lguest: remove reference from Documentation/virtual/00-INDEX
[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 unsigned long 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         return addr;
1310 }
1311
1312 /**
1313  * unmap_vmas - unmap a range of memory covered by a list of vma's
1314  * @tlb: address of the caller's struct mmu_gather
1315  * @vma: the starting vma
1316  * @start_addr: virtual address at which to start unmapping
1317  * @end_addr: virtual address at which to end unmapping
1318  * @nr_accounted: Place number of unmapped pages in vm-accountable vma's here
1319  * @details: details of nonlinear truncation or shared cache invalidation
1320  *
1321  * Returns the end address of the unmapping (restart addr if interrupted).
1322  *
1323  * Unmap all pages in the vma list.
1324  *
1325  * Only addresses between `start' and `end' will be unmapped.
1326  *
1327  * The VMA list must be sorted in ascending virtual address order.
1328  *
1329  * unmap_vmas() assumes that the caller will flush the whole unmapped address
1330  * range after unmap_vmas() returns.  So the only responsibility here is to
1331  * ensure that any thus-far unmapped pages are flushed before unmap_vmas()
1332  * drops the lock and schedules.
1333  */
1334 unsigned long unmap_vmas(struct mmu_gather *tlb,
1335                 struct vm_area_struct *vma, unsigned long start_addr,
1336                 unsigned long end_addr, unsigned long *nr_accounted,
1337                 struct zap_details *details)
1338 {
1339         unsigned long start = start_addr;
1340         struct mm_struct *mm = vma->vm_mm;
1341
1342         mmu_notifier_invalidate_range_start(mm, start_addr, end_addr);
1343         for ( ; vma && vma->vm_start < end_addr; vma = vma->vm_next) {
1344                 unsigned long end;
1345
1346                 start = max(vma->vm_start, start_addr);
1347                 if (start >= vma->vm_end)
1348                         continue;
1349                 end = min(vma->vm_end, end_addr);
1350                 if (end <= vma->vm_start)
1351                         continue;
1352
1353                 if (vma->vm_flags & VM_ACCOUNT)
1354                         *nr_accounted += (end - start) >> PAGE_SHIFT;
1355
1356                 if (unlikely(is_pfn_mapping(vma)))
1357                         untrack_pfn_vma(vma, 0, 0);
1358
1359                 while (start != end) {
1360                         if (unlikely(is_vm_hugetlb_page(vma))) {
1361                                 /*
1362                                  * It is undesirable to test vma->vm_file as it
1363                                  * should be non-null for valid hugetlb area.
1364                                  * However, vm_file will be NULL in the error
1365                                  * cleanup path of do_mmap_pgoff. When
1366                                  * hugetlbfs ->mmap method fails,
1367                                  * do_mmap_pgoff() nullifies vma->vm_file
1368                                  * before calling this function to clean up.
1369                                  * Since no pte has actually been setup, it is
1370                                  * safe to do nothing in this case.
1371                                  */
1372                                 if (vma->vm_file)
1373                                         unmap_hugepage_range(vma, start, end, NULL);
1374
1375                                 start = end;
1376                         } else
1377                                 start = unmap_page_range(tlb, vma, start, end, details);
1378                 }
1379         }
1380
1381         mmu_notifier_invalidate_range_end(mm, start_addr, end_addr);
1382         return start;   /* which is now the end (or restart) address */
1383 }
1384
1385 /**
1386  * zap_page_range - remove user pages in a given range
1387  * @vma: vm_area_struct holding the applicable pages
1388  * @address: starting address of pages to zap
1389  * @size: number of bytes to zap
1390  * @details: details of nonlinear truncation or shared cache invalidation
1391  */
1392 unsigned long zap_page_range(struct vm_area_struct *vma, unsigned long address,
1393                 unsigned long size, struct zap_details *details)
1394 {
1395         struct mm_struct *mm = vma->vm_mm;
1396         struct mmu_gather tlb;
1397         unsigned long end = address + size;
1398         unsigned long nr_accounted = 0;
1399
1400         lru_add_drain();
1401         tlb_gather_mmu(&tlb, mm, 0);
1402         update_hiwater_rss(mm);
1403         end = unmap_vmas(&tlb, vma, address, end, &nr_accounted, details);
1404         tlb_finish_mmu(&tlb, address, end);
1405         return end;
1406 }
1407
1408 /**
1409  * zap_vma_ptes - remove ptes mapping the vma
1410  * @vma: vm_area_struct holding ptes to be zapped
1411  * @address: starting address of pages to zap
1412  * @size: number of bytes to zap
1413  *
1414  * This function only unmaps ptes assigned to VM_PFNMAP vmas.
1415  *
1416  * The entire address range must be fully contained within the vma.
1417  *
1418  * Returns 0 if successful.
1419  */
1420 int zap_vma_ptes(struct vm_area_struct *vma, unsigned long address,
1421                 unsigned long size)
1422 {
1423         if (address < vma->vm_start || address + size > vma->vm_end ||
1424                         !(vma->vm_flags & VM_PFNMAP))
1425                 return -1;
1426         zap_page_range(vma, address, size, NULL);
1427         return 0;
1428 }
1429 EXPORT_SYMBOL_GPL(zap_vma_ptes);
1430
1431 /**
1432  * follow_page - look up a page descriptor from a user-virtual address
1433  * @vma: vm_area_struct mapping @address
1434  * @address: virtual address to look up
1435  * @flags: flags modifying lookup behaviour
1436  *
1437  * @flags can have FOLL_ flags set, defined in <linux/mm.h>
1438  *
1439  * Returns the mapped (struct page *), %NULL if no mapping exists, or
1440  * an error pointer if there is a mapping to something not represented
1441  * by a page descriptor (see also vm_normal_page()).
1442  */
1443 struct page *follow_page(struct vm_area_struct *vma, unsigned long address,
1444                         unsigned int flags)
1445 {
1446         pgd_t *pgd;
1447         pud_t *pud;
1448         pmd_t *pmd;
1449         pte_t *ptep, pte;
1450         spinlock_t *ptl;
1451         struct page *page;
1452         struct mm_struct *mm = vma->vm_mm;
1453
1454         page = follow_huge_addr(mm, address, flags & FOLL_WRITE);
1455         if (!IS_ERR(page)) {
1456                 BUG_ON(flags & FOLL_GET);
1457                 goto out;
1458         }
1459
1460         page = NULL;
1461         pgd = pgd_offset(mm, address);
1462         if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
1463                 goto no_page_table;
1464
1465         pud = pud_offset(pgd, address);
1466         if (pud_none(*pud))
1467                 goto no_page_table;
1468         if (pud_huge(*pud) && vma->vm_flags & VM_HUGETLB) {
1469                 BUG_ON(flags & FOLL_GET);
1470                 page = follow_huge_pud(mm, address, pud, flags & FOLL_WRITE);
1471                 goto out;
1472         }
1473         if (unlikely(pud_bad(*pud)))
1474                 goto no_page_table;
1475
1476         pmd = pmd_offset(pud, address);
1477         if (pmd_none(*pmd))
1478                 goto no_page_table;
1479         if (pmd_huge(*pmd) && vma->vm_flags & VM_HUGETLB) {
1480                 BUG_ON(flags & FOLL_GET);
1481                 page = follow_huge_pmd(mm, address, pmd, flags & FOLL_WRITE);
1482                 goto out;
1483         }
1484         if (pmd_trans_huge(*pmd)) {
1485                 if (flags & FOLL_SPLIT) {
1486                         split_huge_page_pmd(mm, pmd);
1487                         goto split_fallthrough;
1488                 }
1489                 spin_lock(&mm->page_table_lock);
1490                 if (likely(pmd_trans_huge(*pmd))) {
1491                         if (unlikely(pmd_trans_splitting(*pmd))) {
1492                                 spin_unlock(&mm->page_table_lock);
1493                                 wait_split_huge_page(vma->anon_vma, pmd);
1494                         } else {
1495                                 page = follow_trans_huge_pmd(mm, address,
1496                                                              pmd, flags);
1497                                 spin_unlock(&mm->page_table_lock);
1498                                 goto out;
1499                         }
1500                 } else
1501                         spin_unlock(&mm->page_table_lock);
1502                 /* fall through */
1503         }
1504 split_fallthrough:
1505         if (unlikely(pmd_bad(*pmd)))
1506                 goto no_page_table;
1507
1508         ptep = pte_offset_map_lock(mm, pmd, address, &ptl);
1509
1510         pte = *ptep;
1511         if (!pte_present(pte))
1512                 goto no_page;
1513         if ((flags & FOLL_WRITE) && !pte_write(pte))
1514                 goto unlock;
1515
1516         page = vm_normal_page(vma, address, pte);
1517         if (unlikely(!page)) {
1518                 if ((flags & FOLL_DUMP) ||
1519                     !is_zero_pfn(pte_pfn(pte)))
1520                         goto bad_page;
1521                 page = pte_page(pte);
1522         }
1523
1524         if (flags & FOLL_GET)
1525                 get_page_foll(page);
1526         if (flags & FOLL_TOUCH) {
1527                 if ((flags & FOLL_WRITE) &&
1528                     !pte_dirty(pte) && !PageDirty(page))
1529                         set_page_dirty(page);
1530                 /*
1531                  * pte_mkyoung() would be more correct here, but atomic care
1532                  * is needed to avoid losing the dirty bit: it is easier to use
1533                  * mark_page_accessed().
1534                  */
1535                 mark_page_accessed(page);
1536         }
1537         if ((flags & FOLL_MLOCK) && (vma->vm_flags & VM_LOCKED)) {
1538                 /*
1539                  * The preliminary mapping check is mainly to avoid the
1540                  * pointless overhead of lock_page on the ZERO_PAGE
1541                  * which might bounce very badly if there is contention.
1542                  *
1543                  * If the page is already locked, we don't need to
1544                  * handle it now - vmscan will handle it later if and
1545                  * when it attempts to reclaim the page.
1546                  */
1547                 if (page->mapping && trylock_page(page)) {
1548                         lru_add_drain();  /* push cached pages to LRU */
1549                         /*
1550                          * Because we lock page here and migration is
1551                          * blocked by the pte's page reference, we need
1552                          * only check for file-cache page truncation.
1553                          */
1554                         if (page->mapping)
1555                                 mlock_vma_page(page);
1556                         unlock_page(page);
1557                 }
1558         }
1559 unlock:
1560         pte_unmap_unlock(ptep, ptl);
1561 out:
1562         return page;
1563
1564 bad_page:
1565         pte_unmap_unlock(ptep, ptl);
1566         return ERR_PTR(-EFAULT);
1567
1568 no_page:
1569         pte_unmap_unlock(ptep, ptl);
1570         if (!pte_none(pte))
1571                 return page;
1572
1573 no_page_table:
1574         /*
1575          * When core dumping an enormous anonymous area that nobody
1576          * has touched so far, we don't want to allocate unnecessary pages or
1577          * page tables.  Return error instead of NULL to skip handle_mm_fault,
1578          * then get_dump_page() will return NULL to leave a hole in the dump.
1579          * But we can only make this optimization where a hole would surely
1580          * be zero-filled if handle_mm_fault() actually did handle it.
1581          */
1582         if ((flags & FOLL_DUMP) &&
1583             (!vma->vm_ops || !vma->vm_ops->fault))
1584                 return ERR_PTR(-EFAULT);
1585         return page;
1586 }
1587
1588 static inline int stack_guard_page(struct vm_area_struct *vma, unsigned long addr)
1589 {
1590         return stack_guard_page_start(vma, addr) ||
1591                stack_guard_page_end(vma, addr+PAGE_SIZE);
1592 }
1593
1594 /**
1595  * __get_user_pages() - pin user pages in memory
1596  * @tsk:        task_struct of target task
1597  * @mm:         mm_struct of target mm
1598  * @start:      starting user address
1599  * @nr_pages:   number of pages from start to pin
1600  * @gup_flags:  flags modifying pin behaviour
1601  * @pages:      array that receives pointers to the pages pinned.
1602  *              Should be at least nr_pages long. Or NULL, if caller
1603  *              only intends to ensure the pages are faulted in.
1604  * @vmas:       array of pointers to vmas corresponding to each page.
1605  *              Or NULL if the caller does not require them.
1606  * @nonblocking: whether waiting for disk IO or mmap_sem contention
1607  *
1608  * Returns number of pages pinned. This may be fewer than the number
1609  * requested. If nr_pages is 0 or negative, returns 0. If no pages
1610  * were pinned, returns -errno. Each page returned must be released
1611  * with a put_page() call when it is finished with. vmas will only
1612  * remain valid while mmap_sem is held.
1613  *
1614  * Must be called with mmap_sem held for read or write.
1615  *
1616  * __get_user_pages walks a process's page tables and takes a reference to
1617  * each struct page that each user address corresponds to at a given
1618  * instant. That is, it takes the page that would be accessed if a user
1619  * thread accesses the given user virtual address at that instant.
1620  *
1621  * This does not guarantee that the page exists in the user mappings when
1622  * __get_user_pages returns, and there may even be a completely different
1623  * page there in some cases (eg. if mmapped pagecache has been invalidated
1624  * and subsequently re faulted). However it does guarantee that the page
1625  * won't be freed completely. And mostly callers simply care that the page
1626  * contains data that was valid *at some point in time*. Typically, an IO
1627  * or similar operation cannot guarantee anything stronger anyway because
1628  * locks can't be held over the syscall boundary.
