mm: kill vma flag VM_INSERTPAGE
[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 int is_cow_mapping(vm_flags_t flags)
716 {
717         return (flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE;
718 }
719
720 #ifndef is_zero_pfn
721 static inline int is_zero_pfn(unsigned long pfn)
722 {
723         return pfn == zero_pfn;
724 }
725 #endif
726
727 #ifndef my_zero_pfn
728 static inline unsigned long my_zero_pfn(unsigned long addr)
729 {
730         return zero_pfn;
731 }
732 #endif
733
734 /*
735  * vm_normal_page -- This function gets the "struct page" associated with a pte.
736  *
737  * "Special" mappings do not wish to be associated with a "struct page" (either
738  * it doesn't exist, or it exists but they don't want to touch it). In this
739  * case, NULL is returned here. "Normal" mappings do have a struct page.
740  *
741  * There are 2 broad cases. Firstly, an architecture may define a pte_special()
742  * pte bit, in which case this function is trivial. Secondly, an architecture
743  * may not have a spare pte bit, which requires a more complicated scheme,
744  * described below.
745  *
746  * A raw VM_PFNMAP mapping (ie. one that is not COWed) is always considered a
747  * special mapping (even if there are underlying and valid "struct pages").
748  * COWed pages of a VM_PFNMAP are always normal.
749  *
750  * The way we recognize COWed pages within VM_PFNMAP mappings is through the
751  * rules set up by "remap_pfn_range()": the vma will have the VM_PFNMAP bit
752  * set, and the vm_pgoff will point to the first PFN mapped: thus every special
753  * mapping will always honor the rule
754  *
755  *      pfn_of_page == vma->vm_pgoff + ((addr - vma->vm_start) >> PAGE_SHIFT)
756  *
757  * And for normal mappings this is false.
758  *
759  * This restricts such mappings to be a linear translation from virtual address
760  * to pfn. To get around this restriction, we allow arbitrary mappings so long
761  * as the vma is not a COW mapping; in that case, we know that all ptes are
762  * special (because none can have been COWed).
763  *
764  *
765  * In order to support COW of arbitrary special mappings, we have VM_MIXEDMAP.
766  *
767  * VM_MIXEDMAP mappings can likewise contain memory with or without "struct
768  * page" backing, however the difference is that _all_ pages with a struct
769  * page (that is, those where pfn_valid is true) are refcounted and considered
770  * normal pages by the VM. The disadvantage is that pages are refcounted
771  * (which can be slower and simply not an option for some PFNMAP users). The
772  * advantage is that we don't have to follow the strict linearity rule of
773  * PFNMAP mappings in order to support COWable mappings.
774  *
775  */
776 #ifdef __HAVE_ARCH_PTE_SPECIAL
777 # define HAVE_PTE_SPECIAL 1
778 #else
779 # define HAVE_PTE_SPECIAL 0
780 #endif
781 struct page *vm_normal_page(struct vm_area_struct *vma, unsigned long addr,
782                                 pte_t pte)
783 {
784         unsigned long pfn = pte_pfn(pte);
785
786         if (HAVE_PTE_SPECIAL) {
787                 if (likely(!pte_special(pte)))
788                         goto check_pfn;
789                 if (vma->vm_flags & (VM_PFNMAP | VM_MIXEDMAP))
790                         return NULL;
791                 if (!is_zero_pfn(pfn))
792                         print_bad_pte(vma, addr, pte, NULL);
793                 return NULL;
794         }
795
796         /* !HAVE_PTE_SPECIAL case follows: */
797
798         if (unlikely(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP))) {
799                 if (vma->vm_flags & VM_MIXEDMAP) {
800                         if (!pfn_valid(pfn))
801                                 return NULL;
802                         goto out;
803                 } else {
804                         unsigned long off;
805                         off = (addr - vma->vm_start) >> PAGE_SHIFT;
806                         if (pfn == vma->vm_pgoff + off)
807                                 return NULL;
808                         if (!is_cow_mapping(vma->vm_flags))
809                                 return NULL;
810                 }
811         }
812
813         if (is_zero_pfn(pfn))
814                 return NULL;
815 check_pfn:
816         if (unlikely(pfn > highest_memmap_pfn)) {
817                 print_bad_pte(vma, addr, pte, NULL);
818                 return NULL;
819         }
820
821         /*
822          * NOTE! We still have PageReserved() pages in the page tables.
823          * eg. VDSO mappings can cause them to exist.
824          */
825 out:
826         return pfn_to_page(pfn);
827 }
828
829 /*
830  * copy one vm_area from one task to the other. Assumes the page tables
831  * already present in the new task to be cleared in the whole range
832  * covered by this vma.
833  */
834
835 static inline unsigned long
836 copy_one_pte(struct mm_struct *dst_mm, struct mm_struct *src_mm,
837                 pte_t *dst_pte, pte_t *src_pte, struct vm_area_struct *vma,
838                 unsigned long addr, int *rss)
839 {
840         unsigned long vm_flags = vma->vm_flags;
841         pte_t pte = *src_pte;
842         struct page *page;
843
844         /* pte contains position in swap or file, so copy. */
845         if (unlikely(!pte_present(pte))) {
846                 if (!pte_file(pte)) {
847                         swp_entry_t entry = pte_to_swp_entry(pte);
848
849                         if (swap_duplicate(entry) < 0)
850                                 return entry.val;
851
852                         /* make sure dst_mm is on swapoff's mmlist. */
853                         if (unlikely(list_empty(&dst_mm->mmlist))) {
854                                 spin_lock(&mmlist_lock);
855                                 if (list_empty(&dst_mm->mmlist))
856                                         list_add(&dst_mm->mmlist,
857                                                  &src_mm->mmlist);
858                                 spin_unlock(&mmlist_lock);
859                         }
860                         if (likely(!non_swap_entry(entry)))
861                                 rss[MM_SWAPENTS]++;
862                         else if (is_migration_entry(entry)) {
863                                 page = migration_entry_to_page(entry);
864
865                                 if (PageAnon(page))
866                                         rss[MM_ANONPAGES]++;
867                                 else
868                                         rss[MM_FILEPAGES]++;
869
870                                 if (is_write_migration_entry(entry) &&
871                                     is_cow_mapping(vm_flags)) {
872                                         /*
873                                          * COW mappings require pages in both
874                                          * parent and child to be set to read.
875                                          */
876                                         make_migration_entry_read(&entry);
877                                         pte = swp_entry_to_pte(entry);
878                                         set_pte_at(src_mm, addr, src_pte, pte);
879                                 }
880                         }
881                 }
882                 goto out_set_pte;
883         }
884
885         /*
886          * If it's a COW mapping, write protect it both
887          * in the parent and the child
888          */
889         if (is_cow_mapping(vm_flags)) {
890                 ptep_set_wrprotect(src_mm, addr, src_pte);
891                 pte = pte_wrprotect(pte);
892         }
893
894         /*
895          * If it's a shared mapping, mark it clean in
896          * the child
897          */
898         if (vm_flags & VM_SHARED)
899                 pte = pte_mkclean(pte);
900         pte = pte_mkold(pte);
901
902         page = vm_normal_page(vma, addr, pte);
903         if (page) {
904                 get_page(page);
905                 page_dup_rmap(page);
906                 if (PageAnon(page))
907                         rss[MM_ANONPAGES]++;
908                 else
909                         rss[MM_FILEPAGES]++;
910         }
911
912 out_set_pte:
913         set_pte_at(dst_mm, addr, dst_pte, pte);
914         return 0;
915 }
916
917 int copy_pte_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
918                    pmd_t *dst_pmd, pmd_t *src_pmd, struct vm_area_struct *vma,
919                    unsigned long addr, unsigned long end)
920 {
921         pte_t *orig_src_pte, *orig_dst_pte;
922         pte_t *src_pte, *dst_pte;
923         spinlock_t *src_ptl, *dst_ptl;
924         int progress = 0;
925         int rss[NR_MM_COUNTERS];
926         swp_entry_t entry = (swp_entry_t){0};
927
928 again:
929         init_rss_vec(rss);
930
931         dst_pte = pte_alloc_map_lock(dst_mm, dst_pmd, addr, &dst_ptl);
932         if (!dst_pte)
933                 return -ENOMEM;
934         src_pte = pte_offset_map(src_pmd, addr);
935         src_ptl = pte_lockptr(src_mm, src_pmd);
936         spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
937         orig_src_pte = src_pte;
938         orig_dst_pte = dst_pte;
939         arch_enter_lazy_mmu_mode();
940
941         do {
942                 /*
943                  * We are holding two locks at this point - either of them
944                  * could generate latencies in another task on another CPU.
945                  */
946                 if (progress >= 32) {
947                         progress = 0;
948                         if (need_resched() ||
949                             spin_needbreak(src_ptl) || spin_needbreak(dst_ptl))
950                                 break;
951                 }
952                 if (pte_none(*src_pte)) {
953                         progress++;
954                         continue;
955                 }
956                 entry.val = copy_one_pte(dst_mm, src_mm, dst_pte, src_pte,
957                                                         vma, addr, rss);
958                 if (entry.val)
959                         break;
960                 progress += 8;
961         } while (dst_pte++, src_pte++, addr += PAGE_SIZE, addr != end);
962
963         arch_leave_lazy_mmu_mode();
964         spin_unlock(src_ptl);
965         pte_unmap(orig_src_pte);
966         add_mm_rss_vec(dst_mm, rss);
967         pte_unmap_unlock(orig_dst_pte, dst_ptl);
968         cond_resched();
969
970         if (entry.val) {
971                 if (add_swap_count_continuation(entry, GFP_KERNEL) < 0)
972                         return -ENOMEM;
973                 progress = 0;
974         }
975         if (addr != end)
976                 goto again;
977         return 0;
978 }
979
980 static inline int copy_pmd_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
981                 pud_t *dst_pud, pud_t *src_pud, struct vm_area_struct *vma,
982                 unsigned long addr, unsigned long end)
983 {
984         pmd_t *src_pmd, *dst_pmd;
985         unsigned long next;
986
987         dst_pmd = pmd_alloc(dst_mm, dst_pud, addr);
988         if (!dst_pmd)
989                 return -ENOMEM;
990         src_pmd = pmd_offset(src_pud, addr);
991         do {
992                 next = pmd_addr_end(addr, end);
993                 if (pmd_trans_huge(*src_pmd)) {
994                         int err;
995                         VM_BUG_ON(next-addr != HPAGE_PMD_SIZE);
996                         err = copy_huge_pmd(dst_mm, src_mm,
997                                             dst_pmd, src_pmd, addr, vma);
998                         if (err == -ENOMEM)
999                                 return -ENOMEM;
1000                         if (!err)
1001                                 continue;
1002                         /* fall through */
1003                 }
1004                 if (pmd_none_or_clear_bad(src_pmd))
1005                         continue;
1006                 if (copy_pte_range(dst_mm, src_mm, dst_pmd, src_pmd,
1007                                                 vma, addr, next))
1008                         return -ENOMEM;
1009         } while (dst_pmd++, src_pmd++, addr = next, addr != end);
1010         return 0;
1011 }
1012
1013 static inline int copy_pud_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
1014                 pgd_t *dst_pgd, pgd_t *src_pgd, struct vm_area_struct *vma,
1015                 unsigned long addr, unsigned long end)
1016 {
1017         pud_t *src_pud, *dst_pud;
1018         unsigned long next;
1019
1020         dst_pud = pud_alloc(dst_mm, dst_pgd, addr);
1021         if (!dst_pud)
1022                 return -ENOMEM;
1023         src_pud = pud_offset(src_pgd, addr);
1024         do {
1025                 next = pud_addr_end(addr, end);
1026                 if (pud_none_or_clear_bad(src_pud))
1027                         continue;
1028                 if (copy_pmd_range(dst_mm, src_mm, dst_pud, src_pud,
1029                                                 vma, addr, next))
1030                         return -ENOMEM;
1031         } while (dst_pud++, src_pud++, addr = next, addr != end);
1032         return 0;
1033 }
1034
1035 int copy_page_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
1036                 struct vm_area_struct *vma)
1037 {
1038         pgd_t *src_pgd, *dst_pgd;
1039         unsigned long next;
1040         unsigned long addr = vma->vm_start;
1041         unsigned long end = vma->vm_end;
1042         int ret;
1043
1044         /*
1045          * Don't copy ptes where a page fault will fill them correctly.
1046          * Fork becomes much lighter when there are big shared or private
1047          * readonly mappings. The tradeoff is that copy_page_range is more
1048          * efficient than faulting.
1049          */
1050         if (!(vma->vm_flags & (VM_HUGETLB | VM_NONLINEAR |
1051                                VM_PFNMAP | VM_MIXEDMAP))) {
1052                 if (!vma->anon_vma)
1053                         return 0;
1054         }
1055
1056         if (is_vm_hugetlb_page(vma))
1057                 return copy_hugetlb_page_range(dst_mm, src_mm, vma);
1058
1059         if (unlikely(vma->vm_flags & VM_PFNMAP)) {
1060                 /*
1061                  * We do not free on error cases below as remove_vma
1062                  * gets called on error from higher level routine
1063                  */
1064                 ret = track_pfn_copy(vma);
1065                 if (ret)
1066                         return ret;
1067         }
1068
1069         /*
1070          * We need to invalidate the secondary MMU mappings only when
1071          * there could be a permission downgrade on the ptes of the
1072          * parent mm. And a permission downgrade will only happen if
1073          * is_cow_mapping() returns true.
1074          */
1075         if (is_cow_mapping(vma->vm_flags))
1076                 mmu_notifier_invalidate_range_start(src_mm, addr, 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_mapping(vma->vm_flags))
1093                 mmu_notifier_invalidate_range_end(src_mm,
1094                                                   vma->vm_start, end);
1095         return ret;
1096 }
1097
1098 static unsigned long zap_pte_range(struct mmu_gather *tlb,
1099                                 struct vm_area_struct *vma, pmd_t *pmd,
1100                                 unsigned long addr, unsigned long end,
1101                                 struct zap_details *details)
1102 {
1103         struct mm_struct *mm = tlb->mm;
1104         int force_flush = 0;
1105         int rss[NR_MM_COUNTERS];
1106         spinlock_t *ptl;
1107         pte_t *start_pte;
1108         pte_t *pte;
1109
1110 again:
1111         init_rss_vec(rss);
1112         start_pte = pte_offset_map_lock(mm, pmd, addr, &ptl);
1113         pte = start_pte;
1114         arch_enter_lazy_mmu_mode();
1115         do {
1116                 pte_t ptent = *pte;
1117                 if (pte_none(ptent)) {
1118                         continue;
1119                 }
1120
1121                 if (pte_present(ptent)) {
1122                         struct page *page;
1123
1124                         page = vm_normal_page(vma, addr, ptent);
1125                         if (unlikely(details) && page) {
1126                                 /*
1127                                  * unmap_shared_mapping_pages() wants to
1128                                  * invalidate cache without truncating:
1129                                  * unmap shared but keep private pages.
1130                                  */
1131                                 if (details->check_mapping &&
1132                                     details->check_mapping != page->mapping)
1133                                         continue;
1134                                 /*
1135                                  * Each page->index must be checked when
1136                                  * invalidating or truncating nonlinear.
