cd8e0daf1912e1b35453a395721cf3dd74998bc8
[linux-3.10.git] / mm / memory.c
1 /*
2  *  linux/mm/memory.c
3  *
4  *  Copyright (C) 1991, 1992, 1993, 1994  Linus Torvalds
5  */
6
7 /*
8  * demand-loading started 01.12.91 - seems it is high on the list of
9  * things wanted, and it should be easy to implement. - Linus
10  */
11
12 /*
13  * Ok, demand-loading was easy, shared pages a little bit tricker. Shared
14  * pages started 02.12.91, seems to work. - Linus.
15  *
16  * Tested sharing by executing about 30 /bin/sh: under the old kernel it
17  * would have taken more than the 6M I have free, but it worked well as
18  * far as I could see.
19  *
20  * Also corrected some "invalidate()"s - I wasn't doing enough of them.
21  */
22
23 /*
24  * Real VM (paging to/from disk) started 18.12.91. Much more work and
25  * thought has to go into this. Oh, well..
26  * 19.12.91  -  works, somewhat. Sometimes I get faults, don't know why.
27  *              Found it. Everything seems to work now.
28  * 20.12.91  -  Ok, making the swap-device changeable like the root.
29  */
30
31 /*
32  * 05.04.94  -  Multi-page memory management added for v1.1.
33  *              Idea by Alex Bligh (alex@cconcepts.co.uk)
34  *
35  * 16.07.99  -  Support of BIGMEM added by Gerhard Wichert, Siemens AG
36  *              (Gerhard.Wichert@pdb.siemens.de)
37  *
38  * Aug/Sep 2004 Changed to four level page tables (Andi Kleen)
39  */
40
41 #include <linux/kernel_stat.h>
42 #include <linux/mm.h>
43 #include <linux/hugetlb.h>
44 #include <linux/mman.h>
45 #include <linux/swap.h>
46 #include <linux/highmem.h>
47 #include <linux/pagemap.h>
48 #include <linux/ksm.h>
49 #include <linux/rmap.h>
50 #include <linux/export.h>
51 #include <linux/delayacct.h>
52 #include <linux/init.h>
53 #include <linux/writeback.h>
54 #include <linux/memcontrol.h>
55 #include <linux/mmu_notifier.h>
56 #include <linux/kallsyms.h>
57 #include <linux/swapops.h>
58 #include <linux/elf.h>
59 #include <linux/gfp.h>
60
61 #include <asm/io.h>
62 #include <asm/pgalloc.h>
63 #include <asm/uaccess.h>
64 #include <asm/tlb.h>
65 #include <asm/tlbflush.h>
66 #include <asm/pgtable.h>
67
68 #include "internal.h"
69
70 #ifndef CONFIG_NEED_MULTIPLE_NODES
71 /* use the per-pgdat data instead for discontigmem - mbligh */
72 unsigned long max_mapnr;
73 struct page *mem_map;
74
75 EXPORT_SYMBOL(max_mapnr);
76 EXPORT_SYMBOL(mem_map);
77 #endif
78
79 unsigned long num_physpages;
80 /*
81  * A number of key systems in x86 including ioremap() rely on the assumption
82  * that high_memory defines the upper bound on direct map memory, then end
83  * of ZONE_NORMAL.  Under CONFIG_DISCONTIG this means that max_low_pfn and
84  * highstart_pfn must be the same; there must be no gap between ZONE_NORMAL
85  * and ZONE_HIGHMEM.
86  */
87 void * high_memory;
88
89 EXPORT_SYMBOL(num_physpages);
90 EXPORT_SYMBOL(high_memory);
91
92 /*
93  * Randomize the address space (stacks, mmaps, brk, etc.).
94  *
95  * ( When CONFIG_COMPAT_BRK=y we exclude brk from randomization,
96  *   as ancient (libc5 based) binaries can segfault. )
97  */
98 int randomize_va_space __read_mostly =
99 #ifdef CONFIG_COMPAT_BRK
100                                         1;
101 #else
102                                         2;
103 #endif
104
105 static int __init disable_randmaps(char *s)
106 {
107         randomize_va_space = 0;
108         return 1;
109 }
110 __setup("norandmaps", disable_randmaps);
111
112 unsigned long zero_pfn __read_mostly;
113 unsigned long highest_memmap_pfn __read_mostly;
114
115 /*
116  * CONFIG_MMU architectures set up ZERO_PAGE in their paging_init()
117  */
118 static int __init init_zero_pfn(void)
119 {
120         zero_pfn = page_to_pfn(ZERO_PAGE(0));
121         return 0;
122 }
123 core_initcall(init_zero_pfn);
124
125
126 #if defined(SPLIT_RSS_COUNTING)
127
128 void sync_mm_rss(struct mm_struct *mm)
129 {
130         int i;
131
132         for (i = 0; i < NR_MM_COUNTERS; i++) {
133                 if (current->rss_stat.count[i]) {
134                         add_mm_counter(mm, i, current->rss_stat.count[i]);
135                         current->rss_stat.count[i] = 0;
136                 }
137         }
138         current->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_mm_rss(task->mm);
161 }
162 #else /* SPLIT_RSS_COUNTING */
163
164 #define inc_mm_counter_fast(mm, member) inc_mm_counter(mm, member)
165 #define dec_mm_counter_fast(mm, member) dec_mm_counter(mm, member)
166
167 static void check_sync_rss_stat(struct task_struct *task)
168 {
169 }
170
171 #endif /* SPLIT_RSS_COUNTING */
172
173 #ifdef HAVE_GENERIC_MMU_GATHER
174
175 static int tlb_next_batch(struct mmu_gather *tlb)
176 {
177         struct mmu_gather_batch *batch;
178
179         batch = tlb->active;
180         if (batch->next) {
181                 tlb->active = batch->next;
182                 return 1;
183         }
184
185         batch = (void *)__get_free_pages(GFP_NOWAIT | __GFP_NOWARN, 0);
186         if (!batch)
187                 return 0;
188
189         batch->next = NULL;
190         batch->nr   = 0;
191         batch->max  = MAX_GATHER_BATCH;
192
193         tlb->active->next = batch;
194         tlb->active = batch;
195
196         return 1;
197 }
198
199 /* tlb_gather_mmu
200  *      Called to initialize an (on-stack) mmu_gather structure for page-table
201  *      tear-down from @mm. The @fullmm argument is used when @mm is without
202  *      users and we're going to destroy the full address space (exit/execve).
203  */
204 void tlb_gather_mmu(struct mmu_gather *tlb, struct mm_struct *mm, bool fullmm)
205 {
206         tlb->mm = mm;
207
208         tlb->fullmm     = fullmm;
209         tlb->start      = -1UL;
210         tlb->end        = 0;
211         tlb->need_flush = 0;
212         tlb->fast_mode  = (num_possible_cpus() == 1);
213         tlb->local.next = NULL;
214         tlb->local.nr   = 0;
215         tlb->local.max  = ARRAY_SIZE(tlb->__pages);
216         tlb->active     = &tlb->local;
217
218 #ifdef CONFIG_HAVE_RCU_TABLE_FREE
219         tlb->batch = NULL;
220 #endif
221 }
222
223 void tlb_flush_mmu(struct mmu_gather *tlb)
224 {
225         struct mmu_gather_batch *batch;
226
227         if (!tlb->need_flush)
228                 return;
229         tlb->need_flush = 0;
230         tlb_flush(tlb);
231 #ifdef CONFIG_HAVE_RCU_TABLE_FREE
232         tlb_table_flush(tlb);
233 #endif
234
235         if (tlb_fast_mode(tlb))
236                 return;
237
238         for (batch = &tlb->local; batch; batch = batch->next) {
239                 free_pages_and_swap_cache(batch->pages, batch->nr);
240                 batch->nr = 0;
241         }
242         tlb->active = &tlb->local;
243 }
244
245 /* tlb_finish_mmu
246  *      Called at the end of the shootdown operation to free up any resources
247  *      that were required.
248  */
249 void tlb_finish_mmu(struct mmu_gather *tlb, unsigned long start, unsigned long end)
250 {
251         struct mmu_gather_batch *batch, *next;
252
253         tlb->start = start;
254         tlb->end   = end;
255         tlb_flush_mmu(tlb);
256
257         /* keep the page table cache within bounds */
258         check_pgt_cache();
259
260         for (batch = tlb->local.next; batch; batch = next) {
261                 next = batch->next;
262                 free_pages((unsigned long)batch, 0);
263         }
264         tlb->local.next = NULL;
265 }
266
267 /* __tlb_remove_page
268  *      Must perform the equivalent to __free_pte(pte_get_and_clear(ptep)), while
269  *      handling the additional races in SMP caused by other CPUs caching valid
270  *      mappings in their TLBs. Returns the number of free page slots left.
271  *      When out of page slots we must call tlb_flush_mmu().
272  */
273 int __tlb_remove_page(struct mmu_gather *tlb, struct page *page)
274 {
275         struct mmu_gather_batch *batch;
276
277         VM_BUG_ON(!tlb->need_flush);
278
279         if (tlb_fast_mode(tlb)) {
280                 free_page_and_swap_cache(page);
281                 return 1; /* avoid calling tlb_flush_mmu() */
282         }
283
284         batch = tlb->active;
285         batch->pages[batch->nr++] = page;
286         if (batch->nr == batch->max) {
287                 if (!tlb_next_batch(tlb))
288                         return 0;
289                 batch = tlb->active;
290         }
291         VM_BUG_ON(batch->nr > batch->max);
292
293         return batch->max - batch->nr;
294 }
295
296 #endif /* HAVE_GENERIC_MMU_GATHER */
297
298 #ifdef CONFIG_HAVE_RCU_TABLE_FREE
299
300 /*
301  * See the comment near struct mmu_table_batch.
302  */
303
304 static void tlb_remove_table_smp_sync(void *arg)
305 {
306         /* Simply deliver the interrupt */
307 }
308
309 static void tlb_remove_table_one(void *table)
310 {
311         /*
312          * This isn't an RCU grace period and hence the page-tables cannot be
313          * assumed to be actually RCU-freed.
314          *
315          * It is however sufficient for software page-table walkers that rely on
316          * IRQ disabling. See the comment near struct mmu_table_batch.
317          */
318         smp_call_function(tlb_remove_table_smp_sync, NULL, 1);
319         __tlb_remove_table(table);
320 }
321
322 static void tlb_remove_table_rcu(struct rcu_head *head)
323 {
324         struct mmu_table_batch *batch;
325         int i;
326
327         batch = container_of(head, struct mmu_table_batch, rcu);
328
329         for (i = 0; i < batch->nr; i++)
330                 __tlb_remove_table(batch->tables[i]);
331
332         free_page((unsigned long)batch);
333 }
334
335 void tlb_table_flush(struct mmu_gather *tlb)
336 {
337         struct mmu_table_batch **batch = &tlb->batch;
338
339         if (*batch) {
340                 call_rcu_sched(&(*batch)->rcu, tlb_remove_table_rcu);
341                 *batch = NULL;
342         }
343 }
344
345 void tlb_remove_table(struct mmu_gather *tlb, void *table)
346 {
347         struct mmu_table_batch **batch = &tlb->batch;
348
349         tlb->need_flush = 1;
350
351         /*
352          * When there's less then two users of this mm there cannot be a
353          * concurrent page-table walk.
354          */
355         if (atomic_read(&tlb->mm->mm_users) < 2) {
356                 __tlb_remove_table(table);
357                 return;
358         }
359
360         if (*batch == NULL) {
361                 *batch = (struct mmu_table_batch *)__get_free_page(GFP_NOWAIT | __GFP_NOWARN);
362                 if (*batch == NULL) {
363                         tlb_remove_table_one(table);
364                         return;
365                 }
366                 (*batch)->nr = 0;
367         }
368         (*batch)->tables[(*batch)->nr++] = table;
369         if ((*batch)->nr == MAX_TABLE_BATCH)
370                 tlb_table_flush(tlb);
371 }
372
373 #endif /* CONFIG_HAVE_RCU_TABLE_FREE */
374
375 /*
376  * If a p?d_bad entry is found while walking page tables, report
377  * the error, before resetting entry to p?d_none.  Usually (but
378  * very seldom) called out from the p?d_none_or_clear_bad macros.
379  */
380
381 void pgd_clear_bad(pgd_t *pgd)
382 {
383         pgd_ERROR(*pgd);
384         pgd_clear(pgd);
385 }
386
387 void pud_clear_bad(pud_t *pud)
388 {
389         pud_ERROR(*pud);
390         pud_clear(pud);
391 }
392
393 void pmd_clear_bad(pmd_t *pmd)
394 {
395         pmd_ERROR(*pmd);
396         pmd_clear(pmd);
397 }
398
399 /*
400  * Note: this doesn't free the actual pages themselves. That
401  * has been handled earlier when unmapping all the memory regions.
402  */
403 static void free_pte_range(struct mmu_gather *tlb, pmd_t *pmd,
404                            unsigned long addr)
405 {
406         pgtable_t token = pmd_pgtable(*pmd);
407         pmd_clear(pmd);
408         pte_free_tlb(tlb, token, addr);
409         tlb->mm->nr_ptes--;
410 }
411
412 static inline void free_pmd_range(struct mmu_gather *tlb, pud_t *pud,
413                                 unsigned long addr, unsigned long end,
414                                 unsigned long floor, unsigned long ceiling)
415 {
416         pmd_t *pmd;
417         unsigned long next;
418         unsigned long start;
419
420         start = addr;
421         pmd = pmd_offset(pud, addr);
422         do {
423                 next = pmd_addr_end(addr, end);
424                 if (pmd_none_or_clear_bad(pmd))
425                         continue;
426                 free_pte_range(tlb, pmd, addr);
427         } while (pmd++, addr = next, addr != end);
428
429         start &= PUD_MASK;
430         if (start < floor)
431                 return;
432         if (ceiling) {
433                 ceiling &= PUD_MASK;
434                 if (!ceiling)
435                         return;
436         }
437         if (end - 1 > ceiling - 1)
438                 return;
439
440         pmd = pmd_offset(pud, start);
441         pud_clear(pud);
442         pmd_free_tlb(tlb, pmd, start);
443 }
444
445 static inline void free_pud_range(struct mmu_gather *tlb, pgd_t *pgd,
446                                 unsigned long addr, unsigned long end,
447                                 unsigned long floor, unsigned long ceiling)
448 {
449         pud_t *pud;
450         unsigned long next;
451         unsigned long start;
452
453         start = addr;
454         pud = pud_offset(pgd, addr);
455         do {
456                 next = pud_addr_end(addr, end);
457                 if (pud_none_or_clear_bad(pud))
458                         continue;
459                 free_pmd_range(tlb, pud, addr, next, floor, ceiling);
460         } while (pud++, addr = next, addr != end);
461
462         start &= PGDIR_MASK;
463         if (start < floor)
464                 return;
465         if (ceiling) {
466                 ceiling &= PGDIR_MASK;
467                 if (!ceiling)
468                         return;
469         }
470         if (end - 1 > ceiling - 1)
471                 return;
472
473         pud = pud_offset(pgd, start);
474         pgd_clear(pgd);
475         pud_free_tlb(tlb, pud, start);
476 }
477
478 /*
479  * This function frees user-level page tables of a process.
480  *
481  * Must be called with pagetable lock held.
482  */
483 void free_pgd_range(struct mmu_gather *tlb,
484                         unsigned long addr, unsigned long end,
485                         unsigned long floor, unsigned long ceiling)
486 {
487         pgd_t *pgd;
488         unsigned long next;
489
490         /*
491          * The next few lines have given us lots of grief...
492          *
493          * Why are we testing PMD* at this top level?  Because often
494          * there will be no work to do at all, and we'd prefer not to
495          * go all the way down to the bottom just to discover that.
496          *
497          * Why all these "- 1"s?  Because 0 represents both the bottom
498          * of the address space and the top of it (using -1 for the
499          * top wouldn't help much: the masks would do the wrong thing).
500          * The rule is that addr 0 and floor 0 refer to the bottom of
501          * the address space, but end 0 and ceiling 0 refer to the top
502          * Comparisons need to use "end - 1" and "ceiling - 1" (though
503          * that end 0 case should be mythical).
504          *
505          * Wherever addr is brought up or ceiling brought down, we must
506          * be careful to reject "the opposite 0" before it confuses the
507          * subsequent tests.  But what about where end is brought down
508          * by PMD_SIZE below? no, end can't go down to 0 there.
509          *
510          * Whereas we round start (addr) and ceiling down, by different
511          * masks at different levels, in order to test whether a table
512          * now has no other vmas using it, so can be freed, we don't
513          * bother to round floor or end up - the tests don't need that.
514          */
515
516         addr &= PMD_MASK;
517         if (addr < floor) {
518                 addr += PMD_SIZE;
519                 if (!addr)
520                         return;
521         }
522         if (ceiling) {
523                 ceiling &= PMD_MASK;
524                 if (!ceiling)
525                         return;
526         }
527         if (end - 1 > ceiling - 1)
528                 end -= PMD_SIZE;
529         if (addr > end - 1)
530                 return;
531
532         pgd = pgd_offset(tlb->mm, addr);
533         do {
534                 next = pgd_addr_end(addr, end);
535                 if (pgd_none_or_clear_bad(pgd))
536                         continue;
537                 free_pud_range(tlb, pgd, addr, next, floor, ceiling);
538         } while (pgd++, addr = next, addr != end);
539 }
540
541 void free_pgtables(struct mmu_gather *tlb, struct vm_area_struct *vma,
542                 unsigned long floor, unsigned long ceiling)
543 {
544         while (vma) {
545                 struct vm_area_struct *next = vma->vm_next;
546                 unsigned long addr = vma->vm_start;
547
548                 /*
549                  * Hide vma from rmap and truncate_pagecache before freeing
550                  * pgtables
551                  */
552                 unlink_anon_vmas(vma);
553                 unlink_file_vma(vma);
554
555                 if (is_vm_hugetlb_page(vma)) {
556                         hugetlb_free_pgd_range(tlb, addr, vma->vm_end,
557                                 floor, next? next->vm_start: ceiling);
558                 } else {
559                         /*
560                          * Optimization: gather nearby vmas into one call down
561                          */
562                         while (next && next->vm_start <= vma->vm_end + PMD_SIZE
563                                && !is_vm_hugetlb_page(next)) {
564                                 vma = next;
565                                 next = vma->vm_next;
566                                 unlink_anon_vmas(vma);
567                                 unlink_file_vma(vma);
568                         }
569                         free_pgd_range(tlb, addr, vma->vm_end,
570                                 floor, next? next->vm_start: ceiling);
571                 }
572                 vma = next;
573         }
574 }
575
576 int __pte_alloc(struct mm_struct *mm, struct vm_area_struct *vma,
577                 pmd_t *pmd, unsigned long address)
578 {
579         pgtable_t new = pte_alloc_one(mm, address);
580         int wait_split_huge_page;
581         if (!new)
582                 return -ENOMEM;
583
584         /*
585          * Ensure all pte setup (eg. pte page lock and page clearing) are
586          * visible before the pte is made visible to other CPUs by being
587          * put into page tables.
