VM: unmap_page_range() can return void
[linux-2.6.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         batch = (void *)__get_free_pages(GFP_NOWAIT | __GFP_NOWARN, 0);
209         if (!batch)
210                 return 0;
211
212         batch->next = NULL;
213         batch->nr   = 0;
214         batch->max  = MAX_GATHER_BATCH;
215
216         tlb->active->next = batch;
217         tlb->active = batch;
218
219         return 1;
220 }
221
222 /* tlb_gather_mmu
223  *      Called to initialize an (on-stack) mmu_gather structure for page-table
224  *      tear-down from @mm. The @fullmm argument is used when @mm is without
225  *      users and we're going to destroy the full address space (exit/execve).
226  */
227 void tlb_gather_mmu(struct mmu_gather *tlb, struct mm_struct *mm, bool fullmm)
228 {
229         tlb->mm = mm;
230
231         tlb->fullmm     = fullmm;
232         tlb->need_flush = 0;
233         tlb->fast_mode  = (num_possible_cpus() == 1);
234         tlb->local.next = NULL;
235         tlb->local.nr   = 0;
236         tlb->local.max  = ARRAY_SIZE(tlb->__pages);
237         tlb->active     = &tlb->local;
238
239 #ifdef CONFIG_HAVE_RCU_TABLE_FREE
240         tlb->batch = NULL;
241 #endif
242 }
243
244 void tlb_flush_mmu(struct mmu_gather *tlb)
245 {
246         struct mmu_gather_batch *batch;
247
248         if (!tlb->need_flush)
249                 return;
250         tlb->need_flush = 0;
251         tlb_flush(tlb);
252 #ifdef CONFIG_HAVE_RCU_TABLE_FREE
253         tlb_table_flush(tlb);
254 #endif
255
256         if (tlb_fast_mode(tlb))
257                 return;
258
259         for (batch = &tlb->local; batch; batch = batch->next) {
260                 free_pages_and_swap_cache(batch->pages, batch->nr);
261                 batch->nr = 0;
262         }
263         tlb->active = &tlb->local;
264 }
265
266 /* tlb_finish_mmu
267  *      Called at the end of the shootdown operation to free up any resources
268  *      that were required.
269  */
270 void tlb_finish_mmu(struct mmu_gather *tlb, unsigned long start, unsigned long end)
271 {
272         struct mmu_gather_batch *batch, *next;
273
274         tlb_flush_mmu(tlb);
275
276         /* keep the page table cache within bounds */
277         check_pgt_cache();
278
279         for (batch = tlb->local.next; batch; batch = next) {
280                 next = batch->next;
281                 free_pages((unsigned long)batch, 0);
282         }
283         tlb->local.next = NULL;
284 }
285
286 /* __tlb_remove_page
287  *      Must perform the equivalent to __free_pte(pte_get_and_clear(ptep)), while
288  *      handling the additional races in SMP caused by other CPUs caching valid
289  *      mappings in their TLBs. Returns the number of free page slots left.
290  *      When out of page slots we must call tlb_flush_mmu().
291  */
292 int __tlb_remove_page(struct mmu_gather *tlb, struct page *page)
293 {
294         struct mmu_gather_batch *batch;
295
296         VM_BUG_ON(!tlb->need_flush);
297
298         if (tlb_fast_mode(tlb)) {
299                 free_page_and_swap_cache(page);
300                 return 1; /* avoid calling tlb_flush_mmu() */
301         }
302
303         batch = tlb->active;
304         batch->pages[batch->nr++] = page;
305         if (batch->nr == batch->max) {
306                 if (!tlb_next_batch(tlb))
307                         return 0;
308                 batch = tlb->active;
309         }
310         VM_BUG_ON(batch->nr > batch->max);
311
312         return batch->max - batch->nr;
313 }
314
315 #endif /* HAVE_GENERIC_MMU_GATHER */
316
317 #ifdef CONFIG_HAVE_RCU_TABLE_FREE
318
319 /*
320  * See the comment near struct mmu_table_batch.
321  */
322
323 static void tlb_remove_table_smp_sync(void *arg)
324 {
325         /* Simply deliver the interrupt */
326 }
327
328 static void tlb_remove_table_one(void *table)
329 {
330         /*
331          * This isn't an RCU grace period and hence the page-tables cannot be
332          * assumed to be actually RCU-freed.
333          *
334          * It is however sufficient for software page-table walkers that rely on
335          * IRQ disabling. See the comment near struct mmu_table_batch.
336          */
337         smp_call_function(tlb_remove_table_smp_sync, NULL, 1);
338         __tlb_remove_table(table);
339 }
340
341 static void tlb_remove_table_rcu(struct rcu_head *head)
342 {
343         struct mmu_table_batch *batch;
344         int i;
345
346         batch = container_of(head, struct mmu_table_batch, rcu);
347
348         for (i = 0; i < batch->nr; i++)
349                 __tlb_remove_table(batch->tables[i]);
350
351         free_page((unsigned long)batch);
352 }
353
354 void tlb_table_flush(struct mmu_gather *tlb)
355 {
356         struct mmu_table_batch **batch = &tlb->batch;
357
358         if (*batch) {
359                 call_rcu_sched(&(*batch)->rcu, tlb_remove_table_rcu);
360                 *batch = NULL;
361         }
362 }
363
364 void tlb_remove_table(struct mmu_gather *tlb, void *table)
365 {
366         struct mmu_table_batch **batch = &tlb->batch;
367
368         tlb->need_flush = 1;
369
370         /*
371          * When there's less then two users of this mm there cannot be a
372          * concurrent page-table walk.
373          */
374         if (atomic_read(&tlb->mm->mm_users) < 2) {
375                 __tlb_remove_table(table);
376                 return;
377         }
378
379         if (*batch == NULL) {
380                 *batch = (struct mmu_table_batch *)__get_free_page(GFP_NOWAIT | __GFP_NOWARN);
381                 if (*batch == NULL) {
382                         tlb_remove_table_one(table);
383                         return;
384                 }
385                 (*batch)->nr = 0;
386         }
387         (*batch)->tables[(*batch)->nr++] = table;
388         if ((*batch)->nr == MAX_TABLE_BATCH)
389                 tlb_table_flush(tlb);
390 }
391
392 #endif /* CONFIG_HAVE_RCU_TABLE_FREE */
393
394 /*
395  * If a p?d_bad entry is found while walking page tables, report
396  * the error, before resetting entry to p?d_none.  Usually (but
397  * very seldom) called out from the p?d_none_or_clear_bad macros.
398  */
399
400 void pgd_clear_bad(pgd_t *pgd)
401 {
402         pgd_ERROR(*pgd);
403         pgd_clear(pgd);
404 }
405
406 void pud_clear_bad(pud_t *pud)
407 {
408         pud_ERROR(*pud);
409         pud_clear(pud);
410 }
411
412 void pmd_clear_bad(pmd_t *pmd)
413 {
414         pmd_ERROR(*pmd);
415         pmd_clear(pmd);
416 }
417
418 /*
419  * Note: this doesn't free the actual pages themselves. That
420  * has been handled earlier when unmapping all the memory regions.
421  */
422 static void free_pte_range(struct mmu_gather *tlb, pmd_t *pmd,
423                            unsigned long addr)
424 {
425         pgtable_t token = pmd_pgtable(*pmd);
426         pmd_clear(pmd);
427         pte_free_tlb(tlb, token, addr);
428         tlb->mm->nr_ptes--;
429 }
430
431 static inline void free_pmd_range(struct mmu_gather *tlb, pud_t *pud,
432                                 unsigned long addr, unsigned long end,
433                                 unsigned long floor, unsigned long ceiling)
434 {
435         pmd_t *pmd;
436         unsigned long next;
437         unsigned long start;
438
439         start = addr;
440         pmd = pmd_offset(pud, addr);
441         do {
442                 next = pmd_addr_end(addr, end);
443                 if (pmd_none_or_clear_bad(pmd))
444                         continue;
445                 free_pte_range(tlb, pmd, addr);
446         } while (pmd++, addr = next, addr != end);
447
448         start &= PUD_MASK;
449         if (start < floor)
450                 return;
451         if (ceiling) {
452                 ceiling &= PUD_MASK;
453                 if (!ceiling)
454                         return;
455         }
456         if (end - 1 > ceiling - 1)
457                 return;
458
459         pmd = pmd_offset(pud, start);
460         pud_clear(pud);
461         pmd_free_tlb(tlb, pmd, start);
462 }
463
464 static inline void free_pud_range(struct mmu_gather *tlb, pgd_t *pgd,
465                                 unsigned long addr, unsigned long end,
466                                 unsigned long floor, unsigned long ceiling)
467 {
468         pud_t *pud;
469         unsigned long next;
470         unsigned long start;
471
472         start = addr;
473         pud = pud_offset(pgd, addr);
474         do {
475                 next = pud_addr_end(addr, end);
476                 if (pud_none_or_clear_bad(pud))
477                         continue;
478                 free_pmd_range(tlb, pud, addr, next, floor, ceiling);
479         } while (pud++, addr = next, addr != end);
480
481         start &= PGDIR_MASK;
482         if (start < floor)
483                 return;
484         if (ceiling) {
485                 ceiling &= PGDIR_MASK;
486                 if (!ceiling)
487                         return;
488         }
489         if (end - 1 > ceiling - 1)
490                 return;
491
492         pud = pud_offset(pgd, start);
493         pgd_clear(pgd);
494         pud_free_tlb(tlb, pud, start);
495 }
496
497 /*
498  * This function frees user-level page tables of a process.
499  *
500  * Must be called with pagetable lock held.
501  */
502 void free_pgd_range(struct mmu_gather *tlb,
503                         unsigned long addr, unsigned long end,
504                         unsigned long floor, unsigned long ceiling)
505 {
506         pgd_t *pgd;
507         unsigned long next;
508
509         /*
510          * The next few lines have given us lots of grief...
511          *
512          * Why are we testing PMD* at this top level?  Because often
513          * there will be no work to do at all, and we'd prefer not to
514          * go all the way down to the bottom just to discover that.
515          *
516          * Why all these "- 1"s?  Because 0 represents both the bottom
517          * of the address space and the top of it (using -1 for the
518          * top wouldn't help much: the masks would do the wrong thing).
519          * The rule is that addr 0 and floor 0 refer to the bottom of
520          * the address space, but end 0 and ceiling 0 refer to the top
521          * Comparisons need to use "end - 1" and "ceiling - 1" (though
522          * that end 0 case should be mythical).
523          *
524          * Wherever addr is brought up or ceiling brought down, we must
525          * be careful to reject "the opposite 0" before it confuses the
526          * subsequent tests.  But what about where end is brought down
527          * by PMD_SIZE below? no, end can't go down to 0 there.
528          *
529          * Whereas we round start (addr) and ceiling down, by different
530          * masks at different levels, in order to test whether a table
531          * now has no other vmas using it, so can be freed, we don't
532          * bother to round floor or end up - the tests don't need that.
533          */
534
535         addr &= PMD_MASK;
536         if (addr < floor) {
537                 addr += PMD_SIZE;
538                 if (!addr)
539                         return;
540         }
541         if (ceiling) {
542                 ceiling &= PMD_MASK;
543                 if (!ceiling)
544                         return;
545         }
546         if (end - 1 > ceiling - 1)
547                 end -= PMD_SIZE;
548         if (addr > end - 1)
549                 return;
550
551         pgd = pgd_offset(tlb->mm, addr);
552         do {
553                 next = pgd_addr_end(addr, end);
554                 if (pgd_none_or_clear_bad(pgd))
555                         continue;
556                 free_pud_range(tlb, pgd, addr, next, floor, ceiling);
557         } while (pgd++, addr = next, addr != end);
558 }
559
560 void free_pgtables(struct mmu_gather *tlb, struct vm_area_struct *vma,
561                 unsigned long floor, unsigned long ceiling)
562 {
563         while (vma) {
564                 struct vm_area_struct *next = vma->vm_next;
565                 unsigned long addr = vma->vm_start;
566
567                 /*
568                  * Hide vma from rmap and truncate_pagecache before freeing
569                  * pgtables
570                  */
571                 unlink_anon_vmas(vma);
572                 unlink_file_vma(vma);
573
574                 if (is_vm_hugetlb_page(vma)) {
575                         hugetlb_free_pgd_range(tlb, addr, vma->vm_end,
576                                 floor, next? next->vm_start: ceiling);
577                 } else {
578                         /*
579                          * Optimization: gather nearby vmas into one call down
580                          */
581                         while (next && next->vm_start <= vma->vm_end + PMD_SIZE
582                                && !is_vm_hugetlb_page(next)) {
583                                 vma = next;
584                                 next = vma->vm_next;
585                                 unlink_anon_vmas(vma);
586                                 unlink_file_vma(vma);
587                         }
588                         free_pgd_range(tlb, addr, vma->vm_end,
589                                 floor, next? next->vm_start: ceiling);
590                 }
591                 vma = next;
592         }
593 }
594
595 int __pte_alloc(struct mm_struct *mm, struct vm_area_struct *vma,
596                 pmd_t *pmd, unsigned long address)
597 {
598         pgtable_t new = pte_alloc_one(mm, address);
599         int wait_split_huge_page;
600         if (!new)
601                 return -ENOMEM;
602
603         /*
604          * Ensure all pte setup (eg. pte page lock and page clearing) are
605          * visible before the pte is made visible to other CPUs by being
606          * put into page tables.
607          *
608          * The other side of the story is the pointer chasing in the page
609          * table walking code (when walking the page table without locking;
610          * ie. most of the time). Fortunately, these data accesses consist
611          * of a chain of data-dependent loads, meaning most CPUs (alpha
612          * being the notable exception) will already guarantee loads are
613          * seen in-order. See the alpha page table accessors for the
614          * smp_read_barrier_depends() barriers in page table walking code.
615          */
616         smp_wmb(); /* Could be smp_wmb__xxx(before|after)_spin_lock */
617
618         spin_lock(&mm->page_table_lock);
619         wait_split_huge_page = 0;
620         if (likely(pmd_none(*pmd))) {   /* Has another populated it ? */
621                 mm->nr_ptes++;
622                 pmd_populate(mm, pmd, new);
623                 new = NULL;
624         } else if (unlikely(pmd_trans_splitting(*pmd)))
625                 wait_split_huge_page = 1;
626         spin_unlock(&mm->page_table_lock);
627         if (new)
628                 pte_free(mm, new);
629         if (wait_split_huge_page)
630                 wait_split_huge_page(vma->anon_vma, pmd);
631         return 0;
632 }
633
634 int __pte_alloc_kernel(pmd_t *pmd, unsigned long address)
635 {
636         pte_t *new = pte_alloc_one_kernel(&init_mm, address);
637         if (!new)
638                 return -ENOMEM;
639
640         smp_wmb(); /* See comment in __pte_alloc */
641
642         spin_lock(&init_mm.page_table_lock);
643         if (likely(pmd_none(*pmd))) {   /* Has another populated it ? */
644                 pmd_populate_kernel(&init_mm, pmd, new);
645                 new = NULL;
646         } else
647                 VM_BUG_ON(pmd_trans_splitting(*pmd));
648         spin_unlock(&init_mm.page_table_lock);
649         if (new)
650                 pte_free_kernel(&init_mm, new);
651         return 0;
652 }
653
654 static inline void init_rss_vec(int *rss)
655 {
656         memset(rss, 0, sizeof(int) * NR_MM_COUNTERS);
657 }
658
659 static inline void add_mm_rss_vec(struct mm_struct *mm, int *rss)
660 {
661         int i;
662
663         if (current->mm == mm)
664                 sync_mm_rss(current, mm);
665         for (i = 0; i < NR_MM_COUNTERS; i++)
666                 if (rss[i])
667                         add_mm_counter(mm, i, rss[i]);
668 }
669
670 /*
671  * This function is called to print an error when a bad pte
672  * is found. For example, we might have a PFN-mapped pte in
673  * a region that doesn't allow it.
674  *
675  * The calling function must still handle the error.
676  */
677 static void print_bad_pte(struct vm_area_struct *vma, unsigned long addr,
678                           pte_t pte, struct page *page)
679 {
680         pgd_t *pgd = pgd_offset(vma->vm_mm, addr);
681         pud_t *pud = pud_offset(pgd, addr);
682         pmd_t *pmd = pmd_offset(pud, addr);
683         struct address_space *mapping;
684         pgoff_t index;
685         static unsigned long resume;
686         static unsigned long nr_shown;
687         static unsigned long nr_unshown;
688
689         /*
690          * Allow a burst of 60 reports, then keep quiet for that minute;
691          * or allow a steady drip of one report per second.
692          */
693         if (nr_shown == 60) {
694                 if (time_before(jiffies, resume)) {
695                         nr_unshown++;
696                         return;
697                 }
698                 if (nr_unshown) {
699                         printk(KERN_ALERT
700                                 "BUG: Bad page map: %lu messages suppressed\n",
701                                 nr_unshown);
702                         nr_unshown = 0;
703                 }
704                 nr_shown = 0;
705         }
706         if (nr_shown++ == 0)
707                 resume = jiffies + 60 * HZ;
708
709         mapping = vma->vm_file ? vma->vm_file->f_mapping : NULL;
710         index = linear_page_index(vma, addr);
711
712         printk(KERN_ALERT
713                 "BUG: Bad page map in process %s  pte:%08llx pmd:%08llx\n",
714                 current->comm,
715                 (long long)pte_val(pte), (long long)pmd_val(*pmd));
716         if (page)
717                 dump_page(page);
718         printk(KERN_ALERT
719                 "addr:%p vm_flags:%08lx anon_vma:%p mapping:%p index:%lx\n",
720                 (void *)addr, vma->vm_flags, vma->anon_vma, mapping, index);
721         /*
722          * Choose text because data symbols depend on CONFIG_KALLSYMS_ALL=y
723          */
724         if (vma->vm_ops)
725                 print_symbol(KERN_ALERT "vma->vm_ops->fault: %s\n",
726                                 (unsigned long)vma->vm_ops->fault);
727         if (vma->vm_file && vma->vm_file->f_op)
728                 print_symbol(KERN_ALERT "vma->vm_file->f_op->mmap: %s\n",
729                                 (unsigned long)vma->vm_file->f_op->mmap);
730         dump_stack();
731         add_taint(TAINT_BAD_PAGE);
732 }
733
734 static inline int is_cow_mapping(vm_flags_t flags)
735 {
736         return (flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE;
737 }
738
739 #ifndef is_zero_pfn
740 static inline int is_zero_pfn(unsigned long pfn)
741 {
742         return pfn == zero_pfn;
743 }
744 #endif
745
746 #ifndef my_zero_pfn
747 static inline unsigned long my_zero_pfn(unsigned long addr)
748 {
749         return zero_pfn;
750 }
751 #endif
752
753 /*
754  * vm_normal_page -- This function gets the "struct page" associated with a pte.
