page_alloc: remove unused find_zone_movable_pfns_for_nodes() argument
[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->need_flush = 0;
210         tlb->fast_mode  = (num_possible_cpus() == 1);
211         tlb->local.next = NULL;
212         tlb->local.nr   = 0;
213         tlb->local.max  = ARRAY_SIZE(tlb->__pages);
214         tlb->active     = &tlb->local;
215
216 #ifdef CONFIG_HAVE_RCU_TABLE_FREE
217         tlb->batch = NULL;
218 #endif
219 }
220
221 void tlb_flush_mmu(struct mmu_gather *tlb)
222 {
223         struct mmu_gather_batch *batch;
224
225         if (!tlb->need_flush)
226                 return;
227         tlb->need_flush = 0;
228         tlb_flush(tlb);
229 #ifdef CONFIG_HAVE_RCU_TABLE_FREE
230         tlb_table_flush(tlb);
231 #endif
232
233         if (tlb_fast_mode(tlb))
234                 return;
235
236         for (batch = &tlb->local; batch; batch = batch->next) {
237                 free_pages_and_swap_cache(batch->pages, batch->nr);
238                 batch->nr = 0;
239         }
240         tlb->active = &tlb->local;
241 }
242
243 /* tlb_finish_mmu
244  *      Called at the end of the shootdown operation to free up any resources
245  *      that were required.
246  */
247 void tlb_finish_mmu(struct mmu_gather *tlb, unsigned long start, unsigned long end)
248 {
249         struct mmu_gather_batch *batch, *next;
250
251         tlb_flush_mmu(tlb);
252
253         /* keep the page table cache within bounds */
254         check_pgt_cache();
255
256         for (batch = tlb->local.next; batch; batch = next) {
257                 next = batch->next;
258                 free_pages((unsigned long)batch, 0);
259         }
260         tlb->local.next = NULL;
261 }
262
263 /* __tlb_remove_page
264  *      Must perform the equivalent to __free_pte(pte_get_and_clear(ptep)), while
265  *      handling the additional races in SMP caused by other CPUs caching valid
266  *      mappings in their TLBs. Returns the number of free page slots left.
267  *      When out of page slots we must call tlb_flush_mmu().
268  */
269 int __tlb_remove_page(struct mmu_gather *tlb, struct page *page)
270 {
271         struct mmu_gather_batch *batch;
272
273         VM_BUG_ON(!tlb->need_flush);
274
275         if (tlb_fast_mode(tlb)) {
276                 free_page_and_swap_cache(page);
277                 return 1; /* avoid calling tlb_flush_mmu() */
278         }
279
280         batch = tlb->active;
281         batch->pages[batch->nr++] = page;
282         if (batch->nr == batch->max) {
283                 if (!tlb_next_batch(tlb))
284                         return 0;
285                 batch = tlb->active;
286         }
287         VM_BUG_ON(batch->nr > batch->max);
288
289         return batch->max - batch->nr;
290 }
291
292 #endif /* HAVE_GENERIC_MMU_GATHER */
293
294 #ifdef CONFIG_HAVE_RCU_TABLE_FREE
295
296 /*
297  * See the comment near struct mmu_table_batch.
298  */
299
300 static void tlb_remove_table_smp_sync(void *arg)
301 {
302         /* Simply deliver the interrupt */
303 }
304
305 static void tlb_remove_table_one(void *table)
306 {
307         /*
308          * This isn't an RCU grace period and hence the page-tables cannot be
309          * assumed to be actually RCU-freed.
310          *
311          * It is however sufficient for software page-table walkers that rely on
312          * IRQ disabling. See the comment near struct mmu_table_batch.
313          */
314         smp_call_function(tlb_remove_table_smp_sync, NULL, 1);
315         __tlb_remove_table(table);
316 }
317
318 static void tlb_remove_table_rcu(struct rcu_head *head)
319 {
320         struct mmu_table_batch *batch;
321         int i;
322
323         batch = container_of(head, struct mmu_table_batch, rcu);
324
325         for (i = 0; i < batch->nr; i++)
326                 __tlb_remove_table(batch->tables[i]);
327
328         free_page((unsigned long)batch);
329 }
330
331 void tlb_table_flush(struct mmu_gather *tlb)
332 {
333         struct mmu_table_batch **batch = &tlb->batch;
334
335         if (*batch) {
336                 call_rcu_sched(&(*batch)->rcu, tlb_remove_table_rcu);
337                 *batch = NULL;
338         }
339 }
340
341 void tlb_remove_table(struct mmu_gather *tlb, void *table)
342 {
343         struct mmu_table_batch **batch = &tlb->batch;
344
345         tlb->need_flush = 1;
346
347         /*
348          * When there's less then two users of this mm there cannot be a
349          * concurrent page-table walk.
350          */
351         if (atomic_read(&tlb->mm->mm_users) < 2) {
352                 __tlb_remove_table(table);
353                 return;
354         }
355
356         if (*batch == NULL) {
357                 *batch = (struct mmu_table_batch *)__get_free_page(GFP_NOWAIT | __GFP_NOWARN);
358                 if (*batch == NULL) {
359                         tlb_remove_table_one(table);
360                         return;
361                 }
362                 (*batch)->nr = 0;
363         }
364         (*batch)->tables[(*batch)->nr++] = table;
365         if ((*batch)->nr == MAX_TABLE_BATCH)
366                 tlb_table_flush(tlb);
367 }
368
369 #endif /* CONFIG_HAVE_RCU_TABLE_FREE */
370
371 /*
372  * If a p?d_bad entry is found while walking page tables, report
373  * the error, before resetting entry to p?d_none.  Usually (but
374  * very seldom) called out from the p?d_none_or_clear_bad macros.
375  */
376
377 void pgd_clear_bad(pgd_t *pgd)
378 {
379         pgd_ERROR(*pgd);
380         pgd_clear(pgd);
381 }
382
383 void pud_clear_bad(pud_t *pud)
384 {
385         pud_ERROR(*pud);
386         pud_clear(pud);
387 }
388
389 void pmd_clear_bad(pmd_t *pmd)
390 {
391         pmd_ERROR(*pmd);
392         pmd_clear(pmd);
393 }
394
395 /*
396  * Note: this doesn't free the actual pages themselves. That
397  * has been handled earlier when unmapping all the memory regions.
398  */
399 static void free_pte_range(struct mmu_gather *tlb, pmd_t *pmd,
400                            unsigned long addr)
401 {
402         pgtable_t token = pmd_pgtable(*pmd);
403         pmd_clear(pmd);
404         pte_free_tlb(tlb, token, addr);
405         tlb->mm->nr_ptes--;
406 }
407
408 static inline void free_pmd_range(struct mmu_gather *tlb, pud_t *pud,
409                                 unsigned long addr, unsigned long end,
410                                 unsigned long floor, unsigned long ceiling)
411 {
412         pmd_t *pmd;
413         unsigned long next;
414         unsigned long start;
415
416         start = addr;
417         pmd = pmd_offset(pud, addr);
418         do {
419                 next = pmd_addr_end(addr, end);
420                 if (pmd_none_or_clear_bad(pmd))
421                         continue;
422                 free_pte_range(tlb, pmd, addr);
423         } while (pmd++, addr = next, addr != end);
424
425         start &= PUD_MASK;
426         if (start < floor)
427                 return;
428         if (ceiling) {
429                 ceiling &= PUD_MASK;
430                 if (!ceiling)
431                         return;
432         }
433         if (end - 1 > ceiling - 1)
434                 return;
435
436         pmd = pmd_offset(pud, start);
437         pud_clear(pud);
438         pmd_free_tlb(tlb, pmd, start);
439 }
440
441 static inline void free_pud_range(struct mmu_gather *tlb, pgd_t *pgd,
442                                 unsigned long addr, unsigned long end,
443                                 unsigned long floor, unsigned long ceiling)
444 {
445         pud_t *pud;
446         unsigned long next;
447         unsigned long start;
448
449         start = addr;
450         pud = pud_offset(pgd, addr);
451         do {
452                 next = pud_addr_end(addr, end);
453                 if (pud_none_or_clear_bad(pud))
454                         continue;
455                 free_pmd_range(tlb, pud, addr, next, floor, ceiling);
456         } while (pud++, addr = next, addr != end);
457
458         start &= PGDIR_MASK;
459         if (start < floor)
460                 return;
461         if (ceiling) {
462                 ceiling &= PGDIR_MASK;
463                 if (!ceiling)
464                         return;
465         }
466         if (end - 1 > ceiling - 1)
467                 return;
468
469         pud = pud_offset(pgd, start);
470         pgd_clear(pgd);
471         pud_free_tlb(tlb, pud, start);
472 }
473
474 /*
475  * This function frees user-level page tables of a process.
476  *
477  * Must be called with pagetable lock held.
478  */
479 void free_pgd_range(struct mmu_gather *tlb,
480                         unsigned long addr, unsigned long end,
481                         unsigned long floor, unsigned long ceiling)
482 {
483         pgd_t *pgd;
484         unsigned long next;
485
486         /*
487          * The next few lines have given us lots of grief...
488          *
489          * Why are we testing PMD* at this top level?  Because often
490          * there will be no work to do at all, and we'd prefer not to
491          * go all the way down to the bottom just to discover that.
492          *
493          * Why all these "- 1"s?  Because 0 represents both the bottom
494          * of the address space and the top of it (using -1 for the
495          * top wouldn't help much: the masks would do the wrong thing).
496          * The rule is that addr 0 and floor 0 refer to the bottom of
497          * the address space, but end 0 and ceiling 0 refer to the top
498          * Comparisons need to use "end - 1" and "ceiling - 1" (though
499          * that end 0 case should be mythical).
500          *
501          * Wherever addr is brought up or ceiling brought down, we must
502          * be careful to reject "the opposite 0" before it confuses the
503          * subsequent tests.  But what about where end is brought down
504          * by PMD_SIZE below? no, end can't go down to 0 there.
505          *
506          * Whereas we round start (addr) and ceiling down, by different
507          * masks at different levels, in order to test whether a table
508          * now has no other vmas using it, so can be freed, we don't
509          * bother to round floor or end up - the tests don't need that.
510          */
511
512         addr &= PMD_MASK;
513         if (addr < floor) {
514                 addr += PMD_SIZE;
515                 if (!addr)
516                         return;
517         }
518         if (ceiling) {
519                 ceiling &= PMD_MASK;
520                 if (!ceiling)
521                         return;
522         }
523         if (end - 1 > ceiling - 1)
524                 end -= PMD_SIZE;
525         if (addr > end - 1)
526                 return;
527
528         pgd = pgd_offset(tlb->mm, addr);
529         do {
530                 next = pgd_addr_end(addr, end);
531                 if (pgd_none_or_clear_bad(pgd))
532                         continue;
533                 free_pud_range(tlb, pgd, addr, next, floor, ceiling);
534         } while (pgd++, addr = next, addr != end);
535 }
536
537 void free_pgtables(struct mmu_gather *tlb, struct vm_area_struct *vma,
538                 unsigned long floor, unsigned long ceiling)
539 {
540         while (vma) {
541                 struct vm_area_struct *next = vma->vm_next;
542                 unsigned long addr = vma->vm_start;
543
544                 /*
545                  * Hide vma from rmap and truncate_pagecache before freeing
546                  * pgtables
547                  */
548                 unlink_anon_vmas(vma);
549                 unlink_file_vma(vma);
550
551                 if (is_vm_hugetlb_page(vma)) {
552                         hugetlb_free_pgd_range(tlb, addr, vma->vm_end,
553                                 floor, next? next->vm_start: ceiling);
554                 } else {
555                         /*
556                          * Optimization: gather nearby vmas into one call down
557                          */
558                         while (next && next->vm_start <= vma->vm_end + PMD_SIZE
559                                && !is_vm_hugetlb_page(next)) {
560                                 vma = next;
561                                 next = vma->vm_next;
562                                 unlink_anon_vmas(vma);
563                                 unlink_file_vma(vma);
564                         }
565                         free_pgd_range(tlb, addr, vma->vm_end,
566                                 floor, next? next->vm_start: ceiling);
567                 }
568                 vma = next;
569         }
570 }
571
572 int __pte_alloc(struct mm_struct *mm, struct vm_area_struct *vma,
573                 pmd_t *pmd, unsigned long address)
574 {
575         pgtable_t new = pte_alloc_one(mm, address);
576         int wait_split_huge_page;
577         if (!new)
578                 return -ENOMEM;
579
580         /*
581          * Ensure all pte setup (eg. pte page lock and page clearing) are
582          * visible before the pte is made visible to other CPUs by being
583          * put into page tables.
584          *
585          * The other side of the story is the pointer chasing in the page
586          * table walking code (when walking the page table without locking;
587          * ie. most of the time). Fortunately, these data accesses consist
588          * of a chain of data-dependent loads, meaning most CPUs (alpha
589          * being the notable exception) will already guarantee loads are
590          * seen in-order. See the alpha page table accessors for the
591          * smp_read_barrier_depends() barriers in page table walking code.
592          */
593         smp_wmb(); /* Could be smp_wmb__xxx(before|after)_spin_lock */
594
595         spin_lock(&mm->page_table_lock);
596         wait_split_huge_page = 0;
597         if (likely(pmd_none(*pmd))) {   /* Has another populated it ? */
598                 mm->nr_ptes++;
599                 pmd_populate(mm, pmd, new);
600                 new = NULL;
601         } else if (unlikely(pmd_trans_splitting(*pmd)))
602                 wait_split_huge_page = 1;
603         spin_unlock(&mm->page_table_lock);
604         if (new)
605                 pte_free(mm, new);
606         if (wait_split_huge_page)
607                 wait_split_huge_page(vma->anon_vma, pmd);
608         return 0;
609 }
610
611 int __pte_alloc_kernel(pmd_t *pmd, unsigned long address)
612 {
613         pte_t *new = pte_alloc_one_kernel(&init_mm, address);
614         if (!new)
615                 return -ENOMEM;
616
617         smp_wmb(); /* See comment in __pte_alloc */
618
619         spin_lock(&init_mm.page_table_lock);
620         if (likely(pmd_none(*pmd))) {   /* Has another populated it ? */
621                 pmd_populate_kernel(&init_mm, pmd, new);
622                 new = NULL;
623         } else
624                 VM_BUG_ON(pmd_trans_splitting(*pmd));
625         spin_unlock(&init_mm.page_table_lock);
626         if (new)
627                 pte_free_kernel(&init_mm, new);
628         return 0;
629 }
630
631 static inline void init_rss_vec(int *rss)
632 {
633         memset(rss, 0, sizeof(int) * NR_MM_COUNTERS);
634 }
635
636 static inline void add_mm_rss_vec(struct mm_struct *mm, int *rss)
637 {
638         int i;
639
640         if (current->mm == mm)
641                 sync_mm_rss(mm);
642         for (i = 0; i < NR_MM_COUNTERS; i++)
643                 if (rss[i])
644                         add_mm_counter(mm, i, rss[i]);
645 }
646
647 /*
648  * This function is called to print an error when a bad pte
649  * is found. For example, we might have a PFN-mapped pte in
650  * a region that doesn't allow it.
651  *
652  * The calling function must still handle the error.
653  */
654 static void print_bad_pte(struct vm_area_struct *vma, unsigned long addr,
655                           pte_t pte, struct page *page)
656 {
657         pgd_t *pgd = pgd_offset(vma->vm_mm, addr);
658         pud_t *pud = pud_offset(pgd, addr);
659         pmd_t *pmd = pmd_offset(pud, addr);
660         struct address_space *mapping;
661         pgoff_t index;
662         static unsigned long resume;
663         static unsigned long nr_shown;
664         static unsigned long nr_unshown;
665
666         /*
667          * Allow a burst of 60 reports, then keep quiet for that minute;
668          * or allow a steady drip of one report per second.
669          */
670         if (nr_shown == 60) {
671                 if (time_before(jiffies, resume)) {
672                         nr_unshown++;
673                         return;
674                 }
675                 if (nr_unshown) {
676                         printk(KERN_ALERT
677                                 "BUG: Bad page map: %lu messages suppressed\n",
678                                 nr_unshown);
679                         nr_unshown = 0;
680                 }
681                 nr_shown = 0;
682         }
683         if (nr_shown++ == 0)
684                 resume = jiffies + 60 * HZ;
685
686         mapping = vma->vm_file ? vma->vm_file->f_mapping : NULL;
687         index = linear_page_index(vma, addr);
688
689         printk(KERN_ALERT
690                 "BUG: Bad page map in process %s  pte:%08llx pmd:%08llx\n",
691                 current->comm,
692                 (long long)pte_val(pte), (long long)pmd_val(*pmd));
693         if (page)
694                 dump_page(page);
695         printk(KERN_ALERT
696                 "addr:%p vm_flags:%08lx anon_vma:%p mapping:%p index:%lx\n",
697                 (void *)addr, vma->vm_flags, vma->anon_vma, mapping, index);
698         /*
699          * Choose text because data symbols depend on CONFIG_KALLSYMS_ALL=y
700          */
701         if (vma->vm_ops)
702                 print_symbol(KERN_ALERT "vma->vm_ops->fault: %s\n",
703                                 (unsigned long)vma->vm_ops->fault);
704         if (vma->vm_file && vma->vm_file->f_op)
705                 print_symbol(KERN_ALERT "vma->vm_file->f_op->mmap: %s\n",
706                                 (unsigned long)vma->vm_file->f_op->mmap);
707         dump_stack();
708         add_taint(TAINT_BAD_PAGE);
709 }
710
711 static inline int is_cow_mapping(vm_flags_t flags)
712 {
713         return (flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE;
714 }
715
716 #ifndef is_zero_pfn
717 static inline int is_zero_pfn(unsigned long pfn)
718 {
719         return pfn == zero_pfn;
720 }
721 #endif
722
723 #ifndef my_zero_pfn
724 static inline unsigned long my_zero_pfn(unsigned long addr)
725 {
726         return zero_pfn;
727 }
728 #endif
729
730 /*
731  * vm_normal_page -- This function gets the "struct page" associated with a pte.
732  *
733  * "Special" mappings do not wish to be associated with a "struct page" (either
734  * it doesn't exist, or it exists but they don't want to touch it). In this
735  * case, NULL is returned here. "Normal" mappings do have a struct page.
