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