755  *
756  * "Special" mappings do not wish to be associated with a "struct page" (either
757  * it doesn't exist, or it exists but they don't want to touch it). In this
758  * case, NULL is returned here. "Normal" mappings do have a struct page.
759  *
760  * There are 2 broad cases. Firstly, an architecture may define a pte_special()
761  * pte bit, in which case this function is trivial. Secondly, an architecture
762  * may not have a spare pte bit, which requires a more complicated scheme,
763  * described below.
764  *
765  * A raw VM_PFNMAP mapping (ie. one that is not COWed) is always considered a
766  * special mapping (even if there are underlying and valid "struct pages").
767  * COWed pages of a VM_PFNMAP are always normal.
768  *
769  * The way we recognize COWed pages within VM_PFNMAP mappings is through the
770  * rules set up by "remap_pfn_range()": the vma will have the VM_PFNMAP bit
771  * set, and the vm_pgoff will point to the first PFN mapped: thus every special
772  * mapping will always honor the rule
773  *
774  *      pfn_of_page == vma->vm_pgoff + ((addr - vma->vm_start) >> PAGE_SHIFT)
775  *
776  * And for normal mappings this is false.
777  *
778  * This restricts such mappings to be a linear translation from virtual address
779  * to pfn. To get around this restriction, we allow arbitrary mappings so long
780  * as the vma is not a COW mapping; in that case, we know that all ptes are
781  * special (because none can have been COWed).
782  *
783  *
784  * In order to support COW of arbitrary special mappings, we have VM_MIXEDMAP.
785  *
786  * VM_MIXEDMAP mappings can likewise contain memory with or without "struct
787  * page" backing, however the difference is that _all_ pages with a struct
788  * page (that is, those where pfn_valid is true) are refcounted and considered
789  * normal pages by the VM. The disadvantage is that pages are refcounted
790  * (which can be slower and simply not an option for some PFNMAP users). The
791  * advantage is that we don't have to follow the strict linearity rule of
792  * PFNMAP mappings in order to support COWable mappings.
793  *
794  */
795 #ifdef __HAVE_ARCH_PTE_SPECIAL
796 # define HAVE_PTE_SPECIAL 1
797 #else
798 # define HAVE_PTE_SPECIAL 0
799 #endif
800 struct page *vm_normal_page(struct vm_area_struct *vma, unsigned long addr,
801                                 pte_t pte)
802 {
803         unsigned long pfn = pte_pfn(pte);
804
805         if (HAVE_PTE_SPECIAL) {
806                 if (likely(!pte_special(pte)))
807                         goto check_pfn;
808                 if (vma->vm_flags & (VM_PFNMAP | VM_MIXEDMAP))
809                         return NULL;
810                 if (!is_zero_pfn(pfn))
811                         print_bad_pte(vma, addr, pte, NULL);
812                 return NULL;
813         }
814
815         /* !HAVE_PTE_SPECIAL case follows: */
816
817         if (unlikely(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP))) {
818                 if (vma->vm_flags & VM_MIXEDMAP) {
819                         if (!pfn_valid(pfn))
820                                 return NULL;
821                         goto out;
822                 } else {
823                         unsigned long off;
824                         off = (addr - vma->vm_start) >> PAGE_SHIFT;
825                         if (pfn == vma->vm_pgoff + off)
826                                 return NULL;
827                         if (!is_cow_mapping(vma->vm_flags))
828                                 return NULL;
829                 }
830         }
831
832         if (is_zero_pfn(pfn))
833                 return NULL;
834 check_pfn:
835         if (unlikely(pfn > highest_memmap_pfn)) {
836                 print_bad_pte(vma, addr, pte, NULL);
837                 return NULL;
838         }
839
840         /*
841          * NOTE! We still have PageReserved() pages in the page tables.
842          * eg. VDSO mappings can cause them to exist.
843          */
844 out:
845         return pfn_to_page(pfn);
846 }
847
848 /*
849  * copy one vm_area from one task to the other. Assumes the page tables
850  * already present in the new task to be cleared in the whole range
851  * covered by this vma.
852  */
853
854 static inline unsigned long
855 copy_one_pte(struct mm_struct *dst_mm, struct mm_struct *src_mm,
856                 pte_t *dst_pte, pte_t *src_pte, struct vm_area_struct *vma,
857                 unsigned long addr, int *rss)
858 {
859         unsigned long vm_flags = vma->vm_flags;
860         pte_t pte = *src_pte;
861         struct page *page;
862
863         /* pte contains position in swap or file, so copy. */
864         if (unlikely(!pte_present(pte))) {
865                 if (!pte_file(pte)) {
866                         swp_entry_t entry = pte_to_swp_entry(pte);
867
868                         if (swap_duplicate(entry) < 0)
869                                 return entry.val;
870
871                         /* make sure dst_mm is on swapoff's mmlist. */
872                         if (unlikely(list_empty(&dst_mm->mmlist))) {
873                                 spin_lock(&mmlist_lock);
874                                 if (list_empty(&dst_mm->mmlist))
875                                         list_add(&dst_mm->mmlist,
876                                                  &src_mm->mmlist);
877                                 spin_unlock(&mmlist_lock);
878                         }
879                         if (likely(!non_swap_entry(entry)))
880                                 rss[MM_SWAPENTS]++;
881                         else if (is_migration_entry(entry)) {
882                                 page = migration_entry_to_page(entry);
883
884                                 if (PageAnon(page))
885                                         rss[MM_ANONPAGES]++;
886                                 else
887                                         rss[MM_FILEPAGES]++;
888
889                                 if (is_write_migration_entry(entry) &&
890                                     is_cow_mapping(vm_flags)) {
891                                         /*
892                                          * COW mappings require pages in both
893                                          * parent and child to be set to read.
894                                          */
895                                         make_migration_entry_read(&entry);
896                                         pte = swp_entry_to_pte(entry);
897                                         set_pte_at(src_mm, addr, src_pte, pte);
898                                 }
899                         }
900                 }
901                 goto out_set_pte;
902         }
903
904         /*
905          * If it's a COW mapping, write protect it both
906          * in the parent and the child
907          */
908         if (is_cow_mapping(vm_flags)) {
909                 ptep_set_wrprotect(src_mm, addr, src_pte);
910                 pte = pte_wrprotect(pte);
911         }
912
913         /*
914          * If it's a shared mapping, mark it clean in
915          * the child
916          */
917         if (vm_flags & VM_SHARED)
918                 pte = pte_mkclean(pte);
919         pte = pte_mkold(pte);
920
921         page = vm_normal_page(vma, addr, pte);
922         if (page) {
923                 get_page(page);
924                 page_dup_rmap(page);
925                 if (PageAnon(page))
926                         rss[MM_ANONPAGES]++;
927                 else
928                         rss[MM_FILEPAGES]++;
929         }
930
931 out_set_pte:
932         set_pte_at(dst_mm, addr, dst_pte, pte);
933         return 0;
934 }
935
936 int copy_pte_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
937                    pmd_t *dst_pmd, pmd_t *src_pmd, struct vm_area_struct *vma,
938                    unsigned long addr, unsigned long end)
939 {
940         pte_t *orig_src_pte, *orig_dst_pte;
941         pte_t *src_pte, *dst_pte;
942         spinlock_t *src_ptl, *dst_ptl;
943         int progress = 0;
944         int rss[NR_MM_COUNTERS];
945         swp_entry_t entry = (swp_entry_t){0};
946
947 again:
948         init_rss_vec(rss);
949
950         dst_pte = pte_alloc_map_lock(dst_mm, dst_pmd, addr, &dst_ptl);
951         if (!dst_pte)
952                 return -ENOMEM;
953         src_pte = pte_offset_map(src_pmd, addr);
954         src_ptl = pte_lockptr(src_mm, src_pmd);
955         spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
956         orig_src_pte = src_pte;
957         orig_dst_pte = dst_pte;
958         arch_enter_lazy_mmu_mode();
959
960         do {
961                 /*
962                  * We are holding two locks at this point - either of them
963                  * could generate latencies in another task on another CPU.
964                  */
965                 if (progress >= 32) {
966                         progress = 0;
967                         if (need_resched() ||
968                             spin_needbreak(src_ptl) || spin_needbreak(dst_ptl))
969                                 break;
970                 }
971                 if (pte_none(*src_pte)) {
972                         progress++;
973                         continue;
974                 }
975                 entry.val = copy_one_pte(dst_mm, src_mm, dst_pte, src_pte,
976                                                         vma, addr, rss);
977                 if (entry.val)
978                         break;
979                 progress += 8;
980         } while (dst_pte++, src_pte++, addr += PAGE_SIZE, addr != end);
981
982         arch_leave_lazy_mmu_mode();
983         spin_unlock(src_ptl);
984         pte_unmap(orig_src_pte);
985         add_mm_rss_vec(dst_mm, rss);
986         pte_unmap_unlock(orig_dst_pte, dst_ptl);
987         cond_resched();
988
989         if (entry.val) {
990                 if (add_swap_count_continuation(entry, GFP_KERNEL) < 0)
991                         return -ENOMEM;
992                 progress = 0;
993         }
994         if (addr != end)
995                 goto again;
996         return 0;
997 }
998
999 static inline int copy_pmd_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
1000                 pud_t *dst_pud, pud_t *src_pud, struct vm_area_struct *vma,
1001                 unsigned long addr, unsigned long end)
1002 {
1003         pmd_t *src_pmd, *dst_pmd;
1004         unsigned long next;
1005
1006         dst_pmd = pmd_alloc(dst_mm, dst_pud, addr);
1007         if (!dst_pmd)
1008                 return -ENOMEM;
1009         src_pmd = pmd_offset(src_pud, addr);
1010         do {
1011                 next = pmd_addr_end(addr, end);
1012                 if (pmd_trans_huge(*src_pmd)) {
1013                         int err;
1014                         VM_BUG_ON(next-addr != HPAGE_PMD_SIZE);
1015                         err = copy_huge_pmd(dst_mm, src_mm,
1016                                             dst_pmd, src_pmd, addr, vma);
1017                         if (err == -ENOMEM)
1018                                 return -ENOMEM;
1019                         if (!err)
1020                                 continue;
1021                         /* fall through */
1022                 }
1023                 if (pmd_none_or_clear_bad(src_pmd))
1024                         continue;
1025                 if (copy_pte_range(dst_mm, src_mm, dst_pmd, src_pmd,
1026                                                 vma, addr, next))
1027                         return -ENOMEM;
1028         } while (dst_pmd++, src_pmd++, addr = next, addr != end);
1029         return 0;
1030 }
1031
1032 static inline int copy_pud_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
1033                 pgd_t *dst_pgd, pgd_t *src_pgd, struct vm_area_struct *vma,
1034                 unsigned long addr, unsigned long end)
1035 {
1036         pud_t *src_pud, *dst_pud;
1037         unsigned long next;
1038
1039         dst_pud = pud_alloc(dst_mm, dst_pgd, addr);
1040         if (!dst_pud)
1041                 return -ENOMEM;
1042         src_pud = pud_offset(src_pgd, addr);
1043         do {
1044                 next = pud_addr_end(addr, end);
1045                 if (pud_none_or_clear_bad(src_pud))
1046                         continue;
1047                 if (copy_pmd_range(dst_mm, src_mm, dst_pud, src_pud,
1048                                                 vma, addr, next))
1049                         return -ENOMEM;
1050         } while (dst_pud++, src_pud++, addr = next, addr != end);
1051         return 0;
1052 }
1053
1054 int copy_page_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
1055                 struct vm_area_struct *vma)
1056 {
1057         pgd_t *src_pgd, *dst_pgd;
1058         unsigned long next;
1059         unsigned long addr = vma->vm_start;
1060         unsigned long end = vma->vm_end;
1061         int ret;
1062
1063         /*
1064          * Don't copy ptes where a page fault will fill them correctly.
1065          * Fork becomes much lighter when there are big shared or private
1066          * readonly mappings. The tradeoff is that copy_page_range is more
1067          * efficient than faulting.
1068          */
1069         if (!(vma->vm_flags & (VM_HUGETLB|VM_NONLINEAR|VM_PFNMAP|VM_INSERTPAGE))) {
1070                 if (!vma->anon_vma)
1071                         return 0;
1072         }
1073
1074         if (is_vm_hugetlb_page(vma))
1075                 return copy_hugetlb_page_range(dst_mm, src_mm, vma);
1076
1077         if (unlikely(is_pfn_mapping(vma))) {
1078                 /*
1079                  * We do not free on error cases below as remove_vma
1080                  * gets called on error from higher level routine
1081                  */
1082                 ret = track_pfn_vma_copy(vma);
1083                 if (ret)
1084                         return ret;
1085         }
1086
1087         /*
1088          * We need to invalidate the secondary MMU mappings only when
1089          * there could be a permission downgrade on the ptes of the
1090          * parent mm. And a permission downgrade will only happen if
1091          * is_cow_mapping() returns true.
1092          */
1093         if (is_cow_mapping(vma->vm_flags))
1094                 mmu_notifier_invalidate_range_start(src_mm, addr, end);
1095
1096         ret = 0;
1097         dst_pgd = pgd_offset(dst_mm, addr);
1098         src_pgd = pgd_offset(src_mm, addr);
1099         do {
1100                 next = pgd_addr_end(addr, end);
1101                 if (pgd_none_or_clear_bad(src_pgd))
1102                         continue;
1103                 if (unlikely(copy_pud_range(dst_mm, src_mm, dst_pgd, src_pgd,
1104                                             vma, addr, next))) {
1105                         ret = -ENOMEM;
1106                         break;
1107                 }
1108         } while (dst_pgd++, src_pgd++, addr = next, addr != end);
1109
1110         if (is_cow_mapping(vma->vm_flags))
1111                 mmu_notifier_invalidate_range_end(src_mm,
1112                                                   vma->vm_start, end);
1113         return ret;
1114 }
1115
1116 static unsigned long zap_pte_range(struct mmu_gather *tlb,
1117                                 struct vm_area_struct *vma, pmd_t *pmd,
1118                                 unsigned long addr, unsigned long end,
1119                                 struct zap_details *details)
1120 {
1121         struct mm_struct *mm = tlb->mm;
1122         int force_flush = 0;
1123         int rss[NR_MM_COUNTERS];
1124         spinlock_t *ptl;
1125         pte_t *start_pte;
1126         pte_t *pte;
1127
1128 again:
1129         init_rss_vec(rss);
1130         start_pte = pte_offset_map_lock(mm, pmd, addr, &ptl);
1131         pte = start_pte;
1132         arch_enter_lazy_mmu_mode();
1133         do {
1134                 pte_t ptent = *pte;
1135                 if (pte_none(ptent)) {
1136                         continue;
1137                 }
1138
1139                 if (pte_present(ptent)) {
1140                         struct page *page;
1141
1142                         page = vm_normal_page(vma, addr, ptent);
1143                         if (unlikely(details) && page) {
1144                                 /*
1145                                  * unmap_shared_mapping_pages() wants to
1146                                  * invalidate cache without truncating:
1147                                  * unmap shared but keep private pages.
1148                                  */
1149                                 if (details->check_mapping &&
1150                                     details->check_mapping != page->mapping)
1151                                         continue;
1152                                 /*
1153                                  * Each page->index must be checked when
1154                                  * invalidating or truncating nonlinear.
1155                                  */
1156                                 if (details->nonlinear_vma &&
1157                                     (page->index < details->first_index ||
1158                                      page->index > details->last_index))
1159                                         continue;
1160                         }
1161                         ptent = ptep_get_and_clear_full(mm, addr, pte,
1162                                                         tlb->fullmm);
1163                         tlb_remove_tlb_entry(tlb, pte, addr);
1164                         if (unlikely(!page))
1165                                 continue;
1166                         if (unlikely(details) && details->nonlinear_vma
1167                             && linear_page_index(details->nonlinear_vma,
1168                                                 addr) != page->index)
1169                                 set_pte_at(mm, addr, pte,
1170                                            pgoff_to_pte(page->index));
1171                         if (PageAnon(page))
1172                                 rss[MM_ANONPAGES]--;
1173                         else {
1174                                 if (pte_dirty(ptent))
1175                                         set_page_dirty(page);
1176                                 if (pte_young(ptent) &&
1177                                     likely(!VM_SequentialReadHint(vma)))
1178                                         mark_page_accessed(page);
1179                                 rss[MM_FILEPAGES]--;
1180                         }
1181                         page_remove_rmap(page);
1182                         if (unlikely(page_mapcount(page) < 0))
1183                                 print_bad_pte(vma, addr, ptent, page);
1184                         force_flush = !__tlb_remove_page(tlb, page);
1185                         if (force_flush)
1186                                 break;
1187                         continue;
1188                 }
1189                 /*
1190                  * If details->check_mapping, we leave swap entries;
1191                  * if details->nonlinear_vma, we leave file entries.
1192                  */
1193                 if (unlikely(details))
1194                         continue;
1195                 if (pte_file(ptent)) {
1196                         if (unlikely(!(vma->vm_flags & VM_NONLINEAR)))
1197                                 print_bad_pte(vma, addr, ptent, NULL);
1198                 } else {
1199                         swp_entry_t entry = pte_to_swp_entry(ptent);
1200
1201                         if (!non_swap_entry(entry))
1202                                 rss[MM_SWAPENTS]--;
1203                         else if (is_migration_entry(entry)) {
1204                                 struct page *page;
1205
1206                                 page = migration_entry_to_page(entry);
1207
1208                                 if (PageAnon(page))
1209                                         rss[MM_ANONPAGES]--;
1210                                 else
1211                                         rss[MM_FILEPAGES]--;
1212                         }
1213                         if (unlikely(!free_swap_and_cache(entry)))
1214                                 print_bad_pte(vma, addr, ptent, NULL);
1215                 }
1216                 pte_clear_not_present_full(mm, addr, pte, tlb->fullmm);
1217         } while (pte++, addr += PAGE_SIZE, addr != end);
1218
1219         add_mm_rss_vec(mm, rss);
1220         arch_leave_lazy_mmu_mode();
1221         pte_unmap_unlock(start_pte, ptl);
1222
1223         /*
1224          * mmu_gather ran out of room to batch pages, we break out of
1225          * the PTE lock to avoid doing the potential expensive TLB invalidate
1226          * and page-free while holding it.