588          *
589          * The other side of the story is the pointer chasing in the page
590          * table walking code (when walking the page table without locking;
591          * ie. most of the time). Fortunately, these data accesses consist
592          * of a chain of data-dependent loads, meaning most CPUs (alpha
593          * being the notable exception) will already guarantee loads are
594          * seen in-order. See the alpha page table accessors for the
595          * smp_read_barrier_depends() barriers in page table walking code.
596          */
597         smp_wmb(); /* Could be smp_wmb__xxx(before|after)_spin_lock */
598
599         spin_lock(&mm->page_table_lock);
600         wait_split_huge_page = 0;
601         if (likely(pmd_none(*pmd))) {   /* Has another populated it ? */
602                 mm->nr_ptes++;
603                 pmd_populate(mm, pmd, new);
604                 new = NULL;
605         } else if (unlikely(pmd_trans_splitting(*pmd)))
606                 wait_split_huge_page = 1;
607         spin_unlock(&mm->page_table_lock);
608         if (new)
609                 pte_free(mm, new);
610         if (wait_split_huge_page)
611                 wait_split_huge_page(vma->anon_vma, pmd);
612         return 0;
613 }
614
615 int __pte_alloc_kernel(pmd_t *pmd, unsigned long address)
616 {
617         pte_t *new = pte_alloc_one_kernel(&init_mm, address);
618         if (!new)
619                 return -ENOMEM;
620
621         smp_wmb(); /* See comment in __pte_alloc */
622
623         spin_lock(&init_mm.page_table_lock);
624         if (likely(pmd_none(*pmd))) {   /* Has another populated it ? */
625                 pmd_populate_kernel(&init_mm, pmd, new);
626                 new = NULL;
627         } else
628                 VM_BUG_ON(pmd_trans_splitting(*pmd));
629         spin_unlock(&init_mm.page_table_lock);
630         if (new)
631                 pte_free_kernel(&init_mm, new);
632         return 0;
633 }
634
635 static inline void init_rss_vec(int *rss)
636 {
637         memset(rss, 0, sizeof(int) * NR_MM_COUNTERS);
638 }
639
640 static inline void add_mm_rss_vec(struct mm_struct *mm, int *rss)
641 {
642         int i;
643
644         if (current->mm == mm)
645                 sync_mm_rss(mm);
646         for (i = 0; i < NR_MM_COUNTERS; i++)
647                 if (rss[i])
648                         add_mm_counter(mm, i, rss[i]);
649 }
650
651 /*
652  * This function is called to print an error when a bad pte
653  * is found. For example, we might have a PFN-mapped pte in
654  * a region that doesn't allow it.
655  *
656  * The calling function must still handle the error.
657  */
658 static void print_bad_pte(struct vm_area_struct *vma, unsigned long addr,
659                           pte_t pte, struct page *page)
660 {
661         pgd_t *pgd = pgd_offset(vma->vm_mm, addr);
662         pud_t *pud = pud_offset(pgd, addr);
663         pmd_t *pmd = pmd_offset(pud, addr);
664         struct address_space *mapping;
665         pgoff_t index;
666         static unsigned long resume;
667         static unsigned long nr_shown;
668         static unsigned long nr_unshown;
669
670         /*
671          * Allow a burst of 60 reports, then keep quiet for that minute;
672          * or allow a steady drip of one report per second.
673          */
674         if (nr_shown == 60) {
675                 if (time_before(jiffies, resume)) {
676                         nr_unshown++;
677                         return;
678                 }
679                 if (nr_unshown) {
680                         printk(KERN_ALERT
681                                 "BUG: Bad page map: %lu messages suppressed\n",
682                                 nr_unshown);
683                         nr_unshown = 0;
684                 }
685                 nr_shown = 0;
686         }
687         if (nr_shown++ == 0)
688                 resume = jiffies + 60 * HZ;
689
690         mapping = vma->vm_file ? vma->vm_file->f_mapping : NULL;
691         index = linear_page_index(vma, addr);
692
693         printk(KERN_ALERT
694                 "BUG: Bad page map in process %s  pte:%08llx pmd:%08llx\n",
695                 current->comm,
696                 (long long)pte_val(pte), (long long)pmd_val(*pmd));
697         if (page)
698                 dump_page(page);
699         printk(KERN_ALERT
700                 "addr:%p vm_flags:%08lx anon_vma:%p mapping:%p index:%lx\n",
701                 (void *)addr, vma->vm_flags, vma->anon_vma, mapping, index);
702         /*
703          * Choose text because data symbols depend on CONFIG_KALLSYMS_ALL=y
704          */
705         if (vma->vm_ops)
706                 print_symbol(KERN_ALERT "vma->vm_ops->fault: %s\n",
707                                 (unsigned long)vma->vm_ops->fault);
708         if (vma->vm_file && vma->vm_file->f_op)
709                 print_symbol(KERN_ALERT "vma->vm_file->f_op->mmap: %s\n",
710                                 (unsigned long)vma->vm_file->f_op->mmap);
711         dump_stack();
712         add_taint(TAINT_BAD_PAGE);
713 }
714
715 static inline bool is_cow_mapping(vm_flags_t flags)
716 {
717         return (flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE;
718 }
719
720 #ifndef is_zero_pfn
721 static inline int is_zero_pfn(unsigned long pfn)
722 {
723         return pfn == zero_pfn;
724 }
725 #endif
726
727 #ifndef my_zero_pfn
728 static inline unsigned long my_zero_pfn(unsigned long addr)
729 {
730         return zero_pfn;
731 }
732 #endif
733
734 /*
735  * vm_normal_page -- This function gets the "struct page" associated with a pte.
736  *
737  * "Special" mappings do not wish to be associated with a "struct page" (either
738  * it doesn't exist, or it exists but they don't want to touch it). In this
739  * case, NULL is returned here. "Normal" mappings do have a struct page.
740  *
741  * There are 2 broad cases. Firstly, an architecture may define a pte_special()
742  * pte bit, in which case this function is trivial. Secondly, an architecture
743  * may not have a spare pte bit, which requires a more complicated scheme,
744  * described below.
745  *
746  * A raw VM_PFNMAP mapping (ie. one that is not COWed) is always considered a
747  * special mapping (even if there are underlying and valid "struct pages").
748  * COWed pages of a VM_PFNMAP are always normal.
749  *
750  * The way we recognize COWed pages within VM_PFNMAP mappings is through the
751  * rules set up by "remap_pfn_range()": the vma will have the VM_PFNMAP bit
752  * set, and the vm_pgoff will point to the first PFN mapped: thus every special
753  * mapping will always honor the rule
754  *
755  *      pfn_of_page == vma->vm_pgoff + ((addr - vma->vm_start) >> PAGE_SHIFT)
756  *
757  * And for normal mappings this is false.
758  *
759  * This restricts such mappings to be a linear translation from virtual address
760  * to pfn. To get around this restriction, we allow arbitrary mappings so long
761  * as the vma is not a COW mapping; in that case, we know that all ptes are
762  * special (because none can have been COWed).
763  *
764  *
765  * In order to support COW of arbitrary special mappings, we have VM_MIXEDMAP.
766  *
767  * VM_MIXEDMAP mappings can likewise contain memory with or without "struct
768  * page" backing, however the difference is that _all_ pages with a struct
769  * page (that is, those where pfn_valid is true) are refcounted and considered
770  * normal pages by the VM. The disadvantage is that pages are refcounted
771  * (which can be slower and simply not an option for some PFNMAP users). The
772  * advantage is that we don't have to follow the strict linearity rule of
773  * PFNMAP mappings in order to support COWable mappings.
774  *
775  */
776 #ifdef __HAVE_ARCH_PTE_SPECIAL
777 # define HAVE_PTE_SPECIAL 1
778 #else
779 # define HAVE_PTE_SPECIAL 0
780 #endif
781 struct page *vm_normal_page(struct vm_area_struct *vma, unsigned long addr,
782                                 pte_t pte)
783 {
784         unsigned long pfn = pte_pfn(pte);
785
786         if (HAVE_PTE_SPECIAL) {
787                 if (likely(!pte_special(pte)))
788                         goto check_pfn;
789                 if (vma->vm_flags & (VM_PFNMAP | VM_MIXEDMAP))
790                         return NULL;
791                 if (!is_zero_pfn(pfn))
792                         print_bad_pte(vma, addr, pte, NULL);
793                 return NULL;
794         }
795
796         /* !HAVE_PTE_SPECIAL case follows: */
797
798         if (unlikely(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP))) {
799                 if (vma->vm_flags & VM_MIXEDMAP) {
800                         if (!pfn_valid(pfn))
801                                 return NULL;
802                         goto out;
803                 } else {
804                         unsigned long off;
805                         off = (addr - vma->vm_start) >> PAGE_SHIFT;
806                         if (pfn == vma->vm_pgoff + off)
807                                 return NULL;
808                         if (!is_cow_mapping(vma->vm_flags))
809                                 return NULL;
810                 }
811         }
812
813         if (is_zero_pfn(pfn))
814                 return NULL;
815 check_pfn:
816         if (unlikely(pfn > highest_memmap_pfn)) {
817                 print_bad_pte(vma, addr, pte, NULL);
818                 return NULL;
819         }
820
821         /*
822          * NOTE! We still have PageReserved() pages in the page tables.
823          * eg. VDSO mappings can cause them to exist.
824          */
825 out:
826         return pfn_to_page(pfn);
827 }
828
829 /*
830  * copy one vm_area from one task to the other. Assumes the page tables
831  * already present in the new task to be cleared in the whole range
832  * covered by this vma.
833  */
834
835 static inline unsigned long
836 copy_one_pte(struct mm_struct *dst_mm, struct mm_struct *src_mm,
837                 pte_t *dst_pte, pte_t *src_pte, struct vm_area_struct *vma,
838                 unsigned long addr, int *rss)
839 {
840         unsigned long vm_flags = vma->vm_flags;
841         pte_t pte = *src_pte;
842         struct page *page;
843
844         /* pte contains position in swap or file, so copy. */
845         if (unlikely(!pte_present(pte))) {
846                 if (!pte_file(pte)) {
847                         swp_entry_t entry = pte_to_swp_entry(pte);
848
849                         if (swap_duplicate(entry) < 0)
850                                 return entry.val;
851
852                         /* make sure dst_mm is on swapoff's mmlist. */
853                         if (unlikely(list_empty(&dst_mm->mmlist))) {
854                                 spin_lock(&mmlist_lock);
855                                 if (list_empty(&dst_mm->mmlist))
856                                         list_add(&dst_mm->mmlist,
857                                                  &src_mm->mmlist);
858                                 spin_unlock(&mmlist_lock);
859                         }
860                         if (likely(!non_swap_entry(entry)))
861                                 rss[MM_SWAPENTS]++;
862                         else if (is_migration_entry(entry)) {
863                                 page = migration_entry_to_page(entry);
864
865                                 if (PageAnon(page))
866                                         rss[MM_ANONPAGES]++;
867                                 else
868                                         rss[MM_FILEPAGES]++;
869
870                                 if (is_write_migration_entry(entry) &&
871                                     is_cow_mapping(vm_flags)) {
872                                         /*
873                                          * COW mappings require pages in both
874                                          * parent and child to be set to read.
875                                          */
876                                         make_migration_entry_read(&entry);
877                                         pte = swp_entry_to_pte(entry);
878                                         set_pte_at(src_mm, addr, src_pte, pte);
879                                 }
880                         }
881                 }
882                 goto out_set_pte;
883         }
884
885         /*
886          * If it's a COW mapping, write protect it both
887          * in the parent and the child
888          */
889         if (is_cow_mapping(vm_flags)) {
890                 ptep_set_wrprotect(src_mm, addr, src_pte);
891                 pte = pte_wrprotect(pte);
892         }
893
894         /*
895          * If it's a shared mapping, mark it clean in
896          * the child
897          */
898         if (vm_flags & VM_SHARED)
899                 pte = pte_mkclean(pte);
900         pte = pte_mkold(pte);
901
902         page = vm_normal_page(vma, addr, pte);
903         if (page) {
904                 get_page(page);
905                 page_dup_rmap(page);
906                 if (PageAnon(page))
907                         rss[MM_ANONPAGES]++;
908                 else
909                         rss[MM_FILEPAGES]++;
910         }
911
912 out_set_pte:
913         set_pte_at(dst_mm, addr, dst_pte, pte);
914         return 0;
915 }
916
917 int copy_pte_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
918                    pmd_t *dst_pmd, pmd_t *src_pmd, struct vm_area_struct *vma,
919                    unsigned long addr, unsigned long end)
920 {
921         pte_t *orig_src_pte, *orig_dst_pte;
922         pte_t *src_pte, *dst_pte;
923         spinlock_t *src_ptl, *dst_ptl;
924         int progress = 0;
925         int rss[NR_MM_COUNTERS];
926         swp_entry_t entry = (swp_entry_t){0};
927
928 again:
929         init_rss_vec(rss);
930
931         dst_pte = pte_alloc_map_lock(dst_mm, dst_pmd, addr, &dst_ptl);
932         if (!dst_pte)
933                 return -ENOMEM;
934         src_pte = pte_offset_map(src_pmd, addr);
935         src_ptl = pte_lockptr(src_mm, src_pmd);
936         spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
937         orig_src_pte = src_pte;
938         orig_dst_pte = dst_pte;
939         arch_enter_lazy_mmu_mode();
940
941         do {
942                 /*
943                  * We are holding two locks at this point - either of them
944                  * could generate latencies in another task on another CPU.
945                  */
946                 if (progress >= 32) {
947                         progress = 0;
948                         if (need_resched() ||
949                             spin_needbreak(src_ptl) || spin_needbreak(dst_ptl))
950                                 break;
951                 }
952                 if (pte_none(*src_pte)) {
953                         progress++;
954                         continue;
955                 }
956                 entry.val = copy_one_pte(dst_mm, src_mm, dst_pte, src_pte,
957                                                         vma, addr, rss);
958                 if (entry.val)
959                         break;
960                 progress += 8;
961         } while (dst_pte++, src_pte++, addr += PAGE_SIZE, addr != end);
962
963         arch_leave_lazy_mmu_mode();
964         spin_unlock(src_ptl);
965         pte_unmap(orig_src_pte);
966         add_mm_rss_vec(dst_mm, rss);
967         pte_unmap_unlock(orig_dst_pte, dst_ptl);
968         cond_resched();
969
970         if (entry.val) {
971                 if (add_swap_count_continuation(entry, GFP_KERNEL) < 0)
972                         return -ENOMEM;
973                 progress = 0;
974         }
975         if (addr != end)
976                 goto again;
977         return 0;
978 }
979
980 static inline int copy_pmd_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
981                 pud_t *dst_pud, pud_t *src_pud, struct vm_area_struct *vma,
982                 unsigned long addr, unsigned long end)
983 {
984         pmd_t *src_pmd, *dst_pmd;
985         unsigned long next;
986
987         dst_pmd = pmd_alloc(dst_mm, dst_pud, addr);
988         if (!dst_pmd)
989                 return -ENOMEM;
990         src_pmd = pmd_offset(src_pud, addr);
991         do {
992                 next = pmd_addr_end(addr, end);
993                 if (pmd_trans_huge(*src_pmd)) {
994                         int err;
995                         VM_BUG_ON(next-addr != HPAGE_PMD_SIZE);
996                         err = copy_huge_pmd(dst_mm, src_mm,
997                                             dst_pmd, src_pmd, addr, vma);
998                         if (err == -ENOMEM)
999                                 return -ENOMEM;
1000                         if (!err)
1001                                 continue;
1002                         /* fall through */
1003                 }
1004                 if (pmd_none_or_clear_bad(src_pmd))
1005                         continue;
1006                 if (copy_pte_range(dst_mm, src_mm, dst_pmd, src_pmd,
1007                                                 vma, addr, next))
1008                         return -ENOMEM;
1009         } while (dst_pmd++, src_pmd++, addr = next, addr != end);
1010         return 0;
1011 }
1012
1013 static inline int copy_pud_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
1014                 pgd_t *dst_pgd, pgd_t *src_pgd, struct vm_area_struct *vma,
1015                 unsigned long addr, unsigned long end)
1016 {
1017         pud_t *src_pud, *dst_pud;
1018         unsigned long next;
1019
1020         dst_pud = pud_alloc(dst_mm, dst_pgd, addr);
1021         if (!dst_pud)
1022                 return -ENOMEM;
1023         src_pud = pud_offset(src_pgd, addr);
1024         do {
1025                 next = pud_addr_end(addr, end);
1026                 if (pud_none_or_clear_bad(src_pud))
1027                         continue;
1028                 if (copy_pmd_range(dst_mm, src_mm, dst_pud, src_pud,
1029                                                 vma, addr, next))
1030                         return -ENOMEM;
1031         } while (dst_pud++, src_pud++, addr = next, addr != end);
1032         return 0;
1033 }
1034
1035 int copy_page_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
1036                 struct vm_area_struct *vma)
1037 {
1038         pgd_t *src_pgd, *dst_pgd;
1039         unsigned long next;
1040         unsigned long addr = vma->vm_start;
1041         unsigned long end = vma->vm_end;
1042         unsigned long mmun_start;       /* For mmu_notifiers */
1043         unsigned long mmun_end;         /* For mmu_notifiers */
1044         bool is_cow;
1045         int ret;
1046
1047         /*
1048          * Don't copy ptes where a page fault will fill them correctly.
1049          * Fork becomes much lighter when there are big shared or private
1050          * readonly mappings. The tradeoff is that copy_page_range is more
1051          * efficient than faulting.
1052          */
1053         if (!(vma->vm_flags & (VM_HUGETLB | VM_NONLINEAR |
1054                                VM_PFNMAP | VM_MIXEDMAP))) {
1055                 if (!vma->anon_vma)
1056                         return 0;
1057         }
1058
1059         if (is_vm_hugetlb_page(vma))
1060                 return copy_hugetlb_page_range(dst_mm, src_mm, vma);
1061
1062         if (unlikely(vma->vm_flags & VM_PFNMAP)) {
1063                 /*
1064                  * We do not free on error cases below as remove_vma
1065                  * gets called on error from higher level routine
1066                  */
1067                 ret = track_pfn_copy(vma);
1068                 if (ret)
1069                         return ret;
1070         }
1071
1072         /*
1073          * We need to invalidate the secondary MMU mappings only when
1074          * there could be a permission downgrade on the ptes of the
1075          * parent mm. And a permission downgrade will only happen if
1076          * is_cow_mapping() returns true.