1629  *
1630  * If @gup_flags & FOLL_WRITE == 0, the page must not be written to. If
1631  * the page is written to, set_page_dirty (or set_page_dirty_lock, as
1632  * appropriate) must be called after the page is finished with, and
1633  * before put_page is called.
1634  *
1635  * If @nonblocking != NULL, __get_user_pages will not wait for disk IO
1636  * or mmap_sem contention, and if waiting is needed to pin all pages,
1637  * *@nonblocking will be set to 0.
1638  *
1639  * In most cases, get_user_pages or get_user_pages_fast should be used
1640  * instead of __get_user_pages. __get_user_pages should be used only if
1641  * you need some special @gup_flags.
1642  */
1643 int __get_user_pages(struct task_struct *tsk, struct mm_struct *mm,
1644                      unsigned long start, int nr_pages, unsigned int gup_flags,
1645                      struct page **pages, struct vm_area_struct **vmas,
1646                      int *nonblocking)
1647 {
1648         int i;
1649         unsigned long vm_flags;
1650
1651         if (nr_pages <= 0)
1652                 return 0;
1653
1654         VM_BUG_ON(!!pages != !!(gup_flags & FOLL_GET));
1655
1656         /* 
1657          * Require read or write permissions.
1658          * If FOLL_FORCE is set, we only require the "MAY" flags.
1659          */
1660         vm_flags  = (gup_flags & FOLL_WRITE) ?
1661                         (VM_WRITE | VM_MAYWRITE) : (VM_READ | VM_MAYREAD);
1662         vm_flags &= (gup_flags & FOLL_FORCE) ?
1663                         (VM_MAYREAD | VM_MAYWRITE) : (VM_READ | VM_WRITE);
1664         i = 0;
1665
1666         do {
1667                 struct vm_area_struct *vma;
1668
1669                 vma = find_extend_vma(mm, start);
1670                 if (!vma && in_gate_area(mm, start)) {
1671                         unsigned long pg = start & PAGE_MASK;
1672                         pgd_t *pgd;
1673                         pud_t *pud;
1674                         pmd_t *pmd;
1675                         pte_t *pte;
1676
1677                         /* user gate pages are read-only */
1678                         if (gup_flags & FOLL_WRITE)
1679                                 return i ? : -EFAULT;
1680                         if (pg > TASK_SIZE)
1681                                 pgd = pgd_offset_k(pg);
1682                         else
1683                                 pgd = pgd_offset_gate(mm, pg);
1684                         BUG_ON(pgd_none(*pgd));
1685                         pud = pud_offset(pgd, pg);
1686                         BUG_ON(pud_none(*pud));
1687                         pmd = pmd_offset(pud, pg);
1688                         if (pmd_none(*pmd))
1689                                 return i ? : -EFAULT;
1690                         VM_BUG_ON(pmd_trans_huge(*pmd));
1691                         pte = pte_offset_map(pmd, pg);
1692                         if (pte_none(*pte)) {
1693                                 pte_unmap(pte);
1694                                 return i ? : -EFAULT;
1695                         }
1696                         vma = get_gate_vma(mm);
1697                         if (pages) {
1698                                 struct page *page;
1699
1700                                 page = vm_normal_page(vma, start, *pte);
1701                                 if (!page) {
1702                                         if (!(gup_flags & FOLL_DUMP) &&
1703                                              is_zero_pfn(pte_pfn(*pte)))
1704                                                 page = pte_page(*pte);
1705                                         else {
1706                                                 pte_unmap(pte);
1707                                                 return i ? : -EFAULT;
1708                                         }
1709                                 }
1710                                 pages[i] = page;
1711                                 get_page(page);
1712                         }
1713                         pte_unmap(pte);
1714                         goto next_page;
1715                 }
1716
1717                 if (!vma ||
1718                     (vma->vm_flags & (VM_IO | VM_PFNMAP)) ||
1719                     !(vm_flags & vma->vm_flags))
1720                         return i ? : -EFAULT;
1721
1722                 if (is_vm_hugetlb_page(vma)) {
1723                         i = follow_hugetlb_page(mm, vma, pages, vmas,
1724                                         &start, &nr_pages, i, gup_flags);
1725                         continue;
1726                 }
1727
1728                 do {
1729                         struct page *page;
1730                         unsigned int foll_flags = gup_flags;
1731
1732                         /*
1733                          * If we have a pending SIGKILL, don't keep faulting
1734                          * pages and potentially allocating memory.
1735                          */
1736                         if (unlikely(fatal_signal_pending(current)))
1737                                 return i ? i : -ERESTARTSYS;
1738
1739                         cond_resched();
1740                         while (!(page = follow_page(vma, start, foll_flags))) {
1741                                 int ret;
1742                                 unsigned int fault_flags = 0;
1743
1744                                 /* For mlock, just skip the stack guard page. */
1745                                 if (foll_flags & FOLL_MLOCK) {
1746                                         if (stack_guard_page(vma, start))
1747                                                 goto next_page;
1748                                 }
1749                                 if (foll_flags & FOLL_WRITE)
1750                                         fault_flags |= FAULT_FLAG_WRITE;
1751                                 if (nonblocking)
1752                                         fault_flags |= FAULT_FLAG_ALLOW_RETRY;
1753                                 if (foll_flags & FOLL_NOWAIT)
1754                                         fault_flags |= (FAULT_FLAG_ALLOW_RETRY | FAULT_FLAG_RETRY_NOWAIT);
1755
1756                                 ret = handle_mm_fault(mm, vma, start,
1757                                                         fault_flags);
1758
1759                                 if (ret & VM_FAULT_ERROR) {
1760                                         if (ret & VM_FAULT_OOM)
1761                                                 return i ? i : -ENOMEM;
1762                                         if (ret & (VM_FAULT_HWPOISON |
1763                                                    VM_FAULT_HWPOISON_LARGE)) {
1764                                                 if (i)
1765                                                         return i;
1766                                                 else if (gup_flags & FOLL_HWPOISON)
1767                                                         return -EHWPOISON;
1768                                                 else
1769                                                         return -EFAULT;
1770                                         }
1771                                         if (ret & VM_FAULT_SIGBUS)
1772                                                 return i ? i : -EFAULT;
1773                                         BUG();
1774                                 }
1775
1776                                 if (tsk) {
1777                                         if (ret & VM_FAULT_MAJOR)
1778                                                 tsk->maj_flt++;
1779                                         else
1780                                                 tsk->min_flt++;
1781                                 }
1782
1783                                 if (ret & VM_FAULT_RETRY) {
1784                                         if (nonblocking)
1785                                                 *nonblocking = 0;
1786                                         return i;
1787                                 }
1788
1789                                 /*
1790                                  * The VM_FAULT_WRITE bit tells us that
1791                                  * do_wp_page has broken COW when necessary,
1792                                  * even if maybe_mkwrite decided not to set
1793                                  * pte_write. We can thus safely do subsequent
1794                                  * page lookups as if they were reads. But only
1795                                  * do so when looping for pte_write is futile:
1796                                  * in some cases userspace may also be wanting
1797                                  * to write to the gotten user page, which a
1798                                  * read fault here might prevent (a readonly
1799                                  * page might get reCOWed by userspace write).
1800                                  */
1801                                 if ((ret & VM_FAULT_WRITE) &&
1802                                     !(vma->vm_flags & VM_WRITE))
1803                                         foll_flags &= ~FOLL_WRITE;
1804
1805                                 cond_resched();
1806                         }
1807                         if (IS_ERR(page))
1808                                 return i ? i : PTR_ERR(page);
1809                         if (pages) {
1810                                 pages[i] = page;
1811
1812                                 flush_anon_page(vma, page, start);
1813                                 flush_dcache_page(page);
1814                         }
1815 next_page:
1816                         if (vmas)
1817                                 vmas[i] = vma;
1818                         i++;
1819                         start += PAGE_SIZE;
1820                         nr_pages--;
1821                 } while (nr_pages && start < vma->vm_end);
1822         } while (nr_pages);
1823         return i;
1824 }
1825 EXPORT_SYMBOL(__get_user_pages);
1826
1827 /*
1828  * fixup_user_fault() - manually resolve a user page fault
1829  * @tsk:        the task_struct to use for page fault accounting, or
1830  *              NULL if faults are not to be recorded.
1831  * @mm:         mm_struct of target mm
1832  * @address:    user address
1833  * @fault_flags:flags to pass down to handle_mm_fault()
1834  *
1835  * This is meant to be called in the specific scenario where for locking reasons
1836  * we try to access user memory in atomic context (within a pagefault_disable()
1837  * section), this returns -EFAULT, and we want to resolve the user fault before
1838  * trying again.
1839  *
1840  * Typically this is meant to be used by the futex code.
1841  *
1842  * The main difference with get_user_pages() is that this function will
1843  * unconditionally call handle_mm_fault() which will in turn perform all the
1844  * necessary SW fixup of the dirty and young bits in the PTE, while
1845  * handle_mm_fault() only guarantees to update these in the struct page.
1846  *
1847  * This is important for some architectures where those bits also gate the
1848  * access permission to the page because they are maintained in software.  On
1849  * such architectures, gup() will not be enough to make a subsequent access
1850  * succeed.
1851  *
1852  * This should be called with the mm_sem held for read.
1853  */
1854 int fixup_user_fault(struct task_struct *tsk, struct mm_struct *mm,
1855                      unsigned long address, unsigned int fault_flags)
1856 {
1857         struct vm_area_struct *vma;
1858         int ret;
1859
1860         vma = find_extend_vma(mm, address);
1861         if (!vma || address < vma->vm_start)
1862                 return -EFAULT;
1863
1864         ret = handle_mm_fault(mm, vma, address, fault_flags);
1865         if (ret & VM_FAULT_ERROR) {
1866                 if (ret & VM_FAULT_OOM)
1867                         return -ENOMEM;
1868                 if (ret & (VM_FAULT_HWPOISON | VM_FAULT_HWPOISON_LARGE))
1869                         return -EHWPOISON;
1870                 if (ret & VM_FAULT_SIGBUS)
1871                         return -EFAULT;
1872                 BUG();
1873         }
1874         if (tsk) {
1875                 if (ret & VM_FAULT_MAJOR)
1876                         tsk->maj_flt++;
1877                 else
1878                         tsk->min_flt++;
1879         }
1880         return 0;
1881 }
1882
1883 /*
1884  * get_user_pages() - pin user pages in memory
1885  * @tsk:        the task_struct to use for page fault accounting, or
1886  *              NULL if faults are not to be recorded.
1887  * @mm:         mm_struct of target mm
1888  * @start:      starting user address
1889  * @nr_pages:   number of pages from start to pin
1890  * @write:      whether pages will be written to by the caller
1891  * @force:      whether to force write access even if user mapping is
1892  *              readonly. This will result in the page being COWed even
1893  *              in MAP_SHARED mappings. You do not want this.
1894  * @pages:      array that receives pointers to the pages pinned.
1895  *              Should be at least nr_pages long. Or NULL, if caller
1896  *              only intends to ensure the pages are faulted in.
1897  * @vmas:       array of pointers to vmas corresponding to each page.
1898  *              Or NULL if the caller does not require them.
1899  *
1900  * Returns number of pages pinned. This may be fewer than the number
1901  * requested. If nr_pages is 0 or negative, returns 0. If no pages
1902  * were pinned, returns -errno. Each page returned must be released
1903  * with a put_page() call when it is finished with. vmas will only
1904  * remain valid while mmap_sem is held.
1905  *
1906  * Must be called with mmap_sem held for read or write.
1907  *
1908  * get_user_pages walks a process's page tables and takes a reference to
1909  * each struct page that each user address corresponds to at a given
1910  * instant. That is, it takes the page that would be accessed if a user
1911  * thread accesses the given user virtual address at that instant.
1912  *
1913  * This does not guarantee that the page exists in the user mappings when
1914  * get_user_pages returns, and there may even be a completely different
1915  * page there in some cases (eg. if mmapped pagecache has been invalidated
1916  * and subsequently re faulted). However it does guarantee that the page
1917  * won't be freed completely. And mostly callers simply care that the page
1918  * contains data that was valid *at some point in time*. Typically, an IO
1919  * or similar operation cannot guarantee anything stronger anyway because
1920  * locks can't be held over the syscall boundary.
1921  *
1922  * If write=0, the page must not be written to. If the page is written to,
1923  * set_page_dirty (or set_page_dirty_lock, as appropriate) must be called
1924  * after the page is finished with, and before put_page is called.
1925  *
1926  * get_user_pages is typically used for fewer-copy IO operations, to get a
1927  * handle on the memory by some means other than accesses via the user virtual
1928  * addresses. The pages may be submitted for DMA to devices or accessed via
1929  * their kernel linear mapping (via the kmap APIs). Care should be taken to
1930  * use the correct cache flushing APIs.
1931  *
1932  * See also get_user_pages_fast, for performance critical applications.
1933  */
1934 int get_user_pages(struct task_struct *tsk, struct mm_struct *mm,
1935                 unsigned long start, int nr_pages, int write, int force,
1936                 struct page **pages, struct vm_area_struct **vmas)
1937 {
1938         int flags = FOLL_TOUCH;
1939
1940         if (pages)
1941                 flags |= FOLL_GET;
1942         if (write)
1943                 flags |= FOLL_WRITE;
1944         if (force)
1945                 flags |= FOLL_FORCE;
1946
1947         return __get_user_pages(tsk, mm, start, nr_pages, flags, pages, vmas,
1948                                 NULL);
1949 }
1950 EXPORT_SYMBOL(get_user_pages);
1951
1952 /**
1953  * get_dump_page() - pin user page in memory while writing it to core dump
1954  * @addr: user address
1955  *
1956  * Returns struct page pointer of user page pinned for dump,
1957  * to be freed afterwards by page_cache_release() or put_page().