1227          */
1228         if (force_flush) {
1229                 force_flush = 0;
1230                 tlb_flush_mmu(tlb);
1231                 if (addr != end)
1232                         goto again;
1233         }
1234
1235         return addr;
1236 }
1237
1238 static inline unsigned long zap_pmd_range(struct mmu_gather *tlb,
1239                                 struct vm_area_struct *vma, pud_t *pud,
1240                                 unsigned long addr, unsigned long end,
1241                                 struct zap_details *details)
1242 {
1243         pmd_t *pmd;
1244         unsigned long next;
1245
1246         pmd = pmd_offset(pud, addr);
1247         do {
1248                 next = pmd_addr_end(addr, end);
1249                 if (pmd_trans_huge(*pmd)) {
1250                         if (next-addr != HPAGE_PMD_SIZE) {
1251                                 VM_BUG_ON(!rwsem_is_locked(&tlb->mm->mmap_sem));
1252                                 split_huge_page_pmd(vma->vm_mm, pmd);
1253                         } else if (zap_huge_pmd(tlb, vma, pmd, addr))
1254                                 continue;
1255                         /* fall through */
1256                 }
1257                 if (pmd_none_or_clear_bad(pmd))
1258                         continue;
1259                 next = zap_pte_range(tlb, vma, pmd, addr, next, details);
1260                 cond_resched();
1261         } while (pmd++, addr = next, addr != end);
1262
1263         return addr;
1264 }
1265
1266 static inline unsigned long zap_pud_range(struct mmu_gather *tlb,
1267                                 struct vm_area_struct *vma, pgd_t *pgd,
1268                                 unsigned long addr, unsigned long end,
1269                                 struct zap_details *details)
1270 {
1271         pud_t *pud;
1272         unsigned long next;
1273
1274         pud = pud_offset(pgd, addr);
1275         do {
1276                 next = pud_addr_end(addr, end);
1277                 if (pud_none_or_clear_bad(pud))
1278                         continue;
1279                 next = zap_pmd_range(tlb, vma, pud, addr, next, details);
1280         } while (pud++, addr = next, addr != end);
1281
1282         return addr;
1283 }
1284
1285 static void unmap_page_range(struct mmu_gather *tlb,
1286                              struct vm_area_struct *vma,
1287                              unsigned long addr, unsigned long end,
1288                              struct zap_details *details)
1289 {
1290         pgd_t *pgd;
1291         unsigned long next;
1292
1293         if (details && !details->check_mapping && !details->nonlinear_vma)
1294                 details = NULL;
1295
1296         BUG_ON(addr >= end);
1297         mem_cgroup_uncharge_start();
1298         tlb_start_vma(tlb, vma);
1299         pgd = pgd_offset(vma->vm_mm, addr);
1300         do {
1301                 next = pgd_addr_end(addr, end);
1302                 if (pgd_none_or_clear_bad(pgd))
1303                         continue;
1304                 next = zap_pud_range(tlb, vma, pgd, addr, next, details);
1305         } while (pgd++, addr = next, addr != end);
1306         tlb_end_vma(tlb, vma);
1307         mem_cgroup_uncharge_end();
1308 }
1309
1310 /**
1311  * unmap_vmas - unmap a range of memory covered by a list of vma's
1312  * @tlb: address of the caller's struct mmu_gather
1313  * @vma: the starting vma
1314  * @start_addr: virtual address at which to start unmapping
1315  * @end_addr: virtual address at which to end unmapping
1316  * @nr_accounted: Place number of unmapped pages in vm-accountable vma's here
1317  * @details: details of nonlinear truncation or shared cache invalidation
1318  *
1319  * Returns the end address of the unmapping (restart addr if interrupted).
1320  *
1321  * Unmap all pages in the vma list.
1322  *
1323  * Only addresses between `start' and `end' will be unmapped.
1324  *
1325  * The VMA list must be sorted in ascending virtual address order.
1326  *
1327  * unmap_vmas() assumes that the caller will flush the whole unmapped address
1328  * range after unmap_vmas() returns.  So the only responsibility here is to
1329  * ensure that any thus-far unmapped pages are flushed before unmap_vmas()
1330  * drops the lock and schedules.
1331  */
1332 unsigned long unmap_vmas(struct mmu_gather *tlb,
1333                 struct vm_area_struct *vma, unsigned long start_addr,
1334                 unsigned long end_addr, unsigned long *nr_accounted,
1335                 struct zap_details *details)
1336 {
1337         unsigned long start = start_addr;
1338         struct mm_struct *mm = vma->vm_mm;
1339
1340         mmu_notifier_invalidate_range_start(mm, start_addr, end_addr);
1341         for ( ; vma && vma->vm_start < end_addr; vma = vma->vm_next) {
1342                 unsigned long end;
1343
1344                 start = max(vma->vm_start, start_addr);
1345                 if (start >= vma->vm_end)
1346                         continue;
1347                 end = min(vma->vm_end, end_addr);
1348                 if (end <= vma->vm_start)
1349                         continue;
1350
1351                 if (vma->vm_flags & VM_ACCOUNT)
1352                         *nr_accounted += (end - start) >> PAGE_SHIFT;
1353
1354                 if (unlikely(is_pfn_mapping(vma)))
1355                         untrack_pfn_vma(vma, 0, 0);
1356
1357                 while (start != end) {
1358                         if (unlikely(is_vm_hugetlb_page(vma))) {
1359                                 /*
1360                                  * It is undesirable to test vma->vm_file as it
1361                                  * should be non-null for valid hugetlb area.
1362                                  * However, vm_file will be NULL in the error
1363                                  * cleanup path of do_mmap_pgoff. When
1364                                  * hugetlbfs ->mmap method fails,
1365                                  * do_mmap_pgoff() nullifies vma->vm_file
1366                                  * before calling this function to clean up.
1367                                  * Since no pte has actually been setup, it is
1368                                  * safe to do nothing in this case.
1369                                  */
1370                                 if (vma->vm_file)
1371                                         unmap_hugepage_range(vma, start, end, NULL);
1372                         } else
1373                                 unmap_page_range(tlb, vma, start, end, details);
1374                         start = end;
1375                 }
1376         }
1377
1378         mmu_notifier_invalidate_range_end(mm, start_addr, end_addr);
1379         return start;   /* which is now the end (or restart) address */
1380 }
1381
1382 /**
1383  * zap_page_range - remove user pages in a given range
1384  * @vma: vm_area_struct holding the applicable pages
1385  * @address: starting address of pages to zap
1386  * @size: number of bytes to zap
1387  * @details: details of nonlinear truncation or shared cache invalidation
1388  */
1389 unsigned long zap_page_range(struct vm_area_struct *vma, unsigned long address,
1390                 unsigned long size, struct zap_details *details)
1391 {
1392         struct mm_struct *mm = vma->vm_mm;
1393         struct mmu_gather tlb;
1394         unsigned long end = address + size;
1395         unsigned long nr_accounted = 0;
1396
1397         lru_add_drain();
1398         tlb_gather_mmu(&tlb, mm, 0);
1399         update_hiwater_rss(mm);
1400         end = unmap_vmas(&tlb, vma, address, end, &nr_accounted, details);
1401         tlb_finish_mmu(&tlb, address, end);
1402         return end;
1403 }
1404
1405 /**
1406  * zap_vma_ptes - remove ptes mapping the vma
1407  * @vma: vm_area_struct holding ptes to be zapped
1408  * @address: starting address of pages to zap
1409  * @size: number of bytes to zap
1410  *
1411  * This function only unmaps ptes assigned to VM_PFNMAP vmas.
1412  *
1413  * The entire address range must be fully contained within the vma.
1414  *
1415  * Returns 0 if successful.
1416  */
1417 int zap_vma_ptes(struct vm_area_struct *vma, unsigned long address,
1418                 unsigned long size)
1419 {
1420         if (address < vma->vm_start || address + size > vma->vm_end ||
1421                         !(vma->vm_flags & VM_PFNMAP))
1422                 return -1;
1423         zap_page_range(vma, address, size, NULL);
1424         return 0;
1425 }
1426 EXPORT_SYMBOL_GPL(zap_vma_ptes);
1427
1428 /**
1429  * follow_page - look up a page descriptor from a user-virtual address
1430  * @vma: vm_area_struct mapping @address
1431  * @address: virtual address to look up
1432  * @flags: flags modifying lookup behaviour
1433  *
1434  * @flags can have FOLL_ flags set, defined in <linux/mm.h>
1435  *
1436  * Returns the mapped (struct page *), %NULL if no mapping exists, or
1437  * an error pointer if there is a mapping to something not represented
1438  * by a page descriptor (see also vm_normal_page()).
1439  */
1440 struct page *follow_page(struct vm_area_struct *vma, unsigned long address,
1441                         unsigned int flags)
1442 {
1443         pgd_t *pgd;
1444         pud_t *pud;
1445         pmd_t *pmd;
1446         pte_t *ptep, pte;
1447         spinlock_t *ptl;
1448         struct page *page;
1449         struct mm_struct *mm = vma->vm_mm;
1450
1451         page = follow_huge_addr(mm, address, flags & FOLL_WRITE);
1452         if (!IS_ERR(page)) {
1453                 BUG_ON(flags & FOLL_GET);
1454                 goto out;
1455         }
1456
1457         page = NULL;
1458         pgd = pgd_offset(mm, address);
1459         if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
1460                 goto no_page_table;
1461
1462         pud = pud_offset(pgd, address);
1463         if (pud_none(*pud))
1464                 goto no_page_table;
1465         if (pud_huge(*pud) && vma->vm_flags & VM_HUGETLB) {
1466                 BUG_ON(flags & FOLL_GET);
1467                 page = follow_huge_pud(mm, address, pud, flags & FOLL_WRITE);
1468                 goto out;
1469         }
1470         if (unlikely(pud_bad(*pud)))
1471                 goto no_page_table;
1472
1473         pmd = pmd_offset(pud, address);
1474         if (pmd_none(*pmd))
1475                 goto no_page_table;
1476         if (pmd_huge(*pmd) && vma->vm_flags & VM_HUGETLB) {
1477                 BUG_ON(flags & FOLL_GET);
1478                 page = follow_huge_pmd(mm, address, pmd, flags & FOLL_WRITE);
1479                 goto out;
1480         }
1481         if (pmd_trans_huge(*pmd)) {
1482                 if (flags & FOLL_SPLIT) {
1483                         split_huge_page_pmd(mm, pmd);
1484                         goto split_fallthrough;
1485                 }
1486                 spin_lock(&mm->page_table_lock);
1487                 if (likely(pmd_trans_huge(*pmd))) {
1488                         if (unlikely(pmd_trans_splitting(*pmd))) {
1489                                 spin_unlock(&mm->page_table_lock);
1490                                 wait_split_huge_page(vma->anon_vma, pmd);
1491                         } else {
1492                                 page = follow_trans_huge_pmd(mm, address,
1493                                                              pmd, flags);
1494                                 spin_unlock(&mm->page_table_lock);
1495                                 goto out;
1496                         }
1497                 } else
1498                         spin_unlock(&mm->page_table_lock);
1499                 /* fall through */
1500         }
1501 split_fallthrough:
1502         if (unlikely(pmd_bad(*pmd)))
1503                 goto no_page_table;
1504
1505         ptep = pte_offset_map_lock(mm, pmd, address, &ptl);
1506
1507         pte = *ptep;
1508         if (!pte_present(pte))
1509                 goto no_page;
1510         if ((flags & FOLL_WRITE) && !pte_write(pte))
1511                 goto unlock;
1512
1513         page = vm_normal_page(vma, address, pte);
1514         if (unlikely(!page)) {
1515                 if ((flags & FOLL_DUMP) ||
1516                     !is_zero_pfn(pte_pfn(pte)))
1517                         goto bad_page;
1518                 page = pte_page(pte);
1519         }
1520
1521         if (flags & FOLL_GET)
1522                 get_page_foll(page);
1523         if (flags & FOLL_TOUCH) {
1524                 if ((flags & FOLL_WRITE) &&
1525                     !pte_dirty(pte) && !PageDirty(page))
1526                         set_page_dirty(page);
1527                 /*
1528                  * pte_mkyoung() would be more correct here, but atomic care
1529                  * is needed to avoid losing the dirty bit: it is easier to use
1530                  * mark_page_accessed().
1531                  */
1532                 mark_page_accessed(page);
1533         }
1534         if ((flags & FOLL_MLOCK) && (vma->vm_flags & VM_LOCKED)) {
1535                 /*
1536                  * The preliminary mapping check is mainly to avoid the
1537                  * pointless overhead of lock_page on the ZERO_PAGE
1538                  * which might bounce very badly if there is contention.
1539                  *
1540                  * If the page is already locked, we don't need to
1541                  * handle it now - vmscan will handle it later if and
1542                  * when it attempts to reclaim the page.
1543                  */
1544                 if (page->mapping && trylock_page(page)) {
1545                         lru_add_drain();  /* push cached pages to LRU */
1546                         /*
1547                          * Because we lock page here and migration is
1548                          * blocked by the pte's page reference, we need
1549                          * only check for file-cache page truncation.
1550                          */
1551                         if (page->mapping)
1552                                 mlock_vma_page(page);
1553                         unlock_page(page);
1554                 }
1555         }
1556 unlock:
1557         pte_unmap_unlock(ptep, ptl);
1558 out:
1559         return page;
1560
1561 bad_page:
1562         pte_unmap_unlock(ptep, ptl);
1563         return ERR_PTR(-EFAULT);
1564
1565 no_page:
1566         pte_unmap_unlock(ptep, ptl);
1567         if (!pte_none(pte))
1568                 return page;
1569
1570 no_page_table:
1571         /*
1572          * When core dumping an enormous anonymous area that nobody
1573          * has touched so far, we don't want to allocate unnecessary pages or
1574          * page tables.  Return error instead of NULL to skip handle_mm_fault,
1575          * then get_dump_page() will return NULL to leave a hole in the dump.
1576          * But we can only make this optimization where a hole would surely
1577          * be zero-filled if handle_mm_fault() actually did handle it.
1578          */
1579         if ((flags & FOLL_DUMP) &&
1580             (!vma->vm_ops || !vma->vm_ops->fault))
1581                 return ERR_PTR(-EFAULT);
1582         return page;
1583 }
1584
1585 static inline int stack_guard_page(struct vm_area_struct *vma, unsigned long addr)
1586 {
1587         return stack_guard_page_start(vma, addr) ||
1588                stack_guard_page_end(vma, addr+PAGE_SIZE);
1589 }
1590
1591 /**
1592  * __get_user_pages() - pin user pages in memory
1593  * @tsk:        task_struct of target task
1594  * @mm:         mm_struct of target mm
1595  * @start:      starting user address
1596  * @nr_pages:   number of pages from start to pin
1597  * @gup_flags:  flags modifying pin behaviour
1598  * @pages:      array that receives pointers to the pages pinned.
1599  *              Should be at least nr_pages long. Or NULL, if caller
1600  *              only intends to ensure the pages are faulted in.
1601  * @vmas:       array of pointers to vmas corresponding to each page.
1602  *              Or NULL if the caller does not require them.
1603  * @nonblocking: whether waiting for disk IO or mmap_sem contention
1604  *
1605  * Returns number of pages pinned. This may be fewer than the number
1606  * requested. If nr_pages is 0 or negative, returns 0. If no pages
1607  * were pinned, returns -errno. Each page returned must be released
1608  * with a put_page() call when it is finished with. vmas will only
1609  * remain valid while mmap_sem is held.
1610  *
1611  * Must be called with mmap_sem held for read or write.
1612  *
1613  * __get_user_pages walks a process's page tables and takes a reference to
1614  * each struct page that each user address corresponds to at a given
1615  * instant. That is, it takes the page that would be accessed if a user
1616  * thread accesses the given user virtual address at that instant.
1617  *
1618  * This does not guarantee that the page exists in the user mappings when
1619  * __get_user_pages returns, and there may even be a completely different
1620  * page there in some cases (eg. if mmapped pagecache has been invalidated
1621  * and subsequently re faulted). However it does guarantee that the page
1622  * won't be freed completely. And mostly callers simply care that the page
1623  * contains data that was valid *at some point in time*. Typically, an IO
1624  * or similar operation cannot guarantee anything stronger anyway because
1625  * locks can't be held over the syscall boundary.
1626  *
1627  * If @gup_flags & FOLL_WRITE == 0, the page must not be written to. If
1628  * the page is written to, set_page_dirty (or set_page_dirty_lock, as
1629  * appropriate) must be called after the page is finished with, and
1630  * before put_page is called.
1631  *
1632  * If @nonblocking != NULL, __get_user_pages will not wait for disk IO
1633  * or mmap_sem contention, and if waiting is needed to pin all pages,
1634  * *@nonblocking will be set to 0.
1635  *
1636  * In most cases, get_user_pages or get_user_pages_fast should be used
1637  * instead of __get_user_pages. __get_user_pages should be used only if
1638  * you need some special @gup_flags.
1639  */
1640 int __get_user_pages(struct task_struct *tsk, struct mm_struct *mm,
1641                      unsigned long start, int nr_pages, unsigned int gup_flags,
1642                      struct page **pages, struct vm_area_struct **vmas,
1643                      int *nonblocking)
1644 {
1645         int i;
1646         unsigned long vm_flags;
1647
1648         if (nr_pages <= 0)
1649                 return 0;
1650
1651         VM_BUG_ON(!!pages != !!(gup_flags & FOLL_GET));
1652
1653         /* 
1654          * Require read or write permissions.
1655          * If FOLL_FORCE is set, we only require the "MAY" flags.
1656          */
1657         vm_flags  = (gup_flags & FOLL_WRITE) ?
1658                         (VM_WRITE | VM_MAYWRITE) : (VM_READ | VM_MAYREAD);
1659         vm_flags &= (gup_flags & FOLL_FORCE) ?
