iommu/amd: Free domain id when free a domain of struct dma_ops_domain
[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
1407 /**
1408  * zap_vma_ptes - remove ptes mapping the vma
1409  * @vma: vm_area_struct holding ptes to be zapped
1410  * @address: starting address of pages to zap
1411  * @size: number of bytes to zap
1412  *
1413  * This function only unmaps ptes assigned to VM_PFNMAP vmas.
1414  *
1415  * The entire address range must be fully contained within the vma.
1416  *
1417  * Returns 0 if successful.
1418  */
1419 int zap_vma_ptes(struct vm_area_struct *vma, unsigned long address,
1420                 unsigned long size)
1421 {
1422         if (address < vma->vm_start || address + size > vma->vm_end ||
1423                         !(vma->vm_flags & VM_PFNMAP))
1424                 return -1;
1425         zap_page_range(vma, address, size, NULL);
1426         return 0;
1427 }
1428 EXPORT_SYMBOL_GPL(zap_vma_ptes);
1429
1430 static inline bool can_follow_write_pte(pte_t pte, struct page *page,
1431                                         unsigned int flags)
1432 {
1433         if (pte_write(pte))
1434                 return true;
1435
1436         /*
1437          * Make sure that we are really following CoWed page. We do not really
1438          * have to care about exclusiveness of the page because we only want
1439          * to ensure that once COWed page hasn't disappeared in the meantime
1440          * or it hasn't been merged to a KSM page.
1441          */
1442         if ((flags & FOLL_FORCE) && (flags & FOLL_COW))
1443                 return page && PageAnon(page) && !PageKsm(page);
1444
1445         return false;
1446 }
1447
1448 /**
1449  * follow_page - look up a page descriptor from a user-virtual address
1450  * @vma: vm_area_struct mapping @address
1451  * @address: virtual address to look up
1452  * @flags: flags modifying lookup behaviour
1453  *
1454  * @flags can have FOLL_ flags set, defined in <linux/mm.h>
1455  *
1456  * Returns the mapped (struct page *), %NULL if no mapping exists, or
1457  * an error pointer if there is a mapping to something not represented
1458  * by a page descriptor (see also vm_normal_page()).
1459  */
1460 struct page *follow_page(struct vm_area_struct *vma, unsigned long address,
1461                         unsigned int flags)
1462 {
1463         pgd_t *pgd;
1464         pud_t *pud;
1465         pmd_t *pmd;
1466         pte_t *ptep, pte;
1467         spinlock_t *ptl;
1468         struct page *page;
1469         struct mm_struct *mm = vma->vm_mm;
1470
1471         page = follow_huge_addr(mm, address, flags & FOLL_WRITE);
1472         if (!IS_ERR(page)) {
1473                 BUG_ON(flags & FOLL_GET);
1474                 goto out;
1475         }
1476
1477         page = NULL;
1478         pgd = pgd_offset(mm, address);
1479         if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
1480                 goto no_page_table;
1481
1482         pud = pud_offset(pgd, address);
1483         if (pud_none(*pud))
1484                 goto no_page_table;
1485         if (pud_huge(*pud) && vma->vm_flags & VM_HUGETLB) {
1486                 BUG_ON(flags & FOLL_GET);
1487                 page = follow_huge_pud(mm, address, pud, flags & FOLL_WRITE);
1488                 goto out;
1489         }
1490         if (unlikely(pud_bad(*pud)))
1491                 goto no_page_table;
1492
1493         pmd = pmd_offset(pud, address);
1494         if (pmd_none(*pmd))
1495                 goto no_page_table;
1496         if (pmd_huge(*pmd) && vma->vm_flags & VM_HUGETLB) {
1497                 BUG_ON(flags & FOLL_GET);
1498                 page = follow_huge_pmd(mm, address, pmd, flags & FOLL_WRITE);
1499                 goto out;
1500         }
1501         if (pmd_trans_huge(*pmd)) {
1502                 if (flags & FOLL_SPLIT) {
1503                         split_huge_page_pmd(mm, pmd);
1504                         goto split_fallthrough;
1505                 }
1506                 spin_lock(&mm->page_table_lock);
1507                 if (likely(pmd_trans_huge(*pmd))) {
1508                         if (unlikely(pmd_trans_splitting(*pmd))) {
1509                                 spin_unlock(&mm->page_table_lock);
1510                                 wait_split_huge_page(vma->anon_vma, pmd);
1511                         } else {
1512                                 page = follow_trans_huge_pmd(mm, address,
1513                                                              pmd, flags);
1514                                 spin_unlock(&mm->page_table_lock);
1515                                 goto out;
1516                         }
1517                 } else
1518                         spin_unlock(&mm->page_table_lock);
1519                 /* fall through */
1520         }
1521 split_fallthrough:
1522         if (unlikely(pmd_bad(*pmd)))
1523                 goto no_page_table;
1524
1525         ptep = pte_offset_map_lock(mm, pmd, address, &ptl);
1526
1527         pte = *ptep;
1528         if (!pte_present(pte))
1529                 goto no_page;
1530
1531         page = vm_normal_page(vma, address, pte);
1532         if ((flags & FOLL_WRITE) && !can_follow_write_pte(pte, page, flags)) {
1533                 pte_unmap_unlock(ptep, ptl);
1534                 return NULL;
1535         }
1536
1537         if (unlikely(!page)) {
1538                 if ((flags & FOLL_DUMP) ||
1539                     !is_zero_pfn(pte_pfn(pte)))
1540                         goto bad_page;
1541                 page = pte_page(pte);
1542         }
1543
1544         if (flags & FOLL_GET)
1545                 get_page_foll(page);
1546         if (flags & FOLL_TOUCH) {
1547                 if ((flags & FOLL_WRITE) &&
1548                     !pte_dirty(pte) && !PageDirty(page))
1549                         set_page_dirty(page);
1550                 /*
1551                  * pte_mkyoung() would be more correct here, but atomic care
1552                  * is needed to avoid losing the dirty bit: it is easier to use
1553                  * mark_page_accessed().
1554                  */
1555                 mark_page_accessed(page);
1556         }
1557         if ((flags & FOLL_MLOCK) && (vma->vm_flags & VM_LOCKED)) {
1558                 /*
1559                  * The preliminary mapping check is mainly to avoid the
1560                  * pointless overhead of lock_page on the ZERO_PAGE
1561                  * which might bounce very badly if there is contention.
1562                  *
1563                  * If the page is already locked, we don't need to
1564                  * handle it now - vmscan will handle it later if and
1565                  * when it attempts to reclaim the page.
1566                  */
1567                 if (page->mapping && trylock_page(page)) {
1568                         lru_add_drain();  /* push cached pages to LRU */
1569                         /*
1570                          * Because we lock page here and migration is
1571                          * blocked by the pte's page reference, we need
1572                          * only check for file-cache page truncation.
1573                          */
1574                         if (page->mapping)
1575                                 mlock_vma_page(page);
1576                         unlock_page(page);
1577                 }
1578         }
1579
1580         pte_unmap_unlock(ptep, ptl);
1581 out:
1582         return page;
1583
1584 bad_page:
1585         pte_unmap_unlock(ptep, ptl);
1586         return ERR_PTR(-EFAULT);
1587
1588 no_page:
1589         pte_unmap_unlock(ptep, ptl);
1590         if (!pte_none(pte))
1591                 return page;
1592
1593 no_page_table:
1594         /*
1595          * When core dumping an enormous anonymous area that nobody
1596          * has touched so far, we don't want to allocate unnecessary pages or
1597          * page tables.  Return error instead of NULL to skip handle_mm_fault,
1598          * then get_dump_page() will return NULL to leave a hole in the dump.
1599          * But we can only make this optimization where a hole would surely
1600          * be zero-filled if handle_mm_fault() actually did handle it.
1601          */
1602         if ((flags & FOLL_DUMP) &&
1603             (!vma->vm_ops || !vma->vm_ops->fault))
1604                 return ERR_PTR(-EFAULT);
1605         return page;
1606 }
1607
1608 static inline int stack_guard_page(struct vm_area_struct *vma, unsigned long addr)
1609 {
1610         return stack_guard_page_start(vma, addr) ||
1611                stack_guard_page_end(vma, addr+PAGE_SIZE);
1612 }
1613
1614 /**
1615  * __get_user_pages() - pin user pages in memory
1616  * @tsk:        task_struct of target task
1617  * @mm:         mm_struct of target mm
1618  * @start:      starting user address
1619  * @nr_pages:   number of pages from start to pin
1620  * @gup_flags:  flags modifying pin behaviour
1621  * @pages:      array that receives pointers to the pages pinned.
1622  *              Should be at least nr_pages long. Or NULL, if caller
1623  *              only intends to ensure the pages are faulted in.
1624  * @vmas:       array of pointers to vmas corresponding to each page.
1625  *              Or NULL if the caller does not require them.
1626  * @nonblocking: whether waiting for disk IO or mmap_sem contention
1627  *
1628  * Returns number of pages pinned. This may be fewer than the number
1629  * requested. If nr_pages is 0 or negative, returns 0. If no pages
1630  * were pinned, returns -errno. Each page returned must be released
1631  * with a put_page() call when it is finished with. vmas will only
1632  * remain valid while mmap_sem is held.
1633  *
1634  * Must be called with mmap_sem held for read or write.
1635  *
1636  * __get_user_pages walks a process's page tables and takes a reference to
1637  * each struct page that each user address corresponds to at a given
1638  * instant. That is, it takes the page that would be accessed if a user
1639  * thread accesses the given user virtual address at that instant.
1640  *
1641  * This does not guarantee that the page exists in the user mappings when
1642  * __get_user_pages returns, and there may even be a completely different
1643  * page there in some cases (eg. if mmapped pagecache has been invalidated
1644  * and subsequently re faulted). However it does guarantee that the page
1645  * won't be freed completely. And mostly callers simply care that the page
1646  * contains data that was valid *at some point in time*. Typically, an IO
1647  * or similar operation cannot guarantee anything stronger anyway because
1648  * locks can't be held over the syscall boundary.
1649  *
1650  * If @gup_flags & FOLL_WRITE == 0, the page must not be written to. If
1651  * the page is written to, set_page_dirty (or set_page_dirty_lock, as
1652  * appropriate) must be called after the page is finished with, and
1653  * before put_page is called.
1654  *
1655  * If @nonblocking != NULL, __get_user_pages will not wait for disk IO
1656  * or mmap_sem contention, and if waiting is needed to pin all pages,
1657  * *@nonblocking will be set to 0.
1658  *
1659  * In most cases, get_user_pages or get_user_pages_fast should be used
1660  * instead of __get_user_pages. __get_user_pages should be used only if
1661  * you need some special @gup_flags.
1662  */
1663 int __get_user_pages(struct task_struct *tsk, struct mm_struct *mm,
1664                      unsigned long start, int nr_pages, unsigned int gup_flags,
1665                      struct page **pages, struct vm_area_struct **vmas,
1666                      int *nonblocking)
1667 {
1668         int i;
1669         unsigned long vm_flags;
1670
1671         if (nr_pages <= 0)
1672                 return 0;
1673
1674         VM_BUG_ON(!!pages != !!(gup_flags & FOLL_GET));
1675
1676         /* 
1677          * Require read or write permissions.
1678          * If FOLL_FORCE is set, we only require the "MAY" flags.
1679          */
1680         vm_flags  = (gup_flags & FOLL_WRITE) ?
1681                         (VM_WRITE | VM_MAYWRITE) : (VM_READ | VM_MAYREAD);
1682         vm_flags &= (gup_flags & FOLL_FORCE) ?
1683                         (VM_MAYREAD | VM_MAYWRITE) : (VM_READ | VM_WRITE);
1684         i = 0;
1685
1686         do {
1687                 struct vm_area_struct *vma;
1688
1689                 vma = find_extend_vma(mm, start);
1690                 if (!vma && in_gate_area(mm, start)) {
1691                         unsigned long pg = start & PAGE_MASK;
1692                         pgd_t *pgd;
1693                         pud_t *pud;
1694                         pmd_t *pmd;
1695                         pte_t *pte;
1696
1697                         /* user gate pages are read-only */
1698                         if (gup_flags & FOLL_WRITE)
1699                                 return i ? : -EFAULT;
1700                         if (pg > TASK_SIZE)
1701                                 pgd = pgd_offset_k(pg);
1702                         else
1703                                 pgd = pgd_offset_gate(mm, pg);
1704                         BUG_ON(pgd_none(*pgd));
1705                         pud = pud_offset(pgd, pg);
1706                         BUG_ON(pud_none(*pud));
1707                         pmd = pmd_offset(pud, pg);
1708                         if (pmd_none(*pmd))
1709                                 return i ? : -EFAULT;
1710                         VM_BUG_ON(pmd_trans_huge(*pmd));
1711                         pte = pte_offset_map(pmd, pg);
1712                         if (pte_none(*pte)) {
1713                                 pte_unmap(pte);
1714                                 return i ? : -EFAULT;
1715                         }
1716                         vma = get_gate_vma(mm);
1717                         if (pages) {
1718                                 struct page *page;
1719
1720                                 page = vm_normal_page(vma, start, *pte);
1721                                 if (!page) {
1722                                         if (!(gup_flags & FOLL_DUMP) &&
1723                                              is_zero_pfn(pte_pfn(*pte)))
1724                                                 page = pte_page(*pte);
1725                                         else {
1726                                                 pte_unmap(pte);
1727                                                 return i ? : -EFAULT;
1728                                         }
1729                                 }
1730                                 pages[i] = page;
1731                                 get_page(page);
1732                         }
1733                         pte_unmap(pte);
1734                         goto next_page;
1735                 }
1736
1737                 if (!vma ||
1738                     (vma->vm_flags & (VM_IO | VM_PFNMAP)) ||
1739                     !(vm_flags & vma->vm_flags))
1740                         return i ? : -EFAULT;
1741
1742                 if (is_vm_hugetlb_page(vma)) {
1743                         i = follow_hugetlb_page(mm, vma, pages, vmas,
1744                                         &start, &nr_pages, i, gup_flags);
1745                         continue;
1746                 }
1747
1748                 do {
1749                         struct page *page;
1750                         unsigned int foll_flags = gup_flags;
1751
1752                         /*
1753                          * If we have a pending SIGKILL, don't keep faulting
1754                          * pages and potentially allocating memory.