1077          */
1078         is_cow = is_cow_mapping(vma->vm_flags);
1079         mmun_start = addr;
1080         mmun_end   = end;
1081         if (is_cow)
1082                 mmu_notifier_invalidate_range_start(src_mm, mmun_start,
1083                                                     mmun_end);
1084
1085         ret = 0;
1086         dst_pgd = pgd_offset(dst_mm, addr);
1087         src_pgd = pgd_offset(src_mm, addr);
1088         do {
1089                 next = pgd_addr_end(addr, end);
1090                 if (pgd_none_or_clear_bad(src_pgd))
1091                         continue;
1092                 if (unlikely(copy_pud_range(dst_mm, src_mm, dst_pgd, src_pgd,
1093                                             vma, addr, next))) {
1094                         ret = -ENOMEM;
1095                         break;
1096                 }
1097         } while (dst_pgd++, src_pgd++, addr = next, addr != end);
1098
1099         if (is_cow)
1100                 mmu_notifier_invalidate_range_end(src_mm, mmun_start, mmun_end);
1101         return ret;
1102 }
1103
1104 static unsigned long zap_pte_range(struct mmu_gather *tlb,
1105                                 struct vm_area_struct *vma, pmd_t *pmd,
1106                                 unsigned long addr, unsigned long end,
1107                                 struct zap_details *details)
1108 {
1109         struct mm_struct *mm = tlb->mm;
1110         int force_flush = 0;
1111         int rss[NR_MM_COUNTERS];
1112         spinlock_t *ptl;
1113         pte_t *start_pte;
1114         pte_t *pte;
1115
1116 again:
1117         init_rss_vec(rss);
1118         start_pte = pte_offset_map_lock(mm, pmd, addr, &ptl);
1119         pte = start_pte;
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                         else if (is_migration_entry(entry)) {
1192                                 struct page *page;
1193
1194                                 page = migration_entry_to_page(entry);
1195
1196                                 if (PageAnon(page))
1197                                         rss[MM_ANONPAGES]--;
1198                                 else
1199                                         rss[MM_FILEPAGES]--;
1200                         }
1201                         if (unlikely(!free_swap_and_cache(entry)))
1202                                 print_bad_pte(vma, addr, ptent, NULL);
1203                 }
1204                 pte_clear_not_present_full(mm, addr, pte, tlb->fullmm);
1205         } while (pte++, addr += PAGE_SIZE, addr != end);
1206
1207         add_mm_rss_vec(mm, rss);
1208         arch_leave_lazy_mmu_mode();
1209         pte_unmap_unlock(start_pte, ptl);
1210
1211         /*
1212          * mmu_gather ran out of room to batch pages, we break out of
1213          * the PTE lock to avoid doing the potential expensive TLB invalidate
1214          * and page-free while holding it.
1215          */
1216         if (force_flush) {
1217                 force_flush = 0;
1218
1219 #ifdef HAVE_GENERIC_MMU_GATHER
1220                 tlb->start = addr;
1221                 tlb->end = end;
1222 #endif
1223                 tlb_flush_mmu(tlb);
1224                 if (addr != end)
1225                         goto again;
1226         }
1227
1228         return addr;
1229 }
1230
1231 static inline unsigned long zap_pmd_range(struct mmu_gather *tlb,
1232                                 struct vm_area_struct *vma, pud_t *pud,
1233                                 unsigned long addr, unsigned long end,
1234                                 struct zap_details *details)
1235 {
1236         pmd_t *pmd;
1237         unsigned long next;
1238
1239         pmd = pmd_offset(pud, addr);
1240         do {
1241                 next = pmd_addr_end(addr, end);
1242                 if (pmd_trans_huge(*pmd)) {
1243                         if (next - addr != HPAGE_PMD_SIZE) {
1244 #ifdef CONFIG_DEBUG_VM
1245                                 if (!rwsem_is_locked(&tlb->mm->mmap_sem)) {
1246                                         pr_err("%s: mmap_sem is unlocked! addr=0x%lx end=0x%lx vma->vm_start=0x%lx vma->vm_end=0x%lx\n",
1247                                                 __func__, addr, end,
1248                                                 vma->vm_start,
1249                                                 vma->vm_end);
1250                                         BUG();
1251                                 }
1252 #endif
1253                                 split_huge_page_pmd(vma->vm_mm, pmd);
1254                         } else if (zap_huge_pmd(tlb, vma, pmd, addr))
1255                                 goto next;
1256                         /* fall through */
1257                 }
1258                 /*
1259                  * Here there can be other concurrent MADV_DONTNEED or
1260                  * trans huge page faults running, and if the pmd is
1261                  * none or trans huge it can change under us. This is
1262                  * because MADV_DONTNEED holds the mmap_sem in read
1263                  * mode.
1264                  */
1265                 if (pmd_none_or_trans_huge_or_clear_bad(pmd))
1266                         goto next;
1267                 next = zap_pte_range(tlb, vma, pmd, addr, next, details);
1268 next:
1269                 cond_resched();
1270         } while (pmd++, addr = next, addr != end);
1271
1272         return addr;
1273 }
1274
1275 static inline unsigned long zap_pud_range(struct mmu_gather *tlb,
1276                                 struct vm_area_struct *vma, pgd_t *pgd,
1277                                 unsigned long addr, unsigned long end,
1278                                 struct zap_details *details)
1279 {
1280         pud_t *pud;
1281         unsigned long next;
1282
1283         pud = pud_offset(pgd, addr);
1284         do {
1285                 next = pud_addr_end(addr, end);
1286                 if (pud_none_or_clear_bad(pud))
1287                         continue;
1288                 next = zap_pmd_range(tlb, vma, pud, addr, next, details);
1289         } while (pud++, addr = next, addr != end);
1290
1291         return addr;
1292 }
1293
1294 static void unmap_page_range(struct mmu_gather *tlb,
1295                              struct vm_area_struct *vma,
1296                              unsigned long addr, unsigned long end,
1297                              struct zap_details *details)
1298 {
1299         pgd_t *pgd;
1300         unsigned long next;
1301
1302         if (details && !details->check_mapping && !details->nonlinear_vma)
1303                 details = NULL;
1304
1305         BUG_ON(addr >= end);
1306         mem_cgroup_uncharge_start();
1307         tlb_start_vma(tlb, vma);
1308         pgd = pgd_offset(vma->vm_mm, addr);
1309         do {
1310                 next = pgd_addr_end(addr, end);
1311                 if (pgd_none_or_clear_bad(pgd))
1312                         continue;
1313                 next = zap_pud_range(tlb, vma, pgd, addr, next, details);
1314         } while (pgd++, addr = next, addr != end);
1315         tlb_end_vma(tlb, vma);
1316         mem_cgroup_uncharge_end();
1317 }
1318
1319
1320 static void unmap_single_vma(struct mmu_gather *tlb,
1321                 struct vm_area_struct *vma, unsigned long start_addr,
1322                 unsigned long end_addr,
1323                 struct zap_details *details)
1324 {
1325         unsigned long start = max(vma->vm_start, start_addr);
1326         unsigned long end;
1327
1328         if (start >= vma->vm_end)
1329                 return;
1330         end = min(vma->vm_end, end_addr);
1331         if (end <= vma->vm_start)
1332                 return;
1333
1334         if (vma->vm_file)
1335                 uprobe_munmap(vma, start, end);
1336
1337         if (unlikely(vma->vm_flags & VM_PFNMAP))
1338                 untrack_pfn(vma, 0, 0);
1339
1340         if (start != end) {
1341                 if (unlikely(is_vm_hugetlb_page(vma))) {
1342                         /*
1343                          * It is undesirable to test vma->vm_file as it
1344                          * should be non-null for valid hugetlb area.
1345                          * However, vm_file will be NULL in the error
1346                          * cleanup path of do_mmap_pgoff. When
1347                          * hugetlbfs ->mmap method fails,
1348                          * do_mmap_pgoff() nullifies vma->vm_file
1349                          * before calling this function to clean up.
1350                          * Since no pte has actually been setup, it is
1351                          * safe to do nothing in this case.
1352                          */
1353                         if (vma->vm_file) {
1354                                 mutex_lock(&vma->vm_file->f_mapping->i_mmap_mutex);
1355                                 __unmap_hugepage_range_final(tlb, vma, start, end, NULL);
1356                                 mutex_unlock(&vma->vm_file->f_mapping->i_mmap_mutex);
1357                         }
1358                 } else
1359                         unmap_page_range(tlb, vma, start, end, details);
1360         }
1361 }
1362
1363 /**
1364  * unmap_vmas - unmap a range of memory covered by a list of vma's
1365  * @tlb: address of the caller's struct mmu_gather
1366  * @vma: the starting vma
1367  * @start_addr: virtual address at which to start unmapping
1368  * @end_addr: virtual address at which to end unmapping
1369  *
1370  * Unmap all pages in the vma list.
1371  *
1372  * Only addresses between `start' and `end' will be unmapped.
1373  *
1374  * The VMA list must be sorted in ascending virtual address order.
1375  *
1376  * unmap_vmas() assumes that the caller will flush the whole unmapped address
1377  * range after unmap_vmas() returns.  So the only responsibility here is to
1378  * ensure that any thus-far unmapped pages are flushed before unmap_vmas()
1379  * drops the lock and schedules.
1380  */
1381 void unmap_vmas(struct mmu_gather *tlb,
1382                 struct vm_area_struct *vma, unsigned long start_addr,
1383                 unsigned long end_addr)
1384 {
1385         struct mm_struct *mm = vma->vm_mm;
1386
1387         mmu_notifier_invalidate_range_start(mm, start_addr, end_addr);
1388         for ( ; vma && vma->vm_start < end_addr; vma = vma->vm_next)
1389                 unmap_single_vma(tlb, vma, start_addr, end_addr, NULL);
1390         mmu_notifier_invalidate_range_end(mm, start_addr, end_addr);
1391 }
1392
1393 /**
1394  * zap_page_range - remove user pages in a given range
1395  * @vma: vm_area_struct holding the applicable pages
1396  * @start: starting address of pages to zap
1397  * @size: number of bytes to zap
1398  * @details: details of nonlinear truncation or shared cache invalidation
1399  *
1400  * Caller must protect the VMA list
1401  */
1402 void zap_page_range(struct vm_area_struct *vma, unsigned long start,
1403                 unsigned long size, struct zap_details *details)
1404 {
1405         struct mm_struct *mm = vma->vm_mm;
1406         struct mmu_gather tlb;
1407         unsigned long end = start + size;
1408
1409         lru_add_drain();
1410         tlb_gather_mmu(&tlb, mm, 0);
1411         update_hiwater_rss(mm);
1412         mmu_notifier_invalidate_range_start(mm, start, end);
1413         for ( ; vma && vma->vm_start < end; vma = vma->vm_next)
1414                 unmap_single_vma(&tlb, vma, start, end, details);
1415         mmu_notifier_invalidate_range_end(mm, start, end);
1416         tlb_finish_mmu(&tlb, start, end);
1417 }
1418
1419 /**
1420  * zap_page_range_single - remove user pages in a given range
1421  * @vma: vm_area_struct holding the applicable pages
1422  * @address: starting address of pages to zap
1423  * @size: number of bytes to zap
1424  * @details: details of nonlinear truncation or shared cache invalidation
1425  *
1426  * The range must fit into one VMA.
1427  */
1428 static void zap_page_range_single(struct vm_area_struct *vma, unsigned long address,
1429                 unsigned long size, struct zap_details *details)
1430 {
1431         struct mm_struct *mm = vma->vm_mm;
1432         struct mmu_gather tlb;
1433         unsigned long end = address + size;
1434
1435         lru_add_drain();
1436         tlb_gather_mmu(&tlb, mm, 0);
1437         update_hiwater_rss(mm);
1438         mmu_notifier_invalidate_range_start(mm, address, end);
1439         unmap_single_vma(&tlb, vma, address, end, details);
1440         mmu_notifier_invalidate_range_end(mm, address, end);
1441         tlb_finish_mmu(&tlb, address, end);
1442 }
1443
1444 /**
1445  * zap_vma_ptes - remove ptes mapping the vma
1446  * @vma: vm_area_struct holding ptes to be zapped
1447  * @address: starting address of pages to zap
1448  * @size: number of bytes to zap
1449  *
1450  * This function only unmaps ptes assigned to VM_PFNMAP vmas.
1451  *
1452  * The entire address range must be fully contained within the vma.
1453  *
1454  * Returns 0 if successful.
1455  */
1456 int zap_vma_ptes(struct vm_area_struct *vma, unsigned long address,
1457                 unsigned long size)
1458 {
1459         if (address < vma->vm_start || address + size > vma->vm_end ||
1460                         !(vma->vm_flags & VM_PFNMAP))
1461                 return -1;
1462         zap_page_range_single(vma, address, size, NULL);
1463         return 0;
1464 }
1465 EXPORT_SYMBOL_GPL(zap_vma_ptes);
1466
1467 /**
1468  * follow_page - look up a page descriptor from a user-virtual address
1469  * @vma: vm_area_struct mapping @address
1470  * @address: virtual address to look up
1471  * @flags: flags modifying lookup behaviour
1472  *
1473  * @flags can have FOLL_ flags set, defined in <linux/mm.h>
1474  *
1475  * Returns the mapped (struct page *), %NULL if no mapping exists, or
1476  * an error pointer if there is a mapping to something not represented
1477  * by a page descriptor (see also vm_normal_page()).
1478  */
1479 struct page *follow_page(struct vm_area_struct *vma, unsigned long address,
1480                         unsigned int flags)
1481 {
1482         pgd_t *pgd;
1483         pud_t *pud;
1484         pmd_t *pmd;
1485         pte_t *ptep, pte;
1486         spinlock_t *ptl;
1487         struct page *page;
1488         struct mm_struct *mm = vma->vm_mm;
1489
1490         page = follow_huge_addr(mm, address, flags & FOLL_WRITE);
1491         if (!IS_ERR(page)) {
1492                 BUG_ON(flags & FOLL_GET);
1493                 goto out;
1494         }
1495
1496         page = NULL;
1497         pgd = pgd_offset(mm, address);
1498         if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
1499                 goto no_page_table;
1500
1501         pud = pud_offset(pgd, address);
1502         if (pud_none(*pud))
1503                 goto no_page_table;
1504         if (pud_huge(*pud) && vma->vm_flags & VM_HUGETLB) {
1505                 BUG_ON(flags & FOLL_GET);
1506                 page = follow_huge_pud(mm, address, pud, flags & FOLL_WRITE);
1507                 goto out;
1508         }
1509         if (unlikely(pud_bad(*pud)))
1510                 goto no_page_table;
1511
1512         pmd = pmd_offset(pud, address);
1513         if (pmd_none(*pmd))
1514                 goto no_page_table;
1515         if (pmd_huge(*pmd) && vma->vm_flags & VM_HUGETLB) {
1516                 BUG_ON(flags & FOLL_GET);
1517                 page = follow_huge_pmd(mm, address, pmd, flags & FOLL_WRITE);
1518                 goto out;
1519         }
1520         if ((flags & FOLL_NUMA) && pmd_numa(*pmd))
1521                 goto no_page_table;
1522         if (pmd_trans_huge(*pmd)) {
1523                 if (flags & FOLL_SPLIT) {
1524                         split_huge_page_pmd(mm, pmd);
1525                         goto split_fallthrough;
1526                 }
1527                 spin_lock(&mm->page_table_lock);
1528                 if (likely(pmd_trans_huge(*pmd))) {
1529                         if (unlikely(pmd_trans_splitting(*pmd))) {
1530                                 spin_unlock(&mm->page_table_lock);
1531                                 wait_split_huge_page(vma->anon_vma, pmd);
1532                         } else {
1533                                 page = follow_trans_huge_pmd(vma, address,
1534                                                              pmd, flags);
1535                                 spin_unlock(&mm->page_table_lock);
1536                                 goto out;
1537                         }
1538                 } else
1539                         spin_unlock(&mm->page_table_lock);
1540                 /* fall through */
1541         }
1542 split_fallthrough:
1543         if (unlikely(pmd_bad(*pmd)))
1544                 goto no_page_table;
1545
1546         ptep = pte_offset_map_lock(mm, pmd, address, &ptl);
1547
1548         pte = *ptep;
1549         if (!pte_present(pte))
1550                 goto no_page;
1551         if ((flags & FOLL_NUMA) && pte_numa(pte))
1552                 goto no_page;
1553         if ((flags & FOLL_WRITE) && !pte_write(pte))
1554                 goto unlock;
1555
1556         page = vm_normal_page(vma, address, pte);
1557         if (unlikely(!page)) {
1558                 if ((flags & FOLL_DUMP) ||
1559                     !is_zero_pfn(pte_pfn(pte)))
1560                         goto bad_page;
1561                 page = pte_page(pte);
1562         }
1563
1564         if (flags & FOLL_GET)
1565                 get_page_foll(page);
1566         if (flags & FOLL_TOUCH) {
1567                 if ((flags & FOLL_WRITE) &&
1568                     !pte_dirty(pte) && !PageDirty(page))
1569                         set_page_dirty(page);
1570                 /*
1571                  * pte_mkyoung() would be more correct here, but atomic care
1572                  * is needed to avoid losing the dirty bit: it is easier to use
1573                  * mark_page_accessed().
1574                  */
1575                 mark_page_accessed(page);
1576         }
1577         if ((flags & FOLL_MLOCK) && (vma->vm_flags & VM_LOCKED)) {
1578                 /*
1579                  * The preliminary mapping check is mainly to avoid the
1580                  * pointless overhead of lock_page on the ZERO_PAGE
1581                  * which might bounce very badly if there is contention.
1582                  *
1583                  * If the page is already locked, we don't need to
1584                  * handle it now - vmscan will handle it later if and
1585                  * when it attempts to reclaim the page.
1586                  */
1587                 if (page->mapping && trylock_page(page)) {
1588                         lru_add_drain();  /* push cached pages to LRU */
1589                         /*
1590                          * Because we lock page here, and migration is
1591                          * blocked by the pte's page reference, and we
1592                          * know the page is still mapped, we don't even
1593                          * need to check for file-cache page truncation.
1594                          */
1595                         mlock_vma_page(page);
1596                         unlock_page(page);
1597                 }
1598         }
1599 unlock:
1600         pte_unmap_unlock(ptep, ptl);
1601 out:
1602         return page;
1603
1604 bad_page:
1605         pte_unmap_unlock(ptep, ptl);
1606         return ERR_PTR(-EFAULT);
1607
1608 no_page:
1609         pte_unmap_unlock(ptep, ptl);
1610         if (!pte_none(pte))
1611                 return page;
1612
1613 no_page_table:
1614         /*
1615          * When core dumping an enormous anonymous area that nobody
1616          * has touched so far, we don't want to allocate unnecessary pages or
1617          * page tables.  Return error instead of NULL to skip handle_mm_fault,
1618          * then get_dump_page() will return NULL to leave a hole in the dump.
1619          * But we can only make this optimization where a hole would surely
1620          * be zero-filled if handle_mm_fault() actually did handle it.
1621          */
1622         if ((flags & FOLL_DUMP) &&
1623             (!vma->vm_ops || !vma->vm_ops->fault))
1624                 return ERR_PTR(-EFAULT);
1625         return page;
1626 }
1627
1628 static inline int stack_guard_page(struct vm_area_struct *vma, unsigned long addr)
1629 {
1630         return stack_guard_page_start(vma, addr) ||
1631                stack_guard_page_end(vma, addr+PAGE_SIZE);
1632 }
1633
1634 /**
1635  * __get_user_pages() - pin user pages in memory
1636  * @tsk:        task_struct of target task
1637  * @mm:         mm_struct of target mm
1638  * @start:      starting user address
1639  * @nr_pages:   number of pages from start to pin
1640  * @gup_flags:  flags modifying pin behaviour
1641  * @pages:      array that receives pointers to the pages pinned.