736  *
737  * There are 2 broad cases. Firstly, an architecture may define a pte_special()
738  * pte bit, in which case this function is trivial. Secondly, an architecture
739  * may not have a spare pte bit, which requires a more complicated scheme,
740  * described below.
741  *
742  * A raw VM_PFNMAP mapping (ie. one that is not COWed) is always considered a
743  * special mapping (even if there are underlying and valid "struct pages").
744  * COWed pages of a VM_PFNMAP are always normal.
745  *
746  * The way we recognize COWed pages within VM_PFNMAP mappings is through the
747  * rules set up by "remap_pfn_range()": the vma will have the VM_PFNMAP bit
748  * set, and the vm_pgoff will point to the first PFN mapped: thus every special
749  * mapping will always honor the rule
750  *
751  *      pfn_of_page == vma->vm_pgoff + ((addr - vma->vm_start) >> PAGE_SHIFT)
752  *
753  * And for normal mappings this is false.
754  *
755  * This restricts such mappings to be a linear translation from virtual address
756  * to pfn. To get around this restriction, we allow arbitrary mappings so long
757  * as the vma is not a COW mapping; in that case, we know that all ptes are
758  * special (because none can have been COWed).
759  *
760  *
761  * In order to support COW of arbitrary special mappings, we have VM_MIXEDMAP.
762  *
763  * VM_MIXEDMAP mappings can likewise contain memory with or without "struct
764  * page" backing, however the difference is that _all_ pages with a struct
765  * page (that is, those where pfn_valid is true) are refcounted and considered
766  * normal pages by the VM. The disadvantage is that pages are refcounted
767  * (which can be slower and simply not an option for some PFNMAP users). The
768  * advantage is that we don't have to follow the strict linearity rule of
769  * PFNMAP mappings in order to support COWable mappings.
770  *
771  */
772 #ifdef __HAVE_ARCH_PTE_SPECIAL
773 # define HAVE_PTE_SPECIAL 1
774 #else
775 # define HAVE_PTE_SPECIAL 0
776 #endif
777 struct page *vm_normal_page(struct vm_area_struct *vma, unsigned long addr,
778                                 pte_t pte)
779 {
780         unsigned long pfn = pte_pfn(pte);
781
782         if (HAVE_PTE_SPECIAL) {
783                 if (likely(!pte_special(pte)))
784                         goto check_pfn;
785                 if (vma->vm_flags & (VM_PFNMAP | VM_MIXEDMAP))
786                         return NULL;
787                 if (!is_zero_pfn(pfn))
788                         print_bad_pte(vma, addr, pte, NULL);
789                 return NULL;
790         }
791
792         /* !HAVE_PTE_SPECIAL case follows: */
793
794         if (unlikely(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP))) {
795                 if (vma->vm_flags & VM_MIXEDMAP) {
796                         if (!pfn_valid(pfn))
797                                 return NULL;
798                         goto out;
799                 } else {
800                         unsigned long off;
801                         off = (addr - vma->vm_start) >> PAGE_SHIFT;
802                         if (pfn == vma->vm_pgoff + off)
803                                 return NULL;
804                         if (!is_cow_mapping(vma->vm_flags))
805                                 return NULL;
806                 }
807         }
808
809         if (is_zero_pfn(pfn))
810                 return NULL;
811 check_pfn:
812         if (unlikely(pfn > highest_memmap_pfn)) {
813                 print_bad_pte(vma, addr, pte, NULL);
814                 return NULL;
815         }
816
817         /*
818          * NOTE! We still have PageReserved() pages in the page tables.
819          * eg. VDSO mappings can cause them to exist.
820          */
821 out:
822         return pfn_to_page(pfn);
823 }
824
825 /*
826  * copy one vm_area from one task to the other. Assumes the page tables
827  * already present in the new task to be cleared in the whole range
828  * covered by this vma.
829  */
830
831 static inline unsigned long
832 copy_one_pte(struct mm_struct *dst_mm, struct mm_struct *src_mm,
833                 pte_t *dst_pte, pte_t *src_pte, struct vm_area_struct *vma,
834                 unsigned long addr, int *rss)
835 {
836         unsigned long vm_flags = vma->vm_flags;
837         pte_t pte = *src_pte;
838         struct page *page;
839
840         /* pte contains position in swap or file, so copy. */
841         if (unlikely(!pte_present(pte))) {
842                 if (!pte_file(pte)) {
843                         swp_entry_t entry = pte_to_swp_entry(pte);
844
845                         if (swap_duplicate(entry) < 0)
846                                 return entry.val;
847
848                         /* make sure dst_mm is on swapoff's mmlist. */
849                         if (unlikely(list_empty(&dst_mm->mmlist))) {
850                                 spin_lock(&mmlist_lock);
851                                 if (list_empty(&dst_mm->mmlist))
852                                         list_add(&dst_mm->mmlist,
853                                                  &src_mm->mmlist);
854                                 spin_unlock(&mmlist_lock);
855                         }
856                         if (likely(!non_swap_entry(entry)))
857                                 rss[MM_SWAPENTS]++;
858                         else if (is_migration_entry(entry)) {
859                                 page = migration_entry_to_page(entry);
860
861                                 if (PageAnon(page))
862                                         rss[MM_ANONPAGES]++;
863                                 else
864                                         rss[MM_FILEPAGES]++;
865
866                                 if (is_write_migration_entry(entry) &&
867                                     is_cow_mapping(vm_flags)) {
868                                         /*
869                                          * COW mappings require pages in both
870                                          * parent and child to be set to read.
871                                          */
872                                         make_migration_entry_read(&entry);
873                                         pte = swp_entry_to_pte(entry);
874                                         set_pte_at(src_mm, addr, src_pte, pte);
875                                 }
876                         }
877                 }
878                 goto out_set_pte;
879         }
880
881         /*
882          * If it's a COW mapping, write protect it both
883          * in the parent and the child
884          */
885         if (is_cow_mapping(vm_flags)) {
886                 ptep_set_wrprotect(src_mm, addr, src_pte);
887                 pte = pte_wrprotect(pte);
888         }
889
890         /*
891          * If it's a shared mapping, mark it clean in
892          * the child
893          */
894         if (vm_flags & VM_SHARED)
895                 pte = pte_mkclean(pte);
896         pte = pte_mkold(pte);
897
898         page = vm_normal_page(vma, addr, pte);
899         if (page) {
900                 get_page(page);
901                 page_dup_rmap(page);
902                 if (PageAnon(page))
903                         rss[MM_ANONPAGES]++;
904                 else
905                         rss[MM_FILEPAGES]++;
906         }
907
908 out_set_pte:
909         set_pte_at(dst_mm, addr, dst_pte, pte);
910         return 0;
911 }
912
913 int copy_pte_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
914                    pmd_t *dst_pmd, pmd_t *src_pmd, struct vm_area_struct *vma,
915                    unsigned long addr, unsigned long end)
916 {
917         pte_t *orig_src_pte, *orig_dst_pte;
918         pte_t *src_pte, *dst_pte;
919         spinlock_t *src_ptl, *dst_ptl;
920         int progress = 0;
921         int rss[NR_MM_COUNTERS];
922         swp_entry_t entry = (swp_entry_t){0};
923
924 again:
925         init_rss_vec(rss);
926
927         dst_pte = pte_alloc_map_lock(dst_mm, dst_pmd, addr, &dst_ptl);
928         if (!dst_pte)
929                 return -ENOMEM;
930         src_pte = pte_offset_map(src_pmd, addr);
931         src_ptl = pte_lockptr(src_mm, src_pmd);
932         spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
933         orig_src_pte = src_pte;
934         orig_dst_pte = dst_pte;
935         arch_enter_lazy_mmu_mode();
936
937         do {
938                 /*
939                  * We are holding two locks at this point - either of them
940                  * could generate latencies in another task on another CPU.
941                  */
942                 if (progress >= 32) {
943                         progress = 0;
944                         if (need_resched() ||
945                             spin_needbreak(src_ptl) || spin_needbreak(dst_ptl))
946                                 break;
947                 }
948                 if (pte_none(*src_pte)) {
949                         progress++;
950                         continue;
951                 }
952                 entry.val = copy_one_pte(dst_mm, src_mm, dst_pte, src_pte,
953                                                         vma, addr, rss);
954                 if (entry.val)
955                         break;
956                 progress += 8;
957         } while (dst_pte++, src_pte++, addr += PAGE_SIZE, addr != end);
958
959         arch_leave_lazy_mmu_mode();
960         spin_unlock(src_ptl);
961         pte_unmap(orig_src_pte);
962         add_mm_rss_vec(dst_mm, rss);
963         pte_unmap_unlock(orig_dst_pte, dst_ptl);
964         cond_resched();
965
966         if (entry.val) {
967                 if (add_swap_count_continuation(entry, GFP_KERNEL) < 0)
968                         return -ENOMEM;
969                 progress = 0;
970         }
971         if (addr != end)
972                 goto again;
973         return 0;
974 }
975
976 static inline int copy_pmd_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
977                 pud_t *dst_pud, pud_t *src_pud, struct vm_area_struct *vma,
978                 unsigned long addr, unsigned long end)
979 {
980         pmd_t *src_pmd, *dst_pmd;
981         unsigned long next;
982
983         dst_pmd = pmd_alloc(dst_mm, dst_pud, addr);
984         if (!dst_pmd)
985                 return -ENOMEM;
986         src_pmd = pmd_offset(src_pud, addr);
987         do {
988                 next = pmd_addr_end(addr, end);
989                 if (pmd_trans_huge(*src_pmd)) {
990                         int err;
991                         VM_BUG_ON(next-addr != HPAGE_PMD_SIZE);
992                         err = copy_huge_pmd(dst_mm, src_mm,
993                                             dst_pmd, src_pmd, addr, vma);
994                         if (err == -ENOMEM)
995                                 return -ENOMEM;
996                         if (!err)
997                                 continue;
998                         /* fall through */
999                 }
1000                 if (pmd_none_or_clear_bad(src_pmd))
1001                         continue;
1002                 if (copy_pte_range(dst_mm, src_mm, dst_pmd, src_pmd,
1003                                                 vma, addr, next))
1004                         return -ENOMEM;
1005         } while (dst_pmd++, src_pmd++, addr = next, addr != end);
1006         return 0;
1007 }
1008
1009 static inline int copy_pud_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
1010                 pgd_t *dst_pgd, pgd_t *src_pgd, struct vm_area_struct *vma,
1011                 unsigned long addr, unsigned long end)
1012 {
1013         pud_t *src_pud, *dst_pud;
1014         unsigned long next;
1015
1016         dst_pud = pud_alloc(dst_mm, dst_pgd, addr);
1017         if (!dst_pud)
1018                 return -ENOMEM;
1019         src_pud = pud_offset(src_pgd, addr);
1020         do {
1021                 next = pud_addr_end(addr, end);
1022                 if (pud_none_or_clear_bad(src_pud))
1023                         continue;
1024                 if (copy_pmd_range(dst_mm, src_mm, dst_pud, src_pud,
1025                                                 vma, addr, next))
1026                         return -ENOMEM;
1027         } while (dst_pud++, src_pud++, addr = next, addr != end);
1028         return 0;
1029 }
1030
1031 int copy_page_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
1032                 struct vm_area_struct *vma)
1033 {
1034         pgd_t *src_pgd, *dst_pgd;
1035         unsigned long next;
1036         unsigned long addr = vma->vm_start;
1037         unsigned long end = vma->vm_end;
1038         int ret;
1039
1040         /*
1041          * Don't copy ptes where a page fault will fill them correctly.
1042          * Fork becomes much lighter when there are big shared or private
1043          * readonly mappings. The tradeoff is that copy_page_range is more
1044          * efficient than faulting.
1045          */
1046         if (!(vma->vm_flags & (VM_HUGETLB|VM_NONLINEAR|VM_PFNMAP|VM_INSERTPAGE))) {
1047                 if (!vma->anon_vma)
1048                         return 0;
1049         }
1050
1051         if (is_vm_hugetlb_page(vma))
1052                 return copy_hugetlb_page_range(dst_mm, src_mm, vma);
1053
1054         if (unlikely(is_pfn_mapping(vma))) {
1055                 /*
1056                  * We do not free on error cases below as remove_vma
1057                  * gets called on error from higher level routine
1058                  */
1059                 ret = track_pfn_vma_copy(vma);
1060                 if (ret)
1061                         return ret;
1062         }
1063
1064         /*
1065          * We need to invalidate the secondary MMU mappings only when
1066          * there could be a permission downgrade on the ptes of the
1067          * parent mm. And a permission downgrade will only happen if
1068          * is_cow_mapping() returns true.
1069          */
1070         if (is_cow_mapping(vma->vm_flags))
1071                 mmu_notifier_invalidate_range_start(src_mm, addr, end);
1072
1073         ret = 0;
1074         dst_pgd = pgd_offset(dst_mm, addr);
1075         src_pgd = pgd_offset(src_mm, addr);
1076         do {
1077                 next = pgd_addr_end(addr, end);
1078                 if (pgd_none_or_clear_bad(src_pgd))
1079                         continue;
1080                 if (unlikely(copy_pud_range(dst_mm, src_mm, dst_pgd, src_pgd,
1081                                             vma, addr, next))) {
1082                         ret = -ENOMEM;
1083                         break;
1084                 }
1085         } while (dst_pgd++, src_pgd++, addr = next, addr != end);
1086
1087         if (is_cow_mapping(vma->vm_flags))
1088                 mmu_notifier_invalidate_range_end(src_mm,
1089                                                   vma->vm_start, end);
1090         return ret;
1091 }
1092
1093 static unsigned long zap_pte_range(struct mmu_gather *tlb,
1094                                 struct vm_area_struct *vma, pmd_t *pmd,
1095                                 unsigned long addr, unsigned long end,
1096                                 struct zap_details *details)
1097 {
1098         struct mm_struct *mm = tlb->mm;
1099         int force_flush = 0;
1100         int rss[NR_MM_COUNTERS];
1101         spinlock_t *ptl;
1102         pte_t *start_pte;
1103         pte_t *pte;
1104
1105 again:
1106         init_rss_vec(rss);
1107         start_pte = pte_offset_map_lock(mm, pmd, addr, &ptl);
1108         pte = start_pte;
1109         arch_enter_lazy_mmu_mode();
1110         do {
1111                 pte_t ptent = *pte;
1112                 if (pte_none(ptent)) {
1113                         continue;
1114                 }
1115
1116                 if (pte_present(ptent)) {
1117                         struct page *page;
1118
1119                         page = vm_normal_page(vma, addr, ptent);
1120                         if (unlikely(details) && page) {
1121                                 /*
1122                                  * unmap_shared_mapping_pages() wants to
1123                                  * invalidate cache without truncating:
1124                                  * unmap shared but keep private pages.
1125                                  */
1126                                 if (details->check_mapping &&
1127                                     details->check_mapping != page->mapping)
1128                                         continue;
1129                                 /*
1130                                  * Each page->index must be checked when
1131                                  * invalidating or truncating nonlinear.
1132                                  */
1133                                 if (details->nonlinear_vma &&
1134                                     (page->index < details->first_index ||
1135                                      page->index > details->last_index))
1136                                         continue;
1137                         }
1138                         ptent = ptep_get_and_clear_full(mm, addr, pte,
1139                                                         tlb->fullmm);
1140                         tlb_remove_tlb_entry(tlb, pte, addr);
1141                         if (unlikely(!page))
1142                                 continue;
1143                         if (unlikely(details) && details->nonlinear_vma
1144                             && linear_page_index(details->nonlinear_vma,
1145                                                 addr) != page->index)
1146                                 set_pte_at(mm, addr, pte,
1147                                            pgoff_to_pte(page->index));
1148                         if (PageAnon(page))
1149                                 rss[MM_ANONPAGES]--;
1150                         else {
1151                                 if (pte_dirty(ptent))
1152                                         set_page_dirty(page);
1153                                 if (pte_young(ptent) &&
1154                                     likely(!VM_SequentialReadHint(vma)))
1155                                         mark_page_accessed(page);
1156                                 rss[MM_FILEPAGES]--;
1157                         }
1158                         page_remove_rmap(page);
1159                         if (unlikely(page_mapcount(page) < 0))
1160                                 print_bad_pte(vma, addr, ptent, page);
1161                         force_flush = !__tlb_remove_page(tlb, page);
1162                         if (force_flush)
1163                                 break;
1164                         continue;
1165                 }
1166                 /*
1167                  * If details->check_mapping, we leave swap entries;
1168                  * if details->nonlinear_vma, we leave file entries.
1169                  */
1170                 if (unlikely(details))
1171                         continue;
1172                 if (pte_file(ptent)) {
1173                         if (unlikely(!(vma->vm_flags & VM_NONLINEAR)))
1174                                 print_bad_pte(vma, addr, ptent, NULL);
1175                 } else {
1176                         swp_entry_t entry = pte_to_swp_entry(ptent);
1177
1178                         if (!non_swap_entry(entry))
1179                                 rss[MM_SWAPENTS]--;
1180                         else if (is_migration_entry(entry)) {
1181                                 struct page *page;
1182
1183                                 page = migration_entry_to_page(entry);
1184
1185                                 if (PageAnon(page))
1186                                         rss[MM_ANONPAGES]--;
1187                                 else
1188                                         rss[MM_FILEPAGES]--;
1189                         }
1190                         if (unlikely(!free_swap_and_cache(entry)))
1191                                 print_bad_pte(vma, addr, ptent, NULL);
1192                 }
1193                 pte_clear_not_present_full(mm, addr, pte, tlb->fullmm);
1194         } while (pte++, addr += PAGE_SIZE, addr != end);
1195
1196         add_mm_rss_vec(mm, rss);
1197         arch_leave_lazy_mmu_mode();
1198         pte_unmap_unlock(start_pte, ptl);
1199
1200         /*
1201          * mmu_gather ran out of room to batch pages, we break out of
1202          * the PTE lock to avoid doing the potential expensive TLB invalidate
1203          * and page-free while holding it.