1660                         (VM_MAYREAD | VM_MAYWRITE) : (VM_READ | VM_WRITE);
1661         i = 0;
1662
1663         do {
1664                 struct vm_area_struct *vma;
1665
1666                 vma = find_extend_vma(mm, start);
1667                 if (!vma && in_gate_area(mm, start)) {
1668                         unsigned long pg = start & PAGE_MASK;
1669                         pgd_t *pgd;
1670                         pud_t *pud;
1671                         pmd_t *pmd;
1672                         pte_t *pte;
1673
1674                         /* user gate pages are read-only */
1675                         if (gup_flags & FOLL_WRITE)
1676                                 return i ? : -EFAULT;
1677                         if (pg > TASK_SIZE)
1678                                 pgd = pgd_offset_k(pg);
1679                         else
1680                                 pgd = pgd_offset_gate(mm, pg);
1681                         BUG_ON(pgd_none(*pgd));
1682                         pud = pud_offset(pgd, pg);
1683                         BUG_ON(pud_none(*pud));
1684                         pmd = pmd_offset(pud, pg);
1685                         if (pmd_none(*pmd))
1686                                 return i ? : -EFAULT;
1687                         VM_BUG_ON(pmd_trans_huge(*pmd));
1688                         pte = pte_offset_map(pmd, pg);
1689                         if (pte_none(*pte)) {
1690                                 pte_unmap(pte);
1691                                 return i ? : -EFAULT;
1692                         }
1693                         vma = get_gate_vma(mm);
1694                         if (pages) {
1695                                 struct page *page;
1696
1697                                 page = vm_normal_page(vma, start, *pte);
1698                                 if (!page) {
1699                                         if (!(gup_flags & FOLL_DUMP) &&
1700                                              is_zero_pfn(pte_pfn(*pte)))
1701                                                 page = pte_page(*pte);
1702                                         else {
1703                                                 pte_unmap(pte);
1704                                                 return i ? : -EFAULT;
1705                                         }
1706                                 }
1707                                 pages[i] = page;
1708                                 get_page(page);
1709                         }
1710                         pte_unmap(pte);
1711                         goto next_page;
1712                 }
1713
1714                 if (!vma ||
1715                     (vma->vm_flags & (VM_IO | VM_PFNMAP)) ||
1716                     !(vm_flags & vma->vm_flags))
1717                         return i ? : -EFAULT;
1718
1719                 if (is_vm_hugetlb_page(vma)) {
1720                         i = follow_hugetlb_page(mm, vma, pages, vmas,
1721                                         &start, &nr_pages, i, gup_flags);
1722                         continue;
1723                 }
1724
1725                 do {
1726                         struct page *page;
1727                         unsigned int foll_flags = gup_flags;
1728
1729                         /*
1730                          * If we have a pending SIGKILL, don't keep faulting
1731                          * pages and potentially allocating memory.
1732                          */
1733                         if (unlikely(fatal_signal_pending(current)))
1734                                 return i ? i : -ERESTARTSYS;
1735
1736                         cond_resched();
1737                         while (!(page = follow_page(vma, start, foll_flags))) {
1738                                 int ret;
1739                                 unsigned int fault_flags = 0;
1740
1741                                 /* For mlock, just skip the stack guard page. */
1742                                 if (foll_flags & FOLL_MLOCK) {
1743                                         if (stack_guard_page(vma, start))
1744                                                 goto next_page;
1745                                 }
1746                                 if (foll_flags & FOLL_WRITE)
1747                                         fault_flags |= FAULT_FLAG_WRITE;
1748                                 if (nonblocking)
1749                                         fault_flags |= FAULT_FLAG_ALLOW_RETRY;
1750                                 if (foll_flags & FOLL_NOWAIT)
1751                                         fault_flags |= (FAULT_FLAG_ALLOW_RETRY | FAULT_FLAG_RETRY_NOWAIT);
1752
1753                                 ret = handle_mm_fault(mm, vma, start,
1754                                                         fault_flags);
1755
1756                                 if (ret & VM_FAULT_ERROR) {
1757                                         if (ret & VM_FAULT_OOM)
1758                                                 return i ? i : -ENOMEM;
1759                                         if (ret & (VM_FAULT_HWPOISON |
1760                                                    VM_FAULT_HWPOISON_LARGE)) {
1761                                                 if (i)
1762                                                         return i;
1763                                                 else if (gup_flags & FOLL_HWPOISON)
1764                                                         return -EHWPOISON;
1765                                                 else
1766                                                         return -EFAULT;
1767                                         }
1768                                         if (ret & VM_FAULT_SIGBUS)
1769                                                 return i ? i : -EFAULT;
1770                                         BUG();
1771                                 }
1772
1773                                 if (tsk) {
1774                                         if (ret & VM_FAULT_MAJOR)
1775                                                 tsk->maj_flt++;
1776                                         else
1777                                                 tsk->min_flt++;
1778                                 }
1779
1780                                 if (ret & VM_FAULT_RETRY) {
1781                                         if (nonblocking)
1782                                                 *nonblocking = 0;
1783                                         return i;
1784                                 }
1785
1786                                 /*
1787                                  * The VM_FAULT_WRITE bit tells us that
1788                                  * do_wp_page has broken COW when necessary,
1789                                  * even if maybe_mkwrite decided not to set
1790                                  * pte_write. We can thus safely do subsequent
1791                                  * page lookups as if they were reads. But only
1792                                  * do so when looping for pte_write is futile:
1793                                  * in some cases userspace may also be wanting
1794                                  * to write to the gotten user page, which a
1795                                  * read fault here might prevent (a readonly
1796                                  * page might get reCOWed by userspace write).
1797                                  */
1798                                 if ((ret & VM_FAULT_WRITE) &&
1799                                     !(vma->vm_flags & VM_WRITE))
1800                                         foll_flags &= ~FOLL_WRITE;
1801
1802                                 cond_resched();
1803                         }
1804                         if (IS_ERR(page))
1805                                 return i ? i : PTR_ERR(page);
1806                         if (pages) {
1807                                 pages[i] = page;
1808
1809                                 flush_anon_page(vma, page, start);
1810                                 flush_dcache_page(page);
1811                         }
1812 next_page:
1813                         if (vmas)
1814                                 vmas[i] = vma;
1815                         i++;
1816                         start += PAGE_SIZE;
1817                         nr_pages--;
1818                 } while (nr_pages && start < vma->vm_end);
1819         } while (nr_pages);
1820         return i;
1821 }
1822 EXPORT_SYMBOL(__get_user_pages);
1823
1824 /*
1825  * fixup_user_fault() - manually resolve a user page fault
1826  * @tsk:        the task_struct to use for page fault accounting, or
1827  *              NULL if faults are not to be recorded.
1828  * @mm:         mm_struct of target mm
1829  * @address:    user address
1830  * @fault_flags:flags to pass down to handle_mm_fault()
1831  *
1832  * This is meant to be called in the specific scenario where for locking reasons
1833  * we try to access user memory in atomic context (within a pagefault_disable()
1834  * section), this returns -EFAULT, and we want to resolve the user fault before
1835  * trying again.
1836  *
1837  * Typically this is meant to be used by the futex code.
1838  *
1839  * The main difference with get_user_pages() is that this function will
1840  * unconditionally call handle_mm_fault() which will in turn perform all the
1841  * necessary SW fixup of the dirty and young bits in the PTE, while
1842  * handle_mm_fault() only guarantees to update these in the struct page.
1843  *
1844  * This is important for some architectures where those bits also gate the
1845  * access permission to the page because they are maintained in software.  On
1846  * such architectures, gup() will not be enough to make a subsequent access
1847  * succeed.
1848  *
1849  * This should be called with the mm_sem held for read.
1850  */
1851 int fixup_user_fault(struct task_struct *tsk, struct mm_struct *mm,
1852                      unsigned long address, unsigned int fault_flags)
1853 {
1854         struct vm_area_struct *vma;
1855         int ret;
1856
1857         vma = find_extend_vma(mm, address);
1858         if (!vma || address < vma->vm_start)
1859                 return -EFAULT;
1860
1861         ret = handle_mm_fault(mm, vma, address, fault_flags);
1862         if (ret & VM_FAULT_ERROR) {
1863                 if (ret & VM_FAULT_OOM)
1864                         return -ENOMEM;
1865                 if (ret & (VM_FAULT_HWPOISON | VM_FAULT_HWPOISON_LARGE))
1866                         return -EHWPOISON;
1867                 if (ret & VM_FAULT_SIGBUS)
1868                         return -EFAULT;
1869                 BUG();
1870         }
1871         if (tsk) {
1872                 if (ret & VM_FAULT_MAJOR)
1873                         tsk->maj_flt++;
1874                 else
1875                         tsk->min_flt++;
1876         }
1877         return 0;
1878 }
1879
1880 /*
1881  * get_user_pages() - pin user pages in memory
1882  * @tsk:        the task_struct to use for page fault accounting, or
1883  *              NULL if faults are not to be recorded.
1884  * @mm:         mm_struct of target mm
1885  * @start:      starting user address
1886  * @nr_pages:   number of pages from start to pin
1887  * @write:      whether pages will be written to by the caller
1888  * @force:      whether to force write access even if user mapping is
1889  *              readonly. This will result in the page being COWed even
1890  *              in MAP_SHARED mappings. You do not want this.
1891  * @pages:      array that receives pointers to the pages pinned.
1892  *              Should be at least nr_pages long. Or NULL, if caller
1893  *              only intends to ensure the pages are faulted in.
1894  * @vmas:       array of pointers to vmas corresponding to each page.
1895  *              Or NULL if the caller does not require them.
1896  *
1897  * Returns number of pages pinned. This may be fewer than the number
1898  * requested. If nr_pages is 0 or negative, returns 0. If no pages
1899  * were pinned, returns -errno. Each page returned must be released
1900  * with a put_page() call when it is finished with. vmas will only
1901  * remain valid while mmap_sem is held.
1902  *
1903  * Must be called with mmap_sem held for read or write.
1904  *
1905  * get_user_pages walks a process's page tables and takes a reference to
1906  * each struct page that each user address corresponds to at a given
1907  * instant. That is, it takes the page that would be accessed if a user
1908  * thread accesses the given user virtual address at that instant.
1909  *
1910  * This does not guarantee that the page exists in the user mappings when
1911  * get_user_pages returns, and there may even be a completely different
1912  * page there in some cases (eg. if mmapped pagecache has been invalidated
1913  * and subsequently re faulted). However it does guarantee that the page
1914  * won't be freed completely. And mostly callers simply care that the page
1915  * contains data that was valid *at some point in time*. Typically, an IO
1916  * or similar operation cannot guarantee anything stronger anyway because
1917  * locks can't be held over the syscall boundary.
1918  *
1919  * If write=0, the page must not be written to. If the page is written to,
1920  * set_page_dirty (or set_page_dirty_lock, as appropriate) must be called
1921  * after the page is finished with, and before put_page is called.
1922  *
1923  * get_user_pages is typically used for fewer-copy IO operations, to get a
1924  * handle on the memory by some means other than accesses via the user virtual
1925  * addresses. The pages may be submitted for DMA to devices or accessed via
1926  * their kernel linear mapping (via the kmap APIs). Care should be taken to
1927  * use the correct cache flushing APIs.
1928  *
1929  * See also get_user_pages_fast, for performance critical applications.
1930  */
1931 int get_user_pages(struct task_struct *tsk, struct mm_struct *mm,
1932                 unsigned long start, int nr_pages, int write, int force,
1933                 struct page **pages, struct vm_area_struct **vmas)
1934 {
1935         int flags = FOLL_TOUCH;
1936
1937         if (pages)
1938                 flags |= FOLL_GET;
1939         if (write)
1940                 flags |= FOLL_WRITE;
1941         if (force)
1942                 flags |= FOLL_FORCE;
1943
1944         return __get_user_pages(tsk, mm, start, nr_pages, flags, pages, vmas,
1945                                 NULL);
1946 }
1947 EXPORT_SYMBOL(get_user_pages);
1948
1949 /**
1950  * get_dump_page() - pin user page in memory while writing it to core dump
1951  * @addr: user address
1952  *
1953  * Returns struct page pointer of user page pinned for dump,
1954  * to be freed afterwards by page_cache_release() or put_page().
1955  *
1956  * Returns NULL on any kind of failure - a hole must then be inserted into
1957  * the corefile, to preserve alignment with its headers; and also returns
1958  * NULL wherever the ZERO_PAGE, or an anonymous pte_none, has been found -
1959  * allowing a hole to be left in the corefile to save diskspace.
1960  *
1961  * Called without mmap_sem, but after all other threads have been killed.
1962  */
1963 #ifdef CONFIG_ELF_CORE
1964 struct page *get_dump_page(unsigned long addr)
1965 {
1966         struct vm_area_struct *vma;
1967         struct page *page;
1968
1969         if (__get_user_pages(current, current->mm, addr, 1,
1970                              FOLL_FORCE | FOLL_DUMP | FOLL_GET, &page, &vma,
1971                              NULL) < 1)
1972                 return NULL;
1973         flush_cache_page(vma, addr, page_to_pfn(page));
1974         return page;
1975 }
1976 #endif /* CONFIG_ELF_CORE */
1977
1978 pte_t *__get_locked_pte(struct mm_struct *mm, unsigned long addr,
1979                         spinlock_t **ptl)
1980 {
1981         pgd_t * pgd = pgd_offset(mm, addr);
1982         pud_t * pud = pud_alloc(mm, pgd, addr);
1983         if (pud) {
1984                 pmd_t * pmd = pmd_alloc(mm, pud, addr);
1985                 if (pmd) {
1986                         VM_BUG_ON(pmd_trans_huge(*pmd));
1987                         return pte_alloc_map_lock(mm, pmd, addr, ptl);
1988                 }
1989         }
1990         return NULL;
1991 }
1992
1993 /*
1994  * This is the old fallback for page remapping.
1995  *
1996  * For historical reasons, it only allows reserved pages. Only
1997  * old drivers should use this, and they needed to mark their
1998  * pages reserved for the old functions anyway.
1999  */
2000 static int insert_page(struct vm_area_struct *vma, unsigned long addr,
2001                         struct page *page, pgprot_t prot)
2002 {
2003         struct mm_struct *mm = vma->vm_mm;
2004         int retval;
2005         pte_t *pte;
2006         spinlock_t *ptl;
2007
2008         retval = -EINVAL;
2009         if (PageAnon(page))
2010                 goto out;
2011         retval = -ENOMEM;
2012         flush_dcache_page(page);
2013         pte = get_locked_pte(mm, addr, &ptl);
2014         if (!pte)
2015                 goto out;
2016         retval = -EBUSY;
2017         if (!pte_none(*pte))
2018                 goto out_unlock;
2019
2020         /* Ok, finally just insert the thing.. */
2021         get_page(page);
2022         inc_mm_counter_fast(mm, MM_FILEPAGES);
2023         page_add_file_rmap(page);
2024         set_pte_at(mm, addr, pte, mk_pte(page, prot));
2025
2026         retval = 0;
2027         pte_unmap_unlock(pte, ptl);
2028         return retval;
2029 out_unlock:
2030         pte_unmap_unlock(pte, ptl);
2031 out:
2032         return retval;
2033 }
2034
2035 /**
2036  * vm_insert_page - insert single page into user vma
2037  * @vma: user vma to map to
2038  * @addr: target user address of this page
2039  * @page: source kernel page
2040  *
2041  * This allows drivers to insert individual pages they've allocated
2042  * into a user vma.
2043  *
2044  * The page has to be a nice clean _individual_ kernel allocation.
2045  * If you allocate a compound page, you need to have marked it as
2046  * such (__GFP_COMP), or manually just split the page up yourself
2047  * (see split_page()).
2048  *
2049  * NOTE! Traditionally this was done with "remap_pfn_range()" which
2050  * took an arbitrary page protection parameter. This doesn't allow
2051  * that. Your vma protection will have to be set up correctly, which
2052  * means that if you want a shared writable mapping, you'd better
2053  * ask for a shared writable mapping!
2054  *
2055  * The page does not need to be reserved.
2056  */
2057 int vm_insert_page(struct vm_area_struct *vma, unsigned long addr,
2058                         struct page *page)
2059 {
2060         if (addr < vma->vm_start || addr >= vma->vm_end)
2061                 return -EFAULT;
2062         if (!page_count(page))
2063                 return -EINVAL;
2064         vma->vm_flags |= VM_INSERTPAGE;
2065         return insert_page(vma, addr, page, vma->vm_page_prot);
2066 }
2067 EXPORT_SYMBOL(vm_insert_page);
2068
2069 static int insert_pfn(struct vm_area_struct *vma, unsigned long addr,
2070                         unsigned long pfn, pgprot_t prot)
2071 {
2072         struct mm_struct *mm = vma->vm_mm;
2073         int retval;
2074         pte_t *pte, entry;
2075         spinlock_t *ptl;
2076
2077         retval = -ENOMEM;
2078         pte = get_locked_pte(mm, addr, &ptl);
2079         if (!pte)
2080                 goto out;
2081         retval = -EBUSY;
2082         if (!pte_none(*pte))
2083                 goto out_unlock;
2084
2085         /* Ok, finally just insert the thing.. */
2086         entry = pte_mkspecial(pfn_pte(pfn, prot));
2087         set_pte_at(mm, addr, pte, entry);
2088         update_mmu_cache(vma, addr, pte); /* XXX: why not for insert_page? */
2089
2090         retval = 0;
2091 out_unlock:
2092         pte_unmap_unlock(pte, ptl);
2093 out:
2094         return retval;
2095 }
2096
2097 /**
2098  * vm_insert_pfn - insert single pfn into user vma
2099  * @vma: user vma to map to
2100  * @addr: target user address of this page
2101  * @pfn: source kernel pfn
2102  *
2103  * Similar to vm_inert_page, this allows drivers to insert individual pages
2104  * they've allocated into a user vma. Same comments apply.
2105  *
2106  * This function should only be called from a vm_ops->fault handler, and
2107  * in that case the handler should return NULL.
2108  *
2109  * vma cannot be a COW mapping.
2110  *
2111  * As this is called only for pages that do not currently exist, we
2112  * do not need to flush old virtual caches or the TLB.