1755                          */
1756                         if (unlikely(fatal_signal_pending(current)))
1757                                 return i ? i : -ERESTARTSYS;
1758
1759                         cond_resched();
1760                         while (!(page = follow_page(vma, start, foll_flags))) {
1761                                 int ret;
1762                                 unsigned int fault_flags = 0;
1763
1764                                 /* For mlock, just skip the stack guard page. */
1765                                 if (foll_flags & FOLL_MLOCK) {
1766                                         if (stack_guard_page(vma, start))
1767                                                 goto next_page;
1768                                 }
1769                                 if (foll_flags & FOLL_WRITE)
1770                                         fault_flags |= FAULT_FLAG_WRITE;
1771                                 if (nonblocking)
1772                                         fault_flags |= FAULT_FLAG_ALLOW_RETRY;
1773                                 if (foll_flags & FOLL_NOWAIT)
1774                                         fault_flags |= (FAULT_FLAG_ALLOW_RETRY | FAULT_FLAG_RETRY_NOWAIT);
1775
1776                                 ret = handle_mm_fault(mm, vma, start,
1777                                                         fault_flags);
1778
1779                                 if (ret & VM_FAULT_ERROR) {
1780                                         if (ret & VM_FAULT_OOM)
1781                                                 return i ? i : -ENOMEM;
1782                                         if (ret & (VM_FAULT_HWPOISON |
1783                                                    VM_FAULT_HWPOISON_LARGE)) {
1784                                                 if (i)
1785                                                         return i;
1786                                                 else if (gup_flags & FOLL_HWPOISON)
1787                                                         return -EHWPOISON;
1788                                                 else
1789                                                         return -EFAULT;
1790                                         }
1791                                         if (ret & (VM_FAULT_SIGBUS | VM_FAULT_SIGSEGV))
1792                                                 return i ? i : -EFAULT;
1793                                         BUG();
1794                                 }
1795
1796                                 if (tsk) {
1797                                         if (ret & VM_FAULT_MAJOR)
1798                                                 tsk->maj_flt++;
1799                                         else
1800                                                 tsk->min_flt++;
1801                                 }
1802
1803                                 if (ret & VM_FAULT_RETRY) {
1804                                         if (nonblocking)
1805                                                 *nonblocking = 0;
1806                                         return i;
1807                                 }
1808
1809                                 /*
1810                                  * The VM_FAULT_WRITE bit tells us that
1811                                  * do_wp_page has broken COW when necessary,
1812                                  * even if maybe_mkwrite decided not to set
1813                                  * pte_write. We cannot simply drop FOLL_WRITE
1814                                  * here because the COWed page might be gone by
1815                                  * the time we do the subsequent page lookups.
1816                                  */
1817                                 if ((ret & VM_FAULT_WRITE) &&
1818                                     !(vma->vm_flags & VM_WRITE))
1819                                         foll_flags |= FOLL_COW;
1820
1821                                 cond_resched();
1822                         }
1823                         if (IS_ERR(page))
1824                                 return i ? i : PTR_ERR(page);
1825                         if (pages) {
1826                                 pages[i] = page;
1827
1828                                 flush_anon_page(vma, page, start);
1829                                 flush_dcache_page(page);
1830                         }
1831 next_page:
1832                         if (vmas)
1833                                 vmas[i] = vma;
1834                         i++;
1835                         start += PAGE_SIZE;
1836                         nr_pages--;
1837                 } while (nr_pages && start < vma->vm_end);
1838         } while (nr_pages);
1839         return i;
1840 }
1841 EXPORT_SYMBOL(__get_user_pages);
1842
1843 /*
1844  * fixup_user_fault() - manually resolve a user page fault
1845  * @tsk:        the task_struct to use for page fault accounting, or
1846  *              NULL if faults are not to be recorded.
1847  * @mm:         mm_struct of target mm
1848  * @address:    user address
1849  * @fault_flags:flags to pass down to handle_mm_fault()
1850  *
1851  * This is meant to be called in the specific scenario where for locking reasons
1852  * we try to access user memory in atomic context (within a pagefault_disable()
1853  * section), this returns -EFAULT, and we want to resolve the user fault before
1854  * trying again.
1855  *
1856  * Typically this is meant to be used by the futex code.
1857  *
1858  * The main difference with get_user_pages() is that this function will
1859  * unconditionally call handle_mm_fault() which will in turn perform all the
1860  * necessary SW fixup of the dirty and young bits in the PTE, while
1861  * handle_mm_fault() only guarantees to update these in the struct page.
1862  *
1863  * This is important for some architectures where those bits also gate the
1864  * access permission to the page because they are maintained in software.  On
1865  * such architectures, gup() will not be enough to make a subsequent access
1866  * succeed.
1867  *
1868  * This should be called with the mm_sem held for read.
1869  */
1870 int fixup_user_fault(struct task_struct *tsk, struct mm_struct *mm,
1871                      unsigned long address, unsigned int fault_flags)
1872 {
1873         struct vm_area_struct *vma;
1874         vm_flags_t vm_flags;
1875         int ret;
1876
1877         vma = find_extend_vma(mm, address);
1878         if (!vma || address < vma->vm_start)
1879                 return -EFAULT;
1880
1881         vm_flags = (fault_flags & FAULT_FLAG_WRITE) ? VM_WRITE : VM_READ;
1882         if (!(vm_flags & vma->vm_flags))
1883                 return -EFAULT;
1884
1885         ret = handle_mm_fault(mm, vma, address, fault_flags);
1886         if (ret & VM_FAULT_ERROR) {
1887                 if (ret & VM_FAULT_OOM)
1888                         return -ENOMEM;
1889                 if (ret & (VM_FAULT_HWPOISON | VM_FAULT_HWPOISON_LARGE))
1890                         return -EHWPOISON;
1891                 if (ret & (VM_FAULT_SIGBUS | VM_FAULT_SIGSEGV))
1892                         return -EFAULT;
1893                 BUG();
1894         }
1895         if (tsk) {
1896                 if (ret & VM_FAULT_MAJOR)
1897                         tsk->maj_flt++;
1898                 else
1899                         tsk->min_flt++;
1900         }
1901         return 0;
1902 }
1903
1904 /*
1905  * get_user_pages() - pin user pages in memory
1906  * @tsk:        the task_struct to use for page fault accounting, or
1907  *              NULL if faults are not to be recorded.
1908  * @mm:         mm_struct of target mm
1909  * @start:      starting user address
1910  * @nr_pages:   number of pages from start to pin
1911  * @write:      whether pages will be written to by the caller
1912  * @force:      whether to force write access even if user mapping is
1913  *              readonly. This will result in the page being COWed even
1914  *              in MAP_SHARED mappings. You do not want this.
1915  * @pages:      array that receives pointers to the pages pinned.
1916  *              Should be at least nr_pages long. Or NULL, if caller
1917  *              only intends to ensure the pages are faulted in.
1918  * @vmas:       array of pointers to vmas corresponding to each page.
1919  *              Or NULL if the caller does not require them.
1920  *
1921  * Returns number of pages pinned. This may be fewer than the number
1922  * requested. If nr_pages is 0 or negative, returns 0. If no pages
1923  * were pinned, returns -errno. Each page returned must be released
1924  * with a put_page() call when it is finished with. vmas will only
1925  * remain valid while mmap_sem is held.
1926  *
1927  * Must be called with mmap_sem held for read or write.
1928  *
1929  * get_user_pages walks a process's page tables and takes a reference to
1930  * each struct page that each user address corresponds to at a given
1931  * instant. That is, it takes the page that would be accessed if a user
1932  * thread accesses the given user virtual address at that instant.
1933  *
1934  * This does not guarantee that the page exists in the user mappings when
1935  * get_user_pages returns, and there may even be a completely different
1936  * page there in some cases (eg. if mmapped pagecache has been invalidated
1937  * and subsequently re faulted). However it does guarantee that the page
1938  * won't be freed completely. And mostly callers simply care that the page
1939  * contains data that was valid *at some point in time*. Typically, an IO
1940  * or similar operation cannot guarantee anything stronger anyway because
1941  * locks can't be held over the syscall boundary.
1942  *
1943  * If write=0, the page must not be written to. If the page is written to,
1944  * set_page_dirty (or set_page_dirty_lock, as appropriate) must be called
1945  * after the page is finished with, and before put_page is called.
1946  *
1947  * get_user_pages is typically used for fewer-copy IO operations, to get a
1948  * handle on the memory by some means other than accesses via the user virtual
1949  * addresses. The pages may be submitted for DMA to devices or accessed via
1950  * their kernel linear mapping (via the kmap APIs). Care should be taken to
1951  * use the correct cache flushing APIs.
1952  *
1953  * See also get_user_pages_fast, for performance critical applications.
1954  */
1955 int get_user_pages(struct task_struct *tsk, struct mm_struct *mm,
1956                 unsigned long start, int nr_pages, int write, int force,
1957                 struct page **pages, struct vm_area_struct **vmas)
1958 {
1959         int flags = FOLL_TOUCH;
1960
1961         if (pages)
1962                 flags |= FOLL_GET;
1963         if (write)
1964                 flags |= FOLL_WRITE;
1965         if (force)
1966                 flags |= FOLL_FORCE;
1967
1968         return __get_user_pages(tsk, mm, start, nr_pages, flags, pages, vmas,
1969                                 NULL);
1970 }
1971 EXPORT_SYMBOL(get_user_pages);
1972
1973 /**
1974  * get_dump_page() - pin user page in memory while writing it to core dump
1975  * @addr: user address
1976  *
1977  * Returns struct page pointer of user page pinned for dump,
1978  * to be freed afterwards by page_cache_release() or put_page().
1979  *
1980  * Returns NULL on any kind of failure - a hole must then be inserted into
1981  * the corefile, to preserve alignment with its headers; and also returns
1982  * NULL wherever the ZERO_PAGE, or an anonymous pte_none, has been found -
1983  * allowing a hole to be left in the corefile to save diskspace.
1984  *
1985  * Called without mmap_sem, but after all other threads have been killed.
1986  */
1987 #ifdef CONFIG_ELF_CORE
1988 struct page *get_dump_page(unsigned long addr)
1989 {
1990         struct vm_area_struct *vma;
1991         struct page *page;
1992
1993         if (__get_user_pages(current, current->mm, addr, 1,
1994                              FOLL_FORCE | FOLL_DUMP | FOLL_GET, &page, &vma,
1995                              NULL) < 1)
1996                 return NULL;
1997         flush_cache_page(vma, addr, page_to_pfn(page));
1998         return page;
1999 }
2000 #endif /* CONFIG_ELF_CORE */
2001
2002 pte_t *__get_locked_pte(struct mm_struct *mm, unsigned long addr,
2003                         spinlock_t **ptl)
2004 {
2005         pgd_t * pgd = pgd_offset(mm, addr);
2006         pud_t * pud = pud_alloc(mm, pgd, addr);
2007         if (pud) {
2008                 pmd_t * pmd = pmd_alloc(mm, pud, addr);
2009                 if (pmd) {
2010                         VM_BUG_ON(pmd_trans_huge(*pmd));
2011                         return pte_alloc_map_lock(mm, pmd, addr, ptl);
2012                 }
2013         }
2014         return NULL;
2015 }
2016
2017 /*
2018  * This is the old fallback for page remapping.
2019  *
2020  * For historical reasons, it only allows reserved pages. Only
2021  * old drivers should use this, and they needed to mark their
2022  * pages reserved for the old functions anyway.
2023  */
2024 static int insert_page(struct vm_area_struct *vma, unsigned long addr,
2025                         struct page *page, pgprot_t prot)
2026 {
2027         struct mm_struct *mm = vma->vm_mm;
2028         int retval;
2029         pte_t *pte;
2030         spinlock_t *ptl;
2031
2032         retval = -EINVAL;
2033         if (PageAnon(page))
2034                 goto out;
2035         retval = -ENOMEM;
2036         flush_dcache_page(page);
2037         pte = get_locked_pte(mm, addr, &ptl);
2038         if (!pte)
2039                 goto out;
2040         retval = -EBUSY;
2041         if (!pte_none(*pte))
2042                 goto out_unlock;
2043
2044         /* Ok, finally just insert the thing.. */
2045         get_page(page);
2046         inc_mm_counter_fast(mm, MM_FILEPAGES);
2047         page_add_file_rmap(page);
2048         set_pte_at(mm, addr, pte, mk_pte(page, prot));
2049
2050         retval = 0;
2051         pte_unmap_unlock(pte, ptl);
2052         return retval;
2053 out_unlock:
2054         pte_unmap_unlock(pte, ptl);
2055 out:
2056         return retval;
2057 }
2058
2059 /**
2060  * vm_insert_page - insert single page into user vma
2061  * @vma: user vma to map to
2062  * @addr: target user address of this page
2063  * @page: source kernel page
2064  *
2065  * This allows drivers to insert individual pages they've allocated
2066  * into a user vma.