1137                                  */
1138                                 if (details->nonlinear_vma &&
1139                                     (page->index < details->first_index ||
1140                                      page->index > details->last_index))
1141                                         continue;
1142                         }
1143                         ptent = ptep_get_and_clear_full(mm, addr, pte,
1144                                                         tlb->fullmm);
1145                         tlb_remove_tlb_entry(tlb, pte, addr);
1146                         if (unlikely(!page))
1147                                 continue;
1148                         if (unlikely(details) && details->nonlinear_vma
1149                             && linear_page_index(details->nonlinear_vma,
1150                                                 addr) != page->index)
1151                                 set_pte_at(mm, addr, pte,
1152                                            pgoff_to_pte(page->index));
1153                         if (PageAnon(page))
1154                                 rss[MM_ANONPAGES]--;
1155                         else {
1156                                 if (pte_dirty(ptent))
1157                                         set_page_dirty(page);
1158                                 if (pte_young(ptent) &&
1159                                     likely(!VM_SequentialReadHint(vma)))
1160                                         mark_page_accessed(page);
1161                                 rss[MM_FILEPAGES]--;
1162                         }
1163                         page_remove_rmap(page);
1164                         if (unlikely(page_mapcount(page) < 0))
1165                                 print_bad_pte(vma, addr, ptent, page);
1166                         force_flush = !__tlb_remove_page(tlb, page);
1167                         if (force_flush)
1168                                 break;
1169                         continue;
1170                 }
1171                 /*
1172                  * If details->check_mapping, we leave swap entries;
1173                  * if details->nonlinear_vma, we leave file entries.
1174                  */
1175                 if (unlikely(details))
1176                         continue;
1177                 if (pte_file(ptent)) {
1178                         if (unlikely(!(vma->vm_flags & VM_NONLINEAR)))
1179                                 print_bad_pte(vma, addr, ptent, NULL);
1180                 } else {
1181                         swp_entry_t entry = pte_to_swp_entry(ptent);
1182
1183                         if (!non_swap_entry(entry))
1184                                 rss[MM_SWAPENTS]--;
1185                         else if (is_migration_entry(entry)) {
1186                                 struct page *page;
1187
1188                                 page = migration_entry_to_page(entry);
1189
1190                                 if (PageAnon(page))
1191                                         rss[MM_ANONPAGES]--;
1192                                 else
1193                                         rss[MM_FILEPAGES]--;
1194                         }
1195                         if (unlikely(!free_swap_and_cache(entry)))
1196                                 print_bad_pte(vma, addr, ptent, NULL);
1197                 }
1198                 pte_clear_not_present_full(mm, addr, pte, tlb->fullmm);
1199         } while (pte++, addr += PAGE_SIZE, addr != end);
1200
1201         add_mm_rss_vec(mm, rss);
1202         arch_leave_lazy_mmu_mode();
1203         pte_unmap_unlock(start_pte, ptl);
1204
1205         /*
1206          * mmu_gather ran out of room to batch pages, we break out of
1207          * the PTE lock to avoid doing the potential expensive TLB invalidate
1208          * and page-free while holding it.
1209          */
1210         if (force_flush) {
1211                 force_flush = 0;
1212
1213 #ifdef HAVE_GENERIC_MMU_GATHER
1214                 tlb->start = addr;
1215                 tlb->end = end;
1216 #endif
1217                 tlb_flush_mmu(tlb);
1218                 if (addr != end)
1219                         goto again;
1220         }
1221
1222         return addr;
1223 }
1224
1225 static inline unsigned long zap_pmd_range(struct mmu_gather *tlb,
1226                                 struct vm_area_struct *vma, pud_t *pud,
1227                                 unsigned long addr, unsigned long end,
1228                                 struct zap_details *details)
1229 {
1230         pmd_t *pmd;
1231         unsigned long next;
1232
1233         pmd = pmd_offset(pud, addr);
1234         do {
1235                 next = pmd_addr_end(addr, end);
1236                 if (pmd_trans_huge(*pmd)) {
1237                         if (next - addr != HPAGE_PMD_SIZE) {
1238 #ifdef CONFIG_DEBUG_VM
1239                                 if (!rwsem_is_locked(&tlb->mm->mmap_sem)) {
1240                                         pr_err("%s: mmap_sem is unlocked! addr=0x%lx end=0x%lx vma->vm_start=0x%lx vma->vm_end=0x%lx\n",
1241                                                 __func__, addr, end,
1242                                                 vma->vm_start,
1243                                                 vma->vm_end);
1244                                         BUG();
1245                                 }
1246 #endif
1247                                 split_huge_page_pmd(vma->vm_mm, pmd);
1248                         } else if (zap_huge_pmd(tlb, vma, pmd, addr))
1249                                 goto next;
1250                         /* fall through */
1251                 }
1252                 /*
1253                  * Here there can be other concurrent MADV_DONTNEED or
1254                  * trans huge page faults running, and if the pmd is
1255                  * none or trans huge it can change under us. This is
1256                  * because MADV_DONTNEED holds the mmap_sem in read
1257                  * mode.
1258                  */
1259                 if (pmd_none_or_trans_huge_or_clear_bad(pmd))
1260                         goto next;
1261                 next = zap_pte_range(tlb, vma, pmd, addr, next, details);
1262 next:
1263                 cond_resched();
1264         } while (pmd++, addr = next, addr != end);
1265
1266         return addr;
1267 }
1268
1269 static inline unsigned long zap_pud_range(struct mmu_gather *tlb,
1270                                 struct vm_area_struct *vma, pgd_t *pgd,
1271                                 unsigned long addr, unsigned long end,
1272                                 struct zap_details *details)
1273 {
1274         pud_t *pud;
1275         unsigned long next;
1276
1277         pud = pud_offset(pgd, addr);
1278         do {
1279                 next = pud_addr_end(addr, end);
1280                 if (pud_none_or_clear_bad(pud))
1281                         continue;
1282                 next = zap_pmd_range(tlb, vma, pud, addr, next, details);
1283         } while (pud++, addr = next, addr != end);
1284
1285         return addr;
1286 }
1287
1288 static void unmap_page_range(struct mmu_gather *tlb,
1289                              struct vm_area_struct *vma,
1290                              unsigned long addr, unsigned long end,
1291                              struct zap_details *details)
1292 {
1293         pgd_t *pgd;
1294         unsigned long next;
1295
1296         if (details && !details->check_mapping && !details->nonlinear_vma)
1297                 details = NULL;
1298
1299         BUG_ON(addr >= end);
1300         mem_cgroup_uncharge_start();
1301         tlb_start_vma(tlb, vma);
1302         pgd = pgd_offset(vma->vm_mm, addr);
1303         do {
1304                 next = pgd_addr_end(addr, end);
1305                 if (pgd_none_or_clear_bad(pgd))
1306                         continue;
1307                 next = zap_pud_range(tlb, vma, pgd, addr, next, details);
1308         } while (pgd++, addr = next, addr != end);
1309         tlb_end_vma(tlb, vma);
1310         mem_cgroup_uncharge_end();
1311 }
1312
1313
1314 static void unmap_single_vma(struct mmu_gather *tlb,
1315                 struct vm_area_struct *vma, unsigned long start_addr,
1316                 unsigned long end_addr,
1317                 struct zap_details *details)
1318 {
1319         unsigned long start = max(vma->vm_start, start_addr);
1320         unsigned long end;
1321
1322         if (start >= vma->vm_end)
1323                 return;
1324         end = min(vma->vm_end, end_addr);
1325         if (end <= vma->vm_start)
1326                 return;
1327
1328         if (vma->vm_file)
1329                 uprobe_munmap(vma, start, end);
1330
1331         if (unlikely(vma->vm_flags & VM_PFNMAP))
1332                 untrack_pfn(vma, 0, 0);
1333
1334         if (start != end) {
1335                 if (unlikely(is_vm_hugetlb_page(vma))) {
1336                         /*
1337                          * It is undesirable to test vma->vm_file as it
1338                          * should be non-null for valid hugetlb area.
1339                          * However, vm_file will be NULL in the error
1340                          * cleanup path of do_mmap_pgoff. When
1341                          * hugetlbfs ->mmap method fails,
1342                          * do_mmap_pgoff() nullifies vma->vm_file
1343                          * before calling this function to clean up.
1344                          * Since no pte has actually been setup, it is
1345                          * safe to do nothing in this case.
1346                          */
1347                         if (vma->vm_file) {
1348                                 mutex_lock(&vma->vm_file->f_mapping->i_mmap_mutex);
1349                                 __unmap_hugepage_range_final(tlb, vma, start, end, NULL);
1350                                 mutex_unlock(&vma->vm_file->f_mapping->i_mmap_mutex);
1351                         }
1352                 } else
1353                         unmap_page_range(tlb, vma, start, end, details);
1354         }
1355 }
1356
1357 /**
1358  * unmap_vmas - unmap a range of memory covered by a list of vma's
1359  * @tlb: address of the caller's struct mmu_gather
1360  * @vma: the starting vma
1361  * @start_addr: virtual address at which to start unmapping
1362  * @end_addr: virtual address at which to end unmapping
1363  *
1364  * Unmap all pages in the vma list.
1365  *
1366  * Only addresses between `start' and `end' will be unmapped.
1367  *
1368  * The VMA list must be sorted in ascending virtual address order.
1369  *
1370  * unmap_vmas() assumes that the caller will flush the whole unmapped address
1371  * range after unmap_vmas() returns.  So the only responsibility here is to
1372  * ensure that any thus-far unmapped pages are flushed before unmap_vmas()
1373  * drops the lock and schedules.
1374  */
1375 void unmap_vmas(struct mmu_gather *tlb,
1376                 struct vm_area_struct *vma, unsigned long start_addr,
1377                 unsigned long end_addr)
1378 {
1379         struct mm_struct *mm = vma->vm_mm;
1380
1381         mmu_notifier_invalidate_range_start(mm, start_addr, end_addr);
1382         for ( ; vma && vma->vm_start < end_addr; vma = vma->vm_next)
1383                 unmap_single_vma(tlb, vma, start_addr, end_addr, NULL);
1384         mmu_notifier_invalidate_range_end(mm, start_addr, end_addr);
1385 }
1386
1387 /**
1388  * zap_page_range - remove user pages in a given range
1389  * @vma: vm_area_struct holding the applicable pages
1390  * @start: starting address of pages to zap
1391  * @size: number of bytes to zap
1392  * @details: details of nonlinear truncation or shared cache invalidation
1393  *
1394  * Caller must protect the VMA list
1395  */
1396 void zap_page_range(struct vm_area_struct *vma, unsigned long start,
1397                 unsigned long size, struct zap_details *details)
1398 {
1399         struct mm_struct *mm = vma->vm_mm;
1400         struct mmu_gather tlb;
1401         unsigned long end = start + size;
1402
1403         lru_add_drain();
1404         tlb_gather_mmu(&tlb, mm, 0);
1405         update_hiwater_rss(mm);
1406         mmu_notifier_invalidate_range_start(mm, start, end);
1407         for ( ; vma && vma->vm_start < end; vma = vma->vm_next)
1408                 unmap_single_vma(&tlb, vma, start, end, details);
1409         mmu_notifier_invalidate_range_end(mm, start, end);
1410         tlb_finish_mmu(&tlb, start, end);
1411 }
1412
1413 /**
1414  * zap_page_range_single - remove user pages in a given range
1415  * @vma: vm_area_struct holding the applicable pages
1416  * @address: starting address of pages to zap
1417  * @size: number of bytes to zap
1418  * @details: details of nonlinear truncation or shared cache invalidation
1419  *
1420  * The range must fit into one VMA.
1421  */
1422 static void zap_page_range_single(struct vm_area_struct *vma, unsigned long address,
1423                 unsigned long size, struct zap_details *details)
1424 {
1425         struct mm_struct *mm = vma->vm_mm;
1426         struct mmu_gather tlb;
1427         unsigned long end = address + size;
1428
1429         lru_add_drain();
1430         tlb_gather_mmu(&tlb, mm, 0);
1431         update_hiwater_rss(mm);
1432         mmu_notifier_invalidate_range_start(mm, address, end);
1433         unmap_single_vma(&tlb, vma, address, end, details);
1434         mmu_notifier_invalidate_range_end(mm, address, end);
1435         tlb_finish_mmu(&tlb, address, end);
1436 }
1437
1438 /**
1439  * zap_vma_ptes - remove ptes mapping the vma
1440  * @vma: vm_area_struct holding ptes to be zapped
1441  * @address: starting address of pages to zap
1442  * @size: number of bytes to zap
1443  *
1444  * This function only unmaps ptes assigned to VM_PFNMAP vmas.
1445  *
1446  * The entire address range must be fully contained within the vma.
1447  *
1448  * Returns 0 if successful.
1449  */
1450 int zap_vma_ptes(struct vm_area_struct *vma, unsigned long address,
1451                 unsigned long size)
1452 {
1453         if (address < vma->vm_start || address + size > vma->vm_end ||
1454                         !(vma->vm_flags & VM_PFNMAP))
1455                 return -1;
1456         zap_page_range_single(vma, address, size, NULL);
1457         return 0;
1458 }
1459 EXPORT_SYMBOL_GPL(zap_vma_ptes);
1460
1461 /**
1462  * follow_page - look up a page descriptor from a user-virtual address
1463  * @vma: vm_area_struct mapping @address
1464  * @address: virtual address to look up
1465  * @flags: flags modifying lookup behaviour
1466  *
1467  * @flags can have FOLL_ flags set, defined in <linux/mm.h>
1468  *
1469  * Returns the mapped (struct page *), %NULL if no mapping exists, or
1470  * an error pointer if there is a mapping to something not represented
1471  * by a page descriptor (see also vm_normal_page()).
1472  */
1473 struct page *follow_page(struct vm_area_struct *vma, unsigned long address,
1474                         unsigned int flags)
1475 {
1476         pgd_t *pgd;
1477         pud_t *pud;
1478         pmd_t *pmd;
1479         pte_t *ptep, pte;
1480         spinlock_t *ptl;
1481         struct page *page;
1482         struct mm_struct *mm = vma->vm_mm;
1483
1484         page = follow_huge_addr(mm, address, flags & FOLL_WRITE);
1485         if (!IS_ERR(page)) {
1486                 BUG_ON(flags & FOLL_GET);
1487                 goto out;
1488         }
1489
1490         page = NULL;
1491         pgd = pgd_offset(mm, address);
1492         if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
1493                 goto no_page_table;
1494
1495         pud = pud_offset(pgd, address);
1496         if (pud_none(*pud))
1497                 goto no_page_table;
1498         if (pud_huge(*pud) && vma->vm_flags & VM_HUGETLB) {
1499                 BUG_ON(flags & FOLL_GET);
1500                 page = follow_huge_pud(mm, address, pud, flags & FOLL_WRITE);
1501                 goto out;
1502         }
1503         if (unlikely(pud_bad(*pud)))
1504                 goto no_page_table;
1505
1506         pmd = pmd_offset(pud, address);
1507         if (pmd_none(*pmd))
1508                 goto no_page_table;
1509         if (pmd_huge(*pmd) && vma->vm_flags & VM_HUGETLB) {
1510                 BUG_ON(flags & FOLL_GET);
1511                 page = follow_huge_pmd(mm, address, pmd, flags & FOLL_WRITE);
1512                 goto out;
1513         }
1514         if (pmd_trans_huge(*pmd)) {
1515                 if (flags & FOLL_SPLIT) {
1516                         split_huge_page_pmd(mm, pmd);
1517                         goto split_fallthrough;
1518                 }
1519                 spin_lock(&mm->page_table_lock);
1520                 if (likely(pmd_trans_huge(*pmd))) {
1521                         if (unlikely(pmd_trans_splitting(*pmd))) {
1522                                 spin_unlock(&mm->page_table_lock);
1523                                 wait_split_huge_page(vma->anon_vma, pmd);
1524                         } else {
1525                                 page = follow_trans_huge_pmd(mm, address,
1526                                                              pmd, flags);
1527                                 spin_unlock(&mm->page_table_lock);
1528                                 goto out;
1529                         }
1530                 } else
1531                         spin_unlock(&mm->page_table_lock);
1532                 /* fall through */
1533         }
1534 split_fallthrough:
1535         if (unlikely(pmd_bad(*pmd)))
1536                 goto no_page_table;
1537
1538         ptep = pte_offset_map_lock(mm, pmd, address, &ptl);
1539
1540         pte = *ptep;
1541         if (!pte_present(pte))
1542                 goto no_page;
1543         if ((flags & FOLL_WRITE) && !pte_write(pte))
1544                 goto unlock;
1545
1546         page = vm_normal_page(vma, address, pte);
1547         if (unlikely(!page)) {
1548                 if ((flags & FOLL_DUMP) ||
1549                     !is_zero_pfn(pte_pfn(pte)))
1550                         goto bad_page;
1551                 page = pte_page(pte);
1552         }
1553
1554         if (flags & FOLL_GET)
1555                 get_page_foll(page);
1556         if (flags & FOLL_TOUCH) {
1557                 if ((flags & FOLL_WRITE) &&
1558                     !pte_dirty(pte) && !PageDirty(page))
1559                         set_page_dirty(page);
1560                 /*
1561                  * pte_mkyoung() would be more correct here, but atomic care
1562                  * is needed to avoid losing the dirty bit: it is easier to use
1563                  * mark_page_accessed().
