6105f475fa8633edf5180792b2cf0c5288734f08
[linux-2.6.git] / mm / memory.c
1 /*
2  *  linux/mm/memory.c
3  *
4  *  Copyright (C) 1991, 1992, 1993, 1994  Linus Torvalds
5  */
6
7 /*
8  * demand-loading started 01.12.91 - seems it is high on the list of
9  * things wanted, and it should be easy to implement. - Linus
10  */
11
12 /*
13  * Ok, demand-loading was easy, shared pages a little bit tricker. Shared
14  * pages started 02.12.91, seems to work. - Linus.
15  *
16  * Tested sharing by executing about 30 /bin/sh: under the old kernel it
17  * would have taken more than the 6M I have free, but it worked well as
18  * far as I could see.
19  *
20  * Also corrected some "invalidate()"s - I wasn't doing enough of them.
21  */
22
23 /*
24  * Real VM (paging to/from disk) started 18.12.91. Much more work and
25  * thought has to go into this. Oh, well..
26  * 19.12.91  -  works, somewhat. Sometimes I get faults, don't know why.
27  *              Found it. Everything seems to work now.
28  * 20.12.91  -  Ok, making the swap-device changeable like the root.
29  */
30
31 /*
32  * 05.04.94  -  Multi-page memory management added for v1.1.
33  *              Idea by Alex Bligh (alex@cconcepts.co.uk)
34  *
35  * 16.07.99  -  Support of BIGMEM added by Gerhard Wichert, Siemens AG
36  *              (Gerhard.Wichert@pdb.siemens.de)
37  *
38  * Aug/Sep 2004 Changed to four level page tables (Andi Kleen)
39  */
40
41 #include <linux/kernel_stat.h>
42 #include <linux/mm.h>
43 #include <linux/hugetlb.h>
44 #include <linux/mman.h>
45 #include <linux/swap.h>
46 #include <linux/highmem.h>
47 #include <linux/pagemap.h>
48 #include <linux/ksm.h>
49 #include <linux/rmap.h>
50 #include <linux/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 void 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
1295
1296 static void unmap_single_vma(struct mmu_gather *tlb,
1297                 struct vm_area_struct *vma, unsigned long start_addr,
1298                 unsigned long end_addr, unsigned long *nr_accounted,
1299                 struct zap_details *details)
1300 {
1301         unsigned long start = max(vma->vm_start, start_addr);
1302         unsigned long end;
1303
1304         if (start >= vma->vm_end)
1305                 return;
1306         end = min(vma->vm_end, end_addr);
1307         if (end <= vma->vm_start)
1308                 return;
1309
1310         if (vma->vm_flags & VM_ACCOUNT)
1311                 *nr_accounted += (end - start) >> PAGE_SHIFT;
1312
1313         if (unlikely(is_pfn_mapping(vma)))
1314                 untrack_pfn_vma(vma, 0, 0);
1315
1316         if (start != end) {
1317                 if (unlikely(is_vm_hugetlb_page(vma))) {
1318                         /*
1319                          * It is undesirable to test vma->vm_file as it
1320                          * should be non-null for valid hugetlb area.
1321                          * However, vm_file will be NULL in the error
1322                          * cleanup path of do_mmap_pgoff. When
1323                          * hugetlbfs ->mmap method fails,
1324                          * do_mmap_pgoff() nullifies vma->vm_file
1325                          * before calling this function to clean up.
1326                          * Since no pte has actually been setup, it is
1327                          * safe to do nothing in this case.
1328                          */
1329                         if (vma->vm_file)
1330                                 unmap_hugepage_range(vma, start, end, NULL);
1331                 } else
1332                         unmap_page_range(tlb, vma, start, end, details);
1333         }
1334 }
1335
1336 /**
1337  * unmap_vmas - unmap a range of memory covered by a list of vma's
1338  * @tlb: address of the caller's struct mmu_gather
1339  * @vma: the starting vma
1340  * @start_addr: virtual address at which to start unmapping
1341  * @end_addr: virtual address at which to end unmapping
1342  * @nr_accounted: Place number of unmapped pages in vm-accountable vma's here
1343  * @details: details of nonlinear truncation or shared cache invalidation
1344  *
1345  * Unmap all pages in the vma list.
1346  *
1347  * Only addresses between `start' and `end' will be unmapped.
1348  *
1349  * The VMA list must be sorted in ascending virtual address order.
1350  *
1351  * unmap_vmas() assumes that the caller will flush the whole unmapped address
1352  * range after unmap_vmas() returns.  So the only responsibility here is to
1353  * ensure that any thus-far unmapped pages are flushed before unmap_vmas()
1354  * drops the lock and schedules.
1355  */
1356 void unmap_vmas(struct mmu_gather *tlb,
1357                 struct vm_area_struct *vma, unsigned long start_addr,
1358                 unsigned long end_addr, unsigned long *nr_accounted,
1359                 struct zap_details *details)
1360 {
1361         struct mm_struct *mm = vma->vm_mm;
1362
1363         mmu_notifier_invalidate_range_start(mm, start_addr, end_addr);
1364         for ( ; vma && vma->vm_start < end_addr; vma = vma->vm_next)
1365                 unmap_single_vma(tlb, vma, start_addr, end_addr, nr_accounted,
1366                                  details);
1367         mmu_notifier_invalidate_range_end(mm, start_addr, end_addr);
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  * Caller must protect the VMA list
1378  */
1379 void zap_page_range(struct vm_area_struct *vma, unsigned long address,
1380                 unsigned long size, struct zap_details *details)
1381 {
1382         struct mm_struct *mm = vma->vm_mm;
1383         struct mmu_gather tlb;
1384         unsigned long end = address + size;
1385         unsigned long nr_accounted = 0;
1386
1387         lru_add_drain();
1388         tlb_gather_mmu(&tlb, mm, 0);
1389         update_hiwater_rss(mm);
1390         unmap_vmas(&tlb, vma, address, end, &nr_accounted, details);
1391         tlb_finish_mmu(&tlb, address, end);
1392 }
1393
1394 /**
1395  * zap_page_range_single - remove user pages in a given range
1396  * @vma: vm_area_struct holding the applicable pages
1397  * @address: starting address of pages to zap
1398  * @size: number of bytes to zap
1399  * @details: details of nonlinear truncation or shared cache invalidation
1400  *
1401  * The range must fit into one VMA.
1402  */
1403 static void zap_page_range_single(struct vm_area_struct *vma, unsigned long address,
1404                 unsigned long size, struct zap_details *details)
1405 {
1406         struct mm_struct *mm = vma->vm_mm;
1407         struct mmu_gather tlb;
1408         unsigned long end = address + size;
1409         unsigned long nr_accounted = 0;
1410
1411         lru_add_drain();
1412         tlb_gather_mmu(&tlb, mm, 0);
1413         update_hiwater_rss(mm);
1414         mmu_notifier_invalidate_range_start(mm, address, end);
1415         unmap_single_vma(&tlb, vma, address, end, &nr_accounted, details);
1416         mmu_notifier_invalidate_range_end(mm, address, end);
1417         tlb_finish_mmu(&tlb, address, end);
1418 }
1419
1420 /**
1421  * zap_vma_ptes - remove ptes mapping the vma
1422  * @vma: vm_area_struct holding ptes to be zapped
1423  * @address: starting address of pages to zap
1424  * @size: number of bytes to zap
1425  *
1426  * This function only unmaps ptes assigned to VM_PFNMAP vmas.
1427  *
1428  * The entire address range must be fully contained within the vma.
1429  *
1430  * Returns 0 if successful.
1431  */
1432 int zap_vma_ptes(struct vm_area_struct *vma, unsigned long address,
1433                 unsigned long size)
1434 {
1435         if (address < vma->vm_start || address + size > vma->vm_end ||
1436                         !(vma->vm_flags & VM_PFNMAP))
1437                 return -1;
1438         zap_page_range_single(vma, address, size, NULL);
1439         return 0;
1440 }
1441 EXPORT_SYMBOL_GPL(zap_vma_ptes);
1442
1443 /**
1444  * follow_page - look up a page descriptor from a user-virtual address
1445  * @vma: vm_area_struct mapping @address
1446  * @address: virtual address to look up
1447  * @flags: flags modifying lookup behaviour
1448  *
1449  * @flags can have FOLL_ flags set, defined in <linux/mm.h>
1450  *
1451  * Returns the mapped (struct page *), %NULL if no mapping exists, or
1452  * an error pointer if there is a mapping to something not represented
1453  * by a page descriptor (see also vm_normal_page()).
1454  */
1455 struct page *follow_page(struct vm_area_struct *vma, unsigned long address,
1456                         unsigned int flags)
1457 {
1458         pgd_t *pgd;
1459         pud_t *pud;
1460         pmd_t *pmd;
1461         pte_t *ptep, pte;
1462         spinlock_t *ptl;
1463         struct page *page;
1464         struct mm_struct *mm = vma->vm_mm;
1465
1466         page = follow_huge_addr(mm, address, flags & FOLL_WRITE);
1467         if (!IS_ERR(page)) {
1468                 BUG_ON(flags & FOLL_GET);
1469                 goto out;
1470         }
1471
1472         page = NULL;
1473         pgd = pgd_offset(mm, address);
1474         if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
1475                 goto no_page_table;
1476
1477         pud = pud_offset(pgd, address);
1478         if (pud_none(*pud))
1479                 goto no_page_table;
1480         if (pud_huge(*pud) && vma->vm_flags & VM_HUGETLB) {
1481                 BUG_ON(flags & FOLL_GET);
1482                 page = follow_huge_pud(mm, address, pud, flags & FOLL_WRITE);
1483                 goto out;
1484         }
1485         if (unlikely(pud_bad(*pud)))
1486                 goto no_page_table;
1487
1488         pmd = pmd_offset(pud, address);
1489         if (pmd_none(*pmd))
1490                 goto no_page_table;
1491         if (pmd_huge(*pmd) && vma->vm_flags & VM_HUGETLB) {
1492                 BUG_ON(flags & FOLL_GET);
1493                 page = follow_huge_pmd(mm, address, pmd, flags & FOLL_WRITE);
1494                 goto out;
1495         }
1496         if (pmd_trans_huge(*pmd)) {
1497                 if (flags & FOLL_SPLIT) {
1498                         split_huge_page_pmd(mm, pmd);
1499                         goto split_fallthrough;
1500                 }
1501                 spin_lock(&mm->page_table_lock);
1502                 if (likely(pmd_trans_huge(*pmd))) {
1503                         if (unlikely(pmd_trans_splitting(*pmd))) {
1504                                 spin_unlock(&mm->page_table_lock);
1505                                 wait_split_huge_page(vma->anon_vma, pmd);
1506                         } else {
1507                                 page = follow_trans_huge_pmd(mm, address,
1508                                                              pmd, flags);
1509                                 spin_unlock(&mm->page_table_lock);
1510                                 goto out;
1511                         }
1512                 } else
1513                         spin_unlock(&mm->page_table_lock);
1514                 /* fall through */
1515         }
1516 split_fallthrough:
1517         if (unlikely(pmd_bad(*pmd)))
1518                 goto no_page_table;
1519
1520         ptep = pte_offset_map_lock(mm, pmd, address, &ptl);
1521
1522         pte = *ptep;
1523         if (!pte_present(pte))
1524                 goto no_page;
1525         if ((flags & FOLL_WRITE) && !pte_write(pte))
1526                 goto unlock;
1527
1528         page = vm_normal_page(vma, address, pte);
1529         if (unlikely(!page)) {
1530                 if ((flags & FOLL_DUMP) ||
1531                     !is_zero_pfn(pte_pfn(pte)))
1532                         goto bad_page;
1533                 page = pte_page(pte);
1534         }
1535
1536         if (flags & FOLL_GET)
1537                 get_page_foll(page);
1538         if (flags & FOLL_TOUCH) {
1539                 if ((flags & FOLL_WRITE) &&
1540                     !pte_dirty(pte) && !PageDirty(page))
1541                         set_page_dirty(page);
1542                 /*
1543                  * pte_mkyoung() would be more correct here, but atomic care
1544                  * is needed to avoid losing the dirty bit: it is easier to use
1545                  * mark_page_accessed().
1546                  */
1547                 mark_page_accessed(page);
1548         }
1549         if ((flags & FOLL_MLOCK) && (vma->vm_flags & VM_LOCKED)) {
1550                 /*
1551                  * The preliminary mapping check is mainly to avoid the
1552                  * pointless overhead of lock_page on the ZERO_PAGE
1553                  * which might bounce very badly if there is contention.
1554                  *
1555                  * If the page is already locked, we don't need to
1556                  * handle it now - vmscan will handle it later if and
1557                  * when it attempts to reclaim the page.
1558                  */
1559                 if (page->mapping && trylock_page(page)) {
1560                         lru_add_drain();  /* push cached pages to LRU */
1561                         /*
1562                          * Because we lock page here and migration is
1563                          * blocked by the pte's page reference, we need
1564                          * only check for file-cache page truncation.
1565                          */
1566                         if (page->mapping)
1567                                 mlock_vma_page(page);
1568                         unlock_page(page);
1569                 }
1570         }
1571 unlock:
1572         pte_unmap_unlock(ptep, ptl);
1573 out:
1574         return page;
1575
1576 bad_page:
1577         pte_unmap_unlock(ptep, ptl);
1578         return ERR_PTR(-EFAULT);
1579
1580 no_page:
1581         pte_unmap_unlock(ptep, ptl);
1582         if (!pte_none(pte))
1583                 return page;
1584
1585 no_page_table:
1586         /*
1587          * When core dumping an enormous anonymous area that nobody
1588          * has touched so far, we don't want to allocate unnecessary pages or
1589          * page tables.  Return error instead of NULL to skip handle_mm_fault,
1590          * then get_dump_page() will return NULL to leave a hole in the dump.