1204          */
1205         if (force_flush) {
1206                 force_flush = 0;
1207                 tlb_flush_mmu(tlb);
1208                 if (addr != end)
1209                         goto again;
1210         }
1211
1212         return addr;
1213 }
1214
1215 static inline unsigned long zap_pmd_range(struct mmu_gather *tlb,
1216                                 struct vm_area_struct *vma, pud_t *pud,
1217                                 unsigned long addr, unsigned long end,
1218                                 struct zap_details *details)
1219 {
1220         pmd_t *pmd;
1221         unsigned long next;
1222
1223         pmd = pmd_offset(pud, addr);
1224         do {
1225                 next = pmd_addr_end(addr, end);
1226                 if (pmd_trans_huge(*pmd)) {
1227                         if (next - addr != HPAGE_PMD_SIZE) {
1228                                 VM_BUG_ON(!rwsem_is_locked(&tlb->mm->mmap_sem));
1229                                 split_huge_page_pmd(vma->vm_mm, pmd);
1230                         } else if (zap_huge_pmd(tlb, vma, pmd, addr))
1231                                 goto next;
1232                         /* fall through */
1233                 }
1234                 /*
1235                  * Here there can be other concurrent MADV_DONTNEED or
1236                  * trans huge page faults running, and if the pmd is
1237                  * none or trans huge it can change under us. This is
1238                  * because MADV_DONTNEED holds the mmap_sem in read
1239                  * mode.
1240                  */
1241                 if (pmd_none_or_trans_huge_or_clear_bad(pmd))
1242                         goto next;
1243                 next = zap_pte_range(tlb, vma, pmd, addr, next, details);
1244 next:
1245                 cond_resched();
1246         } while (pmd++, addr = next, addr != end);
1247
1248         return addr;
1249 }
1250
1251 static inline unsigned long zap_pud_range(struct mmu_gather *tlb,
1252                                 struct vm_area_struct *vma, pgd_t *pgd,
1253                                 unsigned long addr, unsigned long end,
1254                                 struct zap_details *details)
1255 {
1256         pud_t *pud;
1257         unsigned long next;
1258
1259         pud = pud_offset(pgd, addr);
1260         do {
1261                 next = pud_addr_end(addr, end);
1262                 if (pud_none_or_clear_bad(pud))
1263                         continue;
1264                 next = zap_pmd_range(tlb, vma, pud, addr, next, details);
1265         } while (pud++, addr = next, addr != end);
1266
1267         return addr;
1268 }
1269
1270 static unsigned long unmap_page_range(struct mmu_gather *tlb,
1271                                 struct vm_area_struct *vma,
1272                                 unsigned long addr, unsigned long end,
1273                                 struct zap_details *details)
1274 {
1275         pgd_t *pgd;
1276         unsigned long next;
1277
1278         if (details && !details->check_mapping && !details->nonlinear_vma)
1279                 details = NULL;
1280
1281         BUG_ON(addr >= end);
1282         mem_cgroup_uncharge_start();
1283         tlb_start_vma(tlb, vma);
1284         pgd = pgd_offset(vma->vm_mm, addr);
1285         do {
1286                 next = pgd_addr_end(addr, end);
1287                 if (pgd_none_or_clear_bad(pgd))
1288                         continue;
1289                 next = zap_pud_range(tlb, vma, pgd, addr, next, details);
1290         } while (pgd++, addr = next, addr != end);
1291         tlb_end_vma(tlb, vma);
1292         mem_cgroup_uncharge_end();
1293
1294         return addr;
1295 }
1296
1297 /**
1298  * unmap_vmas - unmap a range of memory covered by a list of vma's
1299  * @tlb: address of the caller's struct mmu_gather
1300  * @vma: the starting vma
1301  * @start_addr: virtual address at which to start unmapping
1302  * @end_addr: virtual address at which to end unmapping
1303  * @nr_accounted: Place number of unmapped pages in vm-accountable vma's here
1304  * @details: details of nonlinear truncation or shared cache invalidation
1305  *
1306  * Returns the end address of the unmapping (restart addr if interrupted).
1307  *
1308  * Unmap all pages in the vma list.
1309  *
1310  * Only addresses between `start' and `end' will be unmapped.
1311  *
1312  * The VMA list must be sorted in ascending virtual address order.
1313  *
1314  * unmap_vmas() assumes that the caller will flush the whole unmapped address
1315  * range after unmap_vmas() returns.  So the only responsibility here is to
1316  * ensure that any thus-far unmapped pages are flushed before unmap_vmas()
1317  * drops the lock and schedules.
1318  */
1319 unsigned long unmap_vmas(struct mmu_gather *tlb,
1320                 struct vm_area_struct *vma, unsigned long start_addr,
1321                 unsigned long end_addr, unsigned long *nr_accounted,
1322                 struct zap_details *details)
1323 {
1324         unsigned long start = start_addr;
1325         struct mm_struct *mm = vma->vm_mm;
1326
1327         mmu_notifier_invalidate_range_start(mm, start_addr, end_addr);
1328         for ( ; vma && vma->vm_start < end_addr; vma = vma->vm_next) {
1329                 unsigned long end;
1330
1331                 start = max(vma->vm_start, start_addr);
1332                 if (start >= vma->vm_end)
1333                         continue;
1334                 end = min(vma->vm_end, end_addr);
1335                 if (end <= vma->vm_start)
1336                         continue;
1337
1338                 if (vma->vm_flags & VM_ACCOUNT)
1339                         *nr_accounted += (end - start) >> PAGE_SHIFT;
1340
1341                 if (unlikely(is_pfn_mapping(vma)))
1342                         untrack_pfn_vma(vma, 0, 0);
1343
1344                 while (start != end) {
1345                         if (unlikely(is_vm_hugetlb_page(vma))) {
1346                                 /*
1347                                  * It is undesirable to test vma->vm_file as it
1348                                  * should be non-null for valid hugetlb area.
1349                                  * However, vm_file will be NULL in the error
1350                                  * cleanup path of do_mmap_pgoff. When
1351                                  * hugetlbfs ->mmap method fails,
1352                                  * do_mmap_pgoff() nullifies vma->vm_file
1353                                  * before calling this function to clean up.
1354                                  * Since no pte has actually been setup, it is
1355                                  * safe to do nothing in this case.
1356                                  */
1357                                 if (vma->vm_file)
1358                                         unmap_hugepage_range(vma, start, end, NULL);
1359
1360                                 start = end;
1361                         } else
1362                                 start = unmap_page_range(tlb, vma, start, end, details);
1363                 }
1364         }
1365
1366         mmu_notifier_invalidate_range_end(mm, start_addr, end_addr);
1367         return start;   /* which is now the end (or restart) address */
1368 }
1369
1370 /**
1371  * zap_page_range - remove user pages in a given range
1372  * @vma: vm_area_struct holding the applicable pages
1373  * @address: starting address of pages to zap
1374  * @size: number of bytes to zap
1375  * @details: details of nonlinear truncation or shared cache invalidation
1376  */
1377 unsigned long zap_page_range(struct vm_area_struct *vma, unsigned long address,
1378                 unsigned long size, struct zap_details *details)
1379 {
1380         struct mm_struct *mm = vma->vm_mm;
1381         struct mmu_gather tlb;
1382         unsigned long end = address + size;
1383         unsigned long nr_accounted = 0;
1384
1385         lru_add_drain();
1386         tlb_gather_mmu(&tlb, mm, 0);
1387         update_hiwater_rss(mm);
1388         end = unmap_vmas(&tlb, vma, address, end, &nr_accounted, details);
1389         tlb_finish_mmu(&tlb, address, end);
1390         return end;
1391 }
1392
1393 /**
1394  * zap_vma_ptes - remove ptes mapping the vma
1395  * @vma: vm_area_struct holding ptes to be zapped
1396  * @address: starting address of pages to zap
1397  * @size: number of bytes to zap
1398  *
1399  * This function only unmaps ptes assigned to VM_PFNMAP vmas.
1400  *
1401  * The entire address range must be fully contained within the vma.
1402  *
1403  * Returns 0 if successful.
1404  */
1405 int zap_vma_ptes(struct vm_area_struct *vma, unsigned long address,
1406                 unsigned long size)
1407 {
1408         if (address < vma->vm_start || address + size > vma->vm_end ||
1409                         !(vma->vm_flags & VM_PFNMAP))
1410                 return -1;
1411         zap_page_range(vma, address, size, NULL);
1412         return 0;
1413 }
1414 EXPORT_SYMBOL_GPL(zap_vma_ptes);
1415
1416 /**
1417  * follow_page - look up a page descriptor from a user-virtual address
1418  * @vma: vm_area_struct mapping @address
1419  * @address: virtual address to look up
1420  * @flags: flags modifying lookup behaviour
1421  *
1422  * @flags can have FOLL_ flags set, defined in <linux/mm.h>
1423  *
1424  * Returns the mapped (struct page *), %NULL if no mapping exists, or
1425  * an error pointer if there is a mapping to something not represented
1426  * by a page descriptor (see also vm_normal_page()).
1427  */
1428 struct page *follow_page(struct vm_area_struct *vma, unsigned long address,
1429                         unsigned int flags)
1430 {
1431         pgd_t *pgd;
1432         pud_t *pud;
1433         pmd_t *pmd;
1434         pte_t *ptep, pte;
1435         spinlock_t *ptl;
1436         struct page *page;
1437         struct mm_struct *mm = vma->vm_mm;
1438
1439         page = follow_huge_addr(mm, address, flags & FOLL_WRITE);
1440         if (!IS_ERR(page)) {
1441                 BUG_ON(flags & FOLL_GET);
1442                 goto out;
1443         }
1444
1445         page = NULL;
1446         pgd = pgd_offset(mm, address);
1447         if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
1448                 goto no_page_table;
1449
1450         pud = pud_offset(pgd, address);
1451         if (pud_none(*pud))
1452                 goto no_page_table;
1453         if (pud_huge(*pud) && vma->vm_flags & VM_HUGETLB) {
1454                 BUG_ON(flags & FOLL_GET);
1455                 page = follow_huge_pud(mm, address, pud, flags & FOLL_WRITE);
1456                 goto out;
1457         }
1458         if (unlikely(pud_bad(*pud)))
1459                 goto no_page_table;
1460
1461         pmd = pmd_offset(pud, address);
1462         if (pmd_none(*pmd))
1463                 goto no_page_table;
1464         if (pmd_huge(*pmd) && vma->vm_flags & VM_HUGETLB) {
1465                 BUG_ON(flags & FOLL_GET);
1466                 page = follow_huge_pmd(mm, address, pmd, flags & FOLL_WRITE);
1467                 goto out;
1468         }
1469         if (pmd_trans_huge(*pmd)) {
1470                 if (flags & FOLL_SPLIT) {
1471                         split_huge_page_pmd(mm, pmd);
1472                         goto split_fallthrough;
1473                 }
1474                 spin_lock(&mm->page_table_lock);
1475                 if (likely(pmd_trans_huge(*pmd))) {
1476                         if (unlikely(pmd_trans_splitting(*pmd))) {
1477                                 spin_unlock(&mm->page_table_lock);
1478                                 wait_split_huge_page(vma->anon_vma, pmd);
1479                         } else {
1480                                 page = follow_trans_huge_pmd(mm, address,
1481                                                              pmd, flags);
1482                                 spin_unlock(&mm->page_table_lock);
1483                                 goto out;
1484                         }
1485                 } else
1486                         spin_unlock(&mm->page_table_lock);
1487                 /* fall through */
1488         }
1489 split_fallthrough:
1490         if (unlikely(pmd_bad(*pmd)))
1491                 goto no_page_table;
1492
1493         ptep = pte_offset_map_lock(mm, pmd, address, &ptl);
1494
1495         pte = *ptep;
1496         if (!pte_present(pte))
1497                 goto no_page;
1498         if ((flags & FOLL_WRITE) && !pte_write(pte))
1499                 goto unlock;
1500
1501         page = vm_normal_page(vma, address, pte);
1502         if (unlikely(!page)) {
1503                 if ((flags & FOLL_DUMP) ||
1504                     !is_zero_pfn(pte_pfn(pte)))
1505                         goto bad_page;
1506                 page = pte_page(pte);
1507         }
1508
1509         if (flags & FOLL_GET)
1510                 get_page_foll(page);
1511         if (flags & FOLL_TOUCH) {
1512                 if ((flags & FOLL_WRITE) &&
1513                     !pte_dirty(pte) && !PageDirty(page))
1514                         set_page_dirty(page);
1515                 /*
1516                  * pte_mkyoung() would be more correct here, but atomic care
1517                  * is needed to avoid losing the dirty bit: it is easier to use
1518                  * mark_page_accessed().
1519                  */
1520                 mark_page_accessed(page);
1521         }
1522         if ((flags & FOLL_MLOCK) && (vma->vm_flags & VM_LOCKED)) {
1523                 /*
1524                  * The preliminary mapping check is mainly to avoid the
1525                  * pointless overhead of lock_page on the ZERO_PAGE
1526                  * which might bounce very badly if there is contention.
1527                  *
1528                  * If the page is already locked, we don't need to
1529                  * handle it now - vmscan will handle it later if and
1530                  * when it attempts to reclaim the page.
1531                  */
1532                 if (page->mapping && trylock_page(page)) {
1533                         lru_add_drain();  /* push cached pages to LRU */
1534                         /*
1535                          * Because we lock page here and migration is
1536                          * blocked by the pte's page reference, we need
1537                          * only check for file-cache page truncation.
1538                          */
1539                         if (page->mapping)
1540                                 mlock_vma_page(page);
1541                         unlock_page(page);
1542                 }
1543         }
1544 unlock:
1545         pte_unmap_unlock(ptep, ptl);
1546 out:
1547         return page;
1548
1549 bad_page:
1550         pte_unmap_unlock(ptep, ptl);
1551         return ERR_PTR(-EFAULT);
1552
1553 no_page:
1554         pte_unmap_unlock(ptep, ptl);
1555         if (!pte_none(pte))
1556                 return page;
1557
1558 no_page_table:
1559         /*
1560          * When core dumping an enormous anonymous area that nobody
1561          * has touched so far, we don't want to allocate unnecessary pages or
1562          * page tables.  Return error instead of NULL to skip handle_mm_fault,
1563          * then get_dump_page() will return NULL to leave a hole in the dump.
1564          * But we can only make this optimization where a hole would surely
1565          * be zero-filled if handle_mm_fault() actually did handle it.
1566          */
1567         if ((flags & FOLL_DUMP) &&
1568             (!vma->vm_ops || !vma->vm_ops->fault))
1569                 return ERR_PTR(-EFAULT);
1570         return page;
1571 }
1572
1573 static inline int stack_guard_page(struct vm_area_struct *vma, unsigned long addr)
1574 {
1575         return stack_guard_page_start(vma, addr) ||
1576                stack_guard_page_end(vma, addr+PAGE_SIZE);
1577 }
1578
1579 /**
1580  * __get_user_pages() - pin user pages in memory
1581  * @tsk:        task_struct of target task
1582  * @mm:         mm_struct of target mm
1583  * @start:      starting user address
1584  * @nr_pages:   number of pages from start to pin
1585  * @gup_flags:  flags modifying pin behaviour
1586  * @pages:      array that receives pointers to the pages pinned.
1587  *              Should be at least nr_pages long. Or NULL, if caller
1588  *              only intends to ensure the pages are faulted in.
1589  * @vmas:       array of pointers to vmas corresponding to each page.
1590  *              Or NULL if the caller does not require them.
1591  * @nonblocking: whether waiting for disk IO or mmap_sem contention
1592  *
1593  * Returns number of pages pinned. This may be fewer than the number
1594  * requested. If nr_pages is 0 or negative, returns 0. If no pages
1595  * were pinned, returns -errno. Each page returned must be released
1596  * with a put_page() call when it is finished with. vmas will only
1597  * remain valid while mmap_sem is held.
1598  *
1599  * Must be called with mmap_sem held for read or write.
1600  *
1601  * __get_user_pages walks a process's page tables and takes a reference to
1602  * each struct page that each user address corresponds to at a given
1603  * instant. That is, it takes the page that would be accessed if a user
1604  * thread accesses the given user virtual address at that instant.
1605  *
1606  * This does not guarantee that the page exists in the user mappings when
1607  * __get_user_pages returns, and there may even be a completely different
1608  * page there in some cases (eg. if mmapped pagecache has been invalidated
1609  * and subsequently re faulted). However it does guarantee that the page
1610  * won't be freed completely. And mostly callers simply care that the page
1611  * contains data that was valid *at some point in time*. Typically, an IO
1612  * or similar operation cannot guarantee anything stronger anyway because
1613  * locks can't be held over the syscall boundary.
1614  *
1615  * If @gup_flags & FOLL_WRITE == 0, the page must not be written to. If
1616  * the page is written to, set_page_dirty (or set_page_dirty_lock, as
1617  * appropriate) must be called after the page is finished with, and
1618  * before put_page is called.
1619  *
1620  * If @nonblocking != NULL, __get_user_pages will not wait for disk IO
1621  * or mmap_sem contention, and if waiting is needed to pin all pages,
1622  * *@nonblocking will be set to 0.
1623  *
1624  * In most cases, get_user_pages or get_user_pages_fast should be used
1625  * instead of __get_user_pages. __get_user_pages should be used only if
1626  * you need some special @gup_flags.
1627  */
1628 int __get_user_pages(struct task_struct *tsk, struct mm_struct *mm,
1629                      unsigned long start, int nr_pages, unsigned int gup_flags,
1630                      struct page **pages, struct vm_area_struct **vmas,
1631                      int *nonblocking)
1632 {
1633         int i;
1634         unsigned long vm_flags;
1635
1636         if (nr_pages <= 0)
1637                 return 0;
1638
1639         VM_BUG_ON(!!pages != !!(gup_flags & FOLL_GET));
1640
1641         /* 
1642          * Require read or write permissions.
1643          * If FOLL_FORCE is set, we only require the "MAY" flags.
1644          */
1645         vm_flags  = (gup_flags & FOLL_WRITE) ?
1646                         (VM_WRITE | VM_MAYWRITE) : (VM_READ | VM_MAYREAD);
1647         vm_flags &= (gup_flags & FOLL_FORCE) ?
