_dst_pte = dst_pte;
954         arch_enter_lazy_mmu_mode();
955
956         do {
957                 /*
958                  * We are holding two locks at this point - either of them
959                  * could generate latencies in another task on another CPU.
960                  */
961                 if (progress >= 32) {
962                         progress = 0;
963                         if (need_resched() ||
964                             spin_needbreak(src_ptl) || spin_needbreak(dst_ptl))
965                                 break;
966                 }
967                 if (pte_none(*src_pte)) {
968                         progress++;
969                         continue;
970                 }
971                 entry.val = copy_one_pte(dst_mm, src_mm, dst_pte, src_pte,
972                                                         vma, addr, rss);
973                 if (entry.val)
974                         break;
975                 progress += 8;
976         } while (dst_pte++, src_pte++, addr += PAGE_SIZE, addr != end);
977
978         arch_leave_lazy_mmu_mode();
979         spin_unlock(src_ptl);
980         pte_unmap(orig_src_pte);
981         add_mm_rss_vec(dst_mm, rss);
982         pte_unmap_unlock(orig_dst_pte, dst_ptl);
983         cond_resched();
984
985         if (entry.val) {
986                 if (add_swap_count_continuation(entry, GFP_KERNEL) < 0)
987                         return -ENOMEM;
988                 progress = 0;
989         }
990         if (addr != end)
991                 goto again;
992         return 0;
993 }
994
995 static inline int copy_pmd_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
996                 pud_t *dst_pud, pud_t *src_pud, struct vm_area_struct *vma,
997                 unsigned long addr, unsigned long end)
998 {
999         pmd_t *src_pmd, *dst_pmd;
1000         unsigned long next;
1001
1002         dst_pmd = pmd_alloc(dst_mm, dst_pud, addr);
1003         if (!dst_pmd)
1004                 return -ENOMEM;
1005         src_pmd = pmd_offset(src_pud, addr);
1006         do {
1007                 next = pmd_addr_end(addr, end);
1008                 if (pmd_trans_huge(*src_pmd)) {
1009                         int err;
1010                         VM_BUG_ON(next-addr != HPAGE_PMD_SIZE);
1011                         err = copy_huge_pmd(dst_mm, src_mm,
1012                                             dst_pmd, src_pmd, addr, vma);
1013                         if (err == -ENOMEM)
1014                                 return -ENOMEM;
1015                         if (!err)
1016                                 continue;
1017                         /* fall through */
1018                 }
1019                 if (pmd_none_or_clear_bad(src_pmd))
1020                         continue;
1021                 if (copy_pte_range(dst_mm, src_mm, dst_pmd, src_pmd,
1022                                                 vma, addr, next))
1023                         return -ENOMEM;
1024         } while (dst_pmd++, src_pmd++, addr = next, addr != end);
1025         return 0;
1026 }
1027
1028 static inline int copy_pud_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
1029                 pgd_t *dst_pgd, pgd_t *src_pgd, struct vm_area_struct *vma,
1030                 unsigned long addr, unsigned long end)
1031 {
1032         pud_t *src_pud, *dst_pud;
1033         unsigned long next;
1034
1035         dst_pud = pud_alloc(dst_mm, dst_pgd, addr);
1036         if (!dst_pud)
1037                 return -ENOMEM;
1038         src_pud = pud_offset(src_pgd, addr);
1039         do {
1040                 next = pud_addr_end(addr, end);
1041                 if (pud_none_or_clear_bad(src_pud))
1042                         continue;
1043                 if (copy_pmd_range(dst_mm, src_mm, dst_pud, src_pud,
1044                                                 vma, addr, next))
1045                         return -ENOMEM;
1046         } while (dst_pud++, src_pud++, addr = next, addr != end);
1047         return 0;
1048 }
1049
1050 int copy_page_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
1051                 struct vm_area_struct *vma)
1052 {
1053         pgd_t *src_pgd, *dst_pgd;
1054         unsigned long next;
1055         unsigned long addr = vma->vm_start;
1056         unsigned long end = vma->vm_end;
1057         int ret;
1058
1059         /*
1060          * Don't copy ptes where a page fault will fill them correctly.
1061          * Fork becomes much lighter when there are big shared or private
1062          * readonly mappings. The tradeoff is that copy_page_range is more
1063          * efficient than faulting.
1064          */
1065         if (!(vma->vm_flags & (VM_HUGETLB|VM_NONLINEAR|VM_PFNMAP|VM_INSERTPAGE))) {
1066                 if (!vma->anon_vma)
1067                         return 0;
1068         }
1069
1070         if (is_vm_hugetlb_page(vma))
1071                 return copy_hugetlb_page_range(dst_mm, src_mm, vma);
1072
1073         if (unlikely(is_pfn_mapping(vma))) {
1074                 /*
1075                  * We do not free on error cases below as remove_vma
1076                  * gets called on error from higher level routine
1077                  */
1078                 ret = track_pfn_vma_copy(vma);
1079                 if (ret)
1080                         return ret;
1081         }
1082
1083         /*
1084          * We need to invalidate the secondary MMU mappings only when
1085          * there could be a permission downgrade on the ptes of the
1086          * parent mm. And a permission downgrade will only happen if
1087          * is_cow_mapping() returns true.
1088          */
1089         if (is_cow_mapping(vma->vm_flags))
1090                 mmu_notifier_invalidate_range_start(src_mm, addr, end);
1091
1092         ret = 0;
1093         dst_pgd = pgd_offset(dst_mm, addr);
1094         src_pgd = pgd_offset(src_mm, addr);
1095         do {
1096                 next = pgd_addr_end(addr, end);
1097                 if (pgd_none_or_clear_bad(src_pgd))
1098                         continue;
1099                 if (unlikely(copy_pud_range(dst_mm, src_mm, dst_pgd, src_pgd,
1100                                             vma, addr, next))) {
1101                         ret = -ENOMEM;
1102                         break;
1103                 }
1104         } while (dst_pgd++, src_pgd++, addr = next, addr != end);
1105
1106         if (is_cow_mapping(vma->vm_flags))
1107                 mmu_notifier_invalidate_range_end(src_mm,
1108                                                   vma->vm_start, end);
1109         return ret;
1110 }
1111
1112 static unsigned long zap_pte_range(struct mmu_gather *tlb,
1113                                 struct vm_area_struct *vma, pmd_t *pmd,
1114                                 unsigned long addr, unsigned long end,
1115                                 struct zap_details *details)
1116 {
1117         struct mm_struct *mm = tlb->mm;
1118         int force_flush = 0;
1119         int rss[NR_MM_COUNTERS];
1120         spinlock_t *ptl;
1121         pte_t *start_pte;
1122         pte_t *pte;
1123
1124 again:
1125         init_rss_vec(rss);
1126         start_pte = pte_offset_map_lock(mm, pmd, addr, &ptl);
1127         pte = start_pte;
1128         arch_enter_lazy_mmu_mode();
1129         do {
1130                 pte_t ptent = *pte;
1131                 if (pte_none(ptent)) {
1132                         continue;
1133                 }
1134
1135                 if (pte_present(ptent)) {
1136                         struct page *page;
1137
1138                         page = vm_normal_page(vma, addr, ptent);
1139                         if (unlikely(details) && page) {
1140                                 /*
1141                                  * unmap_shared_mapping_pages() wants to
1142                                  * invalidate cache without truncating:
1143                                  * unmap shared but keep private pages.
1144                                  */
1145                                 if (details->check_mapping &&
1146                                     details->check_mapping != page->mapping)
1147                                         continue;
1148                                 /*
1149                                  * Each page->index must be checked when
1150                                  * invalidating or truncating nonlinear.
1151                                  */
1152                                 if (details->nonlinear_vma &&
1153                                     (page->index < details->first_index ||
1154                                      page->index > details->last_index))
1155                                         continue;
1156                         }
1157                         ptent = ptep_get_and_clear_full(mm, addr, pte,
1158                                                         tlb->fullmm);
1159                         tlb_remove_tlb_entry(tlb, pte, addr);
1160                         if (unlikely(!page))
1161                                 continue;
1162                         if (unlikely(details) && details->nonlinear_vma
1163                             && linear_page_index(details->nonlinear_vma,
1164                                                 addr) != page->index)
1165                                 set_pte_at(mm, addr, pte,
1166                                            pgoff_to_pte(page->index));
1167                         if (PageAnon(page))
1168                                 rss[MM_ANONPAGES]--;
1169                         else {
1170                                 if (pte_dirty(ptent))
1171                                         set_page_dirty(page);
1172                                 if (pte_young(ptent) &&
1173                                     likely(!VM_SequentialReadHint(vma)))
1174                                         mark_page_accessed(page);
1175                                 rss[MM_FILEPAGES]--;
1176                         }
1177                         page_remove_rmap(page);
1178                         if (unlikely(page_mapcount(page) < 0))
1179                                 print_bad_pte(vma, addr, ptent, page);
1180                         force_flush = !__tlb_remove_page(tlb, page);
1181                         if (force_flush)
1182                                 break;
1183                         continue;
1184                 }
1185                 /*
1186                  * If details->check_mapping, we leave swap entries;
1187                  * if details->nonlinear_vma, we leave file entries.
1188                  */
1189                 if (unlikely(details))
1190                         continue;
1191                 if (pte_file(ptent)) {
1192                         if (unlikely(!(vma->vm_flags & VM_NONLINEAR)))
1193                                 print_bad_pte(vma, addr, ptent, NULL);
1194                 } else {
1195                         swp_entry_t entry = pte_to_swp_entry(ptent);
1196
1197                         if (!non_swap_entry(entry))
1198                                 rss[MM_SWAPENTS]--;
1199                         if (unlikely(!free_swap_and_cache(entry)))
1200                                 print_bad_pte(vma, addr, ptent, NULL);
1201                 }
1202                 pte_clear_not_present_full(mm, addr, pte, tlb->fullmm);
1203         } while (pte++, addr += PAGE_SIZE, addr != end);
1204
1205         add_mm_rss_vec(mm, rss);
1206         arch_leave_lazy_mmu_mode();
1207         pte_unmap_unlock(start_pte, ptl);
1208
1209         /*
1210          * mmu_gather ran out of room to batch pages, we break out of
1211          * the PTE lock to avoid doing the potential expensive TLB invalidate
1212          * and page-free while holding it.
1213          */
1214         if (force_flush) {
1215                 force_flush = 0;
1216                 tlb_flush_mmu(tlb);
1217                 if (addr != end)
1218                         goto again;
1219         }
1220
1221         return addr;
1222 }
1223
1224 static inline unsigned long zap_pmd_range(struct mmu_gather *tlb,
1225                                 struct vm_area_struct *vma, pud_t *pud,
1226                                 unsigned long addr, unsigned long end,
1227                                 struct zap_details *details)
1228 {
1229         pmd_t *pmd;
1230         unsigned long next;
1231
1232         pmd = pmd_offset(pud, addr);
1233         do {
1234                 next = pmd_addr_end(addr, end);
1235                 if (pmd_trans_huge(*pmd)) {
1236                         if (next - addr != HPAGE_PMD_SIZE) {
1237                                 VM_BUG_ON(!rwsem_is_locked(&tlb->mm->mmap_sem));
1238                                 split_huge_page_pmd(vma->vm_mm, pmd);
1239                         } else if (zap_huge_pmd(tlb, vma, pmd))
1240                                 goto next;
1241                         /* fall through */
1242                 }
1243                 /*
1244                  * Here there can be other concurrent MADV_DONTNEED or
1245                  * trans huge page faults running, and if the pmd is
1246                  * none or trans huge it can change under us. This is
1247                  * because MADV_DONTNEED holds the mmap_sem in read
1248                  * mode.