2113  */
2114 int vm_insert_pfn(struct vm_area_struct *vma, unsigned long addr,
2115                         unsigned long pfn)
2116 {
2117         int ret;
2118         pgprot_t pgprot = vma->vm_page_prot;
2119         /*
2120          * Technically, architectures with pte_special can avoid all these
2121          * restrictions (same for remap_pfn_range).  However we would like
2122          * consistency in testing and feature parity among all, so we should
2123          * try to keep these invariants in place for everybody.
2124          */
2125         BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)));
2126         BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) ==
2127                                                 (VM_PFNMAP|VM_MIXEDMAP));
2128         BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags));
2129         BUG_ON((vma->vm_flags & VM_MIXEDMAP) && pfn_valid(pfn));
2130
2131         if (addr < vma->vm_start || addr >= vma->vm_end)
2132                 return -EFAULT;
2133         if (track_pfn_vma_new(vma, &pgprot, pfn, PAGE_SIZE))
2134                 return -EINVAL;
2135
2136         ret = insert_pfn(vma, addr, pfn, pgprot);
2137
2138         if (ret)
2139                 untrack_pfn_vma(vma, pfn, PAGE_SIZE);
2140
2141         return ret;
2142 }
2143 EXPORT_SYMBOL(vm_insert_pfn);
2144
2145 int vm_insert_mixed(struct vm_area_struct *vma, unsigned long addr,
2146                         unsigned long pfn)
2147 {
2148         BUG_ON(!(vma->vm_flags & VM_MIXEDMAP));
2149
2150         if (addr < vma->vm_start || addr >= vma->vm_end)
2151                 return -EFAULT;
2152
2153         /*
2154          * If we don't have pte special, then we have to use the pfn_valid()
2155          * based VM_MIXEDMAP scheme (see vm_normal_page), and thus we *must*
2156          * refcount the page if pfn_valid is true (hence insert_page rather
2157          * than insert_pfn).  If a zero_pfn were inserted into a VM_MIXEDMAP
2158          * without pte special, it would there be refcounted as a normal page.
2159          */
2160         if (!HAVE_PTE_SPECIAL && pfn_valid(pfn)) {
2161                 struct page *page;
2162
2163                 page = pfn_to_page(pfn);
2164                 return insert_page(vma, addr, page, vma->vm_page_prot);
2165         }
2166         return insert_pfn(vma, addr, pfn, vma->vm_page_prot);
2167 }
2168 EXPORT_SYMBOL(vm_insert_mixed);
2169
2170 /*
2171  * maps a range of physical memory into the requested pages. the old
2172  * mappings are removed. any references to nonexistent pages results
2173  * in null mappings (currently treated as "copy-on-access")
2174  */
2175 static int remap_pte_range(struct mm_struct *mm, pmd_t *pmd,
2176                         unsigned long addr, unsigned long end,
2177                         unsigned long pfn, pgprot_t prot)
2178 {
2179         pte_t *pte;
2180         spinlock_t *ptl;
2181
2182         pte = pte_alloc_map_lock(mm, pmd, addr, &ptl);
2183         if (!pte)
2184                 return -ENOMEM;
2185         arch_enter_lazy_mmu_mode();
2186         do {
2187                 BUG_ON(!pte_none(*pte));
2188                 set_pte_at(mm, addr, pte, pte_mkspecial(pfn_pte(pfn, prot)));
2189                 pfn++;
2190         } while (pte++, addr += PAGE_SIZE, addr != end);
2191         arch_leave_lazy_mmu_mode();
2192         pte_unmap_unlock(pte - 1, ptl);
2193         return 0;
2194 }
2195
2196 static inline int remap_pmd_range(struct mm_struct *mm, pud_t *pud,
2197                         unsigned long addr, unsigned long end,
2198                         unsigned long pfn, pgprot_t prot)
2199 {
2200         pmd_t *pmd;
2201         unsigned long next;
2202
2203         pfn -= addr >> PAGE_SHIFT;
2204         pmd = pmd_alloc(mm, pud, addr);
2205         if (!pmd)
2206                 return -ENOMEM;
2207         VM_BUG_ON(pmd_trans_huge(*pmd));
2208         do {
2209                 next = pmd_addr_end(addr, end);
2210                 if (remap_pte_range(mm, pmd, addr, next,
2211                                 pfn + (addr >> PAGE_SHIFT), prot))
2212                         return -ENOMEM;
2213         } while (pmd++, addr = next, addr != end);
2214         return 0;
2215 }
2216
2217 static inline int remap_pud_range(struct mm_struct *mm, pgd_t *pgd,
2218                         unsigned long addr, unsigned long end,
2219                         unsigned long pfn, pgprot_t prot)
2220 {
2221         pud_t *pud;
2222         unsigned long next;
2223
2224         pfn -= addr >> PAGE_SHIFT;
2225         pud = pud_alloc(mm, pgd, addr);
2226         if (!pud)
2227                 return -ENOMEM;
2228         do {
2229                 next = pud_addr_end(addr, end);
2230                 if (remap_pmd_range(mm, pud, addr, next,
2231                                 pfn + (addr >> PAGE_SHIFT), prot))
2232                         return -ENOMEM;
2233         } while (pud++, addr = next, addr != end);
2234         return 0;
2235 }
2236
2237 /**
2238  * remap_pfn_range - remap kernel memory to userspace
2239  * @vma: user vma to map to
2240  * @addr: target user address to start at
2241  * @pfn: physical address of kernel memory
2242  * @size: size of map area
2243  * @prot: page protection flags for this mapping
2244  *
2245  *  Note: this is only safe if the mm semaphore is held when called.
2246  */
2247 int remap_pfn_range(struct vm_area_struct *vma, unsigned long addr,
2248                     unsigned long pfn, unsigned long size, pgprot_t prot)
2249 {
2250         pgd_t *pgd;
2251         unsigned long next;
2252         unsigned long end = addr + PAGE_ALIGN(size);
2253         struct mm_struct *mm = vma->vm_mm;
2254         int err;
2255
2256         /*
2257          * Physically remapped pages are special. Tell the
2258          * rest of the world about it:
2259          *   VM_IO tells people not to look at these pages
2260          *      (accesses can have side effects).
2261          *   VM_RESERVED is specified all over the place, because
2262          *      in 2.4 it kept swapout's vma scan off this vma; but
2263          *      in 2.6 the LRU scan won't even find its pages, so this
2264          *      flag means no more than count its pages in reserved_vm,
2265          *      and omit it from core dump, even when VM_IO turned off.
2266          *   VM_PFNMAP tells the core MM that the base pages are just
2267          *      raw PFN mappings, and do not have a "struct page" associated
2268          *      with them.
2269          *
2270          * There's a horrible special case to handle copy-on-write
2271          * behaviour that some programs depend on. We mark the "original"
2272          * un-COW'ed pages by matching them up with "vma->vm_pgoff".
2273          */
2274         if (addr == vma->vm_start && end == vma->vm_end) {
2275                 vma->vm_pgoff = pfn;
2276                 vma->vm_flags |= VM_PFN_AT_MMAP;
2277         } else if (is_cow_mapping(vma->vm_flags))
2278                 return -EINVAL;
2279
2280         vma->vm_flags |= VM_IO | VM_RESERVED | VM_PFNMAP;
2281
2282         err = track_pfn_vma_new(vma, &prot, pfn, PAGE_ALIGN(size));
2283         if (err) {
2284                 /*
2285                  * To indicate that track_pfn related cleanup is not
2286                  * needed from higher level routine calling unmap_vmas
2287                  */
2288                 vma->vm_flags &= ~(VM_IO | VM_RESERVED | VM_PFNMAP);
2289                 vma->vm_flags &= ~VM_PFN_AT_MMAP;
2290                 return -EINVAL;
2291         }
2292
2293         BUG_ON(addr >= end);
2294         pfn -= addr >> PAGE_SHIFT;
2295         pgd = pgd_offset(mm, addr);
2296         flush_cache_range(vma, addr, end);
2297         do {
2298                 next = pgd_addr_end(addr, end);
2299                 err = remap_pud_range(mm, pgd, addr, next,
2300                                 pfn + (addr >> PAGE_SHIFT), prot);
2301                 if (err)
2302                         break;
2303         } while (pgd++, addr = next, addr != end);
2304
2305         if (err)
2306                 untrack_pfn_vma(vma, pfn, PAGE_ALIGN(size));
2307
2308         return err;
2309 }
2310 EXPORT_SYMBOL(remap_pfn_range);
2311
2312 static int apply_to_pte_range(struct mm_struct *mm, pmd_t *pmd,
2313                                      unsigned long addr, unsigned long end,
2314                                      pte_fn_t fn, void *data)
2315 {
2316         pte_t *pte;
2317         int err;
2318         pgtable_t token;
2319         spinlock_t *uninitialized_var(ptl);
2320
2321         pte = (mm == &init_mm) ?
2322                 pte_alloc_kernel(pmd, addr) :
2323                 pte_alloc_map_lock(mm, pmd, addr, &ptl);
2324         if (!pte)
2325                 return -ENOMEM;
2326
2327         BUG_ON(pmd_huge(*pmd));
2328
2329         arch_enter_lazy_mmu_mode();
2330
2331         token = pmd_pgtable(*pmd);
2332
2333         do {
2334                 err = fn(pte++, token, addr, data);
2335                 if (err)
2336                         break;
2337         } while (addr += PAGE_SIZE, addr != end);
2338
2339         arch_leave_lazy_mmu_mode();
2340
2341         if (mm != &init_mm)
2342                 pte_unmap_unlock(pte-1, ptl);
2343         return err;
2344 }
2345
2346 static int apply_to_pmd_range(struct mm_struct *mm, pud_t *pud,
2347                                      unsigned long addr, unsigned long end,
2348                                      pte_fn_t fn, void *data)
2349 {
2350         pmd_t *pmd;
2351         unsigned long next;
2352         int err;
2353
2354         BUG_ON(pud_huge(*pud));
2355
2356         pmd = pmd_alloc(mm, pud, addr);
2357         if (!pmd)
2358                 return -ENOMEM;
2359         do {
2360                 next = pmd_addr_end(addr, end);
2361                 err = apply_to_pte_range(mm, pmd, addr, next, fn, data);
2362                 if (err)
2363                         break;
2364         } while (pmd++, addr = next, addr != end);
2365         return err;
2366 }
2367
2368 static int apply_to_pud_range(struct mm_struct *mm, pgd_t *pgd,
2369                                      unsigned long addr, unsigned long end,
2370                                      pte_fn_t fn, void *data)
2371 {
2372         pud_t *pud;
2373         unsigned long next;
2374         int err;
2375
2376         pud = pud_alloc(mm, pgd, addr);
2377         if (!pud)
2378                 return -ENOMEM;
2379         do {
2380                 next = pud_addr_end(addr, end);
2381                 err = apply_to_pmd_range(mm, pud, addr, next, fn, data);
2382                 if (err)
2383                         break;
2384         } while (pud++, addr = next, addr != end);
2385         return err;
2386 }
2387
2388 /*
2389  * Scan a region of virtual memory, filling in page tables as necessary
2390  * and calling a provided function on each leaf page table.
2391  */
2392 int apply_to_page_range(struct mm_struct *mm, unsigned long addr,
2393                         unsigned long size, pte_fn_t fn, void *data)
2394 {
2395         pgd_t *pgd;
2396         unsigned long next;
2397         unsigned long end = addr + size;
2398         int err;
2399
2400         BUG_ON(addr >= end);
2401         pgd = pgd_offset(mm, addr);
2402         do {
2403                 next = pgd_addr_end(addr, end);
2404                 err = apply_to_pud_range(mm, pgd, addr, next, fn, data);
2405                 if (err)
2406                         break;
2407         } while (pgd++, addr = next, addr != end);
2408
2409         return err;
2410 }
2411 EXPORT_SYMBOL_GPL(apply_to_page_range);
2412
2413 /*
2414  * handle_pte_fault chooses page fault handler according to an entry
2415  * which was read non-atomically.  Before making any commitment, on
2416  * those architectures or configurations (e.g. i386 with PAE) which
2417  * might give a mix of unmatched parts, do_swap_page and do_nonlinear_fault
2418  * must check under lock before unmapping the pte and proceeding
2419  * (but do_wp_page is only called after already making such a check;
2420  * and do_anonymous_page can safely check later on).
2421  */
2422 static inline int pte_unmap_same(struct mm_struct *mm, pmd_t *pmd,
2423                                 pte_t *page_table, pte_t orig_pte)
2424 {
2425         int same = 1;
2426 #if defined(CONFIG_SMP) || defined(CONFIG_PREEMPT)
2427         if (sizeof(pte_t) > sizeof(unsigned long)) {
2428                 spinlock_t *ptl = pte_lockptr(mm, pmd);
2429                 spin_lock(ptl);
2430                 same = pte_same(*page_table, orig_pte);
2431                 spin_unlock(ptl);
2432         }
2433 #endif
2434         pte_unmap(page_table);
2435         return same;
2436 }
2437
2438 static inline void cow_user_page(struct page *dst, struct page *src, unsigned long va, struct vm_area_struct *vma)
2439 {
2440         /*
2441          * If the source page was a PFN mapping, we don't have
2442          * a "struct page" for it. We do a best-effort copy by
2443          * just copying from the original user address. If that
2444          * fails, we just zero-fill it. Live with it.
2445          */
2446         if (unlikely(!src)) {
2447                 void *kaddr = kmap_atomic(dst, KM_USER0);
2448                 void __user *uaddr = (void __user *)(va & PAGE_MASK);
2449
2450                 /*
2451                  * This really shouldn't fail, because the page is there
2452                  * in the page tables. But it might just be unreadable,
2453                  * in which case we just give up and fill the result with
2454                  * zeroes.
2455                  */
2456                 if (__copy_from_user_inatomic(kaddr, uaddr, PAGE_SIZE))
2457                         clear_page(kaddr);
2458                 kunmap_atomic(kaddr, KM_USER0);
2459                 flush_dcache_page(dst);
2460         } else
2461                 copy_user_highpage(dst, src, va, vma);
2462 }
2463
2464 /*
2465  * This routine handles present pages, when users try to write
2466  * to a shared page. It is done by copying the page to a new address
2467  * and decrementing the shared-page counter for the old page.
2468  *
2469  * Note that this routine assumes that the protection checks have been
2470  * done by the caller (the low-level page fault routine in most cases).
2471  * Thus we can safely just mark it writable once we've done any necessary
2472  * COW.
2473  *
2474  * We also mark the page dirty at this point even though the page will
2475  * change only once the write actually happens. This avoids a few races,
2476  * and potentially makes it more efficient.
2477  *
2478  * We enter with non-exclusive mmap_sem (to exclude vma changes,
2479  * but allow concurrent faults), with pte both mapped and locked.
2480  * We return with mmap_sem still held, but pte unmapped and unlocked.
2481  */
2482 static int do_wp_page(struct mm_struct *mm, struct vm_area_struct *vma,
2483                 unsigned long address, pte_t *page_table, pmd_t *pmd,
2484                 spinlock_t *ptl, pte_t orig_pte)
2485         __releases(ptl)
2486 {
2487         struct page *old_page, *new_page;
2488         pte_t entry;
2489         int ret = 0;
2490         int page_mkwrite = 0;
2491         struct page *dirty_page = NULL;
2492
2493         old_page = vm_normal_page(vma, address, orig_pte);
2494         if (!old_page) {
2495                 /*
2496                  * VM_MIXEDMAP !pfn_valid() case
2497                  *
2498                  * We should not cow pages in a shared writeable mapping.
2499                  * Just mark the pages writable as we can't do any dirty
2500                  * accounting on raw pfn maps.
2501                  */
2502                 if ((vma->vm_flags & (VM_WRITE|VM_SHARED)) ==
2503                                      (VM_WRITE|VM_SHARED))
2504                         goto reuse;
2505                 goto gotten;
2506         }
2507
2508         /*
2509          * Take out anonymous pages first, anonymous shared vmas are
2510          * not dirty accountable.
2511          */
2512         if (PageAnon(old_page) && !PageKsm(old_page)) {
2513                 if (!trylock_page(old_page)) {
2514                         page_cache_get(old_page);
2515                         pte_unmap_unlock(page_table, ptl);
2516                         lock_page(old_page);
2517                         page_table = pte_offset_map_lock(mm, pmd, address,
2518                                                          &ptl);
2519                         if (!pte_same(*page_table, orig_pte)) {
2520                                 unlock_page(old_page);
2521                                 goto unlock;
2522                         }
2523                         page_cache_release(old_page);
2524                 }
2525                 if (reuse_swap_page(old_page)) {
2526                         /*
2527                          * The page is all ours.  Move it to our anon_vma so
2528                          * the rmap code will not search our parent or siblings.
2529                          * Protected against the rmap code by the page lock.
2530                          */
2531                         page_move_anon_rmap(old_page, vma, address);
2532                         unlock_page(old_page);
2533                         goto reuse;
2534                 }
2535                 unlock_page(old_page);
2536         } else if (unlikely((vma->vm_flags & (VM_WRITE|VM_SHARED)) ==
2537                                         (VM_WRITE|VM_SHARED))) {
2538                 /*
2539                  * Only catch write-faults on shared writable pages,
2540                  * read-only shared pages can get COWed by
2541                  * get_user_pages(.write=1, .force=1).
2542                  */
2543                 if (vma->vm_ops && vma->vm_ops->page_mkwrite) {
2544                         struct vm_fault vmf;
2545                         int tmp;
2546
2547                         vmf.virtual_address = (void __user *)(address &
2548                                                                 PAGE_MASK);
2549                         vmf.pgoff = old_page->index;
2550                         vmf.flags = FAULT_FLAG_WRITE|FAULT_FLAG_MKWRITE;
2551                         vmf.page = old_page;
2552
2553                         /*
2554                          * Notify the address space that the page is about to
2555                          * become writable so that it can prohibit this or wait
2556                          * for the page to get into an appropriate state.
2557                          *
2558                          * We do this without the lock held, so that it can
2559                          * sleep if it needs to.