1564                  */
1565                 mark_page_accessed(page);
1566         }
1567         if ((flags & FOLL_MLOCK) && (vma->vm_flags & VM_LOCKED)) {
1568                 /*
1569                  * The preliminary mapping check is mainly to avoid the
1570                  * pointless overhead of lock_page on the ZERO_PAGE
1571                  * which might bounce very badly if there is contention.
1572                  *
1573                  * If the page is already locked, we don't need to
1574                  * handle it now - vmscan will handle it later if and
1575                  * when it attempts to reclaim the page.
1576                  */
1577                 if (page->mapping && trylock_page(page)) {
1578                         lru_add_drain();  /* push cached pages to LRU */
1579                         /*
1580                          * Because we lock page here and migration is
1581                          * blocked by the pte's page reference, we need
1582                          * only check for file-cache page truncation.
1583                          */
1584                         if (page->mapping)
1585                                 mlock_vma_page(page);
1586                         unlock_page(page);
1587                 }
1588         }
1589 unlock:
1590         pte_unmap_unlock(ptep, ptl);
1591 out:
1592         return page;
1593
1594 bad_page:
1595         pte_unmap_unlock(ptep, ptl);
1596         return ERR_PTR(-EFAULT);
1597
1598 no_page:
1599         pte_unmap_unlock(ptep, ptl);
1600         if (!pte_none(pte))
1601                 return page;
1602
1603 no_page_table:
1604         /*
1605          * When core dumping an enormous anonymous area that nobody
1606          * has touched so far, we don't want to allocate unnecessary pages or
1607          * page tables.  Return error instead of NULL to skip handle_mm_fault,
1608          * then get_dump_page() will return NULL to leave a hole in the dump.
1609          * But we can only make this optimization where a hole would surely
1610          * be zero-filled if handle_mm_fault() actually did handle it.
1611          */
1612         if ((flags & FOLL_DUMP) &&
1613             (!vma->vm_ops || !vma->vm_ops->fault))
1614                 return ERR_PTR(-EFAULT);
1615         return page;
1616 }
1617
1618 static inline int stack_guard_page(struct vm_area_struct *vma, unsigned long addr)
1619 {
1620         return stack_guard_page_start(vma, addr) ||
1621                stack_guard_page_end(vma, addr+PAGE_SIZE);
1622 }
1623
1624 /**
1625  * __get_user_pages() - pin user pages in memory
1626  * @tsk:        task_struct of target task
1627  * @mm:         mm_struct of target mm
1628  * @start:      starting user address
1629  * @nr_pages:   number of pages from start to pin
1630  * @gup_flags:  flags modifying pin behaviour
1631  * @pages:      array that receives pointers to the pages pinned.
1632  *              Should be at least nr_pages long. Or NULL, if caller
1633  *              only intends to ensure the pages are faulted in.
1634  * @vmas:       array of pointers to vmas corresponding to each page.
1635  *              Or NULL if the caller does not require them.
1636  * @nonblocking: whether waiting for disk IO or mmap_sem contention
1637  *
1638  * Returns number of pages pinned. This may be fewer than the number
1639  * requested. If nr_pages is 0 or negative, returns 0. If no pages
1640  * were pinned, returns -errno. Each page returned must be released
1641  * with a put_page() call when it is finished with. vmas will only
1642  * remain valid while mmap_sem is held.
1643  *
1644  * Must be called with mmap_sem held for read or write.
1645  *
1646  * __get_user_pages walks a process's page tables and takes a reference to
1647  * each struct page that each user address corresponds to at a given
1648  * instant. That is, it takes the page that would be accessed if a user
1649  * thread accesses the given user virtual address at that instant.
1650  *
1651  * This does not guarantee that the page exists in the user mappings when
1652  * __get_user_pages returns, and there may even be a completely different
1653  * page there in some cases (eg. if mmapped pagecache has been invalidated
1654  * and subsequently re faulted). However it does guarantee that the page
1655  * won't be freed completely. And mostly callers simply care that the page
1656  * contains data that was valid *at some point in time*. Typically, an IO
1657  * or similar operation cannot guarantee anything stronger anyway because
1658  * locks can't be held over the syscall boundary.
1659  *
1660  * If @gup_flags & FOLL_WRITE == 0, the page must not be written to. If
1661  * the page is written to, set_page_dirty (or set_page_dirty_lock, as
1662  * appropriate) must be called after the page is finished with, and
1663  * before put_page is called.
1664  *
1665  * If @nonblocking != NULL, __get_user_pages will not wait for disk IO
1666  * or mmap_sem contention, and if waiting is needed to pin all pages,
1667  * *@nonblocking will be set to 0.
1668  *
1669  * In most cases, get_user_pages or get_user_pages_fast should be used
1670  * instead of __get_user_pages. __get_user_pages should be used only if
1671  * you need some special @gup_flags.
1672  */
1673 int __get_user_pages(struct task_struct *tsk, struct mm_struct *mm,
1674                      unsigned long start, int nr_pages, unsigned int gup_flags,
1675                      struct page **pages, struct vm_area_struct **vmas,
1676                      int *nonblocking)
1677 {
1678         int i;
1679         unsigned long vm_flags;
1680
1681         if (nr_pages <= 0)
1682                 return 0;
1683
1684         VM_BUG_ON(!!pages != !!(gup_flags & FOLL_GET));
1685
1686         /* 
1687          * Require read or write permissions.
1688          * If FOLL_FORCE is set, we only require the "MAY" flags.
1689          */
1690         vm_flags  = (gup_flags & FOLL_WRITE) ?
1691                         (VM_WRITE | VM_MAYWRITE) : (VM_READ | VM_MAYREAD);
1692         vm_flags &= (gup_flags & FOLL_FORCE) ?
1693                         (VM_MAYREAD | VM_MAYWRITE) : (VM_READ | VM_WRITE);
1694         i = 0;
1695
1696         do {
1697                 struct vm_area_struct *vma;
1698
1699                 vma = find_extend_vma(mm, start);
1700                 if (!vma && in_gate_area(mm, start)) {
1701                         unsigned long pg = start & PAGE_MASK;
1702                         pgd_t *pgd;
1703                         pud_t *pud;
1704                         pmd_t *pmd;
1705                         pte_t *pte;
1706
1707                         /* user gate pages are read-only */
1708                         if (gup_flags & FOLL_WRITE)
1709                                 return i ? : -EFAULT;
1710                         if (pg > TASK_SIZE)
1711                                 pgd = pgd_offset_k(pg);
1712                         else
1713                                 pgd = pgd_offset_gate(mm, pg);
1714                         BUG_ON(pgd_none(*pgd));
1715                         pud = pud_offset(pgd, pg);
1716                         BUG_ON(pud_none(*pud));
1717                         pmd = pmd_offset(pud, pg);
1718                         if (pmd_none(*pmd))
1719                                 return i ? : -EFAULT;
1720                         VM_BUG_ON(pmd_trans_huge(*pmd));
1721                         pte = pte_offset_map(pmd, pg);
1722                         if (pte_none(*pte)) {
1723                                 pte_unmap(pte);
1724                                 return i ? : -EFAULT;
1725                         }
1726                         vma = get_gate_vma(mm);
1727                         if (pages) {
1728                                 struct page *page;
1729
1730                                 page = vm_normal_page(vma, start, *pte);
1731                                 if (!page) {
1732                                         if (!(gup_flags & FOLL_DUMP) &&
1733                                              is_zero_pfn(pte_pfn(*pte)))
1734                                                 page = pte_page(*pte);
1735                                         else {
1736                                                 pte_unmap(pte);
1737                                                 return i ? : -EFAULT;
1738                                         }
1739                                 }
1740                                 pages[i] = page;
1741                                 get_page(page);
1742                         }
1743                         pte_unmap(pte);
1744                         goto next_page;
1745                 }
1746
1747                 if (!vma ||
1748                     (vma->vm_flags & (VM_IO | VM_PFNMAP)) ||
1749                     !(vm_flags & vma->vm_flags))
1750                         return i ? : -EFAULT;
1751
1752                 if (is_vm_hugetlb_page(vma)) {
1753                         i = follow_hugetlb_page(mm, vma, pages, vmas,
1754                                         &start, &nr_pages, i, gup_flags);
1755                         continue;
1756                 }
1757
1758                 do {
1759                         struct page *page;
1760                         unsigned int foll_flags = gup_flags;
1761
1762                         /*
1763                          * If we have a pending SIGKILL, don't keep faulting
1764                          * pages and potentially allocating memory.
1765                          */
1766                         if (unlikely(fatal_signal_pending(current)))
1767                                 return i ? i : -ERESTARTSYS;
1768
1769                         cond_resched();
1770                         while (!(page = follow_page(vma, start, foll_flags))) {
1771                                 int ret;
1772                                 unsigned int fault_flags = 0;
1773
1774                                 /* For mlock, just skip the stack guard page. */
1775                                 if (foll_flags & FOLL_MLOCK) {
1776                                         if (stack_guard_page(vma, start))
1777                                                 goto next_page;
1778                                 }
1779                                 if (foll_flags & FOLL_WRITE)
1780                                         fault_flags |= FAULT_FLAG_WRITE;
1781                                 if (nonblocking)
1782                                         fault_flags |= FAULT_FLAG_ALLOW_RETRY;
1783                                 if (foll_flags & FOLL_NOWAIT)
1784                                         fault_flags |= (FAULT_FLAG_ALLOW_RETRY | FAULT_FLAG_RETRY_NOWAIT);
1785
1786                                 ret = handle_mm_fault(mm, vma, start,
1787                                                         fault_flags);
1788
1789                                 if (ret & VM_FAULT_ERROR) {
1790                                         if (ret & VM_FAULT_OOM)
1791                                                 return i ? i : -ENOMEM;
1792                                         if (ret & (VM_FAULT_HWPOISON |
1793                                                    VM_FAULT_HWPOISON_LARGE)) {
1794                                                 if (i)
1795                                                         return i;
1796                                                 else if (gup_flags & FOLL_HWPOISON)
1797                                                         return -EHWPOISON;
1798                                                 else
1799                                                         return -EFAULT;
1800                                         }
1801                                         if (ret & VM_FAULT_SIGBUS)
1802                                                 return i ? i : -EFAULT;
1803                                         BUG();
1804                                 }
1805
1806                                 if (tsk) {
1807                                         if (ret & VM_FAULT_MAJOR)
1808                                                 tsk->maj_flt++;
1809                                         else
1810                                                 tsk->min_flt++;
1811                                 }
1812
1813                                 if (ret & VM_FAULT_RETRY) {
1814                                         if (nonblocking)
1815                                                 *nonblocking = 0;
1816                                         return i;
1817                                 }
1818
1819                                 /*
1820                                  * The VM_FAULT_WRITE bit tells us that
1821                                  * do_wp_page has broken COW when necessary,
1822                                  * even if maybe_mkwrite decided not to set
1823                                  * pte_write. We can thus safely do subsequent
1824                                  * page lookups as if they were reads. But only
1825                                  * do so when looping for pte_write is futile:
1826                                  * in some cases userspace may also be wanting
1827                                  * to write to the gotten user page, which a
1828                                  * read fault here might prevent (a readonly
1829                                  * page might get reCOWed by userspace write).
1830                                  */
1831                                 if ((ret & VM_FAULT_WRITE) &&
1832                                     !(vma->vm_flags & VM_WRITE))
1833                                         foll_flags &= ~FOLL_WRITE;
1834
1835                                 cond_resched();
1836                         }
1837                         if (IS_ERR(page))
1838                                 return i ? i : PTR_ERR(page);
1839                         if (pages) {
1840                                 pages[i] = page;
1841
1842                                 flush_anon_page(vma, page, start);
1843                                 flush_dcache_page(page);
1844                         }
1845 next_page:
1846                         if (vmas)
1847                                 vmas[i] = vma;
1848                         i++;
1849                         start += PAGE_SIZE;
1850                         nr_pages--;
1851                 } while (nr_pages && start < vma->vm_end);
1852         } while (nr_pages);
1853         return i;
1854 }
1855 EXPORT_SYMBOL(__get_user_pages);
1856
1857 /*
1858  * fixup_user_fault() - manually resolve a user page fault
1859  * @tsk:        the task_struct to use for page fault accounting, or
1860  *              NULL if faults are not to be recorded.
1861  * @mm:         mm_struct of target mm
1862  * @address:    user address
1863  * @fault_flags:flags to pass down to handle_mm_fault()
1864  *
1865  * This is meant to be called in the specific scenario where for locking reasons
1866  * we try to access user memory in atomic context (within a pagefault_disable()
1867  * section), this returns -EFAULT, and we want to resolve the user fault before
1868  * trying again.
1869  *
1870  * Typically this is meant to be used by the futex code.
1871  *
1872  * The main difference with get_user_pages() is that this function will
1873  * unconditionally call handle_mm_fault() which will in turn perform all the
1874  * necessary SW fixup of the dirty and young bits in the PTE, while
1875  * handle_mm_fault() only guarantees to update these in the struct page.
1876  *
1877  * This is important for some architectures where those bits also gate the
1878  * access permission to the page because they are maintained in software.  On
1879  * such architectures, gup() will not be enough to make a subsequent access
1880  * succeed.
1881  *
1882  * This should be called with the mm_sem held for read.