1249                  */
1250                 if (pmd_none_or_trans_huge_or_clear_bad(pmd))
1251                         goto next;
1252                 next = zap_pte_range(tlb, vma, pmd, addr, next, details);
1253 next:
1254                 cond_resched();
1255         } while (pmd++, addr = next, addr != end);
1256
1257         return addr;
1258 }
1259
1260 static inline unsigned long zap_pud_range(struct mmu_gather *tlb,
1261                                 struct vm_area_struct *vma, pgd_t *pgd,
1262                                 unsigned long addr, unsigned long end,
1263                                 struct zap_details *details)
1264 {
1265         pud_t *pud;
1266         unsigned long next;
1267
1268         pud = pud_offset(pgd, addr);
1269         do {
1270                 next = pud_addr_end(addr, end);
1271                 if (pud_none_or_clear_bad(pud))
1272                         continue;
1273                 next = zap_pmd_range(tlb, vma, pud, addr, next, details);
1274         } while (pud++, addr = next, addr != end);
1275
1276         return addr;
1277 }
1278
1279 static unsigned long unmap_page_range(struct mmu_gather *tlb,
1280                                 struct vm_area_struct *vma,
1281                                 unsigned long addr, unsigned long end,
1282                                 struct zap_details *details)
1283 {
1284         pgd_t *pgd;
1285         unsigned long next;
1286
1287         if (details && !details->check_mapping && !details->nonlinear_vma)
1288                 details = NULL;
1289
1290         BUG_ON(addr >= end);
1291         mem_cgroup_uncharge_start();
1292         tlb_start_vma(tlb, vma);
1293         pgd = pgd_offset(vma->vm_mm, addr);
1294         do {
1295                 next = pgd_addr_end(addr, end);
1296                 if (pgd_none_or_clear_bad(pgd))
1297                         continue;
1298                 next = zap_pud_range(tlb, vma, pgd, addr, next, details);
1299         } while (pgd++, addr = next, addr != end);
1300         tlb_end_vma(tlb, vma);
1301         mem_cgroup_uncharge_end();
1302
1303         return addr;
1304 }
1305
1306 /**
1307  * unmap_vmas - unmap a range of memory covered by a list of vma's
1308  * @tlb: address of the caller's struct mmu_gather
1309  * @vma: the starting vma
1310  * @start_addr: virtual address at which to start unmapping
1311  * @end_addr: virtual address at which to end unmapping
1312  * @nr_accounted: Place number of unmapped pages in vm-accountable vma's here
1313  * @details: details of nonlinear truncation or shared cache invalidation
1314  *
1315  * Returns the end address of the unmapping (restart addr if interrupted).
1316  *
1317  * Unmap all pages in the vma list.
1318  *
1319  * Only addresses between `start' and `end' will be unmapped.
1320  *
1321  * The VMA list must be sorted in ascending virtual address order.
1322  *
1323  * unmap_vmas() assumes that the caller will flush the whole unmapped address
1324  * range after unmap_vmas() returns.  So the only responsibility here is to
1325  * ensure that any thus-far unmapped pages are flushed before unmap_vmas()
1326  * drops the lock and schedules.
1327  */
1328 unsigned long unmap_vmas(struct mmu_gather *tlb,
1329                 struct vm_area_struct *vma, unsigned long start_addr,
1330                 unsigned long end_addr, unsigned long *nr_accounted,
1331                 struct zap_details *details)
1332 {
1333         unsigned long start = start_addr;
1334         struct mm_struct *mm = vma->vm_mm;
1335
1336         mmu_notifier_invalidate_range_start(mm, start_addr, end_addr);
1337         for ( ; vma && vma->vm_start < end_addr; vma = vma->vm_next) {
1338                 unsigned long end;
1339
1340                 start = max(vma->vm_start, start_addr);
1341                 if (start >= vma->vm_end)
1342                         continue;
1343                 end = min(vma->vm_end, end_addr);
1344                 if (end <= vma->vm_start)
1345                         continue;
1346
1347                 if (vma->vm_flags & VM_ACCOUNT)
1348                         *nr_accounted += (end - start) >> PAGE_SHIFT;
1349
1350                 if (unlikely(is_pfn_mapping(vma)))
1351                         untrack_pfn_vma(vma, 0, 0);
1352
1353                 while (start != end) {
1354                         if (unlikely(is_vm_hugetlb_page(vma))) {
1355                                 /*
1356                                  * It is undesirable to test vma->vm_file as it
1357                                  * should be non-null for valid hugetlb area.
1358                                  * However, vm_file will be NULL in the error
1359                                  * cleanup path of do_mmap_pgoff. When
1360                                  * hugetlbfs ->mmap method fails,
1361                                  * do_mmap_pgoff() nullifies vma->vm_file
1362                                  * before calling this function to clean up.
1363                                  * Since no pte has actually been setup, it is
1364                                  * safe to do nothing in this case.
1365                                  */
1366                                 if (vma->vm_file) {
1367                                         mutex_lock(&vma->vm_file->f_mapping->i_mmap_mutex);
1368                                         __unmap_hugepage_range_final(vma, start, end, NULL);
1369                                         mutex_unlock(&vma->vm_file->f_mapping->i_mmap_mutex);
1370                                 }
1371
1372                                 start = end;
1373                         } else
1374                                 start = unmap_page_range(tlb, vma, start, end, details);
1375                 }
1376         }
1377
1378         mmu_notifier_invalidate_range_end(mm, start_addr, end_addr);
1379         return start;   /* which is now the end (or restart) address */
1380 }
1381
1382 /**
1383  * zap_page_range - remove user pages in a given range
1384  * @vma: vm_area_struct holding the applicable pages
1385  * @address: starting address of pages to zap
1386  * @size: number of bytes to zap
1387  * @details: details of nonlinear truncation or shared cache invalidation
1388  */
1389 unsigned long zap_page_range(struct vm_area_struct *vma, unsigned long address,
1390                 unsigned long size, struct zap_details *details)
1391 {
1392         struct mm_struct *mm = vma->vm_mm;
1393         struct mmu_gather tlb;
1394         unsigned long end = address + size;
1395         unsigned long nr_accounted = 0;
1396
1397         lru_add_drain();
1398         tlb_gather_mmu(&tlb, mm, 0);
1399         update_hiwater_rss(mm);
1400         end = unmap_vmas(&tlb, vma, address, end, &nr_accounted, details);
1401         tlb_finish_mmu(&tlb, address, end);
1402         return end;
1403 }
1404 EXPORT_SYMBOL_GPL(zap_page_range);
1405
1406 /**
1407  * zap_vma_ptes - remove ptes mapping the vma
1408  * @vma: vm_area_struct holding ptes to be zapped
1409  * @address: starting address of pages to zap
1410  * @size: number of bytes to zap
1411  *
1412  * This function only unmaps ptes assigned to VM_PFNMAP vmas.
1413  *
1414  * The entire address range must be fully contained within the vma.
1415  *
1416  * Returns 0 if successful.
1417  */
1418 int zap_vma_ptes(struct vm_area_struct *vma, unsigned long address,
1419                 unsigned long size)
1420 {
1421         if (address < vma->vm_start || address + size > vma->vm_end ||
1422                         !(vma->vm_flags & VM_PFNMAP))
1423                 return -1;
1424         zap_page_range(vma, address, size, NULL);
1425         return 0;
1426 }
1427 EXPORT_SYMBOL_GPL(zap_vma_ptes);
1428
1429 /**
1430  * follow_page - look up a page descriptor from a user-virtual address
1431  * @vma: vm_area_struct mapping @address
1432  * @address: virtual address to look up
1433  * @flags: flags modifying lookup behaviour
1434  *
1435  * @flags can have FOLL_ flags set, defined in <linux/mm.h>
1436  *
1437  * Returns the mapped (struct page *), %NULL if no mapping exists, or
1438  * an error pointer if there is a mapping to something not represented
1439  * by a page descriptor (see also vm_normal_page()).
1440  */
1441 struct page *follow_page(struct vm_area_struct *vma, unsigned long address,
1442                         unsigned int flags)
1443 {
1444         pgd_t *pgd;
1445         pud_t *pud;
1446         pmd_t *pmd;
1447         pte_t *ptep, pte;
1448         spinlock_t *ptl;
1449         struct page *page;
1450         struct mm_struct *mm = vma->vm_mm;
1451
1452         page = follow_huge_addr(mm, address, flags & FOLL_WRITE);
1453         if (!IS_ERR(page)) {
1454                 BUG_ON(flags & FOLL_GET);
1455                 goto out;
1456         }
1457
1458         page = NULL;
1459         pgd = pgd_offset(mm, address);
1460         if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
1461                 goto no_page_table;
1462
1463         pud = pud_offset(pgd, address);
1464         if (pud_none(*pud))
1465                 goto no_page_table;
1466         if (pud_huge(*pud) && vma->vm_flags & VM_HUGETLB) {
1467                 BUG_ON(flags & FOLL_GET);
1468                 page = follow_huge_pud(mm, address, pud, flags & FOLL_WRITE);
1469                 goto out;
1470         }
1471         if (unlikely(pud_bad(*pud)))
1472                 goto no_page_table;
1473
1474         pmd = pmd_offset(pud, address);
1475         if (pmd_none(*pmd))
1476                 goto no_page_table;
1477         if (pmd_huge(*pmd) && vma->vm_flags & VM_HUGETLB) {
1478                 BUG_ON(flags & FOLL_GET);
1479                 page = follow_huge_pmd(mm, address, pmd, flags & FOLL_WRITE);
1480                 goto out;
1481         }
1482         if (pmd_trans_huge(*pmd)) {
1483                 if (flags & FOLL_SPLIT) {
1484                         split_huge_page_pmd(mm, pmd);
1485                         goto split_fallthrough;
1486                 }
1487                 spin_lock(&mm->page_table_lock);
1488                 if (likely(pmd_trans_huge(*pmd))) {
1489                         if (unlikely(pmd_trans_splitting(*pmd))) {
1490                                 spin_unlock(&mm->page_table_lock);
1491                                 wait_split_huge_page(vma->anon_vma, pmd);
1492                         } else {
1493                                 page = follow_trans_huge_pmd(mm, address,
1494                                                              pmd, flags);
1495                                 spin_unlock(&mm->page_table_lock);
1496                                 goto out;
1497                         }
1498                 } else
1499                         spin_unlock(&mm->page_table_lock);
1500                 /* fall through */
1501         }
1502 split_fallthrough:
1503         if (unlikely(pmd_bad(*pmd)))
1504                 goto no_page_table;
1505
1506         ptep = pte_offset_map_lock(mm, pmd, address, &ptl);
1507
1508         pte = *ptep;
1509         if (!pte_present(pte))
1510                 goto no_page;
1511         if ((flags & FOLL_WRITE) && !pte_write(pte))
1512                 goto unlock;
1513
1514         page = vm_normal_page(vma, address, pte);
1515         if (unlikely(!page)) {
1516                 if ((flags & FOLL_DUMP) ||
1517                     !is_zero_pfn(pte_pfn(pte)))
1518                         goto bad_page;
1519                 page = pte_page(pte);
1520         }
1521
1522         if (flags & FOLL_GET)
1523                 get_page_foll(page);
1524         if (flags & FOLL_TOUCH) {
1525                 if ((flags & FOLL_WRITE) &&
1526                     !pte_dirty(pte) && !PageDirty(page))
1527                         set_page_dirty(page);
1528                 /*
1529                  * pte_mkyoung() would be more correct here, but atomic care
1530                  * is needed to avoid losing the dirty bit: it is easier to use
1531                  * mark_page_accessed().