1591          * But we can only make this optimization where a hole would surely
1592          * be zero-filled if handle_mm_fault() actually did handle it.
1593          */
1594         if ((flags & FOLL_DUMP) &&
1595             (!vma->vm_ops || !vma->vm_ops->fault))
1596                 return ERR_PTR(-EFAULT);
1597         return page;
1598 }
1599
1600 static inline int stack_guard_page(struct vm_area_struct *vma, unsigned long addr)
1601 {
1602         return stack_guard_page_start(vma, addr) ||
1603                stack_guard_page_end(vma, addr+PAGE_SIZE);
1604 }
1605
1606 /**
1607  * __get_user_pages() - pin user pages in memory
1608  * @tsk:        task_struct of target task
1609  * @mm:         mm_struct of target mm
1610  * @start:      starting user address
1611  * @nr_pages:   number of pages from start to pin
1612  * @gup_flags:  flags modifying pin behaviour
1613  * @pages:      array that receives pointers to the pages pinned.
1614  *              Should be at least nr_pages long. Or NULL, if caller
1615  *              only intends to ensure the pages are faulted in.
1616  * @vmas:       array of pointers to vmas corresponding to each page.
1617  *              Or NULL if the caller does not require them.
1618  * @nonblocking: whether waiting for disk IO or mmap_sem contention
1619  *
1620  * Returns number of pages pinned. This may be fewer than the number
1621  * requested. If nr_pages is 0 or negative, returns 0. If no pages
1622  * were pinned, returns -errno. Each page returned must be released
1623  * with a put_page() call when it is finished with. vmas will only
1624  * remain valid while mmap_sem is held.
1625  *
1626  * Must be called with mmap_sem held for read or write.
1627  *
1628  * __get_user_pages walks a process's page tables and takes a reference to
1629  * each struct page that each user address corresponds to at a given
1630  * instant. That is, it takes the page that would be accessed if a user
1631  * thread accesses the given user virtual address at that instant.
1632  *
1633  * This does not guarantee that the page exists in the user mappings when
1634  * __get_user_pages returns, and there may even be a completely different
1635  * page there in some cases (eg. if mmapped pagecache has been invalidated
1636  * and subsequently re faulted). However it does guarantee that the page
1637  * won't be freed completely. And mostly callers simply care that the page
1638  * contains data that was valid *at some point in time*. Typically, an IO
1639  * or similar operation cannot guarantee anything stronger anyway because
1640  * locks can't be held over the syscall boundary.
1641  *
1642  * If @gup_flags & FOLL_WRITE == 0, the page must not be written to. If
1643  * the page is written to, set_page_dirty (or set_page_dirty_lock, as
1644  * appropriate) must be called after the page is finished with, and
1645  * before put_page is called.
1646  *
1647  * If @nonblocking != NULL, __get_user_pages will not wait for disk IO
1648  * or mmap_sem contention, and if waiting is needed to pin all pages,
1649  * *@nonblocking will be set to 0.
1650  *
1651  * In most cases, get_user_pages or get_user_pages_fast should be used
1652  * instead of __get_user_pages. __get_user_pages should be used only if
1653  * you need some special @gup_flags.
1654  */
1655 int __get_user_pages(struct task_struct *tsk, struct mm_struct *mm,
1656                      unsigned long start, int nr_pages, unsigned int gup_flags,
1657                      struct page **pages, struct vm_area_struct **vmas,
1658                      int *nonblocking)
1659 {
1660         int i;
1661         unsigned long vm_flags;
1662
1663         if (nr_pages <= 0)
1664                 return 0;
1665
1666         VM_BUG_ON(!!pages != !!(gup_flags & FOLL_GET));
1667
1668         /* 
1669          * Require read or write permissions.
1670          * If FOLL_FORCE is set, we only require the "MAY" flags.
1671          */
1672         vm_flags  = (gup_flags & FOLL_WRITE) ?
1673                         (VM_WRITE | VM_MAYWRITE) : (VM_READ | VM_MAYREAD);
1674         vm_flags &= (gup_flags & FOLL_FORCE) ?
1675                         (VM_MAYREAD | VM_MAYWRITE) : (VM_READ | VM_WRITE);
1676         i = 0;
1677
1678         do {
1679                 struct vm_area_struct *vma;
1680
1681                 vma = find_extend_vma(mm, start);
1682                 if (!vma && in_gate_area(mm, start)) {
1683                         unsigned long pg = start & PAGE_MASK;
1684                         pgd_t *pgd;
1685                         pud_t *pud;
1686                         pmd_t *pmd;
1687                         pte_t *pte;
1688
1689                         /* user gate pages are read-only */
1690                         if (gup_flags & FOLL_WRITE)
1691                                 return i ? : -EFAULT;
1692                         if (pg > TASK_SIZE)
1693                                 pgd = pgd_offset_k(pg);
1694                         else
1695                                 pgd = pgd_offset_gate(mm, pg);
1696                         BUG_ON(pgd_none(*pgd));
1697                         pud = pud_offset(pgd, pg);
1698                         BUG_ON(pud_none(*pud));
1699                         pmd = pmd_offset(pud, pg);
1700                         if (pmd_none(*pmd))
1701                                 return i ? : -EFAULT;
1702                         VM_BUG_ON(pmd_trans_huge(*pmd));
1703                         pte = pte_offset_map(pmd, pg);
1704                         if (pte_none(*pte)) {
1705                                 pte_unmap(pte);
1706                                 return i ? : -EFAULT;
1707                         }
1708                         vma = get_gate_vma(mm);
1709                         if (pages) {
1710                                 struct page *page;
1711
1712                                 page = vm_normal_page(vma, start, *pte);
1713                                 if (!page) {
1714                                         if (!(gup_flags & FOLL_DUMP) &&
1715                                              is_zero_pfn(pte_pfn(*pte)))
1716                                                 page = pte_page(*pte);
1717                                         else {
1718                                                 pte_unmap(pte);
1719                                                 return i ? : -EFAULT;
1720                                         }
1721                                 }
1722                                 pages[i] = page;
1723                                 get_page(page);
1724                         }
1725                         pte_unmap(pte);
1726                         goto next_page;
1727                 }
1728
1729                 if (!vma ||
1730                     (vma->vm_flags & (VM_IO | VM_PFNMAP)) ||
1731                     !(vm_flags & vma->vm_flags))
1732                         return i ? : -EFAULT;
1733
1734                 if (is_vm_hugetlb_page(vma)) {
1735                         i = follow_hugetlb_page(mm, vma, pages, vmas,
1736                                         &start, &nr_pages, i, gup_flags);
1737                         continue;
1738                 }
1739
1740                 do {
1741                         struct page *page;
1742                         unsigned int foll_flags = gup_flags;
1743
1744                         /*
1745                          * If we have a pending SIGKILL, don't keep faulting
1746                          * pages and potentially allocating memory.
1747                          */
1748                         if (unlikely(fatal_signal_pending(current)))
1749                                 return i ? i : -ERESTARTSYS;
1750
1751                         cond_resched();
1752                         while (!(page = follow_page(vma, start, foll_flags))) {
1753                                 int ret;
1754                                 unsigned int fault_flags = 0;
1755
1756                                 /* For mlock, just skip the stack guard page. */
1757                                 if (foll_flags & FOLL_MLOCK) {
1758                                         if (stack_guard_page(vma, start))
1759                                                 goto next_page;
1760                                 }
1761                                 if (foll_flags & FOLL_WRITE)
1762                                         fault_flags |= FAULT_FLAG_WRITE;
1763                                 if (nonblocking)
1764                                         fault_flags |= FAULT_FLAG_ALLOW_RETRY;
1765                                 if (foll_flags & FOLL_NOWAIT)
1766                                         fault_flags |= (FAULT_FLAG_ALLOW_RETRY | FAULT_FLAG_RETRY_NOWAIT);
1767
1768                                 ret = handle_mm_fault(mm, vma, start,
1769                                                         fault_flags);
1770
1771                                 if (ret & VM_FAULT_ERROR) {
1772                                         if (ret & VM_FAULT_OOM)
1773                                                 return i ? i : -ENOMEM;
1774                                         if (ret & (VM_FAULT_HWPOISON |
1775                                                    VM_FAULT_HWPOISON_LARGE)) {
1776                                                 if (i)
1777                                                         return i;
1778                                                 else if (gup_flags & FOLL_HWPOISON)
1779                                                         return -EHWPOISON;
1780                                                 else
1781                                                         return -EFAULT;
1782                                         }
1783                                         if (ret & VM_FAULT_SIGBUS)
1784                                                 return i ? i : -EFAULT;
1785                                         BUG();
1786                                 }
1787
1788                                 if (tsk) {
1789                                         if (ret & VM_FAULT_MAJOR)
1790                                                 tsk->maj_flt++;
1791                                         else
1792                                                 tsk->min_flt++;
1793                                 }
1794
1795                                 if (ret & VM_FAULT_RETRY) {
1796                                         if (nonblocking)
1797                                                 *nonblocking = 0;
1798                                         return i;
1799                                 }
1800
1801                                 /*
1802                                  * The VM_FAULT_WRITE bit tells us that
1803                                  * do_wp_page has broken COW when necessary,
1804                                  * even if maybe_mkwrite decided not to set
1805                                  * pte_write. We can thus safely do subsequent
1806                                  * page lookups as if they were reads. But only
1807                                  * do so when looping for pte_write is futile:
1808                                  * in some cases userspace may also be wanting
1809                                  * to write to the gotten user page, which a
1810                                  * read fault here might prevent (a readonly
1811                                  * page might get reCOWed by userspace write).
1812                                  */
1813                                 if ((ret & VM_FAULT_WRITE) &&
1814                                     !(vma->vm_flags & VM_WRITE))
1815                                         foll_flags &= ~FOLL_WRITE;
1816
1817                                 cond_resched();
1818                         }
1819                         if (IS_ERR(page))
1820                                 return i ? i : PTR_ERR(page);
1821                         if (pages) {
1822                                 pages[i] = page;
1823
1824                                 flush_anon_page(vma, page, start);
1825                                 flush_dcache_page(page);
1826                         }
1827 next_page:
1828                         if (vmas)
1829                                 vmas[i] = vma;
1830                         i++;
1831                         start += PAGE_SIZE;
1832                         nr_pages--;
1833                 } while (nr_pages && start < vma->vm_end);
1834         } while (nr_pages);
1835         return i;
1836 }
1837 EXPORT_SYMBOL(__get_user_pages);
1838
1839 /*
1840  * fixup_user_fault() - manually resolve a user page fault
1841  * @tsk:        the task_struct to use for page fault accounting, or
1842  *              NULL if faults are not to be recorded.
1843  * @mm:         mm_struct of target mm
1844  * @address:    user address
1845  * @fault_flags:flags to pass down to handle_mm_fault()
1846  *
1847  * This is meant to be called in the specific scenario where for locking reasons
1848  * we try to access user memory in atomic context (within a pagefault_disable()
1849  * section), this returns -EFAULT, and we want to resolve the user fault before
1850  * trying again.
1851  *
1852  * Typically this is meant to be used by the futex code.
1853  *
1854  * The main difference with get_user_pages() is that this function will
1855  * unconditionally call handle_mm_fault() which will in turn perform all the
1856  * necessary SW fixup of the dirty and young bits in the PTE, while
1857  * handle_mm_fault() only guarantees to update these in the struct page.
1858  *
1859  * This is important for some architectures where those bits also gate the
1860  * access permission to the page because they are maintained in software.  On
1861  * such architectures, gup() will not be enough to make a subsequent access
1862  * succeed.
1863  *
1864  * This should be called with the mm_sem held for read.
1865  */
1866 int fixup_user_fault(struct task_struct *tsk, struct mm_struct *mm,
1867                      unsigned long address, unsigned int fault_flags)
1868 {
1869         struct vm_area_struct *vma;
1870         int ret;
1871
1872         vma = find_extend_vma(mm, address);
1873         if (!vma || address < vma->vm_start)
1874                 return -EFAULT;
1875
1876         ret = handle_mm_fault(mm, vma, address, fault_flags);
1877         if (ret & VM_FAULT_ERROR) {
1878                 if (ret & VM_FAULT_OOM)
1879                         return -ENOMEM;
1880                 if (ret & (VM_FAULT_HWPOISON | VM_FAULT_HWPOISON_LARGE))
1881                         return -EHWPOISON;
1882                 if (ret & VM_FAULT_SIGBUS)
1883                         return -EFAULT;
1884                 BUG();
1885         }
1886         if (tsk) {
1887                 if (ret & VM_FAULT_MAJOR)
1888                         tsk->maj_flt++;
1889                 else
1890                         tsk->min_flt++;
1891         }
1892         return 0;
1893 }
1894
1895 /*
1896  * get_user_pages() - pin user pages in memory
1897  * @tsk:        the task_struct to use for page fault accounting, or
1898  *              NULL if faults are not to be recorded.
1899  * @mm:         mm_struct of target mm
1900  * @start:      starting user address
1901  * @nr_pages:   number of pages from start to pin
1902  * @write:      whether pages will be written to by the caller
1903  * @force:      whether to force write access even if user mapping is
1904  *              readonly. This will result in the page being COWed even
1905  *              in MAP_SHARED mappings. You do not want this.
1906  * @pages:      array that receives pointers to the pages pinned.
1907  *              Should be at least nr_pages long. Or NULL, if caller
1908  *              only intends to ensure the pages are faulted in.
1909  * @vmas:       array of pointers to vmas corresponding to each page.
1910  *              Or NULL if the caller does not require them.