1958  *
1959  * Returns NULL on any kind of failure - a hole must then be inserted into
1960  * the corefile, to preserve alignment with its headers; and also returns
1961  * NULL wherever the ZERO_PAGE, or an anonymous pte_none, has been found -
1962  * allowing a hole to be left in the corefile to save diskspace.
1963  *
1964  * Called without mmap_sem, but after all other threads have been killed.
1965  */
1966 #ifdef CONFIG_ELF_CORE
1967 struct page *get_dump_page(unsigned long addr)
1968 {
1969         struct vm_area_struct *vma;
1970         struct page *page;
1971
1972         if (__get_user_pages(current, current->mm, addr, 1,
1973                              FOLL_FORCE | FOLL_DUMP | FOLL_GET, &page, &vma,
1974                              NULL) < 1)
1975                 return NULL;
1976         flush_cache_page(vma, addr, page_to_pfn(page));
1977         return page;
1978 }
1979 #endif /* CONFIG_ELF_CORE */
1980
1981 pte_t *__get_locked_pte(struct mm_struct *mm, unsigned long addr,
1982                         spinlock_t **ptl)
1983 {
1984         pgd_t * pgd = pgd_offset(mm, addr);
1985         pud_t * pud = pud_alloc(mm, pgd, addr);
1986         if (pud) {
1987                 pmd_t * pmd = pmd_alloc(mm, pud, addr);
1988                 if (pmd) {
1989                         VM_BUG_ON(pmd_trans_huge(*pmd));
1990                         return pte_alloc_map_lock(mm, pmd, addr, ptl);
1991                 }
1992         }
1993         return NULL;
1994 }
1995
1996 /*
1997  * This is the old fallback for page remapping.
1998  *
1999  * For historical reasons, it only allows reserved pages. Only
2000  * old drivers should use this, and they needed to mark their
2001  * pages reserved for the old functions anyway.
2002  */
2003 static int insert_page(struct vm_area_struct *vma, unsigned long addr,
2004                         struct page *page, pgprot_t prot)
2005 {
2006         struct mm_struct *mm = vma->vm_mm;
2007         int retval;
2008         pte_t *pte;
2009         spinlock_t *ptl;
2010
2011         retval = -EINVAL;
2012         if (PageAnon(page))
2013                 goto out;
2014         retval = -ENOMEM;
2015         flush_dcache_page(page);
2016         pte = get_locked_pte(mm, addr, &ptl);
2017         if (!pte)
2018                 goto out;
2019         retval = -EBUSY;
2020         if (!pte_none(*pte))
2021                 goto out_unlock;
2022
2023         /* Ok, finally just insert the thing.. */
2024         get_page(page);
2025         inc_mm_counter_fast(mm, MM_FILEPAGES);
2026         page_add_file_rmap(page);
2027         set_pte_at(mm, addr, pte, mk_pte(page, prot));
2028
2029         retval = 0;
2030         pte_unmap_unlock(pte, ptl);
2031         return retval;
2032 out_unlock:
2033         pte_unmap_unlock(pte, ptl);
2034 out:
2035         return retval;
2036 }
2037
2038 /**
2039  * vm_insert_page - insert single page into user vma
2040  * @vma: user vma to map to
2041  * @addr: target user address of this page
2042  * @page: source kernel page
2043  *
2044  * This allows drivers to insert individual pages they've allocated
2045  * into a user vma.
2046  *
2047  * The page has to be a nice clean _individual_ kernel allocation.
2048  * If you allocate a compound page, you need to have marked it as
2049  * such (__GFP_COMP), or manually just split the page up yourself
2050  * (see split_page()).
2051  *
2052  * NOTE! Traditionally this was done with "remap_pfn_range()" which
2053  * took an arbitrary page protection parameter. This doesn't allow
2054  * that. Your vma protection will have to be set up correctly, which
2055  * means that if you want a shared writable mapping, you'd better
2056  * ask for a shared writable mapping!
2057  *
2058  * The page does not need to be reserved.
2059  */
2060 int vm_insert_page(struct vm_area_struct *vma, unsigned long addr,
2061                         struct page *page)
2062 {
2063         if (addr < vma->vm_start || addr >= vma->vm_end)
2064                 return -EFAULT;
2065         if (!page_count(page))
2066                 return -EINVAL;
2067         vma->vm_flags |= VM_INSERTPAGE;
2068         return insert_page(vma, addr, page, vma->vm_page_prot);
2069 }
2070 EXPORT_SYMBOL(vm_insert_page);
2071
2072 static int insert_pfn(struct vm_area_struct *vma, unsigned long addr,
2073                         unsigned long pfn, pgprot_t prot)
2074 {
2075         struct mm_struct *mm = vma->vm_mm;
2076         int retval;
2077         pte_t *pte, entry;
2078         spinlock_t *ptl;
2079
2080         retval = -ENOMEM;
2081         pte = get_locked_pte(mm, addr, &ptl);
2082         if (!pte)
2083                 goto out;
2084         retval = -EBUSY;
2085         if (!pte_none(*pte))
2086                 goto out_unlock;
2087
2088         /* Ok, finally just insert the thing.. */
2089         entry = pte_mkspecial(pfn_pte(pfn, prot));
2090         set_pte_at(mm, addr, pte, entry);
2091         update_mmu_cache(vma, addr, pte); /* XXX: why not for insert_page? */
2092
2093         retval = 0;
2094 out_unlock:
2095         pte_unmap_unlock(pte, ptl);
2096 out:
2097         return retval;
2098 }
2099
2100 /**
2101  * vm_insert_pfn - insert single pfn into user vma
2102  * @vma: user vma to map to
2103  * @addr: target user address of this page
2104  * @pfn: source kernel pfn
2105  *
2106  * Similar to vm_inert_page, this allows drivers to insert individual pages
2107  * they've allocated into a user vma. Same comments apply.
2108  *
2109  * This function should only be called from a vm_ops->fault handler, and
2110  * in that case the handler should return NULL.
2111  *
2112  * vma cannot be a COW mapping.
2113  *
2114  * As this is called only for pages that do not currently exist, we
2115  * do not need to flush old virtual caches or the TLB.
2116  */
2117 int vm_insert_pfn(struct vm_area_struct *vma, unsigned long addr,
2118                         unsigned long pfn)
2119 {
2120         int ret;
2121         pgprot_t pgprot = vma->vm_page_prot;
2122         /*
2123          * Technically, architectures with pte_special can avoid all these
2124          * restrictions (same for remap_pfn_range).  However we would like
2125          * consistency in testing and feature parity among all, so we should
2126          * try to keep these invariants in place for everybody.
2127          */
2128         BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)));
2129         BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) ==
2130                                                 (VM_PFNMAP|VM_MIXEDMAP));
2131         BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags));
2132         BUG_ON((vma->vm_flags & VM_MIXEDMAP) && pfn_valid(pfn));
2133
2134         if (addr < vma->vm_start || addr >= vma->vm_end)
2135                 return -EFAULT;
2136         if (track_pfn_vma_new(vma, &pgprot, pfn, PAGE_SIZE))
2137                 return -EINVAL;
2138
2139         ret = insert_pfn(vma, addr, pfn, pgprot);
2140
2141         if (ret)
2142                 untrack_pfn_vma(vma, pfn, PAGE_SIZE);
2143
2144         return ret;
2145 }
2146 EXPORT_SYMBOL(vm_insert_pfn);
2147
2148 int vm_insert_mixed(struct vm_area_struct *vma, unsigned long addr,
2149                         unsigned long pfn)
2150 {
2151         BUG_ON(!(vma->vm_flags & VM_MIXEDMAP));
2152
2153         if (addr < vma->vm_start || addr >= vma->vm_end)
2154                 return -EFAULT;
2155
2156         /*
2157          * If we don't have pte special, then we have to use the pfn_valid()
2158          * based VM_MIXEDMAP scheme (see vm_normal_page), and thus we *must*
2159          * refcount the page if pfn_valid is true (hence insert_page rather
2160          * than insert_pfn).  If a zero_pfn were inserted into a VM_MIXEDMAP
2161          * without pte special, it would there be refcounted as a normal page.
2162          */
2163         if (!HAVE_PTE_SPECIAL && pfn_valid(pfn)) {
2164                 struct page *page;
2165
2166                 page = pfn_to_page(pfn);
2167                 return insert_page(vma, addr, page, vma->vm_page_prot);
2168         }
2169         return insert_pfn(vma, addr, pfn, vma->vm_page_prot);
2170 }
2171 EXPORT_SYMBOL(vm_insert_mixed);
2172
2173 /*
2174  * maps a range of physical memory into the requested pages. the old
2175  * mappings are removed. any references to nonexistent pages results
2176  * in null mappings (currently treated as "copy-on-access")
2177  */
2178 static int remap_pte_range(struct mm_struct *mm, pmd_t *pmd,
2179                         unsigned long addr, unsigned long end,
2180                         unsigned long pfn, pgprot_t prot)
2181 {
2182         pte_t *pte;
2183         spinlock_t *ptl;
2184
2185         pte = pte_alloc_map_lock(mm, pmd, addr, &ptl);
2186         if (!pte)
2187                 return -ENOMEM;
2188         arch_enter_lazy_mmu_mode();
2189         do {
2190                 BUG_ON(!pte_none(*pte));
2191                 set_pte_at(mm, addr, pte, pte_mkspecial(pfn_pte(pfn, prot)));
2192                 pfn++;
2193         } while (pte++, addr += PAGE_SIZE, addr != end);
2194         arch_leave_lazy_mmu_mode();
2195         pte_unmap_unlock(pte - 1, ptl);
2196         return 0;
2197 }
2198
2199 static inline int remap_pmd_range(struct mm_struct *mm, pud_t *pud,
2200                         unsigned long addr, unsigned long end,
2201                         unsigned long pfn, pgprot_t prot)
2202 {
2203         pmd_t *pmd;
2204         unsigned long next;
2205
2206         pfn -= addr >> PAGE_SHIFT;
2207         pmd = pmd_alloc(mm, pud, addr);
2208         if (!pmd)
2209                 return -ENOMEM;
2210         VM_BUG_ON(pmd_trans_huge(*pmd));
2211         do {
2212                 next = pmd_addr_end(addr, end);
2213                 if (remap_pte_range(mm, pmd, addr, next,
2214                                 pfn + (addr >> PAGE_SHIFT), prot))
2215                         return -ENOMEM;
2216         } while (pmd++, addr = next, addr != end);
2217         return 0;
2218 }
2219
2220 static inline int remap_pud_range(struct mm_struct *mm, pgd_t *pgd,
2221                         unsigned long addr, unsigned long end,
2222                         unsigned long pfn, pgprot_t prot)
2223 {
2224         pud_t *pud;
2225         unsigned long next;
2226
2227         pfn -= addr >> PAGE_SHIFT;
2228         pud = pud_alloc(mm, pgd, addr);
2229         if (!pud)
2230                 return -ENOMEM;
2231         do {
2232                 next = pud_addr_end(addr, end);
2233                 if (remap_pmd_range(mm, pud, addr, next,
2234                                 pfn + (addr >> PAGE_SHIFT), prot))
2235                         return -ENOMEM;
2236         } while (pud++, addr = next, addr != end);
2237         return 0;
2238 }
2239
2240 /**
2241  * remap_pfn_range - remap kernel memory to userspace
2242  * @vma: user vma to map to
2243  * @addr: target user address to start at
2244  * @pfn: physical address of kernel memory
2245  * @size: size of map area
2246  * @prot: page protection flags for this mapping
2247  *
2248  *  Note: this is only safe if the mm semaphore is held when called.
2249  */
2250 int remap_pfn_range(struct vm_area_struct *vma, unsigned long addr,
2251                     unsigned long pfn, unsigned long size, pgprot_t prot)
2252 {
2253         pgd_t *pgd;
2254         unsigned long next;
2255         unsigned long end = addr + PAGE_ALIGN(size);
2256         struct mm_struct *mm = vma->vm_mm;
2257         int err;
2258
2259         /*
2260          * Physically remapped pages are special. Tell the
2261          * rest of the world about it:
2262          *   VM_IO tells people not to look at these pages
2263          *      (accesses can have side effects).
2264          *   VM_RESERVED is specified all over the place, because
2265          *      in 2.4 it kept swapout's vma scan off this vma; but
2266          *      in 2.6 the LRU scan won't even find its pages, so this
2267          *      flag means no more than count its pages in reserved_vm,
2268          *      and omit it from core dump, even when VM_IO turned off.
2269          *   VM_PFNMAP tells the core MM that the base pages are just
2270          *      raw PFN mappings, and do not have a "struct page" associated
2271          *      with them.
2272          *
2273          * There's a horrible special case to handle copy-on-write
2274          * behaviour that some programs depend on. We mark the "original"
2275          * un-COW'ed pages by matching them up with "vma->vm_pgoff".
