b9a95f103d492ce2a741ebd10b974c31ab4c07c2
[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 (likely(!non_swap_entry(entry))) {
874                                 if (swap_duplicate(entry) < 0)
875                                         return entry.val;
876
877                                 /* make sure dst_mm is on swapoff's mmlist. */
878                                 if (unlikely(list_empty(&dst_mm->mmlist))) {
879                                         spin_lock(&mmlist_lock);
880                                         if (list_empty(&dst_mm->mmlist))
881                                                 list_add(&dst_mm->mmlist,
882                                                          &src_mm->mmlist);
883                                         spin_unlock(&mmlist_lock);
884                                 }
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                                 addr += PAGE_SIZE;
1183                                 break;
1184                         }
1185                         continue;
1186                 }
1187                 /*
1188                  * If details->check_mapping, we leave swap entries;
1189                  * if details->nonlinear_vma, we leave file entries.
1190                  */
1191                 if (unlikely(details))
1192                         continue;
1193                 if (pte_file(ptent)) {
1194                         if (unlikely(!(vma->vm_flags & VM_NONLINEAR)))
1195                                 print_bad_pte(vma, addr, ptent, NULL);
1196                 } else {
1197                         swp_entry_t entry = pte_to_swp_entry(ptent);
1198
1199                         if (!non_swap_entry(entry))
1200                                 rss[MM_SWAPENTS]--;
1201                         if (unlikely(!free_swap_and_cache(entry)))
1202                                 print_bad_pte(vma, addr, ptent, NULL);
1203                 }
1204                 pte_clear_not_present_full(mm, addr, pte, tlb->fullmm);
1205         } while (pte++, addr += PAGE_SIZE, addr != end);
1206
1207         add_mm_rss_vec(mm, rss);
1208         arch_leave_lazy_mmu_mode();
1209         pte_unmap_unlock(start_pte, ptl);
1210
1211         /*
1212          * mmu_gather ran out of room to batch pages, we break out of
1213          * the PTE lock to avoid doing the potential expensive TLB invalidate
1214          * and page-free while holding it.
1215          */
1216         if (force_flush) {
1217                 force_flush = 0;
1218                 tlb_flush_mmu(tlb);
1219                 if (addr != end)
1220                         goto again;
1221         }
1222
1223         return addr;
1224 }
1225
1226 static inline unsigned long zap_pmd_range(struct mmu_gather *tlb,
1227                                 struct vm_area_struct *vma, pud_t *pud,
1228                                 unsigned long addr, unsigned long end,
1229                                 struct zap_details *details)
1230 {
1231         pmd_t *pmd;
1232         unsigned long next;
1233
1234         pmd = pmd_offset(pud, addr);
1235         do {
1236                 next = pmd_addr_end(addr, end);
1237                 if (pmd_trans_huge(*pmd)) {
1238                         if (next - addr != HPAGE_PMD_SIZE) {
1239                                 VM_BUG_ON(!rwsem_is_locked(&tlb->mm->mmap_sem));
1240                                 split_huge_page_pmd(vma->vm_mm, pmd);
1241                         } else if (zap_huge_pmd(tlb, vma, pmd))
1242                                 goto next;
1243                         /* fall through */
1244                 }
1245                 /*
1246                  * Here there can be other concurrent MADV_DONTNEED or
1247                  * trans huge page faults running, and if the pmd is
1248                  * none or trans huge it can change under us. This is
1249                  * because MADV_DONTNEED holds the mmap_sem in read
1250                  * mode.
1251                  */
1252                 if (pmd_none_or_trans_huge_or_clear_bad(pmd))
1253                         goto next;
1254                 next = zap_pte_range(tlb, vma, pmd, addr, next, details);
1255 next:
1256                 cond_resched();
1257         } while (pmd++, addr = next, addr != end);
1258
1259         return addr;
1260 }
1261
1262 static inline unsigned long zap_pud_range(struct mmu_gather *tlb,
1263                                 struct vm_area_struct *vma, pgd_t *pgd,
1264                                 unsigned long addr, unsigned long end,
1265                                 struct zap_details *details)
1266 {
1267         pud_t *pud;
1268         unsigned long next;
1269
1270         pud = pud_offset(pgd, addr);
1271         do {
1272                 next = pud_addr_end(addr, end);
1273                 if (pud_none_or_clear_bad(pud))
1274                         continue;
1275                 next = zap_pmd_range(tlb, vma, pud, addr, next, details);
1276         } while (pud++, addr = next, addr != end);
1277
1278         return addr;
1279 }
1280
1281 static unsigned long unmap_page_range(struct mmu_gather *tlb,
1282                                 struct vm_area_struct *vma,
1283                                 unsigned long addr, unsigned long end,
1284                                 struct zap_details *details)
1285 {
1286         pgd_t *pgd;
1287         unsigned long next;
1288
1289         if (details && !details->check_mapping && !details->nonlinear_vma)
1290                 details = NULL;
1291
1292         BUG_ON(addr >= end);
1293         mem_cgroup_uncharge_start();
1294         tlb_start_vma(tlb, vma);
1295         pgd = pgd_offset(vma->vm_mm, addr);
1296         do {
1297                 next = pgd_addr_end(addr, end);
1298                 if (pgd_none_or_clear_bad(pgd))
1299                         continue;
1300                 next = zap_pud_range(tlb, vma, pgd, addr, next, details);
1301         } while (pgd++, addr = next, addr != end);
1302         tlb_end_vma(tlb, vma);
1303         mem_cgroup_uncharge_end();
1304
1305         return addr;
1306 }
1307
1308 /**
1309  * unmap_vmas - unmap a range of memory covered by a list of vma's
1310  * @tlb: address of the caller's struct mmu_gather
1311  * @vma: the starting vma
1312  * @start_addr: virtual address at which to start unmapping
1313  * @end_addr: virtual address at which to end unmapping
1314  * @nr_accounted: Place number of unmapped pages in vm-accountable vma's here
1315  * @details: details of nonlinear truncation or shared cache invalidation
1316  *
1317  * Returns the end address of the unmapping (restart addr if interrupted).
1318  *
1319  * Unmap all pages in the vma list.
1320  *
1321  * Only addresses between `start' and `end' will be unmapped.
1322  *
1323  * The VMA list must be sorted in ascending virtual address order.
1324  *
1325  * unmap_vmas() assumes that the caller will flush the whole unmapped address
1326  * range after unmap_vmas() returns.  So the only responsibility here is to
1327  * ensure that any thus-far unmapped pages are flushed before unmap_vmas()
1328  * drops the lock and schedules.
1329  */
1330 unsigned long unmap_vmas(struct mmu_gather *tlb,
1331                 struct vm_area_struct *vma, unsigned long start_addr,
1332                 unsigned long end_addr, unsigned long *nr_accounted,
1333                 struct zap_details *details)
1334 {
1335         unsigned long start = start_addr;
1336         struct mm_struct *mm = vma->vm_mm;
1337
1338         mmu_notifier_invalidate_range_start(mm, start_addr, end_addr);
1339         for ( ; vma && vma->vm_start < end_addr; vma = vma->vm_next) {
1340                 unsigned long end;
1341
1342                 start = max(vma->vm_start, start_addr);
1343                 if (start >= vma->vm_end)
1344                         continue;
1345                 end = min(vma->vm_end, end_addr);
1346                 if (end <= vma->vm_start)
1347                         continue;
1348
1349                 if (vma->vm_flags & VM_ACCOUNT)
1350                         *nr_accounted += (end - start) >> PAGE_SHIFT;
1351
1352                 if (unlikely(is_pfn_mapping(vma)))
1353                         untrack_pfn_vma(vma, 0, 0);
1354
1355                 while (start != end) {
1356                         if (unlikely(is_vm_hugetlb_page(vma))) {
1357                                 /*
1358                                  * It is undesirable to test vma->vm_file as it
1359                                  * should be non-null for valid hugetlb area.
1360                                  * However, vm_file will be NULL in the error
1361                                  * cleanup path of do_mmap_pgoff. When
1362                                  * hugetlbfs ->mmap method fails,
1363                                  * do_mmap_pgoff() nullifies vma->vm_file
1364                                  * before calling this function to clean up.
1365                                  * Since no pte has actually been setup, it is
1366                                  * safe to do nothing in this case.
1367                                  */
1368                                 if (vma->vm_file) {
1369                                         mutex_lock(&vma->vm_file->f_mapping->i_mmap_mutex);
1370                                         __unmap_hugepage_range_final(vma, start, end, NULL);
1371                                         mutex_unlock(&vma->vm_file->f_mapping->i_mmap_mutex);
1372                                 }
1373
1374                                 start = end;
1375                         } else
1376                                 start = unmap_page_range(tlb, vma, start, end, details);
1377                 }
1378         }
1379
1380         mmu_notifier_invalidate_range_end(mm, start_addr, end_addr);
1381         return start;   /* which is now the end (or restart) address */
1382 }
1383
1384 /**
1385  * zap_page_range - remove user pages in a given range
1386  * @vma: vm_area_struct holding the applicable pages
1387  * @address: starting address of pages to zap
1388  * @size: number of bytes to zap
1389  * @details: details of nonlinear truncation or shared cache invalidation
1390  */
1391 unsigned long zap_page_range(struct vm_area_struct *vma, unsigned long address,
1392                 unsigned long size, struct zap_details *details)
1393 {
1394         struct mm_struct *mm = vma->vm_mm;
1395         struct mmu_gather tlb;
1396         unsigned long end = address + size;
1397         unsigned long nr_accounted = 0;
1398
1399         lru_add_drain();
1400         tlb_gather_mmu(&tlb, mm, 0);
1401         update_hiwater_rss(mm);
1402         end = unmap_vmas(&tlb, vma, address, end, &nr_accounted, details);
1403         tlb_finish_mmu(&tlb, address, end);
1404         return end;
1405 }
1406 EXPORT_SYMBOL_GPL(zap_page_range);
1407
1408 /**
1409  * zap_vma_ptes - remove ptes mapping the vma
1410  * @vma: vm_area_struct holding ptes to be zapped
1411  * @address: starting address of pages to zap
1412  * @size: number of bytes to zap
1413  *
1414  * This function only unmaps ptes assigned to VM_PFNMAP vmas.
1415  *
1416  * The entire address range must be fully contained within the vma.
1417  *
1418  * Returns 0 if successful.
1419  */
1420 int zap_vma_ptes(struct vm_area_struct *vma, unsigned long address,
1421                 unsigned long size)
1422 {
1423         if (address < vma->vm_start || address + size > vma->vm_end ||
1424                         !(vma->vm_flags & VM_PFNMAP))
1425                 return -1;
1426         zap_page_range(vma, address, size, NULL);
1427         return 0;
1428 }
1429 EXPORT_SYMBOL_GPL(zap_vma_ptes);
1430
1431 /**
1432  * follow_page - look up a page descriptor from a user-virtual address
1433  * @vma: vm_area_struct mapping @address
1434  * @address: virtual address to look up
1435  * @flags: flags modifying lookup behaviour
1436  *
1437  * @flags can have FOLL_ flags set, defined in <linux/mm.h>
1438  *
1439  * Returns the mapped (struct page *), %NULL if no mapping exists, or
1440  * an error pointer if there is a mapping to something not represented
1441  * by a page descriptor (see also vm_normal_page()).