1532                  */
1533                 mark_page_accessed(page);
1534         }
1535         if ((flags & FOLL_MLOCK) && (vma->vm_flags & VM_LOCKED)) {
1536                 /*
1537                  * The preliminary mapping check is mainly to avoid the
1538                  * pointless overhead of lock_page on the ZERO_PAGE
1539                  * which might bounce very badly if there is contention.
1540                  *
1541                  * If the page is already locked, we don't need to
1542                  * handle it now - vmscan will handle it later if and
1543                  * when it attempts to reclaim the page.
1544                  */
1545                 if (page->mapping && trylock_page(page)) {
1546                         lru_add_drain();  /* push cached pages to LRU */
1547                         /*
1548                          * Because we lock page here and migration is
1549                          * blocked by the pte's page reference, we need
1550                          * only check for file-cache page truncation.
1551                          */
1552                         if (page->mapping)
1553                                 mlock_vma_page(page);
1554                         unlock_page(page);
1555                 }
1556         }
1557 unlock:
1558         pte_unmap_unlock(ptep, ptl);
1559 out:
1560         return page;
1561
1562 bad_page:
1563         pte_unmap_unlock(ptep, ptl);
1564         return ERR_PTR(-EFAULT);
1565
1566 no_page:
1567         pte_unmap_unlock(ptep, ptl);
1568         if (!pte_none(pte))
1569                 return page;
1570
1571 no_page_table:
1572         /*
1573          * When core dumping an enormous anonymous area that nobody
1574          * has touched so far, we don't want to allocate unnecessary pages or
1575          * page tables.  Return error instead of NULL to skip handle_mm_fault,
1576          * then get_dump_page() will return NULL to leave a hole in the dump.
1577          * But we can only make this optimization where a hole would surely
1578          * be zero-filled if handle_mm_fault() actually did handle it.
1579          */
1580         if ((flags & FOLL_DUMP) &&
1581             (!vma->vm_ops || !vma->vm_ops->fault))
1582                 return ERR_PTR(-EFAULT);
1583         return page;
1584 }
1585
1586 static inline int stack_guard_page(struct vm_area_struct *vma, unsigned long addr)
1587 {
1588         return stack_guard_page_start(vma, addr) ||
1589                stack_guard_page_end(vma, addr+PAGE_SIZE);
1590 }
1591
1592 /**
1593  * __get_user_pages() - pin user pages in memory
1594  * @tsk:        task_struct of target task
1595  * @mm:         mm_struct of target mm
1596  * @start:      starting user address
1597  * @nr_pages:   number of pages from start to pin
1598  * @gup_flags:  flags modifying pin behaviour
1599  * @pages:      array that receives pointers to the pages pinned.
1600  *              Should be at least nr_pages long. Or NULL, if caller
1601  *              only intends to ensure the pages are faulted in.
1602  * @vmas:       array of pointers to vmas corresponding to each page.
1603  *              Or NULL if the caller does not require them.
1604  * @nonblocking: whether waiting for disk IO or mmap_sem contention
1605  *
1606  * Returns number of pages pinned. This may be fewer than the number
1607  * requested. If nr_pages is 0 or negative, returns 0. If no pages
1608  * were pinned, returns -errno. Each page returned must be released
1609  * with a put_page() call when it is finished with. vmas will only
1610  * remain valid while mmap_sem is held.
1611  *
1612  * Must be called with mmap_sem held for read or write.
1613  *
1614  * __get_user_pages walks a process's page tables and takes a reference to
1615  * each struct page that each user address corresponds to at a given
1616  * instant. That is, it takes the page that would be accessed if a user
1617  * thread accesses the given user virtual address at that instant.
1618  *
1619  * This does not guarantee that the page exists in the user mappings when
1620  * __get_user_pages returns, and there may even be a completely different
1621  * page there in some cases (eg. if mmapped pagecache has been invalidated
1622  * and subsequently re faulted). However it does guarantee that the page
1623  * won't be freed completely. And mostly callers simply care that the page
1624  * contains data that was valid *at some point in time*. Typically, an IO
1625  * or similar operation cannot guarantee anything stronger anyway because
1626  * locks can't be held over the syscall boundary.
1627  *
1628  * If @gup_flags & FOLL_WRITE == 0, the page must not be written to. If
1629  * the page is written to, set_page_dirty (or set_page_dirty_lock, as
1630  * appropriate) must be called after the page is finished with, and
1631  * before put_page is called.
1632  *
1633  * If @nonblocking != NULL, __get_user_pages will not wait for disk IO
1634  * or mmap_sem contention, and if waiting is needed to pin all pages,
1635  * *@nonblocking will be set to 0.
1636  *
1637  * In most cases, get_user_pages or get_user_pages_fast should be used
1638  * instead of __get_user_pages. __get_user_pages should be used only if
1639  * you need some special @gup_flags.
1640  */
1641 int __get_user_pages(struct task_struct *tsk, struct mm_struct *mm,
1642                      unsigned long start, int nr_pages, unsigned int gup_flags,
1643                      struct page **pages, struct vm_area_struct **vmas,
1644                      int *nonblocking)
1645 {
1646         int i;
1647         unsigned long vm_flags;
1648
1649         if (nr_pages <= 0)
1650                 return 0;
1651
1652         VM_BUG_ON(!!pages != !!(gup_flags & FOLL_GET));
1653
1654         /* 
1655          * Require read or write permissions.
1656          * If FOLL_FORCE is set, we only require the "MAY" flags.
1657          */
1658         vm_flags  = (gup_flags & FOLL_WRITE) ?
1659                         (VM_WRITE | VM_MAYWRITE) : (VM_READ | VM_MAYREAD);
1660         vm_flags &= (gup_flags & FOLL_FORCE) ?
1661                         (VM_MAYREAD | VM_MAYWRITE) : (VM_READ | VM_WRITE);
1662         i = 0;
1663
1664         do {
1665                 struct vm_area_struct *vma;
1666
1667                 vma = find_extend_vma(mm, start);
1668                 if (!vma && in_gate_area(mm, start)) {
1669                         unsigned long pg = start & PAGE_MASK;
1670                         pgd_t *pgd;
1671                         pud_t *pud;
1672                         pmd_t *pmd;
1673                         pte_t *pte;
1674
1675                         /* user gate pages are read-only */
1676                         if (gup_flags & FOLL_WRITE)
1677                                 return i ? : -EFAULT;
1678                         if (pg > TASK_SIZE)
1679                                 pgd = pgd_offset_k(pg);
1680                         else
1681                                 pgd = pgd_offset_gate(mm, pg);
1682                         BUG_ON(pgd_none(*pgd));
1683                         pud = pud_offset(pgd, pg);
1684                         BUG_ON(pud_none(*pud));
1685                         pmd = pmd_offset(pud, pg);
1686                         if (pmd_none(*pmd))
1687                                 return i ? : -EFAULT;
1688                         VM_BUG_ON(pmd_trans_huge(*pmd));
1689                         pte = pte_offset_map(pmd, pg);
1690                         if (pte_none(*pte)) {
1691                                 pte_unmap(pte);
1692                                 return i ? : -EFAULT;
1693                         }
1694                         vma = get_gate_vma(mm);
1695                         if (pages) {
1696                                 struct page *page;
1697
1698                                 page = vm_normal_page(vma, start, *pte);
1699                                 if (!page) {
1700                                         if (!(gup_flags & FOLL_DUMP) &&
1701                                              is_zero_pfn(pte_pfn(*pte)))
1702                                                 page = pte_page(*pte);
1703                                         else {
1704                                                 pte_unmap(pte);
1705                                                 return i ? : -EFAULT;
1706                                         }
1707                                 }
1708                                 pages[i] = page;
1709                                 get_page(page);
1710                         }
1711                         pte_unmap(pte);
1712                         goto next_page;
1713                 }
1714
1715                 if (!vma ||
1716                     (vma->vm_flags & (VM_IO | VM_PFNMAP)) ||
1717                     !(vm_flags & vma->vm_flags))
1718                         return i ? : -EFAULT;
1719
1720                 if (is_vm_hugetlb_page(vma)) {
1721                         i = follow_hugetlb_page(mm, vma, pages, vmas,
1722                                         &start, &nr_pages, i, gup_flags);
1723                         continue;
1724                 }
1725
1726                 do {
1727                         struct page *page;
1728                         unsigned int foll_flags = gup_flags;
1729
1730                         /*
1731                          * If we have a pending SIGKILL, don't keep faulting
1732                          * pages and potentially allocating memory.
1733                          */
1734                         if (unlikely(fatal_signal_pending(current)))
1735                                 return i ? i : -ERESTARTSYS;
1736
1737                         cond_resched();
1738                         while (!(page = follow_page(vma, start, foll_flags))) {
1739                                 int ret;
1740                                 unsigned int fault_flags = 0;
1741
1742                                 /* For mlock, just skip the stack guard page. */
1743                                 if (foll_flags & FOLL_MLOCK) {
1744                                         if (stack_guard_page(vma, start))
1745                                                 goto next_page;
1746                                 }
1747                                 if (foll_flags & FOLL_WRITE)
1748                                         fault_flags |= FAULT_FLAG_WRITE;
1749                                 if (nonblocking)
1750                                         fault_flags |= FAULT_FLAG_ALLOW_RETRY;
1751                                 if (foll_flags & FOLL_NOWAIT)
1752                                         fault_flags |= (FAULT_FLAG_ALLOW_RETRY | FAULT_FLAG_RETRY_NOWAIT);
1753
1754                                 ret = handle_mm_fault(mm, vma, start,
1755                                                         fault_flags);
1756
1757                                 if (ret & VM_FAULT_ERROR) {
1758                                         if (ret & VM_FAULT_OOM)
1759                                                 return i ? i : -ENOMEM;
1760                                         if (ret & (VM_FAULT_HWPOISON |
1761                                                    VM_FAULT_HWPOISON_LARGE)) {
1762                                                 if (i)
1763                                                         return i;
1764                                                 else if (gup_flags & FOLL_HWPOISON)
1765                                                         return -EHWPOISON;
1766                                                 else
1767                                                         return -EFAULT;
1768                                         }
1769                                         if (ret & VM_FAULT_SIGBUS)
1770                                                 return i ? i : -EFAULT;
1771                                         BUG();
1772                                 }
1773
1774                                 if (tsk) {
1775                                         if (ret & VM_FAULT_MAJOR)
1776                                                 tsk->maj_flt++;
1777                                         else
1778                                                 tsk->min_flt++;
1779                                 }
1780
1781                                 if (ret & VM_FAULT_RETRY) {
1782                                         if (nonblocking)
1783                                                 *nonblocking = 0;
1784                                         return i;
1785                                 }
1786
1787                                 /*
1788                                  * The VM_FAULT_WRITE bit tells us that
1789                                  * do_wp_page has broken COW when necessary,
1790                                  * even if maybe_mkwrite decided not to set
1791                                  * pte_write. We can thus safely do subsequent
1792                                  * page lookups as if they were reads. But only
1793                                  * do so when looping for pte_write is futile:
1794                                  * in some cases userspace may also be wanting
1795                                  * to write to the gotten user page, which a
1796                                  * read fault here might prevent (a readonly
1797                                  * page might get reCOWed by userspace write).
1798                                  */
1799                                 if ((ret & VM_FAULT_WRITE) &&
1800                                     !(vma->vm_flags & VM_WRITE))
1801                                         foll_flags &= ~FOLL_WRITE;
1802
1803                                 cond_resched();
1804                         }
1805                         if (IS_ERR(page))
1806                                 return i ? i : PTR_ERR(page);
1807                         if (pages) {
1808                                 pages[i] = page;
1809
1810                                 flush_anon_page(vma, page, start);
1811                                 flush_dcache_page(page);
1812                         }
1813 next_page:
1814                         if (vmas)
1815                                 vmas[i] = vma;
1816                         i++;
1817                         start += PAGE_SIZE;
1818                         nr_pages--;
1819                 } while (nr_pages && start < vma->vm_end);
1820         } while (nr_pages);
1821         return i;
1822 }
1823 EXPORT_SYMBOL(__get_user_pages);
1824
1825 /*
1826  * fixup_user_fault() - manually resolve a user page fault
1827  * @tsk:        the task_struct to use for page fault accounting, or
1828  *              NULL if faults are not to be recorded.