1648                         (VM_MAYREAD | VM_MAYWRITE) : (VM_READ | VM_WRITE);
1649         i = 0;
1650
1651         do {
1652                 struct vm_area_struct *vma;
1653
1654                 vma = find_extend_vma(mm, start);
1655                 if (!vma && in_gate_area(mm, start)) {
1656                         unsigned long pg = start & PAGE_MASK;
1657                         pgd_t *pgd;
1658                         pud_t *pud;
1659                         pmd_t *pmd;
1660                         pte_t *pte;
1661
1662                         /* user gate pages are read-only */
1663                         if (gup_flags & FOLL_WRITE)
1664                                 return i ? : -EFAULT;
1665                         if (pg > TASK_SIZE)
1666                                 pgd = pgd_offset_k(pg);
1667                         else
1668                                 pgd = pgd_offset_gate(mm, pg);
1669                         BUG_ON(pgd_none(*pgd));
1670                         pud = pud_offset(pgd, pg);
1671                         BUG_ON(pud_none(*pud));
1672                         pmd = pmd_offset(pud, pg);
1673                         if (pmd_none(*pmd))
1674                                 return i ? : -EFAULT;
1675                         VM_BUG_ON(pmd_trans_huge(*pmd));
1676                         pte = pte_offset_map(pmd, pg);
1677                         if (pte_none(*pte)) {
1678                                 pte_unmap(pte);
1679                                 return i ? : -EFAULT;
1680                         }
1681                         vma = get_gate_vma(mm);
1682                         if (pages) {
1683                                 struct page *page;
1684
1685                                 page = vm_normal_page(vma, start, *pte);
1686                                 if (!page) {
1687                                         if (!(gup_flags & FOLL_DUMP) &&
1688                                              is_zero_pfn(pte_pfn(*pte)))
1689                                                 page = pte_page(*pte);
1690                                         else {
1691                                                 pte_unmap(pte);
1692                                                 return i ? : -EFAULT;
1693                                         }
1694                                 }
1695                                 pages[i] = page;
1696                                 get_page(page);
1697                         }
1698                         pte_unmap(pte);
1699                         goto next_page;
1700                 }
1701
1702                 if (!vma ||
1703                     (vma->vm_flags & (VM_IO | VM_PFNMAP)) ||
1704                     !(vm_flags & vma->vm_flags))
1705                         return i ? : -EFAULT;
1706
1707                 if (is_vm_hugetlb_page(vma)) {
1708                         i = follow_hugetlb_page(mm, vma, pages, vmas,
1709                                         &start, &nr_pages, i, gup_flags);
1710                         continue;
1711                 }
1712
1713                 do {
1714                         struct page *page;
1715                         unsigned int foll_flags = gup_flags;
1716
1717                         /*
1718                          * If we have a pending SIGKILL, don't keep faulting
1719                          * pages and potentially allocating memory.
1720                          */
1721                         if (unlikely(fatal_signal_pending(current)))
1722                                 return i ? i : -ERESTARTSYS;
1723
1724                         cond_resched();
1725                         while (!(page = follow_page(vma, start, foll_flags))) {
1726                                 int ret;
1727                                 unsigned int fault_flags = 0;
1728
1729                                 /* For mlock, just skip the stack guard page. */
1730                                 if (foll_flags & FOLL_MLOCK) {
1731                                         if (stack_guard_page(vma, start))
1732                                                 goto next_page;
1733                                 }
1734                                 if (foll_flags & FOLL_WRITE)
1735                                         fault_flags |= FAULT_FLAG_WRITE;
1736                                 if (nonblocking)
1737                                         fault_flags |= FAULT_FLAG_ALLOW_RETRY;
1738                                 if (foll_flags & FOLL_NOWAIT)
1739                                         fault_flags |= (FAULT_FLAG_ALLOW_RETRY | FAULT_FLAG_RETRY_NOWAIT);
1740
1741                                 ret = handle_mm_fault(mm, vma, start,
1742                                                         fault_flags);
1743
1744                                 if (ret & VM_FAULT_ERROR) {
1745                                         if (ret & VM_FAULT_OOM)
1746                                                 return i ? i : -ENOMEM;
1747                                         if (ret & (VM_FAULT_HWPOISON |
1748                                                    VM_FAULT_HWPOISON_LARGE)) {
1749                                                 if (i)
1750                                                         return i;
1751                                                 else if (gup_flags & FOLL_HWPOISON)
1752                                                         return -EHWPOISON;
1753                                                 else
1754                                                         return -EFAULT;
1755                                         }
1756                                         if (ret & VM_FAULT_SIGBUS)
1757                                                 return i ? i : -EFAULT;
1758                                         BUG();
1759                                 }
1760
1761                                 if (tsk) {
1762                                         if (ret & VM_FAULT_MAJOR)
1763                                                 tsk->maj_flt++;
1764                                         else
1765                                                 tsk->min_flt++;
1766                                 }
1767
1768                                 if (ret & VM_FAULT_RETRY) {
1769                                         if (nonblocking)
1770                                                 *nonblocking = 0;
1771                                         return i;
1772                                 }
1773
1774                                 /*
1775                                  * The VM_FAULT_WRITE bit tells us that
1776                                  * do_wp_page has broken COW when necessary,
1777                                  * even if maybe_mkwrite decided not to set
1778                                  * pte_write. We can thus safely do subsequent
1779                                  * page lookups as if they were reads. But only
1780                                  * do so when looping for pte_write is futile:
1781                                  * in some cases userspace may also be wanting
1782                                  * to write to the gotten user page, which a
1783                                  * read fault here might prevent (a readonly
1784                                  * page might get reCOWed by userspace write).
1785                                  */
1786                                 if ((ret & VM_FAULT_WRITE) &&
1787                                     !(vma->vm_flags & VM_WRITE))
1788                                         foll_flags &= ~FOLL_WRITE;
1789
1790                                 cond_resched();
1791                         }
1792                         if (IS_ERR(page))
1793                                 return i ? i : PTR_ERR(page);
1794                         if (pages) {
1795                                 pages[i] = page;
1796
1797                                 flush_anon_page(vma, page, start);
1798                                 flush_dcache_page(page);
1799                         }
1800 next_page:
1801                         if (vmas)
1802                                 vmas[i] = vma;
1803                         i++;
1804                         start += PAGE_SIZE;
1805                         nr_pages--;
1806                 } while (nr_pages && start < vma->vm_end);
1807         } while (nr_pages);
1808         return i;
1809 }
1810 EXPORT_SYMBOL(__get_user_pages);
1811
1812 /*
1813  * fixup_user_fault() - manually resolve a user page fault
1814  * @tsk:        the task_struct to use for page fault accounting, or
1815  *              NULL if faults are not to be recorded.
1816  * @mm:         mm_struct of target mm
1817  * @address:    user address
1818  * @fault_flags:flags to pass down to handle_mm_fault()
1819  *
1820  * This is meant to be called in the specific scenario where for locking reasons
1821  * we try to access user memory in atomic context (within a pagefault_disable()
1822  * section), this returns -EFAULT, and we want to resolve the user fault before
1823  * trying again.
1824  *
1825  * Typically this is meant to be used by the futex code.
1826  *
1827  * The main difference with get_user_pages() is that this function will
1828  * unconditionally call handle_mm_fault() which will in turn perform all the
1829  * necessary SW fixup of the dirty and young bits in the PTE, while
1830  * handle_mm_fault() only guarantees to update these in the struct page.
1831  *
1832  * This is important for some architectures where those bits also gate the
1833  * access permission to the page because they are maintained in software.  On
1834  * such architectures, gup() will not be enough to make a subsequent access
1835  * succeed.
1836  *
1837  * This should be called with the mm_sem held for read.
1838  */
1839 int fixup_user_fault(struct task_struct *tsk, struct mm_struct *mm,
1840                      unsigned long address, unsigned int fault_flags)
1841 {
1842         struct vm_area_struct *vma;
1843         int ret;
1844
1845         vma = find_extend_vma(mm, address);
1846         if (!vma || address < vma->vm_start)
1847                 return -EFAULT;
1848
1849         ret = handle_mm_fault(mm, vma, address, fault_flags);
1850         if (ret & VM_FAULT_ERROR) {
1851                 if (ret & VM_FAULT_OOM)
1852                         return -ENOMEM;
1853                 if (ret & (VM_FAULT_HWPOISON | VM_FAULT_HWPOISON_LARGE))
1854                         return -EHWPOISON;
1855                 if (ret & VM_FAULT_SIGBUS)
1856                         return -EFAULT;
1857                 BUG();
1858         }
1859         if (tsk) {
1860                 if (ret & VM_FAULT_MAJOR)
1861                         tsk->maj_flt++;
1862                 else
1863                         tsk->min_flt++;
1864         }
1865         return 0;
1866 }
1867
1868 /*
1869  * get_user_pages() - pin user pages in memory
1870  * @tsk:        the task_struct to use for page fault accounting, or
1871  *              NULL if faults are not to be recorded.
1872  * @mm:         mm_struct of target mm
1873  * @start:      starting user address
1874  * @nr_pages:   number of pages from start to pin
1875  * @write:      whether pages will be written to by the caller
1876  * @force:      whether to force write access even if user mapping is
1877  *              readonly. This will result in the page being COWed even
1878  *              in MAP_SHARED mappings. You do not want this.
1879  * @pages:      array that receives pointers to the pages pinned.
1880  *              Should be at least nr_pages long. Or NULL, if caller
1881  *              only intends to ensure the pages are faulted in.
1882  * @vmas:       array of pointers to vmas corresponding to each page.
1883  *              Or NULL if the caller does not require them.
1884  *
1885  * Returns number of pages pinned. This may be fewer than the number
1886  * requested. If nr_pages is 0 or negative, returns 0. If no pages
1887  * were pinned, returns -errno. Each page returned must be released
1888  * with a put_page() call when it is finished with. vmas will only
1889  * remain valid while mmap_sem is held.
1890  *
1891  * Must be called with mmap_sem held for read or write.
1892  *
1893  * get_user_pages walks a process's page tables and takes a reference to
1894  * each struct page that each user address corresponds to at a given
1895  * instant. That is, it takes the page that would be accessed if a user
1896  * thread accesses the given user virtual address at that instant.
1897  *
1898  * This does not guarantee that the page exists in the user mappings when
1899  * get_user_pages returns, and there may even be a completely different
1900  * page there in some cases (eg. if mmapped pagecache has been invalidated
1901  * and subsequently re faulted). However it does guarantee that the page
1902  * won't be freed completely. And mostly callers simply care that the page
1903  * contains data that was valid *at some point in time*. Typically, an IO
1904  * or similar operation cannot guarantee anything stronger anyway because
1905  * locks can't be held over the syscall boundary.
1906  *
1907  * If write=0, the page must not be written to. If the page is written to,
1908  * set_page_dirty (or set_page_dirty_lock, as appropriate) must be called
1909  * after the page is finished with, and before put_page is called.
1910  *
1911  * get_user_pages is typically used for fewer-copy IO operations, to get a
1912  * handle on the memory by some means other than accesses via the user virtual
1913  * addresses. The pages may be submitted for DMA to devices or accessed via
1914  * their kernel linear mapping (via the kmap APIs). Care should be taken to
1915  * use the correct cache flushing APIs.
1916  *
1917  * See also get_user_pages_fast, for performance critical applications.
1918  */
1919 int get_user_pages(struct task_struct *tsk, struct mm_struct *mm,
1920                 unsigned long start, int nr_pages, int write, int force,
1921                 struct page **pages, struct vm_area_struct **vmas)
1922 {
1923         int flags = FOLL_TOUCH;
1924
1925         if (pages)
1926                 flags |= FOLL_GET;
1927         if (write)
1928                 flags |= FOLL_WRITE;
1929         if (force)
1930                 flags |= FOLL_FORCE;
1931
1932         return __get_user_pages(tsk, mm, start, nr_pages, flags, pages, vmas,
1933                                 NULL);
1934 }
1935 EXPORT_SYMBOL(get_user_pages);
1936
1937 /**
1938  * get_dump_page() - pin user page in memory while writing it to core dump
1939  * @addr: user address
1940  *
1941  * Returns struct page pointer of user page pinned for dump,
1942  * to be freed afterwards by page_cache_release() or put_page().
1943  *
1944  * Returns NULL on any kind of failure - a hole must then be inserted into
1945  * the corefile, to preserve alignment with its headers; and also returns
1946  * NULL wherever the ZERO_PAGE, or an anonymous pte_none, has been found -
1947  * allowing a hole to be left in the corefile to save diskspace.
1948  *
1949  * Called without mmap_sem, but after all other threads have been killed.
1950  */
1951 #ifdef CONFIG_ELF_CORE
1952 struct page *get_dump_page(unsigned long addr)
1953 {
1954         struct vm_area_struct *vma;
1955         struct page *page;
1956
1957         if (__get_user_pages(current, current->mm, addr, 1,
1958                              FOLL_FORCE | FOLL_DUMP | FOLL_GET, &page, &vma,
1959                              NULL) < 1)
1960                 return NULL;
1961         flush_cache_page(vma, addr, page_to_pfn(page));
1962         return page;
1963 }
1964 #endif /* CONFIG_ELF_CORE */
1965
1966 pte_t *__get_locked_pte(struct mm_struct *mm, unsigned long addr,
1967                         spinlock_t **ptl)
1968 {
1969         pgd_t * pgd = pgd_offset(mm, addr);
1970         pud_t * pud = pud_alloc(mm, pgd, addr);
1971         if (pud) {
1972                 pmd_t * pmd = pmd_alloc(mm, pud, addr);
1973                 if (pmd) {
1974                         VM_BUG_ON(pmd_trans_huge(*pmd));
1975                         return pte_alloc_map_lock(mm, pmd, addr, ptl);
1976                 }
1977         }
1978         return NULL;
1979 }
1980
1981 /*
1982  * This is the old fallback for page remapping.
1983  *
1984  * For historical reasons, it only allows reserved pages. Only
1985  * old drivers should use this, and they needed to mark their
1986  * pages reserved for the old functions anyway.
1987  */
1988 static int insert_page(struct vm_area_struct *vma, unsigned long addr,
1989                         struct page *page, pgprot_t prot)
1990 {
1991         struct mm_struct *mm = vma->vm_mm;
1992         int retval;
1993         pte_t *pte;
1994         spinlock_t *ptl;
1995
1996         retval = -EINVAL;
1997         if (PageAnon(page))
1998                 goto out;
1999         retval = -ENOMEM;
2000         flush_dcache_page(page);
2001         pte = get_locked_pte(mm, addr, &ptl);
2002         if (!pte)
2003                 goto out;
2004         retval = -EBUSY;
2005         if (!pte_none(*pte))
2006                 goto out_unlock;
2007
2008         /* Ok, finally just insert the thing.. */
2009         get_page(page);
2010         inc_mm_counter_fast(mm, MM_FILEPAGES);
2011         page_add_file_rmap(page);
2012         set_pte_at(mm, addr, pte, mk_pte(page, prot));
2013
2014         retval = 0;
2015         pte_unmap_unlock(pte, ptl);
2016         return retval;
2017 out_unlock:
2018         pte_unmap_unlock(pte, ptl);
2019 out:
2020         return retval;
2021 }
2022
2023 /**
2024  * vm_insert_page - insert single page into user vma
2025  * @vma: user vma to map to
2026  * @addr: target user address of this page
2027  * @page: source kernel page
2028  *
2029  * This allows drivers to insert individual pages they've allocated
2030  * into a user vma.
2031  *
2032  * The page has to be a nice clean _individual_ kernel allocation.
2033  * If you allocate a compound page, you need to have marked it as
2034  * such (__GFP_COMP), or manually just split the page up yourself
2035  * (see split_page()).
2036  *
2037  * NOTE! Traditionally this was done with "remap_pfn_range()" which
2038  * took an arbitrary page protection parameter. This doesn't allow
2039  * that. Your vma protection will have to be set up correctly, which
2040  * means that if you want a shared writable mapping, you'd better
2041  * ask for a shared writable mapping!
2042  *
2043  * The page does not need to be reserved.
2044  */
2045 int vm_insert_page(struct vm_area_struct *vma, unsigned long addr,
2046                         struct page *page)
2047 {
2048         if (addr < vma->vm_start || addr >= vma->vm_end)
2049                 return -EFAULT;
2050         if (!page_count(page))
2051                 return -EINVAL;
2052         vma->vm_flags |= VM_INSERTPAGE;
2053         return insert_page(vma, addr, page, vma->vm_page_prot);
2054 }
2055 EXPORT_SYMBOL(vm_insert_page);
2056
2057 static int insert_pfn(struct vm_area_struct *vma, unsigned long addr,
2058                         unsigned long pfn, pgprot_t prot)
2059 {
2060         struct mm_struct *mm = vma->vm_mm;
2061         int retval;
2062         pte_t *pte, entry;
2063         spinlock_t *ptl;
2064
2065         retval = -ENOMEM;
2066         pte = get_locked_pte(mm, addr, &ptl);
2067         if (!pte)
2068                 goto out;
2069         retval = -EBUSY;
2070         if (!pte_none(*pte))
2071                 goto out_unlock;
2072
2073         /* Ok, finally just insert the thing.. */
2074         entry = pte_mkspecial(pfn_pte(pfn, prot));
2075         set_pte_at(mm, addr, pte, entry);
2076         update_mmu_cache(vma, addr, pte); /* XXX: why not for insert_page? */
2077
2078         retval = 0;
2079 out_unlock:
2080         pte_unmap_unlock(pte, ptl);
2081 out:
2082         return retval;
2083 }
2084
2085 /**
2086  * vm_insert_pfn - insert single pfn into user vma
2087  * @vma: user vma to map to
2088  * @addr: target user address of this page
2089  * @pfn: source kernel pfn
2090  *
2091  * Similar to vm_inert_page, this allows drivers to insert individual pages
2092  * they've allocated into a user vma. Same comments apply.
2093  *
2094  * This function should only be called from a vm_ops->fault handler, and
2095  * in that case the handler should return NULL.
2096  *
2097  * vma cannot be a COW mapping.
2098  *
2099  * As this is called only for pages that do not currently exist, we
2100  * do not need to flush old virtual caches or the TLB.
2101  */
2102 int vm_insert_pfn(struct vm_area_struct *vma, unsigned long addr,
2103                         unsigned long pfn)
2104 {
2105         int ret;
2106         pgprot_t pgprot = vma->vm_page_prot;
2107         /*
2108          * Technically, architectures with pte_special can avoid all these
2109          * restrictions (same for remap_pfn_range).  However we would like
2110          * consistency in testing and feature parity among all, so we should
2111          * try to keep these invariants in place for everybody.