2067  *
2068  * The page has to be a nice clean _individual_ kernel allocation.
2069  * If you allocate a compound page, you need to have marked it as
2070  * such (__GFP_COMP), or manually just split the page up yourself
2071  * (see split_page()).
2072  *
2073  * NOTE! Traditionally this was done with "remap_pfn_range()" which
2074  * took an arbitrary page protection parameter. This doesn't allow
2075  * that. Your vma protection will have to be set up correctly, which
2076  * means that if you want a shared writable mapping, you'd better
2077  * ask for a shared writable mapping!
2078  *
2079  * The page does not need to be reserved.
2080  */
2081 int vm_insert_page(struct vm_area_struct *vma, unsigned long addr,
2082                         struct page *page)
2083 {
2084         if (addr < vma->vm_start || addr >= vma->vm_end)
2085                 return -EFAULT;
2086         if (!page_count(page))
2087                 return -EINVAL;
2088         vma->vm_flags |= VM_INSERTPAGE;
2089         return insert_page(vma, addr, page, vma->vm_page_prot);
2090 }
2091 EXPORT_SYMBOL(vm_insert_page);
2092
2093 static int insert_pfn(struct vm_area_struct *vma, unsigned long addr,
2094                         unsigned long pfn, pgprot_t prot)
2095 {
2096         struct mm_struct *mm = vma->vm_mm;
2097         int retval;
2098         pte_t *pte, entry;
2099         spinlock_t *ptl;
2100
2101         retval = -ENOMEM;
2102         pte = get_locked_pte(mm, addr, &ptl);
2103         if (!pte)
2104                 goto out;
2105         retval = -EBUSY;
2106         if (!pte_none(*pte))
2107                 goto out_unlock;
2108
2109         /* Ok, finally just insert the thing.. */
2110         entry = pte_mkspecial(pfn_pte(pfn, prot));
2111         set_pte_at(mm, addr, pte, entry);
2112         update_mmu_cache(vma, addr, pte); /* XXX: why not for insert_page? */
2113
2114         retval = 0;
2115 out_unlock:
2116         pte_unmap_unlock(pte, ptl);
2117 out:
2118         return retval;
2119 }
2120
2121 /**
2122  * vm_insert_pfn - insert single pfn into user vma
2123  * @vma: user vma to map to
2124  * @addr: target user address of this page
2125  * @pfn: source kernel pfn
2126  *
2127  * Similar to vm_inert_page, this allows drivers to insert individual pages
2128  * they've allocated into a user vma. Same comments apply.
2129  *
2130  * This function should only be called from a vm_ops->fault handler, and
2131  * in that case the handler should return NULL.
2132  *
2133  * vma cannot be a COW mapping.
2134  *
2135  * As this is called only for pages that do not currently exist, we
2136  * do not need to flush old virtual caches or the TLB.
2137  */
2138 int vm_insert_pfn(struct vm_area_struct *vma, unsigned long addr,
2139                         unsigned long pfn)
2140 {
2141         int ret;
2142         pgprot_t pgprot = vma->vm_page_prot;
2143         /*
2144          * Technically, architectures with pte_special can avoid all these
2145          * restrictions (same for remap_pfn_range).  However we would like
2146          * consistency in testing and feature parity among all, so we should
2147          * try to keep these invariants in place for everybody.
2148          */
2149         BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)));
2150         BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) ==
2151                                                 (VM_PFNMAP|VM_MIXEDMAP));
2152         BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags));
2153         BUG_ON((vma->vm_flags & VM_MIXEDMAP) && pfn_valid(pfn));
2154
2155         if (addr < vma->vm_start || addr >= vma->vm_end)
2156                 return -EFAULT;
2157         if (track_pfn_vma_new(vma, &pgprot, pfn, PAGE_SIZE))
2158                 return -EINVAL;
2159
2160         ret = insert_pfn(vma, addr, pfn, pgprot);
2161
2162         if (ret)
2163                 untrack_pfn_vma(vma, pfn, PAGE_SIZE);
2164
2165         return ret;
2166 }
2167 EXPORT_SYMBOL(vm_insert_pfn);
2168
2169 int vm_insert_mixed(struct vm_area_struct *vma, unsigned long addr,
2170                         unsigned long pfn)
2171 {
2172         BUG_ON(!(vma->vm_flags & VM_MIXEDMAP));
2173
2174         if (addr < vma->vm_start || addr >= vma->vm_end)
2175                 return -EFAULT;
2176
2177         /*
2178          * If we don't have pte special, then we have to use the pfn_valid()
2179          * based VM_MIXEDMAP scheme (see vm_normal_page), and thus we *must*
2180          * refcount the page if pfn_valid is true (hence insert_page rather
2181          * than insert_pfn).  If a zero_pfn were inserted into a VM_MIXEDMAP
2182          * without pte special, it would there be refcounted as a normal page.
2183          */
2184         if (!HAVE_PTE_SPECIAL && pfn_valid(pfn)) {
2185                 struct page *page;
2186
2187                 page = pfn_to_page(pfn);
2188                 return insert_page(vma, addr, page, vma->vm_page_prot);
2189         }
2190         return insert_pfn(vma, addr, pfn, vma->vm_page_prot);
2191 }
2192 EXPORT_SYMBOL(vm_insert_mixed);
2193
2194 /*
2195  * maps a range of physical memory into the requested pages. the old
2196  * mappings are removed. any references to nonexistent pages results
2197  * in null mappings (currently treated as "copy-on-access")
2198  */
2199 static int remap_pte_range(struct mm_struct *mm, pmd_t *pmd,
2200                         unsigned long addr, unsigned long end,
2201                         unsigned long pfn, pgprot_t prot)
2202 {
2203         pte_t *pte;
2204         spinlock_t *ptl;
2205
2206         pte = pte_alloc_map_lock(mm, pmd, addr, &ptl);
2207         if (!pte)
2208                 return -ENOMEM;
2209         arch_enter_lazy_mmu_mode();
2210         do {
2211                 BUG_ON(!pte_none(*pte));
2212                 set_pte_at(mm, addr, pte, pte_mkspecial(pfn_pte(pfn, prot)));
2213                 pfn++;
2214         } while (pte++, addr += PAGE_SIZE, addr != end);
2215         arch_leave_lazy_mmu_mode();
2216         pte_unmap_unlock(pte - 1, ptl);
2217         return 0;
2218 }
2219
2220 static inline int remap_pmd_range(struct mm_struct *mm, pud_t *pud,
2221                         unsigned long addr, unsigned long end,
2222                         unsigned long pfn, pgprot_t prot)
2223 {
2224         pmd_t *pmd;
2225         unsigned long next;
2226
2227         pfn -= addr >> PAGE_SHIFT;
2228         pmd = pmd_alloc(mm, pud, addr);
2229         if (!pmd)
2230                 return -ENOMEM;
2231         VM_BUG_ON(pmd_trans_huge(*pmd));
2232         do {
2233                 next = pmd_addr_end(addr, end);
2234                 if (remap_pte_range(mm, pmd, addr, next,
2235                                 pfn + (addr >> PAGE_SHIFT), prot))
2236                         return -ENOMEM;
2237         } while (pmd++, addr = next, addr != end);
2238         return 0;
2239 }
2240
2241 static inline int remap_pud_range(struct mm_struct *mm, pgd_t *pgd,
2242                         unsigned long addr, unsigned long end,
2243                         unsigned long pfn, pgprot_t prot)
2244 {
2245         pud_t *pud;
2246         unsigned long next;
2247
2248         pfn -= addr >> PAGE_SHIFT;
2249         pud = pud_alloc(mm, pgd, addr);
2250         if (!pud)
2251                 return -ENOMEM;
2252         do {
2253                 next = pud_addr_end(addr, end);
2254                 if (remap_pmd_range(mm, pud, addr, next,
2255                                 pfn + (addr >> PAGE_SHIFT), prot))
2256                         return -ENOMEM;
2257         } while (pud++, addr = next, addr != end);
2258         return 0;
2259 }
2260
2261 /**
2262  * remap_pfn_range - remap kernel memory to userspace
2263  * @vma: user vma to map to
2264  * @addr: target user address to start at
2265  * @pfn: physical address of kernel memory
2266  * @size: size of map area
2267  * @prot: page protection flags for this mapping
2268  *
2269  *  Note: this is only safe if the mm semaphore is held when called.
2270  */
2271 int remap_pfn_range(struct vm_area_struct *vma, unsigned long addr,
2272                     unsigned long pfn, unsigned long size, pgprot_t prot)
2273 {
2274         pgd_t *pgd;
2275         unsigned long next;
2276         unsigned long end = addr + PAGE_ALIGN(size);
2277         struct mm_struct *mm = vma->vm_mm;
2278         int err;
2279
2280         /*
2281          * Physically remapped pages are special. Tell the
2282          * rest of the world about it:
2283          *   VM_IO tells people not to look at these pages
2284          *      (accesses can have side effects).
2285          *   VM_RESERVED is specified all over the place, because
2286          *      in 2.4 it kept swapout's vma scan off this vma; but
2287          *      in 2.6 the LRU scan won't even find its pages, so this
2288          *      flag means no more than count its pages in reserved_vm,
2289          *      and omit it from core dump, even when VM_IO turned off.
2290          *   VM_PFNMAP tells the core MM that the base pages are just
2291          *      raw PFN mappings, and do not have a "struct page" associated
2292          *      with them.
2293          *
2294          * There's a horrible special case to handle copy-on-write
2295          * behaviour that some programs depend on. We mark the "original"
2296          * un-COW'ed pages by matching them up with "vma->vm_pgoff".
2297          */
2298         if (addr == vma->vm_start && end == vma->vm_end) {
2299                 vma->vm_pgoff = pfn;
2300                 vma->vm_flags |= VM_PFN_AT_MMAP;
2301         } else if (is_cow_mapping(vma->vm_flags))
2302                 return -EINVAL;
2303
2304         vma->vm_flags |= VM_IO | VM_RESERVED | VM_PFNMAP;
2305
2306         err = track_pfn_vma_new(vma, &prot, pfn, PAGE_ALIGN(size));
2307         if (err) {
2308                 /*
2309                  * To indicate that track_pfn related cleanup is not
2310                  * needed from higher level routine calling unmap_vmas
2311                  */
2312                 vma->vm_flags &= ~(VM_IO | VM_RESERVED | VM_PFNMAP);
2313                 vma->vm_flags &= ~VM_PFN_AT_MMAP;
2314                 return -EINVAL;
2315         }
2316
2317         BUG_ON(addr >= end);
2318         pfn -= addr >> PAGE_SHIFT;
2319         pgd = pgd_offset(mm, addr);
2320         flush_cache_range(vma, addr, end);
2321         do {
2322                 next = pgd_addr_end(addr, end);
2323                 err = remap_pud_range(mm, pgd, addr, next,
2324                                 pfn + (addr >> PAGE_SHIFT), prot);
2325                 if (err)
2326                         break;
2327         } while (pgd++, addr = next, addr != end);
2328
2329         if (err)
2330                 untrack_pfn_vma(vma, pfn, PAGE_ALIGN(size));
2331
2332         return err;
2333 }
2334 EXPORT_SYMBOL(remap_pfn_range);
2335
2336 /**
2337  * vm_iomap_memory - remap memory to userspace
2338  * @vma: user vma to map to
2339  * @start: start of area
2340  * @len: size of area
2341  *
2342  * This is a simplified io_remap_pfn_range() for common driver use. The
2343  * driver just needs to give us the physical memory range to be mapped,
2344  * we'll figure out the rest from the vma information.
2345  *
2346  * NOTE! Some drivers might want to tweak vma->vm_page_prot first to get
2347  * whatever write-combining details or similar.
2348  */
2349 int vm_iomap_memory(struct vm_area_struct *vma, phys_addr_t start, unsigned long len)
2350 {
2351         unsigned long vm_len, pfn, pages;
2352
2353         /* Check that the physical memory area passed in looks valid */
2354         if (start + len < start)
2355                 return -EINVAL;
2356         /*
2357          * You *really* shouldn't map things that aren't page-aligned,
2358          * but we've historically allowed it because IO memory might
2359          * just have smaller alignment.