1883  */
1884 int fixup_user_fault(struct task_struct *tsk, struct mm_struct *mm,
1885                      unsigned long address, unsigned int fault_flags)
1886 {
1887         struct vm_area_struct *vma;
1888         int ret;
1889
1890         vma = find_extend_vma(mm, address);
1891         if (!vma || address < vma->vm_start)
1892                 return -EFAULT;
1893
1894         ret = handle_mm_fault(mm, vma, address, fault_flags);
1895         if (ret & VM_FAULT_ERROR) {
1896                 if (ret & VM_FAULT_OOM)
1897                         return -ENOMEM;
1898                 if (ret & (VM_FAULT_HWPOISON | VM_FAULT_HWPOISON_LARGE))
1899                         return -EHWPOISON;
1900                 if (ret & VM_FAULT_SIGBUS)
1901                         return -EFAULT;
1902                 BUG();
1903         }
1904         if (tsk) {
1905                 if (ret & VM_FAULT_MAJOR)
1906                         tsk->maj_flt++;
1907                 else
1908                         tsk->min_flt++;
1909         }
1910         return 0;
1911 }
1912
1913 /*
1914  * get_user_pages() - pin user pages in memory
1915  * @tsk:        the task_struct to use for page fault accounting, or
1916  *              NULL if faults are not to be recorded.
1917  * @mm:         mm_struct of target mm
1918  * @start:      starting user address
1919  * @nr_pages:   number of pages from start to pin
1920  * @write:      whether pages will be written to by the caller
1921  * @force:      whether to force write access even if user mapping is
1922  *              readonly. This will result in the page being COWed even
1923  *              in MAP_SHARED mappings. You do not want this.
1924  * @pages:      array that receives pointers to the pages pinned.
1925  *              Should be at least nr_pages long. Or NULL, if caller
1926  *              only intends to ensure the pages are faulted in.
1927  * @vmas:       array of pointers to vmas corresponding to each page.
1928  *              Or NULL if the caller does not require them.
1929  *
1930  * Returns number of pages pinned. This may be fewer than the number
1931  * requested. If nr_pages is 0 or negative, returns 0. If no pages
1932  * were pinned, returns -errno. Each page returned must be released
1933  * with a put_page() call when it is finished with. vmas will only
1934  * remain valid while mmap_sem is held.
1935  *
1936  * Must be called with mmap_sem held for read or write.
1937  *
1938  * get_user_pages walks a process's page tables and takes a reference to
1939  * each struct page that each user address corresponds to at a given
1940  * instant. That is, it takes the page that would be accessed if a user
1941  * thread accesses the given user virtual address at that instant.
1942  *
1943  * This does not guarantee that the page exists in the user mappings when
1944  * get_user_pages returns, and there may even be a completely different
1945  * page there in some cases (eg. if mmapped pagecache has been invalidated
1946  * and subsequently re faulted). However it does guarantee that the page
1947  * won't be freed completely. And mostly callers simply care that the page
1948  * contains data that was valid *at some point in time*. Typically, an IO
1949  * or similar operation cannot guarantee anything stronger anyway because
1950  * locks can't be held over the syscall boundary.
1951  *
1952  * If write=0, the page must not be written to. If the page is written to,
1953  * set_page_dirty (or set_page_dirty_lock, as appropriate) must be called
1954  * after the page is finished with, and before put_page is called.
1955  *
1956  * get_user_pages is typically used for fewer-copy IO operations, to get a
1957  * handle on the memory by some means other than accesses via the user virtual
1958  * addresses. The pages may be submitted for DMA to devices or accessed via
1959  * their kernel linear mapping (via the kmap APIs). Care should be taken to
1960  * use the correct cache flushing APIs.
1961  *
1962  * See also get_user_pages_fast, for performance critical applications.
1963  */
1964 int get_user_pages(struct task_struct *tsk, struct mm_struct *mm,
1965                 unsigned long start, int nr_pages, int write, int force,
1966                 struct page **pages, struct vm_area_struct **vmas)
1967 {
1968         int flags = FOLL_TOUCH;
1969
1970         if (pages)
1971                 flags |= FOLL_GET;
1972         if (write)
1973                 flags |= FOLL_WRITE;
1974         if (force)
1975                 flags |= FOLL_FORCE;
1976
1977         return __get_user_pages(tsk, mm, start, nr_pages, flags, pages, vmas,
1978                                 NULL);
1979 }
1980 EXPORT_SYMBOL(get_user_pages);
1981
1982 /**
1983  * get_dump_page() - pin user page in memory while writing it to core dump
1984  * @addr: user address
1985  *
1986  * Returns struct page pointer of user page pinned for dump,
1987  * to be freed afterwards by page_cache_release() or put_page().
1988  *
1989  * Returns NULL on any kind of failure - a hole must then be inserted into
1990  * the corefile, to preserve alignment with its headers; and also returns
1991  * NULL wherever the ZERO_PAGE, or an anonymous pte_none, has been found -
1992  * allowing a hole to be left in the corefile to save diskspace.
1993  *
1994  * Called without mmap_sem, but after all other threads have been killed.
1995  */
1996 #ifdef CONFIG_ELF_CORE
1997 struct page *get_dump_page(unsigned long addr)
1998 {
1999         struct vm_area_struct *vma;
2000         struct page *page;
2001
2002         if (__get_user_pages(current, current->mm, addr, 1,
2003                              FOLL_FORCE | FOLL_DUMP | FOLL_GET, &page, &vma,
2004                              NULL) < 1)
2005                 return NULL;
2006         flush_cache_page(vma, addr, page_to_pfn(page));
2007         return page;
2008 }
2009 #endif /* CONFIG_ELF_CORE */
2010
2011 pte_t *__get_locked_pte(struct mm_struct *mm, unsigned long addr,
2012                         spinlock_t **ptl)
2013 {
2014         pgd_t * pgd = pgd_offset(mm, addr);
2015         pud_t * pud = pud_alloc(mm, pgd, addr);
2016         if (pud) {
2017                 pmd_t * pmd = pmd_alloc(mm, pud, addr);
2018                 if (pmd) {
2019                         VM_BUG_ON(pmd_trans_huge(*pmd));
2020                         return pte_alloc_map_lock(mm, pmd, addr, ptl);
2021                 }
2022         }
2023         return NULL;
2024 }
2025
2026 /*
2027  * This is the old fallback for page remapping.
2028  *
2029  * For historical reasons, it only allows reserved pages. Only
2030  * old drivers should use this, and they needed to mark their
2031  * pages reserved for the old functions anyway.
2032  */
2033 static int insert_page(struct vm_area_struct *vma, unsigned long addr,
2034                         struct page *page, pgprot_t prot)
2035 {
2036         struct mm_struct *mm = vma->vm_mm;
2037         int retval;
2038         pte_t *pte;
2039         spinlock_t *ptl;
2040
2041         retval = -EINVAL;
2042         if (PageAnon(page))
2043                 goto out;
2044         retval = -ENOMEM;
2045         flush_dcache_page(page);
2046         pte = get_locked_pte(mm, addr, &ptl);
2047         if (!pte)
2048                 goto out;
2049         retval = -EBUSY;
2050         if (!pte_none(*pte))
2051                 goto out_unlock;
2052
2053         /* Ok, finally just insert the thing.. */
2054         get_page(page);
2055         inc_mm_counter_fast(mm, MM_FILEPAGES);
2056         page_add_file_rmap(page);
2057         set_pte_at(mm, addr, pte, mk_pte(page, prot));
2058
2059         retval = 0;
2060         pte_unmap_unlock(pte, ptl);
2061         return retval;
2062 out_unlock:
2063         pte_unmap_unlock(pte, ptl);
2064 out:
2065         return retval;
2066 }
2067
2068 /**
2069  * vm_insert_page - insert single page into user vma
2070  * @vma: user vma to map to
2071  * @addr: target user address of this page
2072  * @page: source kernel page
2073  *
2074  * This allows drivers to insert individual pages they've allocated
2075  * into a user vma.
2076  *
2077  * The page has to be a nice clean _individual_ kernel allocation.
2078  * If you allocate a compound page, you need to have marked it as
2079  * such (__GFP_COMP), or manually just split the page up yourself
2080  * (see split_page()).
2081  *
2082  * NOTE! Traditionally this was done with "remap_pfn_range()" which
2083  * took an arbitrary page protection parameter. This doesn't allow
2084  * that. Your vma protection will have to be set up correctly, which
2085  * means that if you want a shared writable mapping, you'd better
2086  * ask for a shared writable mapping!
2087  *
2088  * The page does not need to be reserved.
2089  *
2090  * Usually this function is called from f_op->mmap() handler
2091  * under mm->mmap_sem write-lock, so it can change vma->vm_flags.
2092  * Caller must set VM_MIXEDMAP on vma if it wants to call this
2093  * function from other places, for example from page-fault handler.
2094  */
2095 int vm_insert_page(struct vm_area_struct *vma, unsigned long addr,
2096                         struct page *page)
2097 {
2098         if (addr < vma->vm_start || addr >= vma->vm_end)
2099                 return -EFAULT;
2100         if (!page_count(page))
2101                 return -EINVAL;
2102         if (!(vma->vm_flags & VM_MIXEDMAP)) {
2103                 BUG_ON(down_read_trylock(&vma->vm_mm->mmap_sem));
2104                 BUG_ON(vma->vm_flags & VM_PFNMAP);
2105                 vma->vm_flags |= VM_MIXEDMAP;
2106         }
2107         return insert_page(vma, addr, page, vma->vm_page_prot);
2108 }
2109 EXPORT_SYMBOL(vm_insert_page);
2110
2111 static int insert_pfn(struct vm_area_struct *vma, unsigned long addr,
2112                         unsigned long pfn, pgprot_t prot)
2113 {
2114         struct mm_struct *mm = vma->vm_mm;
2115         int retval;
2116         pte_t *pte, entry;
2117         spinlock_t *ptl;
2118
2119         retval = -ENOMEM;
2120         pte = get_locked_pte(mm, addr, &ptl);
2121         if (!pte)
2122                 goto out;
2123         retval = -EBUSY;
2124         if (!pte_none(*pte))
2125                 goto out_unlock;
2126
2127         /* Ok, finally just insert the thing.. */
2128         entry = pte_mkspecial(pfn_pte(pfn, prot));
2129         set_pte_at(mm, addr, pte, entry);
2130         update_mmu_cache(vma, addr, pte); /* XXX: why not for insert_page? */
2131
2132         retval = 0;
2133 out_unlock:
2134         pte_unmap_unlock(pte, ptl);
2135 out:
2136         return retval;
2137 }
2138
2139 /**
2140  * vm_insert_pfn - insert single pfn into user vma
2141  * @vma: user vma to map to
2142  * @addr: target user address of this page
2143  * @pfn: source kernel pfn
2144  *
2145  * Similar to vm_inert_page, this allows drivers to insert individual pages
2146  * they've allocated into a user vma. Same comments apply.
2147  *
2148  * This function should only be called from a vm_ops->fault handler, and
2149  * in that case the handler should return NULL.
2150  *
2151  * vma cannot be a COW mapping.
2152  *
2153  * As this is called only for pages that do not currently exist, we
2154  * do not need to flush old virtual caches or the TLB.
2155  */
2156 int vm_insert_pfn(struct vm_area_struct *vma, unsigned long addr,
2157                         unsigned long pfn)
2158 {
2159         int ret;
2160         pgprot_t pgprot = vma->vm_page_prot;
2161         /*
2162          * Technically, architectures with pte_special can avoid all these
2163          * restrictions (same for remap_pfn_range).  However we would like
2164          * consistency in testing and feature parity among all, so we should
2165          * try to keep these invariants in place for everybody.
2166          */
2167         BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)));
2168         BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) ==
2169                                                 (VM_PFNMAP|VM_MIXEDMAP));
2170         BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags));
2171         BUG_ON((vma->vm_flags & VM_MIXEDMAP) && pfn_valid(pfn));
2172
2173         if (addr < vma->vm_start || addr >= vma->vm_end)
2174                 return -EFAULT;
2175         if (track_pfn_insert(vma, &pgprot, pfn))
2176                 return -EINVAL;
2177
2178         ret = insert_pfn(vma, addr, pfn, pgprot);
2179
2180         return ret;
2181 }
2182 EXPORT_SYMBOL(vm_insert_pfn);
2183
2184 int vm_insert_mixed(struct vm_area_struct *vma, unsigned long addr,
2185                         unsigned long pfn)
2186 {
2187         BUG_ON(!(vma->vm_flags & VM_MIXEDMAP));
2188
2189         if (addr < vma->vm_start || addr >= vma->vm_end)
2190                 return -EFAULT;
2191
2192         /*
2193          * If we don't have pte special, then we have to use the pfn_valid()
2194          * based VM_MIXEDMAP scheme (see vm_normal_page), and thus we *must*
2195          * refcount the page if pfn_valid is true (hence insert_page rather
2196          * than insert_pfn).  If a zero_pfn were inserted into a VM_MIXEDMAP
2197          * without pte special, it would there be refcounted as a normal page.
2198          */
2199         if (!HAVE_PTE_SPECIAL && pfn_valid(pfn)) {
2200                 struct page *page;
2201
2202                 page = pfn_to_page(pfn);
2203                 return insert_page(vma, addr, page, vma->vm_page_prot);
2204         }
2205         return insert_pfn(vma, addr, pfn, vma->vm_page_prot);
2206 }
2207 EXPORT_SYMBOL(vm_insert_mixed);
2208
2209 /*
2210  * maps a range of physical memory into the requested pages. the old
2211  * mappings are removed. any references to nonexistent pages results
2212  * in null mappings (currently treated as "copy-on-access")
2213  */
2214 static int remap_pte_range(struct mm_struct *mm, pmd_t *pmd,
2215                         unsigned long addr, unsigned long end,
2216                         unsigned long pfn, pgprot_t prot)
2217 {
2218         pte_t *pte;
2219         spinlock_t *ptl;
2220
2221         pte = pte_alloc_map_lock(mm, pmd, addr, &ptl);
2222         if (!pte)
2223                 return -ENOMEM;
2224         arch_enter_lazy_mmu_mode();
2225         do {
2226                 BUG_ON(!pte_none(*pte));
2227                 set_pte_at(mm, addr, pte, pte_mkspecial(pfn_pte(pfn, prot)));
2228                 pfn++;
2229         } while (pte++, addr += PAGE_SIZE, addr != end);
2230         arch_leave_lazy_mmu_mode();
2231         pte_unmap_unlock(pte - 1, ptl);
2232         return 0;
2233 }
2234
2235 static inline int remap_pmd_range(struct mm_struct *mm, pud_t *pud,
2236                         unsigned long addr, unsigned long end,
2237                         unsigned long pfn, pgprot_t prot)
2238 {
2239         pmd_t *pmd;
2240         unsigned long next;
2241
2242         pfn -= addr >> PAGE_SHIFT;
2243         pmd = pmd_alloc(mm, pud, addr);
2244         if (!pmd)
2245                 return -ENOMEM;
2246         VM_BUG_ON(pmd_trans_huge(*pmd));
2247         do {
2248                 next = pmd_addr_end(addr, end);
2249                 if (remap_pte_range(mm, pmd, addr, next,
2250                                 pfn + (addr >> PAGE_SHIFT), prot))
2251                         return -ENOMEM;
2252         } while (pmd++, addr = next, addr != end);
2253         return 0;
2254 }
2255
2256 static inline int remap_pud_range(struct mm_struct *mm, pgd_t *pgd,
2257                         unsigned long addr, unsigned long end,
2258                         unsigned long pfn, pgprot_t prot)
2259 {
2260         pud_t *pud;
2261         unsigned long next;
2262
2263         pfn -= addr >> PAGE_SHIFT;
2264         pud = pud_alloc(mm, pgd, addr);
2265         if (!pud)
2266                 return -ENOMEM;
2267         do {
2268                 next = pud_addr_end(addr, end);
2269                 if (remap_pmd_range(mm, pud, addr, next,
2270                                 pfn + (addr >> PAGE_SHIFT), prot))
2271                         return -ENOMEM;
2272         } while (pud++, addr = next, addr != end);
2273         return 0;
2274 }
2275
2276 /**
2277  * remap_pfn_range - remap kernel memory to userspace
2278  * @vma: user vma to map to
2279  * @addr: target user address to start at
2280  * @pfn: physical address of kernel memory
2281  * @size: size of map area
2282  * @prot: page protection flags for this mapping
2283  *
2284  *  Note: this is only safe if the mm semaphore is held when called.
