thp: change split_huge_page_pmd() interface
[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 /*
728  * vm_normal_page -- This function gets the "struct page" associated with a pte.
729  *
730  * "Special" mappings do not wish to be associated with a "struct page" (either
731  * it doesn't exist, or it exists but they don't want to touch it). In this
732  * case, NULL is returned here. "Normal" mappings do have a struct page.
733  *
734  * There are 2 broad cases. Firstly, an architecture may define a pte_special()
735  * pte bit, in which case this function is trivial. Secondly, an architecture
736  * may not have a spare pte bit, which requires a more complicated scheme,
737  * described below.
738  *
739  * A raw VM_PFNMAP mapping (ie. one that is not COWed) is always considered a
740  * special mapping (even if there are underlying and valid "struct pages").
741  * COWed pages of a VM_PFNMAP are always normal.
742  *
743  * The way we recognize COWed pages within VM_PFNMAP mappings is through the
744  * rules set up by "remap_pfn_range()": the vma will have the VM_PFNMAP bit
745  * set, and the vm_pgoff will point to the first PFN mapped: thus every special
746  * mapping will always honor the rule
747  *
748  *      pfn_of_page == vma->vm_pgoff + ((addr - vma->vm_start) >> PAGE_SHIFT)
749  *
750  * And for normal mappings this is false.
751  *
752  * This restricts such mappings to be a linear translation from virtual address
753  * to pfn. To get around this restriction, we allow arbitrary mappings so long
754  * as the vma is not a COW mapping; in that case, we know that all ptes are
755  * special (because none can have been COWed).
756  *
757  *
758  * In order to support COW of arbitrary special mappings, we have VM_MIXEDMAP.
759  *
760  * VM_MIXEDMAP mappings can likewise contain memory with or without "struct
761  * page" backing, however the difference is that _all_ pages with a struct
762  * page (that is, those where pfn_valid is true) are refcounted and considered
763  * normal pages by the VM. The disadvantage is that pages are refcounted
764  * (which can be slower and simply not an option for some PFNMAP users). The
765  * advantage is that we don't have to follow the strict linearity rule of
766  * PFNMAP mappings in order to support COWable mappings.
767  *
768  */
769 #ifdef __HAVE_ARCH_PTE_SPECIAL
770 # define HAVE_PTE_SPECIAL 1
771 #else
772 # define HAVE_PTE_SPECIAL 0
773 #endif
774 struct page *vm_normal_page(struct vm_area_struct *vma, unsigned long addr,
775                                 pte_t pte)
776 {
777         unsigned long pfn = pte_pfn(pte);
778
779         if (HAVE_PTE_SPECIAL) {
780                 if (likely(!pte_special(pte)))
781                         goto check_pfn;
782                 if (vma->vm_flags & (VM_PFNMAP | VM_MIXEDMAP))
783                         return NULL;
784                 if (!is_zero_pfn(pfn))
785                         print_bad_pte(vma, addr, pte, NULL);
786                 return NULL;
787         }
788
789         /* !HAVE_PTE_SPECIAL case follows: */
790
791         if (unlikely(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP))) {
792                 if (vma->vm_flags & VM_MIXEDMAP) {
793                         if (!pfn_valid(pfn))
794                                 return NULL;
795                         goto out;
796                 } else {
797                         unsigned long off;
798                         off = (addr - vma->vm_start) >> PAGE_SHIFT;
799                         if (pfn == vma->vm_pgoff + off)
800                                 return NULL;
801                         if (!is_cow_mapping(vma->vm_flags))
802                                 return NULL;
803                 }
804         }
805
806         if (is_zero_pfn(pfn))
807                 return NULL;
808 check_pfn:
809         if (unlikely(pfn > highest_memmap_pfn)) {
810                 print_bad_pte(vma, addr, pte, NULL);
811                 return NULL;
812         }
813
814         /*
815          * NOTE! We still have PageReserved() pages in the page tables.
816          * eg. VDSO mappings can cause them to exist.
817          */
818 out:
819         return pfn_to_page(pfn);
820 }
821
822 /*
823  * copy one vm_area from one task to the other. Assumes the page tables
824  * already present in the new task to be cleared in the whole range
825  * covered by this vma.
826  */
827
828 static inline unsigned long
829 copy_one_pte(struct mm_struct *dst_mm, struct mm_struct *src_mm,
830                 pte_t *dst_pte, pte_t *src_pte, struct vm_area_struct *vma,
831                 unsigned long addr, int *rss)
832 {
833         unsigned long vm_flags = vma->vm_flags;
834         pte_t pte = *src_pte;
835         struct page *page;
836
837         /* pte contains position in swap or file, so copy. */
838         if (unlikely(!pte_present(pte))) {
839                 if (!pte_file(pte)) {
840                         swp_entry_t entry = pte_to_swp_entry(pte);
841
842                         if (swap_duplicate(entry) < 0)
843                                 return entry.val;
844
845                         /* make sure dst_mm is on swapoff's mmlist. */
846                         if (unlikely(list_empty(&dst_mm->mmlist))) {
847                                 spin_lock(&mmlist_lock);
848                                 if (list_empty(&dst_mm->mmlist))
849                                         list_add(&dst_mm->mmlist,
850                                                  &src_mm->mmlist);
851                                 spin_unlock(&mmlist_lock);
852                         }
853                         if (likely(!non_swap_entry(entry)))
854                                 rss[MM_SWAPENTS]++;
855                         else if (is_migration_entry(entry)) {
856                                 page = migration_entry_to_page(entry);
857
858                                 if (PageAnon(page))
859                                         rss[MM_ANONPAGES]++;
860                                 else
861                                         rss[MM_FILEPAGES]++;
862
863                                 if (is_write_migration_entry(entry) &&
864                                     is_cow_mapping(vm_flags)) {
865                                         /*
866                                          * COW mappings require pages in both
867                                          * parent and child to be set to read.
868                                          */
869                                         make_migration_entry_read(&entry);
870                                         pte = swp_entry_to_pte(entry);
871                                         set_pte_at(src_mm, addr, src_pte, pte);
872                                 }
873                         }
874                 }
875                 goto out_set_pte;
876         }
877
878         /*
879          * If it's a COW mapping, write protect it both
880          * in the parent and the child
881          */
882         if (is_cow_mapping(vm_flags)) {
883                 ptep_set_wrprotect(src_mm, addr, src_pte);
884                 pte = pte_wrprotect(pte);
885         }
886
887         /*
888          * If it's a shared mapping, mark it clean in
889          * the child
890          */
891         if (vm_flags & VM_SHARED)
892                 pte = pte_mkclean(pte);
893         pte = pte_mkold(pte);
894
895         page = vm_normal_page(vma, addr, pte);
896         if (page) {
897                 get_page(page);
898                 page_dup_rmap(page);
899                 if (PageAnon(page))
900                         rss[MM_ANONPAGES]++;
901                 else
902                         rss[MM_FILEPAGES]++;
903         }
904
905 out_set_pte:
906         set_pte_at(dst_mm, addr, dst_pte, pte);
907         return 0;
908 }
909
910 int copy_pte_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
911                    pmd_t *dst_pmd, pmd_t *src_pmd, struct vm_area_struct *vma,
912                    unsigned long addr, unsigned long end)
913 {
914         pte_t *orig_src_pte, *orig_dst_pte;
915         pte_t *src_pte, *dst_pte;
916         spinlock_t *src_ptl, *dst_ptl;
917         int progress = 0;
918         int rss[NR_MM_COUNTERS];
919         swp_entry_t entry = (swp_entry_t){0};
920
921 again:
922         init_rss_vec(rss);
923
924         dst_pte = pte_alloc_map_lock(dst_mm, dst_pmd, addr, &dst_ptl);
925         if (!dst_pte)
926                 return -ENOMEM;
927         src_pte = pte_offset_map(src_pmd, addr);
928         src_ptl = pte_lockptr(src_mm, src_pmd);
929         spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
930         orig_src_pte = src_pte;
931         orig_dst_pte = dst_pte;
932         arch_enter_lazy_mmu_mode();
933
934         do {
935                 /*
936                  * We are holding two locks at this point - either of them
937                  * could generate latencies in another task on another CPU.
938                  */
939                 if (progress >= 32) {
940                         progress = 0;
941                         if (need_resched() ||
942                             spin_needbreak(src_ptl) || spin_needbreak(dst_ptl))
943                                 break;
944                 }
945                 if (pte_none(*src_pte)) {
946                         progress++;
947                         continue;
948                 }
949                 entry.val = copy_one_pte(dst_mm, src_mm, dst_pte, src_pte,
950                                                         vma, addr, rss);
951                 if (entry.val)
952                         break;
953                 progress += 8;
954         } while (dst_pte++, src_pte++, addr += PAGE_SIZE, addr != end);
955
956         arch_leave_lazy_mmu_mode();
957         spin_unlock(src_ptl);
958         pte_unmap(orig_src_pte);
959         add_mm_rss_vec(dst_mm, rss);
960         pte_unmap_unlock(orig_dst_pte, dst_ptl);
961         cond_resched();
962
963         if (entry.val) {
964                 if (add_swap_count_continuation(entry, GFP_KERNEL) < 0)
965                         return -ENOMEM;
966                 progress = 0;
967         }
968         if (addr != end)
969                 goto again;
970         return 0;
971 }
972
973 static inline int copy_pmd_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
974                 pud_t *dst_pud, pud_t *src_pud, struct vm_area_struct *vma,
975                 unsigned long addr, unsigned long end)
976 {
977         pmd_t *src_pmd, *dst_pmd;
978         unsigned long next;
979
980         dst_pmd = pmd_alloc(dst_mm, dst_pud, addr);
981         if (!dst_pmd)
982                 return -ENOMEM;
983         src_pmd = pmd_offset(src_pud, addr);
984         do {
985                 next = pmd_addr_end(addr, end);
986                 if (pmd_trans_huge(*src_pmd)) {
987                         int err;
988                         VM_BUG_ON(next-addr != HPAGE_PMD_SIZE);
989                         err = copy_huge_pmd(dst_mm, src_mm,
990                                             dst_pmd, src_pmd, addr, vma);
991                         if (err == -ENOMEM)
992                                 return -ENOMEM;
993                         if (!err)
994                                 continue;
995                         /* fall through */
996                 }
997                 if (pmd_none_or_clear_bad(src_pmd))
998                         continue;
999                 if (copy_pte_range(dst_mm, src_mm, dst_pmd, src_pmd,
1000                                                 vma, addr, next))
1001                         return -ENOMEM;
1002         } while (dst_pmd++, src_pmd++, addr = next, addr != end);
1003         return 0;
1004 }
1005
1006 static inline int copy_pud_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
1007                 pgd_t *dst_pgd, pgd_t *src_pgd, struct vm_area_struct *vma,
1008                 unsigned long addr, unsigned long end)
1009 {
1010         pud_t *src_pud, *dst_pud;
1011         unsigned long next;
1012
1013         dst_pud = pud_alloc(dst_mm, dst_pgd, addr);
1014         if (!dst_pud)
1015                 return -ENOMEM;
1016         src_pud = pud_offset(src_pgd, addr);
1017         do {
1018                 next = pud_addr_end(addr, end);
1019                 if (pud_none_or_clear_bad(src_pud))
1020                         continue;
1021                 if (copy_pmd_range(dst_mm, src_mm, dst_pud, src_pud,
1022                                                 vma, addr, next))
1023                         return -ENOMEM;
1024         } while (dst_pud++, src_pud++, addr = next, addr != end);
1025         return 0;
1026 }
1027
1028 int copy_page_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
1029                 struct vm_area_struct *vma)
1030 {
1031         pgd_t *src_pgd, *dst_pgd;
1032         unsigned long next;
1033         unsigned long addr = vma->vm_start;
1034         unsigned long end = vma->vm_end;
1035         unsigned long mmun_start;       /* For mmu_notifiers */
1036         unsigned long mmun_end;         /* For mmu_notifiers */
1037         bool is_cow;
1038         int ret;
1039
1040         /*
1041          * Don't copy ptes where a page fault will fill them correctly.
1042          * Fork becomes much lighter when there are big shared or private
1043          * readonly mappings. The tradeoff is that copy_page_range is more
1044          * efficient than faulting.
1045          */
1046         if (!(vma->vm_flags & (VM_HUGETLB | VM_NONLINEAR |
1047                                VM_PFNMAP | VM_MIXEDMAP))) {
1048                 if (!vma->anon_vma)
1049                         return 0;
1050         }
1051
1052         if (is_vm_hugetlb_page(vma))
1053                 return copy_hugetlb_page_range(dst_mm, src_mm, vma);
1054
1055         if (unlikely(vma->vm_flags & VM_PFNMAP)) {
1056                 /*
1057                  * We do not free on error cases below as remove_vma
1058                  * gets called on error from higher level routine
1059                  */
1060                 ret = track_pfn_copy(vma);
1061                 if (ret)
1062                         return ret;
1063         }
1064
1065         /*
1066          * We need to invalidate the secondary MMU mappings only when
1067          * there could be a permission downgrade on the ptes of the
1068          * parent mm. And a permission downgrade will only happen if
1069          * is_cow_mapping() returns true.
1070          */
1071         is_cow = is_cow_mapping(vma->vm_flags);
1072         mmun_start = addr;
1073         mmun_end   = end;
1074         if (is_cow)
1075                 mmu_notifier_invalidate_range_start(src_mm, mmun_start,
1076                                                     mmun_end);
1077
1078         ret = 0;
1079         dst_pgd = pgd_offset(dst_mm, addr);
1080         src_pgd = pgd_offset(src_mm, addr);
1081         do {
1082                 next = pgd_addr_end(addr, end);
1083                 if (pgd_none_or_clear_bad(src_pgd))
1084                         continue;
1085                 if (unlikely(copy_pud_range(dst_mm, src_mm, dst_pgd, src_pgd,
1086                                             vma, addr, next))) {
1087                         ret = -ENOMEM;
1088                         break;
1089                 }
1090         } while (dst_pgd++, src_pgd++, addr = next, addr != end);
1091
1092         if (is_cow)
1093                 mmu_notifier_invalidate_range_end(src_mm, mmun_start, mmun_end);
1094         return ret;
1095 }
1096
1097 static unsigned long zap_pte_range(struct mmu_gather *tlb,
1098                                 struct vm_area_struct *vma, pmd_t *pmd,
1099                                 unsigned long addr, unsigned long end,
1100                                 struct zap_details *details)
1101 {
1102         struct mm_struct *mm = tlb->mm;
1103         int force_flush = 0;
1104         int rss[NR_MM_COUNTERS];
1105         spinlock_t *ptl;
1106         pte_t *start_pte;
1107         pte_t *pte;
1108
1109 again:
1110         init_rss_vec(rss);
1111         start_pte = pte_offset_map_lock(mm, pmd, addr, &ptl);
1112         pte = start_pte;
1113         arch_enter_lazy_mmu_mode();
1114         do {
1115                 pte_t ptent = *pte;
1116                 if (pte_none(ptent)) {
1117                         continue;
1118                 }
1119
1120                 if (pte_present(ptent)) {
1121                         struct page *page;
1122
1123                         page = vm_normal_page(vma, addr, ptent);
1124                         if (unlikely(details) && page) {
1125                                 /*
1126                                  * unmap_shared_mapping_pages() wants to
1127                                  * invalidate cache without truncating:
1128                                  * unmap shared but keep private pages.
1129                                  */
1130                                 if (details->check_mapping &&
1131                                     details->check_mapping != page->mapping)
1132                                         continue;
1133                                 /*
1134                                  * Each page->index must be checked when
1135                                  * invalidating or truncating nonlinear.