2360          */
2361         len += start & ~PAGE_MASK;
2362         pfn = start >> PAGE_SHIFT;
2363         pages = (len + ~PAGE_MASK) >> PAGE_SHIFT;
2364         if (pfn + pages < pfn)
2365                 return -EINVAL;
2366
2367         /* We start the mapping 'vm_pgoff' pages into the area */
2368         if (vma->vm_pgoff > pages)
2369                 return -EINVAL;
2370         pfn += vma->vm_pgoff;
2371         pages -= vma->vm_pgoff;
2372
2373         /* Can we fit all of the mapping? */
2374         vm_len = vma->vm_end - vma->vm_start;
2375         if (vm_len >> PAGE_SHIFT > pages)
2376                 return -EINVAL;
2377
2378         /* Ok, let it rip */
2379         return io_remap_pfn_range(vma, vma->vm_start, pfn, vm_len, vma->vm_page_prot);
2380 }
2381 EXPORT_SYMBOL(vm_iomap_memory);
2382
2383 static int apply_to_pte_range(struct mm_struct *mm, pmd_t *pmd,
2384                                      unsigned long addr, unsigned long end,
2385                                      pte_fn_t fn, void *data)
2386 {
2387         pte_t *pte;
2388         int err;
2389         pgtable_t token;
2390         spinlock_t *uninitialized_var(ptl);
2391
2392         pte = (mm == &init_mm) ?
2393                 pte_alloc_kernel(pmd, addr) :
2394                 pte_alloc_map_lock(mm, pmd, addr, &ptl);
2395         if (!pte)
2396                 return -ENOMEM;
2397
2398         BUG_ON(pmd_huge(*pmd));
2399
2400         arch_enter_lazy_mmu_mode();
2401
2402         token = pmd_pgtable(*pmd);
2403
2404         do {
2405                 err = fn(pte++, token, addr, data);
2406                 if (err)
2407                         break;
2408         } while (addr += PAGE_SIZE, addr != end);
2409
2410         arch_leave_lazy_mmu_mode();
2411
2412         if (mm != &init_mm)
2413                 pte_unmap_unlock(pte-1, ptl);
2414         return err;
2415 }
2416
2417 static int apply_to_pmd_range(struct mm_struct *mm, pud_t *pud,
2418                                      unsigned long addr, unsigned long end,
2419                                      pte_fn_t fn, void *data)
2420 {
2421         pmd_t *pmd;
2422         unsigned long next;
2423         int err;
2424
2425         BUG_ON(pud_huge(*pud));
2426
2427         pmd = pmd_alloc(mm, pud, addr);
2428         if (!pmd)
2429                 return -ENOMEM;
2430         do {
2431                 next = pmd_addr_end(addr, end);
2432                 err = apply_to_pte_range(mm, pmd, addr, next, fn, data);
2433                 if (err)
2434                         break;
2435         } while (pmd++, addr = next, addr != end);
2436         return err;
2437 }
2438
2439 static int apply_to_pud_range(struct mm_struct *mm, pgd_t *pgd,
2440                                      unsigned long addr, unsigned long end,
2441                                      pte_fn_t fn, void *data)
2442 {
2443         pud_t *pud;
2444         unsigned long next;
2445         int err;
2446
2447         pud = pud_alloc(mm, pgd, addr);
2448         if (!pud)
2449                 return -ENOMEM;
2450         do {
2451                 next = pud_addr_end(addr, end);
2452                 err = apply_to_pmd_range(mm, pud, addr, next, fn, data);
2453                 if (err)
2454                         break;
2455         } while (pud++, addr = next, addr != end);
2456         return err;
2457 }
2458
2459 /*
2460  * Scan a region of virtual memory, filling in page tables as necessary
2461  * and calling a provided function on each leaf page table.
2462  */
2463 int apply_to_page_range(struct mm_struct *mm, unsigned long addr,
2464                         unsigned long size, pte_fn_t fn, void *data)
2465 {
2466         pgd_t *pgd;
2467         unsigned long next;
2468         unsigned long end = addr + size;
2469         int err;
2470
2471         BUG_ON(addr >= end);
2472         pgd = pgd_offset(mm, addr);
2473         do {
2474                 next = pgd_addr_end(addr, end);
2475                 err = apply_to_pud_range(mm, pgd, addr, next, fn, data);
2476                 if (err)
2477                         break;
2478         } while (pgd++, addr = next, addr != end);
2479
2480         return err;
2481 }
2482 EXPORT_SYMBOL_GPL(apply_to_page_range);
2483
2484 /*
2485  * handle_pte_fault chooses page fault handler according to an entry
2486  * which was read non-atomically.  Before making any commitment, on
2487  * those architectures or configurations (e.g. i386 with PAE) which
2488  * might give a mix of unmatched parts, do_swap_page and do_nonlinear_fault
2489  * must check under lock before unmapping the pte and proceeding
2490  * (but do_wp_page is only called after already making such a check;
2491  * and do_anonymous_page can safely check later on).
2492  */
2493 static inline int pte_unmap_same(struct mm_struct *mm, pmd_t *pmd,
2494                                 pte_t *page_table, pte_t orig_pte)
2495 {
2496         int same = 1;
2497 #if defined(CONFIG_SMP) || defined(CONFIG_PREEMPT)
2498         if (sizeof(pte_t) > sizeof(unsigned long)) {
2499                 spinlock_t *ptl = pte_lockptr(mm, pmd);
2500                 spin_lock(ptl);
2501                 same = pte_same(*page_table, orig_pte);
2502                 spin_unlock(ptl);
2503         }
2504 #endif
2505         pte_unmap(page_table);
2506         return same;
2507 }
2508
2509 static inline void cow_user_page(struct page *dst, struct page *src, unsigned long va, struct vm_area_struct *vma)
2510 {
2511         /*
2512          * If the source page was a PFN mapping, we don't have
2513          * a "struct page" for it. We do a best-effort copy by
2514          * just copying from the original user address. If that
2515          * fails, we just zero-fill it. Live with it.
2516          */
2517         if (unlikely(!src)) {
2518                 void *kaddr = kmap_atomic(dst, KM_USER0);
2519                 void __user *uaddr = (void __user *)(va & PAGE_MASK);
2520
2521                 /*
2522                  * This really shouldn't fail, because the page is there
2523                  * in the page tables. But it might just be unreadable,
2524                  * in which case we just give up and fill the result with
2525                  * zeroes.
2526                  */
2527                 if (__copy_from_user_inatomic(kaddr, uaddr, PAGE_SIZE))
2528                         clear_page(kaddr);
2529                 kunmap_atomic(kaddr, KM_USER0);
2530                 flush_dcache_page(dst);
2531         } else
2532                 copy_user_highpage(dst, src, va, vma);
2533 }
2534
2535 /*
2536  * This routine handles present pages, when users try to write
2537  * to a shared page. It is done by copying the page to a new address
2538  * and decrementing the shared-page counter for the old page.
2539  *
2540  * Note that this routine assumes that the protection checks have been
2541  * done by the caller (the low-level page fault routine in most cases).
2542  * Thus we can safely just mark it writable once we've done any necessary
2543  * COW.
2544  *
2545  * We also mark the page dirty at this point even though the page will
2546  * change only once the write actually happens. This avoids a few races,
2547  * and potentially makes it more efficient.
2548  *
2549  * We enter with non-exclusive mmap_sem (to exclude vma changes,
2550  * but allow concurrent faults), with pte both mapped and locked.
2551  * We return with mmap_sem still held, but pte unmapped and unlocked.
2552  */
2553 static int do_wp_page(struct mm_struct *mm, struct vm_area_struct *vma,
2554                 unsigned long address, pte_t *page_table, pmd_t *pmd,
2555                 spinlock_t *ptl, pte_t orig_pte)
2556         __releases(ptl)
2557 {
2558         struct page *old_page, *new_page;
2559         pte_t entry;
2560         int ret = 0;
2561         int page_mkwrite = 0;
2562         struct page *dirty_page = NULL;
2563
2564         old_page = vm_normal_page(vma, address, orig_pte);
2565         if (!old_page) {
2566                 /*
2567                  * VM_MIXEDMAP !pfn_valid() case
2568                  *
2569                  * We should not cow pages in a shared writeable mapping.
2570                  * Just mark the pages writable as we can't do any dirty
2571                  * accounting on raw pfn maps.
2572                  */
2573                 if ((vma->vm_flags & (VM_WRITE|VM_SHARED)) ==
2574                                      (VM_WRITE|VM_SHARED))
2575                         goto reuse;
2576                 goto gotten;
2577         }
2578
2579         /*
2580          * Take out anonymous pages first, anonymous shared vmas are
2581          * not dirty accountable.
2582          */
2583         if (PageAnon(old_page) && !PageKsm(old_page)) {
2584                 if (!trylock_page(old_page)) {
2585                         page_cache_get(old_page);
2586                         pte_unmap_unlock(page_table, ptl);
2587                         lock_page(old_page);
2588                         page_table = pte_offset_map_lock(mm, pmd, address,
2589                                                          &ptl);
2590                         if (!pte_same(*page_table, orig_pte)) {
2591                                 unlock_page(old_page);
2592                                 goto unlock;
2593                         }
2594                         page_cache_release(old_page);
2595                 }
2596                 if (reuse_swap_page(old_page)) {
2597                         /*
2598                          * The page is all ours.  Move it to our anon_vma so
2599                          * the rmap code will not search our parent or siblings.
2600                          * Protected against the rmap code by the page lock.
2601                          */
2602                         page_move_anon_rmap(old_page, vma, address);
2603                         unlock_page(old_page);
2604                         goto reuse;
2605                 }
2606                 unlock_page(old_page);
2607         } else if (unlikely((vma->vm_flags & (VM_WRITE|VM_SHARED)) ==
2608                                         (VM_WRITE|VM_SHARED))) {
2609                 /*
2610                  * Only catch write-faults on shared writable pages,
2611                  * read-only shared pages can get COWed by
2612                  * get_user_pages(.write=1, .force=1).
2613                  */
2614                 if (vma->vm_ops && vma->vm_ops->page_mkwrite) {
2615                         struct vm_fault vmf;
2616                         int tmp;
2617
2618                         vmf.virtual_address = (void __user *)(address &
2619                                                                 PAGE_MASK);
2620                         vmf.pgoff = old_page->index;
2621                         vmf.flags = FAULT_FLAG_WRITE|FAULT_FLAG_MKWRITE;
2622                         vmf.page = old_page;
2623
2624                         /*
2625                          * Notify the address space that the page is about to
2626                          * become writable so that it can prohibit this or wait
2627                          * for the page to get into an appropriate state.
2628                          *
2629                          * We do this without the lock held, so that it can
2630                          * sleep if it needs to.
2631                          */
2632                         page_cache_get(old_page);
2633                         pte_unmap_unlock(page_table, ptl);
2634
2635                         tmp = vma->vm_ops->page_mkwrite(vma, &vmf);
2636                         if (unlikely(tmp &
2637                                         (VM_FAULT_ERROR | VM_FAULT_NOPAGE))) {
2638                                 ret = tmp;
2639                                 goto unwritable_page;
2640                         }
2641                         if (unlikely(!(tmp & VM_FAULT_LOCKED))) {
2642                                 lock_page(old_page);
2643                                 if (!old_page->mapping) {
2644                                         ret = 0; /* retry the fault */
2645                                         unlock_page(old_page);
2646                                         goto unwritable_page;
2647                                 }
2648                         } else
2649                                 VM_BUG_ON(!PageLocked(old_page));
2650
2651                         /*
2652                          * Since we dropped the lock we need to revalidate
2653                          * the PTE as someone else may have changed it.  If
2654                          * they did, we just return, as we can count on the
2655                          * MMU to tell us if they didn't also make it writable.
2656                          */
2657                         page_table = pte_offset_map_lock(mm, pmd, address,
2658                                                          &ptl);
2659                         if (!pte_same(*page_table, orig_pte)) {
2660                                 unlock_page(old_page);
2661                                 goto unlock;
2662                         }
2663
2664                         page_mkwrite = 1;
2665                 }
2666                 dirty_page = old_page;
2667                 get_page(dirty_page);
2668
2669 reuse:
2670                 flush_cache_page(vma, address, pte_pfn(orig_pte));
2671                 entry = pte_mkyoung(orig_pte);
2672                 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2673                 if (ptep_set_access_flags(vma, address, page_table, entry,1))
2674                         update_mmu_cache(vma, address, page_table);
2675                 pte_unmap_unlock(page_table, ptl);
2676                 ret |= VM_FAULT_WRITE;
2677
2678                 if (!dirty_page)
2679                         return ret;
2680
2681                 if (!page_mkwrite) {
2682                         struct address_space *mapping;
2683                         int dirtied;
2684
2685                         lock_page(dirty_page);
2686                         dirtied = set_page_dirty(dirty_page);
2687                         VM_BUG_ON(PageAnon(dirty_page));
2688                         mapping = dirty_page->mapping;
2689                         unlock_page(dirty_page);
2690
2691                         if (dirtied && 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                 put_page(dirty_page);
2701                 if (page_mkwrite) {
2702                         struct address_space *mapping = dirty_page->mapping;
2703
2704                         set_page_dirty(dirty_page);
2705                         unlock_page(dirty_page);
2706                         page_cache_release(dirty_page);
2707                         if (mapping)    {
2708                                 /*
2709                                  * Some device drivers do not set page.mapping
2710                                  * but still dirty their pages
2711                                  */
2712                                 balance_dirty_pages_ratelimited(mapping);
2713                         }
2714                 }
2715
2716                 /* file_update_time outside page_lock */
2717                 if (vma->vm_file)
2718                         file_update_time(vma->vm_file);
2719
2720                 return ret;
2721         }
2722
2723         /*
2724          * Ok, we need to copy. Oh, well..