1442  */
1443 struct page *follow_page(struct vm_area_struct *vma, unsigned long address,
1444                         unsigned int flags)
1445 {
1446         pgd_t *pgd;
1447         pud_t *pud;
1448         pmd_t *pmd;
1449         pte_t *ptep, pte;
1450         spinlock_t *ptl;
1451         struct page *page;
1452         struct mm_struct *mm = vma->vm_mm;
1453
1454         page = follow_huge_addr(mm, address, flags & FOLL_WRITE);
1455         if (!IS_ERR(page)) {
1456                 BUG_ON(flags & FOLL_GET);
1457                 goto out;
1458         }
1459
1460         page = NULL;
1461         pgd = pgd_offset(mm, address);
1462         if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
1463                 goto no_page_table;
1464
1465         pud = pud_offset(pgd, address);
1466         if (pud_none(*pud))
1467                 goto no_page_table;
1468         if (pud_huge(*pud) && vma->vm_flags & VM_HUGETLB) {
1469                 BUG_ON(flags & FOLL_GET);
1470                 page = follow_huge_pud(mm, address, pud, flags & FOLL_WRITE);
1471                 goto out;
1472         }
1473         if (unlikely(pud_bad(*pud)))
1474                 goto no_page_table;
1475
1476         pmd = pmd_offset(pud, address);
1477         if (pmd_none(*pmd))
1478                 goto no_page_table;
1479         if (pmd_huge(*pmd) && vma->vm_flags & VM_HUGETLB) {
1480                 BUG_ON(flags & FOLL_GET);
1481                 page = follow_huge_pmd(mm, address, pmd, flags & FOLL_WRITE);
1482                 goto out;
1483         }
1484         if (pmd_trans_huge(*pmd)) {
1485                 if (flags & FOLL_SPLIT) {
1486                         split_huge_page_pmd(mm, pmd);
1487                         goto split_fallthrough;
1488                 }
1489                 spin_lock(&mm->page_table_lock);
1490                 if (likely(pmd_trans_huge(*pmd))) {
1491                         if (unlikely(pmd_trans_splitting(*pmd))) {
1492                                 spin_unlock(&mm->page_table_lock);
1493                                 wait_split_huge_page(vma->anon_vma, pmd);
1494                         } else {
1495                                 page = follow_trans_huge_pmd(mm, address,
1496                                                              pmd, flags);
1497                                 spin_unlock(&mm->page_table_lock);
1498                                 goto out;
1499                         }
1500                 } else
1501                         spin_unlock(&mm->page_table_lock);
1502                 /* fall through */
1503         }
1504 split_fallthrough:
1505         if (unlikely(pmd_bad(*pmd)))
1506                 goto no_page_table;
1507
1508         ptep = pte_offset_map_lock(mm, pmd, address, &ptl);
1509
1510         pte = *ptep;
1511         if (!pte_present(pte))
1512                 goto no_page;
1513         if ((flags & FOLL_WRITE) && !pte_write(pte))
1514                 goto unlock;
1515
1516         page = vm_normal_page(vma, address, pte);
1517         if (unlikely(!page)) {
1518                 if ((flags & FOLL_DUMP) ||
1519                     !is_zero_pfn(pte_pfn(pte)))
1520                         goto bad_page;
1521                 page = pte_page(pte);
1522         }
1523
1524         if (flags & FOLL_GET)
1525                 get_page_foll(page);
1526         if (flags & FOLL_TOUCH) {
1527                 if ((flags & FOLL_WRITE) &&
1528                     !pte_dirty(pte) && !PageDirty(page))
1529                         set_page_dirty(page);
1530                 /*
1531                  * pte_mkyoung() would be more correct here, but atomic care
1532                  * is needed to avoid losing the dirty bit: it is easier to use
1533                  * mark_page_accessed().
1534                  */
1535                 mark_page_accessed(page);
1536         }
1537         if ((flags & FOLL_MLOCK) && (vma->vm_flags & VM_LOCKED)) {
1538                 /*
1539                  * The preliminary mapping check is mainly to avoid the
1540                  * pointless overhead of lock_page on the ZERO_PAGE
1541                  * which might bounce very badly if there is contention.
1542                  *
1543                  * If the page is already locked, we don't need to
1544                  * handle it now - vmscan will handle it later if and
1545                  * when it attempts to reclaim the page.
1546                  */
1547                 if (page->mapping && trylock_page(page)) {
1548                         lru_add_drain();  /* push cached pages to LRU */
1549                         /*
1550                          * Because we lock page here and migration is
1551                          * blocked by the pte's page reference, we need
1552                          * only check for file-cache page truncation.
1553                          */
1554                         if (page->mapping)
1555                                 mlock_vma_page(page);
1556                         unlock_page(page);
1557                 }
1558         }
1559 unlock:
1560         pte_unmap_unlock(ptep, ptl);
1561 out:
1562         return page;
1563
1564 bad_page:
1565         pte_unmap_unlock(ptep, ptl);
1566         return ERR_PTR(-EFAULT);
1567
1568 no_page:
1569         pte_unmap_unlock(ptep, ptl);
1570         if (!pte_none(pte))
1571                 return page;
1572
1573 no_page_table:
1574         /*
1575          * When core dumping an enormous anonymous area that nobody
1576          * has touched so far, we don't want to allocate unnecessary pages or
1577          * page tables.  Return error instead of NULL to skip handle_mm_fault,
1578          * then get_dump_page() will return NULL to leave a hole in the dump.
1579          * But we can only make this optimization where a hole would surely
1580          * be zero-filled if handle_mm_fault() actually did handle it.
1581          */
1582         if ((flags & FOLL_DUMP) &&
1583             (!vma->vm_ops || !vma->vm_ops->fault))
1584                 return ERR_PTR(-EFAULT);
1585         return page;
1586 }
1587
1588 static inline int stack_guard_page(struct vm_area_struct *vma, unsigned long addr)
1589 {
1590         return stack_guard_page_start(vma, addr) ||
1591                stack_guard_page_end(vma, addr+PAGE_SIZE);
1592 }
1593
1594 /**
1595  * __get_user_pages() - pin user pages in memory
1596  * @tsk:        task_struct of target task
1597  * @mm:         mm_struct of target mm
1598  * @start:      starting user address
1599  * @nr_pages:   number of pages from start to pin
1600  * @gup_flags:  flags modifying pin behaviour
1601  * @pages:      array that receives pointers to the pages pinned.
1602  *              Should be at least nr_pages long. Or NULL, if caller
1603  *              only intends to ensure the pages are faulted in.
1604  * @vmas:       array of pointers to vmas corresponding to each page.
1605  *              Or NULL if the caller does not require them.
1606  * @nonblocking: whether waiting for disk IO or mmap_sem contention
1607  *
1608  * Returns number of pages pinned. This may be fewer than the number
1609  * requested. If nr_pages is 0 or negative, returns 0. If no pages
1610  * were pinned, returns -errno. Each page returned must be released
1611  * with a put_page() call when it is finished with. vmas will only
1612  * remain valid while mmap_sem is held.
1613  *
1614  * Must be called with mmap_sem held for read or write.
1615  *
1616  * __get_user_pages walks a process's page tables and takes a reference to
1617  * each struct page that each user address corresponds to at a given
1618  * instant. That is, it takes the page that would be accessed if a user
1619  * thread accesses the given user virtual address at that instant.
1620  *
1621  * This does not guarantee that the page exists in the user mappings when
1622  * __get_user_pages returns, and there may even be a completely different
1623  * page there in some cases (eg. if mmapped pagecache has been invalidated
1624  * and subsequently re faulted). However it does guarantee that the page
1625  * won't be freed completely. And mostly callers simply care that the page
1626  * contains data that was valid *at some point in time*. Typically, an IO
1627  * or similar operation cannot guarantee anything stronger anyway because
1628  * locks can't be held over the syscall boundary.
1629  *
1630  * If @gup_flags & FOLL_WRITE == 0, the page must not be written to. If
1631  * the page is written to, set_page_dirty (or set_page_dirty_lock, as
1632  * appropriate) must be called after the page is finished with, and
1633  * before put_page is called.
1634  *
1635  * If @nonblocking != NULL, __get_user_pages will not wait for disk IO
1636  * or mmap_sem contention, and if waiting is needed to pin all pages,
1637  * *@nonblocking will be set to 0.
1638  *
1639  * In most cases, get_user_pages or get_user_pages_fast should be used
1640  * instead of __get_user_pages. __get_user_pages should be used only if
1641  * you need some special @gup_flags.
1642  */
1643 int __get_user_pages(struct task_struct *tsk, struct mm_struct *mm,
1644                      unsigned long start, int nr_pages, unsigned int gup_flags,
1645                      struct page **pages, struct vm_area_struct **vmas,
1646                      int *nonblocking)
1647 {
1648         int i;
1649         unsigned long vm_flags;
1650
1651         if (nr_pages <= 0)
1652                 return 0;
1653
1654         VM_BUG_ON(!!pages != !!(gup_flags & FOLL_GET));
1655
1656         /* 
1657          * Require read or write permissions.
1658          * If FOLL_FORCE is set, we only require the "MAY" flags.
1659          */
1660         vm_flags  = (gup_flags & FOLL_WRITE) ?
1661                         (VM_WRITE | VM_MAYWRITE) : (VM_READ | VM_MAYREAD);
1662         vm_flags &= (gup_flags & FOLL_FORCE) ?
1663                         (VM_MAYREAD | VM_MAYWRITE) : (VM_READ | VM_WRITE);
1664         i = 0;
1665
1666         do {
1667                 struct vm_area_struct *vma;
1668
1669                 vma = find_extend_vma(mm, start);
1670                 if (!vma && in_gate_area(mm, start)) {
1671                         unsigned long pg = start & PAGE_MASK;
1672                         pgd_t *pgd;
1673                         pud_t *pud;
1674                         pmd_t *pmd;
1675                         pte_t *pte;
1676
1677                         /* user gate pages are read-only */
1678                         if (gup_flags & FOLL_WRITE)
1679                                 return i ? : -EFAULT;
1680                         if (pg > TASK_SIZE)
1681                                 pgd = pgd_offset_k(pg);
1682                         else
1683                                 pgd = pgd_offset_gate(mm, pg);
1684                         BUG_ON(pgd_none(*pgd));
1685                         pud = pud_offset(pgd, pg);
1686                         BUG_ON(pud_none(*pud));
1687                         pmd = pmd_offset(pud, pg);
1688                         if (pmd_none(*pmd))
1689                                 return i ? : -EFAULT;
1690                         VM_BUG_ON(pmd_trans_huge(*pmd));
1691                         pte = pte_offset_map(pmd, pg);
1692                         if (pte_none(*pte)) {
1693                                 pte_unmap(pte);
1694                                 return i ? : -EFAULT;
1695                         }
1696                         vma = get_gate_vma(mm);
1697                         if (pages) {
1698                                 struct page *page;
1699
1700                                 page = vm_normal_page(vma, start, *pte);
1701                                 if (!page) {
1702                                         if (!(gup_flags & FOLL_DUMP) &&
1703                                              is_zero_pfn(pte_pfn(*pte)))
1704                                                 page = pte_page(*pte);
1705                                         else {
1706                                                 pte_unmap(pte);
1707                                                 return i ? : -EFAULT;
1708                                         }
1709                                 }
1710                                 pages[i] = page;
1711                                 get_page(page);
1712                         }
1713                         pte_unmap(pte);
1714                         goto next_page;
1715                 }
1716
1717                 if (!vma ||
1718                     (vma->vm_flags & (VM_IO | VM_PFNMAP)) ||
1719                     !(vm_flags & vma->vm_flags))
1720                         return i ? : -EFAULT;
1721
1722                 if (is_vm_hugetlb_page(vma)) {
1723                         i = follow_hugetlb_page(mm, vma, pages, vmas,
1724                                         &start, &nr_pages, i, gup_flags);
1725                         continue;
1726                 }
1727
1728                 do {
1729                         struct page *page;
1730                         unsigned int foll_flags = gup_flags;
1731
1732                         /*
1733                          * If we have a pending SIGKILL, don't keep faulting
1734                          * pages and potentially allocating memory.
1735                          */
1736                         if (unlikely(fatal_signal_pending(current)))
1737                                 return i ? i : -ERESTARTSYS;
1738
1739                         cond_resched();
1740                         while (!(page = follow_page(vma, start, foll_flags))) {
1741                                 int ret;
1742                                 unsigned int fault_flags = 0;
1743
1744                                 /* For mlock, just skip the stack guard page. */
1745                                 if (foll_flags & FOLL_MLOCK) {
1746                                         if (stack_guard_page(vma, start))
1747                                                 goto next_page;
1748                                 }
1749                                 if (foll_flags & FOLL_WRITE)
1750                                         fault_flags |= FAULT_FLAG_WRITE;
1751                                 if (nonblocking)
1752                                         fault_flags |= FAULT_FLAG_ALLOW_RETRY;
1753                                 if (foll_flags & FOLL_NOWAIT)
1754                                         fault_flags |= (FAULT_FLAG_ALLOW_RETRY | FAULT_FLAG_RETRY_NOWAIT);
1755
1756                                 ret = handle_mm_fault(mm, vma, start,
1757                                                         fault_flags);
1758
1759                                 if (ret & VM_FAULT_ERROR) {
1760                                         if (ret & VM_FAULT_OOM)
1761                                                 return i ? i : -ENOMEM;
1762                                         if (ret & (VM_FAULT_HWPOISON |
1763                                                    VM_FAULT_HWPOISON_LARGE)) {
1764                                                 if (i)
1765                                                         return i;
1766                                                 else if (gup_flags & FOLL_HWPOISON)
1767                                                         return -EHWPOISON;
1768                                                 else
1769                                                         return -EFAULT;
1770                                         }
1771                                         if (ret & (VM_FAULT_SIGBUS | VM_FAULT_SIGSEGV))
1772                                                 return i ? i : -EFAULT;
1773                                         BUG();
1774                                 }
1775
1776                                 if (tsk) {
1777                                         if (ret & VM_FAULT_MAJOR)
1778                                                 tsk->maj_flt++;
1779                                         else
1780                                                 tsk->min_flt++;
1781                                 }
1782
1783                                 if (ret & VM_FAULT_RETRY) {
1784                                         if (nonblocking)
1785                                                 *nonblocking = 0;
1786                                         return i;
1787                                 }
1788
1789                                 /*
1790                                  * The VM_FAULT_WRITE bit tells us that
1791                                  * do_wp_page has broken COW when necessary,
1792                                  * even if maybe_mkwrite decided not to set
1793                                  * pte_write. We can thus safely do subsequent
1794                                  * page lookups as if they were reads. But only
1795                                  * do so when looping for pte_write is futile:
1796                                  * in some cases userspace may also be wanting
1797                                  * to write to the gotten user page, which a
1798                                  * read fault here might prevent (a readonly
1799                                  * page might get reCOWed by userspace write).