1642  *              Should be at least nr_pages long. Or NULL, if caller
1643  *              only intends to ensure the pages are faulted in.
1644  * @vmas:       array of pointers to vmas corresponding to each page.
1645  *              Or NULL if the caller does not require them.
1646  * @nonblocking: whether waiting for disk IO or mmap_sem contention
1647  *
1648  * Returns number of pages pinned. This may be fewer than the number
1649  * requested. If nr_pages is 0 or negative, returns 0. If no pages
1650  * were pinned, returns -errno. Each page returned must be released
1651  * with a put_page() call when it is finished with. vmas will only
1652  * remain valid while mmap_sem is held.
1653  *
1654  * Must be called with mmap_sem held for read or write.
1655  *
1656  * __get_user_pages walks a process's page tables and takes a reference to
1657  * each struct page that each user address corresponds to at a given
1658  * instant. That is, it takes the page that would be accessed if a user
1659  * thread accesses the given user virtual address at that instant.
1660  *
1661  * This does not guarantee that the page exists in the user mappings when
1662  * __get_user_pages returns, and there may even be a completely different
1663  * page there in some cases (eg. if mmapped pagecache has been invalidated
1664  * and subsequently re faulted). However it does guarantee that the page
1665  * won't be freed completely. And mostly callers simply care that the page
1666  * contains data that was valid *at some point in time*. Typically, an IO
1667  * or similar operation cannot guarantee anything stronger anyway because
1668  * locks can't be held over the syscall boundary.
1669  *
1670  * If @gup_flags & FOLL_WRITE == 0, the page must not be written to. If
1671  * the page is written to, set_page_dirty (or set_page_dirty_lock, as
1672  * appropriate) must be called after the page is finished with, and
1673  * before put_page is called.
1674  *
1675  * If @nonblocking != NULL, __get_user_pages will not wait for disk IO
1676  * or mmap_sem contention, and if waiting is needed to pin all pages,
1677  * *@nonblocking will be set to 0.
1678  *
1679  * In most cases, get_user_pages or get_user_pages_fast should be used
1680  * instead of __get_user_pages. __get_user_pages should be used only if
1681  * you need some special @gup_flags.
1682  */
1683 int __get_user_pages(struct task_struct *tsk, struct mm_struct *mm,
1684                      unsigned long start, int nr_pages, unsigned int gup_flags,
1685                      struct page **pages, struct vm_area_struct **vmas,
1686                      int *nonblocking)
1687 {
1688         int i;
1689         unsigned long vm_flags;
1690
1691         if (nr_pages <= 0)
1692                 return 0;
1693
1694         VM_BUG_ON(!!pages != !!(gup_flags & FOLL_GET));
1695
1696         /* 
1697          * Require read or write permissions.
1698          * If FOLL_FORCE is set, we only require the "MAY" flags.
1699          */
1700         vm_flags  = (gup_flags & FOLL_WRITE) ?
1701                         (VM_WRITE | VM_MAYWRITE) : (VM_READ | VM_MAYREAD);
1702         vm_flags &= (gup_flags & FOLL_FORCE) ?
1703                         (VM_MAYREAD | VM_MAYWRITE) : (VM_READ | VM_WRITE);
1704
1705         /*
1706          * If FOLL_FORCE and FOLL_NUMA are both set, handle_mm_fault
1707          * would be called on PROT_NONE ranges. We must never invoke
1708          * handle_mm_fault on PROT_NONE ranges or the NUMA hinting
1709          * page faults would unprotect the PROT_NONE ranges if
1710          * _PAGE_NUMA and _PAGE_PROTNONE are sharing the same pte/pmd
1711          * bitflag. So to avoid that, don't set FOLL_NUMA if
1712          * FOLL_FORCE is set.
1713          */
1714         if (!(gup_flags & FOLL_FORCE))
1715                 gup_flags |= FOLL_NUMA;
1716
1717         i = 0;
1718
1719         do {
1720                 struct vm_area_struct *vma;
1721
1722                 vma = find_extend_vma(mm, start);
1723                 if (!vma && in_gate_area(mm, start)) {
1724                         unsigned long pg = start & PAGE_MASK;
1725                         pgd_t *pgd;
1726                         pud_t *pud;
1727                         pmd_t *pmd;
1728                         pte_t *pte;
1729
1730                         /* user gate pages are read-only */
1731                         if (gup_flags & FOLL_WRITE)
1732                                 return i ? : -EFAULT;
1733                         if (pg > TASK_SIZE)
1734                                 pgd = pgd_offset_k(pg);
1735                         else
1736                                 pgd = pgd_offset_gate(mm, pg);
1737                         BUG_ON(pgd_none(*pgd));
1738                         pud = pud_offset(pgd, pg);
1739                         BUG_ON(pud_none(*pud));
1740                         pmd = pmd_offset(pud, pg);
1741                         if (pmd_none(*pmd))
1742                                 return i ? : -EFAULT;
1743                         VM_BUG_ON(pmd_trans_huge(*pmd));
1744                         pte = pte_offset_map(pmd, pg);
1745                         if (pte_none(*pte)) {
1746                                 pte_unmap(pte);
1747                                 return i ? : -EFAULT;
1748                         }
1749                         vma = get_gate_vma(mm);
1750                         if (pages) {
1751                                 struct page *page;
1752
1753                                 page = vm_normal_page(vma, start, *pte);
1754                                 if (!page) {
1755                                         if (!(gup_flags & FOLL_DUMP) &&
1756                                              is_zero_pfn(pte_pfn(*pte)))
1757                                                 page = pte_page(*pte);
1758                                         else {
1759                                                 pte_unmap(pte);
1760                                                 return i ? : -EFAULT;
1761                                         }
1762                                 }
1763                                 pages[i] = page;
1764                                 get_page(page);
1765                         }
1766                         pte_unmap(pte);
1767                         goto next_page;
1768                 }
1769
1770                 if (!vma ||
1771                     (vma->vm_flags & (VM_IO | VM_PFNMAP)) ||
1772                     !(vm_flags & vma->vm_flags))
1773                         return i ? : -EFAULT;
1774
1775                 if (is_vm_hugetlb_page(vma)) {
1776                         i = follow_hugetlb_page(mm, vma, pages, vmas,
1777                                         &start, &nr_pages, i, gup_flags);
1778                         continue;
1779                 }
1780
1781                 do {
1782                         struct page *page;
1783                         unsigned int foll_flags = gup_flags;
1784
1785                         /*
1786                          * If we have a pending SIGKILL, don't keep faulting
1787                          * pages and potentially allocating memory.
1788                          */
1789                         if (unlikely(fatal_signal_pending(current)))
1790                                 return i ? i : -ERESTARTSYS;
1791
1792                         cond_resched();
1793                         while (!(page = follow_page(vma, start, foll_flags))) {
1794                                 int ret;
1795                                 unsigned int fault_flags = 0;
1796
1797                                 /* For mlock, just skip the stack guard page. */
1798                                 if (foll_flags & FOLL_MLOCK) {
1799                                         if (stack_guard_page(vma, start))
1800                                                 goto next_page;
1801                                 }
1802                                 if (foll_flags & FOLL_WRITE)
1803                                         fault_flags |= FAULT_FLAG_WRITE;
1804                                 if (nonblocking)
1805                                         fault_flags |= FAULT_FLAG_ALLOW_RETRY;
1806                                 if (foll_flags & FOLL_NOWAIT)
1807                                         fault_flags |= (FAULT_FLAG_ALLOW_RETRY | FAULT_FLAG_RETRY_NOWAIT);
1808
1809                                 ret = handle_mm_fault(mm, vma, start,
1810                                                         fault_flags);
1811
1812                                 if (ret & VM_FAULT_ERROR) {
1813                                         if (ret & VM_FAULT_OOM)
1814                                                 return i ? i : -ENOMEM;
1815                                         if (ret & (VM_FAULT_HWPOISON |
1816                                                    VM_FAULT_HWPOISON_LARGE)) {
1817                                                 if (i)
1818                                                         return i;
1819                                                 else if (gup_flags & FOLL_HWPOISON)
1820                                                         return -EHWPOISON;
1821                                                 else
1822                                                         return -EFAULT;
1823                                         }
1824                                         if (ret & VM_FAULT_SIGBUS)
1825                                                 return i ? i : -EFAULT;
1826                                         BUG();
1827                                 }
1828
1829                                 if (tsk) {
1830                                         if (ret & VM_FAULT_MAJOR)
1831                                                 tsk->maj_flt++;
1832                                         else
1833                                                 tsk->min_flt++;
1834                                 }
1835
1836                                 if (ret & VM_FAULT_RETRY) {
1837                                         if (nonblocking)
1838                                                 *nonblocking = 0;
1839                                         return i;
1840                                 }
1841
1842                                 /*
1843                                  * The VM_FAULT_WRITE bit tells us that
1844                                  * do_wp_page has broken COW when necessary,
1845                                  * even if maybe_mkwrite decided not to set
1846                                  * pte_write. We can thus safely do subsequent
1847                                  * page lookups as if they were reads. But only
1848                                  * do so when looping for pte_write is futile:
1849                                  * in some cases userspace may also be wanting
1850                                  * to write to the gotten user page, which a
1851                                  * read fault here might prevent (a readonly
1852                                  * page might get reCOWed by userspace write).
1853                                  */
1854                                 if ((ret & VM_FAULT_WRITE) &&
1855                                     !(vma->vm_flags & VM_WRITE))
1856                                         foll_flags &= ~FOLL_WRITE;
1857
1858                                 cond_resched();
1859                         }
1860                         if (IS_ERR(page))
1861                                 return i ? i : PTR_ERR(page);
1862                         if (pages) {
1863                                 pages[i] = page;
1864
1865                                 flush_anon_page(vma, page, start);
1866                                 flush_dcache_page(page);
1867                         }
1868 next_page:
1869                         if (vmas)
1870                                 vmas[i] = vma;
1871                         i++;
1872                         start += PAGE_SIZE;
1873                         nr_pages--;
1874                 } while (nr_pages && start < vma->vm_end);
1875         } while (nr_pages);
1876         return i;
1877 }
1878 EXPORT_SYMBOL(__get_user_pages);
1879
1880 /*
1881  * fixup_user_fault() - manually resolve a user page fault
1882  * @tsk:        the task_struct to use for page fault accounting, or
1883  *              NULL if faults are not to be recorded.
1884  * @mm:         mm_struct of target mm
1885  * @address:    user address
1886  * @fault_flags:flags to pass down to handle_mm_fault()
1887  *
1888  * This is meant to be called in the specific scenario where for locking reasons
1889  * we try to access user memory in atomic context (within a pagefault_disable()
1890  * section), this returns -EFAULT, and we want to resolve the user fault before
1891  * trying again.
1892  *
1893  * Typically this is meant to be used by the futex code.
1894  *
1895  * The main difference with get_user_pages() is that this function will
1896  * unconditionally call handle_mm_fault() which will in turn perform all the
1897  * necessary SW fixup of the dirty and young bits in the PTE, while
1898  * handle_mm_fault() only guarantees to update these in the struct page.
1899  *
1900  * This is important for some architectures where those bits also gate the
1901  * access permission to the page because they are maintained in software.  On
1902  * such architectures, gup() will not be enough to make a subsequent access
1903  * succeed.
1904  *
1905  * This should be called with the mm_sem held for read.
1906  */
1907 int fixup_user_fault(struct task_struct *tsk, struct mm_struct *mm,
1908                      unsigned long address, unsigned int fault_flags)
1909 {
1910         struct vm_area_struct *vma;
1911         int ret;
1912
1913         vma = find_extend_vma(mm, address);
1914         if (!vma || address < vma->vm_start)
1915                 return -EFAULT;
1916
1917         ret = handle_mm_fault(mm, vma, address, fault_flags);
1918         if (ret & VM_FAULT_ERROR) {
1919                 if (ret & VM_FAULT_OOM)
1920                         return -ENOMEM;
1921                 if (ret & (VM_FAULT_HWPOISON | VM_FAULT_HWPOISON_LARGE))
1922                         return -EHWPOISON;
1923                 if (ret & VM_FAULT_SIGBUS)
1924                         return -EFAULT;
1925                 BUG();
1926         }
1927         if (tsk) {
1928                 if (ret & VM_FAULT_MAJOR)
1929                         tsk->maj_flt++;
1930                 else
1931                         tsk->min_flt++;
1932         }
1933         return 0;
1934 }
1935
1936 /*
1937  * get_user_pages() - pin user pages in memory
1938  * @tsk:        the task_struct to use for page fault accounting, or
1939  *              NULL if faults are not to be recorded.
1940  * @mm:         mm_struct of target mm
1941  * @start:      starting user address
1942  * @nr_pages:   number of pages from start to pin
1943  * @write:      whether pages will be written to by the caller
1944  * @force:      whether to force write access even if user mapping is
1945  *              readonly. This will result in the page being COWed even
1946  *              in MAP_SHARED mappings. You do not want this.
1947  * @pages:      array that receives pointers to the pages pinned.
1948  *              Should be at least nr_pages long. Or NULL, if caller
1949  *              only intends to ensure the pages are faulted in.
1950  * @vmas:       array of pointers to vmas corresponding to each page.
1951  *              Or NULL if the caller does not require them.
1952  *
1953  * Returns number of pages pinned. This may be fewer than the number
1954  * requested. If nr_pages is 0 or negative, returns 0. If no pages
1955  * were pinned, returns -errno. Each page returned must be released
1956  * with a put_page() call when it is finished with. vmas will only
1957  * remain valid while mmap_sem is held.
1958  *
1959  * Must be called with mmap_sem held for read or write.
1960  *
1961  * get_user_pages walks a process's page tables and takes a reference to
1962  * each struct page that each user address corresponds to at a given
1963  * instant. That is, it takes the page that would be accessed if a user
1964  * thread accesses the given user virtual address at that instant.
1965  *
1966  * This does not guarantee that the page exists in the user mappings when
1967  * get_user_pages returns, and there may even be a completely different
1968  * page there in some cases (eg. if mmapped pagecache has been invalidated
1969  * and subsequently re faulted). However it does guarantee that the page
1970  * won't be freed completely. And mostly callers simply care that the page
1971  * contains data that was valid *at some point in time*. Typically, an IO
1972  * or similar operation cannot guarantee anything stronger anyway because
1973  * locks can't be held over the syscall boundary.
1974  *
1975  * If write=0, the page must not be written to. If the page is written to,
1976  * set_page_dirty (or set_page_dirty_lock, as appropriate) must be called
1977  * after the page is finished with, and before put_page is called.
1978  *
1979  * get_user_pages is typically used for fewer-copy IO operations, to get a
1980  * handle on the memory by some means other than accesses via the user virtual
1981  * addresses. The pages may be submitted for DMA to devices or accessed via
1982  * their kernel linear mapping (via the kmap APIs). Care should be taken to
1983  * use the correct cache flushing APIs.
1984  *
1985  * See also get_user_pages_fast, for performance critical applications.
1986  */
1987 int get_user_pages(struct task_struct *tsk, struct mm_struct *mm,
1988                 unsigned long start, int nr_pages, int write, int force,
1989                 struct page **pages, struct vm_area_struct **vmas)
1990 {
1991         int flags = FOLL_TOUCH;
1992
1993         if (pages)
1994                 flags |= FOLL_GET;
1995         if (write)
1996                 flags |= FOLL_WRITE;
1997         if (force)
1998                 flags |= FOLL_FORCE;
1999
2000         return __get_user_pages(tsk, mm, start, nr_pages, flags, pages, vmas,
2001                                 NULL);
2002 }
2003 EXPORT_SYMBOL(get_user_pages);
2004
2005 /**
2006  * get_dump_page() - pin user page in memory while writing it to core dump
2007  * @addr: user address
2008  *
2009  * Returns struct page pointer of user page pinned for dump,
2010  * to be freed afterwards by page_cache_release() or put_page().
2011  *
2012  * Returns NULL on any kind of failure - a hole must then be inserted into
2013  * the corefile, to preserve alignment with its headers; and also returns
2014  * NULL wherever the ZERO_PAGE, or an anonymous pte_none, has been found -
2015  * allowing a hole to be left in the corefile to save diskspace.
2016  *
2017  * Called without mmap_sem, but after all other threads have been killed.
2018  */
2019 #ifdef CONFIG_ELF_CORE
2020 struct page *get_dump_page(unsigned long addr)
2021 {
2022         struct vm_area_struct *vma;
2023         struct page *page;
2024
2025         if (__get_user_pages(current, current->mm, addr, 1,
2026                              FOLL_FORCE | FOLL_DUMP | FOLL_GET, &page, &vma,
2027                              NULL) < 1)
2028                 return NULL;
2029         flush_cache_page(vma, addr, page_to_pfn(page));
2030         return page;
2031 }
2032 #endif /* CONFIG_ELF_CORE */
2033
2034 pte_t *__get_locked_pte(struct mm_struct *mm, unsigned long addr,
2035                         spinlock_t **ptl)
2036 {
2037         pgd_t * pgd = pgd_offset(mm, addr);
2038         pud_t * pud = pud_alloc(mm, pgd, addr);
2039         if (pud) {
2040                 pmd_t * pmd = pmd_alloc(mm, pud, addr);
2041                 if (pmd) {
2042                         VM_BUG_ON(pmd_trans_huge(*pmd));
2043                         return pte_alloc_map_lock(mm, pmd, addr, ptl);
2044                 }
2045         }
2046         return NULL;
2047 }
2048
2049 /*
2050  * This is the old fallback for page remapping.
2051  *
2052  * For historical reasons, it only allows reserved pages. Only
2053  * old drivers should use this, and they needed to mark their
2054  * pages reserved for the old functions anyway.
2055  */
2056 static int insert_page(struct vm_area_struct *vma, unsigned long addr,
2057                         struct page *page, pgprot_t prot)
2058 {
2059         struct mm_struct *mm = vma->vm_mm;
2060         int retval;
2061         pte_t *pte;
2062         spinlock_t *ptl;
2063
2064         retval = -EINVAL;
2065         if (PageAnon(page))
2066                 goto out;
2067         retval = -ENOMEM;
2068         flush_dcache_page(page);
2069         pte = get_locked_pte(mm, addr, &ptl);
2070         if (!pte)
2071                 goto out;
2072         retval = -EBUSY;
2073         if (!pte_none(*pte))
2074                 goto out_unlock;
2075
2076         /* Ok, finally just insert the thing.. */
2077         get_page(page);
2078         inc_mm_counter_fast(mm, MM_FILEPAGES);
2079         page_add_file_rmap(page);
2080         set_pte_at(mm, addr, pte, mk_pte(page, prot));
2081
2082         retval = 0;
2083         pte_unmap_unlock(pte, ptl);
2084         return retval;
2085 out_unlock:
2086         pte_unmap_unlock(pte, ptl);
2087 out:
2088         return retval;
2089 }
2090
2091 /**
2092  * vm_insert_page - insert single page into user vma
2093  * @vma: user vma to map to
2094  * @addr: target user address of this page
2095  * @page: source kernel page
2096  *
2097  * This allows drivers to insert individual pages they've allocated
2098  * into a user vma.