2276          */
2277         if (addr == vma->vm_start && end == vma->vm_end) {
2278                 vma->vm_pgoff = pfn;
2279                 vma->vm_flags |= VM_PFN_AT_MMAP;
2280         } else if (is_cow_mapping(vma->vm_flags))
2281                 return -EINVAL;
2282
2283         vma->vm_flags |= VM_IO | VM_RESERVED | VM_PFNMAP;
2284
2285         err = track_pfn_vma_new(vma, &prot, pfn, PAGE_ALIGN(size));
2286         if (err) {
2287                 /*
2288                  * To indicate that track_pfn related cleanup is not
2289                  * needed from higher level routine calling unmap_vmas
2290                  */
2291                 vma->vm_flags &= ~(VM_IO | VM_RESERVED | VM_PFNMAP);
2292                 vma->vm_flags &= ~VM_PFN_AT_MMAP;
2293                 return -EINVAL;
2294         }
2295
2296         BUG_ON(addr >= end);
2297         pfn -= addr >> PAGE_SHIFT;
2298         pgd = pgd_offset(mm, addr);
2299         flush_cache_range(vma, addr, end);
2300         do {
2301                 next = pgd_addr_end(addr, end);
2302                 err = remap_pud_range(mm, pgd, addr, next,
2303                                 pfn + (addr >> PAGE_SHIFT), prot);
2304                 if (err)
2305                         break;
2306         } while (pgd++, addr = next, addr != end);
2307
2308         if (err)
2309                 untrack_pfn_vma(vma, pfn, PAGE_ALIGN(size));
2310
2311         return err;
2312 }
2313 EXPORT_SYMBOL(remap_pfn_range);
2314
2315 static int apply_to_pte_range(struct mm_struct *mm, pmd_t *pmd,
2316                                      unsigned long addr, unsigned long end,
2317                                      pte_fn_t fn, void *data)
2318 {
2319         pte_t *pte;
2320         int err;
2321         pgtable_t token;
2322         spinlock_t *uninitialized_var(ptl);
2323
2324         pte = (mm == &init_mm) ?
2325                 pte_alloc_kernel(pmd, addr) :
2326                 pte_alloc_map_lock(mm, pmd, addr, &ptl);
2327         if (!pte)
2328                 return -ENOMEM;
2329
2330         BUG_ON(pmd_huge(*pmd));
2331
2332         arch_enter_lazy_mmu_mode();
2333
2334         token = pmd_pgtable(*pmd);
2335
2336         do {
2337                 err = fn(pte++, token, addr, data);
2338                 if (err)
2339                         break;
2340         } while (addr += PAGE_SIZE, addr != end);
2341
2342         arch_leave_lazy_mmu_mode();
2343
2344         if (mm != &init_mm)
2345                 pte_unmap_unlock(pte-1, ptl);
2346         return err;
2347 }
2348
2349 static int apply_to_pmd_range(struct mm_struct *mm, pud_t *pud,
2350                                      unsigned long addr, unsigned long end,
2351                                      pte_fn_t fn, void *data)
2352 {
2353         pmd_t *pmd;
2354         unsigned long next;
2355         int err;
2356
2357         BUG_ON(pud_huge(*pud));
2358
2359         pmd = pmd_alloc(mm, pud, addr);
2360         if (!pmd)
2361                 return -ENOMEM;
2362         do {
2363                 next = pmd_addr_end(addr, end);
2364                 err = apply_to_pte_range(mm, pmd, addr, next, fn, data);
2365                 if (err)
2366                         break;
2367         } while (pmd++, addr = next, addr != end);
2368         return err;
2369 }
2370
2371 static int apply_to_pud_range(struct mm_struct *mm, pgd_t *pgd,
2372                                      unsigned long addr, unsigned long end,
2373                                      pte_fn_t fn, void *data)
2374 {
2375         pud_t *pud;
2376         unsigned long next;
2377         int err;
2378
2379         pud = pud_alloc(mm, pgd, addr);
2380         if (!pud)
2381                 return -ENOMEM;
2382         do {
2383                 next = pud_addr_end(addr, end);
2384                 err = apply_to_pmd_range(mm, pud, addr, next, fn, data);
2385                 if (err)
2386                         break;
2387         } while (pud++, addr = next, addr != end);
2388         return err;
2389 }
2390
2391 /*
2392  * Scan a region of virtual memory, filling in page tables as necessary
2393  * and calling a provided function on each leaf page table.
2394  */
2395 int apply_to_page_range(struct mm_struct *mm, unsigned long addr,
2396                         unsigned long size, pte_fn_t fn, void *data)
2397 {
2398         pgd_t *pgd;
2399         unsigned long next;
2400         unsigned long end = addr + size;
2401         int err;
2402
2403         BUG_ON(addr >= end);
2404         pgd = pgd_offset(mm, addr);
2405         do {
2406                 next = pgd_addr_end(addr, end);
2407                 err = apply_to_pud_range(mm, pgd, addr, next, fn, data);
2408                 if (err)
2409                         break;
2410         } while (pgd++, addr = next, addr != end);
2411
2412         return err;
2413 }
2414 EXPORT_SYMBOL_GPL(apply_to_page_range);
2415
2416 /*
2417  * handle_pte_fault chooses page fault handler according to an entry
2418  * which was read non-atomically.  Before making any commitment, on
2419  * those architectures or configurations (e.g. i386 with PAE) which
2420  * might give a mix of unmatched parts, do_swap_page and do_nonlinear_fault
2421  * must check under lock before unmapping the pte and proceeding
2422  * (but do_wp_page is only called after already making such a check;
2423  * and do_anonymous_page can safely check later on).
2424  */
2425 static inline int pte_unmap_same(struct mm_struct *mm, pmd_t *pmd,
2426                                 pte_t *page_table, pte_t orig_pte)
2427 {
2428         int same = 1;
2429 #if defined(CONFIG_SMP) || defined(CONFIG_PREEMPT)
2430         if (sizeof(pte_t) > sizeof(unsigned long)) {
2431                 spinlock_t *ptl = pte_lockptr(mm, pmd);
2432                 spin_lock(ptl);
2433                 same = pte_same(*page_table, orig_pte);
2434                 spin_unlock(ptl);
2435         }
2436 #endif
2437         pte_unmap(page_table);
2438         return same;
2439 }
2440
2441 static inline void cow_user_page(struct page *dst, struct page *src, unsigned long va, struct vm_area_struct *vma)
2442 {
2443         /*
2444          * If the source page was a PFN mapping, we don't have
2445          * a "struct page" for it. We do a best-effort copy by
2446          * just copying from the original user address. If that
2447          * fails, we just zero-fill it. Live with it.
2448          */
2449         if (unlikely(!src)) {
2450                 void *kaddr = kmap_atomic(dst, KM_USER0);
2451                 void __user *uaddr = (void __user *)(va & PAGE_MASK);
2452
2453                 /*
2454                  * This really shouldn't fail, because the page is there
2455                  * in the page tables. But it might just be unreadable,
2456                  * in which case we just give up and fill the result with
2457                  * zeroes.
2458                  */
2459                 if (__copy_from_user_inatomic(kaddr, uaddr, PAGE_SIZE))
2460                         clear_page(kaddr);
2461                 kunmap_atomic(kaddr, KM_USER0);
2462                 flush_dcache_page(dst);
2463         } else
2464                 copy_user_highpage(dst, src, va, vma);
2465 }
2466
2467 /*
2468  * This routine handles present pages, when users try to write
2469  * to a shared page. It is done by copying the page to a new address
2470  * and decrementing the shared-page counter for the old page.
2471  *
2472  * Note that this routine assumes that the protection checks have been
2473  * done by the caller (the low-level page fault routine in most cases).
2474  * Thus we can safely just mark it writable once we've done any necessary
2475  * COW.
2476  *
2477  * We also mark the page dirty at this point even though the page will
2478  * change only once the write actually happens. This avoids a few races,
2479  * and potentially makes it more efficient.
2480  *
2481  * We enter with non-exclusive mmap_sem (to exclude vma changes,
2482  * but allow concurrent faults), with pte both mapped and locked.
2483  * We return with mmap_sem still held, but pte unmapped and unlocked.
2484  */
2485 static int do_wp_page(struct mm_struct *mm, struct vm_area_struct *vma,
2486                 unsigned long address, pte_t *page_table, pmd_t *pmd,
2487                 spinlock_t *ptl, pte_t orig_pte)
2488         __releases(ptl)
2489 {
2490         struct page *old_page, *new_page;
2491         pte_t entry;
2492         int ret = 0;
2493         int page_mkwrite = 0;
2494         struct page *dirty_page = NULL;
2495
2496         old_page = vm_normal_page(vma, address, orig_pte);
2497         if (!old_page) {
2498                 /*
2499                  * VM_MIXEDMAP !pfn_valid() case
2500                  *
2501                  * We should not cow pages in a shared writeable mapping.
2502                  * Just mark the pages writable as we can't do any dirty
2503                  * accounting on raw pfn maps.
2504                  */
2505                 if ((vma->vm_flags & (VM_WRITE|VM_SHARED)) ==
2506                                      (VM_WRITE|VM_SHARED))
2507                         goto reuse;
2508                 goto gotten;
2509         }
2510
2511         /*
2512          * Take out anonymous pages first, anonymous shared vmas are
2513          * not dirty accountable.
2514          */
2515         if (PageAnon(old_page) && !PageKsm(old_page)) {
2516                 if (!trylock_page(old_page)) {
2517                         page_cache_get(old_page);
2518                         pte_unmap_unlock(page_table, ptl);
2519                         lock_page(old_page);
2520                         page_table = pte_offset_map_lock(mm, pmd, address,
2521                                                          &ptl);
2522                         if (!pte_same(*page_table, orig_pte)) {
2523                                 unlock_page(old_page);
2524                                 goto unlock;
2525                         }
2526                         page_cache_release(old_page);
2527                 }
2528                 if (reuse_swap_page(old_page)) {
2529                         /*
2530                          * The page is all ours.  Move it to our anon_vma so
2531                          * the rmap code will not search our parent or siblings.
2532                          * Protected against the rmap code by the page lock.
2533                          */
2534                         page_move_anon_rmap(old_page, vma, address);
2535                         unlock_page(old_page);
2536                         goto reuse;
2537                 }
2538                 unlock_page(old_page);
2539         } else if (unlikely((vma->vm_flags & (VM_WRITE|VM_SHARED)) ==
2540                                         (VM_WRITE|VM_SHARED))) {
2541                 /*
2542                  * Only catch write-faults on shared writable pages,
2543                  * read-only shared pages can get COWed by
2544                  * get_user_pages(.write=1, .force=1).
2545                  */
2546                 if (vma->vm_ops && vma->vm_ops->page_mkwrite) {
2547                         struct vm_fault vmf;
2548                         int tmp;
2549
2550                         vmf.virtual_address = (void __user *)(address &
2551                                                                 PAGE_MASK);
2552                         vmf.pgoff = old_page->index;
2553                         vmf.flags = FAULT_FLAG_WRITE|FAULT_FLAG_MKWRITE;
2554                         vmf.page = old_page;
2555
2556                         /*
2557                          * Notify the address space that the page is about to
2558                          * become writable so that it can prohibit this or wait
2559                          * for the page to get into an appropriate state.
2560                          *
2561                          * We do this without the lock held, so that it can
2562                          * sleep if it needs to.
2563                          */
2564                         page_cache_get(old_page);
2565                         pte_unmap_unlock(page_table, ptl);
2566
2567                         tmp = vma->vm_ops->page_mkwrite(vma, &vmf);
2568                         if (unlikely(tmp &
2569                                         (VM_FAULT_ERROR | VM_FAULT_NOPAGE))) {
2570                                 ret = tmp;
2571                                 goto unwritable_page;
2572                         }
2573                         if (unlikely(!(tmp & VM_FAULT_LOCKED))) {
2574                                 lock_page(old_page);
2575                                 if (!old_page->mapping) {
2576                                         ret = 0; /* retry the fault */
2577                                         unlock_page(old_page);
2578                                         goto unwritable_page;
2579                                 }
2580                         } else
2581                                 VM_BUG_ON(!PageLocked(old_page));
2582
2583                         /*
2584                          * Since we dropped the lock we need to revalidate
2585                          * the PTE as someone else may have changed it.  If
2586                          * they did, we just return, as we can count on the
2587                          * MMU to tell us if they didn't also make it writable.
2588                          */
2589                         page_table = pte_offset_map_lock(mm, pmd, address,
2590                                                          &ptl);
2591                         if (!pte_same(*page_table, orig_pte)) {
2592                                 unlock_page(old_page);
2593                                 goto unlock;
2594                         }
2595
2596                         page_mkwrite = 1;
2597                 }
2598                 dirty_page = old_page;
2599                 get_page(dirty_page);
2600
2601 reuse:
2602                 flush_cache_page(vma, address, pte_pfn(orig_pte));
2603                 entry = pte_mkyoung(orig_pte);
2604                 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2605                 if (ptep_set_access_flags(vma, address, page_table, entry,1))
2606                         update_mmu_cache(vma, address, page_table);
2607                 pte_unmap_unlock(page_table, ptl);
2608                 ret |= VM_FAULT_WRITE;
2609
2610                 if (!dirty_page)
2611                         return ret;
2612
2613                 /*
2614                  * Yes, Virginia, this is actually required to prevent a race
2615                  * with clear_page_dirty_for_io() from clearing the page dirty
2616                  * bit after it clear all dirty ptes, but before a racing
2617                  * do_wp_page installs a dirty pte.