1800                                  */
1801                                 if ((ret & VM_FAULT_WRITE) &&
1802                                     !(vma->vm_flags & VM_WRITE))
1803                                         foll_flags &= ~FOLL_WRITE;
1804
1805                                 cond_resched();
1806                         }
1807                         if (IS_ERR(page))
1808                                 return i ? i : PTR_ERR(page);
1809                         if (pages) {
1810                                 pages[i] = page;
1811
1812                                 flush_anon_page(vma, page, start);
1813                                 flush_dcache_page(page);
1814                         }
1815 next_page:
1816                         if (vmas)
1817                                 vmas[i] = vma;
1818                         i++;
1819                         start += PAGE_SIZE;
1820                         nr_pages--;
1821                 } while (nr_pages && start < vma->vm_end);
1822         } while (nr_pages);
1823         return i;
1824 }
1825 EXPORT_SYMBOL(__get_user_pages);
1826
1827 /*
1828  * fixup_user_fault() - manually resolve a user page fault
1829  * @tsk:        the task_struct to use for page fault accounting, or
1830  *              NULL if faults are not to be recorded.
1831  * @mm:         mm_struct of target mm
1832  * @address:    user address
1833  * @fault_flags:flags to pass down to handle_mm_fault()
1834  *
1835  * This is meant to be called in the specific scenario where for locking reasons
1836  * we try to access user memory in atomic context (within a pagefault_disable()
1837  * section), this returns -EFAULT, and we want to resolve the user fault before
1838  * trying again.
1839  *
1840  * Typically this is meant to be used by the futex code.
1841  *
1842  * The main difference with get_user_pages() is that this function will
1843  * unconditionally call handle_mm_fault() which will in turn perform all the
1844  * necessary SW fixup of the dirty and young bits in the PTE, while
1845  * handle_mm_fault() only guarantees to update these in the struct page.
1846  *
1847  * This is important for some architectures where those bits also gate the
1848  * access permission to the page because they are maintained in software.  On
1849  * such architectures, gup() will not be enough to make a subsequent access
1850  * succeed.
1851  *
1852  * This should be called with the mm_sem held for read.
1853  */
1854 int fixup_user_fault(struct task_struct *tsk, struct mm_struct *mm,
1855                      unsigned long address, unsigned int fault_flags)
1856 {
1857         struct vm_area_struct *vma;
1858         vm_flags_t vm_flags;
1859         int ret;
1860
1861         vma = find_extend_vma(mm, address);
1862         if (!vma || address < vma->vm_start)
1863                 return -EFAULT;
1864
1865         vm_flags = (fault_flags & FAULT_FLAG_WRITE) ? VM_WRITE : VM_READ;
1866         if (!(vm_flags & vma->vm_flags))
1867                 return -EFAULT;
1868
1869         ret = handle_mm_fault(mm, vma, address, fault_flags);
1870         if (ret & VM_FAULT_ERROR) {
1871                 if (ret & VM_FAULT_OOM)
1872                         return -ENOMEM;
1873                 if (ret & (VM_FAULT_HWPOISON | VM_FAULT_HWPOISON_LARGE))
1874                         return -EHWPOISON;
1875                 if (ret & (VM_FAULT_SIGBUS | VM_FAULT_SIGSEGV))
1876                         return -EFAULT;
1877                 BUG();
1878         }
1879         if (tsk) {
1880                 if (ret & VM_FAULT_MAJOR)
1881                         tsk->maj_flt++;
1882                 else
1883                         tsk->min_flt++;
1884         }
1885         return 0;
1886 }
1887
1888 /*
1889  * get_user_pages() - pin user pages in memory
1890  * @tsk:        the task_struct to use for page fault accounting, or
1891  *              NULL if faults are not to be recorded.
1892  * @mm:         mm_struct of target mm
1893  * @start:      starting user address
1894  * @nr_pages:   number of pages from start to pin
1895  * @write:      whether pages will be written to by the caller
1896  * @force:      whether to force write access even if user mapping is
1897  *              readonly. This will result in the page being COWed even
1898  *              in MAP_SHARED mappings. You do not want this.
1899  * @pages:      array that receives pointers to the pages pinned.
1900  *              Should be at least nr_pages long. Or NULL, if caller
1901  *              only intends to ensure the pages are faulted in.
1902  * @vmas:       array of pointers to vmas corresponding to each page.
1903  *              Or NULL if the caller does not require them.
1904  *
1905  * Returns number of pages pinned. This may be fewer than the number
1906  * requested. If nr_pages is 0 or negative, returns 0. If no pages
1907  * were pinned, returns -errno. Each page returned must be released
1908  * with a put_page() call when it is finished with. vmas will only
1909  * remain valid while mmap_sem is held.
1910  *
1911  * Must be called with mmap_sem held for read or write.
1912  *
1913  * get_user_pages walks a process's page tables and takes a reference to
1914  * each struct page that each user address corresponds to at a given
1915  * instant. That is, it takes the page that would be accessed if a user
1916  * thread accesses the given user virtual address at that instant.
1917  *
1918  * This does not guarantee that the page exists in the user mappings when
1919  * get_user_pages returns, and there may even be a completely different
1920  * page there in some cases (eg. if mmapped pagecache has been invalidated
1921  * and subsequently re faulted). However it does guarantee that the page
1922  * won't be freed completely. And mostly callers simply care that the page
1923  * contains data that was valid *at some point in time*. Typically, an IO
1924  * or similar operation cannot guarantee anything stronger anyway because
1925  * locks can't be held over the syscall boundary.
1926  *
1927  * If write=0, the page must not be written to. If the page is written to,
1928  * set_page_dirty (or set_page_dirty_lock, as appropriate) must be called
1929  * after the page is finished with, and before put_page is called.
1930  *
1931  * get_user_pages is typically used for fewer-copy IO operations, to get a
1932  * handle on the memory by some means other than accesses via the user virtual
1933  * addresses. The pages may be submitted for DMA to devices or accessed via
1934  * their kernel linear mapping (via the kmap APIs). Care should be taken to
1935  * use the correct cache flushing APIs.
1936  *
1937  * See also get_user_pages_fast, for performance critical applications.
1938  */
1939 int get_user_pages(struct task_struct *tsk, struct mm_struct *mm,
1940                 unsigned long start, int nr_pages, int write, int force,
1941                 struct page **pages, struct vm_area_struct **vmas)
1942 {
1943         int flags = FOLL_TOUCH;
1944
1945         if (pages)
1946                 flags |= FOLL_GET;
1947         if (write)
1948                 flags |= FOLL_WRITE;
1949         if (force)
1950                 flags |= FOLL_FORCE;
1951
1952         return __get_user_pages(tsk, mm, start, nr_pages, flags, pages, vmas,
1953                                 NULL);
1954 }
1955 EXPORT_SYMBOL(get_user_pages);
1956
1957 /**
1958  * get_dump_page() - pin user page in memory while writing it to core dump
1959  * @addr: user address
1960  *
1961  * Returns struct page pointer of user page pinned for dump,
1962  * to be freed afterwards by page_cache_release() or put_page().
1963  *
1964  * Returns NULL on any kind of failure - a hole must then be inserted into
1965  * the corefile, to preserve alignment with its headers; and also returns
1966  * NULL wherever the ZERO_PAGE, or an anonymous pte_none, has been found -
1967  * allowing a hole to be left in the corefile to save diskspace.
1968  *
1969  * Called without mmap_sem, but after all other threads have been killed.
1970  */
1971 #ifdef CONFIG_ELF_CORE
1972 struct page *get_dump_page(unsigned long addr)
1973 {
1974         struct vm_area_struct *vma;
1975         struct page *page;
1976
1977         if (__get_user_pages(current, current->mm, addr, 1,
1978                              FOLL_FORCE | FOLL_DUMP | FOLL_GET, &page, &vma,
1979                              NULL) < 1)
1980                 return NULL;
1981         flush_cache_page(vma, addr, page_to_pfn(page));
1982         return page;
1983 }
1984 #endif /* CONFIG_ELF_CORE */
1985
1986 pte_t *__get_locked_pte(struct mm_struct *mm, unsigned long addr,
1987                         spinlock_t **ptl)
1988 {
1989         pgd_t * pgd = pgd_offset(mm, addr);
1990         pud_t * pud = pud_alloc(mm, pgd, addr);
1991         if (pud) {
1992                 pmd_t * pmd = pmd_alloc(mm, pud, addr);
1993                 if (pmd) {
1994                         VM_BUG_ON(pmd_trans_huge(*pmd));
1995                         return pte_alloc_map_lock(mm, pmd, addr, ptl);
1996                 }
1997         }
1998         return NULL;
1999 }
2000
2001 /*
2002  * This is the old fallback for page remapping.
2003  *
2004  * For historical reasons, it only allows reserved pages. Only
2005  * old drivers should use this, and they needed to mark their
2006  * pages reserved for the old functions anyway.
2007  */
2008 static int insert_page(struct vm_area_struct *vma, unsigned long addr,
2009                         struct page *page, pgprot_t prot)
2010 {
2011         struct mm_struct *mm = vma->vm_mm;
2012         int retval;
2013         pte_t *pte;
2014         spinlock_t *ptl;
2015
2016         retval = -EINVAL;
2017         if (PageAnon(page))
2018                 goto out;
2019         retval = -ENOMEM;
2020         flush_dcache_page(page);
2021         pte = get_locked_pte(mm, addr, &ptl);
2022         if (!pte)
2023                 goto out;
2024         retval = -EBUSY;
2025         if (!pte_none(*pte))
2026                 goto out_unlock;
2027
2028         /* Ok, finally just insert the thing.. */
2029         get_page(page);
2030         inc_mm_counter_fast(mm, MM_FILEPAGES);
2031         page_add_file_rmap(page);
2032         set_pte_at(mm, addr, pte, mk_pte(page, prot));
2033
2034         retval = 0;
2035         pte_unmap_unlock(pte, ptl);
2036         return retval;
2037 out_unlock:
2038         pte_unmap_unlock(pte, ptl);
2039 out:
2040         return retval;
2041 }
2042
2043 /**
2044  * vm_insert_page - insert single page into user vma
2045  * @vma: user vma to map to
2046  * @addr: target user address of this page
2047  * @page: source kernel page
2048  *
2049  * This allows drivers to insert individual pages they've allocated
2050  * into a user vma.
2051  *
2052  * The page has to be a nice clean _individual_ kernel allocation.
2053  * If you allocate a compound page, you need to have marked it as
2054  * such (__GFP_COMP), or manually just split the page up yourself
2055  * (see split_page()).
2056  *
2057  * NOTE! Traditionally this was done with "remap_pfn_range()" which
2058  * took an arbitrary page protection parameter. This doesn't allow
2059  * that. Your vma protection will have to be set up correctly, which
2060  * means that if you want a shared writable mapping, you'd better
2061  * ask for a shared writable mapping!
2062  *
2063  * The page does not need to be reserved.