2099  *
2100  * The page has to be a nice clean _individual_ kernel allocation.
2101  * If you allocate a compound page, you need to have marked it as
2102  * such (__GFP_COMP), or manually just split the page up yourself
2103  * (see split_page()).
2104  *
2105  * NOTE! Traditionally this was done with "remap_pfn_range()" which
2106  * took an arbitrary page protection parameter. This doesn't allow
2107  * that. Your vma protection will have to be set up correctly, which
2108  * means that if you want a shared writable mapping, you'd better
2109  * ask for a shared writable mapping!
2110  *
2111  * The page does not need to be reserved.
2112  *
2113  * Usually this function is called from f_op->mmap() handler
2114  * under mm->mmap_sem write-lock, so it can change vma->vm_flags.
2115  * Caller must set VM_MIXEDMAP on vma if it wants to call this
2116  * function from other places, for example from page-fault handler.
2117  */
2118 int vm_insert_page(struct vm_area_struct *vma, unsigned long addr,
2119                         struct page *page)
2120 {
2121         if (addr < vma->vm_start || addr >= vma->vm_end)
2122                 return -EFAULT;
2123         if (!page_count(page))
2124                 return -EINVAL;
2125         if (!(vma->vm_flags & VM_MIXEDMAP)) {
2126                 BUG_ON(down_read_trylock(&vma->vm_mm->mmap_sem));
2127                 BUG_ON(vma->vm_flags & VM_PFNMAP);
2128                 vma->vm_flags |= VM_MIXEDMAP;
2129         }
2130         return insert_page(vma, addr, page, vma->vm_page_prot);
2131 }
2132 EXPORT_SYMBOL(vm_insert_page);
2133
2134 static int insert_pfn(struct vm_area_struct *vma, unsigned long addr,
2135                         unsigned long pfn, pgprot_t prot)
2136 {
2137         struct mm_struct *mm = vma->vm_mm;
2138         int retval;
2139         pte_t *pte, entry;
2140         spinlock_t *ptl;
2141
2142         retval = -ENOMEM;
2143         pte = get_locked_pte(mm, addr, &ptl);
2144         if (!pte)
2145                 goto out;
2146         retval = -EBUSY;
2147         if (!pte_none(*pte))
2148                 goto out_unlock;
2149
2150         /* Ok, finally just insert the thing.. */
2151         entry = pte_mkspecial(pfn_pte(pfn, prot));
2152         set_pte_at(mm, addr, pte, entry);
2153         update_mmu_cache(vma, addr, pte); /* XXX: why not for insert_page? */
2154
2155         retval = 0;
2156 out_unlock:
2157         pte_unmap_unlock(pte, ptl);
2158 out:
2159         return retval;
2160 }
2161
2162 /**
2163  * vm_insert_pfn - insert single pfn into user vma
2164  * @vma: user vma to map to
2165  * @addr: target user address of this page
2166  * @pfn: source kernel pfn
2167  *
2168  * Similar to vm_insert_page, this allows drivers to insert individual pages
2169  * they've allocated into a user vma. Same comments apply.
2170  *
2171  * This function should only be called from a vm_ops->fault handler, and
2172  * in that case the handler should return NULL.
2173  *
2174  * vma cannot be a COW mapping.
2175  *
2176  * As this is called only for pages that do not currently exist, we
2177  * do not need to flush old virtual caches or the TLB.
2178  */
2179 int vm_insert_pfn(struct vm_area_struct *vma, unsigned long addr,
2180                         unsigned long pfn)
2181 {
2182         int ret;
2183         pgprot_t pgprot = vma->vm_page_prot;
2184         /*
2185          * Technically, architectures with pte_special can avoid all these
2186          * restrictions (same for remap_pfn_range).  However we would like
2187          * consistency in testing and feature parity among all, so we should
2188          * try to keep these invariants in place for everybody.
2189          */
2190         BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)));
2191         BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) ==
2192                                                 (VM_PFNMAP|VM_MIXEDMAP));
2193         BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags));
2194         BUG_ON((vma->vm_flags & VM_MIXEDMAP) && pfn_valid(pfn));
2195
2196         if (addr < vma->vm_start || addr >= vma->vm_end)
2197                 return -EFAULT;
2198         if (track_pfn_insert(vma, &pgprot, pfn))
2199                 return -EINVAL;
2200
2201         ret = insert_pfn(vma, addr, pfn, pgprot);
2202
2203         return ret;
2204 }
2205 EXPORT_SYMBOL(vm_insert_pfn);
2206
2207 int vm_insert_mixed(struct vm_area_struct *vma, unsigned long addr,
2208                         unsigned long pfn)
2209 {
2210         BUG_ON(!(vma->vm_flags & VM_MIXEDMAP));
2211
2212         if (addr < vma->vm_start || addr >= vma->vm_end)
2213                 return -EFAULT;
2214
2215         /*
2216          * If we don't have pte special, then we have to use the pfn_valid()
2217          * based VM_MIXEDMAP scheme (see vm_normal_page), and thus we *must*
2218          * refcount the page if pfn_valid is true (hence insert_page rather
2219          * than insert_pfn).  If a zero_pfn were inserted into a VM_MIXEDMAP
2220          * without pte special, it would there be refcounted as a normal page.
2221          */
2222         if (!HAVE_PTE_SPECIAL && pfn_valid(pfn)) {
2223                 struct page *page;
2224
2225                 page = pfn_to_page(pfn);
2226                 return insert_page(vma, addr, page, vma->vm_page_prot);
2227         }
2228         return insert_pfn(vma, addr, pfn, vma->vm_page_prot);
2229 }
2230 EXPORT_SYMBOL(vm_insert_mixed);
2231
2232 /*
2233  * maps a range of physical memory into the requested pages. the old
2234  * mappings are removed. any references to nonexistent pages results
2235  * in null mappings (currently treated as "copy-on-access")
2236  */
2237 static int remap_pte_range(struct mm_struct *mm, pmd_t *pmd,
2238                         unsigned long addr, unsigned long end,
2239                         unsigned long pfn, pgprot_t prot)
2240 {
2241         pte_t *pte;
2242         spinlock_t *ptl;
2243
2244         pte = pte_alloc_map_lock(mm, pmd, addr, &ptl);
2245         if (!pte)
2246                 return -ENOMEM;
2247         arch_enter_lazy_mmu_mode();
2248         do {
2249                 BUG_ON(!pte_none(*pte));
2250                 set_pte_at(mm, addr, pte, pte_mkspecial(pfn_pte(pfn, prot)));
2251                 pfn++;
2252         } while (pte++, addr += PAGE_SIZE, addr != end);
2253         arch_leave_lazy_mmu_mode();
2254         pte_unmap_unlock(pte - 1, ptl);
2255         return 0;
2256 }
2257
2258 static inline int remap_pmd_range(struct mm_struct *mm, pud_t *pud,
2259                         unsigned long addr, unsigned long end,
2260                         unsigned long pfn, pgprot_t prot)
2261 {
2262         pmd_t *pmd;
2263         unsigned long next;
2264
2265         pfn -= addr >> PAGE_SHIFT;
2266         pmd = pmd_alloc(mm, pud, addr);
2267         if (!pmd)
2268                 return -ENOMEM;
2269         VM_BUG_ON(pmd_trans_huge(*pmd));
2270         do {
2271                 next = pmd_addr_end(addr, end);
2272                 if (remap_pte_range(mm, pmd, addr, next,
2273                                 pfn + (addr >> PAGE_SHIFT), prot))
2274                         return -ENOMEM;
2275         } while (pmd++, addr = next, addr != end);
2276         return 0;
2277 }
2278
2279 static inline int remap_pud_range(struct mm_struct *mm, pgd_t *pgd,
2280                         unsigned long addr, unsigned long end,
2281                         unsigned long pfn, pgprot_t prot)
2282 {
2283         pud_t *pud;
2284         unsigned long next;
2285
2286         pfn -= addr >> PAGE_SHIFT;
2287         pud = pud_alloc(mm, pgd, addr);
2288         if (!pud)
2289                 return -ENOMEM;
2290         do {
2291                 next = pud_addr_end(addr, end);
2292                 if (remap_pmd_range(mm, pud, addr, next,
2293                                 pfn + (addr >> PAGE_SHIFT), prot))
2294                         return -ENOMEM;
2295         } while (pud++, addr = next, addr != end);
2296         return 0;
2297 }
2298
2299 /**
2300  * remap_pfn_range - remap kernel memory to userspace
2301  * @vma: user vma to map to
2302  * @addr: target user address to start at
2303  * @pfn: physical address of kernel memory
2304  * @size: size of map area
2305  * @prot: page protection flags for this mapping
2306  *
2307  *  Note: this is only safe if the mm semaphore is held when called.
2308  */
2309 int remap_pfn_range(struct vm_area_struct *vma, unsigned long addr,
2310                     unsigned long pfn, unsigned long size, pgprot_t prot)
2311 {
2312         pgd_t *pgd;
2313         unsigned long next;
2314         unsigned long end = addr + PAGE_ALIGN(size);
2315         struct mm_struct *mm = vma->vm_mm;
2316         int err;
2317
2318         /*
2319          * Physically remapped pages are special. Tell the
2320          * rest of the world about it:
2321          *   VM_IO tells people not to look at these pages
2322          *      (accesses can have side effects).
2323          *   VM_PFNMAP tells the core MM that the base pages are just
2324          *      raw PFN mappings, and do not have a "struct page" associated
2325          *      with them.
2326          *   VM_DONTEXPAND
2327          *      Disable vma merging and expanding with mremap().
2328          *   VM_DONTDUMP
2329          *      Omit vma from core dump, even when VM_IO turned off.
2330          *
2331          * There's a horrible special case to handle copy-on-write
2332          * behaviour that some programs depend on. We mark the "original"
2333          * un-COW'ed pages by matching them up with "vma->vm_pgoff".
2334          * See vm_normal_page() for details.
2335          */
2336         if (is_cow_mapping(vma->vm_flags)) {
2337                 if (addr != vma->vm_start || end != vma->vm_end)
2338                         return -EINVAL;
2339                 vma->vm_pgoff = pfn;
2340         }
2341
2342         err = track_pfn_remap(vma, &prot, pfn, addr, PAGE_ALIGN(size));
2343         if (err)
2344                 return -EINVAL;
2345
2346         vma->vm_flags |= VM_IO | VM_PFNMAP | VM_DONTEXPAND | VM_DONTDUMP;
2347
2348         BUG_ON(addr >= end);
2349         pfn -= addr >> PAGE_SHIFT;
2350         pgd = pgd_offset(mm, addr);
2351         flush_cache_range(vma, addr, end);
2352         do {
2353                 next = pgd_addr_end(addr, end);
2354                 err = remap_pud_range(mm, pgd, addr, next,
2355                                 pfn + (addr >> PAGE_SHIFT), prot);
2356                 if (err)
2357                         break;
2358         } while (pgd++, addr = next, addr != end);
2359
2360         if (err)
2361                 untrack_pfn(vma, pfn, PAGE_ALIGN(size));
2362
2363         return err;
2364 }
2365 EXPORT_SYMBOL(remap_pfn_range);
2366
2367 static int apply_to_pte_range(struct mm_struct *mm, pmd_t *pmd,
2368                                      unsigned long addr, unsigned long end,
2369                                      pte_fn_t fn, void *data)
2370 {
2371         pte_t *pte;
2372         int err;
2373         pgtable_t token;
2374         spinlock_t *uninitialized_var(ptl);
2375
2376         pte = (mm == &init_mm) ?
2377                 pte_alloc_kernel(pmd, addr) :
2378                 pte_alloc_map_lock(mm, pmd, addr, &ptl);
2379         if (!pte)
2380                 return -ENOMEM;
2381
2382         BUG_ON(pmd_huge(*pmd));
2383
2384         arch_enter_lazy_mmu_mode();
2385
2386         token = pmd_pgtable(*pmd);
2387
2388         do {
2389                 err = fn(pte++, token, addr, data);
2390                 if (err)
2391                         break;
2392         } while (addr += PAGE_SIZE, addr != end);
2393
2394         arch_leave_lazy_mmu_mode();
2395
2396         if (mm != &init_mm)
2397                 pte_unmap_unlock(pte-1, ptl);
2398         return err;
2399 }
2400
2401 static int apply_to_pmd_range(struct mm_struct *mm, pud_t *pud,
2402                                      unsigned long addr, unsigned long end,
2403                                      pte_fn_t fn, void *data)
2404 {
2405         pmd_t *pmd;
2406         unsigned long next;
2407         int err;
2408
2409         BUG_ON(pud_huge(*pud));
2410
2411         pmd = pmd_alloc(mm, pud, addr);
2412         if (!pmd)
2413                 return -ENOMEM;
2414         do {
2415                 next = pmd_addr_end(addr, end);
2416                 err = apply_to_pte_range(mm, pmd, addr, next, fn, data);
2417                 if (err)
2418                         break;
2419         } while (pmd++, addr = next, addr != end);
2420         return err;
2421 }
2422
2423 static int apply_to_pud_range(struct mm_struct *mm, pgd_t *pgd,
2424                                      unsigned long addr, unsigned long end,
2425                                      pte_fn_t fn, void *data)
2426 {
2427         pud_t *pud;
2428         unsigned long next;
2429         int err;
2430
2431         pud = pud_alloc(mm, pgd, addr);
2432         if (!pud)
2433                 return -ENOMEM;
2434         do {
2435                 next = pud_addr_end(addr, end);
2436                 err = apply_to_pmd_range(mm, pud, addr, next, fn, data);
2437                 if (err)
2438                         break;
2439         } while (pud++, addr = next, addr != end);
2440         return err;
2441 }
2442
2443 /*
2444  * Scan a region of virtual memory, filling in page tables as necessary
2445  * and calling a provided function on each leaf page table.
2446  */
2447 int apply_to_page_range(struct mm_struct *mm, unsigned long addr,
2448                         unsigned long size, pte_fn_t fn, void *data)
2449 {
2450         pgd_t *pgd;
2451         unsigned long next;
2452         unsigned long end = addr + size;
2453         int err;
2454
2455         BUG_ON(addr >= end);
2456         pgd = pgd_offset(mm, addr);
2457         do {
2458                 next = pgd_addr_end(addr, end);
2459                 err = apply_to_pud_range(mm, pgd, addr, next, fn, data);
2460                 if (err)
2461                         break;
2462         } while (pgd++, addr = next, addr != end);
2463
2464         return err;
2465 }
2466 EXPORT_SYMBOL_GPL(apply_to_page_range);
2467
2468 /*
2469  * handle_pte_fault chooses page fault handler according to an entry
2470  * which was read non-atomically.  Before making any commitment, on
2471  * those architectures or configurations (e.g. i386 with PAE) which
2472  * might give a mix of unmatched parts, do_swap_page and do_nonlinear_fault
2473  * must check under lock before unmapping the pte and proceeding
2474  * (but do_wp_page is only called after already making such a check;
2475  * and do_anonymous_page can safely check later on).
2476  */
2477 static inline int pte_unmap_same(struct mm_struct *mm, pmd_t *pmd,
2478                                 pte_t *page_table, pte_t orig_pte)
2479 {
2480         int same = 1;
2481 #if defined(CONFIG_SMP) || defined(CONFIG_PREEMPT)
2482         if (sizeof(pte_t) > sizeof(unsigned long)) {
2483                 spinlock_t *ptl = pte_lockptr(mm, pmd);
2484                 spin_lock(ptl);
2485                 same = pte_same(*page_table, orig_pte);
2486                 spin_unlock(ptl);
2487         }
2488 #endif
2489         pte_unmap(page_table);
2490         return same;
2491 }
2492
2493 static inline void cow_user_page(struct page *dst, struct page *src, unsigned long va, struct vm_area_struct *vma)
2494 {
2495         /*
2496          * If the source page was a PFN mapping, we don't have
2497          * a "struct page" for it. We do a best-effort copy by
2498          * just copying from the original user address. If that
2499          * fails, we just zero-fill it. Live with it.
2500          */
2501         if (unlikely(!src)) {
2502                 void *kaddr = kmap_atomic(dst);
2503                 void __user *uaddr = (void __user *)(va & PAGE_MASK);
2504
2505                 /*
2506                  * This really shouldn't fail, because the page is there
2507                  * in the page tables. But it might just be unreadable,
2508                  * in which case we just give up and fill the result with
2509                  * zeroes.
2510                  */
2511                 if (__copy_from_user_inatomic(kaddr, uaddr, PAGE_SIZE))
2512                         clear_page(kaddr);
2513                 kunmap_atomic(kaddr);
2514                 flush_dcache_page(dst);
2515         } else
2516                 copy_user_highpage(dst, src, va, vma);
2517 }
2518
2519 /*
2520  * This routine handles present pages, when users try to write
2521  * to a shared page. It is done by copying the page to a new address
2522  * and decrementing the shared-page counter for the old page.
2523  *
2524  * Note that this routine assumes that the protection checks have been
2525  * done by the caller (the low-level page fault routine in most cases).
2526  * Thus we can safely just mark it writable once we've done any necessary
2527  * COW.
2528  *
2529  * We also mark the page dirty at this point even though the page will
2530  * change only once the write actually happens. This avoids a few races,
2531  * and potentially makes it more efficient.
2532  *
2533  * We enter with non-exclusive mmap_sem (to exclude vma changes,
2534  * but allow concurrent faults), with pte both mapped and locked.
2535  * We return with mmap_sem still held, but pte unmapped and unlocked.
2536  */
2537 static int do_wp_page(struct mm_struct *mm, struct vm_area_struct *vma,
2538                 unsigned long address, pte_t *page_table, pmd_t *pmd,
2539                 spinlock_t *ptl, pte_t orig_pte)
2540         __releases(ptl)
2541 {
2542         struct page *old_page, *new_page = NULL;
2543         pte_t entry;
2544         int ret = 0;
2545         int page_mkwrite = 0;
2546         struct page *dirty_page = NULL;
2547         unsigned long mmun_start = 0;   /* For mmu_notifiers */
2548         unsigned long mmun_end = 0;     /* For mmu_notifiers */
2549
2550         old_page = vm_normal_page(vma, address, orig_pte);
2551         if (!old_page) {
2552                 /*
2553                  * VM_MIXEDMAP !pfn_valid() case
2554                  *
2555                  * We should not cow pages in a shared writeable mapping.
2556                  * Just mark the pages writable as we can't do any dirty
2557                  * accounting on raw pfn maps.
2558                  */
2559                 if ((vma->vm_flags & (VM_WRITE|VM_SHARED)) ==
2560                                      (VM_WRITE|VM_SHARED))
2561                         goto reuse;
2562                 goto gotten;
2563         }
2564
2565         /*
2566          * Take out anonymous pages first, anonymous shared vmas are
2567          * not dirty accountable.
2568          */
2569         if (PageAnon(old_page) && !PageKsm(old_page)) {
2570                 if (!trylock_page(old_page)) {
2571                         page_cache_get(old_page);
2572                         pte_unmap_unlock(page_table, ptl);
2573                         lock_page(old_page);
2574                         page_table = pte_offset_map_lock(mm, pmd, address,
2575                                                          &ptl);
2576                         if (!pte_same(*page_table, orig_pte)) {
2577                                 unlock_page(old_page);
2578                                 goto unlock;
2579                         }
2580                         page_cache_release(old_page);
2581                 }
2582                 if (reuse_swap_page(old_page)) {
2583                         /*
2584                          * The page is all ours.  Move it to our anon_vma so
2585                          * the rmap code will not search our parent or siblings.