2285  */
2286 int remap_pfn_range(struct vm_area_struct *vma, unsigned long addr,
2287                     unsigned long pfn, unsigned long size, pgprot_t prot)
2288 {
2289         pgd_t *pgd;
2290         unsigned long next;
2291         unsigned long end = addr + PAGE_ALIGN(size);
2292         struct mm_struct *mm = vma->vm_mm;
2293         int err;
2294
2295         /*
2296          * Physically remapped pages are special. Tell the
2297          * rest of the world about it:
2298          *   VM_IO tells people not to look at these pages
2299          *      (accesses can have side effects).
2300          *   VM_RESERVED is specified all over the place, because
2301          *      in 2.4 it kept swapout's vma scan off this vma; but
2302          *      in 2.6 the LRU scan won't even find its pages, so this
2303          *      flag means no more than count its pages in reserved_vm,
2304          *      and omit it from core dump, even when VM_IO turned off.
2305          *   VM_PFNMAP tells the core MM that the base pages are just
2306          *      raw PFN mappings, and do not have a "struct page" associated
2307          *      with them.
2308          *
2309          * There's a horrible special case to handle copy-on-write
2310          * behaviour that some programs depend on. We mark the "original"
2311          * un-COW'ed pages by matching them up with "vma->vm_pgoff".
2312          * See vm_normal_page() for details.
2313          */
2314         if (is_cow_mapping(vma->vm_flags)) {
2315                 if (addr != vma->vm_start || end != vma->vm_end)
2316                         return -EINVAL;
2317                 vma->vm_pgoff = pfn;
2318         }
2319
2320         err = track_pfn_remap(vma, &prot, pfn, addr, PAGE_ALIGN(size));
2321         if (err)
2322                 return -EINVAL;
2323
2324         vma->vm_flags |= VM_IO | VM_RESERVED | VM_PFNMAP;
2325
2326         BUG_ON(addr >= end);
2327         pfn -= addr >> PAGE_SHIFT;
2328         pgd = pgd_offset(mm, addr);
2329         flush_cache_range(vma, addr, end);
2330         do {
2331                 next = pgd_addr_end(addr, end);
2332                 err = remap_pud_range(mm, pgd, addr, next,
2333                                 pfn + (addr >> PAGE_SHIFT), prot);
2334                 if (err)
2335                         break;
2336         } while (pgd++, addr = next, addr != end);
2337
2338         if (err)
2339                 untrack_pfn(vma, pfn, PAGE_ALIGN(size));
2340
2341         return err;
2342 }
2343 EXPORT_SYMBOL(remap_pfn_range);
2344
2345 static int apply_to_pte_range(struct mm_struct *mm, pmd_t *pmd,
2346                                      unsigned long addr, unsigned long end,
2347                                      pte_fn_t fn, void *data)
2348 {
2349         pte_t *pte;
2350         int err;
2351         pgtable_t token;
2352         spinlock_t *uninitialized_var(ptl);
2353
2354         pte = (mm == &init_mm) ?
2355                 pte_alloc_kernel(pmd, addr) :
2356                 pte_alloc_map_lock(mm, pmd, addr, &ptl);
2357         if (!pte)
2358                 return -ENOMEM;
2359
2360         BUG_ON(pmd_huge(*pmd));
2361
2362         arch_enter_lazy_mmu_mode();
2363
2364         token = pmd_pgtable(*pmd);
2365
2366         do {
2367                 err = fn(pte++, token, addr, data);
2368                 if (err)
2369                         break;
2370         } while (addr += PAGE_SIZE, addr != end);
2371
2372         arch_leave_lazy_mmu_mode();
2373
2374         if (mm != &init_mm)
2375                 pte_unmap_unlock(pte-1, ptl);
2376         return err;
2377 }
2378
2379 static int apply_to_pmd_range(struct mm_struct *mm, pud_t *pud,
2380                                      unsigned long addr, unsigned long end,
2381                                      pte_fn_t fn, void *data)
2382 {
2383         pmd_t *pmd;
2384         unsigned long next;
2385         int err;
2386
2387         BUG_ON(pud_huge(*pud));
2388
2389         pmd = pmd_alloc(mm, pud, addr);
2390         if (!pmd)
2391                 return -ENOMEM;
2392         do {
2393                 next = pmd_addr_end(addr, end);
2394                 err = apply_to_pte_range(mm, pmd, addr, next, fn, data);
2395                 if (err)
2396                         break;
2397         } while (pmd++, addr = next, addr != end);
2398         return err;
2399 }
2400
2401 static int apply_to_pud_range(struct mm_struct *mm, pgd_t *pgd,
2402                                      unsigned long addr, unsigned long end,
2403                                      pte_fn_t fn, void *data)
2404 {
2405         pud_t *pud;
2406         unsigned long next;
2407         int err;
2408
2409         pud = pud_alloc(mm, pgd, addr);
2410         if (!pud)
2411                 return -ENOMEM;
2412         do {
2413                 next = pud_addr_end(addr, end);
2414                 err = apply_to_pmd_range(mm, pud, addr, next, fn, data);
2415                 if (err)
2416                         break;
2417         } while (pud++, addr = next, addr != end);
2418         return err;
2419 }
2420
2421 /*
2422  * Scan a region of virtual memory, filling in page tables as necessary
2423  * and calling a provided function on each leaf page table.
2424  */
2425 int apply_to_page_range(struct mm_struct *mm, unsigned long addr,
2426                         unsigned long size, pte_fn_t fn, void *data)
2427 {
2428         pgd_t *pgd;
2429         unsigned long next;
2430         unsigned long end = addr + size;
2431         int err;
2432
2433         BUG_ON(addr >= end);
2434         pgd = pgd_offset(mm, addr);
2435         do {
2436                 next = pgd_addr_end(addr, end);
2437                 err = apply_to_pud_range(mm, pgd, addr, next, fn, data);
2438                 if (err)
2439                         break;
2440         } while (pgd++, addr = next, addr != end);
2441
2442         return err;
2443 }
2444 EXPORT_SYMBOL_GPL(apply_to_page_range);
2445
2446 /*
2447  * handle_pte_fault chooses page fault handler according to an entry
2448  * which was read non-atomically.  Before making any commitment, on
2449  * those architectures or configurations (e.g. i386 with PAE) which
2450  * might give a mix of unmatched parts, do_swap_page and do_nonlinear_fault
2451  * must check under lock before unmapping the pte and proceeding
2452  * (but do_wp_page is only called after already making such a check;
2453  * and do_anonymous_page can safely check later on).
2454  */
2455 static inline int pte_unmap_same(struct mm_struct *mm, pmd_t *pmd,
2456                                 pte_t *page_table, pte_t orig_pte)
2457 {
2458         int same = 1;
2459 #if defined(CONFIG_SMP) || defined(CONFIG_PREEMPT)
2460         if (sizeof(pte_t) > sizeof(unsigned long)) {
2461                 spinlock_t *ptl = pte_lockptr(mm, pmd);
2462                 spin_lock(ptl);
2463                 same = pte_same(*page_table, orig_pte);
2464                 spin_unlock(ptl);
2465         }
2466 #endif
2467         pte_unmap(page_table);
2468         return same;
2469 }
2470
2471 static inline void cow_user_page(struct page *dst, struct page *src, unsigned long va, struct vm_area_struct *vma)
2472 {
2473         /*
2474          * If the source page was a PFN mapping, we don't have
2475          * a "struct page" for it. We do a best-effort copy by
2476          * just copying from the original user address. If that
2477          * fails, we just zero-fill it. Live with it.
2478          */
2479         if (unlikely(!src)) {
2480                 void *kaddr = kmap_atomic(dst);
2481                 void __user *uaddr = (void __user *)(va & PAGE_MASK);
2482
2483                 /*
2484                  * This really shouldn't fail, because the page is there
2485                  * in the page tables. But it might just be unreadable,
2486                  * in which case we just give up and fill the result with
2487                  * zeroes.
2488                  */
2489                 if (__copy_from_user_inatomic(kaddr, uaddr, PAGE_SIZE))
2490                         clear_page(kaddr);
2491                 kunmap_atomic(kaddr);
2492                 flush_dcache_page(dst);
2493         } else
2494                 copy_user_highpage(dst, src, va, vma);
2495 }
2496
2497 /*
2498  * This routine handles present pages, when users try to write
2499  * to a shared page. It is done by copying the page to a new address
2500  * and decrementing the shared-page counter for the old page.
2501  *
2502  * Note that this routine assumes that the protection checks have been
2503  * done by the caller (the low-level page fault routine in most cases).
2504  * Thus we can safely just mark it writable once we've done any necessary
2505  * COW.
2506  *
2507  * We also mark the page dirty at this point even though the page will
2508  * change only once the write actually happens. This avoids a few races,
2509  * and potentially makes it more efficient.
2510  *
2511  * We enter with non-exclusive mmap_sem (to exclude vma changes,
2512  * but allow concurrent faults), with pte both mapped and locked.
2513  * We return with mmap_sem still held, but pte unmapped and unlocked.
2514  */
2515 static int do_wp_page(struct mm_struct *mm, struct vm_area_struct *vma,
2516                 unsigned long address, pte_t *page_table, pmd_t *pmd,
2517                 spinlock_t *ptl, pte_t orig_pte)
2518         __releases(ptl)
2519 {
2520         struct page *old_page, *new_page;
2521         pte_t entry;
2522         int ret = 0;
2523         int page_mkwrite = 0;
2524         struct page *dirty_page = NULL;
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         /*
2703          * Re-check the pte - we dropped the lock
2704          */
2705         page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2706         if (likely(pte_same(*page_table, orig_pte))) {
2707                 if (old_page) {
2708                         if (!PageAnon(old_page)) {
2709                                 dec_mm_counter_fast(mm, MM_FILEPAGES);
2710                                 inc_mm_counter_fast(mm, MM_ANONPAGES);
2711                         }
2712                 } else
2713                         inc_mm_counter_fast(mm, MM_ANONPAGES);
2714                 flush_cache_page(vma, address, pte_pfn(orig_pte));
2715                 entry = mk_pte(new_page, vma->vm_page_prot);
2716                 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2717                 /*
2718                  * Clear the pte entry and flush it first, before updating the
2719                  * pte with the new entry. This will avoid a race condition
2720                  * seen in the presence of one thread doing SMC and another
2721                  * thread doing COW.
2722                  */
2723                 ptep_clear_flush(vma, address, page_table);
2724                 page_add_new_anon_rmap(new_page, vma, address);
2725                 /*
2726                  * We call the notify macro here because, when using secondary
2727                  * mmu page tables (such as kvm shadow page tables), we want the
2728                  * new page to be mapped directly into the secondary page table.
2729                  */
2730                 set_pte_at_notify(mm, address, page_table, entry);
2731                 update_mmu_cache(vma, address, page_table);
2732                 if (old_page) {
2733                         /*
2734                          * Only after switching the pte to the new page may
2735                          * we remove the mapcount here. Otherwise another
2736                          * process may come and find the rmap count decremented
2737                          * before the pte is switched to the new page, and
2738                          * "reuse" the old page writing into it while our pte
2739                          * here still points into it and can be read by other
2740                          * threads.
2741                          *
2742                          * The critical issue is to order this
2743                          * page_remove_rmap with the ptp_clear_flush above.
2744                          * Those stores are ordered by (if nothing else,)
2745                          * the barrier present in the atomic_add_negative
2746                          * in page_remove_rmap.
2747                          *
2748                          * Then the TLB flush in ptep_clear_flush ensures that
2749                          * no process can access the old page before the
2750                          * decremented mapcount is visible. And the old page
2751                          * cannot be reused until after the decremented
2752                          * mapcount is visible. So transitively, TLBs to
2753                          * old page will be flushed before it can be reused.
2754                          */
2755                         page_remove_rmap(old_page);
2756                 }
2757
2758                 /* Free the old page.. */
2759                 new_page = old_page;
2760                 ret |= VM_FAULT_WRITE;
2761         } else
2762                 mem_cgroup_uncharge_page(new_page);
2763
2764         if (new_page)
2765                 page_cache_release(new_page);
2766 unlock:
2767         pte_unmap_unlock(page_table, ptl);
2768         if (old_page) {
2769                 /*
2770                  * Don't let another task, with possibly unlocked vma,
2771                  * keep the mlocked page.
2772                  */
2773                 if ((ret & VM_FAULT_WRITE) && (vma->vm_flags & VM_LOCKED)) {
2774                         lock_page(old_page);    /* LRU manipulation */
2775                         munlock_vma_page(old_page);
2776                         unlock_page(old_page);
2777                 }
2778                 page_cache_release(old_page);
2779         }
2780         return ret;
2781 oom_free_new:
2782         page_cache_release(new_page);
2783 oom:
2784         if (old_page) {
2785                 if (page_mkwrite) {
2786                         unlock_page(old_page);
2787                         page_cache_release(old_page);
2788                 }
2789                 page_cache_release(old_page);
2790         }
2791         return VM_FAULT_OOM;
2792
2793 unwritable_page:
2794         page_cache_release(old_page);
2795         return ret;
2796 }
2797
2798 static void unmap_mapping_range_vma(struct vm_area_struct *vma,
2799                 unsigned long start_addr, unsigned long end_addr,
2800                 struct zap_details *details)
2801 {
2802         zap_page_range_single(vma, start_addr, end_addr - start_addr, details);
2803 }
2804
2805 static inline void unmap_mapping_range_tree(struct prio_tree_root *root,
2806                                             struct zap_details *details)
2807 {
2808         struct vm_area_struct *vma;
2809         struct prio_tree_iter iter;
2810         pgoff_t vba, vea, zba, zea;
2811
2812         vma_prio_tree_foreach(vma, &iter, root,
2813                         details->first_index, details->last_index) {
2814
2815                 vba = vma->vm_pgoff;
2816                 vea = vba + ((vma->vm_end - vma->vm_start) >> PAGE_SHIFT) - 1;
2817                 /* Assume for now that PAGE_CACHE_SHIFT == PAGE_SHIFT */
2818                 zba = details->first_index;
2819                 if (zba < vba)
2820                         zba = vba;
2821                 zea = details->last_index;
2822                 if (zea > vea)
2823                         zea = vea;
2824
2825                 unmap_mapping_range_vma(vma,
2826                         ((zba - vba) << PAGE_SHIFT) + vma->vm_start,
2827                         ((zea - vba + 1) << PAGE_SHIFT) + vma->vm_start,
2828                                 details);
2829         }
2830 }
2831
2832 static inline void unmap_mapping_range_list(struct list_head *head,
2833                                             struct zap_details *details)
2834 {
2835         struct vm_area_struct *vma;
2836
2837         /*
2838          * In nonlinear VMAs there is no correspondence between virtual address
2839          * offset and file offset.  So we must perform an exhaustive search
2840          * across *all* the pages in each nonlinear VMA, not just the pages
2841          * whose virtual address lies outside the file truncation point.
