Merge branch 'stable-3.2' into pandora-3.2
[pandora-kernel.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 static void __sync_task_rss_stat(struct task_struct *task, struct mm_struct *mm)
129 {
130         int i;
131
132         for (i = 0; i < NR_MM_COUNTERS; i++) {
133                 if (task->rss_stat.count[i]) {
134                         add_mm_counter(mm, i, task->rss_stat.count[i]);
135                         task->rss_stat.count[i] = 0;
136                 }
137         }
138         task->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_task_rss_stat(task, task->mm);
161 }
162
163 unsigned long get_mm_counter(struct mm_struct *mm, int member)
164 {
165         long val = 0;
166
167         /*
168          * Don't use task->mm here...for avoiding to use task_get_mm()..
169          * The caller must guarantee task->mm is not invalid.
170          */
171         val = atomic_long_read(&mm->rss_stat.count[member]);
172         /*
173          * counter is updated in asynchronous manner and may go to minus.
174          * But it's never be expected number for users.
175          */
176         if (val < 0)
177                 return 0;
178         return (unsigned long)val;
179 }
180
181 void sync_mm_rss(struct task_struct *task, struct mm_struct *mm)
182 {
183         __sync_task_rss_stat(task, mm);
184 }
185 #else /* SPLIT_RSS_COUNTING */
186
187 #define inc_mm_counter_fast(mm, member) inc_mm_counter(mm, member)
188 #define dec_mm_counter_fast(mm, member) dec_mm_counter(mm, member)
189
190 static void check_sync_rss_stat(struct task_struct *task)
191 {
192 }
193
194 #endif /* SPLIT_RSS_COUNTING */
195
196 #ifdef HAVE_GENERIC_MMU_GATHER
197
198 static int tlb_next_batch(struct mmu_gather *tlb)
199 {
200         struct mmu_gather_batch *batch;
201
202         batch = tlb->active;
203         if (batch->next) {
204                 tlb->active = batch->next;
205                 return 1;
206         }
207
208         batch = (void *)__get_free_pages(GFP_NOWAIT | __GFP_NOWARN, 0);
209         if (!batch)
210                 return 0;
211
212         batch->next = NULL;
213         batch->nr   = 0;
214         batch->max  = MAX_GATHER_BATCH;
215
216         tlb->active->next = batch;
217         tlb->active = batch;
218
219         return 1;
220 }
221
222 /* tlb_gather_mmu
223  *      Called to initialize an (on-stack) mmu_gather structure for page-table
224  *      tear-down from @mm. The @fullmm argument is used when @mm is without
225  *      users and we're going to destroy the full address space (exit/execve).
226  */
227 void tlb_gather_mmu(struct mmu_gather *tlb, struct mm_struct *mm, bool fullmm)
228 {
229         tlb->mm = mm;
230
231         tlb->fullmm     = fullmm;
232         tlb->need_flush = 0;
233         tlb->fast_mode  = (num_possible_cpus() == 1);
234         tlb->local.next = NULL;
235         tlb->local.nr   = 0;
236         tlb->local.max  = ARRAY_SIZE(tlb->__pages);
237         tlb->active     = &tlb->local;
238
239 #ifdef CONFIG_HAVE_RCU_TABLE_FREE
240         tlb->batch = NULL;
241 #endif
242 }
243
244 void tlb_flush_mmu(struct mmu_gather *tlb)
245 {
246         struct mmu_gather_batch *batch;
247
248         if (!tlb->need_flush)
249                 return;
250         tlb->need_flush = 0;
251         tlb_flush(tlb);
252 #ifdef CONFIG_HAVE_RCU_TABLE_FREE
253         tlb_table_flush(tlb);
254 #endif
255
256         if (tlb_fast_mode(tlb))
257                 return;
258
259         for (batch = &tlb->local; batch; batch = batch->next) {
260                 free_pages_and_swap_cache(batch->pages, batch->nr);
261                 batch->nr = 0;
262         }
263         tlb->active = &tlb->local;
264 }
265
266 /* tlb_finish_mmu
267  *      Called at the end of the shootdown operation to free up any resources
268  *      that were required.