2064  */
2065 int vm_insert_page(struct vm_area_struct *vma, unsigned long addr,
2066                         struct page *page)
2067 {
2068         if (addr < vma->vm_start || addr >= vma->vm_end)
2069                 return -EFAULT;
2070         if (!page_count(page))
2071                 return -EINVAL;
2072         vma->vm_flags |= VM_INSERTPAGE;
2073         return insert_page(vma, addr, page, vma->vm_page_prot);
2074 }
2075 EXPORT_SYMBOL(vm_insert_page);
2076
2077 static int insert_pfn(struct vm_area_struct *vma, unsigned long addr,
2078                         unsigned long pfn, pgprot_t prot)
2079 {
2080         struct mm_struct *mm = vma->vm_mm;
2081         int retval;
2082         pte_t *pte, entry;
2083         spinlock_t *ptl;
2084
2085         retval = -ENOMEM;
2086         pte = get_locked_pte(mm, addr, &ptl);
2087         if (!pte)
2088                 goto out;
2089         retval = -EBUSY;
2090         if (!pte_none(*pte))
2091                 goto out_unlock;
2092
2093         /* Ok, finally just insert the thing.. */
2094         entry = pte_mkspecial(pfn_pte(pfn, prot));
2095         set_pte_at(mm, addr, pte, entry);
2096         update_mmu_cache(vma, addr, pte); /* XXX: why not for insert_page? */
2097
2098         retval = 0;
2099 out_unlock:
2100         pte_unmap_unlock(pte, ptl);
2101 out:
2102         return retval;
2103 }
2104
2105 /**
2106  * vm_insert_pfn - insert single pfn into user vma
2107  * @vma: user vma to map to
2108  * @addr: target user address of this page
2109  * @pfn: source kernel pfn
2110  *
2111  * Similar to vm_inert_page, this allows drivers to insert individual pages
2112  * they've allocated into a user vma. Same comments apply.
2113  *
2114  * This function should only be called from a vm_ops->fault handler, and
2115  * in that case the handler should return NULL.
2116  *
2117  * vma cannot be a COW mapping.
2118  *
2119  * As this is called only for pages that do not currently exist, we
2120  * do not need to flush old virtual caches or the TLB.
2121  */
2122 int vm_insert_pfn(struct vm_area_struct *vma, unsigned long addr,
2123                         unsigned long pfn)
2124 {
2125         int ret;
2126         pgprot_t pgprot = vma->vm_page_prot;
2127         /*
2128          * Technically, architectures with pte_special can avoid all these
2129          * restrictions (same for remap_pfn_range).  However we would like
2130          * consistency in testing and feature parity among all, so we should
2131          * try to keep these invariants in place for everybody.
2132          */
2133         BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)));
2134         BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) ==
2135                                                 (VM_PFNMAP|VM_MIXEDMAP));
2136         BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags));
2137         BUG_ON((vma->vm_flags & VM_MIXEDMAP) && pfn_valid(pfn));
2138
2139         if (addr < vma->vm_start || addr >= vma->vm_end)
2140                 return -EFAULT;
2141         if (track_pfn_vma_new(vma, &pgprot, pfn, PAGE_SIZE))
2142                 return -EINVAL;
2143
2144         ret = insert_pfn(vma, addr, pfn, pgprot);
2145
2146         if (ret)
2147                 untrack_pfn_vma(vma, pfn, PAGE_SIZE);
2148
2149         return ret;
2150 }
2151 EXPORT_SYMBOL(vm_insert_pfn);
2152
2153 int vm_insert_mixed(struct vm_area_struct *vma, unsigned long addr,
2154                         unsigned long pfn)
2155 {
2156         BUG_ON(!(vma->vm_flags & VM_MIXEDMAP));
2157
2158         if (addr < vma->vm_start || addr >= vma->vm_end)
2159                 return -EFAULT;
2160
2161         /*
2162          * If we don't have pte special, then we have to use the pfn_valid()
2163          * based VM_MIXEDMAP scheme (see vm_normal_page), and thus we *must*
2164          * refcount the page if pfn_valid is true (hence insert_page rather
2165          * than insert_pfn).  If a zero_pfn were inserted into a VM_MIXEDMAP
2166          * without pte special, it would there be refcounted as a normal page.
2167          */
2168         if (!HAVE_PTE_SPECIAL && pfn_valid(pfn)) {
2169                 struct page *page;
2170
2171                 page = pfn_to_page(pfn);
2172                 return insert_page(vma, addr, page, vma->vm_page_prot);
2173         }
2174         return insert_pfn(vma, addr, pfn, vma->vm_page_prot);
2175 }
2176 EXPORT_SYMBOL(vm_insert_mixed);
2177
2178 /*
2179  * maps a range of physical memory into the requested pages. the old
2180  * mappings are removed. any references to nonexistent pages results
2181  * in null mappings (currently treated as "copy-on-access")
2182  */
2183 static int remap_pte_range(struct mm_struct *mm, pmd_t *pmd,
2184                         unsigned long addr, unsigned long end,
2185                         unsigned long pfn, pgprot_t prot)
2186 {
2187         pte_t *pte;
2188         spinlock_t *ptl;
2189
2190         pte = pte_alloc_map_lock(mm, pmd, addr, &ptl);
2191         if (!pte)
2192                 return -ENOMEM;
2193         arch_enter_lazy_mmu_mode();
2194         do {
2195                 BUG_ON(!pte_none(*pte));
2196                 set_pte_at(mm, addr, pte, pte_mkspecial(pfn_pte(pfn, prot)));
2197                 pfn++;
2198         } while (pte++, addr += PAGE_SIZE, addr != end);
2199         arch_leave_lazy_mmu_mode();
2200         pte_unmap_unlock(pte - 1, ptl);
2201         return 0;
2202 }
2203
2204 static inline int remap_pmd_range(struct mm_struct *mm, pud_t *pud,
2205                         unsigned long addr, unsigned long end,
2206                         unsigned long pfn, pgprot_t prot)
2207 {
2208         pmd_t *pmd;
2209         unsigned long next;
2210
2211         pfn -= addr >> PAGE_SHIFT;
2212         pmd = pmd_alloc(mm, pud, addr);
2213         if (!pmd)
2214                 return -ENOMEM;
2215         VM_BUG_ON(pmd_trans_huge(*pmd));
2216         do {
2217                 next = pmd_addr_end(addr, end);
2218                 if (remap_pte_range(mm, pmd, addr, next,
2219                                 pfn + (addr >> PAGE_SHIFT), prot))
2220                         return -ENOMEM;
2221         } while (pmd++, addr = next, addr != end);
2222         return 0;
2223 }
2224
2225 static inline int remap_pud_range(struct mm_struct *mm, pgd_t *pgd,
2226                         unsigned long addr, unsigned long end,
2227                         unsigned long pfn, pgprot_t prot)
2228 {
2229         pud_t *pud;
2230         unsigned long next;
2231
2232         pfn -= addr >> PAGE_SHIFT;
2233         pud = pud_alloc(mm, pgd, addr);
2234         if (!pud)
2235                 return -ENOMEM;
2236         do {
2237                 next = pud_addr_end(addr, end);
2238                 if (remap_pmd_range(mm, pud, addr, next,
2239                                 pfn + (addr >> PAGE_SHIFT), prot))
2240                         return -ENOMEM;
2241         } while (pud++, addr = next, addr != end);
2242         return 0;
2243 }
2244
2245 /**
2246  * remap_pfn_range - remap kernel memory to userspace
2247  * @vma: user vma to map to
2248  * @addr: target user address to start at
2249  * @pfn: physical address of kernel memory
2250  * @size: size of map area
2251  * @prot: page protection flags for this mapping
2252  *
2253  *  Note: this is only safe if the mm semaphore is held when called.
2254  */
2255 int remap_pfn_range(struct vm_area_struct *vma, unsigned long addr,
2256                     unsigned long pfn, unsigned long size, pgprot_t prot)
2257 {
2258         pgd_t *pgd;
2259         unsigned long next;
2260         unsigned long end = addr + PAGE_ALIGN(size);
2261         struct mm_struct *mm = vma->vm_mm;
2262         int err;
2263
2264         /*
2265          * Physically remapped pages are special. Tell the
2266          * rest of the world about it:
2267          *   VM_IO tells people not to look at these pages
2268          *      (accesses can have side effects).
2269          *   VM_RESERVED is specified all over the place, because
2270          *      in 2.4 it kept swapout's vma scan off this vma; but
2271          *      in 2.6 the LRU scan won't even find its pages, so this
2272          *      flag means no more than count its pages in reserved_vm,
2273          *      and omit it from core dump, even when VM_IO turned off.
2274          *   VM_PFNMAP tells the core MM that the base pages are just
2275          *      raw PFN mappings, and do not have a "struct page" associated
2276          *      with them.
2277          *
2278          * There's a horrible special case to handle copy-on-write
2279          * behaviour that some programs depend on. We mark the "original"
2280          * un-COW'ed pages by matching them up with "vma->vm_pgoff".
2281          */
2282         if (addr == vma->vm_start && end == vma->vm_end) {
2283                 vma->vm_pgoff = pfn;
2284                 vma->vm_flags |= VM_PFN_AT_MMAP;
2285         } else if (is_cow_mapping(vma->vm_flags))
2286                 return -EINVAL;
2287
2288         vma->vm_flags |= VM_IO | VM_RESERVED | VM_PFNMAP;
2289
2290         err = track_pfn_vma_new(vma, &prot, pfn, PAGE_ALIGN(size));
2291         if (err) {
2292                 /*
2293                  * To indicate that track_pfn related cleanup is not
2294                  * needed from higher level routine calling unmap_vmas
2295                  */
2296                 vma->vm_flags &= ~(VM_IO | VM_RESERVED | VM_PFNMAP);
2297                 vma->vm_flags &= ~VM_PFN_AT_MMAP;
2298                 return -EINVAL;
2299         }
2300
2301         BUG_ON(addr >= end);
2302         pfn -= addr >> PAGE_SHIFT;
2303         pgd = pgd_offset(mm, addr);
2304         flush_cache_range(vma, addr, end);
2305         do {
2306                 next = pgd_addr_end(addr, end);
2307                 err = remap_pud_range(mm, pgd, addr, next,
2308                                 pfn + (addr >> PAGE_SHIFT), prot);
2309                 if (err)
2310                         break;
2311         } while (pgd++, addr = next, addr != end);
2312
2313         if (err)
2314                 untrack_pfn_vma(vma, pfn, PAGE_ALIGN(size));
2315
2316         return err;
2317 }
2318 EXPORT_SYMBOL(remap_pfn_range);
2319
2320 /**
2321  * vm_iomap_memory - remap memory to userspace
2322  * @vma: user vma to map to
2323  * @start: start of area
2324  * @len: size of area
2325  *
2326  * This is a simplified io_remap_pfn_range() for common driver use. The
2327  * driver just needs to give us the physical memory range to be mapped,
2328  * we'll figure out the rest from the vma information.
2329  *
2330  * NOTE! Some drivers might want to tweak vma->vm_page_prot first to get
2331  * whatever write-combining details or similar.
2332  */
2333 int vm_iomap_memory(struct vm_area_struct *vma, phys_addr_t start, unsigned long len)
2334 {
2335         unsigned long vm_len, pfn, pages;
2336
2337         /* Check that the physical memory area passed in looks valid */
2338         if (start + len < start)
2339                 return -EINVAL;
2340         /*
2341          * You *really* shouldn't map things that aren't page-aligned,
2342          * but we've historically allowed it because IO memory might
2343          * just have smaller alignment.
2344          */
2345         len += start & ~PAGE_MASK;
2346         pfn = start >> PAGE_SHIFT;
2347         pages = (len + ~PAGE_MASK) >> PAGE_SHIFT;
2348         if (pfn + pages < pfn)
2349                 return -EINVAL;
2350
2351         /* We start the mapping 'vm_pgoff' pages into the area */
2352         if (vma->vm_pgoff > pages)
2353                 return -EINVAL;
2354         pfn += vma->vm_pgoff;
2355         pages -= vma->vm_pgoff;
2356
2357         /* Can we fit all of the mapping? */
2358         vm_len = vma->vm_end - vma->vm_start;
2359         if (vm_len >> PAGE_SHIFT > pages)
2360                 return -EINVAL;
2361
2362         /* Ok, let it rip */
2363         return io_remap_pfn_range(vma, vma->vm_start, pfn, vm_len, vma->vm_page_prot);
2364 }
2365 EXPORT_SYMBOL(vm_iomap_memory);
2366
2367 static int apply_to_pte_range(struct mm_struct *mm, pmd_t *pmd,
2368                                      unsigned long addr, unsigned long end,
2369                                      pte_fn_t fn, void *data)
2370 {
2371         pte_t *pte;
2372         int err;
2373         pgtable_t token;
2374         spinlock_t *uninitialized_var(ptl);
2375
2376         pte = (mm == &init_mm) ?