1911  *
1912  * Returns number of pages pinned. This may be fewer than the number
1913  * requested. If nr_pages is 0 or negative, returns 0. If no pages
1914  * were pinned, returns -errno. Each page returned must be released
1915  * with a put_page() call when it is finished with. vmas will only
1916  * remain valid while mmap_sem is held.
1917  *
1918  * Must be called with mmap_sem held for read or write.
1919  *
1920  * get_user_pages walks a process's page tables and takes a reference to
1921  * each struct page that each user address corresponds to at a given
1922  * instant. That is, it takes the page that would be accessed if a user
1923  * thread accesses the given user virtual address at that instant.
1924  *
1925  * This does not guarantee that the page exists in the user mappings when
1926  * get_user_pages returns, and there may even be a completely different
1927  * page there in some cases (eg. if mmapped pagecache has been invalidated
1928  * and subsequently re faulted). However it does guarantee that the page
1929  * won't be freed completely. And mostly callers simply care that the page
1930  * contains data that was valid *at some point in time*. Typically, an IO
1931  * or similar operation cannot guarantee anything stronger anyway because
1932  * locks can't be held over the syscall boundary.
1933  *
1934  * If write=0, the page must not be written to. If the page is written to,
1935  * set_page_dirty (or set_page_dirty_lock, as appropriate) must be called
1936  * after the page is finished with, and before put_page is called.
1937  *
1938  * get_user_pages is typically used for fewer-copy IO operations, to get a
1939  * handle on the memory by some means other than accesses via the user virtual
1940  * addresses. The pages may be submitted for DMA to devices or accessed via
1941  * their kernel linear mapping (via the kmap APIs). Care should be taken to
1942  * use the correct cache flushing APIs.
1943  *
1944  * See also get_user_pages_fast, for performance critical applications.
1945  */
1946 int get_user_pages(struct task_struct *tsk, struct mm_struct *mm,
1947                 unsigned long start, int nr_pages, int write, int force,
1948                 struct page **pages, struct vm_area_struct **vmas)
1949 {
1950         int flags = FOLL_TOUCH;
1951
1952         if (pages)
1953                 flags |= FOLL_GET;
1954         if (write)
1955                 flags |= FOLL_WRITE;
1956         if (force)
1957                 flags |= FOLL_FORCE;
1958
1959         return __get_user_pages(tsk, mm, start, nr_pages, flags, pages, vmas,
1960                                 NULL);
1961 }
1962 EXPORT_SYMBOL(get_user_pages);
1963
1964 /**
1965  * get_dump_page() - pin user page in memory while writing it to core dump
1966  * @addr: user address
1967  *
1968  * Returns struct page pointer of user page pinned for dump,
1969  * to be freed afterwards by page_cache_release() or put_page().
1970  *
1971  * Returns NULL on any kind of failure - a hole must then be inserted into
1972  * the corefile, to preserve alignment with its headers; and also returns
1973  * NULL wherever the ZERO_PAGE, or an anonymous pte_none, has been found -
1974  * allowing a hole to be left in the corefile to save diskspace.
1975  *
1976  * Called without mmap_sem, but after all other threads have been killed.
1977  */
1978 #ifdef CONFIG_ELF_CORE
1979 struct page *get_dump_page(unsigned long addr)
1980 {
1981         struct vm_area_struct *vma;
1982         struct page *page;
1983
1984         if (__get_user_pages(current, current->mm, addr, 1,
1985                              FOLL_FORCE | FOLL_DUMP | FOLL_GET, &page, &vma,
1986                              NULL) < 1)
1987                 return NULL;
1988         flush_cache_page(vma, addr, page_to_pfn(page));
1989         return page;
1990 }
1991 #endif /* CONFIG_ELF_CORE */
1992
1993 pte_t *__get_locked_pte(struct mm_struct *mm, unsigned long addr,
1994                         spinlock_t **ptl)
1995 {
1996         pgd_t * pgd = pgd_offset(mm, addr);
1997         pud_t * pud = pud_alloc(mm, pgd, addr);
1998         if (pud) {
1999                 pmd_t * pmd = pmd_alloc(mm, pud, addr);
2000                 if (pmd) {
2001                         VM_BUG_ON(pmd_trans_huge(*pmd));
2002                         return pte_alloc_map_lock(mm, pmd, addr, ptl);
2003                 }
2004         }
2005         return NULL;
2006 }
2007
2008 /*
2009  * This is the old fallback for page remapping.
2010  *
2011  * For historical reasons, it only allows reserved pages. Only
2012  * old drivers should use this, and they needed to mark their
2013  * pages reserved for the old functions anyway.
2014  */
2015 static int insert_page(struct vm_area_struct *vma, unsigned long addr,
2016                         struct page *page, pgprot_t prot)
2017 {
2018         struct mm_struct *mm = vma->vm_mm;
2019         int retval;
2020         pte_t *pte;
2021         spinlock_t *ptl;
2022
2023         retval = -EINVAL;
2024         if (PageAnon(page))
2025                 goto out;
2026         retval = -ENOMEM;
2027         flush_dcache_page(page);
2028         pte = get_locked_pte(mm, addr, &ptl);
2029         if (!pte)
2030                 goto out;
2031         retval = -EBUSY;
2032         if (!pte_none(*pte))
2033                 goto out_unlock;
2034
2035         /* Ok, finally just insert the thing.. */
2036         get_page(page);
2037         inc_mm_counter_fast(mm, MM_FILEPAGES);
2038         page_add_file_rmap(page);
2039         set_pte_at(mm, addr, pte, mk_pte(page, prot));
2040
2041         retval = 0;
2042         pte_unmap_unlock(pte, ptl);
2043         return retval;
2044 out_unlock:
2045         pte_unmap_unlock(pte, ptl);
2046 out:
2047         return retval;
2048 }
2049
2050 /**
2051  * vm_insert_page - insert single page into user vma
2052  * @vma: user vma to map to
2053  * @addr: target user address of this page
2054  * @page: source kernel page
2055  *
2056  * This allows drivers to insert individual pages they've allocated
2057  * into a user vma.
2058  *
2059  * The page has to be a nice clean _individual_ kernel allocation.
2060  * If you allocate a compound page, you need to have marked it as
2061  * such (__GFP_COMP), or manually just split the page up yourself
2062  * (see split_page()).
2063  *
2064  * NOTE! Traditionally this was done with "remap_pfn_range()" which
2065  * took an arbitrary page protection parameter. This doesn't allow
2066  * that. Your vma protection will have to be set up correctly, which
2067  * means that if you want a shared writable mapping, you'd better
2068  * ask for a shared writable mapping!
2069  *
2070  * The page does not need to be reserved.
2071  */
2072 int vm_insert_page(struct vm_area_struct *vma, unsigned long addr,
2073                         struct page *page)
2074 {
2075         if (addr < vma->vm_start || addr >= vma->vm_end)
2076                 return -EFAULT;
2077         if (!page_count(page))
2078                 return -EINVAL;
2079         vma->vm_flags |= VM_INSERTPAGE;
2080         return insert_page(vma, addr, page, vma->vm_page_prot);
2081 }
2082 EXPORT_SYMBOL(vm_insert_page);
2083
2084 static int insert_pfn(struct vm_area_struct *vma, unsigned long addr,
2085                         unsigned long pfn, pgprot_t prot)
2086 {
2087         struct mm_struct *mm = vma->vm_mm;
2088         int retval;
2089         pte_t *pte, entry;
2090         spinlock_t *ptl;
2091
2092         retval = -ENOMEM;
2093         pte = get_locked_pte(mm, addr, &ptl);
2094         if (!pte)
2095                 goto out;
2096         retval = -EBUSY;
2097         if (!pte_none(*pte))
2098                 goto out_unlock;
2099
2100         /* Ok, finally just insert the thing.. */
2101         entry = pte_mkspecial(pfn_pte(pfn, prot));
2102         set_pte_at(mm, addr, pte, entry);
2103         update_mmu_cache(vma, addr, pte); /* XXX: why not for insert_page? */
2104
2105         retval = 0;
2106 out_unlock:
2107         pte_unmap_unlock(pte, ptl);
2108 out:
2109         return retval;
2110 }
2111
2112 /**
2113  * vm_insert_pfn - insert single pfn into user vma
2114  * @vma: user vma to map to
2115  * @addr: target user address of this page
2116  * @pfn: source kernel pfn
2117  *
2118  * Similar to vm_inert_page, this allows drivers to insert individual pages
2119  * they've allocated into a user vma. Same comments apply.
2120  *
2121  * This function should only be called from a vm_ops->fault handler, and
2122  * in that case the handler should return NULL.
2123  *
2124  * vma cannot be a COW mapping.
2125  *
2126  * As this is called only for pages that do not currently exist, we
2127  * do not need to flush old virtual caches or the TLB.
2128  */
2129 int vm_insert_pfn(struct vm_area_struct *vma, unsigned long addr,
2130                         unsigned long pfn)
2131 {
2132         int ret;
2133         pgprot_t pgprot = vma->vm_page_prot;
2134         /*
2135          * Technically, architectures with pte_special can avoid all these
2136          * restrictions (same for remap_pfn_range).  However we would like
2137          * consistency in testing and feature parity among all, so we should
2138          * try to keep these invariants in place for everybody.
2139          */
2140         BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)));
2141         BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) ==
2142                                                 (VM_PFNMAP|VM_MIXEDMAP));
2143         BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags));
2144         BUG_ON((vma->vm_flags & VM_MIXEDMAP) && pfn_valid(pfn));
2145
2146         if (addr < vma->vm_start || addr >= vma->vm_end)
2147                 return -EFAULT;
2148         if (track_pfn_vma_new(vma, &pgprot, pfn, PAGE_SIZE))
2149                 return -EINVAL;
2150
2151         ret = insert_pfn(vma, addr, pfn, pgprot);
2152
2153         if (ret)
2154                 untrack_pfn_vma(vma, pfn, PAGE_SIZE);
2155
2156         return ret;
2157 }
2158 EXPORT_SYMBOL(vm_insert_pfn);
2159
2160 int vm_insert_mixed(struct vm_area_struct *vma, unsigned long addr,
2161                         unsigned long pfn)
2162 {
2163         BUG_ON(!(vma->vm_flags & VM_MIXEDMAP));
2164
2165         if (addr < vma->vm_start || addr >= vma->vm_end)
2166                 return -EFAULT;
2167
2168         /*
2169          * If we don't have pte special, then we have to use the pfn_valid()
2170          * based VM_MIXEDMAP scheme (see vm_normal_page), and thus we *must*
2171          * refcount the page if pfn_valid is true (hence insert_page rather
2172          * than insert_pfn).  If a zero_pfn were inserted into a VM_MIXEDMAP
2173          * without pte special, it would there be refcounted as a normal page.
2174          */
2175         if (!HAVE_PTE_SPECIAL && pfn_valid(pfn)) {
2176                 struct page *page;
2177
2178                 page = pfn_to_page(pfn);
2179                 return insert_page(vma, addr, page, vma->vm_page_prot);
2180         }
2181         return insert_pfn(vma, addr, pfn, vma->vm_page_prot);
2182 }
2183 EXPORT_SYMBOL(vm_insert_mixed);
2184
2185 /*
2186  * maps a range of physical memory into the requested pages. the old
2187  * mappings are removed. any references to nonexistent pages results
2188  * in null mappings (currently treated as "copy-on-access")
2189  */
2190 static int remap_pte_range(struct mm_struct *mm, pmd_t *pmd,
2191                         unsigned long addr, unsigned long end,
2192                         unsigned long pfn, pgprot_t prot)
2193 {
2194         pte_t *pte;
2195         spinlock_t *ptl;
2196
2197         pte = pte_alloc_map_lock(mm, pmd, addr, &ptl);
2198         if (!pte)
2199                 return -ENOMEM;
2200         arch_enter_lazy_mmu_mode();
2201         do {
2202                 BUG_ON(!pte_none(*pte));
2203                 set_pte_at(mm, addr, pte, pte_mkspecial(pfn_pte(pfn, prot)));
2204                 pfn++;
2205         } while (pte++, addr += PAGE_SIZE, addr != end);
2206         arch_leave_lazy_mmu_mode();
2207         pte_unmap_unlock(pte - 1, ptl);
2208         return 0;
2209 }
2210
2211 static inline int remap_pmd_range(struct mm_struct *mm, pud_t *pud,
2212                         unsigned long addr, unsigned long end,
2213                         unsigned long pfn, pgprot_t prot)
2214 {
2215         pmd_t *pmd;
2216         unsigned long next;
2217
2218         pfn -= addr >> PAGE_SHIFT;
2219         pmd = pmd_alloc(mm, pud, addr);
2220         if (!pmd)
2221                 return -ENOMEM;
2222         VM_BUG_ON(pmd_trans_huge(*pmd));
2223         do {
2224                 next = pmd_addr_end(addr, end);
2225                 if (remap_pte_range(mm, pmd, addr, next,
2226                                 pfn + (addr >> PAGE_SHIFT), prot))
2227                         return -ENOMEM;
2228         } while (pmd++, addr = next, addr != end);
2229         return 0;
2230 }
2231
2232 static inline int remap_pud_range(struct mm_struct *mm, pgd_t *pgd,
2233                         unsigned long addr, unsigned long end,
2234                         unsigned long pfn, pgprot_t prot)
2235 {
2236         pud_t *pud;
2237         unsigned long next;
2238
2239         pfn -= addr >> PAGE_SHIFT;
2240         pud = pud_alloc(mm, pgd, addr);
2241         if (!pud)
2242                 return -ENOMEM;
2243         do {
2244                 next = pud_addr_end(addr, end);
2245                 if (remap_pmd_range(mm, pud, addr, next,
2246                                 pfn + (addr >> PAGE_SHIFT), prot))
2247                         return -ENOMEM;
2248         } while (pud++, addr = next, addr != end);
2249         return 0;
2250 }
2251
2252 /**
2253  * remap_pfn_range - remap kernel memory to userspace
2254  * @vma: user vma to map to
2255  * @addr: target user address to start at
2256  * @pfn: physical address of kernel memory
2257  * @size: size of map area
2258  * @prot: page protection flags for this mapping
2259  *
2260  *  Note: this is only safe if the mm semaphore is held when called.