1136                                  */
1137                                 if (details->nonlinear_vma &&
1138                                     (page->index < details->first_index ||
1139                                      page->index > details->last_index))
1140                                         continue;
1141                         }
1142                         ptent = ptep_get_and_clear_full(mm, addr, pte,
1143                                                         tlb->fullmm);
1144                         tlb_remove_tlb_entry(tlb, pte, addr);
1145                         if (unlikely(!page))
1146                                 continue;
1147                         if (unlikely(details) && details->nonlinear_vma
1148                             && linear_page_index(details->nonlinear_vma,
1149                                                 addr) != page->index)
1150                                 set_pte_at(mm, addr, pte,
1151                                            pgoff_to_pte(page->index));
1152                         if (PageAnon(page))
1153                                 rss[MM_ANONPAGES]--;
1154                         else {
1155                                 if (pte_dirty(ptent))
1156                                         set_page_dirty(page);
1157                                 if (pte_young(ptent) &&
1158                                     likely(!VM_SequentialReadHint(vma)))
1159                                         mark_page_accessed(page);
1160                                 rss[MM_FILEPAGES]--;
1161                         }
1162                         page_remove_rmap(page);
1163                         if (unlikely(page_mapcount(page) < 0))
1164                                 print_bad_pte(vma, addr, ptent, page);
1165                         force_flush = !__tlb_remove_page(tlb, page);
1166                         if (force_flush)
1167                                 break;
1168                         continue;
1169                 }
1170                 /*
1171                  * If details->check_mapping, we leave swap entries;
1172                  * if details->nonlinear_vma, we leave file entries.
1173                  */
1174                 if (unlikely(details))
1175                         continue;
1176                 if (pte_file(ptent)) {
1177                         if (unlikely(!(vma->vm_flags & VM_NONLINEAR)))
1178                                 print_bad_pte(vma, addr, ptent, NULL);
1179                 } else {
1180                         swp_entry_t entry = pte_to_swp_entry(ptent);
1181
1182                         if (!non_swap_entry(entry))
1183                                 rss[MM_SWAPENTS]--;
1184                         else if (is_migration_entry(entry)) {
1185                                 struct page *page;
1186
1187                                 page = migration_entry_to_page(entry);
1188
1189                                 if (PageAnon(page))
1190                                         rss[MM_ANONPAGES]--;
1191                                 else
1192                                         rss[MM_FILEPAGES]--;
1193                         }
1194                         if (unlikely(!free_swap_and_cache(entry)))
1195                                 print_bad_pte(vma, addr, ptent, NULL);
1196                 }
1197                 pte_clear_not_present_full(mm, addr, pte, tlb->fullmm);
1198         } while (pte++, addr += PAGE_SIZE, addr != end);
1199
1200         add_mm_rss_vec(mm, rss);
1201         arch_leave_lazy_mmu_mode();
1202         pte_unmap_unlock(start_pte, ptl);
1203
1204         /*
1205          * mmu_gather ran out of room to batch pages, we break out of
1206          * the PTE lock to avoid doing the potential expensive TLB invalidate
1207          * and page-free while holding it.
1208          */
1209         if (force_flush) {
1210                 force_flush = 0;
1211
1212 #ifdef HAVE_GENERIC_MMU_GATHER
1213                 tlb->start = addr;
1214                 tlb->end = end;
1215 #endif
1216                 tlb_flush_mmu(tlb);
1217                 if (addr != end)
1218                         goto again;
1219         }
1220
1221         return addr;
1222 }
1223
1224 static inline unsigned long zap_pmd_range(struct mmu_gather *tlb,
1225                                 struct vm_area_struct *vma, pud_t *pud,
1226                                 unsigned long addr, unsigned long end,
1227                                 struct zap_details *details)
1228 {
1229         pmd_t *pmd;
1230         unsigned long next;
1231
1232         pmd = pmd_offset(pud, addr);
1233         do {
1234                 next = pmd_addr_end(addr, end);
1235                 if (pmd_trans_huge(*pmd)) {
1236                         if (next - addr != HPAGE_PMD_SIZE) {
1237 #ifdef CONFIG_DEBUG_VM
1238                                 if (!rwsem_is_locked(&tlb->mm->mmap_sem)) {
1239                                         pr_err("%s: mmap_sem is unlocked! addr=0x%lx end=0x%lx vma->vm_start=0x%lx vma->vm_end=0x%lx\n",
1240                                                 __func__, addr, end,
1241                                                 vma->vm_start,
1242                                                 vma->vm_end);
1243                                         BUG();
1244                                 }
1245 #endif
1246                                 split_huge_page_pmd(vma, addr, pmd);
1247                         } else if (zap_huge_pmd(tlb, vma, pmd, addr))
1248                                 goto next;
1249                         /* fall through */
1250                 }
1251                 /*
1252                  * Here there can be other concurrent MADV_DONTNEED or
1253                  * trans huge page faults running, and if the pmd is
1254                  * none or trans huge it can change under us. This is
1255                  * because MADV_DONTNEED holds the mmap_sem in read
1256                  * mode.
1257                  */
1258                 if (pmd_none_or_trans_huge_or_clear_bad(pmd))
1259                         goto next;
1260                 next = zap_pte_range(tlb, vma, pmd, addr, next, details);
1261 next:
1262                 cond_resched();
1263         } while (pmd++, addr = next, addr != end);
1264
1265         return addr;
1266 }
1267
1268 static inline unsigned long zap_pud_range(struct mmu_gather *tlb,
1269                                 struct vm_area_struct *vma, pgd_t *pgd,
1270                                 unsigned long addr, unsigned long end,
1271                                 struct zap_details *details)
1272 {
1273         pud_t *pud;
1274         unsigned long next;
1275
1276         pud = pud_offset(pgd, addr);
1277         do {
1278                 next = pud_addr_end(addr, end);
1279                 if (pud_none_or_clear_bad(pud))
1280                         continue;
1281                 next = zap_pmd_range(tlb, vma, pud, addr, next, details);
1282         } while (pud++, addr = next, addr != end);
1283
1284         return addr;
1285 }
1286
1287 static void unmap_page_range(struct mmu_gather *tlb,
1288                              struct vm_area_struct *vma,
1289                              unsigned long addr, unsigned long end,
1290                              struct zap_details *details)
1291 {
1292         pgd_t *pgd;
1293         unsigned long next;
1294
1295         if (details && !details->check_mapping && !details->nonlinear_vma)
1296                 details = NULL;
1297
1298         BUG_ON(addr >= end);
1299         mem_cgroup_uncharge_start();
1300         tlb_start_vma(tlb, vma);
1301         pgd = pgd_offset(vma->vm_mm, addr);
1302         do {
1303                 next = pgd_addr_end(addr, end);
1304                 if (pgd_none_or_clear_bad(pgd))
1305                         continue;
1306                 next = zap_pud_range(tlb, vma, pgd, addr, next, details);
1307         } while (pgd++, addr = next, addr != end);
1308         tlb_end_vma(tlb, vma);
1309         mem_cgroup_uncharge_end();
1310 }
1311
1312
1313 static void unmap_single_vma(struct mmu_gather *tlb,
1314                 struct vm_area_struct *vma, unsigned long start_addr,
1315                 unsigned long end_addr,
1316                 struct zap_details *details)
1317 {
1318         unsigned long start = max(vma->vm_start, start_addr);
1319         unsigned long end;
1320
1321         if (start >= vma->vm_end)
1322                 return;
1323         end = min(vma->vm_end, end_addr);
1324         if (end <= vma->vm_start)
1325                 return;
1326
1327         if (vma->vm_file)
1328                 uprobe_munmap(vma, start, end);
1329
1330         if (unlikely(vma->vm_flags & VM_PFNMAP))
1331                 untrack_pfn(vma, 0, 0);
1332
1333         if (start != end) {
1334                 if (unlikely(is_vm_hugetlb_page(vma))) {
1335                         /*
1336                          * It is undesirable to test vma->vm_file as it
1337                          * should be non-null for valid hugetlb area.
1338                          * However, vm_file will be NULL in the error
1339                          * cleanup path of do_mmap_pgoff. When
1340                          * hugetlbfs ->mmap method fails,
1341                          * do_mmap_pgoff() nullifies vma->vm_file
1342                          * before calling this function to clean up.
1343                          * Since no pte has actually been setup, it is
1344                          * safe to do nothing in this case.
1345                          */
1346                         if (vma->vm_file) {
1347                                 mutex_lock(&vma->vm_file->f_mapping->i_mmap_mutex);
1348                                 __unmap_hugepage_range_final(tlb, vma, start, end, NULL);
1349                                 mutex_unlock(&vma->vm_file->f_mapping->i_mmap_mutex);
1350                         }
1351                 } else
1352                         unmap_page_range(tlb, vma, start, end, details);
1353         }
1354 }
1355
1356 /**
1357  * unmap_vmas - unmap a range of memory covered by a list of vma's
1358  * @tlb: address of the caller's struct mmu_gather
1359  * @vma: the starting vma
1360  * @start_addr: virtual address at which to start unmapping
1361  * @end_addr: virtual address at which to end unmapping
1362  *
1363  * Unmap all pages in the vma list.
1364  *
1365  * Only addresses between `start' and `end' will be unmapped.
1366  *
1367  * The VMA list must be sorted in ascending virtual address order.
1368  *
1369  * unmap_vmas() assumes that the caller will flush the whole unmapped address
1370  * range after unmap_vmas() returns.  So the only responsibility here is to
1371  * ensure that any thus-far unmapped pages are flushed before unmap_vmas()
1372  * drops the lock and schedules.
1373  */
1374 void unmap_vmas(struct mmu_gather *tlb,
1375                 struct vm_area_struct *vma, unsigned long start_addr,
1376                 unsigned long end_addr)
1377 {
1378         struct mm_struct *mm = vma->vm_mm;
1379
1380         mmu_notifier_invalidate_range_start(mm, start_addr, end_addr);
1381         for ( ; vma && vma->vm_start < end_addr; vma = vma->vm_next)
1382                 unmap_single_vma(tlb, vma, start_addr, end_addr, NULL);
1383         mmu_notifier_invalidate_range_end(mm, start_addr, end_addr);
1384 }
1385
1386 /**
1387  * zap_page_range - remove user pages in a given range
1388  * @vma: vm_area_struct holding the applicable pages
1389  * @start: starting address of pages to zap
1390  * @size: number of bytes to zap
1391  * @details: details of nonlinear truncation or shared cache invalidation
1392  *
1393  * Caller must protect the VMA list
1394  */
1395 void zap_page_range(struct vm_area_struct *vma, unsigned long start,
1396                 unsigned long size, struct zap_details *details)
1397 {
1398         struct mm_struct *mm = vma->vm_mm;
1399         struct mmu_gather tlb;
1400         unsigned long end = start + size;
1401
1402         lru_add_drain();
1403         tlb_gather_mmu(&tlb, mm, 0);
1404         update_hiwater_rss(mm);
1405         mmu_notifier_invalidate_range_start(mm, start, end);
1406         for ( ; vma && vma->vm_start < end; vma = vma->vm_next)
1407                 unmap_single_vma(&tlb, vma, start, end, details);
1408         mmu_notifier_invalidate_range_end(mm, start, end);
1409         tlb_finish_mmu(&tlb, start, end);
1410 }
1411
1412 /**
1413  * zap_page_range_single - remove user pages in a given range
1414  * @vma: vm_area_struct holding the applicable pages
1415  * @address: starting address of pages to zap
1416  * @size: number of bytes to zap
1417  * @details: details of nonlinear truncation or shared cache invalidation
1418  *
1419  * The range must fit into one VMA.
1420  */
1421 static void zap_page_range_single(struct vm_area_struct *vma, unsigned long address,
1422                 unsigned long size, struct zap_details *details)
1423 {
1424         struct mm_struct *mm = vma->vm_mm;
1425         struct mmu_gather tlb;
1426         unsigned long end = address + size;
1427
1428         lru_add_drain();
1429         tlb_gather_mmu(&tlb, mm, 0);
1430         update_hiwater_rss(mm);
1431         mmu_notifier_invalidate_range_start(mm, address, end);
1432         unmap_single_vma(&tlb, vma, address, end, details);
1433         mmu_notifier_invalidate_range_end(mm, address, end);
1434         tlb_finish_mmu(&tlb, address, end);
1435 }
1436
1437 /**
1438  * zap_vma_ptes - remove ptes mapping the vma
1439  * @vma: vm_area_struct holding ptes to be zapped
1440  * @address: starting address of pages to zap
1441  * @size: number of bytes to zap
1442  *
1443  * This function only unmaps ptes assigned to VM_PFNMAP vmas.
1444  *
1445  * The entire address range must be fully contained within the vma.
1446  *
1447  * Returns 0 if successful.
1448  */
1449 int zap_vma_ptes(struct vm_area_struct *vma, unsigned long address,
1450                 unsigned long size)
1451 {
1452         if (address < vma->vm_start || address + size > vma->vm_end ||
1453                         !(vma->vm_flags & VM_PFNMAP))
1454                 return -1;
1455         zap_page_range_single(vma, address, size, NULL);
1456         return 0;
1457 }
1458 EXPORT_SYMBOL_GPL(zap_vma_ptes);
1459
1460 /**
1461  * follow_page - look up a page descriptor from a user-virtual address
1462  * @vma: vm_area_struct mapping @address
1463  * @address: virtual address to look up
1464  * @flags: flags modifying lookup behaviour
1465  *
1466  * @flags can have FOLL_ flags set, defined in <linux/mm.h>
1467  *
1468  * Returns the mapped (struct page *), %NULL if no mapping exists, or
1469  * an error pointer if there is a mapping to something not represented
1470  * by a page descriptor (see also vm_normal_page()).
1471  */
1472 struct page *follow_page(struct vm_area_struct *vma, unsigned long address,
1473                         unsigned int flags)
1474 {
1475         pgd_t *pgd;
1476         pud_t *pud;
1477         pmd_t *pmd;
1478         pte_t *ptep, pte;
1479         spinlock_t *ptl;
1480         struct page *page;
1481         struct mm_struct *mm = vma->vm_mm;
1482
1483         page = follow_huge_addr(mm, address, flags & FOLL_WRITE);
1484         if (!IS_ERR(page)) {
1485                 BUG_ON(flags & FOLL_GET);
1486                 goto out;
1487         }
1488
1489         page = NULL;
1490         pgd = pgd_offset(mm, address);
1491         if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
1492                 goto no_page_table;
1493
1494         pud = pud_offset(pgd, address);
1495         if (pud_none(*pud))
1496                 goto no_page_table;
1497         if (pud_huge(*pud) && vma->vm_flags & VM_HUGETLB) {
1498                 BUG_ON(flags & FOLL_GET);
1499                 page = follow_huge_pud(mm, address, pud, flags & FOLL_WRITE);
1500                 goto out;
1501         }
1502         if (unlikely(pud_bad(*pud)))
1503                 goto no_page_table;
1504
1505         pmd = pmd_offset(pud, address);
1506         if (pmd_none(*pmd))
1507                 goto no_page_table;
1508         if (pmd_huge(*pmd) && vma->vm_flags & VM_HUGETLB) {
1509                 BUG_ON(flags & FOLL_GET);
1510                 page = follow_huge_pmd(mm, address, pmd, flags & FOLL_WRITE);
1511                 goto out;
1512         }
1513         if (pmd_trans_huge(*pmd)) {
1514                 if (flags & FOLL_SPLIT) {
1515                         split_huge_page_pmd(vma, address, pmd);
1516                         goto split_fallthrough;
1517                 }
1518                 spin_lock(&mm->page_table_lock);
1519                 if (likely(pmd_trans_huge(*pmd))) {
1520                         if (unlikely(pmd_trans_splitting(*pmd))) {
1521                                 spin_unlock(&mm->page_table_lock);
1522                                 wait_split_huge_page(vma->anon_vma, pmd);
1523                         } else {
1524                                 page = follow_trans_huge_pmd(vma, address,
1525                                                              pmd, flags);
1526                                 spin_unlock(&mm->page_table_lock);
1527                                 goto out;
1528                         }
1529                 } else
1530                         spin_unlock(&mm->page_table_lock);
1531                 /* fall through */
1532         }
1533 split_fallthrough:
1534         if (unlikely(pmd_bad(*pmd)))
1535                 goto no_page_table;
1536
1537         ptep = pte_offset_map_lock(mm, pmd, address, &ptl);
1538
1539         pte = *ptep;
1540         if (!pte_present(pte))
1541                 goto no_page;
1542         if ((flags & FOLL_WRITE) && !pte_write(pte))
1543                 goto unlock;
1544
1545         page = vm_normal_page(vma, address, pte);
1546         if (unlikely(!page)) {
1547                 if ((flags & FOLL_DUMP) ||
1548                     !is_zero_pfn(pte_pfn(pte)))
1549                         goto bad_page;
1550                 page = pte_page(pte);
1551         }
1552
1553         if (flags & FOLL_GET)
1554                 get_page_foll(page);
1555         if (flags & FOLL_TOUCH) {
1556                 if ((flags & FOLL_WRITE) &&
1557                     !pte_dirty(pte) && !PageDirty(page))
1558                         set_page_dirty(page);
1559                 /*
1560                  * pte_mkyoung() would be more correct here, but atomic care
1561                  * is needed to avoid losing the dirty bit: it is easier to use
1562                  * mark_page_accessed().