2725          */
2726         page_cache_get(old_page);
2727 gotten:
2728         pte_unmap_unlock(page_table, ptl);
2729
2730         if (unlikely(anon_vma_prepare(vma)))
2731                 goto oom;
2732
2733         if (is_zero_pfn(pte_pfn(orig_pte))) {
2734                 new_page = alloc_zeroed_user_highpage_movable(vma, address);
2735                 if (!new_page)
2736                         goto oom;
2737         } else {
2738                 new_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, address);
2739                 if (!new_page)
2740                         goto oom;
2741                 cow_user_page(new_page, old_page, address, vma);
2742         }
2743         __SetPageUptodate(new_page);
2744
2745         if (mem_cgroup_newpage_charge(new_page, mm, GFP_KERNEL))
2746                 goto oom_free_new;
2747
2748         /*
2749          * Re-check the pte - we dropped the lock
2750          */
2751         page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2752         if (likely(pte_same(*page_table, orig_pte))) {
2753                 if (old_page) {
2754                         if (!PageAnon(old_page)) {
2755                                 dec_mm_counter_fast(mm, MM_FILEPAGES);
2756                                 inc_mm_counter_fast(mm, MM_ANONPAGES);
2757                         }
2758                 } else
2759                         inc_mm_counter_fast(mm, MM_ANONPAGES);
2760                 flush_cache_page(vma, address, pte_pfn(orig_pte));
2761                 entry = mk_pte(new_page, vma->vm_page_prot);
2762                 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2763                 /*
2764                  * Clear the pte entry and flush it first, before updating the
2765                  * pte with the new entry. This will avoid a race condition
2766                  * seen in the presence of one thread doing SMC and another
2767                  * thread doing COW.
2768                  */
2769                 ptep_clear_flush(vma, address, page_table);
2770                 page_add_new_anon_rmap(new_page, vma, address);
2771                 /*
2772                  * We call the notify macro here because, when using secondary
2773                  * mmu page tables (such as kvm shadow page tables), we want the
2774                  * new page to be mapped directly into the secondary page table.
2775                  */
2776                 set_pte_at_notify(mm, address, page_table, entry);
2777                 update_mmu_cache(vma, address, page_table);
2778                 if (old_page) {
2779                         /*
2780                          * Only after switching the pte to the new page may
2781                          * we remove the mapcount here. Otherwise another
2782                          * process may come and find the rmap count decremented
2783                          * before the pte is switched to the new page, and
2784                          * "reuse" the old page writing into it while our pte
2785                          * here still points into it and can be read by other
2786                          * threads.
2787                          *
2788                          * The critical issue is to order this
2789                          * page_remove_rmap with the ptp_clear_flush above.
2790                          * Those stores are ordered by (if nothing else,)
2791                          * the barrier present in the atomic_add_negative
2792                          * in page_remove_rmap.
2793                          *
2794                          * Then the TLB flush in ptep_clear_flush ensures that
2795                          * no process can access the old page before the
2796                          * decremented mapcount is visible. And the old page
2797                          * cannot be reused until after the decremented
2798                          * mapcount is visible. So transitively, TLBs to
2799                          * old page will be flushed before it can be reused.
2800                          */
2801                         page_remove_rmap(old_page);
2802                 }
2803
2804                 /* Free the old page.. */
2805                 new_page = old_page;
2806                 ret |= VM_FAULT_WRITE;
2807         } else
2808                 mem_cgroup_uncharge_page(new_page);
2809
2810         if (new_page)
2811                 page_cache_release(new_page);
2812 unlock:
2813         pte_unmap_unlock(page_table, ptl);
2814         if (old_page) {
2815                 /*
2816                  * Don't let another task, with possibly unlocked vma,
2817                  * keep the mlocked page.
2818                  */
2819                 if ((ret & VM_FAULT_WRITE) && (vma->vm_flags & VM_LOCKED)) {
2820                         lock_page(old_page);    /* LRU manipulation */
2821                         munlock_vma_page(old_page);
2822                         unlock_page(old_page);
2823                 }
2824                 page_cache_release(old_page);
2825         }
2826         return ret;
2827 oom_free_new:
2828         page_cache_release(new_page);
2829 oom:
2830         if (old_page) {
2831                 if (page_mkwrite) {
2832                         unlock_page(old_page);
2833                         page_cache_release(old_page);
2834                 }
2835                 page_cache_release(old_page);
2836         }
2837         return VM_FAULT_OOM;
2838
2839 unwritable_page:
2840         page_cache_release(old_page);
2841         return ret;
2842 }
2843
2844 static void unmap_mapping_range_vma(struct vm_area_struct *vma,
2845                 unsigned long start_addr, unsigned long end_addr,
2846                 struct zap_details *details)
2847 {
2848         zap_page_range(vma, start_addr, end_addr - start_addr, details);
2849 }
2850
2851 static inline void unmap_mapping_range_tree(struct prio_tree_root *root,
2852                                             struct zap_details *details)
2853 {
2854         struct vm_area_struct *vma;
2855         struct prio_tree_iter iter;
2856         pgoff_t vba, vea, zba, zea;
2857
2858         vma_prio_tree_foreach(vma, &iter, root,
2859                         details->first_index, details->last_index) {
2860
2861                 vba = vma->vm_pgoff;
2862                 vea = vba + ((vma->vm_end - vma->vm_start) >> PAGE_SHIFT) - 1;
2863                 /* Assume for now that PAGE_CACHE_SHIFT == PAGE_SHIFT */
2864                 zba = details->first_index;
2865                 if (zba < vba)
2866                         zba = vba;
2867                 zea = details->last_index;
2868                 if (zea > vea)
2869                         zea = vea;
2870
2871                 unmap_mapping_range_vma(vma,
2872                         ((zba - vba) << PAGE_SHIFT) + vma->vm_start,
2873                         ((zea - vba + 1) << PAGE_SHIFT) + vma->vm_start,
2874                                 details);
2875         }
2876 }
2877
2878 static inline void unmap_mapping_range_list(struct list_head *head,
2879                                             struct zap_details *details)
2880 {
2881         struct vm_area_struct *vma;
2882
2883         /*
2884          * In nonlinear VMAs there is no correspondence between virtual address
2885          * offset and file offset.  So we must perform an exhaustive search
2886          * across *all* the pages in each nonlinear VMA, not just the pages
2887          * whose virtual address lies outside the file truncation point.
2888          */
2889         list_for_each_entry(vma, head, shared.vm_set.list) {
2890                 details->nonlinear_vma = vma;
2891                 unmap_mapping_range_vma(vma, vma->vm_start, vma->vm_end, details);
2892         }
2893 }
2894
2895 /**
2896  * unmap_mapping_range - unmap the portion of all mmaps in the specified address_space corresponding to the specified page range in the underlying file.
2897  * @mapping: the address space containing mmaps to be unmapped.
2898  * @holebegin: byte in first page to unmap, relative to the start of
2899  * the underlying file.  This will be rounded down to a PAGE_SIZE
2900  * boundary.  Note that this is different from truncate_pagecache(), which
2901  * must keep the partial page.  In contrast, we must get rid of
2902  * partial pages.
2903  * @holelen: size of prospective hole in bytes.  This will be rounded
2904  * up to a PAGE_SIZE boundary.  A holelen of zero truncates to the
2905  * end of the file.
2906  * @even_cows: 1 when truncating a file, unmap even private COWed pages;
2907  * but 0 when invalidating pagecache, don't throw away private data.
2908  */
2909 void unmap_mapping_range(struct address_space *mapping,
2910                 loff_t const holebegin, loff_t const holelen, int even_cows)
2911 {
2912         struct zap_details details;
2913         pgoff_t hba = holebegin >> PAGE_SHIFT;
2914         pgoff_t hlen = (holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
2915
2916         /* Check for overflow. */
2917         if (sizeof(holelen) > sizeof(hlen)) {
2918                 long long holeend =
2919                         (holebegin + holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
2920                 if (holeend & ~(long long)ULONG_MAX)
2921                         hlen = ULONG_MAX - hba + 1;
2922         }
2923
2924         details.check_mapping = even_cows? NULL: mapping;
2925         details.nonlinear_vma = NULL;
2926         details.first_index = hba;
2927         details.last_index = hba + hlen - 1;
2928         if (details.last_index < details.first_index)
2929                 details.last_index = ULONG_MAX;
2930
2931
2932         mutex_lock(&mapping->i_mmap_mutex);
2933         if (unlikely(!prio_tree_empty(&mapping->i_mmap)))
2934                 unmap_mapping_range_tree(&mapping->i_mmap, &details);
2935         if (unlikely(!list_empty(&mapping->i_mmap_nonlinear)))
2936                 unmap_mapping_range_list(&mapping->i_mmap_nonlinear, &details);
2937         mutex_unlock(&mapping->i_mmap_mutex);
2938 }
2939 EXPORT_SYMBOL(unmap_mapping_range);
2940
2941 /*
2942  * We enter with non-exclusive mmap_sem (to exclude vma changes,
2943  * but allow concurrent faults), and pte mapped but not yet locked.
2944  * We return with mmap_sem still held, but pte unmapped and unlocked.
2945  */
2946 static int do_swap_page(struct mm_struct *mm, struct vm_area_struct *vma,
2947                 unsigned long address, pte_t *page_table, pmd_t *pmd,
2948                 unsigned int flags, pte_t orig_pte)
2949 {
2950         spinlock_t *ptl;
2951         struct page *page, *swapcache = NULL;
2952         swp_entry_t entry;
2953         pte_t pte;
2954         int locked;
2955         struct mem_cgroup *ptr;
2956         int exclusive = 0;
2957         int ret = 0;
2958
2959         if (!pte_unmap_same(mm, pmd, page_table, orig_pte))
2960                 goto out;
2961
2962         entry = pte_to_swp_entry(orig_pte);
2963         if (unlikely(non_swap_entry(entry))) {
2964                 if (is_migration_entry(entry)) {
2965                         migration_entry_wait(mm, pmd, address);
2966                 } else if (is_hwpoison_entry(entry)) {
2967                         ret = VM_FAULT_HWPOISON;
2968                 } else {
2969                         print_bad_pte(vma, address, orig_pte, NULL);
2970                         ret = VM_FAULT_SIGBUS;
2971                 }
2972                 goto out;
2973         }
2974         delayacct_set_flag(DELAYACCT_PF_SWAPIN);
2975         page = lookup_swap_cache(entry);
2976         if (!page) {
2977                 grab_swap_token(mm); /* Contend for token _before_ read-in */
2978                 page = swapin_readahead(entry,
2979                                         GFP_HIGHUSER_MOVABLE, vma, address);
2980                 if (!page) {
2981                         /*
2982                          * Back out if somebody else faulted in this pte
2983                          * while we released the pte lock.
2984                          */
2985                         page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2986                         if (likely(pte_same(*page_table, orig_pte)))
2987                                 ret = VM_FAULT_OOM;
2988                         delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
2989                         goto unlock;
2990                 }
2991
2992                 /* Had to read the page from swap area: Major fault */
2993                 ret = VM_FAULT_MAJOR;
2994                 count_vm_event(PGMAJFAULT);
2995                 mem_cgroup_count_vm_event(mm, PGMAJFAULT);
2996         } else if (PageHWPoison(page)) {
2997                 /*
2998                  * hwpoisoned dirty swapcache pages are kept for killing
2999                  * owner processes (which may be unknown at hwpoison time)
3000                  */
3001                 ret = VM_FAULT_HWPOISON;
3002                 delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
3003                 goto out_release;
3004         }
3005
3006         locked = lock_page_or_retry(page, mm, flags);
3007         delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
3008         if (!locked) {
3009                 ret |= VM_FAULT_RETRY;
3010                 goto out_release;
3011         }
3012
3013         /*
3014          * Make sure try_to_free_swap or reuse_swap_page or swapoff did not
3015          * release the swapcache from under us.  The page pin, and pte_same
3016          * test below, are not enough to exclude that.  Even if it is still
3017          * swapcache, we need to check that the page's swap has not changed.
3018          */
3019         if (unlikely(!PageSwapCache(page) || page_private(page) != entry.val))
3020                 goto out_page;
3021
3022         if (ksm_might_need_to_copy(page, vma, address)) {
3023                 swapcache = page;
3024                 page = ksm_does_need_to_copy(page, vma, address);
3025
3026                 if (unlikely(!page)) {
3027                         ret = VM_FAULT_OOM;
3028                         page = swapcache;
3029                         swapcache = NULL;
3030                         goto out_page;
3031                 }
3032         }
3033
3034         if (mem_cgroup_try_charge_swapin(mm, page, GFP_KERNEL, &ptr)) {
3035                 ret = VM_FAULT_OOM;
3036                 goto out_page;
3037         }
3038
3039         /*
3040          * Back out if somebody else already faulted in this pte.
3041          */
3042         page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
3043         if (unlikely(!pte_same(*page_table, orig_pte)))
3044                 goto out_nomap;
3045
3046         if (unlikely(!PageUptodate(page))) {
3047                 ret = VM_FAULT_SIGBUS;
3048                 goto out_nomap;
3049         }
3050
3051         /*
3052          * The page isn't present yet, go ahead with the fault.
3053          *
3054          * Be careful about the sequence of operations here.
3055          * To get its accounting right, reuse_swap_page() must be called
3056          * while the page is counted on swap but not yet in mapcount i.e.
3057          * before page_add_anon_rmap() and swap_free(); try_to_free_swap()
3058          * must be called after the swap_free(), or it will never succeed.