2377                 pte_alloc_kernel(pmd, addr) :
2378                 pte_alloc_map_lock(mm, pmd, addr, &ptl);
2379         if (!pte)
2380                 return -ENOMEM;
2381
2382         BUG_ON(pmd_huge(*pmd));
2383
2384         arch_enter_lazy_mmu_mode();
2385
2386         token = pmd_pgtable(*pmd);
2387
2388         do {
2389                 err = fn(pte++, token, addr, data);
2390                 if (err)
2391                         break;
2392         } while (addr += PAGE_SIZE, addr != end);
2393
2394         arch_leave_lazy_mmu_mode();
2395
2396         if (mm != &init_mm)
2397                 pte_unmap_unlock(pte-1, ptl);
2398         return err;
2399 }
2400
2401 static int apply_to_pmd_range(struct mm_struct *mm, pud_t *pud,
2402                                      unsigned long addr, unsigned long end,
2403                                      pte_fn_t fn, void *data)
2404 {
2405         pmd_t *pmd;
2406         unsigned long next;
2407         int err;
2408
2409         BUG_ON(pud_huge(*pud));
2410
2411         pmd = pmd_alloc(mm, pud, addr);
2412         if (!pmd)
2413                 return -ENOMEM;
2414         do {
2415                 next = pmd_addr_end(addr, end);
2416                 err = apply_to_pte_range(mm, pmd, addr, next, fn, data);
2417                 if (err)
2418                         break;
2419         } while (pmd++, addr = next, addr != end);
2420         return err;
2421 }
2422
2423 static int apply_to_pud_range(struct mm_struct *mm, pgd_t *pgd,
2424                                      unsigned long addr, unsigned long end,
2425                                      pte_fn_t fn, void *data)
2426 {
2427         pud_t *pud;
2428         unsigned long next;
2429         int err;
2430
2431         pud = pud_alloc(mm, pgd, addr);
2432         if (!pud)
2433                 return -ENOMEM;
2434         do {
2435                 next = pud_addr_end(addr, end);
2436                 err = apply_to_pmd_range(mm, pud, addr, next, fn, data);
2437                 if (err)
2438                         break;
2439         } while (pud++, addr = next, addr != end);
2440         return err;
2441 }
2442
2443 /*
2444  * Scan a region of virtual memory, filling in page tables as necessary
2445  * and calling a provided function on each leaf page table.
2446  */
2447 int apply_to_page_range(struct mm_struct *mm, unsigned long addr,
2448                         unsigned long size, pte_fn_t fn, void *data)
2449 {
2450         pgd_t *pgd;
2451         unsigned long next;
2452         unsigned long end = addr + size;
2453         int err;
2454
2455         BUG_ON(addr >= end);
2456         pgd = pgd_offset(mm, addr);
2457         do {
2458                 next = pgd_addr_end(addr, end);
2459                 err = apply_to_pud_range(mm, pgd, addr, next, fn, data);
2460                 if (err)
2461                         break;
2462         } while (pgd++, addr = next, addr != end);
2463
2464         return err;
2465 }
2466 EXPORT_SYMBOL_GPL(apply_to_page_range);
2467
2468 /*
2469  * handle_pte_fault chooses page fault handler according to an entry
2470  * which was read non-atomically.  Before making any commitment, on
2471  * those architectures or configurations (e.g. i386 with PAE) which
2472  * might give a mix of unmatched parts, do_swap_page and do_nonlinear_fault
2473  * must check under lock before unmapping the pte and proceeding
2474  * (but do_wp_page is only called after already making such a check;
2475  * and do_anonymous_page can safely check later on).
2476  */
2477 static inline int pte_unmap_same(struct mm_struct *mm, pmd_t *pmd,
2478                                 pte_t *page_table, pte_t orig_pte)
2479 {
2480         int same = 1;
2481 #if defined(CONFIG_SMP) || defined(CONFIG_PREEMPT)
2482         if (sizeof(pte_t) > sizeof(unsigned long)) {
2483                 spinlock_t *ptl = pte_lockptr(mm, pmd);
2484                 spin_lock(ptl);
2485                 same = pte_same(*page_table, orig_pte);
2486                 spin_unlock(ptl);
2487         }
2488 #endif
2489         pte_unmap(page_table);
2490         return same;
2491 }
2492
2493 static inline void cow_user_page(struct page *dst, struct page *src, unsigned long va, struct vm_area_struct *vma)
2494 {
2495         /*
2496          * If the source page was a PFN mapping, we don't have
2497          * a "struct page" for it. We do a best-effort copy by
2498          * just copying from the original user address. If that
2499          * fails, we just zero-fill it. Live with it.
2500          */
2501         if (unlikely(!src)) {
2502                 void *kaddr = kmap_atomic(dst, KM_USER0);
2503                 void __user *uaddr = (void __user *)(va & PAGE_MASK);
2504
2505                 /*
2506                  * This really shouldn't fail, because the page is there
2507                  * in the page tables. But it might just be unreadable,
2508                  * in which case we just give up and fill the result with
2509                  * zeroes.
2510                  */
2511                 if (__copy_from_user_inatomic(kaddr, uaddr, PAGE_SIZE))
2512                         clear_page(kaddr);
2513                 kunmap_atomic(kaddr, KM_USER0);
2514                 flush_dcache_page(dst);
2515         } else
2516                 copy_user_highpage(dst, src, va, vma);
2517 }
2518
2519 /*
2520  * This routine handles present pages, when users try to write
2521  * to a shared page. It is done by copying the page to a new address
2522  * and decrementing the shared-page counter for the old page.
2523  *
2524  * Note that this routine assumes that the protection checks have been
2525  * done by the caller (the low-level page fault routine in most cases).
2526  * Thus we can safely just mark it writable once we've done any necessary
2527  * COW.
2528  *
2529  * We also mark the page dirty at this point even though the page will
2530  * change only once the write actually happens. This avoids a few races,
2531  * and potentially makes it more efficient.
2532  *
2533  * We enter with non-exclusive mmap_sem (to exclude vma changes,
2534  * but allow concurrent faults), with pte both mapped and locked.
2535  * We return with mmap_sem still held, but pte unmapped and unlocked.
2536  */
2537 static int do_wp_page(struct mm_struct *mm, struct vm_area_struct *vma,
2538                 unsigned long address, pte_t *page_table, pmd_t *pmd,
2539                 spinlock_t *ptl, pte_t orig_pte)
2540         __releases(ptl)
2541 {
2542         struct page *old_page, *new_page;
2543         pte_t entry;
2544         int ret = 0;
2545         int page_mkwrite = 0;
2546         struct page *dirty_page = NULL;
2547
2548         old_page = vm_normal_page(vma, address, orig_pte);
2549         if (!old_page) {
2550                 /*
2551                  * VM_MIXEDMAP !pfn_valid() case
2552                  *
2553                  * We should not cow pages in a shared writeable mapping.
2554                  * Just mark the pages writable as we can't do any dirty
2555                  * accounting on raw pfn maps.
2556                  */
2557                 if ((vma->vm_flags & (VM_WRITE|VM_SHARED)) ==
2558                                      (VM_WRITE|VM_SHARED))
2559                         goto reuse;
2560                 goto gotten;
2561         }
2562
2563         /*
2564          * Take out anonymous pages first, anonymous shared vmas are
2565          * not dirty accountable.
2566          */
2567         if (PageAnon(old_page) && !PageKsm(old_page)) {
2568                 if (!trylock_page(old_page)) {
2569                         page_cache_get(old_page);
2570                         pte_unmap_unlock(page_table, ptl);
2571                         lock_page(old_page);
2572                         page_table = pte_offset_map_lock(mm, pmd, address,
2573                                                          &ptl);
2574                         if (!pte_same(*page_table, orig_pte)) {
2575                                 unlock_page(old_page);
2576                                 goto unlock;
2577                         }
2578                         page_cache_release(old_page);
2579                 }
2580                 if (reuse_swap_page(old_page)) {
2581                         /*
2582                          * The page is all ours.  Move it to our anon_vma so
2583                          * the rmap code will not search our parent or siblings.
2584                          * Protected against the rmap code by the page lock.
2585                          */
2586                         page_move_anon_rmap(old_page, vma, address);
2587                         unlock_page(old_page);
2588                         goto reuse;
2589                 }
2590                 unlock_page(old_page);
2591         } else if (unlikely((vma->vm_flags & (VM_WRITE|VM_SHARED)) ==
2592                                         (VM_WRITE|VM_SHARED))) {
2593                 /*
2594                  * Only catch write-faults on shared writable pages,
2595                  * read-only shared pages can get COWed by
2596                  * get_user_pages(.write=1, .force=1).
2597                  */
2598                 if (vma->vm_ops && vma->vm_ops->page_mkwrite) {
2599                         struct vm_fault vmf;
2600                         int tmp;
2601
2602                         vmf.virtual_address = (void __user *)(address &
2603                                                                 PAGE_MASK);
2604                         vmf.pgoff = old_page->index;
2605                         vmf.flags = FAULT_FLAG_WRITE|FAULT_FLAG_MKWRITE;
2606                         vmf.page = old_page;
2607
2608                         /*
2609                          * Notify the address space that the page is about to
2610                          * become writable so that it can prohibit this or wait
2611                          * for the page to get into an appropriate state.
2612                          *
2613                          * We do this without the lock held, so that it can
2614                          * sleep if it needs to.
2615                          */
2616                         page_cache_get(old_page);
2617                         pte_unmap_unlock(page_table, ptl);
2618
2619                         tmp = vma->vm_ops->page_mkwrite(vma, &vmf);
2620                         if (unlikely(tmp &
2621                                         (VM_FAULT_ERROR | VM_FAULT_NOPAGE))) {
2622                                 ret = tmp;
2623                                 goto unwritable_page;
2624                         }
2625                         if (unlikely(!(tmp & VM_FAULT_LOCKED))) {
2626                                 lock_page(old_page);
2627                                 if (!old_page->mapping) {
2628                                         ret = 0; /* retry the fault */
2629                                         unlock_page(old_page);
2630                                         goto unwritable_page;
2631                                 }
2632                         } else
2633                                 VM_BUG_ON(!PageLocked(old_page));
2634
2635                         /*
2636                          * Since we dropped the lock we need to revalidate
2637                          * the PTE as someone else may have changed it.  If
2638                          * they did, we just return, as we can count on the
2639                          * MMU to tell us if they didn't also make it writable.
2640                          */
2641                         page_table = pte_offset_map_lock(mm, pmd, address,
2642                                                          &ptl);
2643                         if (!pte_same(*page_table, orig_pte)) {
2644                                 unlock_page(old_page);
2645                                 goto unlock;
2646                         }
2647
2648                         page_mkwrite = 1;
2649                 }
2650                 dirty_page = old_page;
2651                 get_page(dirty_page);
2652
2653 reuse:
2654                 flush_cache_page(vma, address, pte_pfn(orig_pte));
2655                 entry = pte_mkyoung(orig_pte);
2656                 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2657                 if (ptep_set_access_flags(vma, address, page_table, entry,1))
2658                         update_mmu_cache(vma, address, page_table);
2659                 pte_unmap_unlock(page_table, ptl);
2660                 ret |= VM_FAULT_WRITE;
2661
2662                 if (!dirty_page)
2663                         return ret;
2664
2665                 if (!page_mkwrite) {
2666                         struct address_space *mapping;
2667                         int dirtied;
2668
2669                         lock_page(dirty_page);
2670                         dirtied = set_page_dirty(dirty_page);
2671                         VM_BUG_ON(PageAnon(dirty_page));
2672                         mapping = dirty_page->mapping;
2673                         unlock_page(dirty_page);
2674
2675                         if (dirtied && mapping) {
2676                                 /*
2677                                  * Some device drivers do not set page.mapping
2678                                  * but still dirty their pages
2679                                  */
2680                                 balance_dirty_pages_ratelimited(mapping);
2681                         }
2682
2683                 }
2684                 put_page(dirty_page);
2685                 if (page_mkwrite) {
2686                         struct address_space *mapping = dirty_page->mapping;
2687
2688                         set_page_dirty(dirty_page);
2689                         unlock_page(dirty_page);
2690                         page_cache_release(dirty_page);
2691                         if (mapping)    {
2692                                 /*
2693                                  * Some device drivers do not set page.mapping
2694                                  * but still dirty their pages
2695                                  */
2696                                 balance_dirty_pages_ratelimited(mapping);
2697                         }
2698                 }
2699
2700                 /* file_update_time outside page_lock */
2701                 if (vma->vm_file)
2702                         file_update_time(vma->vm_file);
2703
2704                 return ret;
2705         }
2706
2707         /*
2708          * Ok, we need to copy. Oh, well..