2842          */
2843         list_for_each_entry(vma, head, shared.vm_set.list) {
2844                 details->nonlinear_vma = vma;
2845                 unmap_mapping_range_vma(vma, vma->vm_start, vma->vm_end, details);
2846         }
2847 }
2848
2849 /**
2850  * unmap_mapping_range - unmap the portion of all mmaps in the specified address_space corresponding to the specified page range in the underlying file.
2851  * @mapping: the address space containing mmaps to be unmapped.
2852  * @holebegin: byte in first page to unmap, relative to the start of
2853  * the underlying file.  This will be rounded down to a PAGE_SIZE
2854  * boundary.  Note that this is different from truncate_pagecache(), which
2855  * must keep the partial page.  In contrast, we must get rid of
2856  * partial pages.
2857  * @holelen: size of prospective hole in bytes.  This will be rounded
2858  * up to a PAGE_SIZE boundary.  A holelen of zero truncates to the
2859  * end of the file.
2860  * @even_cows: 1 when truncating a file, unmap even private COWed pages;
2861  * but 0 when invalidating pagecache, don't throw away private data.
2862  */
2863 void unmap_mapping_range(struct address_space *mapping,
2864                 loff_t const holebegin, loff_t const holelen, int even_cows)
2865 {
2866         struct zap_details details;
2867         pgoff_t hba = holebegin >> PAGE_SHIFT;
2868         pgoff_t hlen = (holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
2869
2870         /* Check for overflow. */
2871         if (sizeof(holelen) > sizeof(hlen)) {
2872                 long long holeend =
2873                         (holebegin + holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
2874                 if (holeend & ~(long long)ULONG_MAX)
2875                         hlen = ULONG_MAX - hba + 1;
2876         }
2877
2878         details.check_mapping = even_cows? NULL: mapping;
2879         details.nonlinear_vma = NULL;
2880         details.first_index = hba;
2881         details.last_index = hba + hlen - 1;
2882         if (details.last_index < details.first_index)
2883                 details.last_index = ULONG_MAX;
2884
2885
2886         mutex_lock(&mapping->i_mmap_mutex);
2887         if (unlikely(!prio_tree_empty(&mapping->i_mmap)))
2888                 unmap_mapping_range_tree(&mapping->i_mmap, &details);
2889         if (unlikely(!list_empty(&mapping->i_mmap_nonlinear)))
2890                 unmap_mapping_range_list(&mapping->i_mmap_nonlinear, &details);
2891         mutex_unlock(&mapping->i_mmap_mutex);
2892 }
2893 EXPORT_SYMBOL(unmap_mapping_range);
2894
2895 /*
2896  * We enter with non-exclusive mmap_sem (to exclude vma changes,
2897  * but allow concurrent faults), and pte mapped but not yet locked.
2898  * We return with mmap_sem still held, but pte unmapped and unlocked.
2899  */
2900 static int do_swap_page(struct mm_struct *mm, struct vm_area_struct *vma,
2901                 unsigned long address, pte_t *page_table, pmd_t *pmd,
2902                 unsigned int flags, pte_t orig_pte)
2903 {
2904         spinlock_t *ptl;
2905         struct page *page, *swapcache = NULL;
2906         swp_entry_t entry;
2907         pte_t pte;
2908         int locked;
2909         struct mem_cgroup *ptr;
2910         int exclusive = 0;
2911         int ret = 0;
2912
2913         if (!pte_unmap_same(mm, pmd, page_table, orig_pte))
2914                 goto out;
2915
2916         entry = pte_to_swp_entry(orig_pte);
2917         if (unlikely(non_swap_entry(entry))) {
2918                 if (is_migration_entry(entry)) {
2919                         migration_entry_wait(mm, pmd, address);
2920                 } else if (is_hwpoison_entry(entry)) {
2921                         ret = VM_FAULT_HWPOISON;
2922                 } else {
2923                         print_bad_pte(vma, address, orig_pte, NULL);
2924                         ret = VM_FAULT_SIGBUS;
2925                 }
2926                 goto out;
2927         }
2928         delayacct_set_flag(DELAYACCT_PF_SWAPIN);
2929         page = lookup_swap_cache(entry);
2930         if (!page) {
2931                 page = swapin_readahead(entry,
2932                                         GFP_HIGHUSER_MOVABLE, vma, address);
2933                 if (!page) {
2934                         /*
2935                          * Back out if somebody else faulted in this pte
2936                          * while we released the pte lock.
2937                          */
2938                         page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2939                         if (likely(pte_same(*page_table, orig_pte)))
2940                                 ret = VM_FAULT_OOM;
2941                         delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
2942                         goto unlock;
2943                 }
2944
2945                 /* Had to read the page from swap area: Major fault */
2946                 ret = VM_FAULT_MAJOR;
2947                 count_vm_event(PGMAJFAULT);
2948                 mem_cgroup_count_vm_event(mm, PGMAJFAULT);
2949         } else if (PageHWPoison(page)) {
2950                 /*
2951                  * hwpoisoned dirty swapcache pages are kept for killing
2952                  * owner processes (which may be unknown at hwpoison time)
2953                  */
2954                 ret = VM_FAULT_HWPOISON;
2955                 delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
2956                 goto out_release;
2957         }
2958
2959         locked = lock_page_or_retry(page, mm, flags);
2960
2961         delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
2962         if (!locked) {
2963                 ret |= VM_FAULT_RETRY;
2964                 goto out_release;
2965         }
2966
2967         /*
2968          * Make sure try_to_free_swap or reuse_swap_page or swapoff did not
2969          * release the swapcache from under us.  The page pin, and pte_same
2970          * test below, are not enough to exclude that.  Even if it is still
2971          * swapcache, we need to check that the page's swap has not changed.
2972          */
2973         if (unlikely(!PageSwapCache(page) || page_private(page) != entry.val))
2974                 goto out_page;
2975
2976         if (ksm_might_need_to_copy(page, vma, address)) {
2977                 swapcache = page;
2978                 page = ksm_does_need_to_copy(page, vma, address);
2979
2980                 if (unlikely(!page)) {
2981                         ret = VM_FAULT_OOM;
2982                         page = swapcache;
2983                         swapcache = NULL;
2984                         goto out_page;
2985                 }
2986         }
2987
2988         if (mem_cgroup_try_charge_swapin(mm, page, GFP_KERNEL, &ptr)) {
2989                 ret = VM_FAULT_OOM;
2990                 goto out_page;
2991         }
2992
2993         /*
2994          * Back out if somebody else already faulted in this pte.
2995          */
2996         page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2997         if (unlikely(!pte_same(*page_table, orig_pte)))
2998                 goto out_nomap;
2999
3000         if (unlikely(!PageUptodate(page))) {
3001                 ret = VM_FAULT_SIGBUS;
3002                 goto out_nomap;
3003         }
3004
3005         /*
3006          * The page isn't present yet, go ahead with the fault.
3007          *
3008          * Be careful about the sequence of operations here.
3009          * To get its accounting right, reuse_swap_page() must be called
3010          * while the page is counted on swap but not yet in mapcount i.e.
3011          * before page_add_anon_rmap() and swap_free(); try_to_free_swap()
3012          * must be called after the swap_free(), or it will never succeed.
3013          * Because delete_from_swap_page() may be called by reuse_swap_page(),
3014          * mem_cgroup_commit_charge_swapin() may not be able to find swp_entry
3015          * in page->private. In this case, a record in swap_cgroup  is silently
3016          * discarded at swap_free().
3017          */
3018
3019         inc_mm_counter_fast(mm, MM_ANONPAGES);
3020         dec_mm_counter_fast(mm, MM_SWAPENTS);
3021         pte = mk_pte(page, vma->vm_page_prot);
3022         if ((flags & FAULT_FLAG_WRITE) && reuse_swap_page(page)) {
3023                 pte = maybe_mkwrite(pte_mkdirty(pte), vma);
3024                 flags &= ~FAULT_FLAG_WRITE;
3025                 ret |= VM_FAULT_WRITE;
3026                 exclusive = 1;
3027         }
3028         flush_icache_page(vma, page);
3029         set_pte_at(mm, address, page_table, pte);
3030         do_page_add_anon_rmap(page, vma, address, exclusive);
3031         /* It's better to call commit-charge after rmap is established */
3032         mem_cgroup_commit_charge_swapin(page, ptr);
3033
3034         swap_free(entry);
3035         if (vm_swap_full() || (vma->vm_flags & VM_LOCKED) || PageMlocked(page))
3036                 try_to_free_swap(page);
3037         unlock_page(page);
3038         if (swapcache) {
3039                 /*
3040                  * Hold the lock to avoid the swap entry to be reused
3041                  * until we take the PT lock for the pte_same() check
3042                  * (to avoid false positives from pte_same). For
3043                  * further safety release the lock after the swap_free
3044                  * so that the swap count won't change under a
3045                  * parallel locked swapcache.
3046                  */
3047                 unlock_page(swapcache);
3048                 page_cache_release(swapcache);
3049         }
3050
3051         if (flags & FAULT_FLAG_WRITE) {
3052                 ret |= do_wp_page(mm, vma, address, page_table, pmd, ptl, pte);
3053                 if (ret & VM_FAULT_ERROR)
3054                         ret &= VM_FAULT_ERROR;
3055                 goto out;
3056         }
3057
3058         /* No need to invalidate - it was non-present before */
3059         update_mmu_cache(vma, address, page_table);
3060 unlock:
3061         pte_unmap_unlock(page_table, ptl);
3062 out:
3063         return ret;
3064 out_nomap:
3065         mem_cgroup_cancel_charge_swapin(ptr);
3066         pte_unmap_unlock(page_table, ptl);
3067 out_page:
3068         unlock_page(page);
3069 out_release:
3070         page_cache_release(page);
3071         if (swapcache) {
3072                 unlock_page(swapcache);
3073                 page_cache_release(swapcache);
3074         }
3075         return ret;
3076 }
3077
3078 /*
3079  * This is like a special single-page "expand_{down|up}wards()",
3080  * except we must first make sure that 'address{-|+}PAGE_SIZE'
3081  * doesn't hit another vma.
3082  */
3083 static inline int check_stack_guard_page(struct vm_area_struct *vma, unsigned long address)
3084 {
3085         address &= PAGE_MASK;
3086         if ((vma->vm_flags & VM_GROWSDOWN) && address == vma->vm_start) {
3087                 struct vm_area_struct *prev = vma->vm_prev;
3088
3089                 /*
3090                  * Is there a mapping abutting this one below?
3091                  *
3092                  * That's only ok if it's the same stack mapping
3093                  * that has gotten split..
3094                  */
3095                 if (prev && prev->vm_end == address)
3096                         return prev->vm_flags & VM_GROWSDOWN ? 0 : -ENOMEM;
3097
3098                 expand_downwards(vma, address - PAGE_SIZE);
3099         }
3100         if ((vma->vm_flags & VM_GROWSUP) && address + PAGE_SIZE == vma->vm_end) {
3101                 struct vm_area_struct *next = vma->vm_next;
3102
3103                 /* As VM_GROWSDOWN but s/below/above/ */
3104                 if (next && next->vm_start == address + PAGE_SIZE)
3105                         return next->vm_flags & VM_GROWSUP ? 0 : -ENOMEM;
3106
3107                 expand_upwards(vma, address + PAGE_SIZE);
3108         }
3109         return 0;
3110 }
3111
3112 /*
3113  * We enter with non-exclusive mmap_sem (to exclude vma changes,
3114  * but allow concurrent faults), and pte mapped but not yet locked.
3115  * We return with mmap_sem still held, but pte unmapped and unlocked.
3116  */
3117 static int do_anonymous_page(struct mm_struct *mm, struct vm_area_struct *vma,
3118                 unsigned long address, pte_t *page_table, pmd_t *pmd,
3119                 unsigned int flags)
3120 {
3121         struct page *page;
3122         spinlock_t *ptl;
3123         pte_t entry;
3124
3125         pte_unmap(page_table);
3126
3127         /* Check if we need to add a guard page to the stack */
3128         if (check_stack_guard_page(vma, address) < 0)
3129                 return VM_FAULT_SIGBUS;
3130
3131         /* Use the zero-page for reads */
3132         if (!(flags & FAULT_FLAG_WRITE)) {
3133                 entry = pte_mkspecial(pfn_pte(my_zero_pfn(address),
3134                                                 vma->vm_page_prot));
3135                 page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
3136                 if (!pte_none(*page_table))
3137                         goto unlock;
3138                 goto setpte;
3139         }
3140
3141         /* Allocate our own private page. */
3142         if (unlikely(anon_vma_prepare(vma)))
3143                 goto oom;
3144         page = alloc_zeroed_user_highpage_movable(vma, address);
3145         if (!page)
3146                 goto oom;
3147         __SetPageUptodate(page);
3148
3149         if (mem_cgroup_newpage_charge(page, mm, GFP_KERNEL))
3150                 goto oom_free_page;
3151
3152         entry = mk_pte(page, vma->vm_page_prot);
3153         if (vma->vm_flags & VM_WRITE)
3154                 entry = pte_mkwrite(pte_mkdirty(entry));
3155
3156         page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
3157         if (!pte_none(*page_table))
3158                 goto release;
3159
3160         inc_mm_counter_fast(mm, MM_ANONPAGES);
3161         page_add_new_anon_rmap(page, vma, address);
3162 setpte:
3163         set_pte_at(mm, address, page_table, entry);
3164
3165         /* No need to invalidate - it was non-present before */
3166         update_mmu_cache(vma, address, page_table);
3167 unlock:
3168         pte_unmap_unlock(page_table, ptl);
3169         return 0;
3170 release:
3171         mem_cgroup_uncharge_page(page);
3172         page_cache_release(page);
3173         goto unlock;
3174 oom_free_page:
3175         page_cache_release(page);
3176 oom:
3177         return VM_FAULT_OOM;
3178 }
3179
3180 /*
3181  * __do_fault() tries to create a new page mapping. It aggressively
3182  * tries to share with existing pages, but makes a separate copy if
3183  * the FAULT_FLAG_WRITE is set in the flags parameter in order to avoid
3184  * the next page fault.
3185  *
3186  * As this is called only for pages that do not currently exist, we
3187  * do not need to flush old virtual caches or the TLB.
3188  *
3189  * We enter with non-exclusive mmap_sem (to exclude vma changes,
3190  * but allow concurrent faults), and pte neither mapped nor locked.
3191  * We return with mmap_sem still held, but pte unmapped and unlocked.
3192  */
3193 static int __do_fault(struct mm_struct *mm, struct vm_area_struct *vma,
3194                 unsigned long address, pmd_t *pmd,
3195                 pgoff_t pgoff, unsigned int flags, pte_t orig_pte)
3196 {
3197         pte_t *page_table;
3198         spinlock_t *ptl;
3199         struct page *page;
3200         struct page *cow_page;
3201         pte_t entry;
3202         int anon = 0;
3203         struct page *dirty_page = NULL;
3204         struct vm_fault vmf;
3205         int ret;
3206         int page_mkwrite = 0;
3207
3208         /*
3209          * If we do COW later, allocate page befor taking lock_page()
3210          * on the file cache page. This will reduce lock holding time.