2112          */
2113         BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)));
2114         BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) ==
2115                                                 (VM_PFNMAP|VM_MIXEDMAP));
2116         BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags));
2117         BUG_ON((vma->vm_flags & VM_MIXEDMAP) && pfn_valid(pfn));
2118
2119         if (addr < vma->vm_start || addr >= vma->vm_end)
2120                 return -EFAULT;
2121         if (track_pfn_vma_new(vma, &pgprot, pfn, PAGE_SIZE))
2122                 return -EINVAL;
2123
2124         ret = insert_pfn(vma, addr, pfn, pgprot);
2125
2126         if (ret)
2127                 untrack_pfn_vma(vma, pfn, PAGE_SIZE);
2128
2129         return ret;
2130 }
2131 EXPORT_SYMBOL(vm_insert_pfn);
2132
2133 int vm_insert_mixed(struct vm_area_struct *vma, unsigned long addr,
2134                         unsigned long pfn)
2135 {
2136         BUG_ON(!(vma->vm_flags & VM_MIXEDMAP));
2137
2138         if (addr < vma->vm_start || addr >= vma->vm_end)
2139                 return -EFAULT;
2140
2141         /*
2142          * If we don't have pte special, then we have to use the pfn_valid()
2143          * based VM_MIXEDMAP scheme (see vm_normal_page), and thus we *must*
2144          * refcount the page if pfn_valid is true (hence insert_page rather
2145          * than insert_pfn).  If a zero_pfn were inserted into a VM_MIXEDMAP
2146          * without pte special, it would there be refcounted as a normal page.
2147          */
2148         if (!HAVE_PTE_SPECIAL && pfn_valid(pfn)) {
2149                 struct page *page;
2150
2151                 page = pfn_to_page(pfn);
2152                 return insert_page(vma, addr, page, vma->vm_page_prot);
2153         }
2154         return insert_pfn(vma, addr, pfn, vma->vm_page_prot);
2155 }
2156 EXPORT_SYMBOL(vm_insert_mixed);
2157
2158 /*
2159  * maps a range of physical memory into the requested pages. the old
2160  * mappings are removed. any references to nonexistent pages results
2161  * in null mappings (currently treated as "copy-on-access")
2162  */
2163 static int remap_pte_range(struct mm_struct *mm, pmd_t *pmd,
2164                         unsigned long addr, unsigned long end,
2165                         unsigned long pfn, pgprot_t prot)
2166 {
2167         pte_t *pte;
2168         spinlock_t *ptl;
2169
2170         pte = pte_alloc_map_lock(mm, pmd, addr, &ptl);
2171         if (!pte)
2172                 return -ENOMEM;
2173         arch_enter_lazy_mmu_mode();
2174         do {
2175                 BUG_ON(!pte_none(*pte));
2176                 set_pte_at(mm, addr, pte, pte_mkspecial(pfn_pte(pfn, prot)));
2177                 pfn++;
2178         } while (pte++, addr += PAGE_SIZE, addr != end);
2179         arch_leave_lazy_mmu_mode();
2180         pte_unmap_unlock(pte - 1, ptl);
2181         return 0;
2182 }
2183
2184 static inline int remap_pmd_range(struct mm_struct *mm, pud_t *pud,
2185                         unsigned long addr, unsigned long end,
2186                         unsigned long pfn, pgprot_t prot)
2187 {
2188         pmd_t *pmd;
2189         unsigned long next;
2190
2191         pfn -= addr >> PAGE_SHIFT;
2192         pmd = pmd_alloc(mm, pud, addr);
2193         if (!pmd)
2194                 return -ENOMEM;
2195         VM_BUG_ON(pmd_trans_huge(*pmd));
2196         do {
2197                 next = pmd_addr_end(addr, end);
2198                 if (remap_pte_range(mm, pmd, addr, next,
2199                                 pfn + (addr >> PAGE_SHIFT), prot))
2200                         return -ENOMEM;
2201         } while (pmd++, addr = next, addr != end);
2202         return 0;
2203 }
2204
2205 static inline int remap_pud_range(struct mm_struct *mm, pgd_t *pgd,
2206                         unsigned long addr, unsigned long end,
2207                         unsigned long pfn, pgprot_t prot)
2208 {
2209         pud_t *pud;
2210         unsigned long next;
2211
2212         pfn -= addr >> PAGE_SHIFT;
2213         pud = pud_alloc(mm, pgd, addr);
2214         if (!pud)
2215                 return -ENOMEM;
2216         do {
2217                 next = pud_addr_end(addr, end);
2218                 if (remap_pmd_range(mm, pud, addr, next,
2219                                 pfn + (addr >> PAGE_SHIFT), prot))
2220                         return -ENOMEM;
2221         } while (pud++, addr = next, addr != end);
2222         return 0;
2223 }
2224
2225 /**
2226  * remap_pfn_range - remap kernel memory to userspace
2227  * @vma: user vma to map to
2228  * @addr: target user address to start at
2229  * @pfn: physical address of kernel memory
2230  * @size: size of map area
2231  * @prot: page protection flags for this mapping
2232  *
2233  *  Note: this is only safe if the mm semaphore is held when called.
2234  */
2235 int remap_pfn_range(struct vm_area_struct *vma, unsigned long addr,
2236                     unsigned long pfn, unsigned long size, pgprot_t prot)
2237 {
2238         pgd_t *pgd;
2239         unsigned long next;
2240         unsigned long end = addr + PAGE_ALIGN(size);
2241         struct mm_struct *mm = vma->vm_mm;
2242         int err;
2243
2244         /*
2245          * Physically remapped pages are special. Tell the
2246          * rest of the world about it:
2247          *   VM_IO tells people not to look at these pages
2248          *      (accesses can have side effects).
2249          *   VM_RESERVED is specified all over the place, because
2250          *      in 2.4 it kept swapout's vma scan off this vma; but
2251          *      in 2.6 the LRU scan won't even find its pages, so this
2252          *      flag means no more than count its pages in reserved_vm,
2253          *      and omit it from core dump, even when VM_IO turned off.
2254          *   VM_PFNMAP tells the core MM that the base pages are just
2255          *      raw PFN mappings, and do not have a "struct page" associated
2256          *      with them.
2257          *
2258          * There's a horrible special case to handle copy-on-write
2259          * behaviour that some programs depend on. We mark the "original"
2260          * un-COW'ed pages by matching them up with "vma->vm_pgoff".
2261          */
2262         if (addr == vma->vm_start && end == vma->vm_end) {
2263                 vma->vm_pgoff = pfn;
2264                 vma->vm_flags |= VM_PFN_AT_MMAP;
2265         } else if (is_cow_mapping(vma->vm_flags))
2266                 return -EINVAL;
2267
2268         vma->vm_flags |= VM_IO | VM_RESERVED | VM_PFNMAP;
2269
2270         err = track_pfn_vma_new(vma, &prot, pfn, PAGE_ALIGN(size));
2271         if (err) {
2272                 /*
2273                  * To indicate that track_pfn related cleanup is not
2274                  * needed from higher level routine calling unmap_vmas
2275                  */
2276                 vma->vm_flags &= ~(VM_IO | VM_RESERVED | VM_PFNMAP);
2277                 vma->vm_flags &= ~VM_PFN_AT_MMAP;
2278                 return -EINVAL;
2279         }
2280
2281         BUG_ON(addr >= end);
2282         pfn -= addr >> PAGE_SHIFT;
2283         pgd = pgd_offset(mm, addr);
2284         flush_cache_range(vma, addr, end);
2285         do {
2286                 next = pgd_addr_end(addr, end);
2287                 err = remap_pud_range(mm, pgd, addr, next,
2288                                 pfn + (addr >> PAGE_SHIFT), prot);
2289                 if (err)
2290                         break;
2291         } while (pgd++, addr = next, addr != end);
2292
2293         if (err)
2294                 untrack_pfn_vma(vma, pfn, PAGE_ALIGN(size));
2295
2296         return err;
2297 }
2298 EXPORT_SYMBOL(remap_pfn_range);
2299
2300 static int apply_to_pte_range(struct mm_struct *mm, pmd_t *pmd,
2301                                      unsigned long addr, unsigned long end,
2302                                      pte_fn_t fn, void *data)
2303 {
2304         pte_t *pte;
2305         int err;
2306         pgtable_t token;
2307         spinlock_t *uninitialized_var(ptl);
2308
2309         pte = (mm == &init_mm) ?
2310                 pte_alloc_kernel(pmd, addr) :
2311                 pte_alloc_map_lock(mm, pmd, addr, &ptl);
2312         if (!pte)
2313                 return -ENOMEM;
2314
2315         BUG_ON(pmd_huge(*pmd));
2316
2317         arch_enter_lazy_mmu_mode();
2318
2319         token = pmd_pgtable(*pmd);
2320
2321         do {
2322                 err = fn(pte++, token, addr, data);
2323                 if (err)
2324                         break;
2325         } while (addr += PAGE_SIZE, addr != end);
2326
2327         arch_leave_lazy_mmu_mode();
2328
2329         if (mm != &init_mm)
2330                 pte_unmap_unlock(pte-1, ptl);
2331         return err;
2332 }
2333
2334 static int apply_to_pmd_range(struct mm_struct *mm, pud_t *pud,
2335                                      unsigned long addr, unsigned long end,
2336                                      pte_fn_t fn, void *data)
2337 {
2338         pmd_t *pmd;
2339         unsigned long next;
2340         int err;
2341
2342         BUG_ON(pud_huge(*pud));
2343
2344         pmd = pmd_alloc(mm, pud, addr);
2345         if (!pmd)
2346                 return -ENOMEM;
2347         do {
2348                 next = pmd_addr_end(addr, end);
2349                 err = apply_to_pte_range(mm, pmd, addr, next, fn, data);
2350                 if (err)
2351                         break;
2352         } while (pmd++, addr = next, addr != end);
2353         return err;
2354 }
2355
2356 static int apply_to_pud_range(struct mm_struct *mm, pgd_t *pgd,
2357                                      unsigned long addr, unsigned long end,
2358                                      pte_fn_t fn, void *data)
2359 {
2360         pud_t *pud;
2361         unsigned long next;
2362         int err;
2363
2364         pud = pud_alloc(mm, pgd, addr);
2365         if (!pud)
2366                 return -ENOMEM;
2367         do {
2368                 next = pud_addr_end(addr, end);
2369                 err = apply_to_pmd_range(mm, pud, addr, next, fn, data);
2370                 if (err)
2371                         break;
2372         } while (pud++, addr = next, addr != end);
2373         return err;
2374 }
2375
2376 /*
2377  * Scan a region of virtual memory, filling in page tables as necessary
2378  * and calling a provided function on each leaf page table.
2379  */
2380 int apply_to_page_range(struct mm_struct *mm, unsigned long addr,
2381                         unsigned long size, pte_fn_t fn, void *data)
2382 {
2383         pgd_t *pgd;
2384         unsigned long next;
2385         unsigned long end = addr + size;
2386         int err;
2387
2388         BUG_ON(addr >= end);
2389         pgd = pgd_offset(mm, addr);
2390         do {
2391                 next = pgd_addr_end(addr, end);
2392                 err = apply_to_pud_range(mm, pgd, addr, next, fn, data);
2393                 if (err)
2394                         break;
2395         } while (pgd++, addr = next, addr != end);
2396
2397         return err;
2398 }
2399 EXPORT_SYMBOL_GPL(apply_to_page_range);
2400
2401 /*
2402  * handle_pte_fault chooses page fault handler according to an entry
2403  * which was read non-atomically.  Before making any commitment, on
2404  * those architectures or configurations (e.g. i386 with PAE) which
2405  * might give a mix of unmatched parts, do_swap_page and do_nonlinear_fault
2406  * must check under lock before unmapping the pte and proceeding
2407  * (but do_wp_page is only called after already making such a check;
2408  * and do_anonymous_page can safely check later on).
2409  */
2410 static inline int pte_unmap_same(struct mm_struct *mm, pmd_t *pmd,
2411                                 pte_t *page_table, pte_t orig_pte)
2412 {
2413         int same = 1;
2414 #if defined(CONFIG_SMP) || defined(CONFIG_PREEMPT)
2415         if (sizeof(pte_t) > sizeof(unsigned long)) {
2416                 spinlock_t *ptl = pte_lockptr(mm, pmd);
2417                 spin_lock(ptl);
2418                 same = pte_same(*page_table, orig_pte);
2419                 spin_unlock(ptl);
2420         }
2421 #endif
2422         pte_unmap(page_table);
2423         return same;
2424 }
2425
2426 static inline void cow_user_page(struct page *dst, struct page *src, unsigned long va, struct vm_area_struct *vma)
2427 {
2428         /*
2429          * If the source page was a PFN mapping, we don't have
2430          * a "struct page" for it. We do a best-effort copy by
2431          * just copying from the original user address. If that
2432          * fails, we just zero-fill it. Live with it.
2433          */
2434         if (unlikely(!src)) {
2435                 void *kaddr = kmap_atomic(dst);
2436                 void __user *uaddr = (void __user *)(va & PAGE_MASK);
2437
2438                 /*
2439                  * This really shouldn't fail, because the page is there
2440                  * in the page tables. But it might just be unreadable,
2441                  * in which case we just give up and fill the result with
2442                  * zeroes.
2443                  */
2444                 if (__copy_from_user_inatomic(kaddr, uaddr, PAGE_SIZE))
2445                         clear_page(kaddr);
2446                 kunmap_atomic(kaddr);
2447                 flush_dcache_page(dst);
2448         } else
2449                 copy_user_highpage(dst, src, va, vma);
2450 }
2451
2452 /*
2453  * This routine handles present pages, when users try to write
2454  * to a shared page. It is done by copying the page to a new address
2455  * and decrementing the shared-page counter for the old page.
2456  *
2457  * Note that this routine assumes that the protection checks have been
2458  * done by the caller (the low-level page fault routine in most cases).
2459  * Thus we can safely just mark it writable once we've done any necessary
2460  * COW.
2461  *
2462  * We also mark the page dirty at this point even though the page will
2463  * change only once the write actually happens. This avoids a few races,
2464  * and potentially makes it more efficient.
2465  *
2466  * We enter with non-exclusive mmap_sem (to exclude vma changes,
2467  * but allow concurrent faults), with pte both mapped and locked.
2468  * We return with mmap_sem still held, but pte unmapped and unlocked.
2469  */
2470 static int do_wp_page(struct mm_struct *mm, struct vm_area_struct *vma,
2471                 unsigned long address, pte_t *page_table, pmd_t *pmd,
2472                 spinlock_t *ptl, pte_t orig_pte)
2473         __releases(ptl)
2474 {
2475         struct page *old_page, *new_page;
2476         pte_t entry;
2477         int ret = 0;
2478         int page_mkwrite = 0;
2479         struct page *dirty_page = NULL;
2480
2481         old_page = vm_normal_page(vma, address, orig_pte);
2482         if (!old_page) {
2483                 /*
2484                  * VM_MIXEDMAP !pfn_valid() case
2485                  *
2486                  * We should not cow pages in a shared writeable mapping.
2487                  * Just mark the pages writable as we can't do any dirty
2488                  * accounting on raw pfn maps.
2489                  */
2490                 if ((vma->vm_flags & (VM_WRITE|VM_SHARED)) ==
2491                                      (VM_WRITE|VM_SHARED))
2492                         goto reuse;
2493                 goto gotten;
2494         }
2495
2496         /*
2497          * Take out anonymous pages first, anonymous shared vmas are
2498          * not dirty accountable.
2499          */
2500         if (PageAnon(old_page) && !PageKsm(old_page)) {
2501                 if (!trylock_page(old_page)) {
2502                         page_cache_get(old_page);
2503                         pte_unmap_unlock(page_table, ptl);
2504                         lock_page(old_page);
2505                         page_table = pte_offset_map_lock(mm, pmd, address,
2506                                                          &ptl);
2507                         if (!pte_same(*page_table, orig_pte)) {
2508                                 unlock_page(old_page);
2509                                 goto unlock;
2510                         }
2511                         page_cache_release(old_page);
2512                 }
2513                 if (reuse_swap_page(old_page)) {
2514                         /*
2515                          * The page is all ours.  Move it to our anon_vma so
2516                          * the rmap code will not search our parent or siblings.
2517                          * Protected against the rmap code by the page lock.
2518                          */
2519                         page_move_anon_rmap(old_page, vma, address);
2520                         unlock_page(old_page);
2521                         goto reuse;
2522                 }
2523                 unlock_page(old_page);
2524         } else if (unlikely((vma->vm_flags & (VM_WRITE|VM_SHARED)) ==
2525                                         (VM_WRITE|VM_SHARED))) {
2526                 /*
2527                  * Only catch write-faults on shared writable pages,
2528                  * read-only shared pages can get COWed by
2529                  * get_user_pages(.write=1, .force=1).
2530                  */
2531                 if (vma->vm_ops && vma->vm_ops->page_mkwrite) {
2532                         struct vm_fault vmf;
2533                         int tmp;
2534
2535                         vmf.virtual_address = (void __user *)(address &
2536                                                                 PAGE_MASK);
2537                         vmf.pgoff = old_page->index;
2538                         vmf.flags = FAULT_FLAG_WRITE|FAULT_FLAG_MKWRITE;
2539                         vmf.page = old_page;
2540
2541                         /*
2542                          * Notify the address space that the page is about to
2543                          * become writable so that it can prohibit this or wait
2544                          * for the page to get into an appropriate state.
2545                          *
2546                          * We do this without the lock held, so that it can
2547                          * sleep if it needs to.
2548                          */
2549                         page_cache_get(old_page);
2550                         pte_unmap_unlock(page_table, ptl);
2551
2552                         tmp = vma->vm_ops->page_mkwrite(vma, &vmf);
2553                         if (unlikely(tmp &
2554                                         (VM_FAULT_ERROR | VM_FAULT_NOPAGE))) {
2555                                 ret = tmp;
2556                                 goto unwritable_page;
2557                         }
2558                         if (unlikely(!(tmp & VM_FAULT_LOCKED))) {
2559                                 lock_page(old_page);
2560                                 if (!old_page->mapping) {
2561                                         ret = 0; /* retry the fault */
2562                                         unlock_page(old_page);
2563                                         goto unwritable_page;
2564                                 }
2565                         } else
2566                                 VM_BUG_ON(!PageLocked(old_page));
2567
2568                         /*
2569                          * Since we dropped the lock we need to revalidate
2570                          * the PTE as someone else may have changed it.  If
2571                          * they did, we just return, as we can count on the
2572                          * MMU to tell us if they didn't also make it writable.