2709          */
2710         page_cache_get(old_page);
2711 gotten:
2712         pte_unmap_unlock(page_table, ptl);
2713
2714         if (unlikely(anon_vma_prepare(vma)))
2715                 goto oom;
2716
2717         if (is_zero_pfn(pte_pfn(orig_pte))) {
2718                 new_page = alloc_zeroed_user_highpage_movable(vma, address);
2719                 if (!new_page)
2720                         goto oom;
2721         } else {
2722                 new_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, address);
2723                 if (!new_page)
2724                         goto oom;
2725                 cow_user_page(new_page, old_page, address, vma);
2726         }
2727         __SetPageUptodate(new_page);
2728
2729         if (mem_cgroup_newpage_charge(new_page, mm, GFP_KERNEL))
2730                 goto oom_free_new;
2731
2732         /*
2733          * Re-check the pte - we dropped the lock
2734          */
2735         page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2736         if (likely(pte_same(*page_table, orig_pte))) {
2737                 if (old_page) {
2738                         if (!PageAnon(old_page)) {
2739                                 dec_mm_counter_fast(mm, MM_FILEPAGES);
2740                                 inc_mm_counter_fast(mm, MM_ANONPAGES);
2741                         }
2742                 } else
2743                         inc_mm_counter_fast(mm, MM_ANONPAGES);
2744                 flush_cache_page(vma, address, pte_pfn(orig_pte));
2745                 entry = mk_pte(new_page, vma->vm_page_prot);
2746                 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2747                 /*
2748                  * Clear the pte entry and flush it first, before updating the
2749                  * pte with the new entry. This will avoid a race condition
2750                  * seen in the presence of one thread doing SMC and another
2751                  * thread doing COW.
2752                  */
2753                 ptep_clear_flush(vma, address, page_table);
2754                 page_add_new_anon_rmap(new_page, vma, address);
2755                 /*
2756                  * We call the notify macro here because, when using secondary
2757                  * mmu page tables (such as kvm shadow page tables), we want the
2758                  * new page to be mapped directly into the secondary page table.
2759                  */
2760                 set_pte_at_notify(mm, address, page_table, entry);
2761                 update_mmu_cache(vma, address, page_table);
2762                 if (old_page) {
2763                         /*
2764                          * Only after switching the pte to the new page may
2765                          * we remove the mapcount here. Otherwise another
2766                          * process may come and find the rmap count decremented
2767                          * before the pte is switched to the new page, and
2768                          * "reuse" the old page writing into it while our pte
2769                          * here still points into it and can be read by other
2770                          * threads.
2771                          *
2772                          * The critical issue is to order this
2773                          * page_remove_rmap with the ptp_clear_flush above.
2774                          * Those stores are ordered by (if nothing else,)
2775                          * the barrier present in the atomic_add_negative
2776                          * in page_remove_rmap.
2777                          *
2778                          * Then the TLB flush in ptep_clear_flush ensures that
2779                          * no process can access the old page before the
2780                          * decremented mapcount is visible. And the old page
2781                          * cannot be reused until after the decremented
2782                          * mapcount is visible. So transitively, TLBs to
2783                          * old page will be flushed before it can be reused.
2784                          */
2785                         page_remove_rmap(old_page);
2786                 }
2787
2788                 /* Free the old page.. */
2789                 new_page = old_page;
2790                 ret |= VM_FAULT_WRITE;
2791         } else
2792                 mem_cgroup_uncharge_page(new_page);
2793
2794         if (new_page)
2795                 page_cache_release(new_page);
2796 unlock:
2797         pte_unmap_unlock(page_table, ptl);
2798         if (old_page) {
2799                 /*
2800                  * Don't let another task, with possibly unlocked vma,
2801                  * keep the mlocked page.
2802                  */
2803                 if ((ret & VM_FAULT_WRITE) && (vma->vm_flags & VM_LOCKED)) {
2804                         lock_page(old_page);    /* LRU manipulation */
2805                         munlock_vma_page(old_page);
2806                         unlock_page(old_page);
2807                 }
2808                 page_cache_release(old_page);
2809         }
2810         return ret;
2811 oom_free_new:
2812         page_cache_release(new_page);
2813 oom:
2814         if (old_page) {
2815                 if (page_mkwrite) {
2816                         unlock_page(old_page);
2817                         page_cache_release(old_page);
2818                 }
2819                 page_cache_release(old_page);
2820         }
2821         return VM_FAULT_OOM;
2822
2823 unwritable_page:
2824         page_cache_release(old_page);
2825         return ret;
2826 }
2827
2828 static void unmap_mapping_range_vma(struct vm_area_struct *vma,
2829                 unsigned long start_addr, unsigned long end_addr,
2830                 struct zap_details *details)
2831 {
2832         zap_page_range(vma, start_addr, end_addr - start_addr, details);
2833 }
2834
2835 static inline void unmap_mapping_range_tree(struct prio_tree_root *root,
2836                                             struct zap_details *details)
2837 {
2838         struct vm_area_struct *vma;
2839         struct prio_tree_iter iter;
2840         pgoff_t vba, vea, zba, zea;
2841
2842         vma_prio_tree_foreach(vma, &iter, root,
2843                         details->first_index, details->last_index) {
2844
2845                 vba = vma->vm_pgoff;
2846                 vea = vba + ((vma->vm_end - vma->vm_start) >> PAGE_SHIFT) - 1;
2847                 /* Assume for now that PAGE_CACHE_SHIFT == PAGE_SHIFT */
2848                 zba = details->first_index;
2849                 if (zba < vba)
2850                         zba = vba;
2851                 zea = details->last_index;
2852                 if (zea > vea)
2853                         zea = vea;
2854
2855                 unmap_mapping_range_vma(vma,
2856                         ((zba - vba) << PAGE_SHIFT) + vma->vm_start,
2857                         ((zea - vba + 1) << PAGE_SHIFT) + vma->vm_start,
2858                                 details);
2859         }
2860 }
2861
2862 static inline void unmap_mapping_range_list(struct list_head *head,
2863                                             struct zap_details *details)
2864 {
2865         struct vm_area_struct *vma;
2866
2867         /*
2868          * In nonlinear VMAs there is no correspondence between virtual address
2869          * offset and file offset.  So we must perform an exhaustive search
2870          * across *all* the pages in each nonlinear VMA, not just the pages
2871          * whose virtual address lies outside the file truncation point.
2872          */
2873         list_for_each_entry(vma, head, shared.vm_set.list) {
2874                 details->nonlinear_vma = vma;
2875                 unmap_mapping_range_vma(vma, vma->vm_start, vma->vm_end, details);
2876         }
2877 }
2878
2879 /**
2880  * unmap_mapping_range - unmap the portion of all mmaps in the specified address_space corresponding to the specified page range in the underlying file.
2881  * @mapping: the address space containing mmaps to be unmapped.
2882  * @holebegin: byte in first page to unmap, relative to the start of
2883  * the underlying file.  This will be rounded down to a PAGE_SIZE
2884  * boundary.  Note that this is different from truncate_pagecache(), which
2885  * must keep the partial page.  In contrast, we must get rid of
2886  * partial pages.
2887  * @holelen: size of prospective hole in bytes.  This will be rounded
2888  * up to a PAGE_SIZE boundary.  A holelen of zero truncates to the
2889  * end of the file.
2890  * @even_cows: 1 when truncating a file, unmap even private COWed pages;
2891  * but 0 when invalidating pagecache, don't throw away private data.
2892  */
2893 void unmap_mapping_range(struct address_space *mapping,
2894                 loff_t const holebegin, loff_t const holelen, int even_cows)
2895 {
2896         struct zap_details details;
2897         pgoff_t hba = holebegin >> PAGE_SHIFT;
2898         pgoff_t hlen = (holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
2899
2900         /* Check for overflow. */
2901         if (sizeof(holelen) > sizeof(hlen)) {
2902                 long long holeend =
2903                         (holebegin + holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
2904                 if (holeend & ~(long long)ULONG_MAX)
2905                         hlen = ULONG_MAX - hba + 1;
2906         }
2907
2908         details.check_mapping = even_cows? NULL: mapping;
2909         details.nonlinear_vma = NULL;
2910         details.first_index = hba;
2911         details.last_index = hba + hlen - 1;
2912         if (details.last_index < details.first_index)
2913                 details.last_index = ULONG_MAX;
2914
2915
2916         mutex_lock(&mapping->i_mmap_mutex);
2917         if (unlikely(!prio_tree_empty(&mapping->i_mmap)))
2918                 unmap_mapping_range_tree(&mapping->i_mmap, &details);
2919         if (unlikely(!list_empty(&mapping->i_mmap_nonlinear)))
2920                 unmap_mapping_range_list(&mapping->i_mmap_nonlinear, &details);
2921         mutex_unlock(&mapping->i_mmap_mutex);
2922 }
2923 EXPORT_SYMBOL(unmap_mapping_range);
2924
2925 /*
2926  * We enter with non-exclusive mmap_sem (to exclude vma changes,
2927  * but allow concurrent faults), and pte mapped but not yet locked.
2928  * We return with mmap_sem still held, but pte unmapped and unlocked.
2929  */
2930 static int do_swap_page(struct mm_struct *mm, struct vm_area_struct *vma,
2931                 unsigned long address, pte_t *page_table, pmd_t *pmd,
2932                 unsigned int flags, pte_t orig_pte)
2933 {
2934         spinlock_t *ptl;
2935         struct page *page, *swapcache = NULL;
2936         swp_entry_t entry;
2937         pte_t pte;
2938         int locked;
2939         struct mem_cgroup *ptr;
2940         int exclusive = 0;
2941         int ret = 0;
2942
2943         if (!pte_unmap_same(mm, pmd, page_table, orig_pte))
2944                 goto out;
2945
2946         entry = pte_to_swp_entry(orig_pte);
2947         if (unlikely(non_swap_entry(entry))) {
2948                 if (is_migration_entry(entry)) {
2949                         migration_entry_wait(mm, pmd, address);
2950                 } else if (is_hwpoison_entry(entry)) {
2951                         ret = VM_FAULT_HWPOISON;
2952                 } else {
2953                         print_bad_pte(vma, address, orig_pte, NULL);
2954                         ret = VM_FAULT_SIGBUS;
2955                 }
2956                 goto out;
2957         }
2958         delayacct_set_flag(DELAYACCT_PF_SWAPIN);
2959         page = lookup_swap_cache(entry);
2960         if (!page) {
2961                 grab_swap_token(mm); /* Contend for token _before_ read-in */
2962                 page = swapin_readahead(entry,
2963                                         GFP_HIGHUSER_MOVABLE, vma, address);
2964                 if (!page) {
2965                         /*
2966                          * Back out if somebody else faulted in this pte
2967                          * while we released the pte lock.
2968                          */
2969                         page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2970                         if (likely(pte_same(*page_table, orig_pte)))
2971                                 ret = VM_FAULT_OOM;
2972                         delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
2973                         goto unlock;
2974                 }
2975
2976                 /* Had to read the page from swap area: Major fault */
2977                 ret = VM_FAULT_MAJOR;
2978                 count_vm_event(PGMAJFAULT);
2979                 mem_cgroup_count_vm_event(mm, PGMAJFAULT);
2980         } else if (PageHWPoison(page)) {
2981                 /*
2982                  * hwpoisoned dirty swapcache pages are kept for killing
2983                  * owner processes (which may be unknown at hwpoison time)
2984                  */
2985                 ret = VM_FAULT_HWPOISON;
2986                 delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
2987                 goto out_release;
2988         }
2989
2990         locked = lock_page_or_retry(page, mm, flags);
2991         delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
2992         if (!locked) {
2993                 ret |= VM_FAULT_RETRY;
2994                 goto out_release;
2995         }
2996
2997         /*
2998          * Make sure try_to_free_swap or reuse_swap_page or swapoff did not
2999          * release the swapcache from under us.  The page pin, and pte_same
3000          * test below, are not enough to exclude that.  Even if it is still
3001          * swapcache, we need to check that the page's swap has not changed.
3002          */
3003         if (unlikely(!PageSwapCache(page) || page_private(page) != entry.val))
3004                 goto out_page;
3005
3006         if (ksm_might_need_to_copy(page, vma, address)) {
3007                 swapcache = page;
3008                 page = ksm_does_need_to_copy(page, vma, address);
3009
3010                 if (unlikely(!page)) {
3011                         ret = VM_FAULT_OOM;
3012                         page = swapcache;
3013                         swapcache = NULL;
3014                         goto out_page;
3015                 }
3016         }
3017
3018         if (mem_cgroup_try_charge_swapin(mm, page, GFP_KERNEL, &ptr)) {
3019                 ret = VM_FAULT_OOM;
3020                 goto out_page;
3021         }
3022
3023         /*
3024          * Back out if somebody else already faulted in this pte.