2618                  *
2619                  * __do_fault is protected similarly.
2620                  */
2621                 if (!page_mkwrite) {
2622                         wait_on_page_locked(dirty_page);
2623                         set_page_dirty_balance(dirty_page, page_mkwrite);
2624                 }
2625                 put_page(dirty_page);
2626                 if (page_mkwrite) {
2627                         struct address_space *mapping = dirty_page->mapping;
2628
2629                         set_page_dirty(dirty_page);
2630                         unlock_page(dirty_page);
2631                         page_cache_release(dirty_page);
2632                         if (mapping)    {
2633                                 /*
2634                                  * Some device drivers do not set page.mapping
2635                                  * but still dirty their pages
2636                                  */
2637                                 balance_dirty_pages_ratelimited(mapping);
2638                         }
2639                 }
2640
2641                 /* file_update_time outside page_lock */
2642                 if (vma->vm_file)
2643                         file_update_time(vma->vm_file);
2644
2645                 return ret;
2646         }
2647
2648         /*
2649          * Ok, we need to copy. Oh, well..
2650          */
2651         page_cache_get(old_page);
2652 gotten:
2653         pte_unmap_unlock(page_table, ptl);
2654
2655         if (unlikely(anon_vma_prepare(vma)))
2656                 goto oom;
2657
2658         if (is_zero_pfn(pte_pfn(orig_pte))) {
2659                 new_page = alloc_zeroed_user_highpage_movable(vma, address);
2660                 if (!new_page)
2661                         goto oom;
2662         } else {
2663                 new_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, address);
2664                 if (!new_page)
2665                         goto oom;
2666                 cow_user_page(new_page, old_page, address, vma);
2667         }
2668         __SetPageUptodate(new_page);
2669
2670         if (mem_cgroup_newpage_charge(new_page, mm, GFP_KERNEL))
2671                 goto oom_free_new;
2672
2673         /*
2674          * Re-check the pte - we dropped the lock
2675          */
2676         page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2677         if (likely(pte_same(*page_table, orig_pte))) {
2678                 if (old_page) {
2679                         if (!PageAnon(old_page)) {
2680                                 dec_mm_counter_fast(mm, MM_FILEPAGES);
2681                                 inc_mm_counter_fast(mm, MM_ANONPAGES);
2682                         }
2683                 } else
2684                         inc_mm_counter_fast(mm, MM_ANONPAGES);
2685                 flush_cache_page(vma, address, pte_pfn(orig_pte));
2686                 entry = mk_pte(new_page, vma->vm_page_prot);
2687                 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2688                 /*
2689                  * Clear the pte entry and flush it first, before updating the
2690                  * pte with the new entry. This will avoid a race condition
2691                  * seen in the presence of one thread doing SMC and another
2692                  * thread doing COW.
2693                  */
2694                 ptep_clear_flush(vma, address, page_table);
2695                 page_add_new_anon_rmap(new_page, vma, address);
2696                 /*
2697                  * We call the notify macro here because, when using secondary
2698                  * mmu page tables (such as kvm shadow page tables), we want the
2699                  * new page to be mapped directly into the secondary page table.
2700                  */
2701                 set_pte_at_notify(mm, address, page_table, entry);
2702                 update_mmu_cache(vma, address, page_table);
2703                 if (old_page) {
2704                         /*
2705                          * Only after switching the pte to the new page may
2706                          * we remove the mapcount here. Otherwise another
2707                          * process may come and find the rmap count decremented
2708                          * before the pte is switched to the new page, and
2709                          * "reuse" the old page writing into it while our pte
2710                          * here still points into it and can be read by other
2711                          * threads.
2712                          *
2713                          * The critical issue is to order this
2714                          * page_remove_rmap with the ptp_clear_flush above.
2715                          * Those stores are ordered by (if nothing else,)
2716                          * the barrier present in the atomic_add_negative
2717                          * in page_remove_rmap.
2718                          *
2719                          * Then the TLB flush in ptep_clear_flush ensures that
2720                          * no process can access the old page before the
2721                          * decremented mapcount is visible. And the old page
2722                          * cannot be reused until after the decremented
2723                          * mapcount is visible. So transitively, TLBs to
2724                          * old page will be flushed before it can be reused.
2725                          */
2726                         page_remove_rmap(old_page);
2727                 }
2728
2729                 /* Free the old page.. */
2730                 new_page = old_page;
2731                 ret |= VM_FAULT_WRITE;
2732         } else
2733                 mem_cgroup_uncharge_page(new_page);
2734
2735         if (new_page)
2736                 page_cache_release(new_page);
2737 unlock:
2738         pte_unmap_unlock(page_table, ptl);
2739         if (old_page) {
2740                 /*
2741                  * Don't let another task, with possibly unlocked vma,
2742                  * keep the mlocked page.
2743                  */
2744                 if ((ret & VM_FAULT_WRITE) && (vma->vm_flags & VM_LOCKED)) {
2745                         lock_page(old_page);    /* LRU manipulation */
2746                         munlock_vma_page(old_page);
2747                         unlock_page(old_page);
2748                 }
2749                 page_cache_release(old_page);
2750         }
2751         return ret;
2752 oom_free_new:
2753         page_cache_release(new_page);
2754 oom:
2755         if (old_page) {
2756                 if (page_mkwrite) {
2757                         unlock_page(old_page);
2758                         page_cache_release(old_page);
2759                 }
2760                 page_cache_release(old_page);
2761         }
2762         return VM_FAULT_OOM;
2763
2764 unwritable_page:
2765         page_cache_release(old_page);
2766         return ret;
2767 }
2768
2769 static void unmap_mapping_range_vma(struct vm_area_struct *vma,
2770                 unsigned long start_addr, unsigned long end_addr,
2771                 struct zap_details *details)
2772 {
2773         zap_page_range(vma, start_addr, end_addr - start_addr, details);
2774 }
2775
2776 static inline void unmap_mapping_range_tree(struct prio_tree_root *root,
2777                                             struct zap_details *details)
2778 {
2779         struct vm_area_struct *vma;
2780         struct prio_tree_iter iter;
2781         pgoff_t vba, vea, zba, zea;
2782
2783         vma_prio_tree_foreach(vma, &iter, root,
2784                         details->first_index, details->last_index) {
2785
2786                 vba = vma->vm_pgoff;
2787                 vea = vba + ((vma->vm_end - vma->vm_start) >> PAGE_SHIFT) - 1;
2788                 /* Assume for now that PAGE_CACHE_SHIFT == PAGE_SHIFT */
2789                 zba = details->first_index;
2790                 if (zba < vba)
2791                         zba = vba;
2792                 zea = details->last_index;
2793                 if (zea > vea)
2794                         zea = vea;
2795
2796                 unmap_mapping_range_vma(vma,
2797                         ((zba - vba) << PAGE_SHIFT) + vma->vm_start,
2798                         ((zea - vba + 1) << PAGE_SHIFT) + vma->vm_start,
2799                                 details);
2800         }
2801 }
2802
2803 static inline void unmap_mapping_range_list(struct list_head *head,
2804                                             struct zap_details *details)
2805 {
2806         struct vm_area_struct *vma;
2807
2808         /*
2809          * In nonlinear VMAs there is no correspondence between virtual address
2810          * offset and file offset.  So we must perform an exhaustive search
2811          * across *all* the pages in each nonlinear VMA, not just the pages
2812          * whose virtual address lies outside the file truncation point.
2813          */
2814         list_for_each_entry(vma, head, shared.vm_set.list) {
2815                 details->nonlinear_vma = vma;
2816                 unmap_mapping_range_vma(vma, vma->vm_start, vma->vm_end, details);
2817         }
2818 }
2819
2820 /**
2821  * unmap_mapping_range - unmap the portion of all mmaps in the specified address_space corresponding to the specified page range in the underlying file.
2822  * @mapping: the address space containing mmaps to be unmapped.
2823  * @holebegin: byte in first page to unmap, relative to the start of
2824  * the underlying file.  This will be rounded down to a PAGE_SIZE
2825  * boundary.  Note that this is different from truncate_pagecache(), which
2826  * must keep the partial page.  In contrast, we must get rid of
2827  * partial pages.
2828  * @holelen: size of prospective hole in bytes.  This will be rounded
2829  * up to a PAGE_SIZE boundary.  A holelen of zero truncates to the
2830  * end of the file.
2831  * @even_cows: 1 when truncating a file, unmap even private COWed pages;
2832  * but 0 when invalidating pagecache, don't throw away private data.
2833  */
2834 void unmap_mapping_range(struct address_space *mapping,
2835                 loff_t const holebegin, loff_t const holelen, int even_cows)
2836 {
2837         struct zap_details details;
2838         pgoff_t hba = holebegin >> PAGE_SHIFT;
2839         pgoff_t hlen = (holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
2840
2841         /* Check for overflow. */
2842         if (sizeof(holelen) > sizeof(hlen)) {
2843                 long long holeend =
2844                         (holebegin + holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
2845                 if (holeend & ~(long long)ULONG_MAX)
2846                         hlen = ULONG_MAX - hba + 1;
2847         }
2848
2849         details.check_mapping = even_cows? NULL: mapping;
2850         details.nonlinear_vma = NULL;
2851         details.first_index = hba;
2852         details.last_index = hba + hlen - 1;
2853         if (details.last_index < details.first_index)
2854                 details.last_index = ULONG_MAX;
2855
2856
2857         mutex_lock(&mapping->i_mmap_mutex);
2858         if (unlikely(!prio_tree_empty(&mapping->i_mmap)))
2859                 unmap_mapping_range_tree(&mapping->i_mmap, &details);
2860         if (unlikely(!list_empty(&mapping->i_mmap_nonlinear)))
2861                 unmap_mapping_range_list(&mapping->i_mmap_nonlinear, &details);
2862         mutex_unlock(&mapping->i_mmap_mutex);
2863 }
2864 EXPORT_SYMBOL(unmap_mapping_range);
2865
2866 /*
2867  * We enter with non-exclusive mmap_sem (to exclude vma changes,
2868  * but allow concurrent faults), and pte mapped but not yet locked.
2869  * We return with mmap_sem still held, but pte unmapped and unlocked.
2870  */
2871 static int do_swap_page(struct mm_struct *mm, struct vm_area_struct *vma,
2872                 unsigned long address, pte_t *page_table, pmd_t *pmd,
2873                 unsigned int flags, pte_t orig_pte)
2874 {
2875         spinlock_t *ptl;
2876         struct page *page, *swapcache = NULL;
2877         swp_entry_t entry;
2878         pte_t pte;
2879         int locked;
2880         struct mem_cgroup *ptr;
2881         int exclusive = 0;
2882         int ret = 0;
2883
2884         if (!pte_unmap_same(mm, pmd, page_table, orig_pte))
2885                 goto out;
2886
2887         entry = pte_to_swp_entry(orig_pte);
2888         if (unlikely(non_swap_entry(entry))) {
2889                 if (is_migration_entry(entry)) {
2890                         migration_entry_wait(mm, pmd, address);
2891                 } else if (is_hwpoison_entry(entry)) {
2892                         ret = VM_FAULT_HWPOISON;
2893                 } else {
2894                         print_bad_pte(vma, address, orig_pte, NULL);
2895                         ret = VM_FAULT_SIGBUS;
2896                 }
2897                 goto out;
2898         }
2899         delayacct_set_flag(DELAYACCT_PF_SWAPIN);
2900         page = lookup_swap_cache(entry);
2901         if (!page) {
2902                 grab_swap_token(mm); /* Contend for token _before_ read-in */
2903                 page = swapin_readahead(entry,
2904                                         GFP_HIGHUSER_MOVABLE, vma, address);
2905                 if (!page) {
2906                         /*
2907                          * Back out if somebody else faulted in this pte
2908                          * while we released the pte lock.
2909                          */
2910                         page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2911                         if (likely(pte_same(*page_table, orig_pte)))
2912                                 ret = VM_FAULT_OOM;
2913                         delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
2914                         goto unlock;
2915                 }
2916
2917                 /* Had to read the page from swap area: Major fault */
2918                 ret = VM_FAULT_MAJOR;
2919                 count_vm_event(PGMAJFAULT);
2920                 mem_cgroup_count_vm_event(mm, PGMAJFAULT);
2921         } else if (PageHWPoison(page)) {
2922                 /*
2923                  * hwpoisoned dirty swapcache pages are kept for killing
2924                  * owner processes (which may be unknown at hwpoison time)
2925                  */
2926                 ret = VM_FAULT_HWPOISON;
2927                 delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
2928                 goto out_release;
2929         }
2930
2931         locked = lock_page_or_retry(page, mm, flags);
2932         delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
2933         if (!locked) {
2934                 ret |= VM_FAULT_RETRY;
2935                 goto out_release;
2936         }
2937
2938         /*
2939          * Make sure try_to_free_swap or reuse_swap_page or swapoff did not
2940          * release the swapcache from under us.  The page pin, and pte_same
2941          * test below, are not enough to exclude that.  Even if it is still
2942          * swapcache, we need to check that the page's swap has not changed.