2573                          */
2574                         page_table = pte_offset_map_lock(mm, pmd, address,
2575                                                          &ptl);
2576                         if (!pte_same(*page_table, orig_pte)) {
2577                                 unlock_page(old_page);
2578                                 goto unlock;
2579                         }
2580
2581                         page_mkwrite = 1;
2582                 }
2583                 dirty_page = old_page;
2584                 get_page(dirty_page);
2585
2586 reuse:
2587                 flush_cache_page(vma, address, pte_pfn(orig_pte));
2588                 entry = pte_mkyoung(orig_pte);
2589                 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2590                 if (ptep_set_access_flags(vma, address, page_table, entry,1))
2591                         update_mmu_cache(vma, address, page_table);
2592                 pte_unmap_unlock(page_table, ptl);
2593                 ret |= VM_FAULT_WRITE;
2594
2595                 if (!dirty_page)
2596                         return ret;
2597
2598                 /*
2599                  * Yes, Virginia, this is actually required to prevent a race
2600                  * with clear_page_dirty_for_io() from clearing the page dirty
2601                  * bit after it clear all dirty ptes, but before a racing
2602                  * do_wp_page installs a dirty pte.
2603                  *
2604                  * __do_fault is protected similarly.
2605                  */
2606                 if (!page_mkwrite) {
2607                         wait_on_page_locked(dirty_page);
2608                         set_page_dirty_balance(dirty_page, page_mkwrite);
2609                 }
2610                 put_page(dirty_page);
2611                 if (page_mkwrite) {
2612                         struct address_space *mapping = dirty_page->mapping;
2613
2614                         set_page_dirty(dirty_page);
2615                         unlock_page(dirty_page);
2616                         page_cache_release(dirty_page);
2617                         if (mapping)    {
2618                                 /*
2619                                  * Some device drivers do not set page.mapping
2620                                  * but still dirty their pages
2621                                  */
2622                                 balance_dirty_pages_ratelimited(mapping);
2623                         }
2624                 }
2625
2626                 /* file_update_time outside page_lock */
2627                 if (vma->vm_file)
2628                         file_update_time(vma->vm_file);
2629
2630                 return ret;
2631         }
2632
2633         /*
2634          * Ok, we need to copy. Oh, well..
2635          */
2636         page_cache_get(old_page);
2637 gotten:
2638         pte_unmap_unlock(page_table, ptl);
2639
2640         if (unlikely(anon_vma_prepare(vma)))
2641                 goto oom;
2642
2643         if (is_zero_pfn(pte_pfn(orig_pte))) {
2644                 new_page = alloc_zeroed_user_highpage_movable(vma, address);
2645                 if (!new_page)
2646                         goto oom;
2647         } else {
2648                 new_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, address);
2649                 if (!new_page)
2650                         goto oom;
2651                 cow_user_page(new_page, old_page, address, vma);
2652         }
2653         __SetPageUptodate(new_page);
2654
2655         if (mem_cgroup_newpage_charge(new_page, mm, GFP_KERNEL))
2656                 goto oom_free_new;
2657
2658         /*
2659          * Re-check the pte - we dropped the lock
2660          */
2661         page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2662         if (likely(pte_same(*page_table, orig_pte))) {
2663                 if (old_page) {
2664                         if (!PageAnon(old_page)) {
2665                                 dec_mm_counter_fast(mm, MM_FILEPAGES);
2666                                 inc_mm_counter_fast(mm, MM_ANONPAGES);
2667                         }
2668                 } else
2669                         inc_mm_counter_fast(mm, MM_ANONPAGES);
2670                 flush_cache_page(vma, address, pte_pfn(orig_pte));
2671                 entry = mk_pte(new_page, vma->vm_page_prot);
2672                 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2673                 /*
2674                  * Clear the pte entry and flush it first, before updating the
2675                  * pte with the new entry. This will avoid a race condition
2676                  * seen in the presence of one thread doing SMC and another
2677                  * thread doing COW.
2678                  */
2679                 ptep_clear_flush(vma, address, page_table);
2680                 page_add_new_anon_rmap(new_page, vma, address);
2681                 /*
2682                  * We call the notify macro here because, when using secondary
2683                  * mmu page tables (such as kvm shadow page tables), we want the
2684                  * new page to be mapped directly into the secondary page table.
2685                  */
2686                 set_pte_at_notify(mm, address, page_table, entry);
2687                 update_mmu_cache(vma, address, page_table);
2688                 if (old_page) {
2689                         /*
2690                          * Only after switching the pte to the new page may
2691                          * we remove the mapcount here. Otherwise another
2692                          * process may come and find the rmap count decremented
2693                          * before the pte is switched to the new page, and
2694                          * "reuse" the old page writing into it while our pte
2695                          * here still points into it and can be read by other
2696                          * threads.
2697                          *
2698                          * The critical issue is to order this
2699                          * page_remove_rmap with the ptp_clear_flush above.
2700                          * Those stores are ordered by (if nothing else,)
2701                          * the barrier present in the atomic_add_negative
2702                          * in page_remove_rmap.
2703                          *
2704                          * Then the TLB flush in ptep_clear_flush ensures that
2705                          * no process can access the old page before the
2706                          * decremented mapcount is visible. And the old page
2707                          * cannot be reused until after the decremented
2708                          * mapcount is visible. So transitively, TLBs to
2709                          * old page will be flushed before it can be reused.
2710                          */
2711                         page_remove_rmap(old_page);
2712                 }
2713
2714                 /* Free the old page.. */
2715                 new_page = old_page;
2716                 ret |= VM_FAULT_WRITE;
2717         } else
2718                 mem_cgroup_uncharge_page(new_page);
2719
2720         if (new_page)
2721                 page_cache_release(new_page);
2722 unlock:
2723         pte_unmap_unlock(page_table, ptl);
2724         if (old_page) {
2725                 /*
2726                  * Don't let another task, with possibly unlocked vma,
2727                  * keep the mlocked page.
2728                  */
2729                 if ((ret & VM_FAULT_WRITE) && (vma->vm_flags & VM_LOCKED)) {
2730                         lock_page(old_page);    /* LRU manipulation */
2731                         munlock_vma_page(old_page);
2732                         unlock_page(old_page);
2733                 }
2734                 page_cache_release(old_page);
2735         }
2736         return ret;
2737 oom_free_new:
2738         page_cache_release(new_page);
2739 oom:
2740         if (old_page) {
2741                 if (page_mkwrite) {
2742                         unlock_page(old_page);
2743                         page_cache_release(old_page);
2744                 }
2745                 page_cache_release(old_page);
2746         }
2747         return VM_FAULT_OOM;
2748
2749 unwritable_page:
2750         page_cache_release(old_page);
2751         return ret;
2752 }
2753
2754 static void unmap_mapping_range_vma(struct vm_area_struct *vma,
2755                 unsigned long start_addr, unsigned long end_addr,
2756                 struct zap_details *details)
2757 {
2758         zap_page_range(vma, start_addr, end_addr - start_addr, details);
2759 }
2760
2761 static inline void unmap_mapping_range_tree(struct prio_tree_root *root,
2762                                             struct zap_details *details)
2763 {
2764         struct vm_area_struct *vma;
2765         struct prio_tree_iter iter;
2766         pgoff_t vba, vea, zba, zea;
2767
2768         vma_prio_tree_foreach(vma, &iter, root,
2769                         details->first_index, details->last_index) {
2770
2771                 vba = vma->vm_pgoff;
2772                 vea = vba + ((vma->vm_end - vma->vm_start) >> PAGE_SHIFT) - 1;
2773                 /* Assume for now that PAGE_CACHE_SHIFT == PAGE_SHIFT */
2774                 zba = details->first_index;
2775                 if (zba < vba)
2776                         zba = vba;
2777                 zea = details->last_index;
2778                 if (zea > vea)
2779                         zea = vea;
2780
2781                 unmap_mapping_range_vma(vma,
2782                         ((zba - vba) << PAGE_SHIFT) + vma->vm_start,
2783                         ((zea - vba + 1) << PAGE_SHIFT) + vma->vm_start,
2784                                 details);
2785         }
2786 }
2787
2788 static inline void unmap_mapping_range_list(struct list_head *head,
2789                                             struct zap_details *details)
2790 {
2791         struct vm_area_struct *vma;
2792
2793         /*
2794          * In nonlinear VMAs there is no correspondence between virtual address
2795          * offset and file offset.  So we must perform an exhaustive search
2796          * across *all* the pages in each nonlinear VMA, not just the pages
2797          * whose virtual address lies outside the file truncation point.
2798          */
2799         list_for_each_entry(vma, head, shared.vm_set.list) {
2800                 details->nonlinear_vma = vma;
2801                 unmap_mapping_range_vma(vma, vma->vm_start, vma->vm_end, details);
2802         }
2803 }
2804
2805 /**
2806  * unmap_mapping_range - unmap the portion of all mmaps in the specified address_space corresponding to the specified page range in the underlying file.
2807  * @mapping: the address space containing mmaps to be unmapped.
2808  * @holebegin: byte in first page to unmap, relative to the start of
2809  * the underlying file.  This will be rounded down to a PAGE_SIZE
2810  * boundary.  Note that this is different from truncate_pagecache(), which
2811  * must keep the partial page.  In contrast, we must get rid of
2812  * partial pages.
2813  * @holelen: size of prospective hole in bytes.  This will be rounded
2814  * up to a PAGE_SIZE boundary.  A holelen of zero truncates to the
2815  * end of the file.
2816  * @even_cows: 1 when truncating a file, unmap even private COWed pages;
2817  * but 0 when invalidating pagecache, don't throw away private data.
2818  */
2819 void unmap_mapping_range(struct address_space *mapping,
2820                 loff_t const holebegin, loff_t const holelen, int even_cows)
2821 {
2822         struct zap_details details;
2823         pgoff_t hba = holebegin >> PAGE_SHIFT;
2824         pgoff_t hlen = (holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
2825
2826         /* Check for overflow. */
2827         if (sizeof(holelen) > sizeof(hlen)) {
2828                 long long holeend =
2829                         (holebegin + holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
2830                 if (holeend & ~(long long)ULONG_MAX)
2831                         hlen = ULONG_MAX - hba + 1;
2832         }
2833
2834         details.check_mapping = even_cows? NULL: mapping;
2835         details.nonlinear_vma = NULL;
2836         details.first_index = hba;
2837         details.last_index = hba + hlen - 1;
2838         if (details.last_index < details.first_index)
2839                 details.last_index = ULONG_MAX;
2840
2841
2842         mutex_lock(&mapping->i_mmap_mutex);
2843         if (unlikely(!prio_tree_empty(&mapping->i_mmap)))
2844                 unmap_mapping_range_tree(&mapping->i_mmap, &details);
2845         if (unlikely(!list_empty(&mapping->i_mmap_nonlinear)))
2846                 unmap_mapping_range_list(&mapping->i_mmap_nonlinear, &details);
2847         mutex_unlock(&mapping->i_mmap_mutex);
2848 }
2849 EXPORT_SYMBOL(unmap_mapping_range);
2850
2851 /*
2852  * We enter with non-exclusive mmap_sem (to exclude vma changes,
2853  * but allow concurrent faults), and pte mapped but not yet locked.
2854  * We return with mmap_sem still held, but pte unmapped and unlocked.
2855  */
2856 static int do_swap_page(struct mm_struct *mm, struct vm_area_struct *vma,
2857                 unsigned long address, pte_t *page_table, pmd_t *pmd,
2858                 unsigned int flags, pte_t orig_pte)
2859 {
2860         spinlock_t *ptl;
2861         struct page *page, *swapcache = NULL;
2862         swp_entry_t entry;
2863         pte_t pte;
2864         int locked;
2865         struct mem_cgroup *ptr;
2866         int exclusive = 0;
2867         int ret = 0;
2868
2869         if (!pte_unmap_same(mm, pmd, page_table, orig_pte))
2870                 goto out;
2871
2872         entry = pte_to_swp_entry(orig_pte);
2873         if (unlikely(non_swap_entry(entry))) {
2874                 if (is_migration_entry(entry)) {
2875                         migration_entry_wait(mm, pmd, address);
2876                 } else if (is_hwpoison_entry(entry)) {
2877                         ret = VM_FAULT_HWPOISON;
2878                 } else {
2879                         print_bad_pte(vma, address, orig_pte, NULL);
2880                         ret = VM_FAULT_SIGBUS;
2881                 }
2882                 goto out;
2883         }
2884         delayacct_set_flag(DELAYACCT_PF_SWAPIN);
2885         page = lookup_swap_cache(entry);
2886         if (!page) {
2887                 grab_swap_token(mm); /* Contend for token _before_ read-in */
2888                 page = swapin_readahead(entry,
2889                                         GFP_HIGHUSER_MOVABLE, vma, address);
2890                 if (!page) {
2891                         /*
2892                          * Back out if somebody else faulted in this pte
2893                          * while we released the pte lock.
2894                          */
2895                         page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2896                         if (likely(pte_same(*page_table, orig_pte)))
2897                                 ret = VM_FAULT_OOM;
2898                         delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
2899                         goto unlock;
2900                 }
2901
2902                 /* Had to read the page from swap area: Major fault */
2903                 ret = VM_FAULT_MAJOR;
2904                 count_vm_event(PGMAJFAULT);
2905                 mem_cgroup_count_vm_event(mm, PGMAJFAULT);
2906         } else if (PageHWPoison(page)) {
2907                 /*
2908                  * hwpoisoned dirty swapcache pages are kept for killing
2909                  * owner processes (which may be unknown at hwpoison time)
2910                  */
2911                 ret = VM_FAULT_HWPOISON;
2912                 delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
2913                 goto out_release;
2914         }
2915
2916         locked = lock_page_or_retry(page, mm, flags);
2917         delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
2918         if (!locked) {
2919                 ret |= VM_FAULT_RETRY;
2920                 goto out_release;
2921         }
2922
2923         /*
2924          * Make sure try_to_free_swap or reuse_swap_page or swapoff did not
2925          * release the swapcache from under us.  The page pin, and pte_same
2926          * test below, are not enough to exclude that.  Even if it is still
2927          * swapcache, we need to check that the page's swap has not changed.
2928          */
2929         if (unlikely(!PageSwapCache(page) || page_private(page) != entry.val))
2930                 goto out_page;
2931
2932         if (ksm_might_need_to_copy(page, vma, address)) {
2933                 swapcache = page;
2934                 page = ksm_does_need_to_copy(page, vma, address);
2935
2936                 if (unlikely(!page)) {
2937                         ret = VM_FAULT_OOM;
2938                         page = swapcache;
2939                         swapcache = NULL;
2940                         goto out_page;
2941                 }
2942         }
2943
2944         if (mem_cgroup_try_charge_swapin(mm, page, GFP_KERNEL, &ptr)) {
2945                 ret = VM_FAULT_OOM;
2946                 goto out_page;
2947         }
2948
2949         /*
2950          * Back out if somebody else already faulted in this pte.
2951          */
2952         page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2953         if (unlikely(!pte_same(*page_table, orig_pte)))
2954                 goto out_nomap;
2955
2956         if (unlikely(!PageUptodate(page))) {
2957                 ret = VM_FAULT_SIGBUS;
2958                 goto out_nomap;
2959         }
2960
2961         /*
2962          * The page isn't present yet, go ahead with the fault.
2963          *
2964          * Be careful about the sequence of operations here.
2965          * To get its accounting right, reuse_swap_page() must be called
2966          * while the page is counted on swap but not yet in mapcount i.e.
2967          * before page_add_anon_rmap() and swap_free(); try_to_free_swap()
2968          * must be called after the swap_free(), or it will never succeed.
2969          * Because delete_from_swap_page() may be called by reuse_swap_page(),
2970          * mem_cgroup_commit_charge_swapin() may not be able to find swp_entry
2971          * in page->private. In this case, a record in swap_cgroup  is silently
2972          * discarded at swap_free().
2973          */
2974
2975         inc_mm_counter_fast(mm, MM_ANONPAGES);
2976         dec_mm_counter_fast(mm, MM_SWAPENTS);
2977         pte = mk_pte(page, vma->vm_page_prot);
2978         if ((flags & FAULT_FLAG_WRITE) && reuse_swap_page(page)) {
2979                 pte = maybe_mkwrite(pte_mkdirty(pte), vma);
2980                 flags &= ~FAULT_FLAG_WRITE;
2981                 ret |= VM_FAULT_WRITE;
2982                 exclusive = 1;
2983         }
2984         flush_icache_page(vma, page);
2985         set_pte_at(mm, address, page_table, pte);
2986         do_page_add_anon_rmap(page, vma, address, exclusive);
2987         /* It's better to call commit-charge after rmap is established */
2988         mem_cgroup_commit_charge_swapin(page, ptr);
2989
2990         swap_free(entry);
2991         if (vm_swap_full() || (vma->vm_flags & VM_LOCKED) || PageMlocked(page))
2992                 try_to_free_swap(page);
2993         unlock_page(page);
2994         if (swapcache) {
2995                 /*
2996                  * Hold the lock to avoid the swap entry to be reused
2997                  * until we take the PT lock for the pte_same() check
2998                  * (to avoid false positives from pte_same). For
2999                  * further safety release the lock after the swap_free
3000                  * so that the swap count won't change under a
3001                  * parallel locked swapcache.
3002                  */
3003                 unlock_page(swapcache);
3004                 page_cache_release(swapcache);
3005         }
3006
3007         if (flags & FAULT_FLAG_WRITE) {
3008                 ret |= do_wp_page(mm, vma, address, page_table, pmd, ptl, pte);
3009                 if (ret & VM_FAULT_ERROR)
3010                         ret &= VM_FAULT_ERROR;
3011                 goto out;
3012         }
3013
3014         /* No need to invalidate - it was non-present before */
3015         update_mmu_cache(vma, address, page_table);
3016 unlock:
3017         pte_unmap_unlock(page_table, ptl);
3018 out:
3019         return ret;
3020 out_nomap:
3021         mem_cgroup_cancel_charge_swapin(ptr);
3022         pte_unmap_unlock(page_table, ptl);
3023 out_page:
3024         unlock_page(page);
3025 out_release:
3026         page_cache_release(page);
3027         if (swapcache) {
3028                 unlock_page(swapcache);
3029                 page_cache_release(swapcache);
3030         }
3031         return ret;
3032 }
3033
3034 /*
3035  * This is like a special single-page "expand_{down|up}wards()",
3036  * except we must first make sure that 'address{-|+}PAGE_SIZE'
3037  * doesn't hit another vma.