3025          */
3026         page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
3027         if (unlikely(!pte_same(*page_table, orig_pte)))
3028                 goto out_nomap;
3029
3030         if (unlikely(!PageUptodate(page))) {
3031                 ret = VM_FAULT_SIGBUS;
3032                 goto out_nomap;
3033         }
3034
3035         /*
3036          * The page isn't present yet, go ahead with the fault.
3037          *
3038          * Be careful about the sequence of operations here.
3039          * To get its accounting right, reuse_swap_page() must be called
3040          * while the page is counted on swap but not yet in mapcount i.e.
3041          * before page_add_anon_rmap() and swap_free(); try_to_free_swap()
3042          * must be called after the swap_free(), or it will never succeed.
3043          * Because delete_from_swap_page() may be called by reuse_swap_page(),
3044          * mem_cgroup_commit_charge_swapin() may not be able to find swp_entry
3045          * in page->private. In this case, a record in swap_cgroup  is silently
3046          * discarded at swap_free().
3047          */
3048
3049         inc_mm_counter_fast(mm, MM_ANONPAGES);
3050         dec_mm_counter_fast(mm, MM_SWAPENTS);
3051         pte = mk_pte(page, vma->vm_page_prot);
3052         if ((flags & FAULT_FLAG_WRITE) && reuse_swap_page(page)) {
3053                 pte = maybe_mkwrite(pte_mkdirty(pte), vma);
3054                 flags &= ~FAULT_FLAG_WRITE;
3055                 ret |= VM_FAULT_WRITE;
3056                 exclusive = 1;
3057         }
3058         flush_icache_page(vma, page);
3059         set_pte_at(mm, address, page_table, pte);
3060         do_page_add_anon_rmap(page, vma, address, exclusive);
3061         /* It's better to call commit-charge after rmap is established */
3062         mem_cgroup_commit_charge_swapin(page, ptr);
3063
3064         swap_free(entry);
3065         if (vm_swap_full() || (vma->vm_flags & VM_LOCKED) || PageMlocked(page))
3066                 try_to_free_swap(page);
3067         unlock_page(page);
3068         if (swapcache) {
3069                 /*
3070                  * Hold the lock to avoid the swap entry to be reused
3071                  * until we take the PT lock for the pte_same() check
3072                  * (to avoid false positives from pte_same). For
3073                  * further safety release the lock after the swap_free
3074                  * so that the swap count won't change under a
3075                  * parallel locked swapcache.
3076                  */
3077                 unlock_page(swapcache);
3078                 page_cache_release(swapcache);
3079         }
3080
3081         if (flags & FAULT_FLAG_WRITE) {
3082                 ret |= do_wp_page(mm, vma, address, page_table, pmd, ptl, pte);
3083                 if (ret & VM_FAULT_ERROR)
3084                         ret &= VM_FAULT_ERROR;
3085                 goto out;
3086         }
3087
3088         /* No need to invalidate - it was non-present before */
3089         update_mmu_cache(vma, address, page_table);
3090 unlock:
3091         pte_unmap_unlock(page_table, ptl);
3092 out:
3093         return ret;
3094 out_nomap:
3095         mem_cgroup_cancel_charge_swapin(ptr);
3096         pte_unmap_unlock(page_table, ptl);
3097 out_page:
3098         unlock_page(page);
3099 out_release:
3100         page_cache_release(page);
3101         if (swapcache) {
3102                 unlock_page(swapcache);
3103                 page_cache_release(swapcache);
3104         }
3105         return ret;
3106 }
3107
3108 /*
3109  * This is like a special single-page "expand_{down|up}wards()",
3110  * except we must first make sure that 'address{-|+}PAGE_SIZE'
3111  * doesn't hit another vma.
3112  */
3113 static inline int check_stack_guard_page(struct vm_area_struct *vma, unsigned long address)
3114 {
3115         address &= PAGE_MASK;
3116         if ((vma->vm_flags & VM_GROWSDOWN) && address == vma->vm_start) {
3117                 struct vm_area_struct *prev = vma->vm_prev;
3118
3119                 /*
3120                  * Is there a mapping abutting this one below?
3121                  *
3122                  * That's only ok if it's the same stack mapping
3123                  * that has gotten split..
3124                  */
3125                 if (prev && prev->vm_end == address)
3126                         return prev->vm_flags & VM_GROWSDOWN ? 0 : -ENOMEM;
3127
3128                 return expand_downwards(vma, address - PAGE_SIZE);
3129         }
3130         if ((vma->vm_flags & VM_GROWSUP) && address + PAGE_SIZE == vma->vm_end) {
3131                 struct vm_area_struct *next = vma->vm_next;
3132
3133                 /* As VM_GROWSDOWN but s/below/above/ */
3134                 if (next && next->vm_start == address + PAGE_SIZE)
3135                         return next->vm_flags & VM_GROWSUP ? 0 : -ENOMEM;
3136
3137                 return expand_upwards(vma, address + PAGE_SIZE);
3138         }
3139         return 0;
3140 }
3141
3142 /*
3143  * We enter with non-exclusive mmap_sem (to exclude vma changes,
3144  * but allow concurrent faults), and pte mapped but not yet locked.
3145  * We return with mmap_sem still held, but pte unmapped and unlocked.
3146  */
3147 static int do_anonymous_page(struct mm_struct *mm, struct vm_area_struct *vma,
3148                 unsigned long address, pte_t *page_table, pmd_t *pmd,
3149                 unsigned int flags)
3150 {
3151         struct page *page;
3152         spinlock_t *ptl;
3153         pte_t entry;
3154
3155         pte_unmap(page_table);
3156
3157         /* Check if we need to add a guard page to the stack */
3158         if (check_stack_guard_page(vma, address) < 0)
3159                 return VM_FAULT_SIGSEGV;
3160
3161         /* Use the zero-page for reads */
3162         if (!(flags & FAULT_FLAG_WRITE)) {
3163                 entry = pte_mkspecial(pfn_pte(my_zero_pfn(address),
3164                                                 vma->vm_page_prot));
3165                 page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
3166                 if (!pte_none(*page_table))
3167                         goto unlock;
3168                 goto setpte;
3169         }
3170
3171         /* Allocate our own private page. */
3172         if (unlikely(anon_vma_prepare(vma)))
3173                 goto oom;
3174         page = alloc_zeroed_user_highpage_movable(vma, address);
3175         if (!page)
3176                 goto oom;
3177         __SetPageUptodate(page);
3178
3179         if (mem_cgroup_newpage_charge(page, mm, GFP_KERNEL))
3180                 goto oom_free_page;
3181
3182         entry = mk_pte(page, vma->vm_page_prot);
3183         if (vma->vm_flags & VM_WRITE)
3184                 entry = pte_mkwrite(pte_mkdirty(entry));
3185
3186         page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
3187         if (!pte_none(*page_table))
3188                 goto release;
3189
3190         inc_mm_counter_fast(mm, MM_ANONPAGES);
3191         page_add_new_anon_rmap(page, vma, address);
3192 setpte:
3193         set_pte_at(mm, address, page_table, entry);
3194
3195         /* No need to invalidate - it was non-present before */
3196         update_mmu_cache(vma, address, page_table);
3197 unlock:
3198         pte_unmap_unlock(page_table, ptl);
3199         return 0;
3200 release:
3201         mem_cgroup_uncharge_page(page);
3202         page_cache_release(page);
3203         goto unlock;
3204 oom_free_page:
3205         page_cache_release(page);
3206 oom:
3207         return VM_FAULT_OOM;
3208 }
3209
3210 /*
3211  * __do_fault() tries to create a new page mapping. It aggressively
3212  * tries to share with existing pages, but makes a separate copy if
3213  * the FAULT_FLAG_WRITE is set in the flags parameter in order to avoid
3214  * the next page fault.
3215  *
3216  * As this is called only for pages that do not currently exist, we
3217  * do not need to flush old virtual caches or the TLB.
3218  *
3219  * We enter with non-exclusive mmap_sem (to exclude vma changes,
3220  * but allow concurrent faults), and pte neither mapped nor locked.
3221  * We return with mmap_sem still held, but pte unmapped and unlocked.
3222  */
3223 static int __do_fault(struct mm_struct *mm, struct vm_area_struct *vma,
3224                 unsigned long address, pmd_t *pmd,
3225                 pgoff_t pgoff, unsigned int flags, pte_t orig_pte)
3226 {
3227         pte_t *page_table;
3228         spinlock_t *ptl;
3229         struct page *page;
3230         struct page *cow_page;
3231         pte_t entry;
3232         int anon = 0;
3233         struct page *dirty_page = NULL;
3234         struct vm_fault vmf;
3235         int ret;
3236         int page_mkwrite = 0;
3237
3238         /*
3239          * If we do COW later, allocate page befor taking lock_page()
3240          * on the file cache page. This will reduce lock holding time.
3241          */
3242         if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) {
3243
3244                 if (unlikely(anon_vma_prepare(vma)))
3245                         return VM_FAULT_OOM;
3246
3247                 cow_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, address);
3248                 if (!cow_page)
3249                         return VM_FAULT_OOM;
3250
3251                 if (mem_cgroup_newpage_charge(cow_page, mm, GFP_KERNEL)) {
3252                         page_cache_release(cow_page);
3253                         return VM_FAULT_OOM;
3254                 }
3255         } else
3256                 cow_page = NULL;
3257
3258         vmf.virtual_address = (void __user *)(address & PAGE_MASK);
3259         vmf.pgoff = pgoff;
3260         vmf.flags = flags;
3261         vmf.page = NULL;
3262
3263         ret = vma->vm_ops->fault(vma, &vmf);
3264         if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE |
3265                             VM_FAULT_RETRY)))
3266                 goto uncharge_out;
3267
3268         if (unlikely(PageHWPoison(vmf.page))) {
3269                 if (ret & VM_FAULT_LOCKED)
3270                         unlock_page(vmf.page);
3271                 ret = VM_FAULT_HWPOISON;
3272                 goto uncharge_out;
3273         }
3274
3275         /*
3276          * For consistency in subsequent calls, make the faulted page always
3277          * locked.
3278          */
3279         if (unlikely(!(ret & VM_FAULT_LOCKED)))
3280                 lock_page(vmf.page);
3281         else
3282                 VM_BUG_ON(!PageLocked(vmf.page));
3283
3284         /*
3285          * Should we do an early C-O-W break?
3286          */
3287         page = vmf.page;
3288         if (flags & FAULT_FLAG_WRITE) {
3289                 if (!(vma->vm_flags & VM_SHARED)) {
3290                         page = cow_page;
3291                         anon = 1;
3292                         copy_user_highpage(page, vmf.page, address, vma);
3293                         __SetPageUptodate(page);
3294                 } else {
3295                         /*
3296                          * If the page will be shareable, see if the backing
3297                          * address space wants to know that the page is about
3298                          * to become writable
3299                          */
3300                         if (vma->vm_ops->page_mkwrite) {
3301                                 int tmp;
3302
3303                                 unlock_page(page);
3304                                 vmf.flags = FAULT_FLAG_WRITE|FAULT_FLAG_MKWRITE;
3305                                 tmp = vma->vm_ops->page_mkwrite(vma, &vmf);
3306                                 if (unlikely(tmp &
3307                                           (VM_FAULT_ERROR | VM_FAULT_NOPAGE))) {
3308                                         ret = tmp;
3309                                         goto unwritable_page;
3310                                 }
3311                                 if (unlikely(!(tmp & VM_FAULT_LOCKED))) {
3312                                         lock_page(page);
3313                                         if (!page->mapping) {
3314                                                 ret = 0; /* retry the fault */
3315                                                 unlock_page(page);
3316                                                 goto unwritable_page;
3317                                         }
3318                                 } else
3319                                         VM_BUG_ON(!PageLocked(page));
3320                                 page_mkwrite = 1;
3321                         }
3322                 }
3323
3324         }
3325
3326         page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
3327
3328         /*
3329          * This silly early PAGE_DIRTY setting removes a race
3330          * due to the bad i386 page protection. But it's valid
3331          * for other architectures too.
3332          *
3333          * Note that if FAULT_FLAG_WRITE is set, we either now have
3334          * an exclusive copy of the page, or this is a shared mapping,
3335          * so we can make it writable and dirty to avoid having to
3336          * handle that later.