3211          */
3212         if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) {
3213
3214                 if (unlikely(anon_vma_prepare(vma)))
3215                         return VM_FAULT_OOM;
3216
3217                 cow_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, address);
3218                 if (!cow_page)
3219                         return VM_FAULT_OOM;
3220
3221                 if (mem_cgroup_newpage_charge(cow_page, mm, GFP_KERNEL)) {
3222                         page_cache_release(cow_page);
3223                         return VM_FAULT_OOM;
3224                 }
3225         } else
3226                 cow_page = NULL;
3227
3228         vmf.virtual_address = (void __user *)(address & PAGE_MASK);
3229         vmf.pgoff = pgoff;
3230         vmf.flags = flags;
3231         vmf.page = NULL;
3232
3233         ret = vma->vm_ops->fault(vma, &vmf);
3234         if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE |
3235                             VM_FAULT_RETRY)))
3236                 goto uncharge_out;
3237
3238         if (unlikely(PageHWPoison(vmf.page))) {
3239                 if (ret & VM_FAULT_LOCKED)
3240                         unlock_page(vmf.page);
3241                 ret = VM_FAULT_HWPOISON;
3242                 goto uncharge_out;
3243         }
3244
3245         /*
3246          * For consistency in subsequent calls, make the faulted page always
3247          * locked.
3248          */
3249         if (unlikely(!(ret & VM_FAULT_LOCKED)))
3250                 lock_page(vmf.page);
3251         else
3252                 VM_BUG_ON(!PageLocked(vmf.page));
3253
3254         /*
3255          * Should we do an early C-O-W break?
3256          */
3257         page = vmf.page;
3258         if (flags & FAULT_FLAG_WRITE) {
3259                 if (!(vma->vm_flags & VM_SHARED)) {
3260                         page = cow_page;
3261                         anon = 1;
3262                         copy_user_highpage(page, vmf.page, address, vma);
3263                         __SetPageUptodate(page);
3264                 } else {
3265                         /*
3266                          * If the page will be shareable, see if the backing
3267                          * address space wants to know that the page is about
3268                          * to become writable
3269                          */
3270                         if (vma->vm_ops->page_mkwrite) {
3271                                 int tmp;
3272
3273                                 unlock_page(page);
3274                                 vmf.flags = FAULT_FLAG_WRITE|FAULT_FLAG_MKWRITE;
3275                                 tmp = vma->vm_ops->page_mkwrite(vma, &vmf);
3276                                 if (unlikely(tmp &
3277                                           (VM_FAULT_ERROR | VM_FAULT_NOPAGE))) {
3278                                         ret = tmp;
3279                                         goto unwritable_page;
3280                                 }
3281                                 if (unlikely(!(tmp & VM_FAULT_LOCKED))) {
3282                                         lock_page(page);
3283                                         if (!page->mapping) {
3284                                                 ret = 0; /* retry the fault */
3285                                                 unlock_page(page);
3286                                                 goto unwritable_page;
3287                                         }
3288                                 } else
3289                                         VM_BUG_ON(!PageLocked(page));
3290                                 page_mkwrite = 1;
3291                         }
3292                 }
3293
3294         }
3295
3296         page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
3297
3298         /*
3299          * This silly early PAGE_DIRTY setting removes a race
3300          * due to the bad i386 page protection. But it's valid
3301          * for other architectures too.
3302          *
3303          * Note that if FAULT_FLAG_WRITE is set, we either now have
3304          * an exclusive copy of the page, or this is a shared mapping,
3305          * so we can make it writable and dirty to avoid having to
3306          * handle that later.
3307          */
3308         /* Only go through if we didn't race with anybody else... */
3309         if (likely(pte_same(*page_table, orig_pte))) {
3310                 flush_icache_page(vma, page);
3311                 entry = mk_pte(page, vma->vm_page_prot);
3312                 if (flags & FAULT_FLAG_WRITE)
3313                         entry = maybe_mkwrite(pte_mkdirty(entry), vma);
3314                 if (anon) {
3315                         inc_mm_counter_fast(mm, MM_ANONPAGES);
3316                         page_add_new_anon_rmap(page, vma, address);
3317                 } else {
3318                         inc_mm_counter_fast(mm, MM_FILEPAGES);
3319                         page_add_file_rmap(page);
3320                         if (flags & FAULT_FLAG_WRITE) {
3321                                 dirty_page = page;
3322                                 get_page(dirty_page);
3323                         }
3324                 }
3325                 set_pte_at(mm, address, page_table, entry);
3326
3327                 /* no need to invalidate: a not-present page won't be cached */
3328                 update_mmu_cache(vma, address, page_table);
3329         } else {
3330                 if (cow_page)
3331                         mem_cgroup_uncharge_page(cow_page);
3332                 if (anon)
3333                         page_cache_release(page);
3334                 else
3335                         anon = 1; /* no anon but release faulted_page */
3336         }
3337
3338         pte_unmap_unlock(page_table, ptl);
3339
3340         if (dirty_page) {
3341                 struct address_space *mapping = page->mapping;
3342                 int dirtied = 0;
3343
3344                 if (set_page_dirty(dirty_page))
3345                         dirtied = 1;
3346                 unlock_page(dirty_page);
3347                 put_page(dirty_page);
3348                 if ((dirtied || page_mkwrite) && mapping) {
3349                         /*
3350                          * Some device drivers do not set page.mapping but still
3351                          * dirty their pages
3352                          */
3353                         balance_dirty_pages_ratelimited(mapping);
3354                 }
3355
3356                 /* file_update_time outside page_lock */
3357                 if (vma->vm_file && !page_mkwrite)
3358                         file_update_time(vma->vm_file);
3359         } else {
3360                 unlock_page(vmf.page);
3361                 if (anon)
3362                         page_cache_release(vmf.page);
3363         }
3364
3365         return ret;
3366
3367 unwritable_page:
3368         page_cache_release(page);
3369         return ret;
3370 uncharge_out:
3371         /* fs's fault handler get error */
3372         if (cow_page) {
3373                 mem_cgroup_uncharge_page(cow_page);
3374                 page_cache_release(cow_page);
3375         }
3376         return ret;
3377 }
3378
3379 static int do_linear_fault(struct mm_struct *mm, struct vm_area_struct *vma,
3380                 unsigned long address, pte_t *page_table, pmd_t *pmd,
3381                 unsigned int flags, pte_t orig_pte)
3382 {
3383         pgoff_t pgoff = (((address & PAGE_MASK)
3384                         - vma->vm_start) >> PAGE_SHIFT) + vma->vm_pgoff;
3385
3386         pte_unmap(page_table);
3387         return __do_fault(mm, vma, address, pmd, pgoff, flags, orig_pte);
3388 }
3389
3390 /*
3391  * Fault of a previously existing named mapping. Repopulate the pte
3392  * from the encoded file_pte if possible. This enables swappable
3393  * nonlinear vmas.
3394  *
3395  * We enter with non-exclusive mmap_sem (to exclude vma changes,
3396  * but allow concurrent faults), and pte mapped but not yet locked.
3397  * We return with mmap_sem still held, but pte unmapped and unlocked.
3398  */
3399 static int do_nonlinear_fault(struct mm_struct *mm, struct vm_area_struct *vma,
3400                 unsigned long address, pte_t *page_table, pmd_t *pmd,
3401                 unsigned int flags, pte_t orig_pte)
3402 {
3403         pgoff_t pgoff;
3404
3405         flags |= FAULT_FLAG_NONLINEAR;
3406
3407         if (!pte_unmap_same(mm, pmd, page_table, orig_pte))
3408                 return 0;
3409
3410         if (unlikely(!(vma->vm_flags & VM_NONLINEAR))) {
3411                 /*
3412                  * Page table corrupted: show pte and kill process.
3413                  */
3414                 print_bad_pte(vma, address, orig_pte, NULL);
3415                 return VM_FAULT_SIGBUS;
3416         }
3417
3418         pgoff = pte_to_pgoff(orig_pte);
3419         return __do_fault(mm, vma, address, pmd, pgoff, flags, orig_pte);
3420 }
3421
3422 /*
3423  * These routines also need to handle stuff like marking pages dirty
3424  * and/or accessed for architectures that don't do it in hardware (most
3425  * RISC architectures).  The early dirtying is also good on the i386.
3426  *
3427  * There is also a hook called "update_mmu_cache()" that architectures
3428  * with external mmu caches can use to update those (ie the Sparc or
3429  * PowerPC hashed page tables that act as extended TLBs).
3430  *
3431  * We enter with non-exclusive mmap_sem (to exclude vma changes,
3432  * but allow concurrent faults), and pte mapped but not yet locked.
3433  * We return with mmap_sem still held, but pte unmapped and unlocked.
3434  */
3435 int handle_pte_fault(struct mm_struct *mm,
3436                      struct vm_area_struct *vma, unsigned long address,
3437                      pte_t *pte, pmd_t *pmd, unsigned int flags)
3438 {
3439         pte_t entry;
3440         spinlock_t *ptl;
3441
3442         entry = *pte;
3443         if (!pte_present(entry)) {
3444                 if (pte_none(entry)) {
3445                         if (vma->vm_ops) {
3446                                 if (likely(vma->vm_ops->fault))
3447                                         return do_linear_fault(mm, vma, address,
3448                                                 pte, pmd, flags, entry);
3449                         }
3450                         return do_anonymous_page(mm, vma, address,
3451                                                  pte, pmd, flags);
3452                 }
3453                 if (pte_file(entry))
3454                         return do_nonlinear_fault(mm, vma, address,
3455                                         pte, pmd, flags, entry);
3456                 return do_swap_page(mm, vma, address,
3457                                         pte, pmd, flags, entry);
3458         }
3459
3460         ptl = pte_lockptr(mm, pmd);
3461         spin_lock(ptl);
3462         if (unlikely(!pte_same(*pte, entry)))
3463                 goto unlock;
3464         if (flags & FAULT_FLAG_WRITE) {
3465                 if (!pte_write(entry))
3466                         return do_wp_page(mm, vma, address,
3467                                         pte, pmd, ptl, entry);
3468                 entry = pte_mkdirty(entry);
3469         }
3470         entry = pte_mkyoung(entry);
3471         if (ptep_set_access_flags(vma, address, pte, entry, flags & FAULT_FLAG_WRITE)) {
3472                 update_mmu_cache(vma, address, pte);
3473         } else {
3474                 /*
3475                  * This is needed only for protection faults but the arch code
3476                  * is not yet telling us if this is a protection fault or not.
3477                  * This still avoids useless tlb flushes for .text page faults
3478                  * with threads.
3479                  */
3480                 if (flags & FAULT_FLAG_WRITE)
3481                         flush_tlb_fix_spurious_fault(vma, address);
3482         }
3483 unlock:
3484         pte_unmap_unlock(pte, ptl);
3485         return 0;
3486 }
3487
3488 /*
3489  * By the time we get here, we already hold the mm semaphore
3490  */
3491 int handle_mm_fault(struct mm_struct *mm, struct vm_area_struct *vma,
3492                 unsigned long address, unsigned int flags)
3493 {
3494         pgd_t *pgd;
3495         pud_t *pud;
3496         pmd_t *pmd;
3497         pte_t *pte;
3498
3499         __set_current_state(TASK_RUNNING);
3500
3501         count_vm_event(PGFAULT);
3502         mem_cgroup_count_vm_event(mm, PGFAULT);
3503
3504         /* do counter updates before entering really critical section. */
3505         check_sync_rss_stat(current);
3506
3507         if (unlikely(is_vm_hugetlb_page(vma)))
3508                 return hugetlb_fault(mm, vma, address, flags);
3509
3510 retry:
3511         pgd = pgd_offset(mm, address);
3512         pud = pud_alloc(mm, pgd, address);
3513         if (!pud)
3514                 return VM_FAULT_OOM;
3515         pmd = pmd_alloc(mm, pud, address);
3516         if (!pmd)
3517                 return VM_FAULT_OOM;
3518         if (pmd_none(*pmd) && transparent_hugepage_enabled(vma)) {
3519                 if (!vma->vm_ops)
3520                         return do_huge_pmd_anonymous_page(mm, vma, address,
3521                                                           pmd, flags);
3522         } else {
3523                 pmd_t orig_pmd = *pmd;
3524                 int ret;
3525
3526                 barrier();
3527                 if (pmd_trans_huge(orig_pmd)) {
3528                         if (flags & FAULT_FLAG_WRITE &&
3529                             !pmd_write(orig_pmd) &&
3530                             !pmd_trans_splitting(orig_pmd)) {
3531                                 ret = do_huge_pmd_wp_page(mm, vma, address, pmd,
3532                                                           orig_pmd);
3533                                 /*
3534                                  * If COW results in an oom, the huge pmd will
3535                                  * have been split, so retry the fault on the
3536                                  * pte for a smaller charge.
3537                                  */
3538                                 if (unlikely(ret & VM_FAULT_OOM))
3539                                         goto retry;
3540                                 return ret;
3541                         }
3542                         return 0;
3543                 }
3544         }
3545
3546         /*
3547          * Use __pte_alloc instead of pte_alloc_map, because we can't
3548          * run pte_offset_map on the pmd, if an huge pmd could
3549          * materialize from under us from a different thread.
3550          */
3551         if (unlikely(pmd_none(*pmd)) && __pte_alloc(mm, vma, pmd, address))
3552                 return VM_FAULT_OOM;
3553         /* if an huge pmd materialized from under us just retry later */
3554         if (unlikely(pmd_trans_huge(*pmd)))
3555                 return 0;
3556         /*
3557          * A regular pmd is established and it can't morph into a huge pmd
3558          * from under us anymore at this point because we hold the mmap_sem
3559          * read mode and khugepaged takes it in write mode. So now it's
3560          * safe to run pte_offset_map().
3561          */
3562         pte = pte_offset_map(pmd, address);
3563
3564         return handle_pte_fault(mm, vma, address, pte, pmd, flags);
3565 }
3566
3567 #ifndef __PAGETABLE_PUD_FOLDED
3568 /*
3569  * Allocate page upper directory.
3570  * We've already handled the fast-path in-line.
3571  */
3572 int __pud_alloc(struct mm_struct *mm, pgd_t *pgd, unsigned long address)
3573 {
3574         pud_t *new = pud_alloc_one(mm, address);
3575         if (!new)
3576                 return -ENOMEM;
3577
3578         smp_wmb(); /* See comment in __pte_alloc */
3579
3580         spin_lock(&mm->page_table_lock);
3581         if (pgd_present(*pgd))          /* Another has populated it */
3582                 pud_free(mm, new);
3583         else
3584                 pgd_populate(mm, pgd, new);
3585         spin_unlock(&mm->page_table_lock);
3586         return 0;
3587 }
3588 #endif /* __PAGETABLE_PUD_FOLDED */
3589
3590 #ifndef __PAGETABLE_PMD_FOLDED
3591 /*
3592  * Allocate page middle directory.
3593  * We've already handled the fast-path in-line.
3594  */
3595 int __pmd_alloc(struct mm_struct *mm, pud_t *pud, unsigned long address)
3596 {
3597         pmd_t *new = pmd_alloc_one(mm, address);
3598         if (!new)
3599                 return -ENOMEM;
3600
3601         smp_wmb(); /* See comment in __pte_alloc */
3602
3603         spin_lock(&mm->page_table_lock);
3604 #ifndef __ARCH_HAS_4LEVEL_HACK
3605         if (pud_present(*pud))          /* Another has populated it */
3606                 pmd_free(mm, new);
3607         else
3608                 pud_populate(mm, pud, new);
3609 #else
3610         if (pgd_present(*pud))          /* Another has populated it */
3611                 pmd_free(mm, new);
3612         else
3613                 pgd_populate(mm, pud, new);
3614 #endif /* __ARCH_HAS_4LEVEL_HACK */
3615         spin_unlock(&mm->page_table_lock);
3616         return 0;
3617 }
3618 #endif /* __PAGETABLE_PMD_FOLDED */
3619
3620 int make_pages_present(unsigned long addr, unsigned long end)
3621 {
3622         int ret, len, write;
3623         struct vm_area_struct * vma;
3624
3625         vma = find_vma(current->mm, addr);
3626         if (!vma)
3627                 return -ENOMEM;
3628         /*
3629          * We want to touch writable mappings with a write fault in order
3630          * to break COW, except for shared mappings because these don't COW
3631          * and we would not want to dirty them for nothing.