1563                  */
1564                 mark_page_accessed(page);
1565         }
1566         if ((flags & FOLL_MLOCK) && (vma->vm_flags & VM_LOCKED)) {
1567                 /*
1568                  * The preliminary mapping check is mainly to avoid the
1569                  * pointless overhead of lock_page on the ZERO_PAGE
1570                  * which might bounce very badly if there is contention.
1571                  *
1572                  * If the page is already locked, we don't need to
1573                  * handle it now - vmscan will handle it later if and
1574                  * when it attempts to reclaim the page.
1575                  */
1576                 if (page->mapping && trylock_page(page)) {
1577                         lru_add_drain();  /* push cached pages to LRU */
1578                         /*
1579                          * Because we lock page here, and migration is
1580                          * blocked by the pte's page reference, and we
1581                          * know the page is still mapped, we don't even
1582                          * need to check for file-cache page truncation.
1583                          */
1584                         mlock_vma_page(page);
1585                         unlock_page(page);
1586                 }
1587         }
1588 unlock:
1589         pte_unmap_unlock(ptep, ptl);
1590 out:
1591         return page;
1592
1593 bad_page:
1594         pte_unmap_unlock(ptep, ptl);
1595         return ERR_PTR(-EFAULT);
1596
1597 no_page:
1598         pte_unmap_unlock(ptep, ptl);
1599         if (!pte_none(pte))
1600                 return page;
1601
1602 no_page_table:
1603         /*
1604          * When core dumping an enormous anonymous area that nobody
1605          * has touched so far, we don't want to allocate unnecessary pages or
1606          * page tables.  Return error instead of NULL to skip handle_mm_fault,
1607          * then get_dump_page() will return NULL to leave a hole in the dump.
1608          * But we can only make this optimization where a hole would surely
1609          * be zero-filled if handle_mm_fault() actually did handle it.
1610          */
1611         if ((flags & FOLL_DUMP) &&
1612             (!vma->vm_ops || !vma->vm_ops->fault))
1613                 return ERR_PTR(-EFAULT);
1614         return page;
1615 }
1616
1617 static inline int stack_guard_page(struct vm_area_struct *vma, unsigned long addr)
1618 {
1619         return stack_guard_page_start(vma, addr) ||
1620                stack_guard_page_end(vma, addr+PAGE_SIZE);
1621 }
1622
1623 /**
1624  * __get_user_pages() - pin user pages in memory
1625  * @tsk:        task_struct of target task
1626  * @mm:         mm_struct of target mm
1627  * @start:      starting user address
1628  * @nr_pages:   number of pages from start to pin
1629  * @gup_flags:  flags modifying pin behaviour
1630  * @pages:      array that receives pointers to the pages pinned.
1631  *              Should be at least nr_pages long. Or NULL, if caller
1632  *              only intends to ensure the pages are faulted in.
1633  * @vmas:       array of pointers to vmas corresponding to each page.
1634  *              Or NULL if the caller does not require them.
1635  * @nonblocking: whether waiting for disk IO or mmap_sem contention
1636  *
1637  * Returns number of pages pinned. This may be fewer than the number
1638  * requested. If nr_pages is 0 or negative, returns 0. If no pages
1639  * were pinned, returns -errno. Each page returned must be released
1640  * with a put_page() call when it is finished with. vmas will only
1641  * remain valid while mmap_sem is held.
1642  *
1643  * Must be called with mmap_sem held for read or write.
1644  *
1645  * __get_user_pages walks a process's page tables and takes a reference to
1646  * each struct page that each user address corresponds to at a given
1647  * instant. That is, it takes the page that would be accessed if a user
1648  * thread accesses the given user virtual address at that instant.
1649  *
1650  * This does not guarantee that the page exists in the user mappings when
1651  * __get_user_pages returns, and there may even be a completely different
1652  * page there in some cases (eg. if mmapped pagecache has been invalidated
1653  * and subsequently re faulted). However it does guarantee that the page
1654  * won't be freed completely. And mostly callers simply care that the page
1655  * contains data that was valid *at some point in time*. Typically, an IO
1656  * or similar operation cannot guarantee anything stronger anyway because
1657  * locks can't be held over the syscall boundary.
1658  *
1659  * If @gup_flags & FOLL_WRITE == 0, the page must not be written to. If
1660  * the page is written to, set_page_dirty (or set_page_dirty_lock, as
1661  * appropriate) must be called after the page is finished with, and
1662  * before put_page is called.
1663  *
1664  * If @nonblocking != NULL, __get_user_pages will not wait for disk IO
1665  * or mmap_sem contention, and if waiting is needed to pin all pages,
1666  * *@nonblocking will be set to 0.
1667  *
1668  * In most cases, get_user_pages or get_user_pages_fast should be used
1669  * instead of __get_user_pages. __get_user_pages should be used only if
1670  * you need some special @gup_flags.
1671  */
1672 int __get_user_pages(struct task_struct *tsk, struct mm_struct *mm,
1673                      unsigned long start, int nr_pages, unsigned int gup_flags,
1674                      struct page **pages, struct vm_area_struct **vmas,
1675                      int *nonblocking)
1676 {
1677         int i;
1678         unsigned long vm_flags;
1679
1680         if (nr_pages <= 0)
1681                 return 0;
1682
1683         VM_BUG_ON(!!pages != !!(gup_flags & FOLL_GET));
1684
1685         /* 
1686          * Require read or write permissions.
1687          * If FOLL_FORCE is set, we only require the "MAY" flags.
1688          */
1689         vm_flags  = (gup_flags & FOLL_WRITE) ?
1690                         (VM_WRITE | VM_MAYWRITE) : (VM_READ | VM_MAYREAD);
1691         vm_flags &= (gup_flags & FOLL_FORCE) ?
1692                         (VM_MAYREAD | VM_MAYWRITE) : (VM_READ | VM_WRITE);
1693         i = 0;
1694
1695         do {
1696                 struct vm_area_struct *vma;
1697
1698                 vma = find_extend_vma(mm, start);
1699                 if (!vma && in_gate_area(mm, start)) {
1700                         unsigned long pg = start & PAGE_MASK;
1701                         pgd_t *pgd;
1702                         pud_t *pud;
1703                         pmd_t *pmd;
1704                         pte_t *pte;
1705
1706                         /* user gate pages are read-only */
1707                         if (gup_flags & FOLL_WRITE)
1708                                 return i ? : -EFAULT;
1709                         if (pg > TASK_SIZE)
1710                                 pgd = pgd_offset_k(pg);
1711                         else
1712                                 pgd = pgd_offset_gate(mm, pg);
1713                         BUG_ON(pgd_none(*pgd));
1714                         pud = pud_offset(pgd, pg);
1715                         BUG_ON(pud_none(*pud));
1716                         pmd = pmd_offset(pud, pg);
1717                         if (pmd_none(*pmd))
1718                                 return i ? : -EFAULT;
1719                         VM_BUG_ON(pmd_trans_huge(*pmd));
1720                         pte = pte_offset_map(pmd, pg);
1721                         if (pte_none(*pte)) {
1722                                 pte_unmap(pte);
1723                                 return i ? : -EFAULT;
1724                         }
1725                         vma = get_gate_vma(mm);
1726                         if (pages) {
1727                                 struct page *page;
1728
1729                                 page = vm_normal_page(vma, start, *pte);
1730                                 if (!page) {
1731                                         if (!(gup_flags & FOLL_DUMP) &&
1732                                              is_zero_pfn(pte_pfn(*pte)))
1733                                                 page = pte_page(*pte);
1734                                         else {
1735                                                 pte_unmap(pte);
1736                                                 return i ? : -EFAULT;
1737                                         }
1738                                 }
1739                                 pages[i] = page;
1740                                 get_page(page);
1741                         }
1742                         pte_unmap(pte);
1743                         goto next_page;
1744                 }
1745
1746                 if (!vma ||
1747                     (vma->vm_flags & (VM_IO | VM_PFNMAP)) ||
1748                     !(vm_flags & vma->vm_flags))
1749                         return i ? : -EFAULT;
1750
1751                 if (is_vm_hugetlb_page(vma)) {
1752                         i = follow_hugetlb_page(mm, vma, pages, vmas,
1753                                         &start, &nr_pages, i, gup_flags);
1754                         continue;
1755                 }
1756
1757                 do {
1758                         struct page *page;
1759                         unsigned int foll_flags = gup_flags;
1760
1761                         /*
1762                          * If we have a pending SIGKILL, don't keep faulting
1763                          * pages and potentially allocating memory.
1764                          */
1765                         if (unlikely(fatal_signal_pending(current)))
1766                                 return i ? i : -ERESTARTSYS;
1767
1768                         cond_resched();
1769                         while (!(page = follow_page(vma, start, foll_flags))) {
1770                                 int ret;
1771                                 unsigned int fault_flags = 0;
1772
1773                                 /* For mlock, just skip the stack guard page. */
1774                                 if (foll_flags & FOLL_MLOCK) {
1775                                         if (stack_guard_page(vma, start))
1776                                                 goto next_page;
1777                                 }
1778                                 if (foll_flags & FOLL_WRITE)
1779                                         fault_flags |= FAULT_FLAG_WRITE;
1780                                 if (nonblocking)
1781                                         fault_flags |= FAULT_FLAG_ALLOW_RETRY;
1782                                 if (foll_flags & FOLL_NOWAIT)
1783                                         fault_flags |= (FAULT_FLAG_ALLOW_RETRY | FAULT_FLAG_RETRY_NOWAIT);
1784
1785                                 ret = handle_mm_fault(mm, vma, start,
1786                                                         fault_flags);
1787
1788                                 if (ret & VM_FAULT_ERROR) {
1789                                         if (ret & VM_FAULT_OOM)
1790                                                 return i ? i : -ENOMEM;
1791                                         if (ret & (VM_FAULT_HWPOISON |
1792                                                    VM_FAULT_HWPOISON_LARGE)) {
1793                                                 if (i)
1794                                                         return i;
1795                                                 else if (gup_flags & FOLL_HWPOISON)
1796                                                         return -EHWPOISON;
1797                                                 else
1798                                                         return -EFAULT;
1799                                         }
1800                                         if (ret & VM_FAULT_SIGBUS)
1801                                                 return i ? i : -EFAULT;
1802                                         BUG();
1803                                 }
1804
1805                                 if (tsk) {
1806                                         if (ret & VM_FAULT_MAJOR)
1807                                                 tsk->maj_flt++;
1808                                         else
1809                                                 tsk->min_flt++;
1810                                 }
1811
1812                                 if (ret & VM_FAULT_RETRY) {
1813                                         if (nonblocking)
1814                                                 *nonblocking = 0;
1815                                         return i;
1816                                 }
1817
1818                                 /*
1819                                  * The VM_FAULT_WRITE bit tells us that
1820                                  * do_wp_page has broken COW when necessary,
1821                                  * even if maybe_mkwrite decided not to set
1822                                  * pte_write. We can thus safely do subsequent
1823                                  * page lookups as if they were reads. But only
1824                                  * do so when looping for pte_write is futile:
1825                                  * in some cases userspace may also be wanting
1826                                  * to write to the gotten user page, which a
1827                                  * read fault here might prevent (a readonly
1828                                  * page might get reCOWed by userspace write).
1829                                  */
1830                                 if ((ret & VM_FAULT_WRITE) &&
1831                                     !(vma->vm_flags & VM_WRITE))
1832                                         foll_flags &= ~FOLL_WRITE;
1833
1834                                 cond_resched();
1835                         }
1836                         if (IS_ERR(page))
1837                                 return i ? i : PTR_ERR(page);
1838                         if (pages) {
1839                                 pages[i] = page;
1840
1841                                 flush_anon_page(vma, page, start);
1842                                 flush_dcache_page(page);
1843                         }
1844 next_page:
1845                         if (vmas)
1846                                 vmas[i] = vma;
1847                         i++;
1848                         start += PAGE_SIZE;
1849                         nr_pages--;
1850                 } while (nr_pages && start < vma->vm_end);
1851         } while (nr_pages);
1852         return i;
1853 }
1854 EXPORT_SYMBOL(__get_user_pages);
1855
1856 /*
1857  * fixup_user_fault() - manually resolve a user page fault
1858  * @tsk:        the task_struct to use for page fault accounting, or
1859  *              NULL if faults are not to be recorded.
1860  * @mm:         mm_struct of target mm
1861  * @address:    user address
1862  * @fault_flags:flags to pass down to handle_mm_fault()
1863  *
1864  * This is meant to be called in the specific scenario where for locking reasons
1865  * we try to access user memory in atomic context (within a pagefault_disable()
1866  * section), this returns -EFAULT, and we want to resolve the user fault before
1867  * trying again.
1868  *
1869  * Typically this is meant to be used by the futex code.
1870  *
1871  * The main difference with get_user_pages() is that this function will
1872  * unconditionally call handle_mm_fault() which will in turn perform all the
1873  * necessary SW fixup of the dirty and young bits in the PTE, while
1874  * handle_mm_fault() only guarantees to update these in the struct page.
1875  *
1876  * This is important for some architectures where those bits also gate the
1877  * access permission to the page because they are maintained in software.  On
1878  * such architectures, gup() will not be enough to make a subsequent access
1879  * succeed.
1880  *
1881  * This should be called with the mm_sem held for read.
1882  */
1883 int fixup_user_fault(struct task_struct *tsk, struct mm_struct *mm,
1884                      unsigned long address, unsigned int fault_flags)
1885 {
1886         struct vm_area_struct *vma;
1887         int ret;
1888
1889         vma = find_extend_vma(mm, address);
1890         if (!vma || address < vma->vm_start)
1891                 return -EFAULT;
1892
1893         ret = handle_mm_fault(mm, vma, address, fault_flags);
1894         if (ret & VM_FAULT_ERROR) {
1895                 if (ret & VM_FAULT_OOM)
1896                         return -ENOMEM;
1897                 if (ret & (VM_FAULT_HWPOISON | VM_FAULT_HWPOISON_LARGE))
1898                         return -EHWPOISON;
1899                 if (ret & VM_FAULT_SIGBUS)
1900                         return -EFAULT;
1901                 BUG();
1902         }
1903         if (tsk) {
1904                 if (ret & VM_FAULT_MAJOR)
1905                         tsk->maj_flt++;
1906                 else
1907                         tsk->min_flt++;
1908         }
1909         return 0;
1910 }
1911
1912 /*
1913  * get_user_pages() - pin user pages in memory
1914  * @tsk:        the task_struct to use for page fault accounting, or
1915  *              NULL if faults are not to be recorded.
1916  * @mm:         mm_struct of target mm
1917  * @start:      starting user address
1918  * @nr_pages:   number of pages from start to pin
1919  * @write:      whether pages will be written to by the caller
1920  * @force:      whether to force write access even if user mapping is
1921  *              readonly. This will result in the page being COWed even
1922  *              in MAP_SHARED mappings. You do not want this.
1923  * @pages:      array that receives pointers to the pages pinned.
1924  *              Should be at least nr_pages long. Or NULL, if caller
1925  *              only intends to ensure the pages are faulted in.
1926  * @vmas:       array of pointers to vmas corresponding to each page.
1927  *              Or NULL if the caller does not require them.
1928  *
1929  * Returns number of pages pinned. This may be fewer than the number
1930  * requested. If nr_pages is 0 or negative, returns 0. If no pages
1931  * were pinned, returns -errno. Each page returned must be released
1932  * with a put_page() call when it is finished with. vmas will only
1933  * remain valid while mmap_sem is held.
1934  *
1935  * Must be called with mmap_sem held for read or write.
1936  *
1937  * get_user_pages walks a process's page tables and takes a reference to
1938  * each struct page that each user address corresponds to at a given
1939  * instant. That is, it takes the page that would be accessed if a user
1940  * thread accesses the given user virtual address at that instant.
1941  *
1942  * This does not guarantee that the page exists in the user mappings when
1943  * get_user_pages returns, and there may even be a completely different
1944  * page there in some cases (eg. if mmapped pagecache has been invalidated
1945  * and subsequently re faulted). However it does guarantee that the page
1946  * won't be freed completely. And mostly callers simply care that the page
1947  * contains data that was valid *at some point in time*. Typically, an IO
1948  * or similar operation cannot guarantee anything stronger anyway because
1949  * locks can't be held over the syscall boundary.