269  */
270 void tlb_finish_mmu(struct mmu_gather *tlb, unsigned long start, unsigned long end)
271 {
272         struct mmu_gather_batch *batch, *next;
273
274         tlb_flush_mmu(tlb);
275
276         /* keep the page table cache within bounds */
277         check_pgt_cache();
278
279         for (batch = tlb->local.next; batch; batch = next) {
280                 next = batch->next;
281                 free_pages((unsigned long)batch, 0);
282         }
283         tlb->local.next = NULL;
284 }
285
286 /* __tlb_remove_page
287  *      Must perform the equivalent to __free_pte(pte_get_and_clear(ptep)), while
288  *      handling the additional races in SMP caused by other CPUs caching valid
289  *      mappings in their TLBs. Returns the number of free page slots left.
290  *      When out of page slots we must call tlb_flush_mmu().
291  */
292 int __tlb_remove_page(struct mmu_gather *tlb, struct page *page)
293 {
294         struct mmu_gather_batch *batch;
295
296         tlb->need_flush = 1;
297
298         if (tlb_fast_mode(tlb)) {
299                 free_page_and_swap_cache(page);
300                 return 1; /* avoid calling tlb_flush_mmu() */
301         }
302
303         batch = tlb->active;
304         batch->pages[batch->nr++] = page;
305         if (batch->nr == batch->max) {
306                 if (!tlb_next_batch(tlb))
307                         return 0;
308                 batch = tlb->active;
309         }
310         VM_BUG_ON(batch->nr > batch->max);
311
312         return batch->max - batch->nr;
313 }
314
315 #endif /* HAVE_GENERIC_MMU_GATHER */
316
317 #ifdef CONFIG_HAVE_RCU_TABLE_FREE
318
319 /*
320  * See the comment near struct mmu_table_batch.
321  */
322
323 static void tlb_remove_table_smp_sync(void *arg)
324 {
325         /* Simply deliver the interrupt */
326 }
327
328 static void tlb_remove_table_one(void *table)
329 {
330         /*
331          * This isn't an RCU grace period and hence the page-tables cannot be
332          * assumed to be actually RCU-freed.
333          *
334          * It is however sufficient for software page-table walkers that rely on
335          * IRQ disabling. See the comment near struct mmu_table_batch.
336          */
337         smp_call_function(tlb_remove_table_smp_sync, NULL, 1);
338         __tlb_remove_table(table);
339 }
340
341 static void tlb_remove_table_rcu(struct rcu_head *head)
342 {
343         struct mmu_table_batch *batch;
344         int i;
345
346         batch = container_of(head, struct mmu_table_batch, rcu);
347
348         for (i = 0; i < batch->nr; i++)
349                 __tlb_remove_table(batch->tables[i]);
350
351         free_page((unsigned long)batch);
352 }
353
354 void tlb_table_flush(struct mmu_gather *tlb)
355 {
356         struct mmu_table_batch **batch = &tlb->batch;
357
358         if (*batch) {
359                 call_rcu_sched(&(*batch)->rcu, tlb_remove_table_rcu);
360                 *batch = NULL;
361         }
362 }
363
364 void tlb_remove_table(struct mmu_gather *tlb, void *table)
365 {
366         struct mmu_table_batch **batch = &tlb->batch;
367
368         tlb->need_flush = 1;
369
370         /*
371          * When there's less then two users of this mm there cannot be a
372          * concurrent page-table walk.
373          */
374         if (atomic_read(&tlb->mm->mm_users) < 2) {
375                 __tlb_remove_table(table);
376                 return;
377         }
378
379         if (*batch == NULL) {
380                 *batch = (struct mmu_table_batch *)__get_free_page(GFP_NOWAIT | __GFP_NOWARN);
381                 if (*batch == NULL) {
382                         tlb_remove_table_one(table);
383                         return;
384                 }
385                 (*batch)->nr = 0;
386         }
387         (*batch)->tables[(*batch)->nr++] = table;
388         if ((*batch)->nr == MAX_TABLE_BATCH)
389                 tlb_table_flush(tlb);
390 }
391
392 #endif /* CONFIG_HAVE_RCU_TABLE_FREE */
393
394 /*
395  * If a p?d_bad entry is found while walking page tables, report
396  * the error, before resetting entry to p?d_none.  Usually (but
397  * very seldom) called out from the p?d_none_or_clear_bad macros.