3632          */
3633         write = (vma->vm_flags & (VM_WRITE | VM_SHARED)) == VM_WRITE;
3634         BUG_ON(addr >= end);
3635         BUG_ON(end > vma->vm_end);
3636         len = DIV_ROUND_UP(end, PAGE_SIZE) - addr/PAGE_SIZE;
3637         ret = get_user_pages(current, current->mm, addr,
3638                         len, write, 0, NULL, NULL);
3639         if (ret < 0)
3640                 return ret;
3641         return ret == len ? 0 : -EFAULT;
3642 }
3643
3644 #if !defined(__HAVE_ARCH_GATE_AREA)
3645
3646 #if defined(AT_SYSINFO_EHDR)
3647 static struct vm_area_struct gate_vma;
3648
3649 static int __init gate_vma_init(void)
3650 {
3651         gate_vma.vm_mm = NULL;
3652         gate_vma.vm_start = FIXADDR_USER_START;
3653         gate_vma.vm_end = FIXADDR_USER_END;
3654         gate_vma.vm_flags = VM_READ | VM_MAYREAD | VM_EXEC | VM_MAYEXEC;
3655         gate_vma.vm_page_prot = __P101;
3656
3657         return 0;
3658 }
3659 __initcall(gate_vma_init);
3660 #endif
3661
3662 struct vm_area_struct *get_gate_vma(struct mm_struct *mm)
3663 {
3664 #ifdef AT_SYSINFO_EHDR
3665         return &gate_vma;
3666 #else
3667         return NULL;
3668 #endif
3669 }
3670
3671 int in_gate_area_no_mm(unsigned long addr)
3672 {
3673 #ifdef AT_SYSINFO_EHDR
3674         if ((addr >= FIXADDR_USER_START) && (addr < FIXADDR_USER_END))
3675                 return 1;
3676 #endif
3677         return 0;
3678 }
3679
3680 #endif  /* __HAVE_ARCH_GATE_AREA */
3681
3682 static int __follow_pte(struct mm_struct *mm, unsigned long address,
3683                 pte_t **ptepp, spinlock_t **ptlp)
3684 {
3685         pgd_t *pgd;
3686         pud_t *pud;
3687         pmd_t *pmd;
3688         pte_t *ptep;
3689
3690         pgd = pgd_offset(mm, address);
3691         if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
3692                 goto out;
3693
3694         pud = pud_offset(pgd, address);
3695         if (pud_none(*pud) || unlikely(pud_bad(*pud)))
3696                 goto out;
3697
3698         pmd = pmd_offset(pud, address);
3699         VM_BUG_ON(pmd_trans_huge(*pmd));
3700         if (pmd_none(*pmd) || unlikely(pmd_bad(*pmd)))
3701                 goto out;
3702
3703         /* We cannot handle huge page PFN maps. Luckily they don't exist. */
3704         if (pmd_huge(*pmd))
3705                 goto out;
3706
3707         ptep = pte_offset_map_lock(mm, pmd, address, ptlp);
3708         if (!ptep)
3709                 goto out;
3710         if (!pte_present(*ptep))
3711                 goto unlock;
3712         *ptepp = ptep;
3713         return 0;
3714 unlock:
3715         pte_unmap_unlock(ptep, *ptlp);
3716 out:
3717         return -EINVAL;
3718 }
3719
3720 static inline int follow_pte(struct mm_struct *mm, unsigned long address,
3721                              pte_t **ptepp, spinlock_t **ptlp)
3722 {
3723         int res;
3724
3725         /* (void) is needed to make gcc happy */
3726         (void) __cond_lock(*ptlp,
3727                            !(res = __follow_pte(mm, address, ptepp, ptlp)));
3728         return res;
3729 }
3730
3731 /**
3732  * follow_pfn - look up PFN at a user virtual address
3733  * @vma: memory mapping
3734  * @address: user virtual address
3735  * @pfn: location to store found PFN
3736  *
3737  * Only IO mappings and raw PFN mappings are allowed.
3738  *
3739  * Returns zero and the pfn at @pfn on success, -ve otherwise.
3740  */
3741 int follow_pfn(struct vm_area_struct *vma, unsigned long address,
3742         unsigned long *pfn)
3743 {
3744         int ret = -EINVAL;
3745         spinlock_t *ptl;
3746         pte_t *ptep;
3747
3748         if (!(vma->vm_flags & (VM_IO | VM_PFNMAP)))
3749                 return ret;
3750
3751         ret = follow_pte(vma->vm_mm, address, &ptep, &ptl);
3752         if (ret)
3753                 return ret;
3754         *pfn = pte_pfn(*ptep);
3755         pte_unmap_unlock(ptep, ptl);
3756         return 0;
3757 }
3758 EXPORT_SYMBOL(follow_pfn);
3759
3760 #ifdef CONFIG_HAVE_IOREMAP_PROT
3761 int follow_phys(struct vm_area_struct *vma,
3762                 unsigned long address, unsigned int flags,
3763                 unsigned long *prot, resource_size_t *phys)
3764 {
3765         int ret = -EINVAL;
3766         pte_t *ptep, pte;
3767         spinlock_t *ptl;
3768
3769         if (!(vma->vm_flags & (VM_IO | VM_PFNMAP)))
3770                 goto out;
3771
3772         if (follow_pte(vma->vm_mm, address, &ptep, &ptl))
3773                 goto out;
3774         pte = *ptep;
3775
3776         if ((flags & FOLL_WRITE) && !pte_write(pte))
3777                 goto unlock;
3778
3779         *prot = pgprot_val(pte_pgprot(pte));
3780         *phys = (resource_size_t)pte_pfn(pte) << PAGE_SHIFT;
3781
3782         ret = 0;
3783 unlock:
3784         pte_unmap_unlock(ptep, ptl);
3785 out:
3786         return ret;
3787 }
3788
3789 int generic_access_phys(struct vm_area_struct *vma, unsigned long addr,
3790                         void *buf, int len, int write)
3791 {
3792         resource_size_t phys_addr;
3793         unsigned long prot = 0;
3794         void __iomem *maddr;
3795         int offset = addr & (PAGE_SIZE-1);
3796
3797         if (follow_phys(vma, addr, write, &prot, &phys_addr))
3798                 return -EINVAL;
3799
3800         maddr = ioremap_prot(phys_addr, PAGE_SIZE, prot);
3801         if (write)
3802                 memcpy_toio(maddr + offset, buf, len);
3803         else
3804                 memcpy_fromio(buf, maddr + offset, len);
3805         iounmap(maddr);
3806
3807         return len;
3808 }
3809 #endif
3810
3811 /*
3812  * Access another process' address space as given in mm.  If non-NULL, use the
3813  * given task for page fault accounting.
3814  */
3815 static int __access_remote_vm(struct task_struct *tsk, struct mm_struct *mm,
3816                 unsigned long addr, void *buf, int len, int write)
3817 {
3818         struct vm_area_struct *vma;
3819         void *old_buf = buf;
3820
3821         down_read(&mm->mmap_sem);
3822         /* ignore errors, just check how much was successfully transferred */
3823         while (len) {
3824                 int bytes, ret, offset;
3825                 void *maddr;
3826                 struct page *page = NULL;
3827
3828                 ret = get_user_pages(tsk, mm, addr, 1,
3829                                 write, 1, &page, &vma);
3830                 if (ret <= 0) {
3831                         /*
3832                          * Check if this is a VM_IO | VM_PFNMAP VMA, which
3833                          * we can access using slightly different code.
3834                          */
3835 #ifdef CONFIG_HAVE_IOREMAP_PROT
3836                         vma = find_vma(mm, addr);
3837                         if (!vma || vma->vm_start > addr)
3838                                 break;
3839                         if (vma->vm_ops && vma->vm_ops->access)
3840                                 ret = vma->vm_ops->access(vma, addr, buf,
3841                                                           len, write);
3842                         if (ret <= 0)
3843 #endif
3844                                 break;
3845                         bytes = ret;
3846                 } else {
3847                         bytes = len;
3848                         offset = addr & (PAGE_SIZE-1);
3849                         if (bytes > PAGE_SIZE-offset)
3850                                 bytes = PAGE_SIZE-offset;
3851
3852                         maddr = kmap(page);
3853                         if (write) {
3854                                 copy_to_user_page(vma, page, addr,
3855                                                   maddr + offset, buf, bytes);
3856                                 set_page_dirty_lock(page);
3857                         } else {
3858                                 copy_from_user_page(vma, page, addr,
3859                                                     buf, maddr + offset, bytes);
3860                         }
3861                         kunmap(page);
3862                         page_cache_release(page);
3863                 }
3864                 len -= bytes;
3865                 buf += bytes;
3866                 addr += bytes;
3867         }
3868         up_read(&mm->mmap_sem);
3869
3870         return buf - old_buf;
3871 }
3872
3873 /**
3874  * access_remote_vm - access another process' address space
3875  * @mm:         the mm_struct of the target address space
3876  * @addr:       start address to access
3877  * @buf:        source or destination buffer
3878  * @len:        number of bytes to transfer
3879  * @write:      whether the access is a write
3880  *
3881  * The caller must hold a reference on @mm.
3882  */
3883 int access_remote_vm(struct mm_struct *mm, unsigned long addr,
3884                 void *buf, int len, int write)
3885 {
3886         return __access_remote_vm(NULL, mm, addr, buf, len, write);
3887 }
3888
3889 /*
3890  * Access another process' address space.
3891  * Source/target buffer must be kernel space,
3892  * Do not walk the page table directly, use get_user_pages
3893  */
3894 int access_process_vm(struct task_struct *tsk, unsigned long addr,
3895                 void *buf, int len, int write)
3896 {
3897         struct mm_struct *mm;
3898         int ret;
3899
3900         mm = get_task_mm(tsk);
3901         if (!mm)
3902                 return 0;
3903
3904         ret = __access_remote_vm(tsk, mm, addr, buf, len, write);
3905         mmput(mm);
3906
3907         return ret;
3908 }
3909
3910 /*
3911  * Print the name of a VMA.
3912  */
3913 void print_vma_addr(char *prefix, unsigned long ip)
3914 {
3915         struct mm_struct *mm = current->mm;
3916         struct vm_area_struct *vma;
3917
3918         /*
3919          * Do not print if we are in atomic
3920          * contexts (in exception stacks, etc.):
3921          */
3922         if (preempt_count())
3923                 return;
3924
3925         down_read(&mm->mmap_sem);
3926         vma = find_vma(mm, ip);
3927         if (vma && vma->vm_file) {
3928                 struct file *f = vma->vm_file;
3929                 char *buf = (char *)__get_free_page(GFP_KERNEL);
3930                 if (buf) {
3931                         char *p, *s;
3932
3933                         p = d_path(&f->f_path, buf, PAGE_SIZE);
3934                         if (IS_ERR(p))
3935                                 p = "?";
3936                         s = strrchr(p, '/');
3937                         if (s)
3938                                 p = s+1;
3939                         printk("%s%s[%lx+%lx]", prefix, p,
3940                                         vma->vm_start,
3941                                         vma->vm_end - vma->vm_start);
3942                         free_page((unsigned long)buf);
3943                 }
3944         }
3945         up_read(&mm->mmap_sem);
3946 }
3947
3948 #ifdef CONFIG_PROVE_LOCKING
3949 void might_fault(void)
3950 {
3951         /*
3952          * Some code (nfs/sunrpc) uses socket ops on kernel memory while
3953          * holding the mmap_sem, this is safe because kernel memory doesn't
3954          * get paged out, therefore we'll never actually fault, and the
3955          * below annotations will generate false positives.
3956          */
3957         if (segment_eq(get_fs(), KERNEL_DS))
3958                 return;
3959
3960         might_sleep();
3961         /*
3962          * it would be nicer only to annotate paths which are not under
3963          * pagefault_disable, however that requires a larger audit and
3964          * providing helpers like get_user_atomic.
3965          */
3966         if (!in_atomic() && current->mm)
3967                 might_lock_read(&current->mm->mmap_sem);
3968 }
3969 EXPORT_SYMBOL(might_fault);
3970 #endif
3971
3972 #if defined(CONFIG_TRANSPARENT_HUGEPAGE) || defined(CONFIG_HUGETLBFS)
3973 static void clear_gigantic_page(struct page *page,
3974                                 unsigned long addr,
3975                                 unsigned int pages_per_huge_page)
3976 {
3977         int i;
3978         struct page *p = page;
3979
3980         might_sleep();
3981         for (i = 0; i < pages_per_huge_page;
3982              i++, p = mem_map_next(p, page, i)) {
3983                 cond_resched();
3984                 clear_user_highpage(p, addr + i * PAGE_SIZE);
3985         }
3986 }
3987 void clear_huge_page(struct page *page,
3988                      unsigned long addr, unsigned int pages_per_huge_page)
3989 {
3990         int i;
3991
3992         if (unlikely(pages_per_huge_page > MAX_ORDER_NR_PAGES)) {
3993                 clear_gigantic_page(page, addr, pages_per_huge_page);
3994                 return;
3995         }
3996
3997         might_sleep();
3998         for (i = 0; i < pages_per_huge_page; i++) {
3999                 cond_resched();
4000                 clear_user_highpage(page + i, addr + i * PAGE_SIZE);
4001         }
4002 }
4003
4004 static void copy_user_gigantic_page(struct page *dst, struct page *src,
4005                                     unsigned long addr,
4006                                     struct vm_area_struct *vma,
4007                                     unsigned int pages_per_huge_page)
4008 {
4009         int i;
4010         struct page *dst_base = dst;
4011         struct page *src_base = src;
4012
4013         for (i = 0; i < pages_per_huge_page; ) {
4014                 cond_resched();
4015                 copy_user_highpage(dst, src, addr + i*PAGE_SIZE, vma);
4016
4017                 i++;
4018                 dst = mem_map_next(dst, dst_base, i);
4019                 src = mem_map_next(src, src_base, i);
4020         }
4021 }
4022
4023 void copy_user_huge_page(struct page *dst, struct page *src,
4024                          unsigned long addr, struct vm_area_struct *vma,
4025                          unsigned int pages_per_huge_page)
4026 {
4027         int i;
4028
4029         if (unlikely(pages_per_huge_page > MAX_ORDER_NR_PAGES)) {
4030                 copy_user_gigantic_page(dst, src, addr, vma,
4031                                         pages_per_huge_page);
4032                 return;
4033         }
4034
4035         might_sleep();
4036         for (i = 0; i < pages_per_huge_page; i++) {
4037                 cond_resched();
4038                 copy_user_highpage(dst + i, src + i, addr + i*PAGE_SIZE, vma);
4039         }
4040 }
4041 #endif /* CONFIG_TRANSPARENT_HUGEPAGE || CONFIG_HUGETLBFS */