1829  * @mm:         mm_struct of target mm
1830  * @address:    user address
1831  * @fault_flags:flags to pass down to handle_mm_fault()
1832  *
1833  * This is meant to be called in the specific scenario where for locking reasons
1834  * we try to access user memory in atomic context (within a pagefault_disable()
1835  * section), this returns -EFAULT, and we want to resolve the user fault before
1836  * trying again.
1837  *
1838  * Typically this is meant to be used by the futex code.
1839  *
1840  * The main difference with get_user_pages() is that this function will
1841  * unconditionally call handle_mm_fault() which will in turn perform all the
1842  * necessary SW fixup of the dirty and young bits in the PTE, while
1843  * handle_mm_fault() only guarantees to update these in the struct page.
1844  *
1845  * This is important for some architectures where those bits also gate the
1846  * access permission to the page because they are maintained in software.  On
1847  * such architectures, gup() will not be enough to make a subsequent access
1848  * succeed.
1849  *
1850  * This should be called with the mm_sem held for read.
1851  */
1852 int fixup_user_fault(struct task_struct *tsk, struct mm_struct *mm,
1853                      unsigned long address, unsigned int fault_flags)
1854 {
1855         struct vm_area_struct *vma;
1856         int ret;
1857
1858         vma = find_extend_vma(mm, address);
1859         if (!vma || address < vma->vm_start)
1860                 return -EFAULT;
1861
1862         ret = handle_mm_fault(mm, vma, address, fault_flags);
1863         if (ret & VM_FAULT_ERROR) {
1864                 if (ret & VM_FAULT_OOM)
1865                         return -ENOMEM;
1866                 if (ret & (VM_FAULT_HWPOISON | VM_FAULT_HWPOISON_LARGE))
1867                         return -EHWPOISON;
1868                 if (ret & VM_FAULT_SIGBUS)
1869                         return -EFAULT;
1870                 BUG();
1871         }
1872         if (tsk) {
1873                 if (ret & VM_FAULT_MAJOR)
1874                         tsk->maj_flt++;
1875                 else
1876                         tsk->min_flt++;
1877         }
1878         return 0;
1879 }
1880
1881 /*
1882  * get_user_pages() - pin user pages in memory
1883  * @tsk:        the task_struct to use for page fault accounting, or
1884  *              NULL if faults are not to be recorded.
1885  * @mm:         mm_struct of target mm
1886  * @start:      starting user address
1887  * @nr_pages:   number of pages from start to pin
1888  * @write:      whether pages will be written to by the caller
1889  * @force:      whether to force write access even if user mapping is
1890  *              readonly. This will result in the page being COWed even
1891  *              in MAP_SHARED mappings. You do not want this.
1892  * @pages:      array that receives pointers to the pages pinned.
1893  *              Should be at least nr_pages long. Or NULL, if caller
1894  *              only intends to ensure the pages are faulted in.
1895  * @vmas:       array of pointers to vmas corresponding to each page.
1896  *              Or NULL if the caller does not require them.
1897  *
1898  * Returns number of pages pinned. This may be fewer than the number
1899  * requested. If nr_pages is 0 or negative, returns 0. If no pages
1900  * were pinned, returns -errno. Each page returned must be released
1901  * with a put_page() call when it is finished with. vmas will only
1902  * remain valid while mmap_sem is held.
1903  *
1904  * Must be called with mmap_sem held for read or write.
1905  *
1906  * get_user_pages walks a process's page tables and takes a reference to
1907  * each struct page that each user address corresponds to at a given
1908  * instant. That is, it takes the page that would be accessed if a user
1909  * thread accesses the given user virtual address at that instant.
1910  *
1911  * This does not guarantee that the page exists in the user mappings when
1912  * get_user_pages returns, and there may even be a completely different
1913  * page there in some cases (eg. if mmapped pagecache has been invalidated
1914  * and subsequently re faulted). However it does guarantee that the page
1915  * won't be freed completely. And mostly callers simply care that the page
1916  * contains data that was valid *at some point in time*. Typically, an IO
1917  * or similar operation cannot guarantee anything stronger anyway because
1918  * locks can't be held over the syscall boundary.
1919  *
1920  * If write=0, the page must not be written to. If the page is written to,
1921  * set_page_dirty (or set_page_dirty_lock, as appropriate) must be called
1922  * after the page is finished with, and before put_page is called.
1923  *
1924  * get_user_pages is typically used for fewer-copy IO operations, to get a
1925  * handle on the memory by some means other than accesses via the user virtual
1926  * addresses. The pages may be submitted for DMA to devices or accessed via
1927  * their kernel linear mapping (via the kmap APIs). Care should be taken to
1928  * use the correct cache flushing APIs.
1929  *
1930  * See also get_user_pages_fast, for performance critical applications.
1931  */
1932 int get_user_pages(struct task_struct *tsk, struct mm_struct *mm,
1933                 unsigned long start, int nr_pages, int write, int force,
1934                 struct page **pages, struct vm_area_struct **vmas)
1935 {
1936         int flags = FOLL_TOUCH;
1937
1938         if (pages)
1939                 flags |= FOLL_GET;
1940         if (write)
1941                 flags |= FOLL_WRITE;
1942         if (force)
1943                 flags |= FOLL_FORCE;
1944
1945         return __get_user_pages(tsk, mm, start, nr_pages, flags, pages, vmas,
1946                                 NULL);
1947 }
1948 EXPORT_SYMBOL(get_user_pages);
1949
1950 /**
1951  * get_dump_page() - pin user page in memory while writing it to core dump
1952  * @addr: user address
1953  *
1954  * Returns struct page pointer of user page pinned for dump,
1955  * to be freed afterwards by page_cache_release() or put_page().
1956  *
1957  * Returns NULL on any kind of failure - a hole must then be inserted into
1958  * the corefile, to preserve alignment with its headers; and also returns
1959  * NULL wherever the ZERO_PAGE, or an anonymous pte_none, has been found -
1960  * allowing a hole to be left in the corefile to save diskspace.
1961  *
1962  * Called without mmap_sem, but after all other threads have been killed.
1963  */
1964 #ifdef CONFIG_ELF_CORE
1965 struct page *get_dump_page(unsigned long addr)
1966 {
1967         struct vm_area_struct *vma;
1968         struct page *page;
1969
1970         if (__get_user_pages(current, current->mm, addr, 1,
1971                              FOLL_FORCE | FOLL_DUMP | FOLL_GET, &page, &vma,
1972                              NULL) < 1)
1973                 return NULL;
1974         flush_cache_page(vma, addr, page_to_pfn(page));
1975         return page;
1976 }
1977 #endif /* CONFIG_ELF_CORE */
1978
1979 pte_t *__get_locked_pte(struct mm_struct *mm, unsigned long addr,
1980                         spinlock_t **ptl)
1981 {
1982         pgd_t * pgd = pgd_offset(mm, addr);
1983         pud_t * pud = pud_alloc(mm, pgd, addr);
1984         if (pud) {
1985                 pmd_t * pmd = pmd_alloc(mm, pud, addr);
1986                 if (pmd) {
1987                         VM_BUG_ON(pmd_trans_huge(*pmd));
1988                         return pte_alloc_map_lock(mm, pmd, addr, ptl);
1989                 }
1990         }
1991         return NULL;
1992 }
1993
1994 /*
1995  * This is the old fallback for page remapping.
1996  *
1997  * For historical reasons, it only allows reserved pages. Only
1998  * old drivers should use this, and they needed to mark their
1999  * pages reserved for the old functions anyway.
2000  */
2001 static int insert_page(struct vm_area_struct *vma, unsigned long addr,
2002                         struct page *page, pgprot_t prot)
2003 {
2004         struct mm_struct *mm = vma->vm_mm;
2005         int retval;
2006         pte_t *pte;
2007         spinlock_t *ptl;
2008
2009         retval = -EINVAL;
2010         if (PageAnon(page))
2011                 goto out;
2012         retval = -ENOMEM;
2013         flush_dcache_page(page);
2014         pte = get_locked_pte(mm, addr, &ptl);
2015         if (!pte)
2016                 goto out;
2017         retval = -EBUSY;
2018         if (!pte_none(*pte))
2019                 goto out_unlock;
2020
2021         /* Ok, finally just insert the thing.. */
2022         get_page(page);
2023         inc_mm_counter_fast(mm, MM_FILEPAGES);
2024         page_add_file_rmap(page);
2025         set_pte_at(mm, addr, pte, mk_pte(page, prot));
2026
2027         retval = 0;
2028         pte_unmap_unlock(pte, ptl);
2029         return retval;
2030 out_unlock:
2031         pte_unmap_unlock(pte, ptl);
2032 out:
2033         return retval;
2034 }
2035
2036 /**
2037  * vm_insert_page - insert single page into user vma
2038  * @vma: user vma to map to
2039  * @addr: target user address of this page
2040  * @page: source kernel page
2041  *
2042  * This allows drivers to insert individual pages they've allocated
2043  * into a user vma.
2044  *
2045  * The page has to be a nice clean _individual_ kernel allocation.
2046  * If you allocate a compound page, you need to have marked it as
2047  * such (__GFP_COMP), or manually just split the page up yourself
2048  * (see split_page()).
2049  *
2050  * NOTE! Traditionally this was done with "remap_pfn_range()" which
2051  * took an arbitrary page protection parameter. This doesn't allow
2052  * that. Your vma protection will have to be set up correctly, which
2053  * means that if you want a shared writable mapping, you'd better
2054  * ask for a shared writable mapping!
2055  *
2056  * The page does not need to be reserved.
2057  */
2058 int vm_insert_page(struct vm_area_struct *vma, unsigned long addr,
2059                         struct page *page)
2060 {
2061         if (addr < vma->vm_start || addr >= vma->vm_end)
2062                 return -EFAULT;
2063         if (!page_count(page))
2064                 return -EINVAL;
2065         vma->vm_flags |= VM_INSERTPAGE;
2066         return insert_page(vma, addr, page, vma->vm_page_prot);
2067 }
2068 EXPORT_SYMBOL(vm_insert_page);
2069
2070 static int insert_pfn(struct vm_area_struct *vma, unsigned long addr,
2071                         unsigned long pfn, pgprot_t prot)
2072 {
2073         struct mm_struct *mm = vma->vm_mm;
2074         int retval;
2075         pte_t *pte, entry;
2076         spinlock_t *ptl;
2077
2078         retval = -ENOMEM;
2079         pte = get_locked_pte(mm, addr, &ptl);
2080         if (!pte)
2081                 goto out;
2082         retval = -EBUSY;
2083         if (!pte_none(*pte))
2084                 goto out_unlock;
2085
2086         /* Ok, finally just insert the thing.. */
2087         entry = pte_mkspecial(pfn_pte(pfn, prot));
2088         set_pte_at(mm, addr, pte, entry);
2089         update_mmu_cache(vma, addr, pte); /* XXX: why not for insert_page? */
2090
2091         retval = 0;
2092 out_unlock:
2093         pte_unmap_unlock(pte, ptl);
2094 out:
2095         return retval;
2096 }
2097
2098 /**
2099  * vm_insert_pfn - insert single pfn into user vma
2100  * @vma: user vma to map to
2101  * @addr: target user address of this page
2102  * @pfn: source kernel pfn
2103  *
2104  * Similar to vm_inert_page, this allows drivers to insert individual pages
2105  * they've allocated into a user vma. Same comments apply.