2261  */
2262 int remap_pfn_range(struct vm_area_struct *vma, unsigned long addr,
2263                     unsigned long pfn, unsigned long size, pgprot_t prot)
2264 {
2265         pgd_t *pgd;
2266         unsigned long next;
2267         unsigned long end = addr + PAGE_ALIGN(size);
2268         struct mm_struct *mm = vma->vm_mm;
2269         int err;
2270
2271         /*
2272          * Physically remapped pages are special. Tell the
2273          * rest of the world about it:
2274          *   VM_IO tells people not to look at these pages
2275          *      (accesses can have side effects).
2276          *   VM_RESERVED is specified all over the place, because
2277          *      in 2.4 it kept swapout's vma scan off this vma; but
2278          *      in 2.6 the LRU scan won't even find its pages, so this
2279          *      flag means no more than count its pages in reserved_vm,
2280          *      and omit it from core dump, even when VM_IO turned off.
2281          *   VM_PFNMAP tells the core MM that the base pages are just
2282          *      raw PFN mappings, and do not have a "struct page" associated
2283          *      with them.
2284          *
2285          * There's a horrible special case to handle copy-on-write
2286          * behaviour that some programs depend on. We mark the "original"
2287          * un-COW'ed pages by matching them up with "vma->vm_pgoff".
2288          */
2289         if (addr == vma->vm_start && end == vma->vm_end) {
2290                 vma->vm_pgoff = pfn;
2291                 vma->vm_flags |= VM_PFN_AT_MMAP;
2292         } else if (is_cow_mapping(vma->vm_flags))
2293                 return -EINVAL;
2294
2295         vma->vm_flags |= VM_IO | VM_RESERVED | VM_PFNMAP;
2296
2297         err = track_pfn_vma_new(vma, &prot, pfn, PAGE_ALIGN(size));
2298         if (err) {
2299                 /*
2300                  * To indicate that track_pfn related cleanup is not
2301                  * needed from higher level routine calling unmap_vmas
2302                  */
2303                 vma->vm_flags &= ~(VM_IO | VM_RESERVED | VM_PFNMAP);
2304                 vma->vm_flags &= ~VM_PFN_AT_MMAP;
2305                 return -EINVAL;
2306         }
2307
2308         BUG_ON(addr >= end);
2309         pfn -= addr >> PAGE_SHIFT;
2310         pgd = pgd_offset(mm, addr);
2311         flush_cache_range(vma, addr, end);
2312         do {
2313                 next = pgd_addr_end(addr, end);
2314                 err = remap_pud_range(mm, pgd, addr, next,
2315                                 pfn + (addr >> PAGE_SHIFT), prot);
2316                 if (err)
2317                         break;
2318         } while (pgd++, addr = next, addr != end);
2319
2320         if (err)
2321                 untrack_pfn_vma(vma, pfn, PAGE_ALIGN(size));
2322
2323         return err;
2324 }
2325 EXPORT_SYMBOL(remap_pfn_range);
2326
2327 static int apply_to_pte_range(struct mm_struct *mm, pmd_t *pmd,
2328                                      unsigned long addr, unsigned long end,
2329                                      pte_fn_t fn, void *data)
2330 {
2331         pte_t *pte;
2332         int err;
2333         pgtable_t token;
2334         spinlock_t *uninitialized_var(ptl);
2335
2336         pte = (mm == &init_mm) ?
2337                 pte_alloc_kernel(pmd, addr) :
2338                 pte_alloc_map_lock(mm, pmd, addr, &ptl);
2339         if (!pte)
2340                 return -ENOMEM;
2341
2342         BUG_ON(pmd_huge(*pmd));
2343
2344         arch_enter_lazy_mmu_mode();
2345
2346         token = pmd_pgtable(*pmd);
2347
2348         do {
2349                 err = fn(pte++, token, addr, data);
2350                 if (err)
2351                         break;
2352         } while (addr += PAGE_SIZE, addr != end);
2353
2354         arch_leave_lazy_mmu_mode();
2355
2356         if (mm != &init_mm)
2357                 pte_unmap_unlock(pte-1, ptl);
2358         return err;
2359 }
2360
2361 static int apply_to_pmd_range(struct mm_struct *mm, pud_t *pud,
2362                                      unsigned long addr, unsigned long end,
2363                                      pte_fn_t fn, void *data)
2364 {
2365         pmd_t *pmd;
2366         unsigned long next;
2367         int err;
2368
2369         BUG_ON(pud_huge(*pud));
2370
2371         pmd = pmd_alloc(mm, pud, addr);
2372         if (!pmd)
2373                 return -ENOMEM;
2374         do {
2375                 next = pmd_addr_end(addr, end);
2376                 err = apply_to_pte_range(mm, pmd, addr, next, fn, data);
2377                 if (err)
2378                         break;
2379         } while (pmd++, addr = next, addr != end);
2380         return err;
2381 }
2382
2383 static int apply_to_pud_range(struct mm_struct *mm, pgd_t *pgd,
2384                                      unsigned long addr, unsigned long end,
2385                                      pte_fn_t fn, void *data)
2386 {
2387         pud_t *pud;
2388         unsigned long next;
2389         int err;
2390
2391         pud = pud_alloc(mm, pgd, addr);
2392         if (!pud)
2393                 return -ENOMEM;
2394         do {
2395                 next = pud_addr_end(addr, end);
2396                 err = apply_to_pmd_range(mm, pud, addr, next, fn, data);
2397                 if (err)
2398                         break;
2399         } while (pud++, addr = next, addr != end);
2400         return err;
2401 }
2402
2403 /*
2404  * Scan a region of virtual memory, filling in page tables as necessary
2405  * and calling a provided function on each leaf page table.
2406  */
2407 int apply_to_page_range(struct mm_struct *mm, unsigned long addr,
2408                         unsigned long size, pte_fn_t fn, void *data)
2409 {
2410         pgd_t *pgd;
2411         unsigned long next;
2412         unsigned long end = addr + size;
2413         int err;
2414
2415         BUG_ON(addr >= end);
2416         pgd = pgd_offset(mm, addr);
2417         do {
2418                 next = pgd_addr_end(addr, end);
2419                 err = apply_to_pud_range(mm, pgd, addr, next, fn, data);
2420                 if (err)
2421                         break;
2422         } while (pgd++, addr = next, addr != end);
2423
2424         return err;
2425 }
2426 EXPORT_SYMBOL_GPL(apply_to_page_range);
2427
2428 /*
2429  * handle_pte_fault chooses page fault handler according to an entry
2430  * which was read non-atomically.  Before making any commitment, on
2431  * those architectures or configurations (e.g. i386 with PAE) which
2432  * might give a mix of unmatched parts, do_swap_page and do_nonlinear_fault
2433  * must check under lock before unmapping the pte and proceeding
2434  * (but do_wp_page is only called after already making such a check;
2435  * and do_anonymous_page can safely check later on).
2436  */
2437 static inline int pte_unmap_same(struct mm_struct *mm, pmd_t *pmd,
2438                                 pte_t *page_table, pte_t orig_pte)
2439 {
2440         int same = 1;
2441 #if defined(CONFIG_SMP) || defined(CONFIG_PREEMPT)
2442         if (sizeof(pte_t) > sizeof(unsigned long)) {
2443                 spinlock_t *ptl = pte_lockptr(mm, pmd);
2444                 spin_lock(ptl);
2445                 same = pte_same(*page_table, orig_pte);
2446                 spin_unlock(ptl);
2447         }
2448 #endif
2449         pte_unmap(page_table);
2450         return same;
2451 }
2452
2453 static inline void cow_user_page(struct page *dst, struct page *src, unsigned long va, struct vm_area_struct *vma)
2454 {
2455         /*
2456          * If the source page was a PFN mapping, we don't have
2457          * a "struct page" for it. We do a best-effort copy by
2458          * just copying from the original user address. If that
2459          * fails, we just zero-fill it. Live with it.
2460          */
2461         if (unlikely(!src)) {
2462                 void *kaddr = kmap_atomic(dst);
2463                 void __user *uaddr = (void __user *)(va & PAGE_MASK);
2464
2465                 /*
2466                  * This really shouldn't fail, because the page is there
2467                  * in the page tables. But it might just be unreadable,
2468                  * in which case we just give up and fill the result with
2469                  * zeroes.
2470                  */
2471                 if (__copy_from_user_inatomic(kaddr, uaddr, PAGE_SIZE))
2472                         clear_page(kaddr);
2473                 kunmap_atomic(kaddr);
2474                 flush_dcache_page(dst);
2475         } else
2476                 copy_user_highpage(dst, src, va, vma);
2477 }
2478
2479 /*
2480  * This routine handles present pages, when users try to write
2481  * to a shared page. It is done by copying the page to a new address
2482  * and decrementing the shared-page counter for the old page.
2483  *
2484  * Note that this routine assumes that the protection checks have been
2485  * done by the caller (the low-level page fault routine in most cases).
2486  * Thus we can safely just mark it writable once we've done any necessary
2487  * COW.
2488  *
2489  * We also mark the page dirty at this point even though the page will
2490  * change only once the write actually happens. This avoids a few races,
2491  * and potentially makes it more efficient.
2492  *
2493  * We enter with non-exclusive mmap_sem (to exclude vma changes,
2494  * but allow concurrent faults), with pte both mapped and locked.
2495  * We return with mmap_sem still held, but pte unmapped and unlocked.
2496  */
2497 static int do_wp_page(struct mm_struct *mm, struct vm_area_struct *vma,
2498                 unsigned long address, pte_t *page_table, pmd_t *pmd,
2499                 spinlock_t *ptl, pte_t orig_pte)
2500         __releases(ptl)
2501 {
2502         struct page *old_page, *new_page;
2503         pte_t entry;
2504         int ret = 0;
2505         int page_mkwrite = 0;
2506         struct page *dirty_page = NULL;
2507
2508         old_page = vm_normal_page(vma, address, orig_pte);
2509         if (!old_page) {
2510                 /*
2511                  * VM_MIXEDMAP !pfn_valid() case
2512                  *
2513                  * We should not cow pages in a shared writeable mapping.
2514                  * Just mark the pages writable as we can't do any dirty
2515                  * accounting on raw pfn maps.
2516                  */
2517                 if ((vma->vm_flags & (VM_WRITE|VM_SHARED)) ==
2518                                      (VM_WRITE|VM_SHARED))
2519                         goto reuse;
2520                 goto gotten;
2521         }
2522
2523         /*
2524          * Take out anonymous pages first, anonymous shared vmas are
2525          * not dirty accountable.
2526          */
2527         if (PageAnon(old_page) && !PageKsm(old_page)) {
2528                 if (!trylock_page(old_page)) {
2529                         page_cache_get(old_page);
2530                         pte_unmap_unlock(page_table, ptl);
2531                         lock_page(old_page);
2532                         page_table = pte_offset_map_lock(mm, pmd, address,
2533                                                          &ptl);
2534                         if (!pte_same(*page_table, orig_pte)) {
2535                                 unlock_page(old_page);
2536                                 goto unlock;
2537                         }
2538                         page_cache_release(old_page);
2539                 }
2540                 if (reuse_swap_page(old_page)) {
2541                         /*
2542                          * The page is all ours.  Move it to our anon_vma so
2543                          * the rmap code will not search our parent or siblings.
2544                          * Protected against the rmap code by the page lock.
2545                          */
2546                         page_move_anon_rmap(old_page, vma, address);
2547                         unlock_page(old_page);
2548                         goto reuse;
2549                 }
2550                 unlock_page(old_page);
2551         } else if (unlikely((vma->vm_flags & (VM_WRITE|VM_SHARED)) ==
2552                                         (VM_WRITE|VM_SHARED))) {
2553                 /*
2554                  * Only catch write-faults on shared writable pages,
2555                  * read-only shared pages can get COWed by
2556                  * get_user_pages(.write=1, .force=1).
2557                  */
2558                 if (vma->vm_ops && vma->vm_ops->page_mkwrite) {
2559                         struct vm_fault vmf;
2560                         int tmp;
2561
2562                         vmf.virtual_address = (void __user *)(address &
2563                                                                 PAGE_MASK);
2564                         vmf.pgoff = old_page->index;
2565                         vmf.flags = FAULT_FLAG_WRITE|FAULT_FLAG_MKWRITE;
2566                         vmf.page = old_page;
2567
2568                         /*
2569                          * Notify the address space that the page is about to
2570                          * become writable so that it can prohibit this or wait
2571                          * for the page to get into an appropriate state.
2572                          *
2573                          * We do this without the lock held, so that it can
2574                          * sleep if it needs to.
2575                          */
2576                         page_cache_get(old_page);
2577                         pte_unmap_unlock(page_table, ptl);
2578
2579                         tmp = vma->vm_ops->page_mkwrite(vma, &vmf);
2580                         if (unlikely(tmp &
2581                                         (VM_FAULT_ERROR | VM_FAULT_NOPAGE))) {
2582                                 ret = tmp;
2583                                 goto unwritable_page;
2584                         }
2585                         if (unlikely(!(tmp & VM_FAULT_LOCKED))) {
2586                                 lock_page(old_page);
2587                                 if (!old_page->mapping) {
2588                                         ret = 0; /* retry the fault */
2589                                         unlock_page(old_page);
2590                                         goto unwritable_page;
2591                                 }
2592                         } else
2593                                 VM_BUG_ON(!PageLocked(old_page));
2594
2595                         /*
2596                          * Since we dropped the lock we need to revalidate
2597                          * the PTE as someone else may have changed it.  If
2598                          * they did, we just return, as we can count on the
2599                          * MMU to tell us if they didn't also make it writable.