3059          * Because delete_from_swap_page() may be called by reuse_swap_page(),
3060          * mem_cgroup_commit_charge_swapin() may not be able to find swp_entry
3061          * in page->private. In this case, a record in swap_cgroup  is silently
3062          * discarded at swap_free().
3063          */
3064
3065         inc_mm_counter_fast(mm, MM_ANONPAGES);
3066         dec_mm_counter_fast(mm, MM_SWAPENTS);
3067         pte = mk_pte(page, vma->vm_page_prot);
3068         if ((flags & FAULT_FLAG_WRITE) && reuse_swap_page(page)) {
3069                 pte = maybe_mkwrite(pte_mkdirty(pte), vma);
3070                 flags &= ~FAULT_FLAG_WRITE;
3071                 ret |= VM_FAULT_WRITE;
3072                 exclusive = 1;
3073         }
3074         flush_icache_page(vma, page);
3075         set_pte_at(mm, address, page_table, pte);
3076         do_page_add_anon_rmap(page, vma, address, exclusive);
3077         /* It's better to call commit-charge after rmap is established */
3078         mem_cgroup_commit_charge_swapin(page, ptr);
3079
3080         swap_free(entry);
3081         if (vm_swap_full() || (vma->vm_flags & VM_LOCKED) || PageMlocked(page))
3082                 try_to_free_swap(page);
3083         unlock_page(page);
3084         if (swapcache) {
3085                 /*
3086                  * Hold the lock to avoid the swap entry to be reused
3087                  * until we take the PT lock for the pte_same() check
3088                  * (to avoid false positives from pte_same). For
3089                  * further safety release the lock after the swap_free
3090                  * so that the swap count won't change under a
3091                  * parallel locked swapcache.
3092                  */
3093                 unlock_page(swapcache);
3094                 page_cache_release(swapcache);
3095         }
3096
3097         if (flags & FAULT_FLAG_WRITE) {
3098                 ret |= do_wp_page(mm, vma, address, page_table, pmd, ptl, pte);
3099                 if (ret & VM_FAULT_ERROR)
3100                         ret &= VM_FAULT_ERROR;
3101                 goto out;
3102         }
3103
3104         /* No need to invalidate - it was non-present before */
3105         update_mmu_cache(vma, address, page_table);
3106 unlock:
3107         pte_unmap_unlock(page_table, ptl);
3108 out:
3109         return ret;
3110 out_nomap:
3111         mem_cgroup_cancel_charge_swapin(ptr);
3112         pte_unmap_unlock(page_table, ptl);
3113 out_page:
3114         unlock_page(page);
3115 out_release:
3116         page_cache_release(page);
3117         if (swapcache) {
3118                 unlock_page(swapcache);
3119                 page_cache_release(swapcache);
3120         }
3121         return ret;
3122 }
3123
3124 /*
3125  * This is like a special single-page "expand_{down|up}wards()",
3126  * except we must first make sure that 'address{-|+}PAGE_SIZE'
3127  * doesn't hit another vma.
3128  */
3129 static inline int check_stack_guard_page(struct vm_area_struct *vma, unsigned long address)
3130 {
3131         address &= PAGE_MASK;
3132         if ((vma->vm_flags & VM_GROWSDOWN) && address == vma->vm_start) {
3133                 struct vm_area_struct *prev = vma->vm_prev;
3134
3135                 /*
3136                  * Is there a mapping abutting this one below?
3137                  *
3138                  * That's only ok if it's the same stack mapping
3139                  * that has gotten split..
3140                  */
3141                 if (prev && prev->vm_end == address)
3142                         return prev->vm_flags & VM_GROWSDOWN ? 0 : -ENOMEM;
3143
3144                 return expand_downwards(vma, address - PAGE_SIZE);
3145         }
3146         if ((vma->vm_flags & VM_GROWSUP) && address + PAGE_SIZE == vma->vm_end) {
3147                 struct vm_area_struct *next = vma->vm_next;
3148
3149                 /* As VM_GROWSDOWN but s/below/above/ */
3150                 if (next && next->vm_start == address + PAGE_SIZE)
3151                         return next->vm_flags & VM_GROWSUP ? 0 : -ENOMEM;
3152
3153                 return expand_upwards(vma, address + PAGE_SIZE);
3154         }
3155         return 0;
3156 }
3157
3158 /*
3159  * We enter with non-exclusive mmap_sem (to exclude vma changes,
3160  * but allow concurrent faults), and pte mapped but not yet locked.
3161  * We return with mmap_sem still held, but pte unmapped and unlocked.
3162  */
3163 static int do_anonymous_page(struct mm_struct *mm, struct vm_area_struct *vma,
3164                 unsigned long address, pte_t *page_table, pmd_t *pmd,
3165                 unsigned int flags)
3166 {
3167         struct page *page;
3168         spinlock_t *ptl;
3169         pte_t entry;
3170
3171         pte_unmap(page_table);
3172
3173         /* File mapping without ->vm_ops ? */
3174         if (vma->vm_flags & VM_SHARED)
3175                 return VM_FAULT_SIGBUS;
3176
3177         /* Check if we need to add a guard page to the stack */
3178         if (check_stack_guard_page(vma, address) < 0)
3179                 return VM_FAULT_SIGSEGV;
3180
3181         /* Use the zero-page for reads */
3182         if (!(flags & FAULT_FLAG_WRITE)) {
3183                 entry = pte_mkspecial(pfn_pte(my_zero_pfn(address),
3184                                                 vma->vm_page_prot));
3185                 page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
3186                 if (!pte_none(*page_table))
3187                         goto unlock;
3188                 goto setpte;
3189         }
3190
3191         /* Allocate our own private page. */
3192         if (unlikely(anon_vma_prepare(vma)))
3193                 goto oom;
3194         page = alloc_zeroed_user_highpage_movable(vma, address);
3195         if (!page)
3196                 goto oom;
3197         __SetPageUptodate(page);
3198
3199         if (mem_cgroup_newpage_charge(page, mm, GFP_KERNEL))
3200                 goto oom_free_page;
3201
3202         entry = mk_pte(page, vma->vm_page_prot);
3203         if (vma->vm_flags & VM_WRITE)
3204                 entry = pte_mkwrite(pte_mkdirty(entry));
3205
3206         page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
3207         if (!pte_none(*page_table))
3208                 goto release;
3209
3210         inc_mm_counter_fast(mm, MM_ANONPAGES);
3211         page_add_new_anon_rmap(page, vma, address);
3212 setpte:
3213         set_pte_at(mm, address, page_table, entry);
3214
3215         /* No need to invalidate - it was non-present before */
3216         update_mmu_cache(vma, address, page_table);
3217 unlock:
3218         pte_unmap_unlock(page_table, ptl);
3219         return 0;
3220 release:
3221         mem_cgroup_uncharge_page(page);
3222         page_cache_release(page);
3223         goto unlock;
3224 oom_free_page:
3225         page_cache_release(page);
3226 oom:
3227         return VM_FAULT_OOM;
3228 }
3229
3230 /*
3231  * __do_fault() tries to create a new page mapping. It aggressively
3232  * tries to share with existing pages, but makes a separate copy if
3233  * the FAULT_FLAG_WRITE is set in the flags parameter in order to avoid
3234  * the next page fault.
3235  *
3236  * As this is called only for pages that do not currently exist, we
3237  * do not need to flush old virtual caches or the TLB.
3238  *
3239  * We enter with non-exclusive mmap_sem (to exclude vma changes,
3240  * but allow concurrent faults), and pte neither mapped nor locked.
3241  * We return with mmap_sem still held, but pte unmapped and unlocked.
3242  */
3243 static int __do_fault(struct mm_struct *mm, struct vm_area_struct *vma,
3244                 unsigned long address, pmd_t *pmd,
3245                 pgoff_t pgoff, unsigned int flags, pte_t orig_pte)
3246 {
3247         pte_t *page_table;
3248         spinlock_t *ptl;
3249         struct page *page;
3250         struct page *cow_page;
3251         pte_t entry;
3252         int anon = 0;
3253         struct page *dirty_page = NULL;
3254         struct vm_fault vmf;
3255         int ret;
3256         int page_mkwrite = 0;
3257
3258         /*
3259          * If we do COW later, allocate page befor taking lock_page()
3260          * on the file cache page. This will reduce lock holding time.
3261          */
3262         if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) {
3263
3264                 if (unlikely(anon_vma_prepare(vma)))
3265                         return VM_FAULT_OOM;
3266
3267                 cow_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, address);
3268                 if (!cow_page)
3269                         return VM_FAULT_OOM;
3270
3271                 if (mem_cgroup_newpage_charge(cow_page, mm, GFP_KERNEL)) {
3272                         page_cache_release(cow_page);
3273                         return VM_FAULT_OOM;
3274                 }
3275         } else
3276                 cow_page = NULL;
3277
3278         vmf.virtual_address = (void __user *)(address & PAGE_MASK);
3279         vmf.pgoff = pgoff;
3280         vmf.flags = flags;
3281         vmf.page = NULL;
3282
3283         ret = vma->vm_ops->fault(vma, &vmf);
3284         if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE |
3285                             VM_FAULT_RETRY)))
3286                 goto uncharge_out;
3287
3288         if (unlikely(PageHWPoison(vmf.page))) {
3289                 if (ret & VM_FAULT_LOCKED)
3290                         unlock_page(vmf.page);
3291                 ret = VM_FAULT_HWPOISON;
3292                 goto uncharge_out;
3293         }
3294
3295         /*
3296          * For consistency in subsequent calls, make the faulted page always
3297          * locked.
3298          */
3299         if (unlikely(!(ret & VM_FAULT_LOCKED)))
3300                 lock_page(vmf.page);
3301         else
3302                 VM_BUG_ON(!PageLocked(vmf.page));
3303
3304         /*
3305          * Should we do an early C-O-W break?
3306          */
3307         page = vmf.page;
3308         if (flags & FAULT_FLAG_WRITE) {
3309                 if (!(vma->vm_flags & VM_SHARED)) {
3310                         page = cow_page;
3311                         anon = 1;
3312                         copy_user_highpage(page, vmf.page, address, vma);
3313                         __SetPageUptodate(page);
3314                 } else {
3315                         /*
3316                          * If the page will be shareable, see if the backing
3317                          * address space wants to know that the page is about
3318                          * to become writable
3319                          */
3320                         if (vma->vm_ops->page_mkwrite) {
3321                                 int tmp;
3322
3323                                 unlock_page(page);
3324                                 vmf.flags = FAULT_FLAG_WRITE|FAULT_FLAG_MKWRITE;
3325                                 tmp = vma->vm_ops->page_mkwrite(vma, &vmf);
3326                                 if (unlikely(tmp &
3327                                           (VM_FAULT_ERROR | VM_FAULT_NOPAGE))) {
3328                                         ret = tmp;
3329                                         goto unwritable_page;
3330                                 }
3331                                 if (unlikely(!(tmp & VM_FAULT_LOCKED))) {
3332                                         lock_page(page);
3333                                         if (!page->mapping) {
3334                                                 ret = 0; /* retry the fault */
3335                                                 unlock_page(page);
3336                                                 goto unwritable_page;
3337                                         }
3338                                 } else
3339                                         VM_BUG_ON(!PageLocked(page));
3340                                 page_mkwrite = 1;
3341                         }
3342                 }
3343
3344         }
3345
3346         page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
3347
3348         /*
3349          * This silly early PAGE_DIRTY setting removes a race
3350          * due to the bad i386 page protection. But it's valid
3351          * for other architectures too.
3352          *
3353          * Note that if FAULT_FLAG_WRITE is set, we either now have
3354          * an exclusive copy of the page, or this is a shared mapping,
3355          * so we can make it writable and dirty to avoid having to
3356          * handle that later.
3357          */
3358         /* Only go through if we didn't race with anybody else... */
3359         if (likely(pte_same(*page_table, orig_pte))) {
3360                 flush_icache_page(vma, page);
3361                 entry = mk_pte(page, vma->vm_page_prot);
3362                 if (flags & FAULT_FLAG_WRITE)
3363                         entry = maybe_mkwrite(pte_mkdirty(entry), vma);
3364                 if (anon) {
3365                         inc_mm_counter_fast(mm, MM_ANONPAGES);
3366                         page_add_new_anon_rmap(page, vma, address);
3367                 } else {
3368                         inc_mm_counter_fast(mm, MM_FILEPAGES);
3369                         page_add_file_rmap(page);
3370                         if (flags & FAULT_FLAG_WRITE) {
3371                                 dirty_page = page;
3372                                 get_page(dirty_page);
3373                         }
3374                 }
3375                 set_pte_at(mm, address, page_table, entry);
3376
3377                 /* no need to invalidate: a not-present page won't be cached */
3378                 update_mmu_cache(vma, address, page_table);
3379         } else {
3380                 if (cow_page)
3381                         mem_cgroup_uncharge_page(cow_page);
3382                 if (anon)
3383                         page_cache_release(page);
3384                 else
3385                         anon = 1; /* no anon but release faulted_page */
3386         }
3387
3388         pte_unmap_unlock(page_table, ptl);
3389
3390         if (dirty_page) {
3391                 struct address_space *mapping = page->mapping;
3392
3393                 if (set_page_dirty(dirty_page))
3394                         page_mkwrite = 1;
3395                 unlock_page(dirty_page);
3396                 put_page(dirty_page);
3397                 if (page_mkwrite && mapping) {
3398                         /*
3399                          * Some device drivers do not set page.mapping but still
3400                          * dirty their pages
3401                          */
3402                         balance_dirty_pages_ratelimited(mapping);
3403                 }
3404
3405                 /* file_update_time outside page_lock */
3406                 if (vma->vm_file)
3407                         file_update_time(vma->vm_file);
3408         } else {
3409                 unlock_page(vmf.page);
3410                 if (anon)
3411                         page_cache_release(vmf.page);
3412         }
3413
3414         return ret;
3415
3416 unwritable_page:
3417         page_cache_release(page);
3418         return ret;
3419 uncharge_out:
3420         /* fs's fault handler get error */
3421         if (cow_page) {
3422                 mem_cgroup_uncharge_page(cow_page);
3423                 page_cache_release(cow_page);
3424         }
3425         return ret;
3426 }
3427
3428 static int do_linear_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 = (((address & PAGE_MASK)
3433                         - vma->vm_start) >> PAGE_SHIFT) + vma->vm_pgoff;
3434
3435         pte_unmap(page_table);
3436         /* The VMA was not fully populated on mmap() or missing VM_DONTEXPAND */
3437         if (!vma->vm_ops->fault)
3438                 return VM_FAULT_SIGBUS;
3439         return __do_fault(mm, vma, address, pmd, pgoff, flags, orig_pte);
3440 }
3441
3442 /*
3443  * Fault of a previously existing named mapping. Repopulate the pte
3444  * from the encoded file_pte if possible. This enables swappable
3445  * nonlinear vmas.