2106  *
2107  * This function should only be called from a vm_ops->fault handler, and
2108  * in that case the handler should return NULL.
2109  *
2110  * vma cannot be a COW mapping.
2111  *
2112  * As this is called only for pages that do not currently exist, we
2113  * do not need to flush old virtual caches or the TLB.
2114  */
2115 int vm_insert_pfn(struct vm_area_struct *vma, unsigned long addr,
2116                         unsigned long pfn)
2117 {
2118         int ret;
2119         pgprot_t pgprot = vma->vm_page_prot;
2120         /*
2121          * Technically, architectures with pte_special can avoid all these
2122          * restrictions (same for remap_pfn_range).  However we would like
2123          * consistency in testing and feature parity among all, so we should
2124          * try to keep these invariants in place for everybody.
2125          */
2126         BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)));
2127         BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) ==
2128                                                 (VM_PFNMAP|VM_MIXEDMAP));
2129         BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags));
2130         BUG_ON((vma->vm_flags & VM_MIXEDMAP) && pfn_valid(pfn));
2131
2132         if (addr < vma->vm_start || addr >= vma->vm_end)
2133                 return -EFAULT;
2134         if (track_pfn_vma_new(vma, &pgprot, pfn, PAGE_SIZE))
2135                 return -EINVAL;
2136
2137         ret = insert_pfn(vma, addr, pfn, pgprot);
2138
2139         if (ret)
2140                 untrack_pfn_vma(vma, pfn, PAGE_SIZE);
2141
2142         return ret;
2143 }
2144 EXPORT_SYMBOL(vm_insert_pfn);
2145
2146 int vm_insert_mixed(struct vm_area_struct *vma, unsigned long addr,
2147                         unsigned long pfn)
2148 {
2149         BUG_ON(!(vma->vm_flags & VM_MIXEDMAP));
2150
2151         if (addr < vma->vm_start || addr >= vma->vm_end)
2152                 return -EFAULT;
2153
2154         /*
2155          * If we don't have pte special, then we have to use the pfn_valid()
2156          * based VM_MIXEDMAP scheme (see vm_normal_page), and thus we *must*
2157          * refcount the page if pfn_valid is true (hence insert_page rather
2158          * than insert_pfn).  If a zero_pfn were inserted into a VM_MIXEDMAP
2159          * without pte special, it would there be refcounted as a normal page.
2160          */
2161         if (!HAVE_PTE_SPECIAL && pfn_valid(pfn)) {
2162                 struct page *page;
2163
2164                 page = pfn_to_page(pfn);
2165                 return insert_page(vma, addr, page, vma->vm_page_prot);
2166         }
2167         return insert_pfn(vma, addr, pfn, vma->vm_page_prot);
2168 }
2169 EXPORT_SYMBOL(vm_insert_mixed);
2170
2171 /*
2172  * maps a range of physical memory into the requested pages. the old
2173  * mappings are removed. any references to nonexistent pages results
2174  * in null mappings (currently treated as "copy-on-access")
2175  */
2176 static int remap_pte_range(struct mm_struct *mm, pmd_t *pmd,
2177                         unsigned long addr, unsigned long end,
2178                         unsigned long pfn, pgprot_t prot)
2179 {
2180         pte_t *pte;
2181         spinlock_t *ptl;
2182
2183         pte = pte_alloc_map_lock(mm, pmd, addr, &ptl);
2184         if (!pte)
2185                 return -ENOMEM;
2186         arch_enter_lazy_mmu_mode();
2187         do {
2188                 BUG_ON(!pte_none(*pte));
2189                 set_pte_at(mm, addr, pte, pte_mkspecial(pfn_pte(pfn, prot)));
2190                 pfn++;
2191         } while (pte++, addr += PAGE_SIZE, addr != end);
2192         arch_leave_lazy_mmu_mode();
2193         pte_unmap_unlock(pte - 1, ptl);
2194         return 0;
2195 }
2196
2197 static inline int remap_pmd_range(struct mm_struct *mm, pud_t *pud,
2198                         unsigned long addr, unsigned long end,
2199                         unsigned long pfn, pgprot_t prot)
2200 {
2201         pmd_t *pmd;
2202         unsigned long next;
2203
2204         pfn -= addr >> PAGE_SHIFT;
2205         pmd = pmd_alloc(mm, pud, addr);
2206         if (!pmd)
2207                 return -ENOMEM;
2208         VM_BUG_ON(pmd_trans_huge(*pmd));
2209         do {
2210                 next = pmd_addr_end(addr, end);
2211                 if (remap_pte_range(mm, pmd, addr, next,
2212                                 pfn + (addr >> PAGE_SHIFT), prot))
2213                         return -ENOMEM;
2214         } while (pmd++, addr = next, addr != end);
2215         return 0;
2216 }
2217
2218 static inline int remap_pud_range(struct mm_struct *mm, pgd_t *pgd,
2219                         unsigned long addr, unsigned long end,
2220                         unsigned long pfn, pgprot_t prot)
2221 {
2222         pud_t *pud;
2223         unsigned long next;
2224
2225         pfn -= addr >> PAGE_SHIFT;
2226         pud = pud_alloc(mm, pgd, addr);
2227         if (!pud)
2228                 return -ENOMEM;
2229         do {
2230                 next = pud_addr_end(addr, end);
2231                 if (remap_pmd_range(mm, pud, addr, next,
2232                                 pfn + (addr >> PAGE_SHIFT), prot))
2233                         return -ENOMEM;
2234         } while (pud++, addr = next, addr != end);
2235         return 0;
2236 }
2237
2238 /**
2239  * remap_pfn_range - remap kernel memory to userspace
2240  * @vma: user vma to map to
2241  * @addr: target user address to start at
2242  * @pfn: physical address of kernel memory
2243  * @size: size of map area
2244  * @prot: page protection flags for this mapping
2245  *
2246  *  Note: this is only safe if the mm semaphore is held when called.
2247  */
2248 int remap_pfn_range(struct vm_area_struct *vma, unsigned long addr,
2249                     unsigned long pfn, unsigned long size, pgprot_t prot)
2250 {
2251         pgd_t *pgd;
2252         unsigned long next;
2253         unsigned long end = addr + PAGE_ALIGN(size);
2254         struct mm_struct *mm = vma->vm_mm;
2255         int err;
2256
2257         /*
2258          * Physically remapped pages are special. Tell the
2259          * rest of the world about it:
2260          *   VM_IO tells people not to look at these pages
2261          *      (accesses can have side effects).
2262          *   VM_RESERVED is specified all over the place, because
2263          *      in 2.4 it kept swapout's vma scan off this vma; but
2264          *      in 2.6 the LRU scan won't even find its pages, so this
2265          *      flag means no more than count its pages in reserved_vm,
2266          *      and omit it from core dump, even when VM_IO turned off.
2267          *   VM_PFNMAP tells the core MM that the base pages are just
2268          *      raw PFN mappings, and do not have a "struct page" associated
2269          *      with them.
2270          *
2271          * There's a horrible special case to handle copy-on-write
2272          * behaviour that some programs depend on. We mark the "original"
2273          * un-COW'ed pages by matching them up with "vma->vm_pgoff".
2274          */
2275         if (addr == vma->vm_start && end == vma->vm_end) {
2276                 vma->vm_pgoff = pfn;
2277                 vma->vm_flags |= VM_PFN_AT_MMAP;
2278         } else if (is_cow_mapping(vma->vm_flags))
2279                 return -EINVAL;
2280
2281         vma->vm_flags |= VM_IO | VM_RESERVED | VM_PFNMAP;
2282
2283         err = track_pfn_vma_new(vma, &prot, pfn, PAGE_ALIGN(size));
2284         if (err) {
2285                 /*
2286                  * To indicate that track_pfn related cleanup is not
2287                  * needed from higher level routine calling unmap_vmas
2288                  */
2289                 vma->vm_flags &= ~(VM_IO | VM_RESERVED | VM_PFNMAP);
2290                 vma->vm_flags &= ~VM_PFN_AT_MMAP;
2291                 return -EINVAL;
2292         }
2293
2294         BUG_ON(addr >= end);
2295         pfn -= addr >> PAGE_SHIFT;
2296         pgd = pgd_offset(mm, addr);
2297         flush_cache_range(vma, addr, end);
2298         do {
2299                 next = pgd_addr_end(addr, end);
2300                 err = remap_pud_range(mm, pgd, addr, next,
2301                                 pfn + (addr >> PAGE_SHIFT), prot);
2302                 if (err)
2303                         break;
2304         } while (pgd++, addr = next, addr != end);
2305
2306         if (err)
2307                 untrack_pfn_vma(vma, pfn, PAGE_ALIGN(size));
2308
2309         return err;
2310 }
2311 EXPORT_SYMBOL(remap_pfn_range);
2312
2313 static int apply_to_pte_range(struct mm_struct *mm, pmd_t *pmd,
2314                                      unsigned long addr, unsigned long end,
2315                                      pte_fn_t fn, void *data)
2316 {
2317         pte_t *pte;
2318         int err;
2319         pgtable_t token;
2320         spinlock_t *uninitialized_var(ptl);
2321
2322         pte = (mm == &init_mm) ?
2323                 pte_alloc_kernel(pmd, addr) :
2324                 pte_alloc_map_lock(mm, pmd, addr, &ptl);
2325         if (!pte)
2326                 return -ENOMEM;
2327
2328         BUG_ON(pmd_huge(*pmd));
2329
2330         arch_enter_lazy_mmu_mode();
2331
2332         token = pmd_pgtable(*pmd);
2333
2334         do {
2335                 err = fn(pte++, token, addr, data);
2336                 if (err)
2337                         break;
2338         } while (addr += PAGE_SIZE, addr != end);
2339
2340         arch_leave_lazy_mmu_mode();
2341
2342         if (mm != &init_mm)
2343                 pte_unmap_unlock(pte-1, ptl);
2344         return err;
2345 }
2346
2347 static int apply_to_pmd_range(struct mm_struct *mm, pud_t *pud,
2348                                      unsigned long addr, unsigned long end,
2349                                      pte_fn_t fn, void *data)
2350 {
2351         pmd_t *pmd;
2352         unsigned long next;
2353         int err;
2354
2355         BUG_ON(pud_huge(*pud));
2356
2357         pmd = pmd_alloc(mm, pud, addr);
2358         if (!pmd)
2359                 return -ENOMEM;
2360         do {
2361                 next = pmd_addr_end(addr, end);
2362                 err = apply_to_pte_range(mm, pmd, addr, next, fn, data);
2363                 if (err)
2364                         break;
2365         } while (pmd++, addr = next, addr != end);
2366         return err;
2367 }
2368
2369 static int apply_to_pud_range(struct mm_struct *mm, pgd_t *pgd,
2370                                      unsigned long addr, unsigned long end,
2371                                      pte_fn_t fn, void *data)
2372 {
2373         pud_t *pud;
2374         unsigned long next;
2375         int err;
2376
2377         pud = pud_alloc(mm, pgd, addr);
2378         if (!pud)
2379                 return -ENOMEM;
2380         do {
2381                 next = pud_addr_end(addr, end);
2382                 err = apply_to_pmd_range(mm, pud, addr, next, fn, data);
2383                 if (err)
2384                         break;
2385         } while (pud++, addr = next, addr != end);
2386         return err;
2387 }
2388
2389 /*
2390  * Scan a region of virtual memory, filling in page tables as necessary
2391  * and calling a provided function on each leaf page table.