2600                          */
2601                         page_table = pte_offset_map_lock(mm, pmd, address,
2602                                                          &ptl);
2603                         if (!pte_same(*page_table, orig_pte)) {
2604                                 unlock_page(old_page);
2605                                 goto unlock;
2606                         }
2607
2608                         page_mkwrite = 1;
2609                 }
2610                 dirty_page = old_page;
2611                 get_page(dirty_page);
2612
2613 reuse:
2614                 flush_cache_page(vma, address, pte_pfn(orig_pte));
2615                 entry = pte_mkyoung(orig_pte);
2616                 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2617                 if (ptep_set_access_flags(vma, address, page_table, entry,1))
2618                         update_mmu_cache(vma, address, page_table);
2619                 pte_unmap_unlock(page_table, ptl);
2620                 ret |= VM_FAULT_WRITE;
2621
2622                 if (!dirty_page)
2623                         return ret;
2624
2625                 /*
2626                  * Yes, Virginia, this is actually required to prevent a race
2627                  * with clear_page_dirty_for_io() from clearing the page dirty
2628                  * bit after it clear all dirty ptes, but before a racing
2629                  * do_wp_page installs a dirty pte.
2630                  *
2631                  * __do_fault is protected similarly.
2632                  */
2633                 if (!page_mkwrite) {
2634                         wait_on_page_locked(dirty_page);
2635                         set_page_dirty_balance(dirty_page, page_mkwrite);
2636                 }
2637                 put_page(dirty_page);
2638                 if (page_mkwrite) {
2639                         struct address_space *mapping = dirty_page->mapping;
2640
2641                         set_page_dirty(dirty_page);
2642                         unlock_page(dirty_page);
2643                         page_cache_release(dirty_page);
2644                         if (mapping)    {
2645                                 /*
2646                                  * Some device drivers do not set page.mapping
2647                                  * but still dirty their pages
2648                                  */
2649                                 balance_dirty_pages_ratelimited(mapping);
2650                         }
2651                 }
2652
2653                 /* file_update_time outside page_lock */
2654                 if (vma->vm_file)
2655                         file_update_time(vma->vm_file);
2656
2657                 return ret;
2658         }
2659
2660         /*
2661          * Ok, we need to copy. Oh, well..
2662          */
2663         page_cache_get(old_page);
2664 gotten:
2665         pte_unmap_unlock(page_table, ptl);
2666
2667         if (unlikely(anon_vma_prepare(vma)))
2668                 goto oom;
2669
2670         if (is_zero_pfn(pte_pfn(orig_pte))) {
2671                 new_page = alloc_zeroed_user_highpage_movable(vma, address);
2672                 if (!new_page)
2673                         goto oom;
2674         } else {
2675                 new_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, address);
2676                 if (!new_page)
2677                         goto oom;
2678                 cow_user_page(new_page, old_page, address, vma);
2679         }
2680         __SetPageUptodate(new_page);
2681
2682         if (mem_cgroup_newpage_charge(new_page, mm, GFP_KERNEL))
2683                 goto oom_free_new;
2684
2685         /*
2686          * Re-check the pte - we dropped the lock
2687          */
2688         page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2689         if (likely(pte_same(*page_table, orig_pte))) {
2690                 if (old_page) {
2691                         if (!PageAnon(old_page)) {
2692                                 dec_mm_counter_fast(mm, MM_FILEPAGES);
2693                                 inc_mm_counter_fast(mm, MM_ANONPAGES);
2694                         }
2695                 } else
2696                         inc_mm_counter_fast(mm, MM_ANONPAGES);
2697                 flush_cache_page(vma, address, pte_pfn(orig_pte));
2698                 entry = mk_pte(new_page, vma->vm_page_prot);
2699                 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2700                 /*
2701                  * Clear the pte entry and flush it first, before updating the
2702                  * pte with the new entry. This will avoid a race condition
2703                  * seen in the presence of one thread doing SMC and another
2704                  * thread doing COW.
2705                  */
2706                 ptep_clear_flush(vma, address, page_table);
2707                 page_add_new_anon_rmap(new_page, vma, address);
2708                 /*
2709                  * We call the notify macro here because, when using secondary
2710                  * mmu page tables (such as kvm shadow page tables), we want the
2711                  * new page to be mapped directly into the secondary page table.
2712                  */
2713                 set_pte_at_notify(mm, address, page_table, entry);
2714                 update_mmu_cache(vma, address, page_table);
2715                 if (old_page) {
2716                         /*
2717                          * Only after switching the pte to the new page may
2718                          * we remove the mapcount here. Otherwise another
2719                          * process may come and find the rmap count decremented
2720                          * before the pte is switched to the new page, and
2721                          * "reuse" the old page writing into it while our pte
2722                          * here still points into it and can be read by other
2723                          * threads.
2724                          *
2725                          * The critical issue is to order this
2726                          * page_remove_rmap with the ptp_clear_flush above.
2727                          * Those stores are ordered by (if nothing else,)
2728                          * the barrier present in the atomic_add_negative
2729                          * in page_remove_rmap.
2730                          *
2731                          * Then the TLB flush in ptep_clear_flush ensures that
2732                          * no process can access the old page before the
2733                          * decremented mapcount is visible. And the old page
2734                          * cannot be reused until after the decremented
2735                          * mapcount is visible. So transitively, TLBs to
2736                          * old page will be flushed before it can be reused.
2737                          */
2738                         page_remove_rmap(old_page);
2739                 }
2740
2741                 /* Free the old page.. */
2742                 new_page = old_page;
2743                 ret |= VM_FAULT_WRITE;
2744         } else
2745                 mem_cgroup_uncharge_page(new_page);
2746
2747         if (new_page)
2748                 page_cache_release(new_page);
2749 unlock:
2750         pte_unmap_unlock(page_table, ptl);
2751         if (old_page) {
2752                 /*
2753                  * Don't let another task, with possibly unlocked vma,
2754                  * keep the mlocked page.
2755                  */
2756                 if ((ret & VM_FAULT_WRITE) && (vma->vm_flags & VM_LOCKED)) {
2757                         lock_page(old_page);    /* LRU manipulation */
2758                         munlock_vma_page(old_page);
2759                         unlock_page(old_page);
2760                 }
2761                 page_cache_release(old_page);
2762         }
2763         return ret;
2764 oom_free_new:
2765         page_cache_release(new_page);
2766 oom:
2767         if (old_page) {
2768                 if (page_mkwrite) {
2769                         unlock_page(old_page);
2770                         page_cache_release(old_page);
2771                 }
2772                 page_cache_release(old_page);
2773         }
2774         return VM_FAULT_OOM;
2775
2776 unwritable_page:
2777         page_cache_release(old_page);
2778         return ret;
2779 }
2780
2781 static void unmap_mapping_range_vma(struct vm_area_struct *vma,
2782                 unsigned long start_addr, unsigned long end_addr,
2783                 struct zap_details *details)
2784 {
2785         zap_page_range_single(vma, start_addr, end_addr - start_addr, details);
2786 }
2787
2788 static inline void unmap_mapping_range_tree(struct prio_tree_root *root,
2789                                             struct zap_details *details)
2790 {
2791         struct vm_area_struct *vma;
2792         struct prio_tree_iter iter;
2793         pgoff_t vba, vea, zba, zea;
2794
2795         vma_prio_tree_foreach(vma, &iter, root,
2796                         details->first_index, details->last_index) {
2797
2798                 vba = vma->vm_pgoff;
2799                 vea = vba + ((vma->vm_end - vma->vm_start) >> PAGE_SHIFT) - 1;
2800                 /* Assume for now that PAGE_CACHE_SHIFT == PAGE_SHIFT */
2801                 zba = details->first_index;
2802                 if (zba < vba)
2803                         zba = vba;
2804                 zea = details->last_index;
2805                 if (zea > vea)
2806                         zea = vea;
2807
2808                 unmap_mapping_range_vma(vma,
2809                         ((zba - vba) << PAGE_SHIFT) + vma->vm_start,
2810                         ((zea - vba + 1) << PAGE_SHIFT) + vma->vm_start,
2811                                 details);
2812         }
2813 }
2814
2815 static inline void unmap_mapping_range_list(struct list_head *head,
2816                                             struct zap_details *details)
2817 {
2818         struct vm_area_struct *vma;
2819
2820         /*
2821          * In nonlinear VMAs there is no correspondence between virtual address
2822          * offset and file offset.  So we must perform an exhaustive search
2823          * across *all* the pages in each nonlinear VMA, not just the pages
2824          * whose virtual address lies outside the file truncation point.
2825          */
2826         list_for_each_entry(vma, head, shared.vm_set.list) {
2827                 details->nonlinear_vma = vma;
2828                 unmap_mapping_range_vma(vma, vma->vm_start, vma->vm_end, details);
2829         }
2830 }
2831
2832 /**
2833  * unmap_mapping_range - unmap the portion of all mmaps in the specified address_space corresponding to the specified page range in the underlying file.
2834  * @mapping: the address space containing mmaps to be unmapped.
2835  * @holebegin: byte in first page to unmap, relative to the start of
2836  * the underlying file.  This will be rounded down to a PAGE_SIZE
2837  * boundary.  Note that this is different from truncate_pagecache(), which
2838  * must keep the partial page.  In contrast, we must get rid of
2839  * partial pages.
2840  * @holelen: size of prospective hole in bytes.  This will be rounded
2841  * up to a PAGE_SIZE boundary.  A holelen of zero truncates to the
2842  * end of the file.
2843  * @even_cows: 1 when truncating a file, unmap even private COWed pages;
2844  * but 0 when invalidating pagecache, don't throw away private data.
2845  */
2846 void unmap_mapping_range(struct address_space *mapping,
2847                 loff_t const holebegin, loff_t const holelen, int even_cows)
2848 {
2849         struct zap_details details;
2850         pgoff_t hba = holebegin >> PAGE_SHIFT;
2851         pgoff_t hlen = (holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
2852
2853         /* Check for overflow. */
2854         if (sizeof(holelen) > sizeof(hlen)) {
2855                 long long holeend =
2856                         (holebegin + holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
2857                 if (holeend & ~(long long)ULONG_MAX)
2858                         hlen = ULONG_MAX - hba + 1;
2859         }
2860
2861         details.check_mapping = even_cows? NULL: mapping;
2862         details.nonlinear_vma = NULL;
2863         details.first_index = hba;
2864         details.last_index = hba + hlen - 1;
2865         if (details.last_index < details.first_index)
2866                 details.last_index = ULONG_MAX;
2867
2868
2869         mutex_lock(&mapping->i_mmap_mutex);
2870         if (unlikely(!prio_tree_empty(&mapping->i_mmap)))
2871                 unmap_mapping_range_tree(&mapping->i_mmap, &details);
2872         if (unlikely(!list_empty(&mapping->i_mmap_nonlinear)))
2873                 unmap_mapping_range_list(&mapping->i_mmap_nonlinear, &details);
2874         mutex_unlock(&mapping->i_mmap_mutex);
2875 }
2876 EXPORT_SYMBOL(unmap_mapping_range);
2877
2878 /*
2879  * We enter with non-exclusive mmap_sem (to exclude vma changes,
2880  * but allow concurrent faults), and pte mapped but not yet locked.
2881  * We return with mmap_sem still held, but pte unmapped and unlocked.
2882  */
2883 static int do_swap_page(struct mm_struct *mm, struct vm_area_struct *vma,
2884                 unsigned long address, pte_t *page_table, pmd_t *pmd,
2885                 unsigned int flags, pte_t orig_pte)
2886 {
2887         spinlock_t *ptl;
2888         struct page *page, *swapcache = NULL;
2889         swp_entry_t entry;
2890         pte_t pte;
2891         int locked;
2892         struct mem_cgroup *ptr;
2893         int exclusive = 0;
2894         int ret = 0;
2895
2896         if (!pte_unmap_same(mm, pmd, page_table, orig_pte))
2897                 goto out;
2898
2899         entry = pte_to_swp_entry(orig_pte);
2900         if (unlikely(non_swap_entry(entry))) {
2901                 if (is_migration_entry(entry)) {
2902                         migration_entry_wait(mm, pmd, address);
2903                 } else if (is_hwpoison_entry(entry)) {
2904                         ret = VM_FAULT_HWPOISON;
2905                 } else {
2906                         print_bad_pte(vma, address, orig_pte, NULL);
2907                         ret = VM_FAULT_SIGBUS;
2908                 }
2909                 goto out;
2910         }
2911         delayacct_set_flag(DELAYACCT_PF_SWAPIN);
2912         page = lookup_swap_cache(entry);
2913         if (!page) {
2914                 grab_swap_token(mm); /* Contend for token _before_ read-in */
2915                 page = swapin_readahead(entry,
2916                                         GFP_HIGHUSER_MOVABLE, vma, address);
2917                 if (!page) {
2918                         /*
2919                          * Back out if somebody else faulted in this pte
2920                          * while we released the pte lock.