3446  *
3447  * We enter with non-exclusive mmap_sem (to exclude vma changes,
3448  * but allow concurrent faults), and pte mapped but not yet locked.
3449  * We return with mmap_sem still held, but pte unmapped and unlocked.
3450  */
3451 static int do_nonlinear_fault(struct mm_struct *mm, struct vm_area_struct *vma,
3452                 unsigned long address, pte_t *page_table, pmd_t *pmd,
3453                 unsigned int flags, pte_t orig_pte)
3454 {
3455         pgoff_t pgoff;
3456
3457         flags |= FAULT_FLAG_NONLINEAR;
3458
3459         if (!pte_unmap_same(mm, pmd, page_table, orig_pte))
3460                 return 0;
3461
3462         if (unlikely(!(vma->vm_flags & VM_NONLINEAR))) {
3463                 /*
3464                  * Page table corrupted: show pte and kill process.
3465                  */
3466                 print_bad_pte(vma, address, orig_pte, NULL);
3467                 return VM_FAULT_SIGBUS;
3468         }
3469
3470         pgoff = pte_to_pgoff(orig_pte);
3471         return __do_fault(mm, vma, address, pmd, pgoff, flags, orig_pte);
3472 }
3473
3474 /*
3475  * These routines also need to handle stuff like marking pages dirty
3476  * and/or accessed for architectures that don't do it in hardware (most
3477  * RISC architectures).  The early dirtying is also good on the i386.
3478  *
3479  * There is also a hook called "update_mmu_cache()" that architectures
3480  * with external mmu caches can use to update those (ie the Sparc or
3481  * PowerPC hashed page tables that act as extended TLBs).
3482  *
3483  * We enter with non-exclusive mmap_sem (to exclude vma changes,
3484  * but allow concurrent faults), and pte mapped but not yet locked.
3485  * We return with mmap_sem still held, but pte unmapped and unlocked.
3486  */
3487 int handle_pte_fault(struct mm_struct *mm,
3488                      struct vm_area_struct *vma, unsigned long address,
3489                      pte_t *pte, pmd_t *pmd, unsigned int flags)
3490 {
3491         pte_t entry;
3492         spinlock_t *ptl;
3493
3494         entry = *pte;
3495         if (!pte_present(entry)) {
3496                 if (pte_none(entry)) {
3497                         if (vma->vm_ops)
3498                                 return do_linear_fault(mm, vma, address,
3499                                                 pte, pmd, flags, entry);
3500                         return do_anonymous_page(mm, vma, address,
3501                                                  pte, pmd, flags);
3502                 }
3503                 if (pte_file(entry))
3504                         return do_nonlinear_fault(mm, vma, address,
3505                                         pte, pmd, flags, entry);
3506                 return do_swap_page(mm, vma, address,
3507                                         pte, pmd, flags, entry);
3508         }
3509
3510         ptl = pte_lockptr(mm, pmd);
3511         spin_lock(ptl);
3512         if (unlikely(!pte_same(*pte, entry)))
3513                 goto unlock;
3514         if (flags & FAULT_FLAG_WRITE) {
3515                 if (!pte_write(entry))
3516                         return do_wp_page(mm, vma, address,
3517                                         pte, pmd, ptl, entry);
3518                 entry = pte_mkdirty(entry);
3519         }
3520         entry = pte_mkyoung(entry);
3521         if (ptep_set_access_flags(vma, address, pte, entry, flags & FAULT_FLAG_WRITE)) {
3522                 update_mmu_cache(vma, address, pte);
3523         } else {
3524                 /*
3525                  * This is needed only for protection faults but the arch code
3526                  * is not yet telling us if this is a protection fault or not.
3527                  * This still avoids useless tlb flushes for .text page faults
3528                  * with threads.
3529                  */
3530                 if (flags & FAULT_FLAG_WRITE)
3531                         flush_tlb_fix_spurious_fault(vma, address);
3532         }
3533 unlock:
3534         pte_unmap_unlock(pte, ptl);
3535         return 0;
3536 }
3537
3538 /*
3539  * By the time we get here, we already hold the mm semaphore
3540  */
3541 int handle_mm_fault(struct mm_struct *mm, struct vm_area_struct *vma,
3542                 unsigned long address, unsigned int flags)
3543 {
3544         pgd_t *pgd;
3545         pud_t *pud;
3546         pmd_t *pmd;
3547         pte_t *pte;
3548
3549         __set_current_state(TASK_RUNNING);
3550
3551         count_vm_event(PGFAULT);
3552         mem_cgroup_count_vm_event(mm, PGFAULT);
3553
3554         /* do counter updates before entering really critical section. */
3555         check_sync_rss_stat(current);
3556
3557         if (unlikely(is_vm_hugetlb_page(vma)))
3558                 return hugetlb_fault(mm, vma, address, flags);
3559
3560 retry:
3561         pgd = pgd_offset(mm, address);
3562         pud = pud_alloc(mm, pgd, address);
3563         if (!pud)
3564                 return VM_FAULT_OOM;
3565         pmd = pmd_alloc(mm, pud, address);
3566         if (!pmd)
3567                 return VM_FAULT_OOM;
3568         if (pmd_none(*pmd) && transparent_hugepage_enabled(vma)) {
3569                 if (!vma->vm_ops)
3570                         return do_huge_pmd_anonymous_page(mm, vma, address,
3571                                                           pmd, flags);
3572         } else {
3573                 pmd_t orig_pmd = *pmd;
3574                 int ret;
3575
3576                 barrier();
3577                 if (pmd_trans_huge(orig_pmd)) {
3578                         if (flags & FAULT_FLAG_WRITE &&
3579                             !pmd_write(orig_pmd) &&
3580                             !pmd_trans_splitting(orig_pmd)) {
3581                                 ret = do_huge_pmd_wp_page(mm, vma, address, pmd,
3582                                                           orig_pmd);
3583                                 /*
3584                                  * If COW results in an oom, the huge pmd will
3585                                  * have been split, so retry the fault on the
3586                                  * pte for a smaller charge.
3587                                  */
3588                                 if (unlikely(ret & VM_FAULT_OOM))
3589                                         goto retry;
3590                                 return ret;
3591                         }
3592                         return 0;
3593                 }
3594         }
3595
3596         /*
3597          * Use __pte_alloc instead of pte_alloc_map, because we can't
3598          * run pte_offset_map on the pmd, if an huge pmd could
3599          * materialize from under us from a different thread.
3600          */
3601         if (unlikely(pmd_none(*pmd)) && __pte_alloc(mm, vma, pmd, address))
3602                 return VM_FAULT_OOM;
3603         /*
3604          * If a huge pmd materialized under us just retry later.  Use
3605          * pmd_trans_unstable() instead of pmd_trans_huge() to ensure the pmd
3606          * didn't become pmd_trans_huge under us and then back to pmd_none, as
3607          * a result of MADV_DONTNEED running immediately after a huge pmd fault
3608          * in a different thread of this mm, in turn leading to a misleading
3609          * pmd_trans_huge() retval.  All we have to ensure is that it is a
3610          * regular pmd that we can walk with pte_offset_map() and we can do that
3611          * through an atomic read in C, which is what pmd_trans_unstable()
3612          * provides.
3613          */
3614         if (unlikely(pmd_trans_unstable(pmd)))
3615                 return 0;
3616         /*
3617          * A regular pmd is established and it can't morph into a huge pmd
3618          * from under us anymore at this point because we hold the mmap_sem
3619          * read mode and khugepaged takes it in write mode. So now it's
3620          * safe to run pte_offset_map().
3621          */
3622         pte = pte_offset_map(pmd, address);
3623
3624         return handle_pte_fault(mm, vma, address, pte, pmd, flags);
3625 }
3626
3627 #ifndef __PAGETABLE_PUD_FOLDED
3628 /*
3629  * Allocate page upper directory.
3630  * We've already handled the fast-path in-line.
3631  */
3632 int __pud_alloc(struct mm_struct *mm, pgd_t *pgd, unsigned long address)
3633 {
3634         pud_t *new = pud_alloc_one(mm, address);
3635         if (!new)
3636                 return -ENOMEM;
3637
3638         smp_wmb(); /* See comment in __pte_alloc */
3639
3640         spin_lock(&mm->page_table_lock);
3641         if (pgd_present(*pgd))          /* Another has populated it */
3642                 pud_free(mm, new);
3643         else
3644                 pgd_populate(mm, pgd, new);
3645         spin_unlock(&mm->page_table_lock);
3646         return 0;
3647 }
3648 #endif /* __PAGETABLE_PUD_FOLDED */
3649
3650 #ifndef __PAGETABLE_PMD_FOLDED
3651 /*
3652  * Allocate page middle directory.
3653  * We've already handled the fast-path in-line.
3654  */
3655 int __pmd_alloc(struct mm_struct *mm, pud_t *pud, unsigned long address)
3656 {
3657         pmd_t *new = pmd_alloc_one(mm, address);
3658         if (!new)
3659                 return -ENOMEM;
3660
3661         smp_wmb(); /* See comment in __pte_alloc */
3662
3663         spin_lock(&mm->page_table_lock);
3664 #ifndef __ARCH_HAS_4LEVEL_HACK
3665         if (pud_present(*pud))          /* Another has populated it */
3666                 pmd_free(mm, new);
3667         else
3668                 pud_populate(mm, pud, new);
3669 #else
3670         if (pgd_present(*pud))          /* Another has populated it */
3671                 pmd_free(mm, new);
3672         else
3673                 pgd_populate(mm, pud, new);
3674 #endif /* __ARCH_HAS_4LEVEL_HACK */
3675         spin_unlock(&mm->page_table_lock);
3676         return 0;
3677 }
3678 #endif /* __PAGETABLE_PMD_FOLDED */
3679
3680 int make_pages_present(unsigned long addr, unsigned long end)
3681 {
3682         int ret, len, write;
3683         struct vm_area_struct * vma;
3684
3685         vma = find_vma(current->mm, addr);
3686         if (!vma)
3687                 return -ENOMEM;
3688         /*
3689          * We want to touch writable mappings with a write fault in order
3690          * to break COW, except for shared mappings because these don't COW
3691          * and we would not want to dirty them for nothing.
3692          */
3693         write = (vma->vm_flags & (VM_WRITE | VM_SHARED)) == VM_WRITE;
3694         BUG_ON(addr >= end);
3695         BUG_ON(end > vma->vm_end);
3696         len = DIV_ROUND_UP(end, PAGE_SIZE) - addr/PAGE_SIZE;
3697         ret = get_user_pages(current, current->mm, addr,
3698                         len, write, 0, NULL, NULL);
3699         if (ret < 0)
3700                 return ret;
3701         return ret == len ? 0 : -EFAULT;
3702 }
3703
3704 #if !defined(__HAVE_ARCH_GATE_AREA)
3705
3706 #if defined(AT_SYSINFO_EHDR)
3707 static struct vm_area_struct gate_vma;
3708
3709 static int __init gate_vma_init(void)
3710 {
3711         gate_vma.vm_mm = NULL;
3712         gate_vma.vm_start = FIXADDR_USER_START;
3713         gate_vma.vm_end = FIXADDR_USER_END;
3714         gate_vma.vm_flags = VM_READ | VM_MAYREAD | VM_EXEC | VM_MAYEXEC;
3715         gate_vma.vm_page_prot = __P101;
3716         /*
3717          * Make sure the vDSO gets into every core dump.
3718          * Dumping its contents makes post-mortem fully interpretable later
3719          * without matching up the same kernel and hardware config to see
3720          * what PC values meant.