2560                          */
2561                         page_cache_get(old_page);
2562                         pte_unmap_unlock(page_table, ptl);
2563
2564                         tmp = vma->vm_ops->page_mkwrite(vma, &vmf);
2565                         if (unlikely(tmp &
2566                                         (VM_FAULT_ERROR | VM_FAULT_NOPAGE))) {
2567                                 ret = tmp;
2568                                 goto unwritable_page;
2569                         }
2570                         if (unlikely(!(tmp & VM_FAULT_LOCKED))) {
2571                                 lock_page(old_page);
2572                                 if (!old_page->mapping) {
2573                                         ret = 0; /* retry the fault */
2574                                         unlock_page(old_page);
2575                                         goto unwritable_page;
2576                                 }
2577                         } else
2578                                 VM_BUG_ON(!PageLocked(old_page));
2579
2580                         /*
2581                          * Since we dropped the lock we need to revalidate
2582                          * the PTE as someone else may have changed it.  If
2583                          * they did, we just return, as we can count on the
2584                          * MMU to tell us if they didn't also make it writable.
2585                          */
2586                         page_table = pte_offset_map_lock(mm, pmd, address,
2587                                                          &ptl);
2588                         if (!pte_same(*page_table, orig_pte)) {
2589                                 unlock_page(old_page);
2590                                 goto unlock;
2591                         }
2592
2593                         page_mkwrite = 1;
2594                 }
2595                 dirty_page = old_page;
2596                 get_page(dirty_page);
2597
2598 reuse:
2599                 flush_cache_page(vma, address, pte_pfn(orig_pte));
2600                 entry = pte_mkyoung(orig_pte);
2601                 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2602                 if (ptep_set_access_flags(vma, address, page_table, entry,1))
2603                         update_mmu_cache(vma, address, page_table);
2604                 pte_unmap_unlock(page_table, ptl);
2605                 ret |= VM_FAULT_WRITE;
2606
2607                 if (!dirty_page)
2608                         return ret;
2609
2610                 /*
2611                  * Yes, Virginia, this is actually required to prevent a race
2612                  * with clear_page_dirty_for_io() from clearing the page dirty
2613                  * bit after it clear all dirty ptes, but before a racing
2614                  * do_wp_page installs a dirty pte.
2615                  *
2616                  * __do_fault is protected similarly.
2617                  */
2618                 if (!page_mkwrite) {
2619                         wait_on_page_locked(dirty_page);
2620                         set_page_dirty_balance(dirty_page, page_mkwrite);
2621                 }
2622                 put_page(dirty_page);
2623                 if (page_mkwrite) {
2624                         struct address_space *mapping = dirty_page->mapping;
2625
2626                         set_page_dirty(dirty_page);
2627                         unlock_page(dirty_page);
2628                         page_cache_release(dirty_page);
2629                         if (mapping)    {
2630                                 /*
2631                                  * Some device drivers do not set page.mapping
2632                                  * but still dirty their pages
2633                                  */
2634                                 balance_dirty_pages_ratelimited(mapping);
2635                         }
2636                 }
2637
2638                 /* file_update_time outside page_lock */
2639                 if (vma->vm_file)
2640                         file_update_time(vma->vm_file);
2641
2642                 return ret;
2643         }
2644
2645         /*
2646          * Ok, we need to copy. Oh, well..
2647          */
2648         page_cache_get(old_page);
2649 gotten:
2650         pte_unmap_unlock(page_table, ptl);
2651
2652         if (unlikely(anon_vma_prepare(vma)))
2653                 goto oom;
2654
2655         if (is_zero_pfn(pte_pfn(orig_pte))) {
2656                 new_page = alloc_zeroed_user_highpage_movable(vma, address);
2657                 if (!new_page)
2658                         goto oom;
2659         } else {
2660                 new_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, address);
2661                 if (!new_page)
2662                         goto oom;
2663                 cow_user_page(new_page, old_page, address, vma);
2664         }
2665         __SetPageUptodate(new_page);
2666
2667         if (mem_cgroup_newpage_charge(new_page, mm, GFP_KERNEL))
2668                 goto oom_free_new;
2669
2670         /*
2671          * Re-check the pte - we dropped the lock
2672          */
2673         page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2674         if (likely(pte_same(*page_table, orig_pte))) {
2675                 if (old_page) {
2676                         if (!PageAnon(old_page)) {
2677                                 dec_mm_counter_fast(mm, MM_FILEPAGES);
2678                                 inc_mm_counter_fast(mm, MM_ANONPAGES);
2679                         }
2680                 } else
2681                         inc_mm_counter_fast(mm, MM_ANONPAGES);
2682                 flush_cache_page(vma, address, pte_pfn(orig_pte));
2683                 entry = mk_pte(new_page, vma->vm_page_prot);
2684                 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2685                 /*
2686                  * Clear the pte entry and flush it first, before updating the
2687                  * pte with the new entry. This will avoid a race condition
2688                  * seen in the presence of one thread doing SMC and another
2689                  * thread doing COW.
2690                  */
2691                 ptep_clear_flush(vma, address, page_table);
2692                 page_add_new_anon_rmap(new_page, vma, address);
2693                 /*
2694                  * We call the notify macro here because, when using secondary
2695                  * mmu page tables (such as kvm shadow page tables), we want the
2696                  * new page to be mapped directly into the secondary page table.
2697                  */
2698                 set_pte_at_notify(mm, address, page_table, entry);
2699                 update_mmu_cache(vma, address, page_table);
2700                 if (old_page) {
2701                         /*
2702                          * Only after switching the pte to the new page may
2703                          * we remove the mapcount here. Otherwise another
2704                          * process may come and find the rmap count decremented
2705                          * before the pte is switched to the new page, and
2706                          * "reuse" the old page writing into it while our pte
2707                          * here still points into it and can be read by other
2708                          * threads.
2709                          *
2710                          * The critical issue is to order this
2711                          * page_remove_rmap with the ptp_clear_flush above.
2712                          * Those stores are ordered by (if nothing else,)
2713                          * the barrier present in the atomic_add_negative
2714                          * in page_remove_rmap.
2715                          *
2716                          * Then the TLB flush in ptep_clear_flush ensures that
2717                          * no process can access the old page before the
2718                          * decremented mapcount is visible. And the old page
2719                          * cannot be reused until after the decremented
2720                          * mapcount is visible. So transitively, TLBs to
2721                          * old page will be flushed before it can be reused.
2722                          */
2723                         page_remove_rmap(old_page);
2724                 }
2725
2726                 /* Free the old page.. */
2727                 new_page = old_page;
2728                 ret |= VM_FAULT_WRITE;
2729         } else
2730                 mem_cgroup_uncharge_page(new_page);
2731
2732         if (new_page)
2733                 page_cache_release(new_page);
2734 unlock:
2735         pte_unmap_unlock(page_table, ptl);
2736         if (old_page) {
2737                 /*
2738                  * Don't let another task, with possibly unlocked vma,
2739                  * keep the mlocked page.
2740                  */
2741                 if ((ret & VM_FAULT_WRITE) && (vma->vm_flags & VM_LOCKED)) {
2742                         lock_page(old_page);    /* LRU manipulation */
2743                         munlock_vma_page(old_page);
2744                         unlock_page(old_page);
2745                 }
2746                 page_cache_release(old_page);
2747         }
2748         return ret;
2749 oom_free_new:
2750         page_cache_release(new_page);
2751 oom:
2752         if (old_page) {
2753                 if (page_mkwrite) {
2754                         unlock_page(old_page);
2755                         page_cache_release(old_page);
2756                 }
2757                 page_cache_release(old_page);
2758         }
2759         return VM_FAULT_OOM;
2760
2761 unwritable_page:
2762         page_cache_release(old_page);
2763         return ret;
2764 }
2765
2766 static void unmap_mapping_range_vma(struct vm_area_struct *vma,
2767                 unsigned long start_addr, unsigned long end_addr,
2768                 struct zap_details *details)
2769 {
2770         zap_page_range(vma, start_addr, end_addr - start_addr, details);
2771 }
2772
2773 static inline void unmap_mapping_range_tree(struct prio_tree_root *root,
2774                                             struct zap_details *details)
2775 {
2776         struct vm_area_struct *vma;
2777         struct prio_tree_iter iter;
2778         pgoff_t vba, vea, zba, zea;
2779
2780         vma_prio_tree_foreach(vma, &iter, root,
2781                         details->first_index, details->last_index) {
2782
2783                 vba = vma->vm_pgoff;
2784                 vea = vba + ((vma->vm_end - vma->vm_start) >> PAGE_SHIFT) - 1;
2785                 /* Assume for now that PAGE_CACHE_SHIFT == PAGE_SHIFT */
2786                 zba = details->first_index;
2787                 if (zba < vba)
2788                         zba = vba;
2789                 zea = details->last_index;
2790                 if (zea > vea)
2791                         zea = vea;
2792
2793                 unmap_mapping_range_vma(vma,
2794                         ((zba - vba) << PAGE_SHIFT) + vma->vm_start,
2795                         ((zea - vba + 1) << PAGE_SHIFT) + vma->vm_start,
2796                                 details);
2797         }
2798 }
2799
2800 static inline void unmap_mapping_range_list(struct list_head *head,
2801                                             struct zap_details *details)
2802 {
2803         struct vm_area_struct *vma;
2804
2805         /*
2806          * In nonlinear VMAs there is no correspondence between virtual address
2807          * offset and file offset.  So we must perform an exhaustive search
2808          * across *all* the pages in each nonlinear VMA, not just the pages
2809          * whose virtual address lies outside the file truncation point.
2810          */
2811         list_for_each_entry(vma, head, shared.vm_set.list) {
2812                 details->nonlinear_vma = vma;
2813                 unmap_mapping_range_vma(vma, vma->vm_start, vma->vm_end, details);
2814         }
2815 }
2816
2817 /**
2818  * unmap_mapping_range - unmap the portion of all mmaps in the specified address_space corresponding to the specified page range in the underlying file.
2819  * @mapping: the address space containing mmaps to be unmapped.
2820  * @holebegin: byte in first page to unmap, relative to the start of
2821  * the underlying file.  This will be rounded down to a PAGE_SIZE
2822  * boundary.  Note that this is different from truncate_pagecache(), which
2823  * must keep the partial page.  In contrast, we must get rid of
2824  * partial pages.
2825  * @holelen: size of prospective hole in bytes.  This will be rounded
2826  * up to a PAGE_SIZE boundary.  A holelen of zero truncates to the
2827  * end of the file.
2828  * @even_cows: 1 when truncating a file, unmap even private COWed pages;
2829  * but 0 when invalidating pagecache, don't throw away private data.
2830  */
2831 void unmap_mapping_range(struct address_space *mapping,
2832                 loff_t const holebegin, loff_t const holelen, int even_cows)
2833 {
2834         struct zap_details details;
2835         pgoff_t hba = holebegin >> PAGE_SHIFT;
2836         pgoff_t hlen = (holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
2837
2838         /* Check for overflow. */
2839         if (sizeof(holelen) > sizeof(hlen)) {
2840                 long long holeend =
2841                         (holebegin + holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
2842                 if (holeend & ~(long long)ULONG_MAX)
2843                         hlen = ULONG_MAX - hba + 1;
2844         }
2845
2846         details.check_mapping = even_cows? NULL: mapping;
2847         details.nonlinear_vma = NULL;
2848         details.first_index = hba;
2849         details.last_index = hba + hlen - 1;
2850         if (details.last_index < details.first_index)
2851                 details.last_index = ULONG_MAX;
2852
2853
2854         mutex_lock(&mapping->i_mmap_mutex);
2855         if (unlikely(!prio_tree_empty(&mapping->i_mmap)))
2856                 unmap_mapping_range_tree(&mapping->i_mmap, &details);
2857         if (unlikely(!list_empty(&mapping->i_mmap_nonlinear)))
2858                 unmap_mapping_range_list(&mapping->i_mmap_nonlinear, &details);
2859         mutex_unlock(&mapping->i_mmap_mutex);
2860 }
2861 EXPORT_SYMBOL(unmap_mapping_range);
2862
2863 /*
2864  * We enter with non-exclusive mmap_sem (to exclude vma changes,
2865  * but allow concurrent faults), and pte mapped but not yet locked.
2866  * We return with mmap_sem still held, but pte unmapped and unlocked.
2867  */
2868 static int do_swap_page(struct mm_struct *mm, struct vm_area_struct *vma,
2869                 unsigned long address, pte_t *page_table, pmd_t *pmd,
2870                 unsigned int flags, pte_t orig_pte)
2871 {
2872         spinlock_t *ptl;
2873         struct page *page, *swapcache = NULL;
2874         swp_entry_t entry;
2875         pte_t pte;
2876         int locked;
2877         struct mem_cgroup *ptr;
2878         int exclusive = 0;
2879         int ret = 0;
2880
2881         if (!pte_unmap_same(mm, pmd, page_table, orig_pte))
2882                 goto out;
2883
2884         entry = pte_to_swp_entry(orig_pte);
2885         if (unlikely(non_swap_entry(entry))) {
2886                 if (is_migration_entry(entry)) {
2887                         migration_entry_wait(mm, pmd, address);
2888                 } else if (is_hwpoison_entry(entry)) {
2889                         ret = VM_FAULT_HWPOISON;
2890                 } else {
2891                         print_bad_pte(vma, address, orig_pte, NULL);
2892                         ret = VM_FAULT_SIGBUS;
2893                 }
2894                 goto out;
2895         }
2896         delayacct_set_flag(DELAYACCT_PF_SWAPIN);
2897         page = lookup_swap_cache(entry);
2898         if (!page) {
2899                 grab_swap_token(mm); /* Contend for token _before_ read-in */
2900                 page = swapin_readahead(entry,
2901                                         GFP_HIGHUSER_MOVABLE, vma, address);
2902                 if (!page) {
2903                         /*
2904                          * Back out if somebody else faulted in this pte
2905                          * while we released the pte lock.
2906                          */
2907                         page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2908                         if (likely(pte_same(*page_table, orig_pte)))
2909                                 ret = VM_FAULT_OOM;
2910                         delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
2911                         goto unlock;
2912                 }
2913
2914                 /* Had to read the page from swap area: Major fault */
2915                 ret = VM_FAULT_MAJOR;
2916                 count_vm_event(PGMAJFAULT);
2917                 mem_cgroup_count_vm_event(mm, PGMAJFAULT);
2918         } else if (PageHWPoison(page)) {
2919                 /*
2920                  * hwpoisoned dirty swapcache pages are kept for killing
2921                  * owner processes (which may be unknown at hwpoison time)
2922                  */
2923                 ret = VM_FAULT_HWPOISON;
2924                 delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
2925                 goto out_release;
2926         }
2927
2928         locked = lock_page_or_retry(page, mm, flags);
2929         delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
2930         if (!locked) {
2931                 ret |= VM_FAULT_RETRY;
2932                 goto out_release;
2933         }
2934
2935         /*
2936          * Make sure try_to_free_swap or reuse_swap_page or swapoff did not
2937          * release the swapcache from under us.  The page pin, and pte_same
2938          * test below, are not enough to exclude that.  Even if it is still
2939          * swapcache, we need to check that the page's swap has not changed.
2940          */
2941         if (unlikely(!PageSwapCache(page) || page_private(page) != entry.val))
2942                 goto out_page;
2943
2944         if (ksm_might_need_to_copy(page, vma, address)) {
2945                 swapcache = page;
2946                 page = ksm_does_need_to_copy(page, vma, address);
2947
2948                 if (unlikely(!page)) {
2949                         ret = VM_FAULT_OOM;
2950                         page = swapcache;
2951                         swapcache = NULL;
2952                         goto out_page;
2953                 }
2954         }
2955
2956         if (mem_cgroup_try_charge_swapin(mm, page, GFP_KERNEL, &ptr)) {
2957                 ret = VM_FAULT_OOM;
2958                 goto out_page;
2959         }
2960
2961         /*
2962          * Back out if somebody else already faulted in this pte.
2963          */
2964         page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2965         if (unlikely(!pte_same(*page_table, orig_pte)))
2966                 goto out_nomap;
2967
2968         if (unlikely(!PageUptodate(page))) {
2969                 ret = VM_FAULT_SIGBUS;
2970                 goto out_nomap;
2971         }
2972
2973         /*
2974          * The page isn't present yet, go ahead with the fault.
2975          *
2976          * Be careful about the sequence of operations here.
2977          * To get its accounting right, reuse_swap_page() must be called
2978          * while the page is counted on swap but not yet in mapcount i.e.
2979          * before page_add_anon_rmap() and swap_free(); try_to_free_swap()
2980          * must be called after the swap_free(), or it will never succeed.
2981          * Because delete_from_swap_page() may be called by reuse_swap_page(),
2982          * mem_cgroup_commit_charge_swapin() may not be able to find swp_entry
2983          * in page->private. In this case, a record in swap_cgroup  is silently
2984          * discarded at swap_free().
2985          */
2986
2987         inc_mm_counter_fast(mm, MM_ANONPAGES);
2988         dec_mm_counter_fast(mm, MM_SWAPENTS);
2989         pte = mk_pte(page, vma->vm_page_prot);
2990         if ((flags & FAULT_FLAG_WRITE) && reuse_swap_page(page)) {
2991                 pte = maybe_mkwrite(pte_mkdirty(pte), vma);
2992                 flags &= ~FAULT_FLAG_WRITE;
2993                 ret |= VM_FAULT_WRITE;
2994                 exclusive = 1;
2995         }
2996         flush_icache_page(vma, page);
2997         set_pte_at(mm, address, page_table, pte);
2998         do_page_add_anon_rmap(page, vma, address, exclusive);
2999         /* It's better to call commit-charge after rmap is established */
3000         mem_cgroup_commit_charge_swapin(page, ptr);
3001
3002         swap_free(entry);
3003         if (vm_swap_full() || (vma->vm_flags & VM_LOCKED) || PageMlocked(page))
3004                 try_to_free_swap(page);
3005         unlock_page(page);
3006         if (swapcache) {
3007                 /*
3008                  * Hold the lock to avoid the swap entry to be reused
3009                  * until we take the PT lock for the pte_same() check
3010                  * (to avoid false positives from pte_same). For
3011                  * further safety release the lock after the swap_free
3012                  * so that the swap count won't change under a
3013                  * parallel locked swapcache.