2392  */
2393 int apply_to_page_range(struct mm_struct *mm, unsigned long addr,
2394                         unsigned long size, pte_fn_t fn, void *data)
2395 {
2396         pgd_t *pgd;
2397         unsigned long next;
2398         unsigned long end = addr + size;
2399         int err;
2400
2401         BUG_ON(addr >= end);
2402         pgd = pgd_offset(mm, addr);
2403         do {
2404                 next = pgd_addr_end(addr, end);
2405                 err = apply_to_pud_range(mm, pgd, addr, next, fn, data);
2406                 if (err)
2407                         break;
2408         } while (pgd++, addr = next, addr != end);
2409
2410         return err;
2411 }
2412 EXPORT_SYMBOL_GPL(apply_to_page_range);
2413
2414 /*
2415  * handle_pte_fault chooses page fault handler according to an entry
2416  * which was read non-atomically.  Before making any commitment, on
2417  * those architectures or configurations (e.g. i386 with PAE) which
2418  * might give a mix of unmatched parts, do_swap_page and do_nonlinear_fault
2419  * must check under lock before unmapping the pte and proceeding
2420  * (but do_wp_page is only called after already making such a check;
2421  * and do_anonymous_page can safely check later on).
2422  */
2423 static inline int pte_unmap_same(struct mm_struct *mm, pmd_t *pmd,
2424                                 pte_t *page_table, pte_t orig_pte)
2425 {
2426         int same = 1;
2427 #if defined(CONFIG_SMP) || defined(CONFIG_PREEMPT)
2428         if (sizeof(pte_t) > sizeof(unsigned long)) {
2429                 spinlock_t *ptl = pte_lockptr(mm, pmd);
2430                 spin_lock(ptl);
2431                 same = pte_same(*page_table, orig_pte);
2432                 spin_unlock(ptl);
2433         }
2434 #endif
2435         pte_unmap(page_table);
2436         return same;
2437 }
2438
2439 static inline void cow_user_page(struct page *dst, struct page *src, unsigned long va, struct vm_area_struct *vma)
2440 {
2441         /*
2442          * If the source page was a PFN mapping, we don't have
2443          * a "struct page" for it. We do a best-effort copy by
2444          * just copying from the original user address. If that
2445          * fails, we just zero-fill it. Live with it.
2446          */
2447         if (unlikely(!src)) {
2448                 void *kaddr = kmap_atomic(dst, KM_USER0);
2449                 void __user *uaddr = (void __user *)(va & PAGE_MASK);
2450
2451                 /*
2452                  * This really shouldn't fail, because the page is there
2453                  * in the page tables. But it might just be unreadable,
2454                  * in which case we just give up and fill the result with
2455                  * zeroes.
2456                  */
2457                 if (__copy_from_user_inatomic(kaddr, uaddr, PAGE_SIZE))
2458                         clear_page(kaddr);
2459                 kunmap_atomic(kaddr, KM_USER0);
2460                 flush_dcache_page(dst);
2461         } else
2462                 copy_user_highpage(dst, src, va, vma);
2463 }
2464
2465 /*
2466  * This routine handles present pages, when users try to write
2467  * to a shared page. It is done by copying the page to a new address
2468  * and decrementing the shared-page counter for the old page.
2469  *
2470  * Note that this routine assumes that the protection checks have been
2471  * done by the caller (the low-level page fault routine in most cases).
2472  * Thus we can safely just mark it writable once we've done any necessary
2473  * COW.
2474  *
2475  * We also mark the page dirty at this point even though the page will
2476  * change only once the write actually happens. This avoids a few races,
2477  * and potentially makes it more efficient.
2478  *
2479  * We enter with non-exclusive mmap_sem (to exclude vma changes,
2480  * but allow concurrent faults), with pte both mapped and locked.
2481  * We return with mmap_sem still held, but pte unmapped and unlocked.
2482  */
2483 static int do_wp_page(struct mm_struct *mm, struct vm_area_struct *vma,
2484                 unsigned long address, pte_t *page_table, pmd_t *pmd,
2485                 spinlock_t *ptl, pte_t orig_pte)
2486         __releases(ptl)
2487 {
2488         struct page *old_page, *new_page;
2489         pte_t entry;
2490         int ret = 0;
2491         int page_mkwrite = 0;
2492         struct page *dirty_page = NULL;
2493
2494         old_page = vm_normal_page(vma, address, orig_pte);
2495         if (!old_page) {
2496                 /*
2497                  * VM_MIXEDMAP !pfn_valid() case
2498                  *
2499                  * We should not cow pages in a shared writeable mapping.
2500                  * Just mark the pages writable as we can't do any dirty
2501                  * accounting on raw pfn maps.
2502                  */
2503                 if ((vma->vm_flags & (VM_WRITE|VM_SHARED)) ==
2504                                      (VM_WRITE|VM_SHARED))
2505                         goto reuse;
2506                 goto gotten;
2507         }
2508
2509         /*
2510          * Take out anonymous pages first, anonymous shared vmas are
2511          * not dirty accountable.
2512          */
2513         if (PageAnon(old_page) && !PageKsm(old_page)) {
2514                 if (!trylock_page(old_page)) {
2515                         page_cache_get(old_page);
2516                         pte_unmap_unlock(page_table, ptl);
2517                         lock_page(old_page);
2518                         page_table = pte_offset_map_lock(mm, pmd, address,
2519                                                          &ptl);
2520                         if (!pte_same(*page_table, orig_pte)) {
2521                                 unlock_page(old_page);
2522                                 goto unlock;
2523                         }
2524                         page_cache_release(old_page);
2525                 }
2526                 if (reuse_swap_page(old_page)) {
2527                         /*
2528                          * The page is all ours.  Move it to our anon_vma so
2529                          * the rmap code will not search our parent or siblings.
2530                          * Protected against the rmap code by the page lock.
2531                          */
2532                         page_move_anon_rmap(old_page, vma, address);
2533                         unlock_page(old_page);
2534                         goto reuse;
2535                 }
2536                 unlock_page(old_page);
2537         } else if (unlikely((vma->vm_flags & (VM_WRITE|VM_SHARED)) ==
2538                                         (VM_WRITE|VM_SHARED))) {
2539                 /*
2540                  * Only catch write-faults on shared writable pages,
2541                  * read-only shared pages can get COWed by
2542                  * get_user_pages(.write=1, .force=1).
2543                  */
2544                 if (vma->vm_ops && vma->vm_ops->page_mkwrite) {
2545                         struct vm_fault vmf;
2546                         int tmp;
2547
2548                         vmf.virtual_address = (void __user *)(address &
2549                                                                 PAGE_MASK);
2550                         vmf.pgoff = old_page->index;
2551                         vmf.flags = FAULT_FLAG_WRITE|FAULT_FLAG_MKWRITE;
2552                         vmf.page = old_page;
2553
2554                         /*
2555                          * Notify the address space that the page is about to
2556                          * become writable so that it can prohibit this or wait
2557                          * for the page to get into an appropriate state.
2558                          *
2559                          * We do this without the lock held, so that it can
2560                          * sleep if it needs to.
2561                          */
2562                         page_cache_get(old_page);
2563                         pte_unmap_unlock(page_table, ptl);
2564
2565                         tmp = vma->vm_ops->page_mkwrite(vma, &vmf);
2566                         if (unlikely(tmp &
2567                                         (VM_FAULT_ERROR | VM_FAULT_NOPAGE))) {
2568                                 ret = tmp;
2569                                 goto unwritable_page;
2570                         }
2571                         if (unlikely(!(tmp & VM_FAULT_LOCKED))) {
2572                                 lock_page(old_page);
2573                                 if (!old_page->mapping) {
2574                                         ret = 0; /* retry the fault */
2575                                         unlock_page(old_page);
2576                                         goto unwritable_page;
2577                                 }
2578                         } else
2579                                 VM_BUG_ON(!PageLocked(old_page));
2580
2581                         /*
2582                          * Since we dropped the lock we need to revalidate
2583                          * the PTE as someone else may have changed it.  If
2584                          * they did, we just return, as we can count on the
2585                          * MMU to tell us if they didn't also make it writable.
2586                          */
2587                         page_table = pte_offset_map_lock(mm, pmd, address,
2588                                                          &ptl);
2589                         if (!pte_same(*page_table, orig_pte)) {
2590                                 unlock_page(old_page);
2591                                 goto unlock;
2592                         }
2593
2594                         page_mkwrite = 1;
2595                 }
2596                 dirty_page = old_page;
2597                 get_page(dirty_page);
2598
2599 reuse:
2600                 flush_cache_page(vma, address, pte_pfn(orig_pte));
2601                 entry = pte_mkyoung(orig_pte);
2602                 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2603                 if (ptep_set_access_flags(vma, address, page_table, entry,1))
2604                         update_mmu_cache(vma, address, page_table);
2605                 pte_unmap_unlock(page_table, ptl);
2606                 ret |= VM_FAULT_WRITE;
2607
2608                 if (!dirty_page)
2609                         return ret;
2610
2611                 /*
2612                  * Yes, Virginia, this is actually required to prevent a race
2613                  * with clear_page_dirty_for_io() from clearing the page dirty
2614                  * bit after it clear all dirty ptes, but before a racing
2615                  * do_wp_page installs a dirty pte.
2616                  *
2617                  * __do_fault is protected similarly.
2618                  */
2619                 if (!page_mkwrite) {
2620                         wait_on_page_locked(dirty_page);
2621                         set_page_dirty_balance(dirty_page, page_mkwrite);
2622                 }
2623                 put_page(dirty_page);
2624                 if (page_mkwrite) {
2625                         struct address_space *mapping = dirty_page->mapping;
2626
2627                         set_page_dirty(dirty_page);
2628                         unlock_page(dirty_page);
2629                         page_cache_release(dirty_page);
2630                         if (mapping)    {
2631                                 /*
2632                                  * Some device drivers do not set page.mapping
2633                                  * but still dirty their pages
2634                                  */
2635                                 balance_dirty_pages_ratelimited(mapping);
2636                         }
2637                 }
2638
2639                 /* file_update_time outside page_lock */
2640                 if (vma->vm_file)
2641                         file_update_time(vma->vm_file);
2642
2643                 return ret;
2644         }
2645
2646         /*
2647          * Ok, we need to copy. Oh, well..
2648          */
2649         page_cache_get(old_page);
2650 gotten:
2651         pte_unmap_unlock(page_table, ptl);
2652
2653         if (unlikely(anon_vma_prepare(vma)))
2654                 goto oom;
2655
2656         if (is_zero_pfn(pte_pfn(orig_pte))) {
2657                 new_page = alloc_zeroed_user_highpage_movable(vma, address);
2658                 if (!new_page)
2659                         goto oom;
2660         } else {
2661                 new_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, address);
2662                 if (!new_page)
2663                         goto oom;
2664                 cow_user_page(new_page, old_page, address, vma);
2665         }
2666         __SetPageUptodate(new_page);
2667
2668         if (mem_cgroup_newpage_charge(new_page, mm, GFP_KERNEL))
2669                 goto oom_free_new;
2670
2671         /*
2672          * Re-check the pte - we dropped the lock
2673          */
2674         page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2675         if (likely(pte_same(*page_table, orig_pte))) {
2676                 if (old_page) {
2677                         if (!PageAnon(old_page)) {
2678                                 dec_mm_counter_fast(mm, MM_FILEPAGES);
2679                                 inc_mm_counter_fast(mm, MM_ANONPAGES);
2680                         }
2681                 } else
2682                         inc_mm_counter_fast(mm, MM_ANONPAGES);
2683                 flush_cache_page(vma, address, pte_pfn(orig_pte));
2684                 entry = mk_pte(new_page, vma->vm_page_prot);
2685                 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2686                 /*
2687                  * Clear the pte entry and flush it first, before updating the
2688                  * pte with the new entry. This will avoid a race condition
2689                  * seen in the presence of one thread doing SMC and another
2690                  * thread doing COW.