2921                          */
2922                         page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2923                         if (likely(pte_same(*page_table, orig_pte)))
2924                                 ret = VM_FAULT_OOM;
2925                         delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
2926                         goto unlock;
2927                 }
2928
2929                 /* Had to read the page from swap area: Major fault */
2930                 ret = VM_FAULT_MAJOR;
2931                 count_vm_event(PGMAJFAULT);
2932                 mem_cgroup_count_vm_event(mm, PGMAJFAULT);
2933         } else if (PageHWPoison(page)) {
2934                 /*
2935                  * hwpoisoned dirty swapcache pages are kept for killing
2936                  * owner processes (which may be unknown at hwpoison time)
2937                  */
2938                 ret = VM_FAULT_HWPOISON;
2939                 delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
2940                 goto out_release;
2941         }
2942
2943         locked = lock_page_or_retry(page, mm, flags);
2944         delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
2945         if (!locked) {
2946                 ret |= VM_FAULT_RETRY;
2947                 goto out_release;
2948         }
2949
2950         /*
2951          * Make sure try_to_free_swap or reuse_swap_page or swapoff did not
2952          * release the swapcache from under us.  The page pin, and pte_same
2953          * test below, are not enough to exclude that.  Even if it is still
2954          * swapcache, we need to check that the page's swap has not changed.
2955          */
2956         if (unlikely(!PageSwapCache(page) || page_private(page) != entry.val))
2957                 goto out_page;
2958
2959         if (ksm_might_need_to_copy(page, vma, address)) {
2960                 swapcache = page;
2961                 page = ksm_does_need_to_copy(page, vma, address);
2962
2963                 if (unlikely(!page)) {
2964                         ret = VM_FAULT_OOM;
2965                         page = swapcache;
2966                         swapcache = NULL;
2967                         goto out_page;
2968                 }
2969         }
2970
2971         if (mem_cgroup_try_charge_swapin(mm, page, GFP_KERNEL, &ptr)) {
2972                 ret = VM_FAULT_OOM;
2973                 goto out_page;
2974         }
2975
2976         /*
2977          * Back out if somebody else already faulted in this pte.
2978          */
2979         page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2980         if (unlikely(!pte_same(*page_table, orig_pte)))
2981                 goto out_nomap;
2982
2983         if (unlikely(!PageUptodate(page))) {
2984                 ret = VM_FAULT_SIGBUS;
2985                 goto out_nomap;
2986         }
2987
2988         /*
2989          * The page isn't present yet, go ahead with the fault.
2990          *
2991          * Be careful about the sequence of operations here.
2992          * To get its accounting right, reuse_swap_page() must be called
2993          * while the page is counted on swap but not yet in mapcount i.e.
2994          * before page_add_anon_rmap() and swap_free(); try_to_free_swap()
2995          * must be called after the swap_free(), or it will never succeed.
2996          * Because delete_from_swap_page() may be called by reuse_swap_page(),
2997          * mem_cgroup_commit_charge_swapin() may not be able to find swp_entry
2998          * in page->private. In this case, a record in swap_cgroup  is silently
2999          * discarded at swap_free().
3000          */
3001
3002         inc_mm_counter_fast(mm, MM_ANONPAGES);
3003         dec_mm_counter_fast(mm, MM_SWAPENTS);
3004         pte = mk_pte(page, vma->vm_page_prot);
3005         if ((flags & FAULT_FLAG_WRITE) && reuse_swap_page(page)) {
3006                 pte = maybe_mkwrite(pte_mkdirty(pte), vma);
3007                 flags &= ~FAULT_FLAG_WRITE;
3008                 ret |= VM_FAULT_WRITE;
3009                 exclusive = 1;
3010         }
3011         flush_icache_page(vma, page);
3012         set_pte_at(mm, address, page_table, pte);
3013         do_page_add_anon_rmap(page, vma, address, exclusive);
3014         /* It's better to call commit-charge after rmap is established */
3015         mem_cgroup_commit_charge_swapin(page, ptr);
3016
3017         swap_free(entry);
3018         if (vm_swap_full() || (vma->vm_flags & VM_LOCKED) || PageMlocked(page))
3019                 try_to_free_swap(page);
3020         unlock_page(page);
3021         if (swapcache) {
3022                 /*
3023                  * Hold the lock to avoid the swap entry to be reused
3024                  * until we take the PT lock for the pte_same() check
3025                  * (to avoid false positives from pte_same). For
3026                  * further safety release the lock after the swap_free
3027                  * so that the swap count won't change under a
3028                  * parallel locked swapcache.
3029                  */
3030                 unlock_page(swapcache);
3031                 page_cache_release(swapcache);
3032         }
3033
3034         if (flags & FAULT_FLAG_WRITE) {
3035                 ret |= do_wp_page(mm, vma, address, page_table, pmd, ptl, pte);
3036                 if (ret & VM_FAULT_ERROR)
3037                         ret &= VM_FAULT_ERROR;
3038                 goto out;
3039         }
3040
3041         /* No need to invalidate - it was non-present before */
3042         update_mmu_cache(vma, address, page_table);
3043 unlock:
3044         pte_unmap_unlock(page_table, ptl);
3045 out:
3046         return ret;
3047 out_nomap:
3048         mem_cgroup_cancel_charge_swapin(ptr);
3049         pte_unmap_unlock(page_table, ptl);
3050 out_page:
3051         unlock_page(page);
3052 out_release:
3053         page_cache_release(page);
3054         if (swapcache) {
3055                 unlock_page(swapcache);
3056                 page_cache_release(swapcache);
3057         }
3058         return ret;
3059 }
3060
3061 /*
3062  * This is like a special single-page "expand_{down|up}wards()",
3063  * except we must first make sure that 'address{-|+}PAGE_SIZE'
3064  * doesn't hit another vma.
3065  */
3066 static inline int check_stack_guard_page(struct vm_area_struct *vma, unsigned long address)
3067 {
3068         address &= PAGE_MASK;
3069         if ((vma->vm_flags & VM_GROWSDOWN) && address == vma->vm_start) {
3070                 struct vm_area_struct *prev = vma->vm_prev;
3071
3072                 /*
3073                  * Is there a mapping abutting this one below?
3074                  *
3075                  * That's only ok if it's the same stack mapping
3076                  * that has gotten split..
3077                  */
3078                 if (prev && prev->vm_end == address)
3079                         return prev->vm_flags & VM_GROWSDOWN ? 0 : -ENOMEM;
3080
3081                 expand_downwards(vma, address - PAGE_SIZE);
3082         }
3083         if ((vma->vm_flags & VM_GROWSUP) && address + PAGE_SIZE == vma->vm_end) {
3084                 struct vm_area_struct *next = vma->vm_next;
3085
3086                 /* As VM_GROWSDOWN but s/below/above/ */
3087                 if (next && next->vm_start == address + PAGE_SIZE)
3088                         return next->vm_flags & VM_GROWSUP ? 0 : -ENOMEM;
3089
3090                 expand_upwards(vma, address + PAGE_SIZE);
3091         }
3092         return 0;
3093 }
3094
3095 /*
3096  * We enter with non-exclusive mmap_sem (to exclude vma changes,
3097  * but allow concurrent faults), and pte mapped but not yet locked.
3098  * We return with mmap_sem still held, but pte unmapped and unlocked.
3099  */
3100 static int do_anonymous_page(struct mm_struct *mm, struct vm_area_struct *vma,
3101                 unsigned long address, pte_t *page_table, pmd_t *pmd,
3102                 unsigned int flags)
3103 {
3104         struct page *page;
3105         spinlock_t *ptl;
3106         pte_t entry;
3107
3108         pte_unmap(page_table);
3109
3110         /* Check if we need to add a guard page to the stack */
3111         if (check_stack_guard_page(vma, address) < 0)
3112                 return VM_FAULT_SIGBUS;
3113
3114         /* Use the zero-page for reads */
3115         if (!(flags & FAULT_FLAG_WRITE)) {
3116                 entry = pte_mkspecial(pfn_pte(my_zero_pfn(address),
3117                                                 vma->vm_page_prot));
3118                 page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
3119                 if (!pte_none(*page_table))
3120                         goto unlock;
3121                 goto setpte;
3122         }
3123
3124         /* Allocate our own private page. */
3125         if (unlikely(anon_vma_prepare(vma)))
3126                 goto oom;
3127         page = alloc_zeroed_user_highpage_movable(vma, address);
3128         if (!page)
3129                 goto oom;
3130         __SetPageUptodate(page);
3131
3132         if (mem_cgroup_newpage_charge(page, mm, GFP_KERNEL))
3133                 goto oom_free_page;
3134
3135         entry = mk_pte(page, vma->vm_page_prot);
3136         if (vma->vm_flags & VM_WRITE)
3137                 entry = pte_mkwrite(pte_mkdirty(entry));
3138
3139         page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
3140         if (!pte_none(*page_table))
3141                 goto release;
3142
3143         inc_mm_counter_fast(mm, MM_ANONPAGES);
3144         page_add_new_anon_rmap(page, vma, address);
3145 setpte:
3146         set_pte_at(mm, address, page_table, entry);
3147
3148         /* No need to invalidate - it was non-present before */
3149         update_mmu_cache(vma, address, page_table);
3150 unlock:
3151         pte_unmap_unlock(page_table, ptl);
3152         return 0;
3153 release:
3154         mem_cgroup_uncharge_page(page);
3155         page_cache_release(page);
3156         goto unlock;
3157 oom_free_page:
3158         page_cache_release(page);
3159 oom:
3160         return VM_FAULT_OOM;
3161 }
3162
3163 /*
3164  * __do_fault() tries to create a new page mapping. It aggressively
3165  * tries to share with existing pages, but makes a separate copy if
3166  * the FAULT_FLAG_WRITE is set in the flags parameter in order to avoid
3167  * the next page fault.
3168  *
3169  * As this is called only for pages that do not currently exist, we
3170  * do not need to flush old virtual caches or the TLB.
3171  *
3172  * We enter with non-exclusive mmap_sem (to exclude vma changes,
3173  * but allow concurrent faults), and pte neither mapped nor locked.
3174  * We return with mmap_sem still held, but pte unmapped and unlocked.
3175  */
3176 static int __do_fault(struct mm_struct *mm, struct vm_area_struct *vma,
3177                 unsigned long address, pmd_t *pmd,
3178                 pgoff_t pgoff, unsigned int flags, pte_t orig_pte)
3179 {
3180         pte_t *page_table;
3181         spinlock_t *ptl;
3182         struct page *page;
3183         struct page *cow_page;
3184         pte_t entry;
3185         int anon = 0;
3186         struct page *dirty_page = NULL;
3187         struct vm_fault vmf;
3188         int ret;
3189         int page_mkwrite = 0;
3190
3191         /*
3192          * If we do COW later, allocate page befor taking lock_page()
3193          * on the file cache page. This will reduce lock holding time.
3194          */
3195         if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) {
3196
3197                 if (unlikely(anon_vma_prepare(vma)))
3198                         return VM_FAULT_OOM;
3199
3200                 cow_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, address);
3201                 if (!cow_page)
3202                         return VM_FAULT_OOM;
3203
3204                 if (mem_cgroup_newpage_charge(cow_page, mm, GFP_KERNEL)) {
3205                         page_cache_release(cow_page);
3206                         return VM_FAULT_OOM;
3207                 }
3208         } else
3209                 cow_page = NULL;
3210
3211         vmf.virtual_address = (void __user *)(address & PAGE_MASK);
3212         vmf.pgoff = pgoff;
3213         vmf.flags = flags;
3214         vmf.page = NULL;
3215
3216         ret = vma->vm_ops->fault(vma, &vmf);
3217         if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE |
3218                             VM_FAULT_RETRY)))
3219                 goto uncharge_out;
3220
3221         if (unlikely(PageHWPoison(vmf.page))) {
3222                 if (ret & VM_FAULT_LOCKED)
3223                         unlock_page(vmf.page);
3224                 ret = VM_FAULT_HWPOISON;
3225                 goto uncharge_out;
3226         }
3227
3228         /*
3229          * For consistency in subsequent calls, make the faulted page always
3230          * locked.
3231          */
3232         if (unlikely(!(ret & VM_FAULT_LOCKED)))
3233                 lock_page(vmf.page);
3234         else
3235                 VM_BUG_ON(!PageLocked(vmf.page));
3236
3237         /*
3238          * Should we do an early C-O-W break?
3239          */
3240         page = vmf.page;
3241         if (flags & FAULT_FLAG_WRITE) {
3242                 if (!(vma->vm_flags & VM_SHARED)) {
3243                         page = cow_page;
3244                         anon = 1;
3245                         copy_user_highpage(page, vmf.page, address, vma);
3246                         __SetPageUptodate(page);
3247                 } else {
3248                         /*
3249                          * If the page will be shareable, see if the backing
3250                          * address space wants to know that the page is about
3251                          * to become writable
3252                          */
3253                         if (vma->vm_ops->page_mkwrite) {
3254                                 int tmp;
3255
3256                                 unlock_page(page);
3257                                 vmf.flags = FAULT_FLAG_WRITE|FAULT_FLAG_MKWRITE;
3258                                 tmp = vma->vm_ops->page_mkwrite(vma, &vmf);
3259                                 if (unlikely(tmp &
3260                                           (VM_FAULT_ERROR | VM_FAULT_NOPAGE))) {
3261                                         ret = tmp;
3262                                         goto unwritable_page;
3263                                 }
3264                                 if (unlikely(!(tmp & VM_FAULT_LOCKED))) {
3265                                         lock_page(page);
3266                                         if (!page->mapping) {
3267                                                 ret = 0; /* retry the fault */
3268                                                 unlock_page(page);
3269                                                 goto unwritable_page;
3270                                         }
3271                                 } else
3272                                         VM_BUG_ON(!PageLocked(page));
3273                                 page_mkwrite = 1;
3274                         }
3275                 }
3276
3277         }
3278
3279         page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
3280
3281         /*
3282          * This silly early PAGE_DIRTY setting removes a race
3283          * due to the bad i386 page protection. But it's valid
3284          * for other architectures too.