2943          */
2944         if (unlikely(!PageSwapCache(page) || page_private(page) != entry.val))
2945                 goto out_page;
2946
2947         if (ksm_might_need_to_copy(page, vma, address)) {
2948                 swapcache = page;
2949                 page = ksm_does_need_to_copy(page, vma, address);
2950
2951                 if (unlikely(!page)) {
2952                         ret = VM_FAULT_OOM;
2953                         page = swapcache;
2954                         swapcache = NULL;
2955                         goto out_page;
2956                 }
2957         }
2958
2959         if (mem_cgroup_try_charge_swapin(mm, page, GFP_KERNEL, &ptr)) {
2960                 ret = VM_FAULT_OOM;
2961                 goto out_page;
2962         }
2963
2964         /*
2965          * Back out if somebody else already faulted in this pte.
2966          */
2967         page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2968         if (unlikely(!pte_same(*page_table, orig_pte)))
2969                 goto out_nomap;
2970
2971         if (unlikely(!PageUptodate(page))) {
2972                 ret = VM_FAULT_SIGBUS;
2973                 goto out_nomap;
2974         }
2975
2976         /*
2977          * The page isn't present yet, go ahead with the fault.
2978          *
2979          * Be careful about the sequence of operations here.
2980          * To get its accounting right, reuse_swap_page() must be called
2981          * while the page is counted on swap but not yet in mapcount i.e.
2982          * before page_add_anon_rmap() and swap_free(); try_to_free_swap()
2983          * must be called after the swap_free(), or it will never succeed.
2984          * Because delete_from_swap_page() may be called by reuse_swap_page(),
2985          * mem_cgroup_commit_charge_swapin() may not be able to find swp_entry
2986          * in page->private. In this case, a record in swap_cgroup  is silently
2987          * discarded at swap_free().
2988          */
2989
2990         inc_mm_counter_fast(mm, MM_ANONPAGES);
2991         dec_mm_counter_fast(mm, MM_SWAPENTS);
2992         pte = mk_pte(page, vma->vm_page_prot);
2993         if ((flags & FAULT_FLAG_WRITE) && reuse_swap_page(page)) {
2994                 pte = maybe_mkwrite(pte_mkdirty(pte), vma);
2995                 flags &= ~FAULT_FLAG_WRITE;
2996                 ret |= VM_FAULT_WRITE;
2997                 exclusive = 1;
2998         }
2999         flush_icache_page(vma, page);
3000         set_pte_at(mm, address, page_table, pte);
3001         do_page_add_anon_rmap(page, vma, address, exclusive);
3002         /* It's better to call commit-charge after rmap is established */
3003         mem_cgroup_commit_charge_swapin(page, ptr);
3004
3005         swap_free(entry);
3006         if (vm_swap_full() || (vma->vm_flags & VM_LOCKED) || PageMlocked(page))
3007                 try_to_free_swap(page);
3008         unlock_page(page);
3009         if (swapcache) {
3010                 /*
3011                  * Hold the lock to avoid the swap entry to be reused
3012                  * until we take the PT lock for the pte_same() check
3013                  * (to avoid false positives from pte_same). For
3014                  * further safety release the lock after the swap_free
3015                  * so that the swap count won't change under a
3016                  * parallel locked swapcache.
3017                  */
3018                 unlock_page(swapcache);
3019                 page_cache_release(swapcache);
3020         }
3021
3022         if (flags & FAULT_FLAG_WRITE) {
3023                 ret |= do_wp_page(mm, vma, address, page_table, pmd, ptl, pte);
3024                 if (ret & VM_FAULT_ERROR)
3025                         ret &= VM_FAULT_ERROR;
3026                 goto out;
3027         }
3028
3029         /* No need to invalidate - it was non-present before */
3030         update_mmu_cache(vma, address, page_table);
3031 unlock:
3032         pte_unmap_unlock(page_table, ptl);
3033 out:
3034         return ret;
3035 out_nomap:
3036         mem_cgroup_cancel_charge_swapin(ptr);
3037         pte_unmap_unlock(page_table, ptl);
3038 out_page:
3039         unlock_page(page);
3040 out_release:
3041         page_cache_release(page);
3042         if (swapcache) {
3043                 unlock_page(swapcache);
3044                 page_cache_release(swapcache);
3045         }
3046         return ret;
3047 }
3048
3049 /*
3050  * This is like a special single-page "expand_{down|up}wards()",
3051  * except we must first make sure that 'address{-|+}PAGE_SIZE'
3052  * doesn't hit another vma.
3053  */
3054 static inline int check_stack_guard_page(struct vm_area_struct *vma, unsigned long address)
3055 {
3056         address &= PAGE_MASK;
3057         if ((vma->vm_flags & VM_GROWSDOWN) && address == vma->vm_start) {
3058                 struct vm_area_struct *prev = vma->vm_prev;
3059
3060                 /*
3061                  * Is there a mapping abutting this one below?
3062                  *
3063                  * That's only ok if it's the same stack mapping
3064                  * that has gotten split..
3065                  */
3066                 if (prev && prev->vm_end == address)
3067                         return prev->vm_flags & VM_GROWSDOWN ? 0 : -ENOMEM;
3068
3069                 expand_downwards(vma, address - PAGE_SIZE);
3070         }
3071         if ((vma->vm_flags & VM_GROWSUP) && address + PAGE_SIZE == vma->vm_end) {
3072                 struct vm_area_struct *next = vma->vm_next;
3073
3074                 /* As VM_GROWSDOWN but s/below/above/ */
3075                 if (next && next->vm_start == address + PAGE_SIZE)
3076                         return next->vm_flags & VM_GROWSUP ? 0 : -ENOMEM;
3077
3078                 expand_upwards(vma, address + PAGE_SIZE);
3079         }
3080         return 0;
3081 }
3082
3083 /*
3084  * We enter with non-exclusive mmap_sem (to exclude vma changes,
3085  * but allow concurrent faults), and pte mapped but not yet locked.
3086  * We return with mmap_sem still held, but pte unmapped and unlocked.
3087  */
3088 static int do_anonymous_page(struct mm_struct *mm, struct vm_area_struct *vma,
3089                 unsigned long address, pte_t *page_table, pmd_t *pmd,
3090                 unsigned int flags)
3091 {
3092         struct page *page;
3093         spinlock_t *ptl;
3094         pte_t entry;
3095
3096         pte_unmap(page_table);
3097
3098         /* Check if we need to add a guard page to the stack */
3099         if (check_stack_guard_page(vma, address) < 0)
3100                 return VM_FAULT_SIGBUS;
3101
3102         /* Use the zero-page for reads */
3103         if (!(flags & FAULT_FLAG_WRITE)) {
3104                 entry = pte_mkspecial(pfn_pte(my_zero_pfn(address),
3105                                                 vma->vm_page_prot));
3106                 page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
3107                 if (!pte_none(*page_table))
3108                         goto unlock;
3109                 goto setpte;
3110         }
3111
3112         /* Allocate our own private page. */
3113         if (unlikely(anon_vma_prepare(vma)))
3114                 goto oom;
3115         page = alloc_zeroed_user_highpage_movable(vma, address);
3116         if (!page)
3117                 goto oom;
3118         __SetPageUptodate(page);
3119
3120         if (mem_cgroup_newpage_charge(page, mm, GFP_KERNEL))
3121                 goto oom_free_page;
3122
3123         entry = mk_pte(page, vma->vm_page_prot);
3124         if (vma->vm_flags & VM_WRITE)
3125                 entry = pte_mkwrite(pte_mkdirty(entry));
3126
3127         page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
3128         if (!pte_none(*page_table))
3129                 goto release;
3130
3131         inc_mm_counter_fast(mm, MM_ANONPAGES);
3132         page_add_new_anon_rmap(page, vma, address);
3133 setpte:
3134         set_pte_at(mm, address, page_table, entry);
3135
3136         /* No need to invalidate - it was non-present before */
3137         update_mmu_cache(vma, address, page_table);
3138 unlock:
3139         pte_unmap_unlock(page_table, ptl);
3140         return 0;
3141 release:
3142         mem_cgroup_uncharge_page(page);
3143         page_cache_release(page);
3144         goto unlock;
3145 oom_free_page:
3146         page_cache_release(page);
3147 oom:
3148         return VM_FAULT_OOM;
3149 }
3150
3151 /*
3152  * __do_fault() tries to create a new page mapping. It aggressively
3153  * tries to share with existing pages, but makes a separate copy if
3154  * the FAULT_FLAG_WRITE is set in the flags parameter in order to avoid
3155  * the next page fault.
3156  *
3157  * As this is called only for pages that do not currently exist, we
3158  * do not need to flush old virtual caches or the TLB.
3159  *
3160  * We enter with non-exclusive mmap_sem (to exclude vma changes,
3161  * but allow concurrent faults), and pte neither mapped nor locked.
3162  * We return with mmap_sem still held, but pte unmapped and unlocked.
3163  */
3164 static int __do_fault(struct mm_struct *mm, struct vm_area_struct *vma,
3165                 unsigned long address, pmd_t *pmd,
3166                 pgoff_t pgoff, unsigned int flags, pte_t orig_pte)
3167 {
3168         pte_t *page_table;
3169         spinlock_t *ptl;
3170         struct page *page;
3171         struct page *cow_page;
3172         pte_t entry;
3173         int anon = 0;
3174         struct page *dirty_page = NULL;
3175         struct vm_fault vmf;
3176         int ret;
3177         int page_mkwrite = 0;
3178
3179         /*
3180          * If we do COW later, allocate page befor taking lock_page()
3181          * on the file cache page. This will reduce lock holding time.
3182          */
3183         if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) {
3184
3185                 if (unlikely(anon_vma_prepare(vma)))
3186                         return VM_FAULT_OOM;
3187
3188                 cow_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, address);
3189                 if (!cow_page)
3190                         return VM_FAULT_OOM;
3191
3192                 if (mem_cgroup_newpage_charge(cow_page, mm, GFP_KERNEL)) {
3193                         page_cache_release(cow_page);
3194                         return VM_FAULT_OOM;
3195                 }
3196         } else
3197                 cow_page = NULL;
3198
3199         vmf.virtual_address = (void __user *)(address & PAGE_MASK);
3200         vmf.pgoff = pgoff;
3201         vmf.flags = flags;
3202         vmf.page = NULL;
3203
3204         ret = vma->vm_ops->fault(vma, &vmf);
3205         if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE |
3206                             VM_FAULT_RETRY)))
3207                 goto uncharge_out;
3208
3209         if (unlikely(PageHWPoison(vmf.page))) {
3210                 if (ret & VM_FAULT_LOCKED)
3211                         unlock_page(vmf.page);
3212                 ret = VM_FAULT_HWPOISON;
3213                 goto uncharge_out;
3214         }
3215
3216         /*
3217          * For consistency in subsequent calls, make the faulted page always
3218          * locked.
3219          */
3220         if (unlikely(!(ret & VM_FAULT_LOCKED)))
3221                 lock_page(vmf.page);
3222         else
3223                 VM_BUG_ON(!PageLocked(vmf.page));
3224
3225         /*
3226          * Should we do an early C-O-W break?
3227          */
3228         page = vmf.page;
3229         if (flags & FAULT_FLAG_WRITE) {
3230                 if (!(vma->vm_flags & VM_SHARED)) {
3231                         page = cow_page;
3232                         anon = 1;
3233                         copy_user_highpage(page, vmf.page, address, vma);
3234                         __SetPageUptodate(page);
3235                 } else {
3236                         /*
3237                          * If the page will be shareable, see if the backing
3238                          * address space wants to know that the page is about
3239                          * to become writable
3240                          */
3241                         if (vma->vm_ops->page_mkwrite) {
3242                                 int tmp;
3243
3244                                 unlock_page(page);
3245                                 vmf.flags = FAULT_FLAG_WRITE|FAULT_FLAG_MKWRITE;
3246                                 tmp = vma->vm_ops->page_mkwrite(vma, &vmf);
3247                                 if (unlikely(tmp &
3248                                           (VM_FAULT_ERROR | VM_FAULT_NOPAGE))) {
3249                                         ret = tmp;
3250                                         goto unwritable_page;
3251                                 }
3252                                 if (unlikely(!(tmp & VM_FAULT_LOCKED))) {
3253                                         lock_page(page);
3254                                         if (!page->mapping) {
3255                                                 ret = 0; /* retry the fault */
3256                                                 unlock_page(page);
3257                                                 goto unwritable_page;
3258                                         }
3259                                 } else
3260                                         VM_BUG_ON(!PageLocked(page));
3261                                 page_mkwrite = 1;
3262                         }
3263                 }
3264
3265         }
3266
3267         page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
3268
3269         /*
3270          * This silly early PAGE_DIRTY setting removes a race
3271          * due to the bad i386 page protection. But it's valid
3272          * for other architectures too.
3273          *
3274          * Note that if FAULT_FLAG_WRITE is set, we either now have
3275          * an exclusive copy of the page, or this is a shared mapping,
3276          * so we can make it writable and dirty to avoid having to
3277          * handle that later.