1950  *
1951  * If write=0, the page must not be written to. If the page is written to,
1952  * set_page_dirty (or set_page_dirty_lock, as appropriate) must be called
1953  * after the page is finished with, and before put_page is called.
1954  *
1955  * get_user_pages is typically used for fewer-copy IO operations, to get a
1956  * handle on the memory by some means other than accesses via the user virtual
1957  * addresses. The pages may be submitted for DMA to devices or accessed via
1958  * their kernel linear mapping (via the kmap APIs). Care should be taken to
1959  * use the correct cache flushing APIs.
1960  *
1961  * See also get_user_pages_fast, for performance critical applications.
1962  */
1963 int get_user_pages(struct task_struct *tsk, struct mm_struct *mm,
1964                 unsigned long start, int nr_pages, int write, int force,
1965                 struct page **pages, struct vm_area_struct **vmas)
1966 {
1967         int flags = FOLL_TOUCH;
1968
1969         if (pages)
1970                 flags |= FOLL_GET;
1971         if (write)
1972                 flags |= FOLL_WRITE;
1973         if (force)
1974                 flags |= FOLL_FORCE;
1975
1976         return __get_user_pages(tsk, mm, start, nr_pages, flags, pages, vmas,
1977                                 NULL);
1978 }
1979 EXPORT_SYMBOL(get_user_pages);
1980
1981 /**
1982  * get_dump_page() - pin user page in memory while writing it to core dump
1983  * @addr: user address
1984  *
1985  * Returns struct page pointer of user page pinned for dump,
1986  * to be freed afterwards by page_cache_release() or put_page().
1987  *
1988  * Returns NULL on any kind of failure - a hole must then be inserted into
1989  * the corefile, to preserve alignment with its headers; and also returns
1990  * NULL wherever the ZERO_PAGE, or an anonymous pte_none, has been found -
1991  * allowing a hole to be left in the corefile to save diskspace.
1992  *
1993  * Called without mmap_sem, but after all other threads have been killed.
1994  */
1995 #ifdef CONFIG_ELF_CORE
1996 struct page *get_dump_page(unsigned long addr)
1997 {
1998         struct vm_area_struct *vma;
1999         struct page *page;
2000
2001         if (__get_user_pages(current, current->mm, addr, 1,
2002                              FOLL_FORCE | FOLL_DUMP | FOLL_GET, &page, &vma,
2003                              NULL) < 1)
2004                 return NULL;
2005         flush_cache_page(vma, addr, page_to_pfn(page));
2006         return page;
2007 }
2008 #endif /* CONFIG_ELF_CORE */
2009
2010 pte_t *__get_locked_pte(struct mm_struct *mm, unsigned long addr,
2011                         spinlock_t **ptl)
2012 {
2013         pgd_t * pgd = pgd_offset(mm, addr);
2014         pud_t * pud = pud_alloc(mm, pgd, addr);
2015         if (pud) {
2016                 pmd_t * pmd = pmd_alloc(mm, pud, addr);
2017                 if (pmd) {
2018                         VM_BUG_ON(pmd_trans_huge(*pmd));
2019                         return pte_alloc_map_lock(mm, pmd, addr, ptl);
2020                 }
2021         }
2022         return NULL;
2023 }
2024
2025 /*
2026  * This is the old fallback for page remapping.
2027  *
2028  * For historical reasons, it only allows reserved pages. Only
2029  * old drivers should use this, and they needed to mark their
2030  * pages reserved for the old functions anyway.
2031  */
2032 static int insert_page(struct vm_area_struct *vma, unsigned long addr,
2033                         struct page *page, pgprot_t prot)
2034 {
2035         struct mm_struct *mm = vma->vm_mm;
2036         int retval;
2037         pte_t *pte;
2038         spinlock_t *ptl;
2039
2040         retval = -EINVAL;
2041         if (PageAnon(page))
2042                 goto out;
2043         retval = -ENOMEM;
2044         flush_dcache_page(page);
2045         pte = get_locked_pte(mm, addr, &ptl);
2046         if (!pte)
2047                 goto out;
2048         retval = -EBUSY;
2049         if (!pte_none(*pte))
2050                 goto out_unlock;
2051
2052         /* Ok, finally just insert the thing.. */
2053         get_page(page);
2054         inc_mm_counter_fast(mm, MM_FILEPAGES);
2055         page_add_file_rmap(page);
2056         set_pte_at(mm, addr, pte, mk_pte(page, prot));
2057
2058         retval = 0;
2059         pte_unmap_unlock(pte, ptl);
2060         return retval;
2061 out_unlock:
2062         pte_unmap_unlock(pte, ptl);
2063 out:
2064         return retval;
2065 }
2066
2067 /**
2068  * vm_insert_page - insert single page into user vma
2069  * @vma: user vma to map to
2070  * @addr: target user address of this page
2071  * @page: source kernel page
2072  *
2073  * This allows drivers to insert individual pages they've allocated
2074  * into a user vma.
2075  *
2076  * The page has to be a nice clean _individual_ kernel allocation.
2077  * If you allocate a compound page, you need to have marked it as
2078  * such (__GFP_COMP), or manually just split the page up yourself
2079  * (see split_page()).
2080  *
2081  * NOTE! Traditionally this was done with "remap_pfn_range()" which
2082  * took an arbitrary page protection parameter. This doesn't allow
2083  * that. Your vma protection will have to be set up correctly, which
2084  * means that if you want a shared writable mapping, you'd better
2085  * ask for a shared writable mapping!
2086  *
2087  * The page does not need to be reserved.
2088  *
2089  * Usually this function is called from f_op->mmap() handler
2090  * under mm->mmap_sem write-lock, so it can change vma->vm_flags.
2091  * Caller must set VM_MIXEDMAP on vma if it wants to call this
2092  * function from other places, for example from page-fault handler.
2093  */
2094 int vm_insert_page(struct vm_area_struct *vma, unsigned long addr,
2095                         struct page *page)
2096 {
2097         if (addr < vma->vm_start || addr >= vma->vm_end)
2098                 return -EFAULT;
2099         if (!page_count(page))
2100                 return -EINVAL;
2101         if (!(vma->vm_flags & VM_MIXEDMAP)) {
2102                 BUG_ON(down_read_trylock(&vma->vm_mm->mmap_sem));
2103                 BUG_ON(vma->vm_flags & VM_PFNMAP);
2104                 vma->vm_flags |= VM_MIXEDMAP;
2105         }
2106         return insert_page(vma, addr, page, vma->vm_page_prot);
2107 }
2108 EXPORT_SYMBOL(vm_insert_page);
2109
2110 static int insert_pfn(struct vm_area_struct *vma, unsigned long addr,
2111                         unsigned long pfn, pgprot_t prot)
2112 {
2113         struct mm_struct *mm = vma->vm_mm;
2114         int retval;
2115         pte_t *pte, entry;
2116         spinlock_t *ptl;
2117
2118         retval = -ENOMEM;
2119         pte = get_locked_pte(mm, addr, &ptl);
2120         if (!pte)
2121                 goto out;
2122         retval = -EBUSY;
2123         if (!pte_none(*pte))
2124                 goto out_unlock;
2125
2126         /* Ok, finally just insert the thing.. */
2127         entry = pte_mkspecial(pfn_pte(pfn, prot));
2128         set_pte_at(mm, addr, pte, entry);
2129         update_mmu_cache(vma, addr, pte); /* XXX: why not for insert_page? */
2130
2131         retval = 0;
2132 out_unlock:
2133         pte_unmap_unlock(pte, ptl);
2134 out:
2135         return retval;
2136 }
2137
2138 /**
2139  * vm_insert_pfn - insert single pfn into user vma
2140  * @vma: user vma to map to
2141  * @addr: target user address of this page
2142  * @pfn: source kernel pfn
2143  *
2144  * Similar to vm_insert_page, this allows drivers to insert individual pages
2145  * they've allocated into a user vma. Same comments apply.
2146  *
2147  * This function should only be called from a vm_ops->fault handler, and
2148  * in that case the handler should return NULL.
2149  *
2150  * vma cannot be a COW mapping.
2151  *
2152  * As this is called only for pages that do not currently exist, we
2153  * do not need to flush old virtual caches or the TLB.
2154  */
2155 int vm_insert_pfn(struct vm_area_struct *vma, unsigned long addr,
2156                         unsigned long pfn)
2157 {
2158         int ret;
2159         pgprot_t pgprot = vma->vm_page_prot;
2160         /*
2161          * Technically, architectures with pte_special can avoid all these
2162          * restrictions (same for remap_pfn_range).  However we would like
2163          * consistency in testing and feature parity among all, so we should
2164          * try to keep these invariants in place for everybody.
2165          */
2166         BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)));
2167         BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) ==
2168                                                 (VM_PFNMAP|VM_MIXEDMAP));
2169         BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags));
2170         BUG_ON((vma->vm_flags & VM_MIXEDMAP) && pfn_valid(pfn));
2171
2172         if (addr < vma->vm_start || addr >= vma->vm_end)
2173                 return -EFAULT;
2174         if (track_pfn_insert(vma, &pgprot, pfn))
2175                 return -EINVAL;
2176
2177         ret = insert_pfn(vma, addr, pfn, pgprot);
2178
2179         return ret;
2180 }
2181 EXPORT_SYMBOL(vm_insert_pfn);
2182
2183 int vm_insert_mixed(struct vm_area_struct *vma, unsigned long addr,
2184                         unsigned long pfn)
2185 {
2186         BUG_ON(!(vma->vm_flags & VM_MIXEDMAP));
2187
2188         if (addr < vma->vm_start || addr >= vma->vm_end)
2189                 return -EFAULT;
2190
2191         /*
2192          * If we don't have pte special, then we have to use the pfn_valid()
2193          * based VM_MIXEDMAP scheme (see vm_normal_page), and thus we *must*
2194          * refcount the page if pfn_valid is true (hence insert_page rather
2195          * than insert_pfn).  If a zero_pfn were inserted into a VM_MIXEDMAP
2196          * without pte special, it would there be refcounted as a normal page.
2197          */
2198         if (!HAVE_PTE_SPECIAL && pfn_valid(pfn)) {
2199                 struct page *page;
2200
2201                 page = pfn_to_page(pfn);
2202                 return insert_page(vma, addr, page, vma->vm_page_prot);
2203         }
2204         return insert_pfn(vma, addr, pfn, vma->vm_page_prot);
2205 }
2206 EXPORT_SYMBOL(vm_insert_mixed);
2207
2208 /*
2209  * maps a range of physical memory into the requested pages. the old
2210  * mappings are removed. any references to nonexistent pages results
2211  * in null mappings (currently treated as "copy-on-access")
2212  */
2213 static int remap_pte_range(struct mm_struct *mm, pmd_t *pmd,
2214                         unsigned long addr, unsigned long end,
2215                         unsigned long pfn, pgprot_t prot)
2216 {
2217         pte_t *pte;
2218         spinlock_t *ptl;
2219
2220         pte = pte_alloc_map_lock(mm, pmd, addr, &ptl);
2221         if (!pte)
2222                 return -ENOMEM;
2223         arch_enter_lazy_mmu_mode();
2224         do {
2225                 BUG_ON(!pte_none(*pte));
2226                 set_pte_at(mm, addr, pte, pte_mkspecial(pfn_pte(pfn, prot)));
2227                 pfn++;
2228         } while (pte++, addr += PAGE_SIZE, addr != end);
2229         arch_leave_lazy_mmu_mode();
2230         pte_unmap_unlock(pte - 1, ptl);
2231         return 0;
2232 }
2233
2234 static inline int remap_pmd_range(struct mm_struct *mm, pud_t *pud,
2235                         unsigned long addr, unsigned long end,
2236                         unsigned long pfn, pgprot_t prot)
2237 {
2238         pmd_t *pmd;
2239         unsigned long next;
2240
2241         pfn -= addr >> PAGE_SHIFT;
2242         pmd = pmd_alloc(mm, pud, addr);
2243         if (!pmd)
2244                 return -ENOMEM;
2245         VM_BUG_ON(pmd_trans_huge(*pmd));
2246         do {
2247                 next = pmd_addr_end(addr, end);
2248                 if (remap_pte_range(mm, pmd, addr, next,
2249                                 pfn + (addr >> PAGE_SHIFT), prot))
2250                         return -ENOMEM;
2251         } while (pmd++, addr = next, addr != end);
2252         return 0;
2253 }
2254
2255 static inline int remap_pud_range(struct mm_struct *mm, pgd_t *pgd,
2256                         unsigned long addr, unsigned long end,
2257                         unsigned long pfn, pgprot_t prot)
2258 {
2259         pud_t *pud;
2260         unsigned long next;
2261
2262         pfn -= addr >> PAGE_SHIFT;
2263         pud = pud_alloc(mm, pgd, addr);
2264         if (!pud)
2265                 return -ENOMEM;
2266         do {
2267                 next = pud_addr_end(addr, end);
2268                 if (remap_pmd_range(mm, pud, addr, next,
2269                                 pfn + (addr >> PAGE_SHIFT), prot))
2270                         return -ENOMEM;
2271         } while (pud++, addr = next, addr != end);
2272         return 0;
2273 }
2274
2275 /**
2276  * remap_pfn_range - remap kernel memory to userspace
2277  * @vma: user vma to map to
2278  * @addr: target user address to start at
2279  * @pfn: physical address of kernel memory
2280  * @size: size of map area
2281  * @prot: page protection flags for this mapping
2282  *
2283  *  Note: this is only safe if the mm semaphore is held when called.
2284  */
2285 int remap_pfn_range(struct vm_area_struct *vma, unsigned long addr,
2286                     unsigned long pfn, unsigned long size, pgprot_t prot)
2287 {
2288         pgd_t *pgd;
2289         unsigned long next;
2290         unsigned long end = addr + PAGE_ALIGN(size);
2291         struct mm_struct *mm = vma->vm_mm;
2292         int err;
2293
2294         /*
2295          * Physically remapped pages are special. Tell the
2296          * rest of the world about it:
2297          *   VM_IO tells people not to look at these pages
2298          *      (accesses can have side effects).
2299          *   VM_PFNMAP tells the core MM that the base pages are just
2300          *      raw PFN mappings, and do not have a "struct page" associated
2301          *      with them.
2302          *   VM_DONTEXPAND
2303          *      Disable vma merging and expanding with mremap().
2304          *   VM_DONTDUMP
2305          *      Omit vma from core dump, even when VM_IO turned off.
2306          *
2307          * There's a horrible special case to handle copy-on-write
2308          * behaviour that some programs depend on. We mark the "original"
2309          * un-COW'ed pages by matching them up with "vma->vm_pgoff".
2310          * See vm_normal_page() for details.
2311          */
2312         if (is_cow_mapping(vma->vm_flags)) {
2313                 if (addr != vma->vm_start || end != vma->vm_end)
2314                         return -EINVAL;
2315                 vma->vm_pgoff = pfn;
2316         }
2317
2318         err = track_pfn_remap(vma, &prot, pfn, addr, PAGE_ALIGN(size));
2319         if (err)
2320                 return -EINVAL;
2321
2322         vma->vm_flags |= VM_IO | VM_PFNMAP | VM_DONTEXPAND | VM_DONTDUMP;
2323
2324         BUG_ON(addr >= end);
2325         pfn -= addr >> PAGE_SHIFT;
2326         pgd = pgd_offset(mm, addr);
2327         flush_cache_range(vma, addr, end);
2328         do {
2329                 next = pgd_addr_end(addr, end);
2330                 err = remap_pud_range(mm, pgd, addr, next,
2331                                 pfn + (addr >> PAGE_SHIFT), prot);
2332                 if (err)
2333                         break;
2334         } while (pgd++, addr = next, addr != end);
2335
2336         if (err)
2337                 untrack_pfn(vma, pfn, PAGE_ALIGN(size));
2338
2339         return err;
2340 }
2341 EXPORT_SYMBOL(remap_pfn_range);
2342
2343 static int apply_to_pte_range(struct mm_struct *mm, pmd_t *pmd,
2344                                      unsigned long addr, unsigned long end,
2345                                      pte_fn_t fn, void *data)
2346 {
2347         pte_t *pte;
2348         int err;
2349         pgtable_t token;
2350         spinlock_t *uninitialized_var(ptl);
2351
2352         pte = (mm == &init_mm) ?