398  */
399
400 void pgd_clear_bad(pgd_t *pgd)
401 {
402         pgd_ERROR(*pgd);
403         pgd_clear(pgd);
404 }
405
406 void pud_clear_bad(pud_t *pud)
407 {
408         pud_ERROR(*pud);
409         pud_clear(pud);
410 }
411
412 void pmd_clear_bad(pmd_t *pmd)
413 {
414         pmd_ERROR(*pmd);
415         pmd_clear(pmd);
416 }
417
418 /*
419  * Note: this doesn't free the actual pages themselves. That
420  * has been handled earlier when unmapping all the memory regions.
421  */
422 static void free_pte_range(struct mmu_gather *tlb, pmd_t *pmd,
423                            unsigned long addr)
424 {
425         pgtable_t token = pmd_pgtable(*pmd);
426         pmd_clear(pmd);
427         pte_free_tlb(tlb, token, addr);
428         tlb->mm->nr_ptes--;
429 }
430
431 static inline void free_pmd_range(struct mmu_gather *tlb, pud_t *pud,
432                                 unsigned long addr, unsigned long end,
433                                 unsigned long floor, unsigned long ceiling)
434 {
435         pmd_t *pmd;
436         unsigned long next;
437         unsigned long start;
438
439         start = addr;
440         pmd = pmd_offset(pud, addr);
441         do {
442                 next = pmd_addr_end(addr, end);
443                 if (pmd_none_or_clear_bad(pmd))
444                         continue;
445                 free_pte_range(tlb, pmd, addr);
446         } while (pmd++, addr = next, addr != end);
447
448         start &= PUD_MASK;
449         if (start < floor)
450                 return;
451         if (ceiling) {
452                 ceiling &= PUD_MASK;
453                 if (!ceiling)
454                         return;
455         }
456         if (end - 1 > ceiling - 1)
457                 return;
458
459         pmd = pmd_offset(pud, start);
460         pud_clear(pud);
461         pmd_free_tlb(tlb, pmd, start);
462 }
463
464 static inline void free_pud_range(struct mmu_gather *tlb, pgd_t *pgd,
465                                 unsigned long addr, unsigned long end,
466                                 unsigned long floor, unsigned long ceiling)
467 {
468         pud_t *pud;
469         unsigned long next;
470         unsigned long start;
471
472         start = addr;
473         pud = pud_offset(pgd, addr);
474         do {
475                 next = pud_addr_end(addr, end);
476                 if (pud_none_or_clear_bad(pud))
477                         continue;
478                 free_pmd_range(tlb, pud, addr, next, floor, ceiling);
479         } while (pud++, addr = next, addr != end);
480
481         start &= PGDIR_MASK;
482         if (start < floor)
483                 return;
484         if (ceiling) {
485                 ceiling &= PGDIR_MASK;
486                 if (!ceiling)
487                         return;
488         }
489         if (end - 1 > ceiling - 1)
490                 return;
491
492         pud = pud_offset(pgd, start);
493         pgd_clear(pgd);
494         pud_free_tlb(tlb, pud, start);
495 }
496
497 /*
498  * This function frees user-level page tables of a process.
499  *
500  * Must be called with pagetable lock held.
501  */
502 void free_pgd_range(struct mmu_gather *tlb,
503                         unsigned long addr, unsigned long end,
504                         unsigned long floor, unsigned long ceiling)
505 {
506         pgd_t *pgd;
507         unsigned long next;
508
509         /*
510          * The next few lines have given us lots of grief...
511          *
512          * Why are we testing PMD* at this top level?  Because often
513          * there will be no work to do at all, and we'd prefer not to
514          * go all the way down to the bottom just to discover that.
515          *
516          * Why all these "- 1"s?  Because 0 represents both the bottom
517          * of the address space and the top of it (using -1 for the
518          * top wouldn't help much: the masks would do the wrong thing).
519          * The rule is that addr 0 and floor 0 refer to the bottom of
520          * the address space, but end 0 and ceiling 0 refer to the top
521          * Comparisons need to use "end - 1" and "ceiling - 1" (though
522          * that end 0 case should be mythical).