3278          */
3279         /* Only go through if we didn't race with anybody else... */
3280         if (likely(pte_same(*page_table, orig_pte))) {
3281                 flush_icache_page(vma, page);
3282                 entry = mk_pte(page, vma->vm_page_prot);
3283                 if (flags & FAULT_FLAG_WRITE)
3284                         entry = maybe_mkwrite(pte_mkdirty(entry), vma);
3285                 if (anon) {
3286                         inc_mm_counter_fast(mm, MM_ANONPAGES);
3287                         page_add_new_anon_rmap(page, vma, address);
3288                 } else {
3289                         inc_mm_counter_fast(mm, MM_FILEPAGES);
3290                         page_add_file_rmap(page);
3291                         if (flags & FAULT_FLAG_WRITE) {
3292                                 dirty_page = page;
3293                                 get_page(dirty_page);
3294                         }
3295                 }
3296                 set_pte_at(mm, address, page_table, entry);
3297
3298                 /* no need to invalidate: a not-present page won't be cached */
3299                 update_mmu_cache(vma, address, page_table);
3300         } else {
3301                 if (cow_page)
3302                         mem_cgroup_uncharge_page(cow_page);
3303                 if (anon)
3304                         page_cache_release(page);
3305                 else
3306                         anon = 1; /* no anon but release faulted_page */
3307         }
3308
3309         pte_unmap_unlock(page_table, ptl);
3310
3311         if (dirty_page) {
3312                 struct address_space *mapping = page->mapping;
3313
3314                 if (set_page_dirty(dirty_page))
3315                         page_mkwrite = 1;
3316                 unlock_page(dirty_page);
3317                 put_page(dirty_page);
3318                 if (page_mkwrite && mapping) {
3319                         /*
3320                          * Some device drivers do not set page.mapping but still
3321                          * dirty their pages
3322                          */
3323                         balance_dirty_pages_ratelimited(mapping);
3324                 }
3325
3326                 /* file_update_time outside page_lock */
3327                 if (vma->vm_file)
3328                         file_update_time(vma->vm_file);
3329         } else {
3330                 unlock_page(vmf.page);
3331                 if (anon)
3332                         page_cache_release(vmf.page);
3333         }
3334
3335         return ret;
3336
3337 unwritable_page:
3338         page_cache_release(page);
3339         return ret;
3340 uncharge_out:
3341         /* fs's fault handler get error */
3342         if (cow_page) {
3343                 mem_cgroup_uncharge_page(cow_page);
3344                 page_cache_release(cow_page);
3345         }
3346         return ret;
3347 }
3348
3349 static int do_linear_fault(struct mm_struct *mm, struct vm_area_struct *vma,
3350                 unsigned long address, pte_t *page_table, pmd_t *pmd,
3351                 unsigned int flags, pte_t orig_pte)
3352 {
3353         pgoff_t pgoff = (((address & PAGE_MASK)
3354                         - vma->vm_start) >> PAGE_SHIFT) + vma->vm_pgoff;
3355
3356         pte_unmap(page_table);
3357         return __do_fault(mm, vma, address, pmd, pgoff, flags, orig_pte);
3358 }
3359
3360 /*
3361  * Fault of a previously existing named mapping. Repopulate the pte
3362  * from the encoded file_pte if possible. This enables swappable
3363  * nonlinear vmas.
3364  *
3365  * We enter with non-exclusive mmap_sem (to exclude vma changes,
3366  * but allow concurrent faults), and pte mapped but not yet locked.
3367  * We return with mmap_sem still held, but pte unmapped and unlocked.
3368  */
3369 static int do_nonlinear_fault(struct mm_struct *mm, struct vm_area_struct *vma,
3370                 unsigned long address, pte_t *page_table, pmd_t *pmd,
3371                 unsigned int flags, pte_t orig_pte)
3372 {
3373         pgoff_t pgoff;
3374
3375         flags |= FAULT_FLAG_NONLINEAR;
3376
3377         if (!pte_unmap_same(mm, pmd, page_table, orig_pte))
3378                 return 0;
3379
3380         if (unlikely(!(vma->vm_flags & VM_NONLINEAR))) {
3381                 /*
3382                  * Page table corrupted: show pte and kill process.
3383                  */
3384                 print_bad_pte(vma, address, orig_pte, NULL);
3385                 return VM_FAULT_SIGBUS;
3386         }
3387
3388         pgoff = pte_to_pgoff(orig_pte);
3389         return __do_fault(mm, vma, address, pmd, pgoff, flags, orig_pte);
3390 }
3391
3392 /*
3393  * These routines also need to handle stuff like marking pages dirty
3394  * and/or accessed for architectures that don't do it in hardware (most
3395  * RISC architectures).  The early dirtying is also good on the i386.
3396  *
3397  * There is also a hook called "update_mmu_cache()" that architectures
3398  * with external mmu caches can use to update those (ie the Sparc or
3399  * PowerPC hashed page tables that act as extended TLBs).
3400  *
3401  * We enter with non-exclusive mmap_sem (to exclude vma changes,
3402  * but allow concurrent faults), and pte mapped but not yet locked.
3403  * We return with mmap_sem still held, but pte unmapped and unlocked.
3404  */
3405 int handle_pte_fault(struct mm_struct *mm,
3406                      struct vm_area_struct *vma, unsigned long address,
3407                      pte_t *pte, pmd_t *pmd, unsigned int flags)
3408 {
3409         pte_t entry;
3410         spinlock_t *ptl;
3411
3412         entry = *pte;
3413         if (!pte_present(entry)) {
3414                 if (pte_none(entry)) {
3415                         if (vma->vm_ops) {
3416                                 if (likely(vma->vm_ops->fault))
3417                                         return do_linear_fault(mm, vma, address,
3418                                                 pte, pmd, flags, entry);
3419                         }
3420                         return do_anonymous_page(mm, vma, address,
3421                                                  pte, pmd, flags);
3422                 }
3423                 if (pte_file(entry))
3424                         return do_nonlinear_fault(mm, vma, address,
3425                                         pte, pmd, flags, entry);
3426                 return do_swap_page(mm, vma, address,
3427                                         pte, pmd, flags, entry);
3428         }
3429
3430         ptl = pte_lockptr(mm, pmd);
3431         spin_lock(ptl);
3432         if (unlikely(!pte_same(*pte, entry)))
3433                 goto unlock;
3434         if (flags & FAULT_FLAG_WRITE) {
3435                 if (!pte_write(entry))
3436                         return do_wp_page(mm, vma, address,
3437                                         pte, pmd, ptl, entry);
3438                 entry = pte_mkdirty(entry);
3439         }
3440         entry = pte_mkyoung(entry);
3441         if (ptep_set_access_flags(vma, address, pte, entry, flags & FAULT_FLAG_WRITE)) {
3442                 update_mmu_cache(vma, address, pte);
3443         } else {
3444                 /*
3445                  * This is needed only for protection faults but the arch code
3446                  * is not yet telling us if this is a protection fault or not.
3447                  * This still avoids useless tlb flushes for .text page faults
3448                  * with threads.
3449                  */
3450                 if (flags & FAULT_FLAG_WRITE)
3451                         flush_tlb_fix_spurious_fault(vma, address);
3452         }
3453 unlock:
3454         pte_unmap_unlock(pte, ptl);
3455         return 0;
3456 }
3457
3458 /*
3459  * By the time we get here, we already hold the mm semaphore
3460  */
3461 int handle_mm_fault(struct mm_struct *mm, struct vm_area_struct *vma,
3462                 unsigned long address, unsigned int flags)
3463 {
3464         pgd_t *pgd;
3465         pud_t *pud;
3466         pmd_t *pmd;
3467         pte_t *pte;
3468
3469         __set_current_state(TASK_RUNNING);
3470
3471         count_vm_event(PGFAULT);
3472         mem_cgroup_count_vm_event(mm, PGFAULT);
3473
3474         /* do counter updates before entering really critical section. */
3475         check_sync_rss_stat(current);
3476
3477         if (unlikely(is_vm_hugetlb_page(vma)))
3478                 return hugetlb_fault(mm, vma, address, flags);
3479
3480         pgd = pgd_offset(mm, address);
3481         pud = pud_alloc(mm, pgd, address);
3482         if (!pud)
3483                 return VM_FAULT_OOM;
3484         pmd = pmd_alloc(mm, pud, address);
3485         if (!pmd)
3486                 return VM_FAULT_OOM;
3487         if (pmd_none(*pmd) && transparent_hugepage_enabled(vma)) {
3488                 if (!vma->vm_ops)
3489                         return do_huge_pmd_anonymous_page(mm, vma, address,
3490                                                           pmd, flags);
3491         } else {
3492                 pmd_t orig_pmd = *pmd;
3493                 barrier();
3494                 if (pmd_trans_huge(orig_pmd)) {
3495                         if (flags & FAULT_FLAG_WRITE &&
3496                             !pmd_write(orig_pmd) &&
3497                             !pmd_trans_splitting(orig_pmd))
3498                                 return do_huge_pmd_wp_page(mm, vma, address,
3499                                                            pmd, orig_pmd);
3500                         return 0;
3501                 }
3502         }
3503
3504         /*
3505          * Use __pte_alloc instead of pte_alloc_map, because we can't
3506          * run pte_offset_map on the pmd, if an huge pmd could
3507          * materialize from under us from a different thread.
3508          */
3509         if (unlikely(pmd_none(*pmd)) && __pte_alloc(mm, vma, pmd, address))
3510                 return VM_FAULT_OOM;
3511         /* if an huge pmd materialized from under us just retry later */
3512         if (unlikely(pmd_trans_huge(*pmd)))
3513                 return 0;
3514         /*
3515          * A regular pmd is established and it can't morph into a huge pmd
3516          * from under us anymore at this point because we hold the mmap_sem
3517          * read mode and khugepaged takes it in write mode. So now it's
3518          * safe to run pte_offset_map().
3519          */
3520         pte = pte_offset_map(pmd, address);
3521
3522         return handle_pte_fault(mm, vma, address, pte, pmd, flags);
3523 }
3524
3525 #ifndef __PAGETABLE_PUD_FOLDED
3526 /*
3527  * Allocate page upper directory.
3528  * We've already handled the fast-path in-line.
3529  */
3530 int __pud_alloc(struct mm_struct *mm, pgd_t *pgd, unsigned long address)
3531 {
3532         pud_t *new = pud_alloc_one(mm, address);
3533         if (!new)
3534                 return -ENOMEM;
3535
3536         smp_wmb(); /* See comment in __pte_alloc */
3537
3538         spin_lock(&mm->page_table_lock);
3539         if (pgd_present(*pgd))          /* Another has populated it */
3540                 pud_free(mm, new);
3541         else
3542                 pgd_populate(mm, pgd, new);
3543         spin_unlock(&mm->page_table_lock);
3544         return 0;
3545 }
3546 #endif /* __PAGETABLE_PUD_FOLDED */
3547
3548 #ifndef __PAGETABLE_PMD_FOLDED
3549 /*
3550  * Allocate page middle directory.
3551  * We've already handled the fast-path in-line.
3552  */
3553 int __pmd_alloc(struct mm_struct *mm, pud_t *pud, unsigned long address)
3554 {
3555         pmd_t *new = pmd_alloc_one(mm, address);
3556         if (!new)
3557                 return -ENOMEM;
3558
3559         smp_wmb(); /* See comment in __pte_alloc */
3560
3561         spin_lock(&mm->page_table_lock);
3562 #ifndef __ARCH_HAS_4LEVEL_HACK
3563         if (pud_present(*pud))          /* Another has populated it */
3564                 pmd_free(mm, new);
3565         else
3566                 pud_populate(mm, pud, new);
3567 #else
3568         if (pgd_present(*pud))          /* Another has populated it */
3569                 pmd_free(mm, new);
3570         else
3571                 pgd_populate(mm, pud, new);
3572 #endif /* __ARCH_HAS_4LEVEL_HACK */
3573         spin_unlock(&mm->page_table_lock);
3574         return 0;
3575 }
3576 #endif /* __PAGETABLE_PMD_FOLDED */
3577
3578 int make_pages_present(unsigned long addr, unsigned long end)
3579 {
3580         int ret, len, write;
3581         struct vm_area_struct * vma;
3582
3583         vma = find_vma(current->mm, addr);
3584         if (!vma)
3585                 return -ENOMEM;
3586         /*
3587          * We want to touch writable mappings with a write fault in order
3588          * to break COW, except for shared mappings because these don't COW
3589          * and we would not want to dirty them for nothing.
3590          */
3591         write = (vma->vm_flags & (VM_WRITE | VM_SHARED)) == VM_WRITE;
3592         BUG_ON(addr >= end);
3593         BUG_ON(end > vma->vm_end);
3594         len = DIV_ROUND_UP(end, PAGE_SIZE) - addr/PAGE_SIZE;
3595         ret = get_user_pages(current, current->mm, addr,
3596                         len, write, 0, NULL, NULL);
3597         if (ret < 0)
3598                 return ret;
3599         return ret == len ? 0 : -EFAULT;
3600 }
3601
3602 #if !defined(__HAVE_ARCH_GATE_AREA)
3603
3604 #if defined(AT_SYSINFO_EHDR)
3605 static struct vm_area_struct gate_vma;
3606
3607 static int __init gate_vma_init(void)
3608 {
3609         gate_vma.vm_mm = NULL;
3610         gate_vma.vm_start = FIXADDR_USER_START;
3611         gate_vma.vm_end = FIXADDR_USER_END;
3612         gate_vma.vm_flags = VM_READ | VM_MAYREAD | VM_EXEC | VM_MAYEXEC;
3613         gate_vma.vm_page_prot = __P101;
3614         /*
3615          * Make sure the vDSO gets into every core dump.
3616          * Dumping its contents makes post-mortem fully interpretable later
3617          * without matching up the same kernel and hardware config to see
3618          * what PC values meant.