2353                 pte_alloc_kernel(pmd, addr) :
2354                 pte_alloc_map_lock(mm, pmd, addr, &ptl);
2355         if (!pte)
2356                 return -ENOMEM;
2357
2358         BUG_ON(pmd_huge(*pmd));
2359
2360         arch_enter_lazy_mmu_mode();
2361
2362         token = pmd_pgtable(*pmd);
2363
2364         do {
2365                 err = fn(pte++, token, addr, data);
2366                 if (err)
2367                         break;
2368         } while (addr += PAGE_SIZE, addr != end);
2369
2370         arch_leave_lazy_mmu_mode();
2371
2372         if (mm != &init_mm)
2373                 pte_unmap_unlock(pte-1, ptl);
2374         return err;
2375 }
2376
2377 static int apply_to_pmd_range(struct mm_struct *mm, pud_t *pud,
2378                                      unsigned long addr, unsigned long end,
2379                                      pte_fn_t fn, void *data)
2380 {
2381         pmd_t *pmd;
2382         unsigned long next;
2383         int err;
2384
2385         BUG_ON(pud_huge(*pud));
2386
2387         pmd = pmd_alloc(mm, pud, addr);
2388         if (!pmd)
2389                 return -ENOMEM;
2390         do {
2391                 next = pmd_addr_end(addr, end);
2392                 err = apply_to_pte_range(mm, pmd, addr, next, fn, data);
2393                 if (err)
2394                         break;
2395         } while (pmd++, addr = next, addr != end);
2396         return err;
2397 }
2398
2399 static int apply_to_pud_range(struct mm_struct *mm, pgd_t *pgd,
2400                                      unsigned long addr, unsigned long end,
2401                                      pte_fn_t fn, void *data)
2402 {
2403         pud_t *pud;
2404         unsigned long next;
2405         int err;
2406
2407         pud = pud_alloc(mm, pgd, addr);
2408         if (!pud)
2409                 return -ENOMEM;
2410         do {
2411                 next = pud_addr_end(addr, end);
2412                 err = apply_to_pmd_range(mm, pud, addr, next, fn, data);
2413                 if (err)
2414                         break;
2415         } while (pud++, addr = next, addr != end);
2416         return err;
2417 }
2418
2419 /*
2420  * Scan a region of virtual memory, filling in page tables as necessary
2421  * and calling a provided function on each leaf page table.
2422  */
2423 int apply_to_page_range(struct mm_struct *mm, unsigned long addr,
2424                         unsigned long size, pte_fn_t fn, void *data)
2425 {
2426         pgd_t *pgd;
2427         unsigned long next;
2428         unsigned long end = addr + size;
2429         int err;
2430
2431         BUG_ON(addr >= end);
2432         pgd = pgd_offset(mm, addr);
2433         do {
2434                 next = pgd_addr_end(addr, end);
2435                 err = apply_to_pud_range(mm, pgd, addr, next, fn, data);
2436                 if (err)
2437                         break;
2438         } while (pgd++, addr = next, addr != end);
2439
2440         return err;
2441 }
2442 EXPORT_SYMBOL_GPL(apply_to_page_range);
2443
2444 /*
2445  * handle_pte_fault chooses page fault handler according to an entry
2446  * which was read non-atomically.  Before making any commitment, on
2447  * those architectures or configurations (e.g. i386 with PAE) which
2448  * might give a mix of unmatched parts, do_swap_page and do_nonlinear_fault
2449  * must check under lock before unmapping the pte and proceeding
2450  * (but do_wp_page is only called after already making such a check;
2451  * and do_anonymous_page can safely check later on).
2452  */
2453 static inline int pte_unmap_same(struct mm_struct *mm, pmd_t *pmd,
2454                                 pte_t *page_table, pte_t orig_pte)
2455 {
2456         int same = 1;
2457 #if defined(CONFIG_SMP) || defined(CONFIG_PREEMPT)
2458         if (sizeof(pte_t) > sizeof(unsigned long)) {
2459                 spinlock_t *ptl = pte_lockptr(mm, pmd);
2460                 spin_lock(ptl);
2461                 same = pte_same(*page_table, orig_pte);
2462                 spin_unlock(ptl);
2463         }
2464 #endif
2465         pte_unmap(page_table);
2466         return same;
2467 }
2468
2469 static inline void cow_user_page(struct page *dst, struct page *src, unsigned long va, struct vm_area_struct *vma)
2470 {
2471         /*
2472          * If the source page was a PFN mapping, we don't have
2473          * a "struct page" for it. We do a best-effort copy by
2474          * just copying from the original user address. If that
2475          * fails, we just zero-fill it. Live with it.
2476          */
2477         if (unlikely(!src)) {
2478                 void *kaddr = kmap_atomic(dst);
2479                 void __user *uaddr = (void __user *)(va & PAGE_MASK);
2480
2481                 /*
2482                  * This really shouldn't fail, because the page is there
2483                  * in the page tables. But it might just be unreadable,
2484                  * in which case we just give up and fill the result with
2485                  * zeroes.
2486                  */
2487                 if (__copy_from_user_inatomic(kaddr, uaddr, PAGE_SIZE))
2488                         clear_page(kaddr);
2489                 kunmap_atomic(kaddr);
2490                 flush_dcache_page(dst);
2491         } else
2492                 copy_user_highpage(dst, src, va, vma);
2493 }
2494
2495 /*
2496  * This routine handles present pages, when users try to write
2497  * to a shared page. It is done by copying the page to a new address
2498  * and decrementing the shared-page counter for the old page.
2499  *
2500  * Note that this routine assumes that the protection checks have been
2501  * done by the caller (the low-level page fault routine in most cases).
2502  * Thus we can safely just mark it writable once we've done any necessary
2503  * COW.
2504  *
2505  * We also mark the page dirty at this point even though the page will
2506  * change only once the write actually happens. This avoids a few races,
2507  * and potentially makes it more efficient.
2508  *
2509  * We enter with non-exclusive mmap_sem (to exclude vma changes,
2510  * but allow concurrent faults), with pte both mapped and locked.
2511  * We return with mmap_sem still held, but pte unmapped and unlocked.
2512  */
2513 static int do_wp_page(struct mm_struct *mm, struct vm_area_struct *vma,
2514                 unsigned long address, pte_t *page_table, pmd_t *pmd,
2515                 spinlock_t *ptl, pte_t orig_pte)
2516         __releases(ptl)
2517 {
2518         struct page *old_page, *new_page = NULL;
2519         pte_t entry;
2520         int ret = 0;
2521         int page_mkwrite = 0;
2522         struct page *dirty_page = NULL;
2523         unsigned long mmun_start = 0;   /* For mmu_notifiers */
2524         unsigned long mmun_end = 0;     /* For mmu_notifiers */
2525
2526         old_page = vm_normal_page(vma, address, orig_pte);
2527         if (!old_page) {
2528                 /*
2529                  * VM_MIXEDMAP !pfn_valid() case
2530                  *
2531                  * We should not cow pages in a shared writeable mapping.
2532                  * Just mark the pages writable as we can't do any dirty
2533                  * accounting on raw pfn maps.
2534                  */
2535                 if ((vma->vm_flags & (VM_WRITE|VM_SHARED)) ==
2536                                      (VM_WRITE|VM_SHARED))
2537                         goto reuse;
2538                 goto gotten;
2539         }
2540
2541         /*
2542          * Take out anonymous pages first, anonymous shared vmas are
2543          * not dirty accountable.
2544          */
2545         if (PageAnon(old_page) && !PageKsm(old_page)) {
2546                 if (!trylock_page(old_page)) {
2547                         page_cache_get(old_page);
2548                         pte_unmap_unlock(page_table, ptl);
2549                         lock_page(old_page);
2550                         page_table = pte_offset_map_lock(mm, pmd, address,
2551                                                          &ptl);
2552                         if (!pte_same(*page_table, orig_pte)) {
2553                                 unlock_page(old_page);
2554                                 goto unlock;
2555                         }
2556                         page_cache_release(old_page);
2557                 }
2558                 if (reuse_swap_page(old_page)) {
2559                         /*
2560                          * The page is all ours.  Move it to our anon_vma so
2561                          * the rmap code will not search our parent or siblings.
2562                          * Protected against the rmap code by the page lock.
2563                          */
2564                         page_move_anon_rmap(old_page, vma, address);
2565                         unlock_page(old_page);
2566                         goto reuse;
2567                 }
2568                 unlock_page(old_page);
2569         } else if (unlikely((vma->vm_flags & (VM_WRITE|VM_SHARED)) ==
2570                                         (VM_WRITE|VM_SHARED))) {
2571                 /*
2572                  * Only catch write-faults on shared writable pages,
2573                  * read-only shared pages can get COWed by
2574                  * get_user_pages(.write=1, .force=1).
2575                  */
2576                 if (vma->vm_ops && vma->vm_ops->page_mkwrite) {
2577                         struct vm_fault vmf;
2578                         int tmp;
2579
2580                         vmf.virtual_address = (void __user *)(address &
2581                                                                 PAGE_MASK);
2582                         vmf.pgoff = old_page->index;
2583                         vmf.flags = FAULT_FLAG_WRITE|FAULT_FLAG_MKWRITE;
2584                         vmf.page = old_page;
2585
2586                         /*
2587                          * Notify the address space that the page is about to
2588                          * become writable so that it can prohibit this or wait
2589                          * for the page to get into an appropriate state.
2590                          *
2591                          * We do this without the lock held, so that it can
2592                          * sleep if it needs to.
2593                          */
2594                         page_cache_get(old_page);
2595                         pte_unmap_unlock(page_table, ptl);
2596
2597                         tmp = vma->vm_ops->page_mkwrite(vma, &vmf);
2598                         if (unlikely(tmp &
2599                                         (VM_FAULT_ERROR | VM_FAULT_NOPAGE))) {
2600                                 ret = tmp;
2601                                 goto unwritable_page;
2602                         }
2603                         if (unlikely(!(tmp & VM_FAULT_LOCKED))) {
2604                                 lock_page(old_page);
2605                                 if (!old_page->mapping) {
2606                                         ret = 0; /* retry the fault */
2607                                         unlock_page(old_page);
2608                                         goto unwritable_page;
2609                                 }
2610                         } else
2611                                 VM_BUG_ON(!PageLocked(old_page));
2612
2613                         /*
2614                          * Since we dropped the lock we need to revalidate
2615                          * the PTE as someone else may have changed it.  If
2616                          * they did, we just return, as we can count on the
2617                          * MMU to tell us if they didn't also make it writable.
2618                          */
2619                         page_table = pte_offset_map_lock(mm, pmd, address,
2620                                                          &ptl);
2621                         if (!pte_same(*page_table, orig_pte)) {
2622                                 unlock_page(old_page);
2623                                 goto unlock;
2624                         }
2625
2626                         page_mkwrite = 1;
2627                 }
2628                 dirty_page = old_page;
2629                 get_page(dirty_page);
2630
2631 reuse:
2632                 flush_cache_page(vma, address, pte_pfn(orig_pte));
2633                 entry = pte_mkyoung(orig_pte);
2634                 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2635                 if (ptep_set_access_flags(vma, address, page_table, entry,1))
2636                         update_mmu_cache(vma, address, page_table);
2637                 pte_unmap_unlock(page_table, ptl);
2638                 ret |= VM_FAULT_WRITE;
2639
2640                 if (!dirty_page)
2641                         return ret;
2642
2643                 /*
2644                  * Yes, Virginia, this is actually required to prevent a race
2645                  * with clear_page_dirty_for_io() from clearing the page dirty
2646                  * bit after it clear all dirty ptes, but before a racing
2647                  * do_wp_page installs a dirty pte.
2648                  *
2649                  * __do_fault is protected similarly.
2650                  */
2651                 if (!page_mkwrite) {
2652                         wait_on_page_locked(dirty_page);
2653                         set_page_dirty_balance(dirty_page, page_mkwrite);
2654                         /* file_update_time outside page_lock */
2655                         if (vma->vm_file)
2656                                 file_update_time(vma->vm_file);
2657                 }
2658                 put_page(dirty_page);
2659                 if (page_mkwrite) {
2660                         struct address_space *mapping = dirty_page->mapping;
2661
2662                         set_page_dirty(dirty_page);
2663                         unlock_page(dirty_page);
2664                         page_cache_release(dirty_page);
2665                         if (mapping)    {
2666                                 /*
2667                                  * Some device drivers do not set page.mapping
2668                                  * but still dirty their pages
2669                                  */
2670                                 balance_dirty_pages_ratelimited(mapping);
2671                         }
2672                 }
2673
2674                 return ret;
2675         }
2676
2677         /*
2678          * Ok, we need to copy. Oh, well..
2679          */
2680         page_cache_get(old_page);
2681 gotten:
2682         pte_unmap_unlock(page_table, ptl);
2683
2684         if (unlikely(anon_vma_prepare(vma)))
2685                 goto oom;
2686
2687         if (is_zero_pfn(pte_pfn(orig_pte))) {
2688                 new_page = alloc_zeroed_user_highpage_movable(vma, address);
2689                 if (!new_page)
2690                         goto oom;
2691         } else {
2692                 new_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, address);
2693                 if (!new_page)
2694                         goto oom;
2695                 cow_user_page(new_page, old_page, address, vma);
2696         }
2697         __SetPageUptodate(new_page);
2698
2699         if (mem_cgroup_newpage_charge(new_page, mm, GFP_KERNEL))
2700                 goto oom_free_new;
2701
2702         mmun_start  = address & PAGE_MASK;
2703         mmun_end    = mmun_start + PAGE_SIZE;
2704         mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
2705
2706         /*
2707          * Re-check the pte - we dropped the lock
2708          */
2709         page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2710         if (likely(pte_same(*page_table, orig_pte))) {
2711                 if (old_page) {
2712                         if (!PageAnon(old_page)) {
2713                                 dec_mm_counter_fast(mm, MM_FILEPAGES);
2714                                 inc_mm_counter_fast(mm, MM_ANONPAGES);
2715                         }
2716                 } else
2717                         inc_mm_counter_fast(mm, MM_ANONPAGES);
2718                 flush_cache_page(vma, address, pte_pfn(orig_pte));
2719                 entry = mk_pte(new_page, vma->vm_page_prot);
2720                 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2721                 /*
2722                  * Clear the pte entry and flush it first, before updating the
2723                  * pte with the new entry. This will avoid a race condition
2724                  * seen in the presence of one thread doing SMC and another
2725                  * thread doing COW.
2726                  */
2727                 ptep_clear_flush(vma, address, page_table);
2728                 page_add_new_anon_rmap(new_page, vma, address);
2729                 /*
2730                  * We call the notify macro here because, when using secondary
2731                  * mmu page tables (such as kvm shadow page tables), we want the
2732                  * new page to be mapped directly into the secondary page table.
2733                  */
2734                 set_pte_at_notify(mm, address, page_table, entry);
2735                 update_mmu_cache(vma, address, page_table);
2736                 if (old_page) {
2737                         /*
2738                          * Only after switching the pte to the new page may
2739                          * we remove the mapcount here. Otherwise another
2740                          * process may come and find the rmap count decremented
2741                          * before the pte is switched to the new page, and
2742                          * "reuse" the old page writing into it while our pte
2743                          * here still points into it and can be read by other
2744                          * threads.
2745                          *
2746                          * The critical issue is to order this
2747                          * page_remove_rmap with the ptp_clear_flush above.
2748                          * Those stores are ordered by (if nothing else,)
2749                          * the barrier present in the atomic_add_negative
2750                          * in page_remove_rmap.
2751                          *
2752                          * Then the TLB flush in ptep_clear_flush ensures that
2753                          * no process can access the old page before the
2754                          * decremented mapcount is visible. And the old page
2755                          * cannot be reused until after the decremented
2756                          * mapcount is visible. So transitively, TLBs to
2757                          * old page will be flushed before it can be reused.
2758                          */
2759                         page_remove_rmap(old_page);
2760                 }
2761
2762                 /* Free the old page.. */
2763                 new_page = old_page;
2764                 ret |= VM_FAULT_WRITE;
2765         } else
2766                 mem_cgroup_uncharge_page(new_page);
2767
2768         if (new_page)
2769                 page_cache_release(new_page);
2770 unlock:
2771         pte_unmap_unlock(page_table, ptl);
2772         if (mmun_end > mmun_start)
2773                 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
2774         if (old_page) {
2775                 /*
2776                  * Don't let another task, with possibly unlocked vma,
2777                  * keep the mlocked page.