3038  */
3039 static inline int check_stack_guard_page(struct vm_area_struct *vma, unsigned long address)
3040 {
3041         address &= PAGE_MASK;
3042         if ((vma->vm_flags & VM_GROWSDOWN) && address == vma->vm_start) {
3043                 struct vm_area_struct *prev = vma->vm_prev;
3044
3045                 /*
3046                  * Is there a mapping abutting this one below?
3047                  *
3048                  * That's only ok if it's the same stack mapping
3049                  * that has gotten split..
3050                  */
3051                 if (prev && prev->vm_end == address)
3052                         return prev->vm_flags & VM_GROWSDOWN ? 0 : -ENOMEM;
3053
3054                 expand_downwards(vma, address - PAGE_SIZE);
3055         }
3056         if ((vma->vm_flags & VM_GROWSUP) && address + PAGE_SIZE == vma->vm_end) {
3057                 struct vm_area_struct *next = vma->vm_next;
3058
3059                 /* As VM_GROWSDOWN but s/below/above/ */
3060                 if (next && next->vm_start == address + PAGE_SIZE)
3061                         return next->vm_flags & VM_GROWSUP ? 0 : -ENOMEM;
3062
3063                 expand_upwards(vma, address + PAGE_SIZE);
3064         }
3065         return 0;
3066 }
3067
3068 /*
3069  * We enter with non-exclusive mmap_sem (to exclude vma changes,
3070  * but allow concurrent faults), and pte mapped but not yet locked.
3071  * We return with mmap_sem still held, but pte unmapped and unlocked.
3072  */
3073 static int do_anonymous_page(struct mm_struct *mm, struct vm_area_struct *vma,
3074                 unsigned long address, pte_t *page_table, pmd_t *pmd,
3075                 unsigned int flags)
3076 {
3077         struct page *page;
3078         spinlock_t *ptl;
3079         pte_t entry;
3080
3081         pte_unmap(page_table);
3082
3083         /* Check if we need to add a guard page to the stack */
3084         if (check_stack_guard_page(vma, address) < 0)
3085                 return VM_FAULT_SIGBUS;
3086
3087         /* Use the zero-page for reads */
3088         if (!(flags & FAULT_FLAG_WRITE)) {
3089                 entry = pte_mkspecial(pfn_pte(my_zero_pfn(address),
3090                                                 vma->vm_page_prot));
3091                 page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
3092                 if (!pte_none(*page_table))
3093                         goto unlock;
3094                 goto setpte;
3095         }
3096
3097         /* Allocate our own private page. */
3098         if (unlikely(anon_vma_prepare(vma)))
3099                 goto oom;
3100         page = alloc_zeroed_user_highpage_movable(vma, address);
3101         if (!page)
3102                 goto oom;
3103         __SetPageUptodate(page);
3104
3105         if (mem_cgroup_newpage_charge(page, mm, GFP_KERNEL))
3106                 goto oom_free_page;
3107
3108         entry = mk_pte(page, vma->vm_page_prot);
3109         if (vma->vm_flags & VM_WRITE)
3110                 entry = pte_mkwrite(pte_mkdirty(entry));
3111
3112         page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
3113         if (!pte_none(*page_table))
3114                 goto release;
3115
3116         inc_mm_counter_fast(mm, MM_ANONPAGES);
3117         page_add_new_anon_rmap(page, vma, address);
3118 setpte:
3119         set_pte_at(mm, address, page_table, entry);
3120
3121         /* No need to invalidate - it was non-present before */
3122         update_mmu_cache(vma, address, page_table);
3123 unlock:
3124         pte_unmap_unlock(page_table, ptl);
3125         return 0;
3126 release:
3127         mem_cgroup_uncharge_page(page);
3128         page_cache_release(page);
3129         goto unlock;
3130 oom_free_page:
3131         page_cache_release(page);
3132 oom:
3133         return VM_FAULT_OOM;
3134 }
3135
3136 /*
3137  * __do_fault() tries to create a new page mapping. It aggressively
3138  * tries to share with existing pages, but makes a separate copy if
3139  * the FAULT_FLAG_WRITE is set in the flags parameter in order to avoid
3140  * the next page fault.
3141  *
3142  * As this is called only for pages that do not currently exist, we
3143  * do not need to flush old virtual caches or the TLB.
3144  *
3145  * We enter with non-exclusive mmap_sem (to exclude vma changes,
3146  * but allow concurrent faults), and pte neither mapped nor locked.
3147  * We return with mmap_sem still held, but pte unmapped and unlocked.
3148  */
3149 static int __do_fault(struct mm_struct *mm, struct vm_area_struct *vma,
3150                 unsigned long address, pmd_t *pmd,
3151                 pgoff_t pgoff, unsigned int flags, pte_t orig_pte)
3152 {
3153         pte_t *page_table;
3154         spinlock_t *ptl;
3155         struct page *page;
3156         struct page *cow_page;
3157         pte_t entry;
3158         int anon = 0;
3159         struct page *dirty_page = NULL;
3160         struct vm_fault vmf;
3161         int ret;
3162         int page_mkwrite = 0;
3163
3164         /*
3165          * If we do COW later, allocate page befor taking lock_page()
3166          * on the file cache page. This will reduce lock holding time.
3167          */
3168         if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) {
3169
3170                 if (unlikely(anon_vma_prepare(vma)))
3171                         return VM_FAULT_OOM;
3172
3173                 cow_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, address);
3174                 if (!cow_page)
3175                         return VM_FAULT_OOM;
3176
3177                 if (mem_cgroup_newpage_charge(cow_page, mm, GFP_KERNEL)) {
3178                         page_cache_release(cow_page);
3179                         return VM_FAULT_OOM;
3180                 }
3181         } else
3182                 cow_page = NULL;
3183
3184         vmf.virtual_address = (void __user *)(address & PAGE_MASK);
3185         vmf.pgoff = pgoff;
3186         vmf.flags = flags;
3187         vmf.page = NULL;
3188
3189         ret = vma->vm_ops->fault(vma, &vmf);
3190         if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE |
3191                             VM_FAULT_RETRY)))
3192                 goto uncharge_out;
3193
3194         if (unlikely(PageHWPoison(vmf.page))) {
3195                 if (ret & VM_FAULT_LOCKED)
3196                         unlock_page(vmf.page);
3197                 ret = VM_FAULT_HWPOISON;
3198                 goto uncharge_out;
3199         }
3200
3201         /*
3202          * For consistency in subsequent calls, make the faulted page always
3203          * locked.
3204          */
3205         if (unlikely(!(ret & VM_FAULT_LOCKED)))
3206                 lock_page(vmf.page);
3207         else
3208                 VM_BUG_ON(!PageLocked(vmf.page));
3209
3210         /*
3211          * Should we do an early C-O-W break?
3212          */
3213         page = vmf.page;
3214         if (flags & FAULT_FLAG_WRITE) {
3215                 if (!(vma->vm_flags & VM_SHARED)) {
3216                         page = cow_page;
3217                         anon = 1;
3218                         copy_user_highpage(page, vmf.page, address, vma);
3219                         __SetPageUptodate(page);
3220                 } else {
3221                         /*
3222                          * If the page will be shareable, see if the backing
3223                          * address space wants to know that the page is about
3224                          * to become writable
3225                          */
3226                         if (vma->vm_ops->page_mkwrite) {
3227                                 int tmp;
3228
3229                                 unlock_page(page);
3230                                 vmf.flags = FAULT_FLAG_WRITE|FAULT_FLAG_MKWRITE;
3231                                 tmp = vma->vm_ops->page_mkwrite(vma, &vmf);
3232                                 if (unlikely(tmp &
3233                                           (VM_FAULT_ERROR | VM_FAULT_NOPAGE))) {
3234                                         ret = tmp;
3235                                         goto unwritable_page;
3236                                 }
3237                                 if (unlikely(!(tmp & VM_FAULT_LOCKED))) {
3238                                         lock_page(page);
3239                                         if (!page->mapping) {
3240                                                 ret = 0; /* retry the fault */
3241                                                 unlock_page(page);
3242                                                 goto unwritable_page;
3243                                         }
3244                                 } else
3245                                         VM_BUG_ON(!PageLocked(page));
3246                                 page_mkwrite = 1;
3247                         }
3248                 }
3249
3250         }
3251
3252         page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
3253
3254         /*
3255          * This silly early PAGE_DIRTY setting removes a race
3256          * due to the bad i386 page protection. But it's valid
3257          * for other architectures too.
3258          *
3259          * Note that if FAULT_FLAG_WRITE is set, we either now have
3260          * an exclusive copy of the page, or this is a shared mapping,
3261          * so we can make it writable and dirty to avoid having to
3262          * handle that later.
3263          */
3264         /* Only go through if we didn't race with anybody else... */
3265         if (likely(pte_same(*page_table, orig_pte))) {
3266                 flush_icache_page(vma, page);
3267                 entry = mk_pte(page, vma->vm_page_prot);
3268                 if (flags & FAULT_FLAG_WRITE)
3269                         entry = maybe_mkwrite(pte_mkdirty(entry), vma);
3270                 if (anon) {
3271                         inc_mm_counter_fast(mm, MM_ANONPAGES);
3272                         page_add_new_anon_rmap(page, vma, address);
3273                 } else {
3274                         inc_mm_counter_fast(mm, MM_FILEPAGES);
3275                         page_add_file_rmap(page);
3276                         if (flags & FAULT_FLAG_WRITE) {
3277                                 dirty_page = page;
3278                                 get_page(dirty_page);
3279                         }
3280                 }
3281                 set_pte_at(mm, address, page_table, entry);
3282
3283                 /* no need to invalidate: a not-present page won't be cached */
3284                 update_mmu_cache(vma, address, page_table);
3285         } else {
3286                 if (cow_page)
3287                         mem_cgroup_uncharge_page(cow_page);
3288                 if (anon)
3289                         page_cache_release(page);
3290                 else
3291                         anon = 1; /* no anon but release faulted_page */
3292         }
3293
3294         pte_unmap_unlock(page_table, ptl);
3295
3296         if (dirty_page) {
3297                 struct address_space *mapping = page->mapping;
3298
3299                 if (set_page_dirty(dirty_page))
3300                         page_mkwrite = 1;
3301                 unlock_page(dirty_page);
3302                 put_page(dirty_page);
3303                 if (page_mkwrite && mapping) {
3304                         /*
3305                          * Some device drivers do not set page.mapping but still
3306                          * dirty their pages
3307                          */
3308                         balance_dirty_pages_ratelimited(mapping);
3309                 }
3310
3311                 /* file_update_time outside page_lock */
3312                 if (vma->vm_file)
3313                         file_update_time(vma->vm_file);
3314         } else {
3315                 unlock_page(vmf.page);
3316                 if (anon)
3317                         page_cache_release(vmf.page);
3318         }
3319
3320         return ret;
3321
3322 unwritable_page:
3323         page_cache_release(page);
3324         return ret;
3325 uncharge_out:
3326         /* fs's fault handler get error */
3327         if (cow_page) {
3328                 mem_cgroup_uncharge_page(cow_page);
3329                 page_cache_release(cow_page);
3330         }
3331         return ret;
3332 }
3333
3334 static int do_linear_fault(struct mm_struct *mm, struct vm_area_struct *vma,
3335                 unsigned long address, pte_t *page_table, pmd_t *pmd,
3336                 unsigned int flags, pte_t orig_pte)
3337 {
3338         pgoff_t pgoff = (((address & PAGE_MASK)
3339                         - vma->vm_start) >> PAGE_SHIFT) + vma->vm_pgoff;
3340
3341         pte_unmap(page_table);
3342         return __do_fault(mm, vma, address, pmd, pgoff, flags, orig_pte);
3343 }
3344
3345 /*
3346  * Fault of a previously existing named mapping. Repopulate the pte
3347  * from the encoded file_pte if possible. This enables swappable
3348  * nonlinear vmas.
3349  *
3350  * We enter with non-exclusive mmap_sem (to exclude vma changes,
3351  * but allow concurrent faults), and pte mapped but not yet locked.
3352  * We return with mmap_sem still held, but pte unmapped and unlocked.
3353  */
3354 static int do_nonlinear_fault(struct mm_struct *mm, struct vm_area_struct *vma,
3355                 unsigned long address, pte_t *page_table, pmd_t *pmd,
3356                 unsigned int flags, pte_t orig_pte)
3357 {
3358         pgoff_t pgoff;
3359
3360         flags |= FAULT_FLAG_NONLINEAR;
3361
3362         if (!pte_unmap_same(mm, pmd, page_table, orig_pte))
3363                 return 0;
3364
3365         if (unlikely(!(vma->vm_flags & VM_NONLINEAR))) {
3366                 /*
3367                  * Page table corrupted: show pte and kill process.
3368                  */
3369                 print_bad_pte(vma, address, orig_pte, NULL);
3370                 return VM_FAULT_SIGBUS;
3371         }
3372
3373         pgoff = pte_to_pgoff(orig_pte);
3374         return __do_fault(mm, vma, address, pmd, pgoff, flags, orig_pte);
3375 }
3376
3377 /*
3378  * These routines also need to handle stuff like marking pages dirty
3379  * and/or accessed for architectures that don't do it in hardware (most
3380  * RISC architectures).  The early dirtying is also good on the i386.
3381  *
3382  * There is also a hook called "update_mmu_cache()" that architectures
3383  * with external mmu caches can use to update those (ie the Sparc or
3384  * PowerPC hashed page tables that act as extended TLBs).
3385  *
3386  * We enter with non-exclusive mmap_sem (to exclude vma changes,
3387  * but allow concurrent faults), and pte mapped but not yet locked.
3388  * We return with mmap_sem still held, but pte unmapped and unlocked.
3389  */
3390 int handle_pte_fault(struct mm_struct *mm,
3391                      struct vm_area_struct *vma, unsigned long address,
3392                      pte_t *pte, pmd_t *pmd, unsigned int flags)
3393 {
3394         pte_t entry;
3395         spinlock_t *ptl;
3396
3397         entry = *pte;
3398         if (!pte_present(entry)) {
3399                 if (pte_none(entry)) {
3400                         if (vma->vm_ops) {
3401                                 if (likely(vma->vm_ops->fault))
3402                                         return do_linear_fault(mm, vma, address,
3403                                                 pte, pmd, flags, entry);
3404                         }
3405                         return do_anonymous_page(mm, vma, address,
3406                                                  pte, pmd, flags);
3407                 }
3408                 if (pte_file(entry))
3409                         return do_nonlinear_fault(mm, vma, address,
3410                                         pte, pmd, flags, entry);
3411                 return do_swap_page(mm, vma, address,
3412                                         pte, pmd, flags, entry);
3413         }
3414
3415         ptl = pte_lockptr(mm, pmd);
3416         spin_lock(ptl);
3417         if (unlikely(!pte_same(*pte, entry)))
3418                 goto unlock;
3419         if (flags & FAULT_FLAG_WRITE) {
3420                 if (!pte_write(entry))
3421                         return do_wp_page(mm, vma, address,
3422                                         pte, pmd, ptl, entry);
3423                 entry = pte_mkdirty(entry);
3424         }
3425         entry = pte_mkyoung(entry);
3426         if (ptep_set_access_flags(vma, address, pte, entry, flags & FAULT_FLAG_WRITE)) {
3427                 update_mmu_cache(vma, address, pte);
3428         } else {
3429                 /*
3430                  * This is needed only for protection faults but the arch code
3431                  * is not yet telling us if this is a protection fault or not.
3432                  * This still avoids useless tlb flushes for .text page faults
3433                  * with threads.
3434                  */
3435                 if (flags & FAULT_FLAG_WRITE)
3436                         flush_tlb_fix_spurious_fault(vma, address);
3437         }
3438 unlock:
3439         pte_unmap_unlock(pte, ptl);
3440         return 0;
3441 }
3442
3443 /*
3444  * By the time we get here, we already hold the mm semaphore
3445  */
3446 int handle_mm_fault(struct mm_struct *mm, struct vm_area_struct *vma,
3447                 unsigned long address, unsigned int flags)
3448 {
3449         pgd_t *pgd;
3450         pud_t *pud;
3451         pmd_t *pmd;
3452         pte_t *pte;
3453
3454         __set_current_state(TASK_RUNNING);
3455
3456         count_vm_event(PGFAULT);
3457         mem_cgroup_count_vm_event(mm, PGFAULT);
3458
3459         /* do counter updates before entering really critical section. */
3460         check_sync_rss_stat(current);
3461
3462         if (unlikely(is_vm_hugetlb_page(vma)))
3463                 return hugetlb_fault(mm, vma, address, flags);
3464
3465         pgd = pgd_offset(mm, address);
3466         pud = pud_alloc(mm, pgd, address);
3467         if (!pud)
3468                 return VM_FAULT_OOM;
3469         pmd = pmd_alloc(mm, pud, address);
3470         if (!pmd)
3471                 return VM_FAULT_OOM;
3472         if (pmd_none(*pmd) && transparent_hugepage_enabled(vma)) {
3473                 if (!vma->vm_ops)
3474                         return do_huge_pmd_anonymous_page(mm, vma, address,
3475                                                           pmd, flags);
3476         } else {
3477                 pmd_t orig_pmd = *pmd;
3478                 barrier();
3479                 if (pmd_trans_huge(orig_pmd)) {
3480                         if (flags & FAULT_FLAG_WRITE &&
3481                             !pmd_write(orig_pmd) &&
3482                             !pmd_trans_splitting(orig_pmd))
3483                                 return do_huge_pmd_wp_page(mm, vma, address,
3484                                                            pmd, orig_pmd);
3485                         return 0;
3486                 }
3487         }
3488
3489         /*
3490          * Use __pte_alloc instead of pte_alloc_map, because we can't
3491          * run pte_offset_map on the pmd, if an huge pmd could
3492          * materialize from under us from a different thread.