2586                          * Protected against the rmap code by the page lock.
2587                          */
2588                         page_move_anon_rmap(old_page, vma, address);
2589                         unlock_page(old_page);
2590                         goto reuse;
2591                 }
2592                 unlock_page(old_page);
2593         } else if (unlikely((vma->vm_flags & (VM_WRITE|VM_SHARED)) ==
2594                                         (VM_WRITE|VM_SHARED))) {
2595                 /*
2596                  * Only catch write-faults on shared writable pages,
2597                  * read-only shared pages can get COWed by
2598                  * get_user_pages(.write=1, .force=1).
2599                  */
2600                 if (vma->vm_ops && vma->vm_ops->page_mkwrite) {
2601                         struct vm_fault vmf;
2602                         int tmp;
2603
2604                         vmf.virtual_address = (void __user *)(address &
2605                                                                 PAGE_MASK);
2606                         vmf.pgoff = old_page->index;
2607                         vmf.flags = FAULT_FLAG_WRITE|FAULT_FLAG_MKWRITE;
2608                         vmf.page = old_page;
2609
2610                         /*
2611                          * Notify the address space that the page is about to
2612                          * become writable so that it can prohibit this or wait
2613                          * for the page to get into an appropriate state.
2614                          *
2615                          * We do this without the lock held, so that it can
2616                          * sleep if it needs to.
2617                          */
2618                         page_cache_get(old_page);
2619                         pte_unmap_unlock(page_table, ptl);
2620
2621                         tmp = vma->vm_ops->page_mkwrite(vma, &vmf);
2622                         if (unlikely(tmp &
2623                                         (VM_FAULT_ERROR | VM_FAULT_NOPAGE))) {
2624                                 ret = tmp;
2625                                 goto unwritable_page;
2626                         }
2627                         if (unlikely(!(tmp & VM_FAULT_LOCKED))) {
2628                                 lock_page(old_page);
2629                                 if (!old_page->mapping) {
2630                                         ret = 0; /* retry the fault */
2631                                         unlock_page(old_page);
2632                                         goto unwritable_page;
2633                                 }
2634                         } else
2635                                 VM_BUG_ON(!PageLocked(old_page));
2636
2637                         /*
2638                          * Since we dropped the lock we need to revalidate
2639                          * the PTE as someone else may have changed it.  If
2640                          * they did, we just return, as we can count on the
2641                          * MMU to tell us if they didn't also make it writable.
2642                          */
2643                         page_table = pte_offset_map_lock(mm, pmd, address,
2644                                                          &ptl);
2645                         if (!pte_same(*page_table, orig_pte)) {
2646                                 unlock_page(old_page);
2647                                 goto unlock;
2648                         }
2649
2650                         page_mkwrite = 1;
2651                 }
2652                 dirty_page = old_page;
2653                 get_page(dirty_page);
2654
2655 reuse:
2656                 flush_cache_page(vma, address, pte_pfn(orig_pte));
2657                 entry = pte_mkyoung(orig_pte);
2658                 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2659                 if (ptep_set_access_flags(vma, address, page_table, entry,1))
2660                         update_mmu_cache(vma, address, page_table);
2661                 pte_unmap_unlock(page_table, ptl);
2662                 ret |= VM_FAULT_WRITE;
2663
2664                 if (!dirty_page)
2665                         return ret;
2666
2667                 /*
2668                  * Yes, Virginia, this is actually required to prevent a race
2669                  * with clear_page_dirty_for_io() from clearing the page dirty
2670                  * bit after it clear all dirty ptes, but before a racing
2671                  * do_wp_page installs a dirty pte.
2672                  *
2673                  * __do_fault is protected similarly.
2674                  */
2675                 if (!page_mkwrite) {
2676                         wait_on_page_locked(dirty_page);
2677                         set_page_dirty_balance(dirty_page, page_mkwrite);
2678                         /* file_update_time outside page_lock */
2679                         if (vma->vm_file)
2680                                 file_update_time(vma->vm_file);
2681                 }
2682                 put_page(dirty_page);
2683                 if (page_mkwrite) {
2684                         struct address_space *mapping = dirty_page->mapping;
2685
2686                         set_page_dirty(dirty_page);
2687                         unlock_page(dirty_page);
2688                         page_cache_release(dirty_page);
2689                         if (mapping)    {
2690                                 /*
2691                                  * Some device drivers do not set page.mapping
2692                                  * but still dirty their pages
2693                                  */
2694                                 balance_dirty_pages_ratelimited(mapping);
2695                         }
2696                 }
2697
2698                 return ret;
2699         }
2700
2701         /*
2702          * Ok, we need to copy. Oh, well..
2703          */
2704         page_cache_get(old_page);
2705 gotten:
2706         pte_unmap_unlock(page_table, ptl);
2707
2708         if (unlikely(anon_vma_prepare(vma)))
2709                 goto oom;
2710
2711         if (is_zero_pfn(pte_pfn(orig_pte))) {
2712                 new_page = alloc_zeroed_user_highpage_movable(vma, address);
2713                 if (!new_page)
2714                         goto oom;
2715         } else {
2716                 new_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, address);
2717                 if (!new_page)
2718                         goto oom;
2719                 cow_user_page(new_page, old_page, address, vma);
2720         }
2721         __SetPageUptodate(new_page);
2722
2723         if (mem_cgroup_newpage_charge(new_page, mm, GFP_KERNEL))
2724                 goto oom_free_new;
2725
2726         mmun_start  = address & PAGE_MASK;
2727         mmun_end    = mmun_start + PAGE_SIZE;
2728         mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
2729
2730         /*
2731          * Re-check the pte - we dropped the lock
2732          */
2733         page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2734         if (likely(pte_same(*page_table, orig_pte))) {
2735                 if (old_page) {
2736                         if (!PageAnon(old_page)) {
2737                                 dec_mm_counter_fast(mm, MM_FILEPAGES);
2738                                 inc_mm_counter_fast(mm, MM_ANONPAGES);
2739                         }
2740                 } else
2741                         inc_mm_counter_fast(mm, MM_ANONPAGES);
2742                 flush_cache_page(vma, address, pte_pfn(orig_pte));
2743                 entry = mk_pte(new_page, vma->vm_page_prot);
2744                 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2745                 /*
2746                  * Clear the pte entry and flush it first, before updating the
2747                  * pte with the new entry. This will avoid a race condition
2748                  * seen in the presence of one thread doing SMC and another
2749                  * thread doing COW.
2750                  */
2751                 ptep_clear_flush(vma, address, page_table);
2752                 page_add_new_anon_rmap(new_page, vma, address);
2753                 /*
2754                  * We call the notify macro here because, when using secondary
2755                  * mmu page tables (such as kvm shadow page tables), we want the
2756                  * new page to be mapped directly into the secondary page table.
2757                  */
2758                 set_pte_at_notify(mm, address, page_table, entry);
2759                 update_mmu_cache(vma, address, page_table);
2760                 if (old_page) {
2761                         /*
2762                          * Only after switching the pte to the new page may
2763                          * we remove the mapcount here. Otherwise another
2764                          * process may come and find the rmap count decremented
2765                          * before the pte is switched to the new page, and
2766                          * "reuse" the old page writing into it while our pte
2767                          * here still points into it and can be read by other
2768                          * threads.
2769                          *
2770                          * The critical issue is to order this
2771                          * page_remove_rmap with the ptp_clear_flush above.
2772                          * Those stores are ordered by (if nothing else,)
2773                          * the barrier present in the atomic_add_negative
2774                          * in page_remove_rmap.
2775                          *
2776                          * Then the TLB flush in ptep_clear_flush ensures that
2777                          * no process can access the old page before the
2778                          * decremented mapcount is visible. And the old page
2779                          * cannot be reused until after the decremented
2780                          * mapcount is visible. So transitively, TLBs to
2781                          * old page will be flushed before it can be reused.
2782                          */
2783                         page_remove_rmap(old_page);
2784                 }
2785
2786                 /* Free the old page.. */
2787                 new_page = old_page;
2788                 ret |= VM_FAULT_WRITE;
2789         } else
2790                 mem_cgroup_uncharge_page(new_page);
2791
2792         if (new_page)
2793                 page_cache_release(new_page);
2794 unlock:
2795         pte_unmap_unlock(page_table, ptl);
2796         if (mmun_end > mmun_start)
2797                 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
2798         if (old_page) {
2799                 /*
2800                  * Don't let another task, with possibly unlocked vma,
2801                  * keep the mlocked page.
2802                  */
2803                 if ((ret & VM_FAULT_WRITE) && (vma->vm_flags & VM_LOCKED)) {
2804                         lock_page(old_page);    /* LRU manipulation */
2805                         munlock_vma_page(old_page);
2806                         unlock_page(old_page);
2807                 }
2808                 page_cache_release(old_page);
2809         }
2810         return ret;
2811 oom_free_new:
2812         page_cache_release(new_page);
2813 oom:
2814         if (old_page) {
2815                 if (page_mkwrite) {
2816                         unlock_page(old_page);
2817                         page_cache_release(old_page);
2818                 }
2819                 page_cache_release(old_page);
2820         }
2821         return VM_FAULT_OOM;
2822
2823 unwritable_page:
2824         page_cache_release(old_page);
2825         return ret;
2826 }
2827
2828 static void unmap_mapping_range_vma(struct vm_area_struct *vma,
2829                 unsigned long start_addr, unsigned long end_addr,
2830                 struct zap_details *details)
2831 {
2832         zap_page_range_single(vma, start_addr, end_addr - start_addr, details);
2833 }
2834
2835 static inline void unmap_mapping_range_tree(struct rb_root *root,
2836                                             struct zap_details *details)
2837 {
2838         struct vm_area_struct *vma;
2839         pgoff_t vba, vea, zba, zea;
2840
2841         vma_interval_tree_foreach(vma, root,
2842                         details->first_index, details->last_index) {
2843
2844                 vba = vma->vm_pgoff;
2845                 vea = vba + ((vma->vm_end - vma->vm_start) >> PAGE_SHIFT) - 1;
2846                 /* Assume for now that PAGE_CACHE_SHIFT == PAGE_SHIFT */
2847                 zba = details->first_index;
2848                 if (zba < vba)
2849                         zba = vba;
2850                 zea = details->last_index;
2851                 if (zea > vea)
2852                         zea = vea;
2853
2854                 unmap_mapping_range_vma(vma,
2855                         ((zba - vba) << PAGE_SHIFT) + vma->vm_start,
2856                         ((zea - vba + 1) << PAGE_SHIFT) + vma->vm_start,
2857                                 details);
2858         }
2859 }
2860
2861 static inline void unmap_mapping_range_list(struct list_head *head,
2862                                             struct zap_details *details)
2863 {
2864         struct vm_area_struct *vma;
2865
2866         /*
2867          * In nonlinear VMAs there is no correspondence between virtual address
2868          * offset and file offset.  So we must perform an exhaustive search
2869          * across *all* the pages in each nonlinear VMA, not just the pages
2870          * whose virtual address lies outside the file truncation point.
2871          */
2872         list_for_each_entry(vma, head, shared.nonlinear) {
2873                 details->nonlinear_vma = vma;
2874                 unmap_mapping_range_vma(vma, vma->vm_start, vma->vm_end, details);
2875         }
2876 }
2877
2878 /**
2879  * unmap_mapping_range - unmap the portion of all mmaps in the specified address_space corresponding to the specified page range in the underlying file.
2880  * @mapping: the address space containing mmaps to be unmapped.
2881  * @holebegin: byte in first page to unmap, relative to the start of
2882  * the underlying file.  This will be rounded down to a PAGE_SIZE
2883  * boundary.  Note that this is different from truncate_pagecache(), which
2884  * must keep the partial page.  In contrast, we must get rid of
2885  * partial pages.
2886  * @holelen: size of prospective hole in bytes.  This will be rounded
2887  * up to a PAGE_SIZE boundary.  A holelen of zero truncates to the
2888  * end of the file.
2889  * @even_cows: 1 when truncating a file, unmap even private COWed pages;
2890  * but 0 when invalidating pagecache, don't throw away private data.
2891  */
2892 void unmap_mapping_range(struct address_space *mapping,
2893                 loff_t const holebegin, loff_t const holelen, int even_cows)
2894 {
2895         struct zap_details details;
2896         pgoff_t hba = holebegin >> PAGE_SHIFT;
2897         pgoff_t hlen = (holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
2898
2899         /* Check for overflow. */
2900         if (sizeof(holelen) > sizeof(hlen)) {
2901                 long long holeend =
2902                         (holebegin + holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
2903                 if (holeend & ~(long long)ULONG_MAX)
2904                         hlen = ULONG_MAX - hba + 1;
2905         }
2906
2907         details.check_mapping = even_cows? NULL: mapping;
2908         details.nonlinear_vma = NULL;
2909         details.first_index = hba;
2910         details.last_index = hba + hlen - 1;
2911         if (details.last_index < details.first_index)
2912                 details.last_index = ULONG_MAX;
2913
2914
2915         mutex_lock(&mapping->i_mmap_mutex);
2916         if (unlikely(!RB_EMPTY_ROOT(&mapping->i_mmap)))
2917                 unmap_mapping_range_tree(&mapping->i_mmap, &details);
2918         if (unlikely(!list_empty(&mapping->i_mmap_nonlinear)))
2919                 unmap_mapping_range_list(&mapping->i_mmap_nonlinear, &details);
2920         mutex_unlock(&mapping->i_mmap_mutex);
2921 }
2922 EXPORT_SYMBOL(unmap_mapping_range);
2923
2924 /*
2925  * We enter with non-exclusive mmap_sem (to exclude vma changes,
2926  * but allow concurrent faults), and pte mapped but not yet locked.
2927  * We return with mmap_sem still held, but pte unmapped and unlocked.
2928  */
2929 static int do_swap_page(struct mm_struct *mm, struct vm_area_struct *vma,
2930                 unsigned long address, pte_t *page_table, pmd_t *pmd,
2931                 unsigned int flags, pte_t orig_pte)
2932 {
2933         spinlock_t *ptl;
2934         struct page *page, *swapcache = NULL;
2935         swp_entry_t entry;
2936         pte_t pte;
2937         int locked;
2938         struct mem_cgroup *ptr;
2939         int exclusive = 0;
2940         int ret = 0;
2941
2942         if (!pte_unmap_same(mm, pmd, page_table, orig_pte))
2943                 goto out;
2944
2945         entry = pte_to_swp_entry(orig_pte);
2946         if (unlikely(non_swap_entry(entry))) {
2947                 if (is_migration_entry(entry)) {
2948                         migration_entry_wait(mm, pmd, address);
2949                 } else if (is_hwpoison_entry(entry)) {
2950                         ret = VM_FAULT_HWPOISON;
2951                 } else {
2952                         print_bad_pte(vma, address, orig_pte, NULL);
2953                         ret = VM_FAULT_SIGBUS;
2954                 }
2955                 goto out;
2956         }
2957         delayacct_set_flag(DELAYACCT_PF_SWAPIN);
2958         page = lookup_swap_cache(entry);
2959         if (!page) {
2960                 page = swapin_readahead(entry,
2961                                         GFP_HIGHUSER_MOVABLE, vma, address);
2962                 if (!page) {
2963                         /*
2964                          * Back out if somebody else faulted in this pte
2965                          * while we released the pte lock.
2966                          */
2967                         page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2968                         if (likely(pte_same(*page_table, orig_pte)))
2969                                 ret = VM_FAULT_OOM;
2970                         delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
2971                         goto unlock;
2972                 }
2973
2974                 /* Had to read the page from swap area: Major fault */
2975                 ret = VM_FAULT_MAJOR;
2976                 count_vm_event(PGMAJFAULT);
2977                 mem_cgroup_count_vm_event(mm, PGMAJFAULT);
2978         } else if (PageHWPoison(page)) {
2979                 /*
2980                  * hwpoisoned dirty swapcache pages are kept for killing
2981                  * owner processes (which may be unknown at hwpoison time)
2982                  */
2983                 ret = VM_FAULT_HWPOISON;
2984                 delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
2985                 goto out_release;
2986         }
2987
2988         locked = lock_page_or_retry(page, mm, flags);
2989
2990         delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
2991         if (!locked) {
2992                 ret |= VM_FAULT_RETRY;
2993                 goto out_release;
2994         }
2995
2996         /*
2997          * Make sure try_to_free_swap or reuse_swap_page or swapoff did not
2998          * release the swapcache from under us.  The page pin, and pte_same
2999          * test below, are not enough to exclude that.  Even if it is still
3000          * swapcache, we need to check that the page's swap has not changed.
3001          */
3002         if (unlikely(!PageSwapCache(page) || page_private(page) != entry.val))
3003                 goto out_page;
3004
3005         if (ksm_might_need_to_copy(page, vma, address)) {
3006                 swapcache = page;
3007                 page = ksm_does_need_to_copy(page, vma, address);
3008
3009                 if (unlikely(!page)) {
3010                         ret = VM_FAULT_OOM;
3011                         page = swapcache;
3012                         swapcache = NULL;
3013                         goto out_page;
3014                 }
3015         }
3016
3017         if (mem_cgroup_try_charge_swapin(mm, page, GFP_KERNEL, &ptr)) {
3018                 ret = VM_FAULT_OOM;
3019                 goto out_page;
3020         }
3021
3022         /*
3023          * Back out if somebody else already faulted in this pte.
3024          */
3025         page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
3026         if (unlikely(!pte_same(*page_table, orig_pte)))
3027                 goto out_nomap;
3028
3029         if (unlikely(!PageUptodate(page))) {
3030                 ret = VM_FAULT_SIGBUS;
3031                 goto out_nomap;
3032         }
3033
3034         /*
3035          * The page isn't present yet, go ahead with the fault.
3036          *
3037          * Be careful about the sequence of operations here.
3038          * To get its accounting right, reuse_swap_page() must be called
3039          * while the page is counted on swap but not yet in mapcount i.e.
3040          * before page_add_anon_rmap() and swap_free(); try_to_free_swap()
3041          * must be called after the swap_free(), or it will never succeed.
3042          * Because delete_from_swap_page() may be called by reuse_swap_page(),
3043          * mem_cgroup_commit_charge_swapin() may not be able to find swp_entry
3044          * in page->private. In this case, a record in swap_cgroup  is silently
3045          * discarded at swap_free().