2778                  */
2779                 if ((ret & VM_FAULT_WRITE) && (vma->vm_flags & VM_LOCKED)) {
2780                         lock_page(old_page);    /* LRU manipulation */
2781                         munlock_vma_page(old_page);
2782                         unlock_page(old_page);
2783                 }
2784                 page_cache_release(old_page);
2785         }
2786         return ret;
2787 oom_free_new:
2788         page_cache_release(new_page);
2789 oom:
2790         if (old_page) {
2791                 if (page_mkwrite) {
2792                         unlock_page(old_page);
2793                         page_cache_release(old_page);
2794                 }
2795                 page_cache_release(old_page);
2796         }
2797         return VM_FAULT_OOM;
2798
2799 unwritable_page:
2800         page_cache_release(old_page);
2801         return ret;
2802 }
2803
2804 static void unmap_mapping_range_vma(struct vm_area_struct *vma,
2805                 unsigned long start_addr, unsigned long end_addr,
2806                 struct zap_details *details)
2807 {
2808         zap_page_range_single(vma, start_addr, end_addr - start_addr, details);
2809 }
2810
2811 static inline void unmap_mapping_range_tree(struct rb_root *root,
2812                                             struct zap_details *details)
2813 {
2814         struct vm_area_struct *vma;
2815         pgoff_t vba, vea, zba, zea;
2816
2817         vma_interval_tree_foreach(vma, root,
2818                         details->first_index, details->last_index) {
2819
2820                 vba = vma->vm_pgoff;
2821                 vea = vba + ((vma->vm_end - vma->vm_start) >> PAGE_SHIFT) - 1;
2822                 /* Assume for now that PAGE_CACHE_SHIFT == PAGE_SHIFT */
2823                 zba = details->first_index;
2824                 if (zba < vba)
2825                         zba = vba;
2826                 zea = details->last_index;
2827                 if (zea > vea)
2828                         zea = vea;
2829
2830                 unmap_mapping_range_vma(vma,
2831                         ((zba - vba) << PAGE_SHIFT) + vma->vm_start,
2832                         ((zea - vba + 1) << PAGE_SHIFT) + vma->vm_start,
2833                                 details);
2834         }
2835 }
2836
2837 static inline void unmap_mapping_range_list(struct list_head *head,
2838                                             struct zap_details *details)
2839 {
2840         struct vm_area_struct *vma;
2841
2842         /*
2843          * In nonlinear VMAs there is no correspondence between virtual address
2844          * offset and file offset.  So we must perform an exhaustive search
2845          * across *all* the pages in each nonlinear VMA, not just the pages
2846          * whose virtual address lies outside the file truncation point.
2847          */
2848         list_for_each_entry(vma, head, shared.nonlinear) {
2849                 details->nonlinear_vma = vma;
2850                 unmap_mapping_range_vma(vma, vma->vm_start, vma->vm_end, details);
2851         }
2852 }
2853
2854 /**
2855  * unmap_mapping_range - unmap the portion of all mmaps in the specified address_space corresponding to the specified page range in the underlying file.
2856  * @mapping: the address space containing mmaps to be unmapped.
2857  * @holebegin: byte in first page to unmap, relative to the start of
2858  * the underlying file.  This will be rounded down to a PAGE_SIZE
2859  * boundary.  Note that this is different from truncate_pagecache(), which
2860  * must keep the partial page.  In contrast, we must get rid of
2861  * partial pages.
2862  * @holelen: size of prospective hole in bytes.  This will be rounded
2863  * up to a PAGE_SIZE boundary.  A holelen of zero truncates to the
2864  * end of the file.
2865  * @even_cows: 1 when truncating a file, unmap even private COWed pages;
2866  * but 0 when invalidating pagecache, don't throw away private data.
2867  */
2868 void unmap_mapping_range(struct address_space *mapping,
2869                 loff_t const holebegin, loff_t const holelen, int even_cows)
2870 {
2871         struct zap_details details;
2872         pgoff_t hba = holebegin >> PAGE_SHIFT;
2873         pgoff_t hlen = (holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
2874
2875         /* Check for overflow. */
2876         if (sizeof(holelen) > sizeof(hlen)) {
2877                 long long holeend =
2878                         (holebegin + holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
2879                 if (holeend & ~(long long)ULONG_MAX)
2880                         hlen = ULONG_MAX - hba + 1;
2881         }
2882
2883         details.check_mapping = even_cows? NULL: mapping;
2884         details.nonlinear_vma = NULL;
2885         details.first_index = hba;
2886         details.last_index = hba + hlen - 1;
2887         if (details.last_index < details.first_index)
2888                 details.last_index = ULONG_MAX;
2889
2890
2891         mutex_lock(&mapping->i_mmap_mutex);
2892         if (unlikely(!RB_EMPTY_ROOT(&mapping->i_mmap)))
2893                 unmap_mapping_range_tree(&mapping->i_mmap, &details);
2894         if (unlikely(!list_empty(&mapping->i_mmap_nonlinear)))
2895                 unmap_mapping_range_list(&mapping->i_mmap_nonlinear, &details);
2896         mutex_unlock(&mapping->i_mmap_mutex);
2897 }
2898 EXPORT_SYMBOL(unmap_mapping_range);
2899
2900 /*
2901  * We enter with non-exclusive mmap_sem (to exclude vma changes,
2902  * but allow concurrent faults), and pte mapped but not yet locked.
2903  * We return with mmap_sem still held, but pte unmapped and unlocked.
2904  */
2905 static int do_swap_page(struct mm_struct *mm, struct vm_area_struct *vma,
2906                 unsigned long address, pte_t *page_table, pmd_t *pmd,
2907                 unsigned int flags, pte_t orig_pte)
2908 {
2909         spinlock_t *ptl;
2910         struct page *page, *swapcache = NULL;
2911         swp_entry_t entry;
2912         pte_t pte;
2913         int locked;
2914         struct mem_cgroup *ptr;
2915         int exclusive = 0;
2916         int ret = 0;
2917
2918         if (!pte_unmap_same(mm, pmd, page_table, orig_pte))
2919                 goto out;
2920
2921         entry = pte_to_swp_entry(orig_pte);
2922         if (unlikely(non_swap_entry(entry))) {
2923                 if (is_migration_entry(entry)) {
2924                         migration_entry_wait(mm, pmd, address);
2925                 } else if (is_hwpoison_entry(entry)) {
2926                         ret = VM_FAULT_HWPOISON;
2927                 } else {
2928                         print_bad_pte(vma, address, orig_pte, NULL);
2929                         ret = VM_FAULT_SIGBUS;
2930                 }
2931                 goto out;
2932         }
2933         delayacct_set_flag(DELAYACCT_PF_SWAPIN);
2934         page = lookup_swap_cache(entry);
2935         if (!page) {
2936                 page = swapin_readahead(entry,
2937                                         GFP_HIGHUSER_MOVABLE, vma, address);
2938                 if (!page) {
2939                         /*
2940                          * Back out if somebody else faulted in this pte
2941                          * while we released the pte lock.
2942                          */
2943                         page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2944                         if (likely(pte_same(*page_table, orig_pte)))
2945                                 ret = VM_FAULT_OOM;
2946                         delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
2947                         goto unlock;
2948                 }
2949
2950                 /* Had to read the page from swap area: Major fault */
2951                 ret = VM_FAULT_MAJOR;
2952                 count_vm_event(PGMAJFAULT);
2953                 mem_cgroup_count_vm_event(mm, PGMAJFAULT);
2954         } else if (PageHWPoison(page)) {
2955                 /*
2956                  * hwpoisoned dirty swapcache pages are kept for killing
2957                  * owner processes (which may be unknown at hwpoison time)
2958                  */
2959                 ret = VM_FAULT_HWPOISON;
2960                 delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
2961                 goto out_release;
2962         }
2963
2964         locked = lock_page_or_retry(page, mm, flags);
2965
2966         delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
2967         if (!locked) {
2968                 ret |= VM_FAULT_RETRY;
2969                 goto out_release;
2970         }
2971
2972         /*
2973          * Make sure try_to_free_swap or reuse_swap_page or swapoff did not
2974          * release the swapcache from under us.  The page pin, and pte_same
2975          * test below, are not enough to exclude that.  Even if it is still
2976          * swapcache, we need to check that the page's swap has not changed.
2977          */
2978         if (unlikely(!PageSwapCache(page) || page_private(page) != entry.val))
2979                 goto out_page;
2980
2981         if (ksm_might_need_to_copy(page, vma, address)) {
2982                 swapcache = page;
2983                 page = ksm_does_need_to_copy(page, vma, address);
2984
2985                 if (unlikely(!page)) {
2986                         ret = VM_FAULT_OOM;
2987                         page = swapcache;
2988                         swapcache = NULL;
2989                         goto out_page;
2990                 }
2991         }
2992
2993         if (mem_cgroup_try_charge_swapin(mm, page, GFP_KERNEL, &ptr)) {
2994                 ret = VM_FAULT_OOM;
2995                 goto out_page;
2996         }
2997
2998         /*
2999          * Back out if somebody else already faulted in this pte.
3000          */
3001         page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
3002         if (unlikely(!pte_same(*page_table, orig_pte)))
3003                 goto out_nomap;
3004
3005         if (unlikely(!PageUptodate(page))) {
3006                 ret = VM_FAULT_SIGBUS;
3007                 goto out_nomap;
3008         }
3009
3010         /*
3011          * The page isn't present yet, go ahead with the fault.
3012          *
3013          * Be careful about the sequence of operations here.
3014          * To get its accounting right, reuse_swap_page() must be called
3015          * while the page is counted on swap but not yet in mapcount i.e.
3016          * before page_add_anon_rmap() and swap_free(); try_to_free_swap()
3017          * must be called after the swap_free(), or it will never succeed.
3018          * Because delete_from_swap_page() may be called by reuse_swap_page(),
3019          * mem_cgroup_commit_charge_swapin() may not be able to find swp_entry
3020          * in page->private. In this case, a record in swap_cgroup  is silently
3021          * discarded at swap_free().
3022          */
3023
3024         inc_mm_counter_fast(mm, MM_ANONPAGES);
3025         dec_mm_counter_fast(mm, MM_SWAPENTS);
3026         pte = mk_pte(page, vma->vm_page_prot);
3027         if ((flags & FAULT_FLAG_WRITE) && reuse_swap_page(page)) {
3028                 pte = maybe_mkwrite(pte_mkdirty(pte), vma);
3029                 flags &= ~FAULT_FLAG_WRITE;
3030                 ret |= VM_FAULT_WRITE;
3031                 exclusive = 1;
3032         }
3033         flush_icache_page(vma, page);
3034         set_pte_at(mm, address, page_table, pte);
3035         do_page_add_anon_rmap(page, vma, address, exclusive);
3036         /* It's better to call commit-charge after rmap is established */
3037         mem_cgroup_commit_charge_swapin(page, ptr);
3038
3039         swap_free(entry);
3040         if (vm_swap_full() || (vma->vm_flags & VM_LOCKED) || PageMlocked(page))
3041                 try_to_free_swap(page);
3042         unlock_page(page);
3043         if (swapcache) {
3044                 /*
3045                  * Hold the lock to avoid the swap entry to be reused
3046                  * until we take the PT lock for the pte_same() check
3047                  * (to avoid false positives from pte_same). For
3048                  * further safety release the lock after the swap_free
3049                  * so that the swap count won't change under a
3050                  * parallel locked swapcache.
3051                  */
3052                 unlock_page(swapcache);
3053                 page_cache_release(swapcache);
3054         }
3055
3056         if (flags & FAULT_FLAG_WRITE) {
3057                 ret |= do_wp_page(mm, vma, address, page_table, pmd, ptl, pte);
3058                 if (ret & VM_FAULT_ERROR)
3059                         ret &= VM_FAULT_ERROR;
3060                 goto out;
3061         }
3062
3063         /* No need to invalidate - it was non-present before */
3064         update_mmu_cache(vma, address, page_table);
3065 unlock:
3066         pte_unmap_unlock(page_table, ptl);
3067 out:
3068         return ret;
3069 out_nomap:
3070         mem_cgroup_cancel_charge_swapin(ptr);
3071         pte_unmap_unlock(page_table, ptl);
3072 out_page:
3073         unlock_page(page);
3074 out_release:
3075         page_cache_release(page);
3076         if (swapcache) {
3077                 unlock_page(swapcache);
3078                 page_cache_release(swapcache);
3079         }
3080         return ret;
3081 }
3082
3083 /*
3084  * This is like a special single-page "expand_{down|up}wards()",
3085  * except we must first make sure that 'address{-|+}PAGE_SIZE'
3086  * doesn't hit another vma.
3087  */
3088 static inline int check_stack_guard_page(struct vm_area_struct *vma, unsigned long address)
3089 {
3090         address &= PAGE_MASK;
3091         if ((vma->vm_flags & VM_GROWSDOWN) && address == vma->vm_start) {
3092                 struct vm_area_struct *prev = vma->vm_prev;
3093
3094                 /*
3095                  * Is there a mapping abutting this one below?
3096                  *
3097                  * That's only ok if it's the same stack mapping
3098                  * that has gotten split..
3099                  */
3100                 if (prev && prev->vm_end == address)
3101                         return prev->vm_flags & VM_GROWSDOWN ? 0 : -ENOMEM;
3102
3103                 expand_downwards(vma, address - PAGE_SIZE);
3104         }
3105         if ((vma->vm_flags & VM_GROWSUP) && address + PAGE_SIZE == vma->vm_end) {
3106                 struct vm_area_struct *next = vma->vm_next;
3107
3108                 /* As VM_GROWSDOWN but s/below/above/ */
3109                 if (next && next->vm_start == address + PAGE_SIZE)
3110                         return next->vm_flags & VM_GROWSUP ? 0 : -ENOMEM;
3111
3112                 expand_upwards(vma, address + PAGE_SIZE);
3113         }
3114         return 0;
3115 }
3116
3117 /*
3118  * We enter with non-exclusive mmap_sem (to exclude vma changes,
3119  * but allow concurrent faults), and pte mapped but not yet locked.
3120  * We return with mmap_sem still held, but pte unmapped and unlocked.
3121  */
3122 static int do_anonymous_page(struct mm_struct *mm, struct vm_area_struct *vma,
3123                 unsigned long address, pte_t *page_table, pmd_t *pmd,
3124                 unsigned int flags)
3125 {
3126         struct page *page;
3127         spinlock_t *ptl;
3128         pte_t entry;
3129
3130         pte_unmap(page_table);
3131
3132         /* Check if we need to add a guard page to the stack */
3133         if (check_stack_guard_page(vma, address) < 0)
3134                 return VM_FAULT_SIGBUS;
3135
3136         /* Use the zero-page for reads */
3137         if (!(flags & FAULT_FLAG_WRITE)) {
3138                 entry = pte_mkspecial(pfn_pte(my_zero_pfn(address),
3139                                                 vma->vm_page_prot));
3140                 page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
3141                 if (!pte_none(*page_table))
3142                         goto unlock;
3143                 goto setpte;
3144         }
3145
3146         /* Allocate our own private page. */
3147         if (unlikely(anon_vma_prepare(vma)))
3148                 goto oom;
3149         page = alloc_zeroed_user_highpage_movable(vma, address);
3150         if (!page)
3151                 goto oom;
3152         __SetPageUptodate(page);
3153
3154         if (mem_cgroup_newpage_charge(page, mm, GFP_KERNEL))
3155                 goto oom_free_page;
3156
3157         entry = mk_pte(page, vma->vm_page_prot);
3158         if (vma->vm_flags & VM_WRITE)
3159                 entry = pte_mkwrite(pte_mkdirty(entry));
3160
3161         page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
3162         if (!pte_none(*page_table))
3163                 goto release;
3164
3165         inc_mm_counter_fast(mm, MM_ANONPAGES);
3166         page_add_new_anon_rmap(page, vma, address);
3167 setpte:
3168         set_pte_at(mm, address, page_table, entry);
3169
3170         /* No need to invalidate - it was non-present before */
3171         update_mmu_cache(vma, address, page_table);
3172 unlock:
3173         pte_unmap_unlock(page_table, ptl);
3174         return 0;
3175 release:
3176         mem_cgroup_uncharge_page(page);
3177         page_cache_release(page);
3178         goto unlock;
3179 oom_free_page:
3180         page_cache_release(page);
3181 oom:
3182         return VM_FAULT_OOM;
3183 }
3184
3185 /*
3186  * __do_fault() tries to create a new page mapping. It aggressively
3187  * tries to share with existing pages, but makes a separate copy if
3188  * the FAULT_FLAG_WRITE is set in the flags parameter in order to avoid
3189  * the next page fault.