3014                  */
3015                 unlock_page(swapcache);
3016                 page_cache_release(swapcache);
3017         }
3018
3019         if (flags & FAULT_FLAG_WRITE) {
3020                 ret |= do_wp_page(mm, vma, address, page_table, pmd, ptl, pte);
3021                 if (ret & VM_FAULT_ERROR)
3022                         ret &= VM_FAULT_ERROR;
3023                 goto out;
3024         }
3025
3026         /* No need to invalidate - it was non-present before */
3027         update_mmu_cache(vma, address, page_table);
3028 unlock:
3029         pte_unmap_unlock(page_table, ptl);
3030 out:
3031         return ret;
3032 out_nomap:
3033         mem_cgroup_cancel_charge_swapin(ptr);
3034         pte_unmap_unlock(page_table, ptl);
3035 out_page:
3036         unlock_page(page);
3037 out_release:
3038         page_cache_release(page);
3039         if (swapcache) {
3040                 unlock_page(swapcache);
3041                 page_cache_release(swapcache);
3042         }
3043         return ret;
3044 }
3045
3046 /*
3047  * This is like a special single-page "expand_{down|up}wards()",
3048  * except we must first make sure that 'address{-|+}PAGE_SIZE'
3049  * doesn't hit another vma.
3050  */
3051 static inline int check_stack_guard_page(struct vm_area_struct *vma, unsigned long address)
3052 {
3053         address &= PAGE_MASK;
3054         if ((vma->vm_flags & VM_GROWSDOWN) && address == vma->vm_start) {
3055                 struct vm_area_struct *prev = vma->vm_prev;
3056
3057                 /*
3058                  * Is there a mapping abutting this one below?
3059                  *
3060                  * That's only ok if it's the same stack mapping
3061                  * that has gotten split..
3062                  */
3063                 if (prev && prev->vm_end == address)
3064                         return prev->vm_flags & VM_GROWSDOWN ? 0 : -ENOMEM;
3065
3066                 expand_downwards(vma, address - PAGE_SIZE);
3067         }
3068         if ((vma->vm_flags & VM_GROWSUP) && address + PAGE_SIZE == vma->vm_end) {
3069                 struct vm_area_struct *next = vma->vm_next;
3070
3071                 /* As VM_GROWSDOWN but s/below/above/ */
3072                 if (next && next->vm_start == address + PAGE_SIZE)
3073                         return next->vm_flags & VM_GROWSUP ? 0 : -ENOMEM;
3074
3075                 expand_upwards(vma, address + PAGE_SIZE);
3076         }
3077         return 0;
3078 }
3079
3080 /*
3081  * We enter with non-exclusive mmap_sem (to exclude vma changes,
3082  * but allow concurrent faults), and pte mapped but not yet locked.
3083  * We return with mmap_sem still held, but pte unmapped and unlocked.
3084  */
3085 static int do_anonymous_page(struct mm_struct *mm, struct vm_area_struct *vma,
3086                 unsigned long address, pte_t *page_table, pmd_t *pmd,
3087                 unsigned int flags)
3088 {
3089         struct page *page;
3090         spinlock_t *ptl;
3091         pte_t entry;
3092
3093         pte_unmap(page_table);
3094
3095         /* Check if we need to add a guard page to the stack */
3096         if (check_stack_guard_page(vma, address) < 0)
3097                 return VM_FAULT_SIGBUS;
3098
3099         /* Use the zero-page for reads */
3100         if (!(flags & FAULT_FLAG_WRITE)) {
3101                 entry = pte_mkspecial(pfn_pte(my_zero_pfn(address),
3102                                                 vma->vm_page_prot));
3103                 page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
3104                 if (!pte_none(*page_table))
3105                         goto unlock;
3106                 goto setpte;
3107         }
3108
3109         /* Allocate our own private page. */
3110         if (unlikely(anon_vma_prepare(vma)))
3111                 goto oom;
3112         page = alloc_zeroed_user_highpage_movable(vma, address);
3113         if (!page)
3114                 goto oom;
3115         __SetPageUptodate(page);
3116
3117         if (mem_cgroup_newpage_charge(page, mm, GFP_KERNEL))
3118                 goto oom_free_page;
3119
3120         entry = mk_pte(page, vma->vm_page_prot);
3121         if (vma->vm_flags & VM_WRITE)
3122                 entry = pte_mkwrite(pte_mkdirty(entry));
3123
3124         page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
3125         if (!pte_none(*page_table))
3126                 goto release;
3127
3128         inc_mm_counter_fast(mm, MM_ANONPAGES);
3129         page_add_new_anon_rmap(page, vma, address);
3130 setpte:
3131         set_pte_at(mm, address, page_table, entry);
3132
3133         /* No need to invalidate - it was non-present before */
3134         update_mmu_cache(vma, address, page_table);
3135 unlock:
3136         pte_unmap_unlock(page_table, ptl);
3137         return 0;
3138 release:
3139         mem_cgroup_uncharge_page(page);
3140         page_cache_release(page);
3141         goto unlock;
3142 oom_free_page:
3143         page_cache_release(page);
3144 oom:
3145         return VM_FAULT_OOM;
3146 }
3147
3148 /*
3149  * __do_fault() tries to create a new page mapping. It aggressively
3150  * tries to share with existing pages, but makes a separate copy if
3151  * the FAULT_FLAG_WRITE is set in the flags parameter in order to avoid
3152  * the next page fault.
3153  *
3154  * As this is called only for pages that do not currently exist, we
3155  * do not need to flush old virtual caches or the TLB.
3156  *
3157  * We enter with non-exclusive mmap_sem (to exclude vma changes,
3158  * but allow concurrent faults), and pte neither mapped nor locked.
3159  * We return with mmap_sem still held, but pte unmapped and unlocked.
3160  */
3161 static int __do_fault(struct mm_struct *mm, struct vm_area_struct *vma,
3162                 unsigned long address, pmd_t *pmd,
3163                 pgoff_t pgoff, unsigned int flags, pte_t orig_pte)
3164 {
3165         pte_t *page_table;
3166         spinlock_t *ptl;
3167         struct page *page;
3168         struct page *cow_page;
3169         pte_t entry;
3170         int anon = 0;
3171         struct page *dirty_page = NULL;
3172         struct vm_fault vmf;
3173         int ret;
3174         int page_mkwrite = 0;
3175
3176         /*
3177          * If we do COW later, allocate page befor taking lock_page()
3178          * on the file cache page. This will reduce lock holding time.
3179          */
3180         if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) {
3181
3182                 if (unlikely(anon_vma_prepare(vma)))
3183                         return VM_FAULT_OOM;
3184
3185                 cow_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, address);
3186                 if (!cow_page)
3187                         return VM_FAULT_OOM;
3188
3189                 if (mem_cgroup_newpage_charge(cow_page, mm, GFP_KERNEL)) {
3190                         page_cache_release(cow_page);
3191                         return VM_FAULT_OOM;
3192                 }
3193         } else
3194                 cow_page = NULL;
3195
3196         vmf.virtual_address = (void __user *)(address & PAGE_MASK);
3197         vmf.pgoff = pgoff;
3198         vmf.flags = flags;
3199         vmf.page = NULL;
3200
3201         ret = vma->vm_ops->fault(vma, &vmf);
3202         if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE |
3203                             VM_FAULT_RETRY)))
3204                 goto uncharge_out;
3205
3206         if (unlikely(PageHWPoison(vmf.page))) {
3207                 if (ret & VM_FAULT_LOCKED)
3208                         unlock_page(vmf.page);
3209                 ret = VM_FAULT_HWPOISON;
3210                 goto uncharge_out;
3211         }
3212
3213         /*
3214          * For consistency in subsequent calls, make the faulted page always
3215          * locked.
3216          */
3217         if (unlikely(!(ret & VM_FAULT_LOCKED)))
3218                 lock_page(vmf.page);
3219         else
3220                 VM_BUG_ON(!PageLocked(vmf.page));
3221
3222         /*
3223          * Should we do an early C-O-W break?
3224          */
3225         page = vmf.page;
3226         if (flags & FAULT_FLAG_WRITE) {
3227                 if (!(vma->vm_flags & VM_SHARED)) {
3228                         page = cow_page;
3229                         anon = 1;
3230                         copy_user_highpage(page, vmf.page, address, vma);
3231                         __SetPageUptodate(page);
3232                 } else {
3233                         /*
3234                          * If the page will be shareable, see if the backing
3235                          * address space wants to know that the page is about
3236                          * to become writable
3237                          */
3238                         if (vma->vm_ops->page_mkwrite) {
3239                                 int tmp;
3240
3241                                 unlock_page(page);
3242                                 vmf.flags = FAULT_FLAG_WRITE|FAULT_FLAG_MKWRITE;
3243                                 tmp = vma->vm_ops->page_mkwrite(vma, &vmf);
3244                                 if (unlikely(tmp &
3245                                           (VM_FAULT_ERROR | VM_FAULT_NOPAGE))) {
3246                                         ret = tmp;
3247                                         goto unwritable_page;
3248                                 }
3249                                 if (unlikely(!(tmp & VM_FAULT_LOCKED))) {
3250                                         lock_page(page);
3251                                         if (!page->mapping) {
3252                                                 ret = 0; /* retry the fault */
3253                                                 unlock_page(page);
3254                                                 goto unwritable_page;
3255                                         }
3256                                 } else
3257                                         VM_BUG_ON(!PageLocked(page));
3258                                 page_mkwrite = 1;
3259                         }
3260                 }
3261
3262         }
3263
3264         page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
3265
3266         /*
3267          * This silly early PAGE_DIRTY setting removes a race
3268          * due to the bad i386 page protection. But it's valid
3269          * for other architectures too.
3270          *
3271          * Note that if FAULT_FLAG_WRITE is set, we either now have
3272          * an exclusive copy of the page, or this is a shared mapping,
3273          * so we can make it writable and dirty to avoid having to
3274          * handle that later.
3275          */
3276         /* Only go through if we didn't race with anybody else... */
3277         if (likely(pte_same(*page_table, orig_pte))) {
3278                 flush_icache_page(vma, page);
3279                 entry = mk_pte(page, vma->vm_page_prot);
3280                 if (flags & FAULT_FLAG_WRITE)
3281                         entry = maybe_mkwrite(pte_mkdirty(entry), vma);
3282                 if (anon) {
3283                         inc_mm_counter_fast(mm, MM_ANONPAGES);
3284                         page_add_new_anon_rmap(page, vma, address);
3285                 } else {
3286                         inc_mm_counter_fast(mm, MM_FILEPAGES);
3287                         page_add_file_rmap(page);
3288                         if (flags & FAULT_FLAG_WRITE) {
3289                                 dirty_page = page;
3290                                 get_page(dirty_page);
3291                         }
3292                 }
3293                 set_pte_at(mm, address, page_table, entry);
3294
3295                 /* no need to invalidate: a not-present page won't be cached */
3296                 update_mmu_cache(vma, address, page_table);
3297         } else {
3298                 if (cow_page)
3299                         mem_cgroup_uncharge_page(cow_page);
3300                 if (anon)
3301                         page_cache_release(page);
3302                 else
3303                         anon = 1; /* no anon but release faulted_page */
3304         }
3305
3306         pte_unmap_unlock(page_table, ptl);
3307
3308         if (dirty_page) {
3309                 struct address_space *mapping = page->mapping;
3310
3311                 if (set_page_dirty(dirty_page))
3312                         page_mkwrite = 1;
3313                 unlock_page(dirty_page);
3314                 put_page(dirty_page);
3315                 if (page_mkwrite && mapping) {
3316                         /*
3317                          * Some device drivers do not set page.mapping but still
3318                          * dirty their pages
3319                          */
3320                         balance_dirty_pages_ratelimited(mapping);
3321                 }
3322
3323                 /* file_update_time outside page_lock */
3324                 if (vma->vm_file)
3325                         file_update_time(vma->vm_file);
3326         } else {
3327                 unlock_page(vmf.page);
3328                 if (anon)
3329                         page_cache_release(vmf.page);
3330         }
3331
3332         return ret;
3333
3334 unwritable_page:
3335         page_cache_release(page);
3336         return ret;
3337 uncharge_out:
3338         /* fs's fault handler get error */
3339         if (cow_page) {
3340                 mem_cgroup_uncharge_page(cow_page);
3341                 page_cache_release(cow_page);
3342         }
3343         return ret;
3344 }
3345
3346 static int do_linear_fault(struct mm_struct *mm, struct vm_area_struct *vma,
3347                 unsigned long address, pte_t *page_table, pmd_t *pmd,
3348                 unsigned int flags, pte_t orig_pte)
3349 {
3350         pgoff_t pgoff = (((address & PAGE_MASK)
3351                         - vma->vm_start) >> PAGE_SHIFT) + vma->vm_pgoff;
3352
3353         pte_unmap(page_table);
3354         return __do_fault(mm, vma, address, pmd, pgoff, flags, orig_pte);
3355 }
3356
3357 /*
3358  * Fault of a previously existing named mapping. Repopulate the pte
3359  * from the encoded file_pte if possible. This enables swappable
3360  * nonlinear vmas.
3361  *
3362  * We enter with non-exclusive mmap_sem (to exclude vma changes,
3363  * but allow concurrent faults), and pte mapped but not yet locked.
3364  * We return with mmap_sem still held, but pte unmapped and unlocked.
3365  */
3366 static int do_nonlinear_fault(struct mm_struct *mm, struct vm_area_struct *vma,
3367                 unsigned long address, pte_t *page_table, pmd_t *pmd,
3368                 unsigned int flags, pte_t orig_pte)
3369 {
3370         pgoff_t pgoff;
3371
3372         flags |= FAULT_FLAG_NONLINEAR;
3373
3374         if (!pte_unmap_same(mm, pmd, page_table, orig_pte))
3375                 return 0;
3376
3377         if (unlikely(!(vma->vm_flags & VM_NONLINEAR))) {
3378                 /*
3379                  * Page table corrupted: show pte and kill process.
3380                  */
3381                 print_bad_pte(vma, address, orig_pte, NULL);
3382                 return VM_FAULT_SIGBUS;
3383         }
3384
3385         pgoff = pte_to_pgoff(orig_pte);
3386         return __do_fault(mm, vma, address, pmd, pgoff, flags, orig_pte);
3387 }
3388
3389 /*
3390  * These routines also need to handle stuff like marking pages dirty
3391  * and/or accessed for architectures that don't do it in hardware (most
3392  * RISC architectures).  The early dirtying is also good on the i386.
3393  *
3394  * There is also a hook called "update_mmu_cache()" that architectures
3395  * with external mmu caches can use to update those (ie the Sparc or
3396  * PowerPC hashed page tables that act as extended TLBs).
3397  *
3398  * We enter with non-exclusive mmap_sem (to exclude vma changes,
3399  * but allow concurrent faults), and pte mapped but not yet locked.
3400  * We return with mmap_sem still held, but pte unmapped and unlocked.
3401  */
3402 int handle_pte_fault(struct mm_struct *mm,
3403                      struct vm_area_struct *vma, unsigned long address,
3404                      pte_t *pte, pmd_t *pmd, unsigned int flags)
3405 {
3406         pte_t entry;
3407         spinlock_t *ptl;
3408
3409         entry = *pte;
3410         if (!pte_present(entry)) {
3411                 if (pte_none(entry)) {
3412                         if (vma->vm_ops) {
3413                                 if (likely(vma->vm_ops->fault))
3414                                         return do_linear_fault(mm, vma, address,
3415                                                 pte, pmd, flags, entry);
3416                         }
3417                         return do_anonymous_page(mm, vma, address,
3418                                                  pte, pmd, flags);
3419                 }
3420                 if (pte_file(entry))
3421                         return do_nonlinear_fault(mm, vma, address,
3422                                         pte, pmd, flags, entry);
3423                 return do_swap_page(mm, vma, address,
3424                                         pte, pmd, flags, entry);
3425         }
3426
3427         ptl = pte_lockptr(mm, pmd);
3428         spin_lock(ptl);
3429         if (unlikely(!pte_same(*pte, entry)))
3430                 goto unlock;
3431         if (flags & FAULT_FLAG_WRITE) {
3432                 if (!pte_write(entry))
3433                         return do_wp_page(mm, vma, address,
3434                                         pte, pmd, ptl, entry);
3435                 entry = pte_mkdirty(entry);
3436         }
3437         entry = pte_mkyoung(entry);
3438         if (ptep_set_access_flags(vma, address, pte, entry, flags & FAULT_FLAG_WRITE)) {
3439                 update_mmu_cache(vma, address, pte);
3440         } else {
3441                 /*
3442                  * This is needed only for protection faults but the arch code
3443                  * is not yet telling us if this is a protection fault or not.
3444                  * This still avoids useless tlb flushes for .text page faults
3445                  * with threads.
3446                  */
3447                 if (flags & FAULT_FLAG_WRITE)
3448                         flush_tlb_fix_spurious_fault(vma, address);
3449         }
3450 unlock:
3451         pte_unmap_unlock(pte, ptl);
3452         return 0;
3453 }
3454
3455 /*
3456  * By the time we get here, we already hold the mm semaphore
3457  */
3458 int handle_mm_fault(struct mm_struct *mm, struct vm_area_struct *vma,
3459                 unsigned long address, unsigned int flags)
3460 {
3461         pgd_t *pgd;
3462         pud_t *pud;
3463         pmd_t *pmd;
3464         pte_t *pte;
3465
3466         __set_current_state(TASK_RUNNING);
3467
3468         count_vm_event(PGFAULT);
3469         mem_cgroup_count_vm_event(mm, PGFAULT);
3470
3471         /* do counter updates before entering really critical section. */
3472         check_sync_rss_stat(current);
3473
3474         if (unlikely(is_vm_hugetlb_page(vma)))
3475                 return hugetlb_fault(mm, vma, address, flags);
3476
3477         pgd = pgd_offset(mm, address);
3478         pud = pud_alloc(mm, pgd, address);
3479         if (!pud)
3480                 return VM_FAULT_OOM;
3481         pmd = pmd_alloc(mm, pud, address);
3482         if (!pmd)
3483                 return VM_FAULT_OOM;
3484         if (pmd_none(*pmd) && transparent_hugepage_enabled(vma)) {
3485                 if (!vma->vm_ops)
3486                         return do_huge_pmd_anonymous_page(mm, vma, address,
3487                                                           pmd, flags);
3488         } else {
3489                 pmd_t orig_pmd = *pmd;
3490                 barrier();
3491                 if (pmd_trans_huge(orig_pmd)) {
3492                         if (flags & FAULT_FLAG_WRITE &&
3493                             !pmd_write(orig_pmd) &&
3494                             !pmd_trans_splitting(orig_pmd))
3495                                 return do_huge_pmd_wp_page(mm, vma, address,
3496                                                            pmd, orig_pmd);
3497                         return 0;
3498                 }
3499         }
3500
3501         /*
3502          * Use __pte_alloc instead of pte_alloc_map, because we can't
3503          * run pte_offset_map on the pmd, if an huge pmd could
3504          * materialize from under us from a different thread.