3046          */
3047
3048         inc_mm_counter_fast(mm, MM_ANONPAGES);
3049         dec_mm_counter_fast(mm, MM_SWAPENTS);
3050         pte = mk_pte(page, vma->vm_page_prot);
3051         if ((flags & FAULT_FLAG_WRITE) && reuse_swap_page(page)) {
3052                 pte = maybe_mkwrite(pte_mkdirty(pte), vma);
3053                 flags &= ~FAULT_FLAG_WRITE;
3054                 ret |= VM_FAULT_WRITE;
3055                 exclusive = 1;
3056         }
3057         flush_icache_page(vma, page);
3058         set_pte_at(mm, address, page_table, pte);
3059         do_page_add_anon_rmap(page, vma, address, exclusive);
3060         /* It's better to call commit-charge after rmap is established */
3061         mem_cgroup_commit_charge_swapin(page, ptr);
3062
3063         swap_free(entry);
3064         if (vm_swap_full() || (vma->vm_flags & VM_LOCKED) || PageMlocked(page))
3065                 try_to_free_swap(page);
3066         unlock_page(page);
3067         if (swapcache) {
3068                 /*
3069                  * Hold the lock to avoid the swap entry to be reused
3070                  * until we take the PT lock for the pte_same() check
3071                  * (to avoid false positives from pte_same). For
3072                  * further safety release the lock after the swap_free
3073                  * so that the swap count won't change under a
3074                  * parallel locked swapcache.
3075                  */
3076                 unlock_page(swapcache);
3077                 page_cache_release(swapcache);
3078         }
3079
3080         if (flags & FAULT_FLAG_WRITE) {
3081                 ret |= do_wp_page(mm, vma, address, page_table, pmd, ptl, pte);
3082                 if (ret & VM_FAULT_ERROR)
3083                         ret &= VM_FAULT_ERROR;
3084                 goto out;
3085         }
3086
3087         /* No need to invalidate - it was non-present before */
3088         update_mmu_cache(vma, address, page_table);
3089 unlock:
3090         pte_unmap_unlock(page_table, ptl);
3091 out:
3092         return ret;
3093 out_nomap:
3094         mem_cgroup_cancel_charge_swapin(ptr);
3095         pte_unmap_unlock(page_table, ptl);
3096 out_page:
3097         unlock_page(page);
3098 out_release:
3099         page_cache_release(page);
3100         if (swapcache) {
3101                 unlock_page(swapcache);
3102                 page_cache_release(swapcache);
3103         }
3104         return ret;
3105 }
3106
3107 /*
3108  * This is like a special single-page "expand_{down|up}wards()",
3109  * except we must first make sure that 'address{-|+}PAGE_SIZE'
3110  * doesn't hit another vma.
3111  */
3112 static inline int check_stack_guard_page(struct vm_area_struct *vma, unsigned long address)
3113 {
3114         address &= PAGE_MASK;
3115         if ((vma->vm_flags & VM_GROWSDOWN) && address == vma->vm_start) {
3116                 struct vm_area_struct *prev = vma->vm_prev;
3117
3118                 /*
3119                  * Is there a mapping abutting this one below?
3120                  *
3121                  * That's only ok if it's the same stack mapping
3122                  * that has gotten split..
3123                  */
3124                 if (prev && prev->vm_end == address)
3125                         return prev->vm_flags & VM_GROWSDOWN ? 0 : -ENOMEM;
3126
3127                 expand_downwards(vma, address - PAGE_SIZE);
3128         }
3129         if ((vma->vm_flags & VM_GROWSUP) && address + PAGE_SIZE == vma->vm_end) {
3130                 struct vm_area_struct *next = vma->vm_next;
3131
3132                 /* As VM_GROWSDOWN but s/below/above/ */
3133                 if (next && next->vm_start == address + PAGE_SIZE)
3134                         return next->vm_flags & VM_GROWSUP ? 0 : -ENOMEM;
3135
3136                 expand_upwards(vma, address + PAGE_SIZE);
3137         }
3138         return 0;
3139 }
3140
3141 /*
3142  * We enter with non-exclusive mmap_sem (to exclude vma changes,
3143  * but allow concurrent faults), and pte mapped but not yet locked.
3144  * We return with mmap_sem still held, but pte unmapped and unlocked.
3145  */
3146 static int do_anonymous_page(struct mm_struct *mm, struct vm_area_struct *vma,
3147                 unsigned long address, pte_t *page_table, pmd_t *pmd,
3148                 unsigned int flags)
3149 {
3150         struct page *page;
3151         spinlock_t *ptl;
3152         pte_t entry;
3153
3154         pte_unmap(page_table);
3155
3156         /* Check if we need to add a guard page to the stack */
3157         if (check_stack_guard_page(vma, address) < 0)
3158                 return VM_FAULT_SIGBUS;
3159
3160         /* Use the zero-page for reads */
3161         if (!(flags & FAULT_FLAG_WRITE)) {
3162                 entry = pte_mkspecial(pfn_pte(my_zero_pfn(address),
3163                                                 vma->vm_page_prot));
3164                 page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
3165                 if (!pte_none(*page_table))
3166                         goto unlock;
3167                 goto setpte;
3168         }
3169
3170         /* Allocate our own private page. */
3171         if (unlikely(anon_vma_prepare(vma)))
3172                 goto oom;
3173         page = alloc_zeroed_user_highpage_movable(vma, address);
3174         if (!page)
3175                 goto oom;
3176         __SetPageUptodate(page);
3177
3178         if (mem_cgroup_newpage_charge(page, mm, GFP_KERNEL))
3179                 goto oom_free_page;
3180
3181         entry = mk_pte(page, vma->vm_page_prot);
3182         if (vma->vm_flags & VM_WRITE)
3183                 entry = pte_mkwrite(pte_mkdirty(entry));
3184
3185         page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
3186         if (!pte_none(*page_table))
3187                 goto release;
3188
3189         inc_mm_counter_fast(mm, MM_ANONPAGES);
3190         page_add_new_anon_rmap(page, vma, address);
3191 setpte:
3192         set_pte_at(mm, address, page_table, entry);
3193
3194         /* No need to invalidate - it was non-present before */
3195         update_mmu_cache(vma, address, page_table);
3196 unlock:
3197         pte_unmap_unlock(page_table, ptl);
3198         return 0;
3199 release:
3200         mem_cgroup_uncharge_page(page);
3201         page_cache_release(page);
3202         goto unlock;
3203 oom_free_page:
3204         page_cache_release(page);
3205 oom:
3206         return VM_FAULT_OOM;
3207 }
3208
3209 /*
3210  * __do_fault() tries to create a new page mapping. It aggressively
3211  * tries to share with existing pages, but makes a separate copy if
3212  * the FAULT_FLAG_WRITE is set in the flags parameter in order to avoid
3213  * the next page fault.
3214  *
3215  * As this is called only for pages that do not currently exist, we
3216  * do not need to flush old virtual caches or the TLB.
3217  *
3218  * We enter with non-exclusive mmap_sem (to exclude vma changes,
3219  * but allow concurrent faults), and pte neither mapped nor locked.
3220  * We return with mmap_sem still held, but pte unmapped and unlocked.
3221  */
3222 static int __do_fault(struct mm_struct *mm, struct vm_area_struct *vma,
3223                 unsigned long address, pmd_t *pmd,
3224                 pgoff_t pgoff, unsigned int flags, pte_t orig_pte)
3225 {
3226         pte_t *page_table;
3227         spinlock_t *ptl;
3228         struct page *page;
3229         struct page *cow_page;
3230         pte_t entry;
3231         int anon = 0;
3232         struct page *dirty_page = NULL;
3233         struct vm_fault vmf;
3234         int ret;
3235         int page_mkwrite = 0;
3236
3237         /*
3238          * If we do COW later, allocate page befor taking lock_page()
3239          * on the file cache page. This will reduce lock holding time.
3240          */
3241         if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) {
3242
3243                 if (unlikely(anon_vma_prepare(vma)))
3244                         return VM_FAULT_OOM;
3245
3246                 cow_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, address);
3247                 if (!cow_page)
3248                         return VM_FAULT_OOM;
3249
3250                 if (mem_cgroup_newpage_charge(cow_page, mm, GFP_KERNEL)) {
3251                         page_cache_release(cow_page);
3252                         return VM_FAULT_OOM;
3253                 }
3254         } else
3255                 cow_page = NULL;
3256
3257         vmf.virtual_address = (void __user *)(address & PAGE_MASK);
3258         vmf.pgoff = pgoff;
3259         vmf.flags = flags;
3260         vmf.page = NULL;
3261
3262         ret = vma->vm_ops->fault(vma, &vmf);
3263         if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE |
3264                             VM_FAULT_RETRY)))
3265                 goto uncharge_out;
3266
3267         if (unlikely(PageHWPoison(vmf.page))) {
3268                 if (ret & VM_FAULT_LOCKED)
3269                         unlock_page(vmf.page);
3270                 ret = VM_FAULT_HWPOISON;
3271                 goto uncharge_out;
3272         }
3273
3274         /*
3275          * For consistency in subsequent calls, make the faulted page always
3276          * locked.
3277          */
3278         if (unlikely(!(ret & VM_FAULT_LOCKED)))
3279                 lock_page(vmf.page);
3280         else
3281                 VM_BUG_ON(!PageLocked(vmf.page));
3282
3283         /*
3284          * Should we do an early C-O-W break?
3285          */
3286         page = vmf.page;
3287         if (flags & FAULT_FLAG_WRITE) {
3288                 if (!(vma->vm_flags & VM_SHARED)) {
3289                         page = cow_page;
3290                         anon = 1;
3291                         copy_user_highpage(page, vmf.page, address, vma);
3292                         __SetPageUptodate(page);
3293                 } else {
3294                         /*
3295                          * If the page will be shareable, see if the backing
3296                          * address space wants to know that the page is about
3297                          * to become writable
3298                          */
3299                         if (vma->vm_ops->page_mkwrite) {
3300                                 int tmp;
3301
3302                                 unlock_page(page);
3303                                 vmf.flags = FAULT_FLAG_WRITE|FAULT_FLAG_MKWRITE;
3304                                 tmp = vma->vm_ops->page_mkwrite(vma, &vmf);
3305                                 if (unlikely(tmp &
3306                                           (VM_FAULT_ERROR | VM_FAULT_NOPAGE))) {
3307                                         ret = tmp;
3308                                         goto unwritable_page;
3309                                 }
3310                                 if (unlikely(!(tmp & VM_FAULT_LOCKED))) {
3311                                         lock_page(page);
3312                                         if (!page->mapping) {
3313                                                 ret = 0; /* retry the fault */
3314                                                 unlock_page(page);
3315                                                 goto unwritable_page;
3316                                         }
3317                                 } else
3318                                         VM_BUG_ON(!PageLocked(page));
3319                                 page_mkwrite = 1;
3320                         }
3321                 }
3322
3323         }
3324
3325         page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
3326
3327         /*
3328          * This silly early PAGE_DIRTY setting removes a race
3329          * due to the bad i386 page protection. But it's valid
3330          * for other architectures too.
3331          *
3332          * Note that if FAULT_FLAG_WRITE is set, we either now have
3333          * an exclusive copy of the page, or this is a shared mapping,
3334          * so we can make it writable and dirty to avoid having to
3335          * handle that later.
3336          */
3337         /* Only go through if we didn't race with anybody else... */
3338         if (likely(pte_same(*page_table, orig_pte))) {
3339                 flush_icache_page(vma, page);
3340                 entry = mk_pte(page, vma->vm_page_prot);
3341                 if (flags & FAULT_FLAG_WRITE)
3342                         entry = maybe_mkwrite(pte_mkdirty(entry), vma);
3343                 if (anon) {
3344                         inc_mm_counter_fast(mm, MM_ANONPAGES);
3345                         page_add_new_anon_rmap(page, vma, address);
3346                 } else {
3347                         inc_mm_counter_fast(mm, MM_FILEPAGES);
3348                         page_add_file_rmap(page);
3349                         if (flags & FAULT_FLAG_WRITE) {
3350                                 dirty_page = page;
3351                                 get_page(dirty_page);
3352                         }
3353                 }
3354                 set_pte_at(mm, address, page_table, entry);
3355
3356                 /* no need to invalidate: a not-present page won't be cached */
3357                 update_mmu_cache(vma, address, page_table);
3358         } else {
3359                 if (cow_page)
3360                         mem_cgroup_uncharge_page(cow_page);
3361                 if (anon)
3362                         page_cache_release(page);
3363                 else
3364                         anon = 1; /* no anon but release faulted_page */
3365         }
3366
3367         pte_unmap_unlock(page_table, ptl);
3368
3369         if (dirty_page) {
3370                 struct address_space *mapping = page->mapping;
3371                 int dirtied = 0;
3372
3373                 if (set_page_dirty(dirty_page))
3374                         dirtied = 1;
3375                 unlock_page(dirty_page);
3376                 put_page(dirty_page);
3377                 if ((dirtied || page_mkwrite) && mapping) {
3378                         /*
3379                          * Some device drivers do not set page.mapping but still
3380                          * dirty their pages
3381                          */
3382                         balance_dirty_pages_ratelimited(mapping);
3383                 }
3384
3385                 /* file_update_time outside page_lock */
3386                 if (vma->vm_file && !page_mkwrite)
3387                         file_update_time(vma->vm_file);
3388         } else {
3389                 unlock_page(vmf.page);
3390                 if (anon)
3391                         page_cache_release(vmf.page);
3392         }
3393
3394         return ret;
3395
3396 unwritable_page:
3397         page_cache_release(page);
3398         return ret;
3399 uncharge_out:
3400         /* fs's fault handler get error */
3401         if (cow_page) {
3402                 mem_cgroup_uncharge_page(cow_page);
3403                 page_cache_release(cow_page);
3404         }
3405         return ret;
3406 }
3407
3408 static int do_linear_fault(struct mm_struct *mm, struct vm_area_struct *vma,
3409                 unsigned long address, pte_t *page_table, pmd_t *pmd,
3410                 unsigned int flags, pte_t orig_pte)
3411 {
3412         pgoff_t pgoff = (((address & PAGE_MASK)
3413                         - vma->vm_start) >> PAGE_SHIFT) + vma->vm_pgoff;
3414
3415         pte_unmap(page_table);
3416         return __do_fault(mm, vma, address, pmd, pgoff, flags, orig_pte);
3417 }
3418
3419 /*
3420  * Fault of a previously existing named mapping. Repopulate the pte
3421  * from the encoded file_pte if possible. This enables swappable
3422  * nonlinear vmas.
3423  *
3424  * We enter with non-exclusive mmap_sem (to exclude vma changes,
3425  * but allow concurrent faults), and pte mapped but not yet locked.
3426  * We return with mmap_sem still held, but pte unmapped and unlocked.
3427  */
3428 static int do_nonlinear_fault(struct mm_struct *mm, struct vm_area_struct *vma,
3429                 unsigned long address, pte_t *page_table, pmd_t *pmd,
3430                 unsigned int flags, pte_t orig_pte)
3431 {
3432         pgoff_t pgoff;
3433
3434         flags |= FAULT_FLAG_NONLINEAR;
3435
3436         if (!pte_unmap_same(mm, pmd, page_table, orig_pte))
3437                 return 0;
3438
3439         if (unlikely(!(vma->vm_flags & VM_NONLINEAR))) {
3440                 /*
3441                  * Page table corrupted: show pte and kill process.
3442                  */
3443                 print_bad_pte(vma, address, orig_pte, NULL);
3444                 return VM_FAULT_SIGBUS;
3445         }
3446
3447         pgoff = pte_to_pgoff(orig_pte);
3448         return __do_fault(mm, vma, address, pmd, pgoff, flags, orig_pte);
3449 }
3450
3451 /*
3452  * These routines also need to handle stuff like marking pages dirty
3453  * and/or accessed for architectures that don't do it in hardware (most
3454  * RISC architectures).  The early dirtying is also good on the i386.
3455  *
3456  * There is also a hook called "update_mmu_cache()" that architectures
3457  * with external mmu caches can use to update those (ie the Sparc or
3458  * PowerPC hashed page tables that act as extended TLBs).
3459  *
3460  * We enter with non-exclusive mmap_sem (to exclude vma changes,
3461  * but allow concurrent faults), and pte mapped but not yet locked.
3462  * We return with mmap_sem still held, but pte unmapped and unlocked.
3463  */
3464 int handle_pte_fault(struct mm_struct *mm,
3465                      struct vm_area_struct *vma, unsigned long address,
3466                      pte_t *pte, pmd_t *pmd, unsigned int flags)
3467 {
3468         pte_t entry;
3469         spinlock_t *ptl;
3470
3471         entry = *pte;
3472         if (!pte_present(entry)) {
3473                 if (pte_none(entry)) {
3474                         if (vma->vm_ops) {
3475                                 if (likely(vma->vm_ops->fault))
3476                                         return do_linear_fault(mm, vma, address,
3477                                                 pte, pmd, flags, entry);
3478                         }
3479                         return do_anonymous_page(mm, vma, address,
3480                                                  pte, pmd, flags);
3481                 }
3482                 if (pte_file(entry))
3483                         return do_nonlinear_fault(mm, vma, address,
3484                                         pte, pmd, flags, entry);
3485                 return do_swap_page(mm, vma, address,
3486                                         pte, pmd, flags, entry);
3487         }
3488
3489         ptl = pte_lockptr(mm, pmd);
3490         spin_lock(ptl);
3491         if (unlikely(!pte_same(*pte, entry)))
3492                 goto unlock;
3493         if (flags & FAULT_FLAG_WRITE) {
3494                 if (!pte_write(entry))
3495                         return do_wp_page(mm, vma, address,
3496                                         pte, pmd, ptl, entry);
3497                 entry = pte_mkdirty(entry);
3498         }
3499         entry = pte_mkyoung(entry);
3500         if (ptep_set_access_flags(vma, address, pte, entry, flags & FAULT_FLAG_WRITE)) {
3501                 update_mmu_cache(vma, address, pte);
3502         } else {
3503                 /*
3504                  * This is needed only for protection faults but the arch code
3505                  * is not yet telling us if this is a protection fault or not.
3506                  * This still avoids useless tlb flushes for .text page faults
3507                  * with threads.
3508                  */
3509                 if (flags & FAULT_FLAG_WRITE)
3510                         flush_tlb_fix_spurious_fault(vma, address);
3511         }
3512 unlock:
3513         pte_unmap_unlock(pte, ptl);
3514         return 0;
3515 }
3516
3517 /*
3518  * By the time we get here, we already hold the mm semaphore
3519  */
3520 int handle_mm_fault(struct mm_struct *mm, struct vm_area_struct *vma,
3521                 unsigned long address, unsigned int flags)
3522 {
3523         pgd_t *pgd;
3524         pud_t *pud;
3525         pmd_t *pmd;
3526         pte_t *pte;
3527
3528         __set_current_state(TASK_RUNNING);
3529
3530         count_vm_event(PGFAULT);
3531         mem_cgroup_count_vm_event(mm, PGFAULT);
3532
3533         /* do counter updates before entering really critical section. */
3534         check_sync_rss_stat(current);
3535
3536         if (unlikely(is_vm_hugetlb_page(vma)))
3537                 return hugetlb_fault(mm, vma, address, flags);
3538
3539 retry:
3540         pgd = pgd_offset(mm, address);
3541         pud = pud_alloc(mm, pgd, address);
3542         if (!pud)
3543                 return VM_FAULT_OOM;
3544         pmd = pmd_alloc(mm, pud, address);
3545         if (!pmd)
3546                 return VM_FAULT_OOM;
3547         if (pmd_none(*pmd) && transparent_hugepage_enabled(vma)) {
3548                 if (!vma->vm_ops)
3549                         return do_huge_pmd_anonymous_page(mm, vma, address,
3550                                                           pmd, flags);
3551         } else {
3552                 pmd_t orig_pmd = *pmd;
3553                 int ret;
3554
3555                 barrier();
3556                 if (pmd_trans_huge(orig_pmd)) {
3557                         if (flags & FAULT_FLAG_WRITE &&
3558                             !pmd_write(orig_pmd) &&
3559                             !pmd_trans_splitting(orig_pmd)) {
3560                                 ret = do_huge_pmd_wp_page(mm, vma, address, pmd,
3561                                                           orig_pmd);
3562                                 /*
3563                                  * If COW results in an oom, the huge pmd will
3564                                  * have been split, so retry the fault on the
3565                                  * pte for a smaller charge.