3190  *
3191  * As this is called only for pages that do not currently exist, we
3192  * do not need to flush old virtual caches or the TLB.
3193  *
3194  * We enter with non-exclusive mmap_sem (to exclude vma changes,
3195  * but allow concurrent faults), and pte neither mapped nor locked.
3196  * We return with mmap_sem still held, but pte unmapped and unlocked.
3197  */
3198 static int __do_fault(struct mm_struct *mm, struct vm_area_struct *vma,
3199                 unsigned long address, pmd_t *pmd,
3200                 pgoff_t pgoff, unsigned int flags, pte_t orig_pte)
3201 {
3202         pte_t *page_table;
3203         spinlock_t *ptl;
3204         struct page *page;
3205         struct page *cow_page;
3206         pte_t entry;
3207         int anon = 0;
3208         struct page *dirty_page = NULL;
3209         struct vm_fault vmf;
3210         int ret;
3211         int page_mkwrite = 0;
3212
3213         /*
3214          * If we do COW later, allocate page befor taking lock_page()
3215          * on the file cache page. This will reduce lock holding time.
3216          */
3217         if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) {
3218
3219                 if (unlikely(anon_vma_prepare(vma)))
3220                         return VM_FAULT_OOM;
3221
3222                 cow_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, address);
3223                 if (!cow_page)
3224                         return VM_FAULT_OOM;
3225
3226                 if (mem_cgroup_newpage_charge(cow_page, mm, GFP_KERNEL)) {
3227                         page_cache_release(cow_page);
3228                         return VM_FAULT_OOM;
3229                 }
3230         } else
3231                 cow_page = NULL;
3232
3233         vmf.virtual_address = (void __user *)(address & PAGE_MASK);
3234         vmf.pgoff = pgoff;
3235         vmf.flags = flags;
3236         vmf.page = NULL;
3237
3238         ret = vma->vm_ops->fault(vma, &vmf);
3239         if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE |
3240                             VM_FAULT_RETRY)))
3241                 goto uncharge_out;
3242
3243         if (unlikely(PageHWPoison(vmf.page))) {
3244                 if (ret & VM_FAULT_LOCKED)
3245                         unlock_page(vmf.page);
3246                 ret = VM_FAULT_HWPOISON;
3247                 goto uncharge_out;
3248         }
3249
3250         /*
3251          * For consistency in subsequent calls, make the faulted page always
3252          * locked.
3253          */
3254         if (unlikely(!(ret & VM_FAULT_LOCKED)))
3255                 lock_page(vmf.page);
3256         else
3257                 VM_BUG_ON(!PageLocked(vmf.page));
3258
3259         /*
3260          * Should we do an early C-O-W break?
3261          */
3262         page = vmf.page;
3263         if (flags & FAULT_FLAG_WRITE) {
3264                 if (!(vma->vm_flags & VM_SHARED)) {
3265                         page = cow_page;
3266                         anon = 1;
3267                         copy_user_highpage(page, vmf.page, address, vma);
3268                         __SetPageUptodate(page);
3269                 } else {
3270                         /*
3271                          * If the page will be shareable, see if the backing
3272                          * address space wants to know that the page is about
3273                          * to become writable
3274                          */
3275                         if (vma->vm_ops->page_mkwrite) {
3276                                 int tmp;
3277
3278                                 unlock_page(page);
3279                                 vmf.flags = FAULT_FLAG_WRITE|FAULT_FLAG_MKWRITE;
3280                                 tmp = vma->vm_ops->page_mkwrite(vma, &vmf);
3281                                 if (unlikely(tmp &
3282                                           (VM_FAULT_ERROR | VM_FAULT_NOPAGE))) {
3283                                         ret = tmp;
3284                                         goto unwritable_page;
3285                                 }
3286                                 if (unlikely(!(tmp & VM_FAULT_LOCKED))) {
3287                                         lock_page(page);
3288                                         if (!page->mapping) {
3289                                                 ret = 0; /* retry the fault */
3290                                                 unlock_page(page);
3291                                                 goto unwritable_page;
3292                                         }
3293                                 } else
3294                                         VM_BUG_ON(!PageLocked(page));
3295                                 page_mkwrite = 1;
3296                         }
3297                 }
3298
3299         }
3300
3301         page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
3302
3303         /*
3304          * This silly early PAGE_DIRTY setting removes a race
3305          * due to the bad i386 page protection. But it's valid
3306          * for other architectures too.
3307          *
3308          * Note that if FAULT_FLAG_WRITE is set, we either now have
3309          * an exclusive copy of the page, or this is a shared mapping,
3310          * so we can make it writable and dirty to avoid having to
3311          * handle that later.
3312          */
3313         /* Only go through if we didn't race with anybody else... */
3314         if (likely(pte_same(*page_table, orig_pte))) {
3315                 flush_icache_page(vma, page);
3316                 entry = mk_pte(page, vma->vm_page_prot);
3317                 if (flags & FAULT_FLAG_WRITE)
3318                         entry = maybe_mkwrite(pte_mkdirty(entry), vma);
3319                 if (anon) {
3320                         inc_mm_counter_fast(mm, MM_ANONPAGES);
3321                         page_add_new_anon_rmap(page, vma, address);
3322                 } else {
3323                         inc_mm_counter_fast(mm, MM_FILEPAGES);
3324                         page_add_file_rmap(page);
3325                         if (flags & FAULT_FLAG_WRITE) {
3326                                 dirty_page = page;
3327                                 get_page(dirty_page);
3328                         }
3329                 }
3330                 set_pte_at(mm, address, page_table, entry);
3331
3332                 /* no need to invalidate: a not-present page won't be cached */
3333                 update_mmu_cache(vma, address, page_table);
3334         } else {
3335                 if (cow_page)
3336                         mem_cgroup_uncharge_page(cow_page);
3337                 if (anon)
3338                         page_cache_release(page);
3339                 else
3340                         anon = 1; /* no anon but release faulted_page */
3341         }
3342
3343         pte_unmap_unlock(page_table, ptl);
3344
3345         if (dirty_page) {
3346                 struct address_space *mapping = page->mapping;
3347                 int dirtied = 0;
3348
3349                 if (set_page_dirty(dirty_page))
3350                         dirtied = 1;
3351                 unlock_page(dirty_page);
3352                 put_page(dirty_page);
3353                 if ((dirtied || page_mkwrite) && mapping) {
3354                         /*
3355                          * Some device drivers do not set page.mapping but still
3356                          * dirty their pages
3357                          */
3358                         balance_dirty_pages_ratelimited(mapping);
3359                 }
3360
3361                 /* file_update_time outside page_lock */
3362                 if (vma->vm_file && !page_mkwrite)
3363                         file_update_time(vma->vm_file);
3364         } else {
3365                 unlock_page(vmf.page);
3366                 if (anon)
3367                         page_cache_release(vmf.page);
3368         }
3369
3370         return ret;
3371
3372 unwritable_page:
3373         page_cache_release(page);
3374         return ret;
3375 uncharge_out:
3376         /* fs's fault handler get error */
3377         if (cow_page) {
3378                 mem_cgroup_uncharge_page(cow_page);
3379                 page_cache_release(cow_page);
3380         }
3381         return ret;
3382 }
3383
3384 static int do_linear_fault(struct mm_struct *mm, struct vm_area_struct *vma,
3385                 unsigned long address, pte_t *page_table, pmd_t *pmd,
3386                 unsigned int flags, pte_t orig_pte)
3387 {
3388         pgoff_t pgoff = (((address & PAGE_MASK)
3389                         - vma->vm_start) >> PAGE_SHIFT) + vma->vm_pgoff;
3390
3391         pte_unmap(page_table);
3392         return __do_fault(mm, vma, address, pmd, pgoff, flags, orig_pte);
3393 }
3394
3395 /*
3396  * Fault of a previously existing named mapping. Repopulate the pte
3397  * from the encoded file_pte if possible. This enables swappable
3398  * nonlinear vmas.
3399  *
3400  * We enter with non-exclusive mmap_sem (to exclude vma changes,
3401  * but allow concurrent faults), and pte mapped but not yet locked.
3402  * We return with mmap_sem still held, but pte unmapped and unlocked.
3403  */
3404 static int do_nonlinear_fault(struct mm_struct *mm, struct vm_area_struct *vma,
3405                 unsigned long address, pte_t *page_table, pmd_t *pmd,
3406                 unsigned int flags, pte_t orig_pte)
3407 {
3408         pgoff_t pgoff;
3409
3410         flags |= FAULT_FLAG_NONLINEAR;
3411
3412         if (!pte_unmap_same(mm, pmd, page_table, orig_pte))
3413                 return 0;
3414
3415         if (unlikely(!(vma->vm_flags & VM_NONLINEAR))) {
3416                 /*
3417                  * Page table corrupted: show pte and kill process.
3418                  */
3419                 print_bad_pte(vma, address, orig_pte, NULL);
3420                 return VM_FAULT_SIGBUS;
3421         }
3422
3423         pgoff = pte_to_pgoff(orig_pte);
3424         return __do_fault(mm, vma, address, pmd, pgoff, flags, orig_pte);
3425 }
3426
3427 /*
3428  * These routines also need to handle stuff like marking pages dirty
3429  * and/or accessed for architectures that don't do it in hardware (most
3430  * RISC architectures).  The early dirtying is also good on the i386.
3431  *
3432  * There is also a hook called "update_mmu_cache()" that architectures
3433  * with external mmu caches can use to update those (ie the Sparc or
3434  * PowerPC hashed page tables that act as extended TLBs).
3435  *
3436  * We enter with non-exclusive mmap_sem (to exclude vma changes,
3437  * but allow concurrent faults), and pte mapped but not yet locked.
3438  * We return with mmap_sem still held, but pte unmapped and unlocked.
3439  */
3440 int handle_pte_fault(struct mm_struct *mm,
3441                      struct vm_area_struct *vma, unsigned long address,
3442                      pte_t *pte, pmd_t *pmd, unsigned int flags)
3443 {
3444         pte_t entry;
3445         spinlock_t *ptl;
3446
3447         entry = *pte;
3448         if (!pte_present(entry)) {
3449                 if (pte_none(entry)) {
3450                         if (vma->vm_ops) {
3451                                 if (likely(vma->vm_ops->fault))
3452                                         return do_linear_fault(mm, vma, address,
3453                                                 pte, pmd, flags, entry);
3454                         }
3455                         return do_anonymous_page(mm, vma, address,
3456                                                  pte, pmd, flags);
3457                 }
3458                 if (pte_file(entry))
3459                         return do_nonlinear_fault(mm, vma, address,
3460                                         pte, pmd, flags, entry);
3461                 return do_swap_page(mm, vma, address,
3462                                         pte, pmd, flags, entry);
3463         }
3464
3465         ptl = pte_lockptr(mm, pmd);
3466         spin_lock(ptl);
3467         if (unlikely(!pte_same(*pte, entry)))
3468                 goto unlock;
3469         if (flags & FAULT_FLAG_WRITE) {
3470                 if (!pte_write(entry))
3471                         return do_wp_page(mm, vma, address,
3472                                         pte, pmd, ptl, entry);
3473                 entry = pte_mkdirty(entry);
3474         }
3475         entry = pte_mkyoung(entry);
3476         if (ptep_set_access_flags(vma, address, pte, entry, flags & FAULT_FLAG_WRITE)) {
3477                 update_mmu_cache(vma, address, pte);
3478         } else {
3479                 /*
3480                  * This is needed only for protection faults but the arch code
3481                  * is not yet telling us if this is a protection fault or not.
3482                  * This still avoids useless tlb flushes for .text page faults
3483                  * with threads.
3484                  */
3485                 if (flags & FAULT_FLAG_WRITE)
3486                         flush_tlb_fix_spurious_fault(vma, address);
3487         }
3488 unlock:
3489         pte_unmap_unlock(pte, ptl);
3490         return 0;
3491 }
3492
3493 /*
3494  * By the time we get here, we already hold the mm semaphore
3495  */
3496 int handle_mm_fault(struct mm_struct *mm, struct vm_area_struct *vma,
3497                 unsigned long address, unsigned int flags)
3498 {
3499         pgd_t *pgd;
3500         pud_t *pud;
3501         pmd_t *pmd;
3502         pte_t *pte;
3503
3504         __set_current_state(TASK_RUNNING);
3505
3506         count_vm_event(PGFAULT);
3507         mem_cgroup_count_vm_event(mm, PGFAULT);
3508
3509         /* do counter updates before entering really critical section. */
3510         check_sync_rss_stat(current);
3511
3512         if (unlikely(is_vm_hugetlb_page(vma)))
3513                 return hugetlb_fault(mm, vma, address, flags);
3514
3515 retry:
3516         pgd = pgd_offset(mm, address);
3517         pud = pud_alloc(mm, pgd, address);
3518         if (!pud)
3519                 return VM_FAULT_OOM;
3520         pmd = pmd_alloc(mm, pud, address);
3521         if (!pmd)
3522                 return VM_FAULT_OOM;
3523         if (pmd_none(*pmd) && transparent_hugepage_enabled(vma)) {
3524                 if (!vma->vm_ops)
3525                         return do_huge_pmd_anonymous_page(mm, vma, address,
3526                                                           pmd, flags);
3527         } else {
3528                 pmd_t orig_pmd = *pmd;
3529                 int ret;
3530
3531                 barrier();
3532                 if (pmd_trans_huge(orig_pmd)) {
3533                         unsigned int dirty = flags & FAULT_FLAG_WRITE;
3534
3535                         if (dirty && !pmd_write(orig_pmd) &&
3536                             !pmd_trans_splitting(orig_pmd)) {
3537                                 ret = do_huge_pmd_wp_page(mm, vma, address, pmd,
3538                                                           orig_pmd);
3539                                 /*
3540                                  * If COW results in an oom, the huge pmd will
3541                                  * have been split, so retry the fault on the
3542                                  * pte for a smaller charge.
3543                                  */
3544                                 if (unlikely(ret & VM_FAULT_OOM))
3545                                         goto retry;
3546                                 return ret;
3547                         } else {
3548                                 huge_pmd_set_accessed(mm, vma, address, pmd,
3549                                                       orig_pmd, dirty);
3550                         }
3551                         return 0;
3552                 }
3553         }
3554
3555         /*
3556          * Use __pte_alloc instead of pte_alloc_map, because we can't
3557          * run pte_offset_map on the pmd, if an huge pmd could
3558          * materialize from under us from a different thread.
3559          */
3560         if (unlikely(pmd_none(*pmd)) && __pte_alloc(mm, vma, pmd, address))
3561                 return VM_FAULT_OOM;
3562         /* if an huge pmd materialized from under us just retry later */
3563         if (unlikely(pmd_trans_huge(*pmd)))
3564                 return 0;
3565         /*
3566          * A regular pmd is established and it can't morph into a huge pmd
3567          * from under us anymore at this point because we hold the mmap_sem
3568          * read mode and khugepaged takes it in write mode. So now it's
3569          * safe to run pte_offset_map().
3570          */
3571         pte = pte_offset_map(pmd, address);
3572
3573         return handle_pte_fault(mm, vma, address, pte, pmd, flags);
3574 }
3575
3576 #ifndef __PAGETABLE_PUD_FOLDED
3577 /*
3578  * Allocate page upper directory.
3579  * We've already handled the fast-path in-line.
3580  */
3581 int __pud_alloc(struct mm_struct *mm, pgd_t *pgd, unsigned long address)
3582 {
3583         pud_t *new = pud_alloc_one(mm, address);
3584         if (!new)
3585                 return -ENOMEM;
3586
3587         smp_wmb(); /* See comment in __pte_alloc */
3588
3589         spin_lock(&mm->page_table_lock);
3590         if (pgd_present(*pgd))          /* Another has populated it */
3591                 pud_free(mm, new);
3592         else
3593                 pgd_populate(mm, pgd, new);
3594         spin_unlock(&mm->page_table_lock);
3595         return 0;
3596 }
3597 #endif /* __PAGETABLE_PUD_FOLDED */
3598
3599 #ifndef __PAGETABLE_PMD_FOLDED
3600 /*
3601  * Allocate page middle directory.