3619          */
3620         gate_vma.vm_flags |= VM_ALWAYSDUMP;
3621         return 0;
3622 }
3623 __initcall(gate_vma_init);
3624 #endif
3625
3626 struct vm_area_struct *get_gate_vma(struct mm_struct *mm)
3627 {
3628 #ifdef AT_SYSINFO_EHDR
3629         return &gate_vma;
3630 #else
3631         return NULL;
3632 #endif
3633 }
3634
3635 int in_gate_area_no_mm(unsigned long addr)
3636 {
3637 #ifdef AT_SYSINFO_EHDR
3638         if ((addr >= FIXADDR_USER_START) && (addr < FIXADDR_USER_END))
3639                 return 1;
3640 #endif
3641         return 0;
3642 }
3643
3644 #endif  /* __HAVE_ARCH_GATE_AREA */
3645
3646 static int __follow_pte(struct mm_struct *mm, unsigned long address,
3647                 pte_t **ptepp, spinlock_t **ptlp)
3648 {
3649         pgd_t *pgd;
3650         pud_t *pud;
3651         pmd_t *pmd;
3652         pte_t *ptep;
3653
3654         pgd = pgd_offset(mm, address);
3655         if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
3656                 goto out;
3657
3658         pud = pud_offset(pgd, address);
3659         if (pud_none(*pud) || unlikely(pud_bad(*pud)))
3660                 goto out;
3661
3662         pmd = pmd_offset(pud, address);
3663         VM_BUG_ON(pmd_trans_huge(*pmd));
3664         if (pmd_none(*pmd) || unlikely(pmd_bad(*pmd)))
3665                 goto out;
3666
3667         /* We cannot handle huge page PFN maps. Luckily they don't exist. */
3668         if (pmd_huge(*pmd))
3669                 goto out;
3670
3671         ptep = pte_offset_map_lock(mm, pmd, address, ptlp);
3672         if (!ptep)
3673                 goto out;
3674         if (!pte_present(*ptep))
3675                 goto unlock;
3676         *ptepp = ptep;
3677         return 0;
3678 unlock:
3679         pte_unmap_unlock(ptep, *ptlp);
3680 out:
3681         return -EINVAL;
3682 }
3683
3684 static inline int follow_pte(struct mm_struct *mm, unsigned long address,
3685                              pte_t **ptepp, spinlock_t **ptlp)
3686 {
3687         int res;
3688
3689         /* (void) is needed to make gcc happy */
3690         (void) __cond_lock(*ptlp,
3691                            !(res = __follow_pte(mm, address, ptepp, ptlp)));
3692         return res;
3693 }
3694
3695 /**
3696  * follow_pfn - look up PFN at a user virtual address
3697  * @vma: memory mapping
3698  * @address: user virtual address
3699  * @pfn: location to store found PFN
3700  *
3701  * Only IO mappings and raw PFN mappings are allowed.
3702  *
3703  * Returns zero and the pfn at @pfn on success, -ve otherwise.
3704  */
3705 int follow_pfn(struct vm_area_struct *vma, unsigned long address,
3706         unsigned long *pfn)
3707 {
3708         int ret = -EINVAL;
3709         spinlock_t *ptl;
3710         pte_t *ptep;
3711
3712         if (!(vma->vm_flags & (VM_IO | VM_PFNMAP)))
3713                 return ret;
3714
3715         ret = follow_pte(vma->vm_mm, address, &ptep, &ptl);
3716         if (ret)
3717                 return ret;
3718         *pfn = pte_pfn(*ptep);
3719         pte_unmap_unlock(ptep, ptl);
3720         return 0;
3721 }
3722 EXPORT_SYMBOL(follow_pfn);
3723
3724 #ifdef CONFIG_HAVE_IOREMAP_PROT
3725 int follow_phys(struct vm_area_struct *vma,
3726                 unsigned long address, unsigned int flags,
3727                 unsigned long *prot, resource_size_t *phys)
3728 {
3729         int ret = -EINVAL;
3730         pte_t *ptep, pte;
3731         spinlock_t *ptl;
3732
3733         if (!(vma->vm_flags & (VM_IO | VM_PFNMAP)))
3734                 goto out;
3735
3736         if (follow_pte(vma->vm_mm, address, &ptep, &ptl))
3737                 goto out;
3738         pte = *ptep;
3739
3740         if ((flags & FOLL_WRITE) && !pte_write(pte))
3741                 goto unlock;
3742
3743         *prot = pgprot_val(pte_pgprot(pte));
3744         *phys = (resource_size_t)pte_pfn(pte) << PAGE_SHIFT;
3745
3746         ret = 0;
3747 unlock:
3748         pte_unmap_unlock(ptep, ptl);
3749 out:
3750         return ret;
3751 }
3752
3753 int generic_access_phys(struct vm_area_struct *vma, unsigned long addr,
3754                         void *buf, int len, int write)
3755 {
3756         resource_size_t phys_addr;
3757         unsigned long prot = 0;
3758         void __iomem *maddr;
3759         int offset = addr & (PAGE_SIZE-1);
3760
3761         if (follow_phys(vma, addr, write, &prot, &phys_addr))
3762                 return -EINVAL;
3763
3764         maddr = ioremap_prot(phys_addr, PAGE_SIZE, prot);
3765         if (write)
3766                 memcpy_toio(maddr + offset, buf, len);
3767         else
3768                 memcpy_fromio(buf, maddr + offset, len);
3769         iounmap(maddr);
3770
3771         return len;
3772 }
3773 #endif
3774
3775 /*
3776  * Access another process' address space as given in mm.  If non-NULL, use the
3777  * given task for page fault accounting.
3778  */
3779 static int __access_remote_vm(struct task_struct *tsk, struct mm_struct *mm,
3780                 unsigned long addr, void *buf, int len, int write)
3781 {
3782         struct vm_area_struct *vma;
3783         void *old_buf = buf;
3784
3785         down_read(&mm->mmap_sem);
3786         /* ignore errors, just check how much was successfully transferred */
3787         while (len) {
3788                 int bytes, ret, offset;
3789                 void *maddr;
3790                 struct page *page = NULL;
3791
3792                 ret = get_user_pages(tsk, mm, addr, 1,
3793                                 write, 1, &page, &vma);
3794                 if (ret <= 0) {
3795                         /*
3796                          * Check if this is a VM_IO | VM_PFNMAP VMA, which
3797                          * we can access using slightly different code.
3798                          */
3799 #ifdef CONFIG_HAVE_IOREMAP_PROT
3800                         vma = find_vma(mm, addr);
3801                         if (!vma || vma->vm_start > addr)
3802                                 break;
3803                         if (vma->vm_ops && vma->vm_ops->access)
3804                                 ret = vma->vm_ops->access(vma, addr, buf,
3805                                                           len, write);
3806                         if (ret <= 0)
3807 #endif
3808                                 break;
3809                         bytes = ret;
3810                 } else {
3811                         bytes = len;
3812                         offset = addr & (PAGE_SIZE-1);
3813                         if (bytes > PAGE_SIZE-offset)
3814                                 bytes = PAGE_SIZE-offset;
3815
3816                         maddr = kmap(page);
3817                         if (write) {
3818                                 copy_to_user_page(vma, page, addr,
3819                                                   maddr + offset, buf, bytes);
3820                                 set_page_dirty_lock(page);
3821                         } else {
3822                                 copy_from_user_page(vma, page, addr,
3823                                                     buf, maddr + offset, bytes);
3824                         }
3825                         kunmap(page);
3826                         page_cache_release(page);
3827                 }
3828                 len -= bytes;
3829                 buf += bytes;
3830                 addr += bytes;
3831         }
3832         up_read(&mm->mmap_sem);
3833
3834         return buf - old_buf;
3835 }
3836
3837 /**
3838  * access_remote_vm - access another process' address space
3839  * @mm:         the mm_struct of the target address space
3840  * @addr:       start address to access
3841  * @buf:        source or destination buffer
3842  * @len:        number of bytes to transfer
3843  * @write:      whether the access is a write
3844  *
3845  * The caller must hold a reference on @mm.
3846  */
3847 int access_remote_vm(struct mm_struct *mm, unsigned long addr,
3848                 void *buf, int len, int write)
3849 {
3850         return __access_remote_vm(NULL, mm, addr, buf, len, write);
3851 }
3852
3853 /*
3854  * Access another process' address space.
3855  * Source/target buffer must be kernel space,
3856  * Do not walk the page table directly, use get_user_pages
3857  */
3858 int access_process_vm(struct task_struct *tsk, unsigned long addr,
3859                 void *buf, int len, int write)
3860 {
3861         struct mm_struct *mm;
3862         int ret;
3863
3864         mm = get_task_mm(tsk);
3865         if (!mm)
3866                 return 0;
3867
3868         ret = __access_remote_vm(tsk, mm, addr, buf, len, write);
3869         mmput(mm);
3870
3871         return ret;
3872 }
3873
3874 /*
3875  * Print the name of a VMA.
3876  */
3877 void print_vma_addr(char *prefix, unsigned long ip)
3878 {
3879         struct mm_struct *mm = current->mm;
3880         struct vm_area_struct *vma;
3881
3882         /*
3883          * Do not print if we are in atomic
3884          * contexts (in exception stacks, etc.):
3885          */
3886         if (preempt_count())
3887                 return;
3888
3889         down_read(&mm->mmap_sem);
3890         vma = find_vma(mm, ip);
3891         if (vma && vma->vm_file) {
3892                 struct file *f = vma->vm_file;
3893                 char *buf = (char *)__get_free_page(GFP_KERNEL);
3894                 if (buf) {
3895                         char *p, *s;
3896
3897                         p = d_path(&f->f_path, buf, PAGE_SIZE);
3898                         if (IS_ERR(p))
3899                                 p = "?";
3900                         s = strrchr(p, '/');
3901                         if (s)
3902                                 p = s+1;
3903                         printk("%s%s[%lx+%lx]", prefix, p,
3904                                         vma->vm_start,
3905                                         vma->vm_end - vma->vm_start);
3906                         free_page((unsigned long)buf);
3907                 }
3908         }
3909         up_read(&current->mm->mmap_sem);
3910 }
3911
3912 #ifdef CONFIG_PROVE_LOCKING
3913 void might_fault(void)
3914 {
3915         /*
3916          * Some code (nfs/sunrpc) uses socket ops on kernel memory while
3917          * holding the mmap_sem, this is safe because kernel memory doesn't
3918          * get paged out, therefore we'll never actually fault, and the
3919          * below annotations will generate false positives.
3920          */
3921         if (segment_eq(get_fs(), KERNEL_DS))
3922                 return;
3923
3924         might_sleep();
3925         /*
3926          * it would be nicer only to annotate paths which are not under
3927          * pagefault_disable, however that requires a larger audit and
3928          * providing helpers like get_user_atomic.
3929          */
3930         if (!in_atomic() && current->mm)
3931                 might_lock_read(&current->mm->mmap_sem);
3932 }
3933 EXPORT_SYMBOL(might_fault);
3934 #endif
3935
3936 #if defined(CONFIG_TRANSPARENT_HUGEPAGE) || defined(CONFIG_HUGETLBFS)
3937 static void clear_gigantic_page(struct page *page,
3938                                 unsigned long addr,
3939                                 unsigned int pages_per_huge_page)
3940 {
3941         int i;
3942         struct page *p = page;
3943
3944         might_sleep();
3945         for (i = 0; i < pages_per_huge_page;
3946              i++, p = mem_map_next(p, page, i)) {
3947                 cond_resched();
3948                 clear_user_highpage(p, addr + i * PAGE_SIZE);
3949         }
3950 }
3951 void clear_huge_page(struct page *page,
3952                      unsigned long addr, unsigned int pages_per_huge_page)
3953 {
3954         int i;
3955
3956         if (unlikely(pages_per_huge_page > MAX_ORDER_NR_PAGES)) {
3957                 clear_gigantic_page(page, addr, pages_per_huge_page);
3958                 return;
3959         }
3960
3961         might_sleep();
3962         for (i = 0; i < pages_per_huge_page; i++) {
3963                 cond_resched();
3964                 clear_user_highpage(page + i, addr + i * PAGE_SIZE);
3965         }
3966 }
3967
3968 static void copy_user_gigantic_page(struct page *dst, struct page *src,
3969                                     unsigned long addr,
3970                                     struct vm_area_struct *vma,
3971                                     unsigned int pages_per_huge_page)
3972 {
3973         int i;
3974         struct page *dst_base = dst;
3975         struct page *src_base = src;
3976
3977         for (i = 0; i < pages_per_huge_page; ) {
3978                 cond_resched();
3979                 copy_user_highpage(dst, src, addr + i*PAGE_SIZE, vma);
3980
3981                 i++;
3982                 dst = mem_map_next(dst, dst_base, i);
3983                 src = mem_map_next(src, src_base, i);
3984         }
3985 }
3986
3987 void copy_user_huge_page(struct page *dst, struct page *src,
3988                          unsigned long addr, struct vm_area_struct *vma,
3989                          unsigned int pages_per_huge_page)
3990 {
3991         int i;
3992
3993         if (unlikely(pages_per_huge_page > MAX_ORDER_NR_PAGES)) {
3994                 copy_user_gigantic_page(dst, src, addr, vma,
3995                                         pages_per_huge_page);
3996                 return;
3997         }
3998
3999         might_sleep();
4000         for (i = 0; i < pages_per_huge_page; i++) {
4001                 cond_resched();
4002                 copy_user_highpage(dst + i, src + i, addr + i*PAGE_SIZE, vma);
4003         }
4004 }
4005 #endif /* CONFIG_TRANSPARENT_HUGEPAGE || CONFIG_HUGETLBFS */