2691                  */
2692                 ptep_clear_flush(vma, address, page_table);
2693                 page_add_new_anon_rmap(new_page, vma, address);
2694                 /*
2695                  * We call the notify macro here because, when using secondary
2696                  * mmu page tables (such as kvm shadow page tables), we want the
2697                  * new page to be mapped directly into the secondary page table.
2698                  */
2699                 set_pte_at_notify(mm, address, page_table, entry);
2700                 update_mmu_cache(vma, address, page_table);
2701                 if (old_page) {
2702                         /*
2703                          * Only after switching the pte to the new page may
2704                          * we remove the mapcount here. Otherwise another
2705                          * process may come and find the rmap count decremented
2706                          * before the pte is switched to the new page, and
2707                          * "reuse" the old page writing into it while our pte
2708                          * here still points into it and can be read by other
2709                          * threads.
2710                          *
2711                          * The critical issue is to order this
2712                          * page_remove_rmap with the ptp_clear_flush above.
2713                          * Those stores are ordered by (if nothing else,)
2714                          * the barrier present in the atomic_add_negative
2715                          * in page_remove_rmap.
2716                          *
2717                          * Then the TLB flush in ptep_clear_flush ensures that
2718                          * no process can access the old page before the
2719                          * decremented mapcount is visible. And the old page
2720                          * cannot be reused until after the decremented
2721                          * mapcount is visible. So transitively, TLBs to
2722                          * old page will be flushed before it can be reused.
2723                          */
2724                         page_remove_rmap(old_page);
2725                 }
2726
2727                 /* Free the old page.. */
2728                 new_page = old_page;
2729                 ret |= VM_FAULT_WRITE;
2730         } else
2731                 mem_cgroup_uncharge_page(new_page);
2732
2733         if (new_page)
2734                 page_cache_release(new_page);
2735 unlock:
2736         pte_unmap_unlock(page_table, ptl);
2737         if (old_page) {
2738                 /*
2739                  * Don't let another task, with possibly unlocked vma,
2740                  * keep the mlocked page.
2741                  */
2742                 if ((ret & VM_FAULT_WRITE) && (vma->vm_flags & VM_LOCKED)) {
2743                         lock_page(old_page);    /* LRU manipulation */
2744                         munlock_vma_page(old_page);
2745                         unlock_page(old_page);
2746                 }
2747                 page_cache_release(old_page);
2748         }
2749         return ret;
2750 oom_free_new:
2751         page_cache_release(new_page);
2752 oom:
2753         if (old_page) {
2754                 if (page_mkwrite) {
2755                         unlock_page(old_page);
2756                         page_cache_release(old_page);
2757                 }
2758                 page_cache_release(old_page);
2759         }
2760         return VM_FAULT_OOM;
2761
2762 unwritable_page:
2763         page_cache_release(old_page);
2764         return ret;
2765 }
2766
2767 static void unmap_mapping_range_vma(struct vm_area_struct *vma,
2768                 unsigned long start_addr, unsigned long end_addr,
2769                 struct zap_details *details)
2770 {
2771         zap_page_range(vma, start_addr, end_addr - start_addr, details);
2772 }
2773
2774 static inline void unmap_mapping_range_tree(struct prio_tree_root *root,
2775                                             struct zap_details *details)
2776 {
2777         struct vm_area_struct *vma;
2778         struct prio_tree_iter iter;
2779         pgoff_t vba, vea, zba, zea;
2780
2781         vma_prio_tree_foreach(vma, &iter, root,
2782                         details->first_index, details->last_index) {
2783
2784                 vba = vma->vm_pgoff;
2785                 vea = vba + ((vma->vm_end - vma->vm_start) >> PAGE_SHIFT) - 1;
2786                 /* Assume for now that PAGE_CACHE_SHIFT == PAGE_SHIFT */
2787                 zba = details->first_index;
2788                 if (zba < vba)
2789                         zba = vba;
2790                 zea = details->last_index;
2791                 if (zea > vea)
2792                         zea = vea;
2793
2794                 unmap_mapping_range_vma(vma,
2795                         ((zba - vba) << PAGE_SHIFT) + vma->vm_start,
2796                         ((zea - vba + 1) << PAGE_SHIFT) + vma->vm_start,
2797                                 details);
2798         }
2799 }
2800
2801 static inline void unmap_mapping_range_list(struct list_head *head,
2802                                             struct zap_details *details)
2803 {
2804         struct vm_area_struct *vma;
2805
2806         /*
2807          * In nonlinear VMAs there is no correspondence between virtual address
2808          * offset and file offset.  So we must perform an exhaustive search
2809          * across *all* the pages in each nonlinear VMA, not just the pages
2810          * whose virtual address lies outside the file truncation point.
2811          */
2812         list_for_each_entry(vma, head, shared.vm_set.list) {
2813                 details->nonlinear_vma = vma;
2814                 unmap_mapping_range_vma(vma, vma->vm_start, vma->vm_end, details);
2815         }
2816 }
2817
2818 /**
2819  * unmap_mapping_range - unmap the portion of all mmaps in the specified address_space corresponding to the specified page range in the underlying file.
2820  * @mapping: the address space containing mmaps to be unmapped.
2821  * @holebegin: byte in first page to unmap, relative to the start of
2822  * the underlying file.  This will be rounded down to a PAGE_SIZE
2823  * boundary.  Note that this is different from truncate_pagecache(), which
2824  * must keep the partial page.  In contrast, we must get rid of
2825  * partial pages.
2826  * @holelen: size of prospective hole in bytes.  This will be rounded
2827  * up to a PAGE_SIZE boundary.  A holelen of zero truncates to the
2828  * end of the file.
2829  * @even_cows: 1 when truncating a file, unmap even private COWed pages;
2830  * but 0 when invalidating pagecache, don't throw away private data.
2831  */
2832 void unmap_mapping_range(struct address_space *mapping,
2833                 loff_t const holebegin, loff_t const holelen, int even_cows)
2834 {
2835         struct zap_details details;
2836         pgoff_t hba = holebegin >> PAGE_SHIFT;
2837         pgoff_t hlen = (holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
2838
2839         /* Check for overflow. */
2840         if (sizeof(holelen) > sizeof(hlen)) {
2841                 long long holeend =
2842                         (holebegin + holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
2843                 if (holeend & ~(long long)ULONG_MAX)
2844                         hlen = ULONG_MAX - hba + 1;
2845         }
2846
2847         details.check_mapping = even_cows? NULL: mapping;
2848         details.nonlinear_vma = NULL;
2849         details.first_index = hba;
2850         details.last_index = hba + hlen - 1;
2851         if (details.last_index < details.first_index)
2852                 details.last_index = ULONG_MAX;
2853
2854
2855         mutex_lock(&mapping->i_mmap_mutex);
2856         if (unlikely(!prio_tree_empty(&mapping->i_mmap)))
2857                 unmap_mapping_range_tree(&mapping->i_mmap, &details);
2858         if (unlikely(!list_empty(&mapping->i_mmap_nonlinear)))
2859                 unmap_mapping_range_list(&mapping->i_mmap_nonlinear, &details);
2860         mutex_unlock(&mapping->i_mmap_mutex);
2861 }
2862 EXPORT_SYMBOL(unmap_mapping_range);
2863
2864 /*
2865  * We enter with non-exclusive mmap_sem (to exclude vma changes,
2866  * but allow concurrent faults), and pte mapped but not yet locked.
2867  * We return with mmap_sem still held, but pte unmapped and unlocked.
2868  */
2869 static int do_swap_page(struct mm_struct *mm, struct vm_area_struct *vma,
2870                 unsigned long address, pte_t *page_table, pmd_t *pmd,
2871                 unsigned int flags, pte_t orig_pte)
2872 {
2873         spinlock_t *ptl;
2874         struct page *page, *swapcache = NULL;
2875         swp_entry_t entry;
2876         pte_t pte;
2877         int locked;
2878         struct mem_cgroup *ptr;
2879         int exclusive = 0;
2880         int ret = 0;
2881
2882         if (!pte_unmap_same(mm, pmd, page_table, orig_pte))
2883                 goto out;
2884
2885         entry = pte_to_swp_entry(orig_pte);
2886         if (unlikely(non_swap_entry(entry))) {
2887                 if (is_migration_entry(entry)) {
2888                         migration_entry_wait(mm, pmd, address);
2889                 } else if (is_hwpoison_entry(entry)) {
2890                         ret = VM_FAULT_HWPOISON;
2891                 } else {
2892                         print_bad_pte(vma, address, orig_pte, NULL);
2893                         ret = VM_FAULT_SIGBUS;
2894                 }
2895                 goto out;
2896         }
2897         delayacct_set_flag(DELAYACCT_PF_SWAPIN);
2898         page = lookup_swap_cache(entry);
2899         if (!page) {
2900                 grab_swap_token(mm); /* Contend for token _before_ read-in */
2901                 page = swapin_readahead(entry,
2902                                         GFP_HIGHUSER_MOVABLE, vma, address);
2903                 if (!page) {
2904                         /*
2905                          * Back out if somebody else faulted in this pte
2906                          * while we released the pte lock.
2907                          */
2908                         page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2909                         if (likely(pte_same(*page_table, orig_pte)))
2910                                 ret = VM_FAULT_OOM;
2911                         delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
2912                         goto unlock;
2913                 }
2914
2915                 /* Had to read the page from swap area: Major fault */
2916                 ret = VM_FAULT_MAJOR;
2917                 count_vm_event(PGMAJFAULT);
2918                 mem_cgroup_count_vm_event(mm, PGMAJFAULT);
2919         } else if (PageHWPoison(page)) {
2920                 /*
2921                  * hwpoisoned dirty swapcache pages are kept for killing
2922                  * owner processes (which may be unknown at hwpoison time)
2923                  */
2924                 ret = VM_FAULT_HWPOISON;
2925                 delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
2926                 goto out_release;
2927         }
2928
2929         locked = lock_page_or_retry(page, mm, flags);
2930         delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
2931         if (!locked) {
2932                 ret |= VM_FAULT_RETRY;
2933                 goto out_release;
2934         }
2935
2936         /*
2937          * Make sure try_to_free_swap or reuse_swap_page or swapoff did not
2938          * release the swapcache from under us.  The page pin, and pte_same
2939          * test below, are not enough to exclude that.  Even if it is still
2940          * swapcache, we need to check that the page's swap has not changed.
2941          */
2942         if (unlikely(!PageSwapCache(page) || page_private(page) != entry.val))
2943                 goto out_page;
2944
2945         if (ksm_might_need_to_copy(page, vma, address)) {
2946                 swapcache = page;
2947                 page = ksm_does_need_to_copy(page, vma, address);
2948
2949                 if (unlikely(!page)) {
2950                         ret = VM_FAULT_OOM;
2951                         page = swapcache;
2952                         swapcache = NULL;
2953                         goto out_page;
2954                 }
2955         }
2956
2957         if (mem_cgroup_try_charge_swapin(mm, page, GFP_KERNEL, &ptr)) {
2958                 ret = VM_FAULT_OOM;
2959                 goto out_page;
2960         }
2961
2962         /*
2963          * Back out if somebody else already faulted in this pte.
2964          */
2965         page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2966         if (unlikely(!pte_same(*page_table, orig_pte)))
2967                 goto out_nomap;
2968
2969         if (unlikely(!PageUptodate(page))) {
2970                 ret = VM_FAULT_SIGBUS;
2971                 goto out_nomap;
2972         }
2973
2974         /*
2975          * The page isn't present yet, go ahead with the fault.