3285          *
3286          * Note that if FAULT_FLAG_WRITE is set, we either now have
3287          * an exclusive copy of the page, or this is a shared mapping,
3288          * so we can make it writable and dirty to avoid having to
3289          * handle that later.
3290          */
3291         /* Only go through if we didn't race with anybody else... */
3292         if (likely(pte_same(*page_table, orig_pte))) {
3293                 flush_icache_page(vma, page);
3294                 entry = mk_pte(page, vma->vm_page_prot);
3295                 if (flags & FAULT_FLAG_WRITE)
3296                         entry = maybe_mkwrite(pte_mkdirty(entry), vma);
3297                 if (anon) {
3298                         inc_mm_counter_fast(mm, MM_ANONPAGES);
3299                         page_add_new_anon_rmap(page, vma, address);
3300                 } else {
3301                         inc_mm_counter_fast(mm, MM_FILEPAGES);
3302                         page_add_file_rmap(page);
3303                         if (flags & FAULT_FLAG_WRITE) {
3304                                 dirty_page = page;
3305                                 get_page(dirty_page);
3306                         }
3307                 }
3308                 set_pte_at(mm, address, page_table, entry);
3309
3310                 /* no need to invalidate: a not-present page won't be cached */
3311                 update_mmu_cache(vma, address, page_table);
3312         } else {
3313                 if (cow_page)
3314                         mem_cgroup_uncharge_page(cow_page);
3315                 if (anon)
3316                         page_cache_release(page);
3317                 else
3318                         anon = 1; /* no anon but release faulted_page */
3319         }
3320
3321         pte_unmap_unlock(page_table, ptl);
3322
3323         if (dirty_page) {
3324                 struct address_space *mapping = page->mapping;
3325
3326                 if (set_page_dirty(dirty_page))
3327                         page_mkwrite = 1;
3328                 unlock_page(dirty_page);
3329                 put_page(dirty_page);
3330                 if (page_mkwrite && mapping) {
3331                         /*
3332                          * Some device drivers do not set page.mapping but still
3333                          * dirty their pages
3334                          */
3335                         balance_dirty_pages_ratelimited(mapping);
3336                 }
3337
3338                 /* file_update_time outside page_lock */
3339                 if (vma->vm_file)
3340                         file_update_time(vma->vm_file);
3341         } else {
3342                 unlock_page(vmf.page);
3343                 if (anon)
3344                         page_cache_release(vmf.page);
3345         }
3346
3347         return ret;
3348
3349 unwritable_page:
3350         page_cache_release(page);
3351         return ret;
3352 uncharge_out:
3353         /* fs's fault handler get error */
3354         if (cow_page) {
3355                 mem_cgroup_uncharge_page(cow_page);
3356                 page_cache_release(cow_page);
3357         }
3358         return ret;
3359 }
3360
3361 static int do_linear_fault(struct mm_struct *mm, struct vm_area_struct *vma,
3362                 unsigned long address, pte_t *page_table, pmd_t *pmd,
3363                 unsigned int flags, pte_t orig_pte)
3364 {
3365         pgoff_t pgoff = (((address & PAGE_MASK)
3366                         - vma->vm_start) >> PAGE_SHIFT) + vma->vm_pgoff;
3367
3368         pte_unmap(page_table);
3369         return __do_fault(mm, vma, address, pmd, pgoff, flags, orig_pte);
3370 }
3371
3372 /*
3373  * Fault of a previously existing named mapping. Repopulate the pte
3374  * from the encoded file_pte if possible. This enables swappable
3375  * nonlinear vmas.
3376  *
3377  * We enter with non-exclusive mmap_sem (to exclude vma changes,
3378  * but allow concurrent faults), and pte mapped but not yet locked.
3379  * We return with mmap_sem still held, but pte unmapped and unlocked.
3380  */
3381 static int do_nonlinear_fault(struct mm_struct *mm, struct vm_area_struct *vma,
3382                 unsigned long address, pte_t *page_table, pmd_t *pmd,
3383                 unsigned int flags, pte_t orig_pte)
3384 {
3385         pgoff_t pgoff;
3386
3387         flags |= FAULT_FLAG_NONLINEAR;
3388
3389         if (!pte_unmap_same(mm, pmd, page_table, orig_pte))
3390                 return 0;
3391
3392         if (unlikely(!(vma->vm_flags & VM_NONLINEAR))) {
3393                 /*
3394                  * Page table corrupted: show pte and kill process.
3395                  */
3396                 print_bad_pte(vma, address, orig_pte, NULL);
3397                 return VM_FAULT_SIGBUS;
3398         }
3399
3400         pgoff = pte_to_pgoff(orig_pte);
3401         return __do_fault(mm, vma, address, pmd, pgoff, flags, orig_pte);
3402 }
3403
3404 /*
3405  * These routines also need to handle stuff like marking pages dirty
3406  * and/or accessed for architectures that don't do it in hardware (most
3407  * RISC architectures).  The early dirtying is also good on the i386.
3408  *
3409  * There is also a hook called "update_mmu_cache()" that architectures
3410  * with external mmu caches can use to update those (ie the Sparc or
3411  * PowerPC hashed page tables that act as extended TLBs).
3412  *
3413  * We enter with non-exclusive mmap_sem (to exclude vma changes,
3414  * but allow concurrent faults), and pte mapped but not yet locked.
3415  * We return with mmap_sem still held, but pte unmapped and unlocked.
3416  */
3417 int handle_pte_fault(struct mm_struct *mm,
3418                      struct vm_area_struct *vma, unsigned long address,
3419                      pte_t *pte, pmd_t *pmd, unsigned int flags)
3420 {
3421         pte_t entry;
3422         spinlock_t *ptl;
3423
3424         entry = *pte;
3425         if (!pte_present(entry)) {
3426                 if (pte_none(entry)) {
3427                         if (vma->vm_ops) {
3428                                 if (likely(vma->vm_ops->fault))
3429                                         return do_linear_fault(mm, vma, address,
3430                                                 pte, pmd, flags, entry);
3431                         }
3432                         return do_anonymous_page(mm, vma, address,
3433                                                  pte, pmd, flags);
3434                 }
3435                 if (pte_file(entry))
3436                         return do_nonlinear_fault(mm, vma, address,
3437                                         pte, pmd, flags, entry);
3438                 return do_swap_page(mm, vma, address,
3439                                         pte, pmd, flags, entry);
3440         }
3441
3442         ptl = pte_lockptr(mm, pmd);
3443         spin_lock(ptl);
3444         if (unlikely(!pte_same(*pte, entry)))
3445                 goto unlock;
3446         if (flags & FAULT_FLAG_WRITE) {
3447                 if (!pte_write(entry))
3448                         return do_wp_page(mm, vma, address,
3449                                         pte, pmd, ptl, entry);
3450                 entry = pte_mkdirty(entry);
3451         }
3452         entry = pte_mkyoung(entry);
3453         if (ptep_set_access_flags(vma, address, pte, entry, flags & FAULT_FLAG_WRITE)) {
3454                 update_mmu_cache(vma, address, pte);
3455         } else {
3456                 /*
3457                  * This is needed only for protection faults but the arch code
3458                  * is not yet telling us if this is a protection fault or not.
3459                  * This still avoids useless tlb flushes for .text page faults
3460                  * with threads.
3461                  */
3462                 if (flags & FAULT_FLAG_WRITE)
3463                         flush_tlb_fix_spurious_fault(vma, address);
3464         }
3465 unlock:
3466         pte_unmap_unlock(pte, ptl);
3467         return 0;
3468 }
3469
3470 /*
3471  * By the time we get here, we already hold the mm semaphore
3472  */
3473 int handle_mm_fault(struct mm_struct *mm, struct vm_area_struct *vma,
3474                 unsigned long address, unsigned int flags)
3475 {
3476         pgd_t *pgd;
3477         pud_t *pud;
3478         pmd_t *pmd;
3479         pte_t *pte;
3480
3481         __set_current_state(TASK_RUNNING);
3482
3483         count_vm_event(PGFAULT);
3484         mem_cgroup_count_vm_event(mm, PGFAULT);
3485
3486         /* do counter updates before entering really critical section. */
3487         check_sync_rss_stat(current);
3488
3489         if (unlikely(is_vm_hugetlb_page(vma)))
3490                 return hugetlb_fault(mm, vma, address, flags);
3491
3492         pgd = pgd_offset(mm, address);
3493         pud = pud_alloc(mm, pgd, address);
3494         if (!pud)
3495                 return VM_FAULT_OOM;
3496         pmd = pmd_alloc(mm, pud, address);
3497         if (!pmd)
3498                 return VM_FAULT_OOM;
3499         if (pmd_none(*pmd) && transparent_hugepage_enabled(vma)) {
3500                 if (!vma->vm_ops)
3501                         return do_huge_pmd_anonymous_page(mm, vma, address,
3502                                                           pmd, flags);
3503         } else {
3504                 pmd_t orig_pmd = *pmd;
3505                 barrier();
3506                 if (pmd_trans_huge(orig_pmd)) {
3507                         if (flags & FAULT_FLAG_WRITE &&
3508                             !pmd_write(orig_pmd) &&
3509                             !pmd_trans_splitting(orig_pmd))
3510                                 return do_huge_pmd_wp_page(mm, vma, address,
3511                                                            pmd, orig_pmd);
3512                         return 0;
3513                 }
3514         }
3515
3516         /*
3517          * Use __pte_alloc instead of pte_alloc_map, because we can't
3518          * run pte_offset_map on the pmd, if an huge pmd could
3519          * materialize from under us from a different thread.
3520          */
3521         if (unlikely(pmd_none(*pmd)) && __pte_alloc(mm, vma, pmd, address))
3522                 return VM_FAULT_OOM;
3523         /* if an huge pmd materialized from under us just retry later */
3524         if (unlikely(pmd_trans_huge(*pmd)))
3525                 return 0;
3526         /*
3527          * A regular pmd is established and it can't morph into a huge pmd
3528          * from under us anymore at this point because we hold the mmap_sem
3529          * read mode and khugepaged takes it in write mode. So now it's
3530          * safe to run pte_offset_map().
3531          */
3532         pte = pte_offset_map(pmd, address);
3533
3534         return handle_pte_fault(mm, vma, address, pte, pmd, flags);
3535 }
3536
3537 #ifndef __PAGETABLE_PUD_FOLDED
3538 /*
3539  * Allocate page upper directory.
3540  * We've already handled the fast-path in-line.
3541  */
3542 int __pud_alloc(struct mm_struct *mm, pgd_t *pgd, unsigned long address)
3543 {
3544         pud_t *new = pud_alloc_one(mm, address);
3545         if (!new)
3546                 return -ENOMEM;
3547
3548         smp_wmb(); /* See comment in __pte_alloc */
3549
3550         spin_lock(&mm->page_table_lock);
3551         if (pgd_present(*pgd))          /* Another has populated it */
3552                 pud_free(mm, new);
3553         else
3554                 pgd_populate(mm, pgd, new);
3555         spin_unlock(&mm->page_table_lock);
3556         return 0;
3557 }
3558 #endif /* __PAGETABLE_PUD_FOLDED */
3559
3560 #ifndef __PAGETABLE_PMD_FOLDED
3561 /*
3562  * Allocate page middle directory.
3563  * We've already handled the fast-path in-line.
3564  */
3565 int __pmd_alloc(struct mm_struct *mm, pud_t *pud, unsigned long address)
3566 {
3567         pmd_t *new = pmd_alloc_one(mm, address);
3568         if (!new)
3569                 return -ENOMEM;
3570
3571         smp_wmb(); /* See comment in __pte_alloc */
3572
3573         spin_lock(&mm->page_table_lock);
3574 #ifndef __ARCH_HAS_4LEVEL_HACK
3575         if (pud_present(*pud))          /* Another has populated it */
3576                 pmd_free(mm, new);
3577         else
3578                 pud_populate(mm, pud, new);
3579 #else
3580         if (pgd_present(*pud))          /* Another has populated it */
3581                 pmd_free(mm, new);
3582         else
3583                 pgd_populate(mm, pud, new);
3584 #endif /* __ARCH_HAS_4LEVEL_HACK */
3585         spin_unlock(&mm->page_table_lock);
3586         return 0;
3587 }
3588 #endif /* __PAGETABLE_PMD_FOLDED */
3589
3590 int make_pages_present(unsigned long addr, unsigned long end)
3591 {
3592         int ret, len, write;
3593         struct vm_area_struct * vma;
3594
3595         vma = find_vma(current->mm, addr);
3596         if (!vma)
3597                 return -ENOMEM;
3598         /*
3599          * We want to touch writable mappings with a write fault in order
3600          * to break COW, except for shared mappings because these don't COW
3601          * and we would not want to dirty them for nothing.