3721          */
3722         gate_vma.vm_flags |= VM_ALWAYSDUMP;
3723         return 0;
3724 }
3725 __initcall(gate_vma_init);
3726 #endif
3727
3728 struct vm_area_struct *get_gate_vma(struct mm_struct *mm)
3729 {
3730 #ifdef AT_SYSINFO_EHDR
3731         return &gate_vma;
3732 #else
3733         return NULL;
3734 #endif
3735 }
3736
3737 int in_gate_area_no_mm(unsigned long addr)
3738 {
3739 #ifdef AT_SYSINFO_EHDR
3740         if ((addr >= FIXADDR_USER_START) && (addr < FIXADDR_USER_END))
3741                 return 1;
3742 #endif
3743         return 0;
3744 }
3745
3746 #endif  /* __HAVE_ARCH_GATE_AREA */
3747
3748 static int __follow_pte(struct mm_struct *mm, unsigned long address,
3749                 pte_t **ptepp, spinlock_t **ptlp)
3750 {
3751         pgd_t *pgd;
3752         pud_t *pud;
3753         pmd_t *pmd;
3754         pte_t *ptep;
3755
3756         pgd = pgd_offset(mm, address);
3757         if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
3758                 goto out;
3759
3760         pud = pud_offset(pgd, address);
3761         if (pud_none(*pud) || unlikely(pud_bad(*pud)))
3762                 goto out;
3763
3764         pmd = pmd_offset(pud, address);
3765         VM_BUG_ON(pmd_trans_huge(*pmd));
3766         if (pmd_none(*pmd) || unlikely(pmd_bad(*pmd)))
3767                 goto out;
3768
3769         /* We cannot handle huge page PFN maps. Luckily they don't exist. */
3770         if (pmd_huge(*pmd))
3771                 goto out;
3772
3773         ptep = pte_offset_map_lock(mm, pmd, address, ptlp);
3774         if (!ptep)
3775                 goto out;
3776         if (!pte_present(*ptep))
3777                 goto unlock;
3778         *ptepp = ptep;
3779         return 0;
3780 unlock:
3781         pte_unmap_unlock(ptep, *ptlp);
3782 out:
3783         return -EINVAL;
3784 }
3785
3786 static inline int follow_pte(struct mm_struct *mm, unsigned long address,
3787                              pte_t **ptepp, spinlock_t **ptlp)
3788 {
3789         int res;
3790
3791         /* (void) is needed to make gcc happy */
3792         (void) __cond_lock(*ptlp,
3793                            !(res = __follow_pte(mm, address, ptepp, ptlp)));
3794         return res;
3795 }
3796
3797 /**
3798  * follow_pfn - look up PFN at a user virtual address
3799  * @vma: memory mapping
3800  * @address: user virtual address
3801  * @pfn: location to store found PFN
3802  *
3803  * Only IO mappings and raw PFN mappings are allowed.
3804  *
3805  * Returns zero and the pfn at @pfn on success, -ve otherwise.
3806  */
3807 int follow_pfn(struct vm_area_struct *vma, unsigned long address,
3808         unsigned long *pfn)
3809 {
3810         int ret = -EINVAL;
3811         spinlock_t *ptl;
3812         pte_t *ptep;
3813
3814         if (!(vma->vm_flags & (VM_IO | VM_PFNMAP)))
3815                 return ret;
3816
3817         ret = follow_pte(vma->vm_mm, address, &ptep, &ptl);
3818         if (ret)
3819                 return ret;
3820         *pfn = pte_pfn(*ptep);
3821         pte_unmap_unlock(ptep, ptl);
3822         return 0;
3823 }
3824 EXPORT_SYMBOL(follow_pfn);
3825
3826 #ifdef CONFIG_HAVE_IOREMAP_PROT
3827 int follow_phys(struct vm_area_struct *vma,
3828                 unsigned long address, unsigned int flags,
3829                 unsigned long *prot, resource_size_t *phys)
3830 {
3831         int ret = -EINVAL;
3832         pte_t *ptep, pte;
3833         spinlock_t *ptl;
3834
3835         if (!(vma->vm_flags & (VM_IO | VM_PFNMAP)))
3836                 goto out;
3837
3838         if (follow_pte(vma->vm_mm, address, &ptep, &ptl))
3839                 goto out;
3840         pte = *ptep;
3841
3842         if ((flags & FOLL_WRITE) && !pte_write(pte))
3843                 goto unlock;
3844
3845         *prot = pgprot_val(pte_pgprot(pte));
3846         *phys = (resource_size_t)pte_pfn(pte) << PAGE_SHIFT;
3847
3848         ret = 0;
3849 unlock:
3850         pte_unmap_unlock(ptep, ptl);
3851 out:
3852         return ret;
3853 }
3854
3855 int generic_access_phys(struct vm_area_struct *vma, unsigned long addr,
3856                         void *buf, int len, int write)
3857 {
3858         resource_size_t phys_addr;
3859         unsigned long prot = 0;
3860         void __iomem *maddr;
3861         int offset = addr & (PAGE_SIZE-1);
3862
3863         if (follow_phys(vma, addr, write, &prot, &phys_addr))
3864                 return -EINVAL;
3865
3866         maddr = ioremap_prot(phys_addr, PAGE_ALIGN(len + offset), prot);
3867         if (write)
3868                 memcpy_toio(maddr + offset, buf, len);
3869         else
3870                 memcpy_fromio(buf, maddr + offset, len);
3871         iounmap(maddr);
3872
3873         return len;
3874 }
3875 #endif
3876
3877 /*
3878  * Access another process' address space as given in mm.  If non-NULL, use the
3879  * given task for page fault accounting.
3880  */
3881 static int __access_remote_vm(struct task_struct *tsk, struct mm_struct *mm,
3882                 unsigned long addr, void *buf, int len, int write)
3883 {
3884         struct vm_area_struct *vma;
3885         void *old_buf = buf;
3886
3887         down_read(&mm->mmap_sem);
3888         /* ignore errors, just check how much was successfully transferred */
3889         while (len) {
3890                 int bytes, ret, offset;
3891                 void *maddr;
3892                 struct page *page = NULL;
3893
3894                 ret = get_user_pages(tsk, mm, addr, 1,
3895                                 write, 1, &page, &vma);
3896                 if (ret <= 0) {
3897                         /*
3898                          * Check if this is a VM_IO | VM_PFNMAP VMA, which
3899                          * we can access using slightly different code.
3900                          */
3901 #ifdef CONFIG_HAVE_IOREMAP_PROT
3902                         vma = find_vma(mm, addr);
3903                         if (!vma || vma->vm_start > addr)
3904                                 break;
3905                         if (vma->vm_ops && vma->vm_ops->access)
3906                                 ret = vma->vm_ops->access(vma, addr, buf,
3907                                                           len, write);
3908                         if (ret <= 0)
3909 #endif
3910                                 break;
3911                         bytes = ret;
3912                 } else {
3913                         bytes = len;
3914                         offset = addr & (PAGE_SIZE-1);
3915                         if (bytes > PAGE_SIZE-offset)
3916                                 bytes = PAGE_SIZE-offset;
3917
3918                         maddr = kmap(page);
3919                         if (write) {
3920                                 copy_to_user_page(vma, page, addr,
3921                                                   maddr + offset, buf, bytes);
3922                                 set_page_dirty_lock(page);
3923                         } else {
3924                                 copy_from_user_page(vma, page, addr,
3925                                                     buf, maddr + offset, bytes);
3926                         }
3927                         kunmap(page);
3928                         page_cache_release(page);
3929                 }
3930                 len -= bytes;
3931                 buf += bytes;
3932                 addr += bytes;
3933         }
3934         up_read(&mm->mmap_sem);
3935
3936         return buf - old_buf;
3937 }
3938
3939 /**
3940  * access_remote_vm - access another process' address space
3941  * @mm:         the mm_struct of the target address space
3942  * @addr:       start address to access
3943  * @buf:        source or destination buffer
3944  * @len:        number of bytes to transfer
3945  * @write:      whether the access is a write
3946  *
3947  * The caller must hold a reference on @mm.
3948  */
3949 int access_remote_vm(struct mm_struct *mm, unsigned long addr,
3950                 void *buf, int len, int write)
3951 {
3952         return __access_remote_vm(NULL, mm, addr, buf, len, write);
3953 }
3954
3955 /*
3956  * Access another process' address space.
3957  * Source/target buffer must be kernel space,
3958  * Do not walk the page table directly, use get_user_pages
3959  */
3960 int access_process_vm(struct task_struct *tsk, unsigned long addr,
3961                 void *buf, int len, int write)
3962 {
3963         struct mm_struct *mm;
3964         int ret;
3965
3966         mm = get_task_mm(tsk);
3967         if (!mm)
3968                 return 0;
3969
3970         ret = __access_remote_vm(tsk, mm, addr, buf, len, write);
3971         mmput(mm);
3972
3973         return ret;
3974 }
3975
3976 /*
3977  * Print the name of a VMA.
3978  */
3979 void print_vma_addr(char *prefix, unsigned long ip)
3980 {
3981         struct mm_struct *mm = current->mm;
3982         struct vm_area_struct *vma;
3983
3984         /*
3985          * Do not print if we are in atomic
3986          * contexts (in exception stacks, etc.):
3987          */
3988         if (preempt_count())
3989                 return;
3990
3991         down_read(&mm->mmap_sem);
3992         vma = find_vma(mm, ip);
3993         if (vma && vma->vm_file) {
3994                 struct file *f = vma->vm_file;
3995                 char *buf = (char *)__get_free_page(GFP_KERNEL);
3996                 if (buf) {
3997                         char *p, *s;
3998
3999                         p = d_path(&f->f_path, buf, PAGE_SIZE);
4000                         if (IS_ERR(p))
4001                                 p = "?";
4002                         s = strrchr(p, '/');
4003                         if (s)
4004                                 p = s+1;
4005                         printk("%s%s[%lx+%lx]", prefix, p,
4006                                         vma->vm_start,
4007                                         vma->vm_end - vma->vm_start);
4008                         free_page((unsigned long)buf);
4009                 }
4010         }
4011         up_read(&current->mm->mmap_sem);
4012 }
4013
4014 #ifdef CONFIG_PROVE_LOCKING
4015 void might_fault(void)
4016 {
4017         /*
4018          * Some code (nfs/sunrpc) uses socket ops on kernel memory while
4019          * holding the mmap_sem, this is safe because kernel memory doesn't
4020          * get paged out, therefore we'll never actually fault, and the
4021          * below annotations will generate false positives.
4022          */
4023         if (segment_eq(get_fs(), KERNEL_DS))
4024                 return;
4025
4026         might_sleep();
4027         /*
4028          * it would be nicer only to annotate paths which are not under
4029          * pagefault_disable, however that requires a larger audit and
4030          * providing helpers like get_user_atomic.
4031          */
4032         if (!in_atomic() && current->mm)
4033                 might_lock_read(&current->mm->mmap_sem);
4034 }
4035 EXPORT_SYMBOL(might_fault);
4036 #endif
4037
4038 #if defined(CONFIG_TRANSPARENT_HUGEPAGE) || defined(CONFIG_HUGETLBFS)
4039 static void clear_gigantic_page(struct page *page,
4040                                 unsigned long addr,
4041                                 unsigned int pages_per_huge_page)
4042 {
4043         int i;
4044         struct page *p = page;
4045
4046         might_sleep();
4047         for (i = 0; i < pages_per_huge_page;
4048              i++, p = mem_map_next(p, page, i)) {
4049                 cond_resched();
4050                 clear_user_highpage(p, addr + i * PAGE_SIZE);
4051         }
4052 }
4053 void clear_huge_page(struct page *page,
4054                      unsigned long addr, unsigned int pages_per_huge_page)
4055 {
4056         int i;
4057
4058         if (unlikely(pages_per_huge_page > MAX_ORDER_NR_PAGES)) {
4059                 clear_gigantic_page(page, addr, pages_per_huge_page);
4060                 return;
4061         }
4062
4063         might_sleep();
4064         for (i = 0; i < pages_per_huge_page; i++) {
4065                 cond_resched();
4066                 clear_user_highpage(page + i, addr + i * PAGE_SIZE);
4067         }
4068 }
4069
4070 static void copy_user_gigantic_page(struct page *dst, struct page *src,
4071                                     unsigned long addr,
4072                                     struct vm_area_struct *vma,
4073                                     unsigned int pages_per_huge_page)
4074 {
4075         int i;
4076         struct page *dst_base = dst;
4077         struct page *src_base = src;
4078
4079         for (i = 0; i < pages_per_huge_page; ) {
4080                 cond_resched();
4081                 copy_user_highpage(dst, src, addr + i*PAGE_SIZE, vma);
4082
4083                 i++;
4084                 dst = mem_map_next(dst, dst_base, i);
4085                 src = mem_map_next(src, src_base, i);
4086         }
4087 }
4088
4089 void copy_user_huge_page(struct page *dst, struct page *src,
4090                          unsigned long addr, struct vm_area_struct *vma,
4091                          unsigned int pages_per_huge_page)
4092 {
4093         int i;
4094
4095         if (unlikely(pages_per_huge_page > MAX_ORDER_NR_PAGES)) {
4096                 copy_user_gigantic_page(dst, src, addr, vma,
4097                                         pages_per_huge_page);
4098                 return;
4099         }
4100
4101         might_sleep();
4102         for (i = 0; i < pages_per_huge_page; i++) {
4103                 cond_resched();
4104                 copy_user_highpage(dst + i, src + i, addr + i*PAGE_SIZE, vma);
4105         }
4106 }
4107 #endif /* CONFIG_TRANSPARENT_HUGEPAGE || CONFIG_HUGETLBFS */