3337          */
3338         /* Only go through if we didn't race with anybody else... */
3339         if (likely(pte_same(*page_table, orig_pte))) {
3340                 flush_icache_page(vma, page);
3341                 entry = mk_pte(page, vma->vm_page_prot);
3342                 if (flags & FAULT_FLAG_WRITE)
3343                         entry = maybe_mkwrite(pte_mkdirty(entry), vma);
3344                 if (anon) {
3345                         inc_mm_counter_fast(mm, MM_ANONPAGES);
3346                         page_add_new_anon_rmap(page, vma, address);
3347                 } else {
3348                         inc_mm_counter_fast(mm, MM_FILEPAGES);
3349                         page_add_file_rmap(page);
3350                         if (flags & FAULT_FLAG_WRITE) {
3351                                 dirty_page = page;
3352                                 get_page(dirty_page);
3353                         }
3354                 }
3355                 set_pte_at(mm, address, page_table, entry);
3356
3357                 /* no need to invalidate: a not-present page won't be cached */
3358                 update_mmu_cache(vma, address, page_table);
3359         } else {
3360                 if (cow_page)
3361                         mem_cgroup_uncharge_page(cow_page);
3362                 if (anon)
3363                         page_cache_release(page);
3364                 else
3365                         anon = 1; /* no anon but release faulted_page */
3366         }
3367
3368         pte_unmap_unlock(page_table, ptl);
3369
3370         if (dirty_page) {
3371                 struct address_space *mapping = page->mapping;
3372
3373                 if (set_page_dirty(dirty_page))
3374                         page_mkwrite = 1;
3375                 unlock_page(dirty_page);
3376                 put_page(dirty_page);
3377                 if (page_mkwrite && mapping) {
3378                         /*
3379                          * Some device drivers do not set page.mapping but still
3380                          * dirty their pages
3381                          */
3382                         balance_dirty_pages_ratelimited(mapping);
3383                 }
3384
3385                 /* file_update_time outside page_lock */
3386                 if (vma->vm_file)
3387                         file_update_time(vma->vm_file);
3388         } else {
3389                 unlock_page(vmf.page);
3390                 if (anon)
3391                         page_cache_release(vmf.page);
3392         }
3393
3394         return ret;
3395
3396 unwritable_page:
3397         page_cache_release(page);
3398         return ret;
3399 uncharge_out:
3400         /* fs's fault handler get error */
3401         if (cow_page) {
3402                 mem_cgroup_uncharge_page(cow_page);
3403                 page_cache_release(cow_page);
3404         }
3405         return ret;
3406 }
3407
3408 static int do_linear_fault(struct mm_struct *mm, struct vm_area_struct *vma,
3409                 unsigned long address, pte_t *page_table, pmd_t *pmd,
3410                 unsigned int flags, pte_t orig_pte)
3411 {
3412         pgoff_t pgoff = (((address & PAGE_MASK)
3413                         - vma->vm_start) >> PAGE_SHIFT) + vma->vm_pgoff;
3414
3415         pte_unmap(page_table);
3416         return __do_fault(mm, vma, address, pmd, pgoff, flags, orig_pte);
3417 }
3418
3419 /*
3420  * Fault of a previously existing named mapping. Repopulate the pte
3421  * from the encoded file_pte if possible. This enables swappable
3422  * nonlinear vmas.
3423  *
3424  * We enter with non-exclusive mmap_sem (to exclude vma changes,
3425  * but allow concurrent faults), and pte mapped but not yet locked.
3426  * We return with mmap_sem still held, but pte unmapped and unlocked.
3427  */
3428 static int do_nonlinear_fault(struct mm_struct *mm, struct vm_area_struct *vma,
3429                 unsigned long address, pte_t *page_table, pmd_t *pmd,
3430                 unsigned int flags, pte_t orig_pte)
3431 {
3432         pgoff_t pgoff;
3433
3434         flags |= FAULT_FLAG_NONLINEAR;
3435
3436         if (!pte_unmap_same(mm, pmd, page_table, orig_pte))
3437                 return 0;
3438
3439         if (unlikely(!(vma->vm_flags & VM_NONLINEAR))) {
3440                 /*
3441                  * Page table corrupted: show pte and kill process.
3442                  */
3443                 print_bad_pte(vma, address, orig_pte, NULL);
3444                 return VM_FAULT_SIGBUS;
3445         }
3446
3447         pgoff = pte_to_pgoff(orig_pte);
3448         return __do_fault(mm, vma, address, pmd, pgoff, flags, orig_pte);
3449 }
3450
3451 /*
3452  * These routines also need to handle stuff like marking pages dirty
3453  * and/or accessed for architectures that don't do it in hardware (most
3454  * RISC architectures).  The early dirtying is also good on the i386.
3455  *
3456  * There is also a hook called "update_mmu_cache()" that architectures
3457  * with external mmu caches can use to update those (ie the Sparc or
3458  * PowerPC hashed page tables that act as extended TLBs).
3459  *
3460  * We enter with non-exclusive mmap_sem (to exclude vma changes,
3461  * but allow concurrent faults), and pte mapped but not yet locked.
3462  * We return with mmap_sem still held, but pte unmapped and unlocked.
3463  */
3464 int handle_pte_fault(struct mm_struct *mm,
3465                      struct vm_area_struct *vma, unsigned long address,
3466                      pte_t *pte, pmd_t *pmd, unsigned int flags)
3467 {
3468         pte_t entry;
3469         spinlock_t *ptl;
3470
3471         entry = *pte;
3472         if (!pte_present(entry)) {
3473                 if (pte_none(entry)) {
3474                         if (vma->vm_ops) {
3475                                 if (likely(vma->vm_ops->fault))
3476                                         return do_linear_fault(mm, vma, address,
3477                                                 pte, pmd, flags, entry);
3478                         }
3479                         return do_anonymous_page(mm, vma, address,
3480                                                  pte, pmd, flags);
3481                 }
3482                 if (pte_file(entry))
3483                         return do_nonlinear_fault(mm, vma, address,
3484                                         pte, pmd, flags, entry);
3485                 return do_swap_page(mm, vma, address,
3486                                         pte, pmd, flags, entry);
3487         }
3488
3489         ptl = pte_lockptr(mm, pmd);
3490         spin_lock(ptl);
3491         if (unlikely(!pte_same(*pte, entry)))
3492                 goto unlock;
3493         if (flags & FAULT_FLAG_WRITE) {
3494                 if (!pte_write(entry))
3495                         return do_wp_page(mm, vma, address,
3496                                         pte, pmd, ptl, entry);
3497                 entry = pte_mkdirty(entry);
3498         }
3499         entry = pte_mkyoung(entry);
3500         if (ptep_set_access_flags(vma, address, pte, entry, flags & FAULT_FLAG_WRITE)) {
3501                 update_mmu_cache(vma, address, pte);
3502         } else {
3503                 /*
3504                  * This is needed only for protection faults but the arch code
3505                  * is not yet telling us if this is a protection fault or not.
3506                  * This still avoids useless tlb flushes for .text page faults
3507                  * with threads.
3508                  */
3509                 if (flags & FAULT_FLAG_WRITE)
3510                         flush_tlb_fix_spurious_fault(vma, address);
3511         }
3512 unlock:
3513         pte_unmap_unlock(pte, ptl);
3514         return 0;
3515 }
3516
3517 /*
3518  * By the time we get here, we already hold the mm semaphore
3519  */
3520 int handle_mm_fault(struct mm_struct *mm, struct vm_area_struct *vma,
3521                 unsigned long address, unsigned int flags)
3522 {
3523         pgd_t *pgd;
3524         pud_t *pud;
3525         pmd_t *pmd;
3526         pte_t *pte;
3527
3528         __set_current_state(TASK_RUNNING);
3529
3530         count_vm_event(PGFAULT);
3531         mem_cgroup_count_vm_event(mm, PGFAULT);
3532
3533         /* do counter updates before entering really critical section. */
3534         check_sync_rss_stat(current);
3535
3536         if (unlikely(is_vm_hugetlb_page(vma)))
3537                 return hugetlb_fault(mm, vma, address, flags);
3538
3539 retry:
3540         pgd = pgd_offset(mm, address);
3541         pud = pud_alloc(mm, pgd, address);
3542         if (!pud)
3543                 return VM_FAULT_OOM;
3544         pmd = pmd_alloc(mm, pud, address);
3545         if (!pmd)
3546                 return VM_FAULT_OOM;
3547         if (pmd_none(*pmd) && transparent_hugepage_enabled(vma)) {
3548                 if (!vma->vm_ops)
3549                         return do_huge_pmd_anonymous_page(mm, vma, address,
3550                                                           pmd, flags);
3551         } else {
3552                 pmd_t orig_pmd = *pmd;
3553                 int ret;
3554
3555                 barrier();
3556                 if (pmd_trans_huge(orig_pmd)) {
3557                         unsigned int dirty = flags & FAULT_FLAG_WRITE;
3558
3559                         if (dirty && !pmd_write(orig_pmd) &&
3560                             !pmd_trans_splitting(orig_pmd)) {
3561                                 ret = do_huge_pmd_wp_page(mm, vma, address, pmd,
3562                                                           orig_pmd);
3563                                 /*
3564                                  * If COW results in an oom, the huge pmd will
3565                                  * have been split, so retry the fault on the
3566                                  * pte for a smaller charge.
3567                                  */
3568                                 if (unlikely(ret & VM_FAULT_OOM))
3569                                         goto retry;
3570                                 return ret;
3571                         } else {
3572                                 huge_pmd_set_accessed(mm, vma, address, pmd,
3573                                                       orig_pmd, dirty);
3574                         }
3575                         return 0;
3576                 }
3577         }
3578
3579         /*
3580          * Use __pte_alloc instead of pte_alloc_map, because we can't
3581          * run pte_offset_map on the pmd, if an huge pmd could
3582          * materialize from under us from a different thread.
3583          */
3584         if (unlikely(pmd_none(*pmd)) && __pte_alloc(mm, vma, pmd, address))
3585                 return VM_FAULT_OOM;
3586         /* if an huge pmd materialized from under us just retry later */
3587         if (unlikely(pmd_trans_huge(*pmd)))
3588                 return 0;
3589         /*
3590          * A regular pmd is established and it can't morph into a huge pmd
3591          * from under us anymore at this point because we hold the mmap_sem
3592          * read mode and khugepaged takes it in write mode. So now it's
3593          * safe to run pte_offset_map().
3594          */
3595         pte = pte_offset_map(pmd, address);
3596
3597         return handle_pte_fault(mm, vma, address, pte, pmd, flags);
3598 }
3599
3600 #ifndef __PAGETABLE_PUD_FOLDED
3601 /*
3602  * Allocate page upper directory.
3603  * We've already handled the fast-path in-line.
3604  */
3605 int __pud_alloc(struct mm_struct *mm, pgd_t *pgd, unsigned long address)
3606 {
3607         pud_t *new = pud_alloc_one(mm, address);
3608         if (!new)
3609                 return -ENOMEM;
3610
3611         smp_wmb(); /* See comment in __pte_alloc */
3612
3613         spin_lock(&mm->page_table_lock);
3614         if (pgd_present(*pgd))          /* Another has populated it */
3615                 pud_free(mm, new);
3616         else
3617                 pgd_populate(mm, pgd, new);
3618         spin_unlock(&mm->page_table_lock);
3619         return 0;
3620 }
3621 #endif /* __PAGETABLE_PUD_FOLDED */
3622
3623 #ifndef __PAGETABLE_PMD_FOLDED
3624 /*
3625  * Allocate page middle directory.
3626  * We've already handled the fast-path in-line.
3627  */
3628 int __pmd_alloc(struct mm_struct *mm, pud_t *pud, unsigned long address)
3629 {
3630         pmd_t *new = pmd_alloc_one(mm, address);
3631         if (!new)
3632                 return -ENOMEM;
3633
3634         smp_wmb(); /* See comment in __pte_alloc */
3635
3636         spin_lock(&mm->page_table_lock);
3637 #ifndef __ARCH_HAS_4LEVEL_HACK
3638         if (pud_present(*pud))          /* Another has populated it */
3639                 pmd_free(mm, new);
3640         else
3641                 pud_populate(mm, pud, new);
3642 #else
3643         if (pgd_present(*pud))          /* Another has populated it */
3644                 pmd_free(mm, new);
3645         else
3646                 pgd_populate(mm, pud, new);
3647 #endif /* __ARCH_HAS_4LEVEL_HACK */
3648         spin_unlock(&mm->page_table_lock);
3649         return 0;
3650 }
3651 #endif /* __PAGETABLE_PMD_FOLDED */
3652
3653 int make_pages_present(unsigned long addr, unsigned long end)
3654 {
3655         int ret, len, write;
3656         struct vm_area_struct * vma;
3657
3658         vma = find_vma(current->mm, addr);
3659         if (!vma)
3660                 return -ENOMEM;
3661         /*
3662          * We want to touch writable mappings with a write fault in order
3663          * to break COW, except for shared mappings because these don't COW
3664          * and we wou