3505          */
3506         if (unlikely(pmd_none(*pmd)) && __pte_alloc(mm, vma, pmd, address))
3507                 return VM_FAULT_OOM;
3508         /* if an huge pmd materialized from under us just retry later */
3509         if (unlikely(pmd_trans_huge(*pmd)))
3510                 return 0;
3511         /*
3512          * A regular pmd is established and it can't morph into a huge pmd
3513          * from under us anymore at this point because we hold the mmap_sem
3514          * read mode and khugepaged takes it in write mode. So now it's
3515          * safe to run pte_offset_map().
3516          */
3517         pte = pte_offset_map(pmd, address);
3518
3519         return handle_pte_fault(mm, vma, address, pte, pmd, flags);
3520 }
3521
3522 #ifndef __PAGETABLE_PUD_FOLDED
3523 /*
3524  * Allocate page upper directory.
3525  * We've already handled the fast-path in-line.
3526  */
3527 int __pud_alloc(struct mm_struct *mm, pgd_t *pgd, unsigned long address)
3528 {
3529         pud_t *new = pud_alloc_one(mm, address);
3530         if (!new)
3531                 return -ENOMEM;
3532
3533         smp_wmb(); /* See comment in __pte_alloc */
3534
3535         spin_lock(&mm->page_table_lock);
3536         if (pgd_present(*pgd))          /* Another has populated it */
3537                 pud_free(mm, new);
3538         else
3539                 pgd_populate(mm, pgd, new);
3540         spin_unlock(&mm->page_table_lock);
3541         return 0;
3542 }
3543 #endif /* __PAGETABLE_PUD_FOLDED */
3544
3545 #ifndef __PAGETABLE_PMD_FOLDED
3546 /*
3547  * Allocate page middle directory.
3548  * We've already handled the fast-path in-line.
3549  */
3550 int __pmd_alloc(struct mm_struct *mm, pud_t *pud, unsigned long address)
3551 {
3552         pmd_t *new = pmd_alloc_one(mm, address);
3553         if (!new)
3554                 return -ENOMEM;
3555
3556         smp_wmb(); /* See comment in __pte_alloc */
3557
3558         spin_lock(&mm->page_table_lock);
3559 #ifndef __ARCH_HAS_4LEVEL_HACK
3560         if (pud_present(*pud))          /* Another has populated it */
3561                 pmd_free(mm, new);
3562         else
3563                 pud_populate(mm, pud, new);
3564 #else
3565         if (pgd_present(*pud))          /* Another has populated it */
3566                 pmd_free(mm, new);
3567         else
3568                 pgd_populate(mm, pud, new);
3569 #endif /* __ARCH_HAS_4LEVEL_HACK */
3570         spin_unlock(&mm->page_table_lock);
3571         return 0;
3572 }
3573 #endif /* __PAGETABLE_PMD_FOLDED */
3574
3575 int make_pages_present(unsigned long addr, unsigned long end)
3576 {
3577         int ret, len, write;
3578         struct vm_area_struct * vma;
3579
3580         vma = find_vma(current->mm, addr);
3581         if (!vma)
3582                 return -ENOMEM;
3583         /*
3584          * We want to touch writable mappings with a write fault in order
3585          * to break COW, except for shared mappings because these don't COW
3586          * and we would not want to dirty them for nothing.
3587          */
3588         write = (vma->vm_flags & (VM_WRITE | VM_SHARED)) == VM_WRITE;
3589         BUG_ON(addr >= end);
3590         BUG_ON(end > vma->vm_end);
3591         len = DIV_ROUND_UP(end, PAGE_SIZE) - addr/PAGE_SIZE;
3592         ret = get_user_pages(current, current->mm, addr,
3593                         len, write, 0, NULL, NULL);
3594         if (ret < 0)
3595                 return ret;
3596         return ret == len ? 0 : -EFAULT;
3597 }
3598
3599 #if !defined(__HAVE_ARCH_GATE_AREA)
3600
3601 #if defined(AT_SYSINFO_EHDR)
3602 static struct vm_area_struct gate_vma;
3603
3604 static int __init gate_vma_init(void)
3605 {
3606         gate_vma.vm_mm = NULL;
3607         gate_vma.vm_start = FIXADDR_USER_START;
3608         gate_vma.vm_end = FIXADDR_USER_END;
3609         gate_vma.vm_flags = VM_READ | VM_MAYREAD | VM_EXEC | VM_MAYEXEC;
3610         gate_vma.vm_page_prot = __P101;
3611         /*
3612          * Make sure the vDSO gets into every core dump.
3613          * Dumping its contents makes post-mortem fully interpretable later
3614          * without matching up the same kernel and hardware config to see
3615          * what PC values meant.
3616          */
3617         gate_vma.vm_flags |= VM_ALWAYSDUMP;
3618         return 0;
3619 }
3620 __initcall(gate_vma_init);
3621 #endif
3622
3623 struct vm_area_struct *get_gate_vma(struct mm_struct *mm)
3624 {
3625 #ifdef AT_SYSINFO_EHDR
3626         return &gate_vma;
3627 #else
3628         return NULL;
3629 #endif
3630 }
3631
3632 int in_gate_area_no_mm(unsigned long addr)
3633 {
3634 #ifdef AT_SYSINFO_EHDR
3635         if ((addr >= FIXADDR_USER_START) && (addr < FIXADDR_USER_END))
3636                 return 1;
3637 #endif
3638         return 0;
3639 }
3640
3641 #endif  /* __HAVE_ARCH_GATE_AREA */
3642
3643 static int __follow_pte(struct mm_struct *mm, unsigned long address,
3644                 pte_t **ptepp, spinlock_t **ptlp)
3645 {
3646         pgd_t *pgd;
3647         pud_t *pud;
3648         pmd_t *pmd;
3649         pte_t *ptep;
3650
3651         pgd = pgd_offset(mm, address);
3652         if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
3653                 goto out;
3654
3655         pud = pud_offset(pgd, address);
3656         if (pud_none(*pud) || unlikely(pud_bad(*pud)))
3657                 goto out;
3658
3659         pmd = pmd_offset(pud, address);
3660         VM_BUG_ON(pmd_trans_huge(*pmd));
3661         if (pmd_none(*pmd) || unlikely(pmd_bad(*pmd)))
3662                 goto out;
3663
3664         /* We cannot handle huge page PFN maps. Luckily they don't exist. */
3665         if (pmd_huge(*pmd))
3666                 goto out;
3667
3668         ptep = pte_offset_map_lock(mm, pmd, address, ptlp);
3669         if (!ptep)
3670                 goto out;
3671         if (!pte_present(*ptep))
3672                 goto unlock;
3673         *ptepp = ptep;
3674         return 0;
3675 unlock:
3676         pte_unmap_unlock(ptep, *ptlp);
3677 out:
3678         return -EINVAL;
3679 }
3680
3681 static inline int follow_pte(struct mm_struct *mm, unsigned long address,
3682                              pte_t **ptepp, spinlock_t **ptlp)
3683 {
3684         int res;
3685
3686         /* (void) is needed to make gcc happy */
3687         (void) __cond_lock(*ptlp,
3688                            !(res = __follow_pte(mm, address, ptepp, ptlp)));
3689         return res;
3690 }
3691
3692 /**
3693  * follow_pfn - look up PFN at a user virtual address
3694  * @vma: memory mapping
3695  * @address: user virtual address
3696  * @pfn: location to store found PFN
3697  *
3698  * Only IO mappings and raw PFN mappings are allowed.
3699  *
3700  * Returns zero and the pfn at @pfn on success, -ve otherwise.
3701  */
3702 int follow_pfn(struct vm_area_struct *vma, unsigned long address,
3703         unsigned long *pfn)
3704 {
3705         int ret = -EINVAL;
3706         spinlock_t *ptl;
3707         pte_t *ptep;
3708
3709         if (!(vma->vm_flags & (VM_IO | VM_PFNMAP)))
3710                 return ret;
3711
3712         ret = follow_pte(vma->vm_mm, address, &ptep, &ptl);
3713         if (ret)
3714                 return ret;
3715         *pfn = pte_pfn(*ptep);
3716         pte_unmap_unlock(ptep, ptl);
3717         return 0;
3718 }
3719 EXPORT_SYMBOL(follow_pfn);
3720
3721 #ifdef CONFIG_HAVE_IOREMAP_PROT
3722 int follow_phys(struct vm_area_struct *vma,
3723                 unsigned long address, unsigned int flags,
3724                 unsigned long *prot, resource_size_t *phys)
3725 {
3726         int ret = -EINVAL;
3727         pte_t *ptep, pte;
3728         spinlock_t *ptl;
3729
3730         if (!(vma->vm_flags & (VM_IO | VM_PFNMAP)))
3731                 goto out;
3732
3733         if (follow_pte(vma->vm_mm, address, &ptep, &ptl))
3734                 goto out;
3735         pte = *ptep;
3736
3737         if ((flags & FOLL_WRITE) && !pte_write(pte))
3738                 goto unlock;
3739
3740         *prot = pgprot_val(pte_pgprot(pte));
3741         *phys = (resource_size_t)pte_pfn(pte) << PAGE_SHIFT;
3742
3743         ret = 0;
3744 unlock:
3745         pte_unmap_unlock(ptep, ptl);
3746 out:
3747         return ret;
3748 }
3749
3750 int generic_access_phys(struct vm_area_struct *vma, unsigned long addr,
3751                         void *buf, int len, int write)
3752 {
3753         resource_size_t phys_addr;
3754         unsigned long prot = 0;
3755         void __iomem *maddr;
3756         int offset = addr & (PAGE_SIZE-1);
3757
3758         if (follow_phys(vma, addr, write, &prot, &phys_addr))
3759                 return -EINVAL;
3760
3761         maddr = ioremap_prot(phys_addr, PAGE_SIZE, prot);
3762         if (write)
3763                 memcpy_toio(maddr + offset, buf, len);
3764         else
3765                 memcpy_fromio(buf, maddr + offset, len);
3766         iounmap(maddr);
3767
3768         return len;
3769 }
3770 #endif
3771
3772 /*
3773  * Access another process' address space as given in mm.  If non-NULL, use the
3774  * given task for page fault accounting.
3775  */
3776 static int __access_remote_vm(struct task_struct *tsk, struct mm_struct *mm,
3777                 unsigned long addr, void *buf, int len, int write)
3778 {
3779         struct vm_area_struct *vma;
3780         void *old_buf = buf;
3781
3782         down_read(&mm->mmap_sem);
3783         /* ignore errors, just check how much was successfully transferred */
3784         while (len) {
3785                 int bytes, ret, offset;
3786                 void *maddr;
3787                 struct page *page = NULL;
3788
3789                 ret = get_user_pages(tsk, mm, addr, 1,
3790                                 write, 1, &page, &vma);
3791                 if (ret <= 0) {
3792                         /*
3793                          * Check if this is a VM_IO | VM_PFNMAP VMA, which
3794                          * we can access using slightly different code.
3795                          */
3796 #ifdef CONFIG_HAVE_IOREMAP_PROT
3797                         vma = find_vma(mm, addr);
3798                         if (!vma || vma->vm_start > addr)
3799                                 break;
3800                         if (vma->vm_ops && vma->vm_ops->access)
3801                                 ret = vma->vm_ops->access(vma, addr, buf,
3802                                                           len, write);
3803                         if (ret <= 0)
3804 #endif
3805                                 break;
3806                         bytes = ret;
3807                 } else {
3808                         bytes = len;
3809                         offset = addr & (PAGE_SIZE-1);
3810                         if (bytes > PAGE_SIZE-offset)
3811                                 bytes = PAGE_SIZE-offset;
3812
3813                         maddr = kmap(page);
3814                         if (write) {
3815                                 copy_to_user_page(vma, page, addr,
3816                                                   maddr + offset, buf, bytes);
3817                                 set_page_dirty_lock(page);
3818                         } else {
3819                                 copy_from_user_page(vma, page, addr,
3820                                                     buf, maddr + offset, bytes);
3821                         }
3822                         kunmap(page);
3823                         page_cache_release(page);
3824                 }
3825                 len -= bytes;
3826                 buf += bytes;
3827                 addr += bytes;
3828         }
3829         up_read(&mm->mmap_sem);
3830
3831         return buf - old_buf;
3832 }
3833
3834 /**
3835  * access_remote_vm - access another process' address space
3836  * @mm:         the mm_struct of the target address space
3837  * @addr:       start address to access
3838  * @buf:        source or destination buffer
3839  * @len:        number of bytes to transfer
3840  * @write:      whether the access is a write
3841  *
3842  * The caller must hold a reference on @mm.
3843  */
3844 int access_remote_vm(struct mm_struct *mm, unsigned long addr,
3845                 void *buf, int len, int write)
3846 {
3847         return __access_remote_vm(NULL, mm, addr, buf, len, write);
3848 }
3849
3850 /*
3851  * Access another process' address space.
3852  * Source/target buffer must be kernel space,
3853  * Do not walk the page table directly, use get_user_pages
3854  */
3855 int access_process_vm(struct task_struct *tsk, unsigned long addr,
3856                 void *buf, int len, int write)
3857 {
3858         struct mm_struct *mm;
3859         int ret;
3860
3861         mm = get_task_mm(tsk);
3862         if (!mm)
3863                 return 0;
3864
3865         ret = __access_remote_vm(tsk, mm, addr, buf, len, write);
3866         mmput(mm);
3867
3868         return ret;
3869 }
3870
3871 /*
3872  * Print the name of a VMA.
3873  */
3874 void print_vma_addr(char *prefix, unsigned long ip)
3875 {
3876         struct mm_struct *mm = current->mm;
3877         struct vm_area_struct *vma;
3878
3879         /*
3880          * Do not print if we are in atomic
3881          * contexts (in exception stacks, etc.):
3882          */
3883         if (preempt_count())
3884                 return;
3885
3886         down_read(&mm->mmap_sem);
3887         vma = find_vma(mm, ip);
3888         if (vma && vma->vm_file) {
3889                 struct file *f = vma->vm_file;
3890                 char *buf = (char *)__get_free_page(GFP_KERNEL);
3891                 if (buf) {
3892                         char *p, *s;
3893
3894                         p = d_path(&f->f_path, buf, PAGE_SIZE);
3895                         if (IS_ERR(p))
3896                                 p = "?";
3897                         s = strrchr(p, '/');
3898                         if (s)
3899                                 p = s+1;
3900                         printk("%s%s[%lx+%lx]", prefix, p,
3901                                         vma->vm_start,
3902                                         vma->vm_end - vma->vm_start);
3903                         free_page((unsigned long)buf);
3904                 }
3905         }
3906         up_read(&current->mm->mmap_sem);
3907 }
3908
3909 #ifdef CONFIG_PROVE_LOCKING
3910 void might_fault(void)
3911 {
3912         /*
3913          * Some code (nfs/sunrpc) uses socket ops on kernel memory while
3914          * holding the mmap_sem, this is safe because kernel memory doesn't
3915          * get paged out, therefore we'll never actually fault, and the
3916          * below annotations will generate false positives.
3917          */
3918         if (segment_eq(get_fs(), KERNEL_DS))
3919                 return;
3920
3921         might_sleep();
3922         /*
3923          * it would be nicer only to annotate paths which are not under
3924          * pagefault_disable, however that requires a larger audit and
3925          * providing helpers like get_user_atomic.
3926          */
3927         if (!in_atomic() && current->mm)
3928                 might_lock_read(&current->mm->mmap_sem);
3929 }
3930 EXPORT_SYMBOL(might_fault);
3931 #endif
3932
3933 #if defined(CONFIG_TRANSPARENT_HUGEPAGE) || defined(CONFIG_HUGETLBFS)
3934 static void clear_gigantic_page(struct page *page,
3935                                 unsigned long addr,
3936                                 unsigned int pages_per_huge_page)
3937 {
3938         int i;
3939         struct page *p = page;
3940
3941         might_sleep();
3942         for (i = 0; i < pages_per_huge_page;
3943              i++, p = mem_map_next(p, page, i)) {
3944                 cond_resched();
3945                 clear_user_highpage(p, addr + i * PAGE_SIZE);
3946         }
3947 }
3948 void clear_huge_page(struct page *page,
3949                      unsigned long addr, unsigned int pages_per_huge_page)
3950 {
3951         int i;
3952
3953         if (unlikely(pages_per_huge_page > MAX_ORDER_NR_PAGES)) {
3954                 clear_gigantic_page(page, addr, pages_per_huge_page);
3955                 return;
3956         }
3957
3958         might_sleep();
3959         for (i = 0; i < pages_per_huge_page; i++) {
3960                 cond_resched();
3961                 clear_user_highpage(page + i, addr + i * PAGE_SIZE);
3962         }
3963 }
3964
3965 static void copy_user_gigantic_page(struct page *dst, struct page *src,
3966                                     unsigned long addr,
3967                                     struct vm_area_struct *vma,
3968                                     unsigned int pages_per_huge_page)
3969 {
3970         int i;
3971         struct page *dst_base = dst;
3972         struct page *src_base = src;
3973
3974         for (i = 0; i < pages_per_huge_page; ) {
3975                 cond_resched();
3976                 copy_user_highpage(dst, src, addr + i*PAGE_SIZE, vma);
3977
3978                 i++;
3979                 dst = mem_map_next(dst, dst_base, i);
3980                 src = mem_map_next(src, src_base, i);
3981         }
3982 }
3983
3984 void copy_user_huge_page(struct page *dst, struct page *src,
3985                          unsigned long addr, struct vm_area_struct *vma,
3986                          unsigned int pages_per_huge_page)
3987 {
3988         int i;
3989
3990         if (unlikely(pages_per_huge_page > MAX_ORDER_NR_PAGES)) {
3991                 copy_user_gigantic_page(dst, src, addr, vma,
3992                                         pages_per_huge_page);
3993                 return;
3994         }
3995
3996         might_sleep();
3997         for (i = 0; i < pages_per_huge_page; i++) {
3998                 cond_resched();
3999                 copy_user_highpage(dst + i, src + i, addr + i*PAGE_SIZE, vma);
4000         }
4001 }
4002 #endif /* CONFIG_TRANSPARENT_HUGEPAGE || CONFIG_HUGETLBFS */