3602          */
3603         write = (vma->vm_flags & (VM_WRITE | VM_SHARED)) == VM_WRITE;
3604         BUG_ON(addr >= end);
3605         BUG_ON(end > vma->vm_end);
3606         len = DIV_ROUND_UP(end, PAGE_SIZE) - addr/PAGE_SIZE;
3607         ret = get_user_pages(current, current->mm, addr,
3608                         len, write, 0, NULL, NULL);
3609         if (ret < 0)
3610                 return ret;
3611         return ret == len ? 0 : -EFAULT;
3612 }
3613
3614 #if !defined(__HAVE_ARCH_GATE_AREA)
3615
3616 #if defined(AT_SYSINFO_EHDR)
3617 static struct vm_area_struct gate_vma;
3618
3619 static int __init gate_vma_init(void)
3620 {
3621         gate_vma.vm_mm = NULL;
3622         gate_vma.vm_start = FIXADDR_USER_START;
3623         gate_vma.vm_end = FIXADDR_USER_END;
3624         gate_vma.vm_flags = VM_READ | VM_MAYREAD | VM_EXEC | VM_MAYEXEC;
3625         gate_vma.vm_page_prot = __P101;
3626
3627         return 0;
3628 }
3629 __initcall(gate_vma_init);
3630 #endif
3631
3632 struct vm_area_struct *get_gate_vma(struct mm_struct *mm)
3633 {
3634 #ifdef AT_SYSINFO_EHDR
3635         return &gate_vma;
3636 #else
3637         return NULL;
3638 #endif
3639 }
3640
3641 int in_gate_area_no_mm(unsigned long addr)
3642 {
3643 #ifdef AT_SYSINFO_EHDR
3644         if ((addr >= FIXADDR_USER_START) && (addr < FIXADDR_USER_END))
3645                 return 1;
3646 #endif
3647         return 0;
3648 }
3649
3650 #endif  /* __HAVE_ARCH_GATE_AREA */
3651
3652 static int __follow_pte(struct mm_struct *mm, unsigned long address,
3653                 pte_t **ptepp, spinlock_t **ptlp)
3654 {
3655         pgd_t *pgd;
3656         pud_t *pud;
3657         pmd_t *pmd;
3658         pte_t *ptep;
3659
3660         pgd = pgd_offset(mm, address);
3661         if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
3662                 goto out;
3663
3664         pud = pud_offset(pgd, address);
3665         if (pud_none(*pud) || unlikely(pud_bad(*pud)))
3666                 goto out;
3667
3668         pmd = pmd_offset(pud, address);
3669         VM_BUG_ON(pmd_trans_huge(*pmd));
3670         if (pmd_none(*pmd) || unlikely(pmd_bad(*pmd)))
3671                 goto out;
3672
3673         /* We cannot handle huge page PFN maps. Luckily they don't exist. */
3674         if (pmd_huge(*pmd))
3675                 goto out;
3676
3677         ptep = pte_offset_map_lock(mm, pmd, address, ptlp);
3678         if (!ptep)
3679                 goto out;
3680         if (!pte_present(*ptep))
3681                 goto unlock;
3682         *ptepp = ptep;
3683         return 0;
3684 unlock:
3685         pte_unmap_unlock(ptep, *ptlp);
3686 out:
3687         return -EINVAL;
3688 }
3689
3690 static inline int follow_pte(struct mm_struct *mm, unsigned long address,
3691                              pte_t **ptepp, spinlock_t **ptlp)
3692 {
3693         int res;
3694
3695         /* (void) is needed to make gcc happy */
3696         (void) __cond_lock(*ptlp,
3697                            !(res = __follow_pte(mm, address, ptepp, ptlp)));
3698         return res;
3699 }
3700
3701 /**
3702  * follow_pfn - look up PFN at a user virtual address
3703  * @vma: memory mapping
3704  * @address: user virtual address
3705  * @pfn: location to store found PFN
3706  *
3707  * Only IO mappings and raw PFN mappings are allowed.
3708  *
3709  * Returns zero and the pfn at @pfn on success, -ve otherwise.
3710  */
3711 int follow_pfn(struct vm_area_struct *vma, unsigned long address,
3712         unsigned long *pfn)
3713 {
3714         int ret = -EINVAL;
3715         spinlock_t *ptl;
3716         pte_t *ptep;
3717
3718         if (!(vma->vm_flags & (VM_IO | VM_PFNMAP)))
3719                 return ret;
3720
3721         ret = follow_pte(vma->vm_mm, address, &ptep, &ptl);
3722         if (ret)
3723                 return ret;
3724         *pfn = pte_pfn(*ptep);
3725         pte_unmap_unlock(ptep, ptl);
3726         return 0;
3727 }
3728 EXPORT_SYMBOL(follow_pfn);
3729
3730 #ifdef CONFIG_HAVE_IOREMAP_PROT
3731 int follow_phys(struct vm_area_struct *vma,
3732                 unsigned long address, unsigned int flags,
3733                 unsigned long *prot, resource_size_t *phys)
3734 {
3735         int ret = -EINVAL;
3736         pte_t *ptep, pte;
3737         spinlock_t *ptl;
3738
3739         if (!(vma->vm_flags & (VM_IO | VM_PFNMAP)))
3740                 goto out;
3741
3742         if (follow_pte(vma->vm_mm, address, &ptep, &ptl))
3743                 goto out;
3744         pte = *ptep;
3745
3746         if ((flags & FOLL_WRITE) && !pte_write(pte))
3747                 goto unlock;
3748
3749         *prot = pgprot_val(pte_pgprot(pte));
3750         *phys = (resource_size_t)pte_pfn(pte) << PAGE_SHIFT;
3751
3752         ret = 0;
3753 unlock:
3754         pte_unmap_unlock(ptep, ptl);
3755 out:
3756         return ret;
3757 }
3758
3759 int generic_access_phys(struct vm_area_struct *vma, unsigned long addr,
3760                         void *buf, int len, int write)
3761 {
3762         resource_size_t phys_addr;
3763         unsigned long prot = 0;
3764         void __iomem *maddr;
3765         int offset = addr & (PAGE_SIZE-1);
3766
3767         if (follow_phys(vma, addr, write, &prot, &phys_addr))
3768                 return -EINVAL;
3769
3770         maddr = ioremap_prot(phys_addr, PAGE_SIZE, prot);
3771         if (write)
3772                 memcpy_toio(maddr + offset, buf, len);
3773         else
3774                 memcpy_fromio(buf, maddr + offset, len);
3775         iounmap(maddr);
3776
3777         return len;
3778 }
3779 #endif
3780
3781 /*
3782  * Access another process' address space as given in mm.  If non-NULL, use the
3783  * given task for page fault accounting.
3784  */
3785 static int __access_remote_vm(struct task_struct *tsk, struct mm_struct *mm,
3786                 unsigned long addr, void *buf, int len, int write)
3787 {
3788         struct vm_area_struct *vma;
3789         void *old_buf = buf;
3790
3791         down_read(&mm->mmap_sem);
3792         /* ignore errors, just check how much was successfully transferred */
3793         while (len) {
3794                 int bytes, ret, offset;
3795                 void *maddr;
3796                 struct page *page = NULL;
3797
3798                 ret = get_user_pages(tsk, mm, addr, 1,
3799                                 write, 1, &page, &vma);
3800                 if (ret <= 0) {
3801                         /*
3802                          * Check if this is a VM_IO | VM_PFNMAP VMA, which
3803                          * we can access using slightly different code.
3804                          */
3805 #ifdef CONFIG_HAVE_IOREMAP_PROT
3806                         vma = find_vma(mm, addr);
3807                         if (!vma || vma->vm_start > addr)
3808                                 break;
3809                         if (vma->vm_ops && vma->vm_ops->access)
3810                                 ret = vma->vm_ops->access(vma, addr, buf,
3811                                                           len, write);
3812                         if (ret <= 0)
3813 #endif
3814                                 break;
3815                         bytes = ret;
3816                 } else {
3817                         bytes = len;
3818                         offset = addr & (PAGE_SIZE-1);
3819                         if (bytes > PAGE_SIZE-offset)
3820                                 bytes = PAGE_SIZE-offset;
3821
3822                         maddr = kmap(page);
3823                         if (write) {
3824                                 copy_to_user_page(vma, page, addr,
3825                                                   maddr + offset, buf, bytes);
3826                                 set_page_dirty_lock(page);
3827                         } else {
3828                                 copy_from_user_page(vma, page, addr,
3829                                                     buf, maddr + offset, bytes);
3830                         }
3831                         kunmap(page);
3832                         page_cache_release(page);
3833                 }
3834                 len -= bytes;
3835                 buf += bytes;
3836                 addr += bytes;
3837         }
3838         up_read(&mm->mmap_sem);
3839
3840         return buf - old_buf;
3841 }
3842
3843 /**
3844  * access_remote_vm - access another process' address space
3845  * @mm:         the mm_struct of the target address space
3846  * @addr:       start address to access
3847  * @buf:        source or destination buffer
3848  * @len:        number of bytes to transfer
3849  * @write:      whether the access is a write
3850  *
3851  * The caller must hold a reference on @mm.
3852  */
3853 int access_remote_vm(struct mm_struct *mm, unsigned long addr,
3854                 void *buf, int len, int write)
3855 {
3856         return __access_remote_vm(NULL, mm, addr, buf, len, write);
3857 }
3858
3859 /*
3860  * Access another process' address space.
3861  * Source/target buffer must be kernel space,
3862  * Do not walk the page table directly, use get_user_pages
3863  */
3864 int access_process_vm(struct task_struct *tsk, unsigned long addr,
3865                 void *buf, int len, int write)
3866 {
3867         struct mm_struct *mm;
3868         int ret;
3869
3870         mm = get_task_mm(tsk);
3871         if (!mm)
3872                 return 0;
3873
3874         ret = __access_remote_vm(tsk, mm, addr, buf, len, write);
3875         mmput(mm);
3876
3877         return ret;
3878 }
3879
3880 /*
3881  * Print the name of a VMA.
3882  */
3883 void print_vma_addr(char *prefix, unsigned long ip)
3884 {
3885         struct mm_struct *mm = current->mm;
3886         struct vm_area_struct *vma;
3887
3888         /*
3889          * Do not print if we are in atomic
3890          * contexts (in exception stacks, etc.):
3891          */
3892         if (preempt_count())
3893                 return;
3894
3895         down_read(&mm->mmap_sem);
3896         vma = find_vma(mm, ip);
3897         if (vma && vma->vm_file) {
3898                 struct file *f = vma->vm_file;
3899                 char *buf = (char *)__get_free_page(GFP_KERNEL);
3900                 if (buf) {
3901                         char *p, *s;
3902
3903                         p = d_path(&f->f_path, buf, PAGE_SIZE);
3904                         if (IS_ERR(p))
3905                                 p = "?";
3906                         s = strrchr(p, '/');
3907                         if (s)
3908                                 p = s+1;
3909                         printk("%s%s[%lx+%lx]", prefix, p,
3910                                         vma->vm_start,
3911                                         vma->vm_end - vma->vm_start);
3912                         free_page((unsigned long)buf);
3913                 }
3914         }
3915         up_read(&current->mm->mmap_sem);
3916 }
3917
3918 #ifdef CONFIG_PROVE_LOCKING
3919 void might_fault(void)
3920 {
3921         /*
3922          * Some code (nfs/sunrpc) uses socket ops on kernel memory while
3923          * holding the mmap_sem, this is safe because kernel memory doesn't
3924          * get paged out, therefore we'll never actually fault, and the
3925          * below annotations will generate false positives.
3926          */
3927         if (segment_eq(get_fs(), KERNEL_DS))
3928                 return;
3929
3930         might_sleep();
3931         /*
3932          * it would be nicer only to annotate paths which are not under
3933          * pagefault_disable, however that requires a larger audit and
3934          * providing helpers like get_user_atomic.
3935          */
3936         if (!in_atomic() && current->mm)
3937                 might_lock_read(&current->mm->mmap_sem);
3938 }
3939 EXPORT_SYMBOL(might_fault);
3940 #endif
3941
3942 #if defined(CONFIG_TRANSPARENT_HUGEPAGE) || defined(CONFIG_HUGETLBFS)
3943 static void clear_gigantic_page(struct page *page,
3944                                 unsigned long addr,
3945                                 unsigned int pages_per_huge_page)
3946 {
3947         int i;
3948         struct page *p = page;
3949
3950         might_sleep();
3951         for (i = 0; i < pages_per_huge_page;
3952              i++, p = mem_map_next(p, page, i)) {
3953                 cond_resched();
3954                 clear_user_highpage(p, addr + i * PAGE_SIZE);
3955         }
3956 }
3957 void clear_huge_page(struct page *page,
3958                      unsigned long addr, unsigned int pages_per_huge_page)
3959 {
3960         int i;
3961
3962         if (unlikely(pages_per_huge_page > MAX_ORDER_NR_PAGES)) {
3963                 clear_gigantic_page(page, addr, pages_per_huge_page);
3964                 return;
3965         }
3966
3967         might_sleep();
3968         for (i = 0; i < pages_per_huge_page; i++) {
3969                 cond_resched();
3970                 clear_user_highpage(page + i, addr + i * PAGE_SIZE);
3971         }
3972 }
3973
3974 static void copy_user_gigantic_page(struct page *dst, struct page *src,
3975                                     unsigned long addr,
3976                                     struct vm_area_struct *vma,
3977                                     unsigned int pages_per_huge_page)
3978 {
3979         int i;
3980         struct page *dst_base = dst;
3981         struct page *src_base = src;
3982
3983         for (i = 0; i < pages_per_huge_page; ) {
3984                 cond_resched();
3985                 copy_user_highpage(dst, src, addr + i*PAGE_SIZE, vma);
3986
3987                 i++;
3988                 dst = mem_map_next(dst, dst_base, i);
3989                 src = mem_map_next(src, src_base, i);
3990         }
3991 }
3992
3993 void copy_user_huge_page(struct page *dst, struct page *src,
3994                          unsigned long addr, struct vm_area_struct *vma,
3995                          unsigned int pages_per_huge_page)
3996 {
3997         int i;
3998
3999         if (unlikely(pages_per_huge_page > MAX_ORDER_NR_PAGES)) {
4000                 copy_user_gigantic_page(dst, src, addr, vma,
4001                                         pages_per_huge_page);
4002                 return;
4003         }
4004
4005         might_sleep();
4006         for (i = 0; i < pages_per_huge_page; i++) {
4007                 cond_resched();
4008                 copy_user_highpage(dst + i, src + i, addr + i*PAGE_SIZE, vma);
4009         }
4010 }
4011 #endif /* CONFIG_TRANSPARENT_HUGEPAGE || CONFIG_HUGETLBFS */