3566                                  */
3567                                 if (unlikely(ret & VM_FAULT_OOM))
3568                                         goto retry;
3569                                 return ret;
3570                         }
3571                         return 0;
3572                 }
3573         }
3574
3575         /*
3576          * Use __pte_alloc instead of pte_alloc_map, because we can't
3577          * run pte_offset_map on the pmd, if an huge pmd could
3578          * materialize from under us from a different thread.
3579          */
3580         if (unlikely(pmd_none(*pmd)) &&
3581             unlikely(__pte_alloc(mm, vma, pmd, address)))
3582                 return VM_FAULT_OOM;
3583         /* if an huge pmd materialized from under us just retry later */
3584         if (unlikely(pmd_trans_huge(*pmd)))
3585                 return 0;
3586         /*
3587          * A regular pmd is established and it can't morph into a huge pmd
3588          * from under us anymore at this point because we hold the mmap_sem
3589          * read mode and khugepaged takes it in write mode. So now it's
3590          * safe to run pte_offset_map().
3591          */
3592         pte = pte_offset_map(pmd, address);
3593
3594         return handle_pte_fault(mm, vma, address, pte, pmd, flags);
3595 }
3596
3597 #ifndef __PAGETABLE_PUD_FOLDED
3598 /*
3599  * Allocate page upper directory.
3600  * We've already handled the fast-path in-line.
3601  */
3602 int __pud_alloc(struct mm_struct *mm, pgd_t *pgd, unsigned long address)
3603 {
3604         pud_t *new = pud_alloc_one(mm, address);
3605         if (!new)
3606                 return -ENOMEM;
3607
3608         smp_wmb(); /* See comment in __pte_alloc */
3609
3610         spin_lock(&mm->page_table_lock);
3611         if (pgd_present(*pgd))          /* Another has populated it */
3612                 pud_free(mm, new);
3613         else
3614                 pgd_populate(mm, pgd, new);
3615         spin_unlock(&mm->page_table_lock);
3616         return 0;
3617 }
3618 #endif /* __PAGETABLE_PUD_FOLDED */
3619
3620 #ifndef __PAGETABLE_PMD_FOLDED
3621 /*
3622  * Allocate page middle directory.
3623  * We've already handled the fast-path in-line.
3624  */
3625 int __pmd_alloc(struct mm_struct *mm, pud_t *pud, unsigned long address)
3626 {
3627         pmd_t *new = pmd_alloc_one(mm, address);
3628         if (!new)
3629                 return -ENOMEM;
3630
3631         smp_wmb(); /* See comment in __pte_alloc */
3632
3633         spin_lock(&mm->page_table_lock);
3634 #ifndef __ARCH_HAS_4LEVEL_HACK
3635         if (pud_present(*pud))          /* Another has populated it */
3636                 pmd_free(mm, new);
3637         else
3638                 pud_populate(mm, pud, new);
3639 #else
3640         if (pgd_present(*pud))          /* Another has populated it */
3641                 pmd_free(mm, new);
3642         else
3643                 pgd_populate(mm, pud, new);
3644 #endif /* __ARCH_HAS_4LEVEL_HACK */
3645         spin_unlock(&mm->page_table_lock);
3646         return 0;
3647 }
3648 #endif /* __PAGETABLE_PMD_FOLDED */
3649
3650 int make_pages_present(unsigned long addr, unsigned long end)
3651 {
3652         int ret, len, write;
3653         struct vm_area_struct * vma;
3654
3655         vma = find_vma(current->mm, addr);
3656         if (!vma)
3657                 return -ENOMEM;
3658         /*
3659          * We want to touch writable mappings with a write fault in order
3660          * to break COW, except for shared mappings because these don't COW
3661          * and we would not want to dirty them for nothing.
3662          */
3663         write = (vma->vm_flags & (VM_WRITE | VM_SHARED)) == VM_WRITE;
3664         BUG_ON(addr >= end);
3665         BUG_ON(end > vma->vm_end);
3666         len = DIV_ROUND_UP(end, PAGE_SIZE) - addr/PAGE_SIZE;
3667         ret = get_user_pages(current, current->mm, addr,
3668                         len, write, 0, NULL, NULL);
3669         if (ret < 0)
3670                 return ret;
3671         return ret == len ? 0 : -EFAULT;
3672 }
3673
3674 #if !defined(__HAVE_ARCH_GATE_AREA)
3675
3676 #if defined(AT_SYSINFO_EHDR)
3677 static struct vm_area_struct gate_vma;
3678
3679 static int __init gate_vma_init(void)
3680 {
3681         gate_vma.vm_mm = NULL;
3682         gate_vma.vm_start = FIXADDR_USER_START;
3683         gate_vma.vm_end = FIXADDR_USER_END;
3684         gate_vma.vm_flags = VM_READ | VM_MAYREAD | VM_EXEC | VM_MAYEXEC;
3685         gate_vma.vm_page_prot = __P101;
3686
3687         return 0;
3688 }
3689 __initcall(gate_vma_init);
3690 #endif
3691
3692 struct vm_area_struct *get_gate_vma(struct mm_struct *mm)
3693 {
3694 #ifdef AT_SYSINFO_EHDR
3695         return &gate_vma;
3696 #else
3697         return NULL;
3698 #endif
3699 }
3700
3701 int in_gate_area_no_mm(unsigned long addr)
3702 {
3703 #ifdef AT_SYSINFO_EHDR
3704         if ((addr >= FIXADDR_USER_START) && (addr < FIXADDR_USER_END))
3705                 return 1;
3706 #endif
3707         return 0;
3708 }
3709
3710 #endif  /* __HAVE_ARCH_GATE_AREA */
3711
3712 static int __follow_pte(struct mm_struct *mm, unsigned long address,
3713                 pte_t **ptepp, spinlock_t **ptlp)
3714 {
3715         pgd_t *pgd;
3716         pud_t *pud;
3717         pmd_t *pmd;
3718         pte_t *ptep;
3719
3720         pgd = pgd_offset(mm, address);
3721         if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
3722                 goto out;
3723
3724         pud = pud_offset(pgd, address);
3725         if (pud_none(*pud) || unlikely(pud_bad(*pud)))
3726                 goto out;
3727
3728         pmd = pmd_offset(pud, address);
3729         VM_BUG_ON(pmd_trans_huge(*pmd));
3730         if (pmd_none(*pmd) || unlikely(pmd_bad(*pmd)))
3731                 goto out;
3732
3733         /* We cannot handle huge page PFN maps. Luckily they don't exist. */
3734         if (pmd_huge(*pmd))
3735                 goto out;
3736
3737         ptep = pte_offset_map_lock(mm, pmd, address, ptlp);
3738         if (!ptep)
3739                 goto out;
3740         if (!pte_present(*ptep))
3741                 goto unlock;
3742         *ptepp = ptep;
3743         return 0;
3744 unlock:
3745         pte_unmap_unlock(ptep, *ptlp);
3746 out:
3747         return -EINVAL;
3748 }
3749
3750 static inline int follow_pte(struct mm_struct *mm, unsigned long address,
3751                              pte_t **ptepp, spinlock_t **ptlp)
3752 {
3753         int res;
3754
3755         /* (void) is needed to make gcc happy */
3756         (void) __cond_lock(*ptlp,
3757                            !(res = __follow_pte(mm, address, ptepp, ptlp)));
3758         return res;
3759 }
3760
3761 /**
3762  * follow_pfn - look up PFN at a user virtual address
3763  * @vma: memory mapping
3764  * @address: user virtual address
3765  * @pfn: location to store found PFN
3766  *
3767  * Only IO mappings and raw PFN mappings are allowed.
3768  *
3769  * Returns zero and the pfn at @pfn on success, -ve otherwise.
3770  */
3771 int follow_pfn(struct vm_area_struct *vma, unsigned long address,
3772         unsigned long *pfn)
3773 {
3774         int ret = -EINVAL;
3775         spinlock_t *ptl;
3776         pte_t *ptep;
3777
3778         if (!(vma->vm_flags & (VM_IO | VM_PFNMAP)))
3779                 return ret;
3780
3781         ret = follow_pte(vma->vm_mm, address, &ptep, &ptl);
3782         if (ret)
3783                 return ret;
3784         *pfn = pte_pfn(*ptep);
3785         pte_unmap_unlock(ptep, ptl);
3786         return 0;
3787 }
3788 EXPORT_SYMBOL(follow_pfn);
3789
3790 #ifdef CONFIG_HAVE_IOREMAP_PROT
3791 int follow_phys(struct vm_area_struct *vma,
3792                 unsigned long address, unsigned int flags,
3793                 unsigned long *prot, resource_size_t *phys)
3794 {
3795         int ret = -EINVAL;
3796         pte_t *ptep, pte;
3797         spinlock_t *ptl;
3798
3799         if (!(vma->vm_flags & (VM_IO | VM_PFNMAP)))
3800                 goto out;
3801
3802         if (follow_pte(vma->vm_mm, address, &ptep, &ptl))
3803                 goto out;
3804         pte = *ptep;
3805
3806         if ((flags & FOLL_WRITE) && !pte_write(pte))
3807                 goto unlock;
3808
3809         *prot = pgprot_val(pte_pgprot(pte));
3810         *phys = (resource_size_t)pte_pfn(pte) << PAGE_SHIFT;
3811
3812         ret = 0;
3813 unlock:
3814         pte_unmap_unlock(ptep, ptl);
3815 out:
3816         return ret;
3817 }
3818
3819 int generic_access_phys(struct vm_area_struct *vma, unsigned long addr,
3820                         void *buf, int len, int write)
3821 {
3822         resource_size_t phys_addr;
3823         unsigned long prot = 0;
3824         void __iomem *maddr;
3825         int offset = addr & (PAGE_SIZE-1);
3826
3827         if (follow_phys(vma, addr, write, &prot, &phys_addr))
3828                 return -EINVAL;
3829
3830         maddr = ioremap_prot(phys_addr, PAGE_SIZE, prot);
3831         if (write)
3832                 memcpy_toio(maddr + offset, buf, len);
3833         else
3834                 memcpy_fromio(buf, maddr + offset, len);
3835         iounmap(maddr);
3836
3837         return len;
3838 }
3839 #endif
3840
3841 /*
3842  * Access another process' address space as given in mm.  If non-NULL, use the
3843  * given task for page fault accounting.
3844  */
3845 static int __access_remote_vm(struct task_struct *tsk, struct mm_struct *mm,
3846                 unsigned long addr, void *buf, int len, int write)
3847 {
3848         struct vm_area_struct *vma;
3849         void *old_buf = buf;
3850
3851         down_read(&mm->mmap_sem);
3852         /* ignore errors, just check how much was successfully transferred */
3853         while (len) {
3854                 int bytes, ret, offset;
3855                 void *maddr;
3856                 struct page *page = NULL;
3857
3858                 ret = get_user_pages(tsk, mm, addr, 1,
3859                                 write, 1, &page, &vma);
3860                 if (ret <= 0) {
3861                         /*
3862                          * Check if this is a VM_IO | VM_PFNMAP VMA, which
3863                          * we can access using slightly different code.
3864                          */
3865 #ifdef CONFIG_HAVE_IOREMAP_PROT
3866                         vma = find_vma(mm, addr);
3867                         if (!vma || vma->vm_start > addr)
3868                                 break;
3869                         if (vma->vm_ops && vma->vm_ops->access)
3870                                 ret = vma->vm_ops->access(vma, addr, buf,
3871                                                           len, write);
3872                         if (ret <= 0)
3873 #endif
3874                                 break;
3875                         bytes = ret;
3876                 } else {
3877                         bytes = len;
3878                         offset = addr & (PAGE_SIZE-1);
3879                         if (bytes > PAGE_SIZE-offset)
3880                                 bytes = PAGE_SIZE-offset;
3881
3882                         maddr = kmap(page);
3883                         if (write) {
3884                                 copy_to_user_page(vma, page, addr,
3885                                                   maddr + offset, buf, bytes);
3886                                 set_page_dirty_lock(page);
3887                         } else {
3888                                 copy_from_user_page(vma, page, addr,
3889                                                     buf, maddr + offset, bytes);
3890                         }
3891                         kunmap(page);
3892                         page_cache_release(page);
3893                 }
3894                 len -= bytes;
3895                 buf += bytes;
3896                 addr += bytes;
3897         }
3898         up_read(&mm->mmap_sem);
3899
3900         return buf - old_buf;
3901 }
3902
3903 /**
3904  * access_remote_vm - access another process' address space
3905  * @mm:         the mm_struct of the target address space
3906  * @addr:       start address to access
3907  * @buf:        source or destination buffer
3908  * @len:        number of bytes to transfer
3909  * @write:      whether the access is a write
3910  *
3911  * The caller must hold a reference on @mm.
3912  */
3913 int access_remote_vm(struct mm_struct *mm, unsigned long addr,
3914                 void *buf, int len, int write)
3915 {
3916         return __access_remote_vm(NULL, mm, addr, buf, len, write);
3917 }
3918
3919 /*
3920  * Access another process' address space.
3921  * Source/target buffer must be kernel space,
3922  * Do not walk the page table directly, use get_user_pages
3923  */
3924 int access_process_vm(struct task_struct *tsk, unsigned long addr,
3925                 void *buf, int len, int write)
3926 {
3927         struct mm_struct *mm;
3928         int ret;
3929
3930         mm = get_task_mm(tsk);
3931         if (!mm)
3932                 return 0;
3933
3934         ret = __access_remote_vm(tsk, mm, addr, buf, len, write);
3935         mmput(mm);
3936
3937         return ret;
3938 }
3939
3940 /*
3941  * Print the name of a VMA.
3942  */
3943 void print_vma_addr(char *prefix, unsigned long ip)
3944 {
3945         struct mm_struct *mm = current->mm;
3946         struct vm_area_struct *vma;
3947
3948         /*
3949          * Do not print if we are in atomic
3950          * contexts (in exception stacks, etc.):
3951          */
3952         if (preempt_count())
3953                 return;
3954
3955         down_read(&mm->mmap_sem);
3956         vma = find_vma(mm, ip);
3957         if (vma && vma->vm_file) {
3958                 struct file *f = vma->vm_file;
3959                 char *buf = (char *)__get_free_page(GFP_KERNEL);
3960                 if (buf) {
3961                         char *p, *s;
3962
3963                         p = d_path(&f->f_path, buf, PAGE_SIZE);
3964                         if (IS_ERR(p))
3965                                 p = "?";
3966                         s = strrchr(p, '/');
3967                         if (s)
3968                                 p = s+1;
3969                         printk("%s%s[%lx+%lx]", prefix, p,
3970                                         vma->vm_start,
3971                                         vma->vm_end - vma->vm_start);
3972                         free_page((unsigned long)buf);
3973                 }
3974         }
3975         up_read(&mm->mmap_sem);
3976 }
3977
3978 #ifdef CONFIG_PROVE_LOCKING
3979 void might_fault(void)
3980 {
3981         /*
3982          * Some code (nfs/sunrpc) uses socket ops on kernel memory while
3983          * holding the mmap_sem, this is safe because kernel memory doesn't
3984          * get paged out, therefore we'll never actually fault, and the
3985          * below annotations will generate false positives.
3986          */
3987         if (segment_eq(get_fs(), KERNEL_DS))
3988                 return;
3989
3990         might_sleep();
3991         /*
3992          * it would be nicer only to annotate paths which are not under
3993          * pagefault_disable, however that requires a larger audit and
3994          * providing helpers like get_user_atomic.
3995          */
3996         if (!in_atomic() && current->mm)
3997                 might_lock_read(&current->mm->mmap_sem);
3998 }
3999 EXPORT_SYMBOL(might_fault);
4000 #endif
4001
4002 #if defined(CONFIG_TRANSPARENT_HUGEPAGE) || defined(CONFIG_HUGETLBFS)
4003 static void clear_gigantic_page(struct page *page,
4004                                 unsigned long addr,
4005                                 unsigned int pages_per_huge_page)
4006 {
4007         int i;
4008         struct page *p = page;
4009
4010         might_sleep();
4011         for (i = 0; i < pages_per_huge_page;
4012              i++, p = mem_map_next(p, page, i)) {
4013                 cond_resched();
4014                 clear_user_highpage(p, addr + i * PAGE_SIZE);
4015         }
4016 }
4017 void clear_huge_page(struct page *page,
4018                      unsigned long addr, unsigned int pages_per_huge_page)
4019 {
4020         int i;
4021
4022         if (unlikely(pages_per_huge_page > MAX_ORDER_NR_PAGES)) {
4023                 clear_gigantic_page(page, addr, pages_per_huge_page);
4024                 return;
4025         }
4026
4027         might_sleep();
4028         for (i = 0; i < pages_per_huge_page; i++) {
4029                 cond_resched();
4030                 clear_user_highpage(page + i, addr + i * PAGE_SIZE);
4031         }
4032 }
4033
4034 static void copy_user_gigantic_page(struct page *dst, struct page *src,
4035                                     unsigned long addr,
4036                                     struct vm_area_struct *vma,
4037                                     unsigned int pages_per_huge_page)
4038 {
4039         int i;
4040         struct page *dst_base = dst;
4041         struct page *src_base = src;
4042
4043         for (i = 0; i < pages_per_huge_page; ) {
4044                 cond_resched();
4045                 copy_user_highpage(dst, src, addr + i*PAGE_SIZE, vma);
4046
4047                 i++;
4048                 dst = mem_map_next(dst, dst_base, i);
4049                 src = mem_map_next(src, src_base, i);
4050         }
4051 }
4052
4053 void copy_user_huge_page(struct page *dst, struct page *src,
4054                          unsigned long addr, struct vm_area_struct *vma,
4055                          unsigned int pages_per_huge_page)
4056 {
4057         int i;
4058
4059         if (unlikely(pages_per_huge_page > MAX_ORDER_NR_PAGES)) {
4060                 copy_user_gigantic_page(dst, src, addr, vma,
4061                                         pages_per_huge_page);
4062                 return;
4063         }
4064
4065         might_sleep();
4066         for (i = 0; i < pages_per_huge_page; i++) {
4067                 cond_resched();
4068                 copy_user_highpage(dst + i, src + i, addr + i*PAGE_SIZE, vma);
4069         }
4070 }
4071 #endif /* CONFIG_TRANSPARENT_HUGEPAGE || CONFIG_HUGETLBFS */