2976          *
2977          * Be careful about the sequence of operations here.
2978          * To get its accounting right, reuse_swap_page() must be called
2979          * while the page is counted on swap but not yet in mapcount i.e.
2980          * before page_add_anon_rmap() and swap_free(); try_to_free_swap()
2981          * must be called after the swap_free(), or it will never succeed.
2982          * Because delete_from_swap_page() may be called by reuse_swap_page(),
2983          * mem_cgroup_commit_charge_swapin() may not be able to find swp_entry
2984          * in page->private. In this case, a record in swap_cgroup  is silently
2985          * discarded at swap_free().
2986          */
2987
2988         inc_mm_counter_fast(mm, MM_ANONPAGES);
2989         dec_mm_counter_fast(mm, MM_SWAPENTS);
2990         pte = mk_pte(page, vma->vm_page_prot);
2991         if ((flags & FAULT_FLAG_WRITE) && reuse_swap_page(page)) {
2992                 pte = maybe_mkwrite(pte_mkdirty(pte), vma);
2993                 flags &= ~FAULT_FLAG_WRITE;
2994                 ret |= VM_FAULT_WRITE;
2995                 exclusive = 1;
2996         }
2997         flush_icache_page(vma, page);
2998         set_pte_at(mm, address, page_table, pte);
2999         do_page_add_anon_rmap(page, vma, address, exclusive);
3000         /* It's better to call commit-charge after rmap is established */
3001         mem_cgroup_commit_charge_swapin(page, ptr);
3002
3003         swap_free(entry);
3004         if (vm_swap_full() || (vma->vm_flags & VM_LOCKED) || PageMlocked(page))
3005                 try_to_free_swap(page);
3006         unlock_page(page);
3007         if (swapcache) {
3008                 /*
3009                  * Hold the lock to avoid the swap entry to be reused
3010                  * until we take the PT lock for the pte_same() check
3011                  * (to avoid false positives from pte_same). For
3012                  * further safety release the lock after the swap_free
3013                  * so that the swap count won't change under a
3014                  * parallel locked swapcache.
3015                  */
3016                 unlock_page(swapcache);
3017                 page_cache_release(swapcache);
3018         }
3019
3020         if (flags & FAULT_FLAG_WRITE) {
3021                 ret |= do_wp_page(mm, vma, address, page_table, pmd, ptl, pte);
3022                 if (ret & VM_FAULT_ERROR)
3023                         ret &= VM_FAULT_ERROR;
3024                 goto out;
3025         }
3026
3027         /* No need to invalidate - it was non-present before */
3028         update_mmu_cache(vma, address, page_table);
3029 unlock:
3030         pte_unmap_unlock(page_table, ptl);
3031 out:
3032         return ret;
3033 out_nomap:
3034         mem_cgroup_cancel_charge_swapin(ptr);
3035         pte_unmap_unlock(page_table, ptl);
3036 out_page:
3037         unlock_page(page);
3038 out_release:
3039         page_cache_release(page);
3040         if (swapcache) {
3041                 unlock_page(swapcache);
3042                 page_cache_release(swapcache);
3043         }
3044         return ret;
3045 }
3046
3047 /*
3048  * This is like a special single-page "expand_{down|up}wards()",
3049  * except we must first make sure that 'address{-|+}PAGE_SIZE'
3050  * doesn't hit another vma.
3051  */
3052 static inline int check_stack_guard_page(struct vm_area_struct *vma, unsigned long address)
3053 {
3054         address &= PAGE_MASK;
3055         if ((vma->vm_flags & VM_GROWSDOWN) && address == vma->vm_start) {
3056                 struct vm_area_struct *prev = vma->vm_prev;
3057
3058                 /*
3059                  * Is there a mapping abutting this one below?
3060                  *
3061                  * That's only ok if it's the same stack mapping
3062                  * that has gotten split..
3063                  */
3064                 if (prev && prev->vm_end == address)
3065                         return prev->vm_flags & VM_GROWSDOWN ? 0 : -ENOMEM;
3066
3067                 expand_downwards(vma, address - PAGE_SIZE);
3068         }
3069         if ((vma->vm_flags & VM_GROWSUP) && address + PAGE_SIZE == vma->vm_end) {
3070                 struct vm_area_struct *next = vma->vm_next;
3071
3072                 /* As VM_GROWSDOWN but s/below/above/ */
3073                 if (next && next->vm_start == address + PAGE_SIZE)
3074                         return next->vm_flags & VM_GROWSUP ? 0 : -ENOMEM;
3075
3076                 expand_upwards(vma, address + PAGE_SIZE);
3077         }
3078         return 0;
3079 }
3080
3081 /*
3082  * We enter with non-exclusive mmap_sem (to exclude vma changes,
3083  * but allow concurrent faults), and pte mapped but not yet locked.
3084  * We return with mmap_sem still held, but pte unmapped and unlocked.
3085  */
3086 static int do_anonymous_page(struct mm_struct *mm, struct vm_area_struct *vma,
3087                 unsigned long address, pte_t *page_table, pmd_t *pmd,
3088                 unsigned int flags)
3089 {
3090         struct page *page;
3091         spinlock_t *ptl;
3092         pte_t entry;
3093
3094         pte_unmap(page_table);
3095
3096         /* Check if we need to add a guard page to the stack */
3097         if (check_stack_guard_page(vma, address) < 0)
3098                 return VM_FAULT_SIGBUS;
3099
3100         /* Use the zero-page for reads */
3101         if (!(flags & FAULT_FLAG_WRITE)) {
3102                 entry = pte_mkspecial(pfn_pte(my_zero_pfn(address),
3103                                                 vma->vm_page_prot));
3104                 page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
3105                 if (!pte_none(*page_table))
3106                         goto unlock;
3107                 goto setpte;
3108         }
3109
3110         /* Allocate our own private page. */
3111         if (unlikely(anon_vma_prepare(vma)))
3112                 goto oom;
3113         page = alloc_zeroed_user_highpage_movable(vma, address);
3114         if (!page)
3115                 goto oom;
3116         __SetPageUptodate(page);
3117
3118         if (mem_cgroup_newpage_charge(page, mm, GFP_KERNEL))
3119                 goto oom_free_page;
3120
3121         entry = mk_pte(page, vma->vm_page_prot);
3122         if (vma->vm_flags & VM_WRITE)
3123                 entry = pte_mkwrite(pte_mkdirty(entry));
3124
3125         page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
3126         if (!pte_none(*page_table))
3127                 goto release;
3128
3129         inc_mm_counter_fast(mm, MM_ANONPAGES);
3130         page_add_new_anon_rmap(page, vma, address);
3131 setpte:
3132         set_pte_at(mm, address, page_table, entry);
3133
3134         /* No need to invalidate - it was non-present before */
3135         update_mmu_cache(vma, address, page_table);
3136 unlock:
3137         pte_unmap_unlock(page_table, ptl);
3138         return 0;
3139 release:
3140         mem_cgroup_uncharge_page(page);
3141         page_cache_release(page);
3142         goto unlock;
3143 oom_free_page:
3144         page_cache_release(page);
3145 oom:
3146         return VM_FAULT_OOM;
3147 }
3148
3149 /*
3150  * __do_fault() tries to create a new page mapping. It aggressively
3151  * tries to share with existing pages, but makes a separate copy if
3152  * the FAULT_FLAG_WRITE is set in the flags parameter in order to avoid
3153  * the next page fault.
3154  *
3155  * As this is called only for pages that do not currently exist, we
3156  * do not need to flush old virtual caches or the TLB.
3157  *
3158  * We enter with non-exclusive mmap_sem (to exclude vma changes,
3159  * but allow concurrent faults), and pte neither mapped nor locked.
3160  * We return with mmap_sem still held, but pte unmapped and unlocked.
3161  */
3162 static int __do_fault(struct mm_struct *mm, struct vm_area_struct *vma,
3163                 unsigned long address, pmd_t *pmd,
3164                 pgoff_t pgoff, unsigned int flags, pte_t orig_pte)
3165 {
3166         pte_t *page_table;
3167         spinlock_t *ptl;
3168         struct page *page;
3169         struct page *cow_page;
3170         pte_t entry;
3171         int anon = 0;
3172         struct page *dirty_page = NULL;
3173         struct vm_fault vmf;
3174         int ret;
3175         int page_mkwrite = 0;
3176
3177         /*
3178          * If we do COW later, allocate page befor taking lock_page()
3179          * on the file cache page. This will reduce lock holding time.
3180          */
3181         if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) {
3182
3183                 if (unlikely(anon_vma_prepare(vma)))
3184                         return VM_FAULT_OOM;
3185
3186                 cow_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, address);
3187                 if (!cow_page)
3188                         return VM_FAULT_OOM;
3189
3190                 if (mem_cgroup_newpage_charge(cow_page, mm, GFP_KERNEL)) {
3191                         page_cache_release(cow_page);
3192                         return VM_FAULT_OOM;
3193                 }
3194         } else
3195                 cow_page = NULL;
3196
3197         vmf.virtual_address = (void __user *)(address & PAGE_MASK);
3198         vmf.pgoff = pgoff;
3199         vmf.flags = flags;
3200         vmf.page = NULL;
3201
3202         ret = vma->vm_ops->fault(vma, &vmf);
3203         if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE |
3204                             VM_FAULT_RETRY)))
3205                 goto uncharge_out;
3206
3207         if (unlikely(PageHWPoison(vmf.page))) {
3208                 if (ret & VM_FAULT_LOCKED)
3209                         unlock_page(vmf.page);
3210                 ret = VM_FAULT_HWPOISON;
3211                 goto uncharge_out;
3212         }
3213
3214         /*
3215          * For consistency in subsequent calls, make the faulted page always
3216          * locked.
3217          */
3218         if (unlikely(!(ret & VM_FAULT_LOCKED)))
3219                 lock_page(vmf.page);
3220         else
3221                 VM_BUG_ON(!PageLocked(vmf.page));
3222
3223         /*
3224          * Should we do an early C-O-W break?
3225          */
3226         page = vmf.page;
3227         if (flags & FAULT_FLAG_WRITE) {
3228                 if (!(vma->vm_flags & VM_SHARED)) {
3229                         page = cow_page;
3230                         anon = 1;
3231                         copy_user_highpage(page, vmf.page, address, vma);
3232                         __SetPageUptodate(page);
3233                 } else {
3234                         /*
3235                          * If the page will be shareable, see if the backing
3236                          * address space wants to know that the page is about
3237                          * to become writable
3238                          */
3239                         if (vma->vm_ops->page_mkwrite) {
3240                                 int tmp;
3241
3242                                 unlock_page(page);
3243                                 vmf.flags = FAULT_FLAG_WRITE|FAULT_FLAG_MKWRITE;
3244                                 tmp = vma->vm_ops->page_mkwrite(vma, &vmf);
3245                                 if (unlikely(tmp &
3246                                           (VM_FAULT_ERROR | VM_FAULT_NOPAGE))) {
3247                                         ret = tmp;
3248                                         goto unwritable_page;
3249                                 }
3250                                 if (unlikely(!(tmp & VM_FAULT_LOCKED))) {
3251                                         lock_page(page);
3252                                         if (!page->mapping) {
3253                                                 ret = 0; /* retry the fault */
3254                                                 unlock_page(page);
3255                                                 goto unwritable_page;
3256                                         }
3257                                 } else
3258                                         VM_BUG_ON(!PageLocked(page));
3259                                 page_mkwrite = 1;
3260                         }
3261                 }
3262
3263         }
3264
3265         page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
3266
3267         /*
3268          * This silly early PAGE_DIRTY setting removes a race
3269          * due to the bad i386 page protection. But it's valid