3493          */
3494         if (unlikely(pmd_none(*pmd)) && __pte_alloc(mm, vma, pmd, address))
3495                 return VM_FAULT_OOM;
3496         /* if an huge pmd materialized from under us just retry later */
3497         if (unlikely(pmd_trans_huge(*pmd)))
3498                 return 0;
3499         /*
3500          * A regular pmd is established and it can't morph into a huge pmd
3501          * from under us anymore at this point because we hold the mmap_sem
3502          * read mode and khugepaged takes it in write mode. So now it's
3503          * safe to run pte_offset_map().
3504          */
3505         pte = pte_offset_map(pmd, address);
3506
3507         return handle_pte_fault(mm, vma, address, pte, pmd, flags);
3508 }
3509
3510 #ifndef __PAGETABLE_PUD_FOLDED
3511 /*
3512  * Allocate page upper directory.
3513  * We've already handled the fast-path in-line.
3514  */
3515 int __pud_alloc(struct mm_struct *mm, pgd_t *pgd, unsigned long address)
3516 {
3517         pud_t *new = pud_alloc_one(mm, address);
3518         if (!new)
3519                 return -ENOMEM;
3520
3521         smp_wmb(); /* See comment in __pte_alloc */
3522
3523         spin_lock(&mm->page_table_lock);
3524         if (pgd_present(*pgd))          /* Another has populated it */
3525                 pud_free(mm, new);
3526         else
3527                 pgd_populate(mm, pgd, new);
3528         spin_unlock(&mm->page_table_lock);
3529         return 0;
3530 }
3531 #endif /* __PAGETABLE_PUD_FOLDED */
3532
3533 #ifndef __PAGETABLE_PMD_FOLDED
3534 /*
3535  * Allocate page middle directory.
3536  * We've already handled the fast-path in-line.
3537  */
3538 int __pmd_alloc(struct mm_struct *mm, pud_t *pud, unsigned long address)
3539 {
3540         pmd_t *new = pmd_alloc_one(mm, address);
3541         if (!new)
3542                 return -ENOMEM;
3543
3544         smp_wmb(); /* See comment in __pte_alloc */
3545
3546         spin_lock(&mm->page_table_lock);
3547 #ifndef __ARCH_HAS_4LEVEL_HACK
3548         if (pud_present(*pud))          /* Another has populated it */
3549                 pmd_free(mm, new);
3550         else
3551                 pud_populate(mm, pud, new);
3552 #else
3553         if (pgd_present(*pud))          /* Another has populated it */
3554                 pmd_free(mm, new);
3555         else
3556                 pgd_populate(mm, pud, new);
3557 #endif /* __ARCH_HAS_4LEVEL_HACK */
3558         spin_unlock(&mm->page_table_lock);
3559         return 0;
3560 }
3561 #endif /* __PAGETABLE_PMD_FOLDED */
3562
3563 int make_pages_present(unsigned long addr, unsigned long end)
3564 {
3565         int ret, len, write;
3566         struct vm_area_struct * vma;
3567
3568         vma = find_vma(current->mm, addr);
3569         if (!vma)
3570                 return -ENOMEM;
3571         /*
3572          * We want to touch writable mappings with a write fault in order
3573          * to break COW, except for shared mappings because these don't COW
3574          * and we would not want to dirty them for nothing.
3575          */
3576         write = (vma->vm_flags & (VM_WRITE | VM_SHARED)) == VM_WRITE;
3577         BUG_ON(addr >= end);
3578         BUG_ON(end > vma->vm_end);
3579         len = DIV_ROUND_UP(end, PAGE_SIZE) - addr/PAGE_SIZE;
3580         ret = get_user_pages(current, current->mm, addr,
3581                         len, write, 0, NULL, NULL);
3582         if (ret < 0)
3583                 return ret;
3584         return ret == len ? 0 : -EFAULT;
3585 }
3586
3587 #if !defined(__HAVE_ARCH_GATE_AREA)
3588
3589 #if defined(AT_SYSINFO_EHDR)
3590 static struct vm_area_struct gate_vma;
3591
3592 static int __init gate_vma_init(void)
3593 {
3594         gate_vma.vm_mm = NULL;
3595         gate_vma.vm_start = FIXADDR_USER_START;
3596         gate_vma.vm_end = FIXADDR_USER_END;
3597         gate_vma.vm_flags = VM_READ | VM_MAYREAD | VM_EXEC | VM_MAYEXEC;
3598         gate_vma.vm_page_prot = __P101;
3599         /*
3600          * Make sure the vDSO gets into every core dump.
3601          * Dumping its contents makes post-mortem fully interpretable later
3602          * without matching up the same kernel and hardware config to see
3603          * what PC values meant.
3604          */
3605         gate_vma.vm_flags |= VM_ALWAYSDUMP;
3606         return 0;
3607 }
3608 __initcall(gate_vma_init);
3609 #endif
3610
3611 struct vm_area_struct *get_gate_vma(struct mm_struct *mm)
3612 {
3613 #ifdef AT_SYSINFO_EHDR
3614         return &gate_vma;
3615 #else
3616         return NULL;
3617 #endif
3618 }
3619
3620 int in_gate_area_no_mm(unsigned long addr)
3621 {
3622 #ifdef AT_SYSINFO_EHDR
3623         if ((addr >= FIXADDR_USER_START) && (addr < FIXADDR_USER_END))
3624                 return 1;
3625 #endif
3626         return 0;
3627 }
3628
3629 #endif  /* __HAVE_ARCH_GATE_AREA */
3630
3631 static int __follow_pte(struct mm_struct *mm, unsigned long address,
3632                 pte_t **ptepp, spinlock_t **ptlp)
3633 {
3634         pgd_t *pgd;
3635         pud_t *pud;
3636         pmd_t *pmd;
3637         pte_t *ptep;
3638
3639         pgd = pgd_offset(mm, address);
3640         if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
3641                 goto out;
3642
3643         pud = pud_offset(pgd, address);
3644         if (pud_none(*pud) || unlikely(pud_bad(*pud)))
3645                 goto out;
3646
3647         pmd = pmd_offset(pud, address);
3648         VM_BUG_ON(pmd_trans_huge(*pmd));
3649         if (pmd_none(*pmd) || unlikely(pmd_bad(*pmd)))
3650                 goto out;
3651
3652         /* We cannot handle huge page PFN maps. Luckily they don't exist. */
3653         if (pmd_huge(*pmd))
3654                 goto out;
3655
3656         ptep = pte_offset_map_lock(mm, pmd, address, ptlp);
3657         if (!ptep)
3658                 goto out;
3659         if (!pte_present(*ptep))
3660                 goto unlock;
3661         *ptepp = ptep;
3662         return 0;
3663 unlock:
3664         pte_unmap_unlock(ptep, *ptlp);
3665 out:
3666         return -EINVAL;
3667 }
3668
3669 static inline int follow_pte(struct mm_struct *mm, unsigned long address,
3670                              pte_t **ptepp, spinlock_t **ptlp)
3671 {
3672         int res;
3673
3674         /* (void) is needed to make gcc happy */
3675         (void) __cond_lock(*ptlp,
3676                            !(res = __follow_pte(mm, address, ptepp, ptlp)));
3677         return res;
3678 }
3679
3680 /**
3681  * follow_pfn - look up PFN at a user virtual address
3682  * @vma: memory mapping
3683  * @address: user virtual address
3684  * @pfn: location to store found PFN
3685  *
3686  * Only IO mappings and raw PFN mappings are allowed.
3687  *
3688  * Returns zero and the pfn at @pfn on success, -ve otherwise.
3689  */
3690 int follow_pfn(struct vm_area_struct *vma, unsigned long address,
3691         unsigned long *pfn)
3692 {
3693         int ret = -EINVAL;
3694         spinlock_t *ptl;
3695         pte_t *ptep;
3696
3697         if (!(vma->vm_flags & (VM_IO | VM_PFNMAP)))
3698                 return ret;
3699
3700         ret = follow_pte(vma->vm_mm, address, &ptep, &ptl);
3701         if (ret)
3702                 return ret;
3703         *pfn = pte_pfn(*ptep);
3704         pte_unmap_unlock(ptep, ptl);
3705         return 0;
3706 }
3707 EXPORT_SYMBOL(follow_pfn);
3708
3709 #ifdef CONFIG_HAVE_IOREMAP_PROT
3710 int follow_phys(struct vm_area_struct *vma,
3711                 unsigned long address, unsigned int flags,
3712                 unsigned long *prot, resource_size_t *phys)
3713 {
3714         int ret = -EINVAL;
3715         pte_t *ptep, pte;
3716         spinlock_t *ptl;
3717
3718         if (!(vma->vm_flags & (VM_IO | VM_PFNMAP)))
3719                 goto out;
3720
3721         if (follow_pte(vma->vm_mm, address, &ptep, &ptl))
3722                 goto out;
3723         pte = *ptep;
3724
3725         if ((flags & FOLL_WRITE) && !pte_write(pte))
3726                 goto unlock;
3727
3728         *prot = pgprot_val(pte_pgprot(pte));
3729         *phys = (resource_size_t)pte_pfn(pte) << PAGE_SHIFT;
3730
3731         ret = 0;
3732 unlock:
3733         pte_unmap_unlock(ptep, ptl);
3734 out:
3735         return ret;
3736 }
3737
3738 int generic_access_phys(struct vm_area_struct *vma, unsigned long addr,
3739                         void *buf, int len, int write)
3740 {
3741         resource_size_t phys_addr;
3742         unsigned long prot = 0;
3743         void __iomem *maddr;
3744         int offset = addr & (PAGE_SIZE-1);
3745
3746         if (follow_phys(vma, addr, write, &prot, &phys_addr))
3747                 return -EINVAL;
3748
3749         maddr = ioremap_prot(phys_addr, PAGE_SIZE, prot);
3750         if (write)
3751                 memcpy_toio(maddr + offset, buf, len);
3752         else
3753                 memcpy_fromio(buf, maddr + offset, len);
3754         iounmap(maddr);
3755
3756         return len;
3757 }
3758 #endif
3759
3760 /*
3761  * Access another process' address space as given in mm.  If non-NULL, use the
3762  * given task for page fault accounting.
3763  */
3764 static int __access_remote_vm(struct task_struct *tsk, struct mm_struct *mm,
3765                 unsigned long addr, void *buf, int len, int write)
3766 {
3767         struct vm_area_struct *vma;
3768         void *old_buf = buf;
3769
3770         down_read(&mm->mmap_sem);
3771         /* ignore errors, just check how much was successfully transferred */
3772         while (len) {
3773                 int bytes, ret, offset;
3774                 void *maddr;
3775                 struct page *page = NULL;
3776
3777                 ret = get_user_pages(tsk, mm, addr, 1,
3778                                 write, 1, &page, &vma);
3779                 if (ret <= 0) {
3780                         /*
3781                          * Check if this is a VM_IO | VM_PFNMAP VMA, which
3782                          * we can access using slightly different code.
3783                          */
3784 #ifdef CONFIG_HAVE_IOREMAP_PROT
3785                         vma = find_vma(mm, addr);
3786                         if (!vma || vma->vm_start > addr)
3787                                 break;
3788                         if (vma->vm_ops && vma->vm_ops->access)
3789                                 ret = vma->vm_ops->access(vma, addr, buf,
3790                                                           len, write);
3791                         if (ret <= 0)
3792 #endif
3793                                 break;
3794                         bytes = ret;
3795                 } else {
3796                         bytes = len;
3797                         offset = addr & (PAGE_SIZE-1);
3798                         if (bytes > PAGE_SIZE-offset)
3799                                 bytes = PAGE_SIZE-offset;
3800
3801                         maddr = kmap(page);
3802                         if (write) {
3803                                 copy_to_user_page(vma, page, addr,
3804                                                   maddr + offset, buf, bytes);
3805                                 set_page_dirty_lock(page);
3806                         } else {
3807                                 copy_from_user_page(vma, page, addr,
3808                                                     buf, maddr + offset, bytes);
3809                         }
3810                         kunmap(page);
3811                         page_cache_release(page);
3812                 }
3813                 len -= bytes;
3814                 buf += bytes;
3815                 addr += bytes;
3816         }
3817         up_read(&mm->mmap_sem);
3818
3819         return buf - old_buf;
3820 }
3821
3822 /**
3823  * access_remote_vm - access another process' address space
3824  * @mm:         the mm_struct of the target address space
3825  * @addr:       start address to access
3826  * @buf:        source or destination buffer
3827  * @len:        number of bytes to transfer
3828  * @write:      whether the access is a write
3829  *
3830  * The caller must hold a reference on @mm.
3831  */
3832 int access_remote_vm(struct mm_struct *mm, unsigned long addr,
3833                 void *buf, int len, int write)
3834 {
3835         return __access_remote_vm(NULL, mm, addr, buf, len, write);
3836 }
3837
3838 /*
3839  * Access another process' address space.
3840  * Source/target buffer must be kernel space,
3841  * Do not walk the page table directly, use get_user_pages
3842  */
3843 int access_process_vm(struct task_struct *tsk, unsigned long addr,
3844                 void *buf, int len, int write)
3845 {
3846         struct mm_struct *mm;
3847         int ret;
3848
3849         mm = get_task_mm(tsk);
3850         if (!mm)
3851                 return 0;
3852
3853         ret = __access_remote_vm(tsk, mm, addr, buf, len, write);
3854         mmput(mm);
3855
3856         return ret;
3857 }
3858
3859 /*
3860  * Print the name of a VMA.
3861  */
3862 void print_vma_addr(char *prefix, unsigned long ip)
3863 {
3864         struct mm_struct *mm = current->mm;
3865         struct vm_area_struct *vma;
3866
3867         /*
3868          * Do not print if we are in atomic
3869          * contexts (in exception stacks, etc.):
3870          */
3871         if (preempt_count())
3872                 return;
3873
3874         down_read(&mm->mmap_sem);
3875         vma = find_vma(mm, ip);
3876         if (vma && vma->vm_file) {
3877                 struct file *f = vma->vm_file;
3878                 char *buf = (char *)__get_free_page(GFP_KERNEL);
3879                 if (buf) {
3880                         char *p, *s;
3881
3882                         p = d_path(&f->f_path, buf, PAGE_SIZE);
3883                         if (IS_ERR(p))
3884                                 p = "?";
3885                         s = strrchr(p, '/');
3886                         if (s)
3887                                 p = s+1;
3888                         printk("%s%s[%lx+%lx]", prefix, p,
3889                                         vma->vm_start,
3890                                         vma->vm_end - vma->vm_start);
3891                         free_page((unsigned long)buf);
3892                 }
3893         }
3894         up_read(&current->mm->mmap_sem);
3895 }
3896
3897 #ifdef CONFIG_PROVE_LOCKING
3898 void might_fault(void)
3899 {
3900         /*
3901          * Some code (nfs/sunrpc) uses socket ops on kernel memory while
3902          * holding the mmap_sem, this is safe because kernel memory doesn't
3903          * get paged out, therefore we'll never actually fault, and the
3904          * below annotations will generate false positives.
3905          */
3906         if (segment_eq(get_fs(), KERNEL_DS))
3907                 return;
3908
3909         might_sleep();
3910         /*
3911          * it would be nicer only to annotate paths which are not under
3912          * pagefault_disable, however that requires a larger audit and
3913          * providing helpers like get_user_atomic.
3914          */
3915         if (!in_atomic() && current->mm)
3916                 might_lock_read(&current->mm->mmap_sem);
3917 }
3918 EXPORT_SYMBOL(might_fault);
3919 #endif
3920
3921 #if defined(CONFIG_TRANSPARENT_HUGEPAGE) || defined(CONFIG_HUGETLBFS)
3922 static void clear_gigantic_page(struct page *page,
3923                                 unsigned long addr,
3924                                 unsigned int pages_per_huge_page)
3925 {
3926         int i;
3927         struct page *p = page;
3928
3929         might_sleep();
3930         for (i = 0; i < pages_per_huge_page;
3931              i++, p = mem_map_next(p, page, i)) {
3932                 cond_resched();
3933                 clear_user_highpage(p, addr + i * PAGE_SIZE);
3934         }
3935 }
3936 void clear_huge_page(struct page *page,
3937                      unsigned long addr, unsigned int pages_per_huge_page)
3938 {
3939         int i;
3940
3941         if (unlikely(pages_per_huge_page > MAX_ORDER_NR_PAGES)) {
3942                 clear_gigantic_page(page, addr, pages_per_huge_page);
3943                 return;
3944         }
3945
3946         might_sleep();
3947         for (i = 0; i < pages_per_huge_page; i++) {
3948                 cond_resched();
3949                 clear_user_highpage(page + i, addr + i * PAGE_SIZE);
3950         }
3951 }
3952
3953 static void copy_user_gigantic_page(struct page *dst, struct page *src,
3954                                     unsigned long addr,
3955                                     struct vm_area_struct *vma,
3956                                     unsigned int pages_per_huge_page)
3957 {
3958         int i;
3959         struct page *dst_base = dst;
3960         struct page *src_base = src;
3961
3962         for (i = 0; i < pages_per_huge_page; ) {
3963                 cond_resched();
3964                 copy_user_highpage(dst, src, addr + i*PAGE_SIZE, vma);
3965
3966                 i++;
3967                 dst = mem_map_next(dst, dst_base, i);
3968                 src = mem_map_next(src, src_base, i);
3969         }
3970 }
3971
3972 void copy_user_huge_page(struct page *dst, struct page *src,
3973                          unsigned long addr, struct vm_area_struct *vma,
3974                          unsigned int pages_per_huge_page)
3975 {
3976         int i;
3977
3978         if (unlikely(pages_per_huge_page > MAX_ORDER_NR_PAGES)) {
3979                 copy_user_gigantic_page(dst, src, addr, vma,
3980                                         pages_per_huge_page);
3981                 return;
3982         }
3983
3984         might_sleep();
3985         for (i = 0; i < pages_per_huge_page; i++) {
3986                 cond_resched();
3987                 copy_user_highpage(dst + i, src + i, addr + i*PAGE_SIZE, vma);
3988         }
3989 }
3990 #endif /* CONFIG_TRANSPARENT_HUGEPAGE || CONFIG_HUGETLBFS */