3602  * We've already handled the fast-path in-line.
3603  */
3604 int __pmd_alloc(struct mm_struct *mm, pud_t *pud, unsigned long address)
3605 {
3606         pmd_t *new = pmd_alloc_one(mm, address);
3607         if (!new)
3608                 return -ENOMEM;
3609
3610         smp_wmb(); /* See comment in __pte_alloc */
3611
3612         spin_lock(&mm->page_table_lock);
3613 #ifndef __ARCH_HAS_4LEVEL_HACK
3614         if (pud_present(*pud))          /* Another has populated it */
3615                 pmd_free(mm, new);
3616         else
3617                 pud_populate(mm, pud, new);
3618 #else
3619         if (pgd_present(*pud))          /* Another has populated it */
3620                 pmd_free(mm, new);
3621         else
3622                 pgd_populate(mm, pud, new);
3623 #endif /* __ARCH_HAS_4LEVEL_HACK */
3624         spin_unlock(&mm->page_table_lock);
3625         return 0;
3626 }
3627 #endif /* __PAGETABLE_PMD_FOLDED */
3628
3629 int make_pages_present(unsigned long addr, unsigned long end)
3630 {
3631         int ret, len, write;
3632         struct vm_area_struct * vma;
3633
3634         vma = find_vma(current->mm, addr);
3635         if (!vma)
3636                 return -ENOMEM;
3637         /*
3638          * We want to touch writable mappings with a write fault in order
3639          * to break COW, except for shared mappings because these don't COW
3640          * and we would not want to dirty them for nothing.
3641          */
3642         write = (vma->vm_flags & (VM_WRITE | VM_SHARED)) == VM_WRITE;
3643         BUG_ON(addr >= end);
3644         BUG_ON(end > vma->vm_end);
3645         len = DIV_ROUND_UP(end, PAGE_SIZE) - addr/PAGE_SIZE;
3646         ret = get_user_pages(current, current->mm, addr,
3647                         len, write, 0, NULL, NULL);
3648         if (ret < 0)
3649                 return ret;
3650         return ret == len ? 0 : -EFAULT;
3651 }
3652
3653 #if !defined(__HAVE_ARCH_GATE_AREA)
3654
3655 #if defined(AT_SYSINFO_EHDR)
3656 static struct vm_area_struct gate_vma;
3657
3658 static int __init gate_vma_init(void)
3659 {
3660         gate_vma.vm_mm = NULL;
3661         gate_vma.vm_start = FIXADDR_USER_START;
3662         gate_vma.vm_end = FIXADDR_USER_END;
3663         gate_vma.vm_flags = VM_READ | VM_MAYREAD | VM_EXEC | VM_MAYEXEC;
3664         gate_vma.vm_page_prot = __P101;
3665
3666         return 0;
3667 }
3668 __initcall(gate_vma_init);
3669 #endif
3670
3671 struct vm_area_struct *get_gate_vma(struct mm_struct *mm)
3672 {
3673 #ifdef AT_SYSINFO_EHDR
3674         return &gate_vma;
3675 #else
3676         return NULL;
3677 #endif
3678 }
3679
3680 int in_gate_area_no_mm(unsigned long addr)
3681 {
3682 #ifdef AT_SYSINFO_EHDR
3683         if ((addr >= FIXADDR_USER_START) && (addr < FIXADDR_USER_END))
3684                 return 1;
3685 #endif
3686         return 0;
3687 }
3688
3689 #endif  /* __HAVE_ARCH_GATE_AREA */
3690
3691 static int __follow_pte(struct mm_struct *mm, unsigned long address,
3692                 pte_t **ptepp, spinlock_t **ptlp)
3693 {
3694         pgd_t *pgd;
3695         pud_t *pud;
3696         pmd_t *pmd;
3697         pte_t *ptep;
3698
3699         pgd = pgd_offset(mm, address);
3700         if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
3701                 goto out;
3702
3703         pud = pud_offset(pgd, address);
3704         if (pud_none(*pud) || unlikely(pud_bad(*pud)))
3705                 goto out;
3706
3707         pmd = pmd_offset(pud, address);
3708         VM_BUG_ON(pmd_trans_huge(*pmd));
3709         if (pmd_none(*pmd) || unlikely(pmd_bad(*pmd)))
3710                 goto out;
3711
3712         /* We cannot handle huge page PFN maps. Luckily they don't exist. */
3713         if (pmd_huge(*pmd))
3714                 goto out;
3715
3716         ptep = pte_offset_map_lock(mm, pmd, address, ptlp);
3717         if (!ptep)
3718                 goto out;
3719         if (!pte_present(*ptep))
3720                 goto unlock;
3721         *ptepp = ptep;
3722         return 0;
3723 unlock:
3724         pte_unmap_unlock(ptep, *ptlp);
3725 out:
3726         return -EINVAL;
3727 }
3728
3729 static inline int follow_pte(struct mm_struct *mm, unsigned long address,
3730                              pte_t **ptepp, spinlock_t **ptlp)
3731 {
3732         int res;
3733
3734         /* (void) is needed to make gcc happy */
3735         (void) __cond_lock(*ptlp,
3736                            !(res = __follow_pte(mm, address, ptepp, ptlp)));
3737         return res;
3738 }
3739
3740 /**
3741  * follow_pfn - look up PFN at a user virtual address
3742  * @vma: memory mapping
3743  * @address: user virtual address
3744  * @pfn: location to store found PFN
3745  *
3746  * Only IO mappings and raw PFN mappings are allowed.
3747  *
3748  * Returns zero and the pfn at @pfn on success, -ve otherwise.
3749  */
3750 int follow_pfn(struct vm_area_struct *vma, unsigned long address,
3751         unsigned long *pfn)
3752 {
3753         int ret = -EINVAL;
3754         spinlock_t *ptl;
3755         pte_t *ptep;
3756
3757         if (!(vma->vm_flags & (VM_IO | VM_PFNMAP)))
3758                 return ret;
3759
3760         ret = follow_pte(vma->vm_mm, address, &ptep, &ptl);
3761         if (ret)
3762                 return ret;
3763         *pfn = pte_pfn(*ptep);
3764         pte_unmap_unlock(ptep, ptl);
3765         return 0;
3766 }
3767 EXPORT_SYMBOL(follow_pfn);
3768
3769 #ifdef CONFIG_HAVE_IOREMAP_PROT
3770 int follow_phys(struct vm_area_struct *vma,
3771                 unsigned long address, unsigned int flags,
3772                 unsigned long *prot, resource_size_t *phys)
3773 {
3774         int ret = -EINVAL;
3775         pte_t *ptep, pte;
3776         spinlock_t *ptl;
3777
3778         if (!(vma->vm_flags & (VM_IO | VM_PFNMAP)))
3779                 goto out;
3780
3781         if (follow_pte(vma->vm_mm, address, &ptep, &ptl))
3782                 goto out;
3783         pte = *ptep;
3784
3785         if ((flags & FOLL_WRITE) && !pte_write(pte))
3786                 goto unlock;
3787
3788         *prot = pgprot_val(pte_pgprot(pte));
3789         *phys = (resource_size_t)pte_pfn(pte) << PAGE_SHIFT;
3790
3791         ret = 0;
3792 unlock:
3793         pte_unmap_unlock(ptep, ptl);
3794 out:
3795         return ret;
3796 }
3797
3798 int generic_access_phys(struct vm_area_struct *vma, unsigned long addr,
3799                         void *buf, int len, int write)
3800 {
3801         resource_size_t phys_addr;
3802         unsigned long prot = 0;
3803         void __iomem *maddr;
3804         int offset = addr & (PAGE_SIZE-1);
3805
3806         if (follow_phys(vma, addr, write, &prot, &phys_addr))
3807                 return -EINVAL;
3808
3809         maddr = ioremap_prot(phys_addr, PAGE_SIZE, prot);
3810         if (write)
3811                 memcpy_toio(maddr + offset, buf, len);
3812         else
3813                 memcpy_fromio(buf, maddr + offset, len);
3814         iounmap(maddr);
3815
3816         return len;
3817 }
3818 #endif
3819
3820 /*
3821  * Access another process' address space as given in mm.  If non-NULL, use the
3822  * given task for page fault accounting.
3823  */
3824 static int __access_remote_vm(struct task_struct *tsk, struct mm_struct *mm,
3825                 unsigned long addr, void *buf, int len, int write)
3826 {
3827         struct vm_area_struct *vma;
3828         void *old_buf = buf;
3829
3830         down_read(&mm->mmap_sem);
3831         /* ignore errors, just check how much was successfully transferred */
3832         while (len) {
3833                 int bytes, ret, offset;
3834                 void *maddr;
3835                 struct page *page = NULL;
3836
3837                 ret = get_user_pages(tsk, mm, addr, 1,
3838                                 write, 1, &page, &vma);
3839                 if (ret <= 0) {
3840                         /*
3841                          * Check if this is a VM_IO | VM_PFNMAP VMA, which
3842                          * we can access using slightly different code.
3843                          */
3844 #ifdef CONFIG_HAVE_IOREMAP_PROT
3845                         vma = find_vma(mm, addr);
3846                         if (!vma || vma->vm_start > addr)
3847                                 break;
3848                         if (vma->vm_ops && vma->vm_ops->access)
3849                                 ret = vma->vm_ops->access(vma, addr, buf,
3850                                                           len, write);
3851                         if (ret <= 0)
3852 #endif
3853                                 break;
3854                         bytes = ret;
3855                 } else {
3856                         bytes = len;
3857                         offset = addr & (PAGE_SIZE-1);
3858                         if (bytes > PAGE_SIZE-offset)
3859                                 bytes = PAGE_SIZE-offset;
3860
3861                         maddr = kmap(page);
3862                         if (write) {
3863                                 copy_to_user_page(vma, page, addr,
3864                                                   maddr + offset, buf, bytes);
3865                                 set_page_dirty_lock(page);
3866                         } else {
3867                                 copy_from_user_page(vma, page, addr,
3868                                                     buf, maddr + offset, bytes);
3869                         }
3870                         kunmap(page);
3871                         page_cache_release(page);
3872                 }
3873                 len -= bytes;
3874                 buf += bytes;
3875                 addr += bytes;
3876         }
3877         up_read(&mm->mmap_sem);
3878
3879         return buf - old_buf;
3880 }
3881
3882 /**
3883  * access_remote_vm - access another process' address space
3884  * @mm:         the mm_struct of the target address space
3885  * @addr:       start address to access
3886  * @buf:        source or destination buffer
3887  * @len:        number of bytes to transfer
3888  * @write:      whether the access is a write
3889  *
3890  * The caller must hold a reference on @mm.
3891  */
3892 int access_remote_vm(struct mm_struct *mm, unsigned long addr,
3893                 void *buf, int len, int write)
3894 {
3895         return __access_remote_vm(NULL, mm, addr, buf, len, write);
3896 }
3897
3898 /*
3899  * Access another process' address space.
3900  * Source/target buffer must be kernel space,
3901  * Do not walk the page table directly, use get_user_pages
3902  */
3903 int access_process_vm(struct task_struct *tsk, unsigned long addr,
3904                 void *buf, int len, int write)
3905 {
3906         struct mm_struct *mm;
3907         int ret;
3908
3909         mm = get_task_mm(tsk);
3910         if (!mm)
3911                 return 0;
3912
3913         ret = __access_remote_vm(tsk, mm, addr, buf, len, write);
3914         mmput(mm);
3915
3916         return ret;
3917 }
3918
3919 /*
3920  * Print the name of a VMA.
3921  */
3922 void print_vma_addr(char *prefix, unsigned long ip)
3923 {
3924         struct mm_struct *mm = current->mm;
3925         struct vm_area_struct *vma;
3926
3927         /*
3928          * Do not print if we are in atomic
3929          * contexts (in exception stacks, etc.):
3930          */
3931         if (preempt_count())
3932                 return;
3933
3934         down_read(&mm->mmap_sem);
3935         vma = find_vma(mm, ip);
3936         if (vma && vma->vm_file) {
3937                 struct file *f = vma->vm_file;
3938                 char *buf = (char *)__get_free_page(GFP_KERNEL);
3939                 if (buf) {
3940                         char *p, *s;
3941
3942                         p = d_path(&f->f_path, buf, PAGE_SIZE);
3943                         if (IS_ERR(p))
3944                                 p = "?";
3945                         s = strrchr(p, '/');
3946                         if (s)
3947                                 p = s+1;
3948                         printk("%s%s[%lx+%lx]", prefix, p,
3949                                         vma->vm_start,
3950                                         vma->vm_end - vma->vm_start);
3951                         free_page((unsigned long)buf);
3952                 }
3953         }
3954         up_read(&mm->mmap_sem);
3955 }
3956
3957 #ifdef CONFIG_PROVE_LOCKING
3958 void might_fault(void)
3959 {
3960         /*
3961          * Some code (nfs/sunrpc) uses socket ops on kernel memory while
3962          * holding the mmap_sem, this is safe because kernel memory doesn't
3963          * get paged out, therefore we'll never actually fault, and the
3964          * below annotations will generate false positives.
3965          */
3966         if (segment_eq(get_fs(), KERNEL_DS))
3967                 return;
3968
3969         might_sleep();
3970         /*
3971          * it would be nicer only to annotate paths which are not under
3972          * pagefault_disable, however that requires a larger audit and
3973          * providing helpers like get_user_atomic.
3974          */
3975         if (!in_atomic() && current->mm)
3976                 might_lock_read(&current->mm->mmap_sem);
3977 }
3978 EXPORT_SYMBOL(might_fault);
3979 #endif
3980
3981 #if defined(CONFIG_TRANSPARENT_HUGEPAGE) || defined(CONFIG_HUGETLBFS)
3982 static void clear_gigantic_page(struct page *page,
3983                                 unsigned long addr,
3984                                 unsigned int pages_per_huge_page)
3985 {
3986         int i;
3987         struct page *p = page;
3988
3989         might_sleep();
3990         for (i = 0; i < pages_per_huge_page;
3991              i++, p = mem_map_next(p, page, i)) {
3992                 cond_resched();
3993                 clear_user_highpage(p, addr + i * PAGE_SIZE);
3994         }
3995 }
3996 void clear_huge_page(struct page *page,
3997                      unsigned long addr, unsigned int pages_per_huge_page)
3998 {
3999         int i;
4000
4001         if (unlikely(pages_per_huge_page > MAX_ORDER_NR_PAGES)) {
4002                 clear_gigantic_page(page, addr, pages_per_huge_page);
4003                 return;
4004         }
4005
4006         might_sleep();
4007         for (i = 0; i < pages_per_huge_page; i++) {
4008                 cond_resched();
4009                 clear_user_highpage(page + i, addr + i * PAGE_SIZE);
4010         }
4011 }
4012
4013 static void copy_user_gigantic_page(struct page *dst, struct page *src,
4014                                     unsigned long addr,
4015                                     struct vm_area_struct *vma,
4016                                     unsigned int pages_per_huge_page)
4017 {
4018         int i;
4019         struct page *dst_base = dst;
4020         struct page *src_base = src;
4021
4022         for (i = 0; i < pages_per_huge_page; ) {
4023                 cond_resched();
4024                 copy_user_highpage(dst, src, addr + i*PAGE_SIZE, vma);
4025
4026                 i++;
4027                 dst = mem_map_next(dst, dst_base, i);
4028                 src = mem_map_next(src, src_base, i);
4029         }
4030 }
4031
4032 void copy_user_huge_page(struct page *dst, struct page *src,
4033                          unsigned long addr, struct vm_area_struct *vma,
4034                          unsigned int pages_per_huge_page)
4035 {
4036         int i;
4037
4038         if (unlikely(pages_per_huge_page > MAX_ORDER_NR_PAGES)) {
4039                 copy_user_gigantic_page(dst, src, addr, vma,
4040                                         pages_per_huge_page);
4041                 return;
4042         }
4043
4044         might_sleep();
4045         for (i = 0; i < pages_per_huge_page; i++) {
4046                 cond_resched();
4047                 copy_user_highpage(dst + i, src + i, addr + i*PAGE_SIZE, vma);
4048         }
4049 }
4050 #endif /* CONFIG_TRANSPARENT_HUGEPAGE || CONFIG_HUGETLBFS */