523          *
524          * Wherever addr is brought up or ceiling brought down, we must
525          * be careful to reject "the opposite 0" before it confuses the
526          * subsequent tests.  But what about where end is brought down
527          * by PMD_SIZE below? no, end can't go down to 0 there.
528          *
529          * Whereas we round start (addr) and ceiling down, by different
530          * masks at different levels, in order to test whether a table
531          * now has no other vmas using it, so can be freed, we don't
532          * bother to round floor or end up - the tests don't need that.
533          */
534
535         addr &= PMD_MASK;
536         if (addr < floor) {
537                 addr += PMD_SIZE;
538                 if (!addr)
539                         return;
540         }
541         if (ceiling) {
542                 ceiling &= PMD_MASK;
543                 if (!ceiling)
544                         return;
545         }
546         if (end - 1 > ceiling - 1)
547                 end -= PMD_SIZE;
548         if (addr > end - 1)
549                 return;
550
551         pgd = pgd_offset(tlb->mm, addr);
552         do {
553                 next = pgd_addr_end(addr, end);
554                 if (pgd_none_or_clear_bad(pgd))
555                         continue;
556                 free_pud_range(tlb, pgd, addr, next, floor, ceiling);
557         } while (pgd++, addr = next, addr != end);
558 }
559
560 void free_pgtables(struct mmu_gather *tlb, struct vm_area_struct *vma,
561                 unsigned long floor, unsigned long ceiling)
562 {
563         while (vma) {
564                 struct vm_area_struct *next = vma->vm_next;
565                 unsigned long addr = vma->vm_start;
566
567                 /*
568                  * Hide vma from rmap and truncate_pagecache before freeing
569                  * pgtables
570                  */
571                 unlink_anon_vmas(vma);
572                 unlink_file_vma(vma);
573
574                 if (is_vm_hugetlb_page(vma)) {
575                         hugetlb_free_pgd_range(tlb, addr, vma->vm_end,
576                                 floor, next? next->vm_start: ceiling);
577                 } else {
578                         /*
579                          * Optimization: gather nearby vmas into one call down
580                          */
581                         while (next && next->vm_start <= vma->vm_end + PMD_SIZE
582                                && !is_vm_hugetlb_page(next)) {
583                                 vma = next;
584                                 next = vma->vm_next;
585                                 unlink_anon_vmas(vma);
586                                 unlink_file_vma(vma);
587                         }
588                         free_pgd_range(tlb, addr, vma->vm_end,
589                                 floor, next? next->vm_start: ceiling);
590                 }
591                 vma = next;
592         }
593 }
594
595 int __pte_alloc(struct mm_struct *mm, struct vm_area_struct *vma,
596                 pmd_t *pmd, unsigned long address)
597 {
598         pgtable_t new = pte_alloc_one(mm, address);
599         int wait_split_huge_page;
600         if (!new)
601                 return -ENOMEM;
602
603         /*
604          * Ensure all pte setup (eg. pte page lock and page clearing) are
605          * visible before the pte is made visible to other CPUs by being
606          * put into page tables.
607          *
608          * The other side of the story is the pointer chasing in the page
609          * table walking code (when walking the page table without locking;
610          * ie. most of the time). Fortunately, these data accesses consist
611          * of a chain of data-dependent loads, meaning most CPUs (alpha
612          * being the notable exception) will already guarantee loads are
613          * seen in-order. See the alpha page table accessors for the
614          * smp_read_barrier_depends() barriers in page table walking code.
615          */
616         smp_wmb(); /* Could be smp_wmb__xxx(before|after)_spin_lock */
617
618         spin_lock(&mm->page_table_lock);
619         wait_split_huge_page = 0;
620         if (likely(pmd_none(*pmd))) {   /* Has another populated it ? */
621                 mm->nr_ptes++;
622                 pmd_populate(mm, pmd, new);
623                 new = NULL;
624         } else if (unlikely(pmd_trans_splitting(*pmd)))
625                 wait_split_huge_page = 1;
626         spin_unlock(&mm->page_table_lock);
627         if (new)
628                 pte_free(mm, new);
629         if (wait_split_huge_page)
630                 wait_split_huge_page(vma->anon_vma, pmd);
631         return 0;
632 }
633
634 int __pte_alloc_kernel(pmd_t *pmd, unsigned long address)
635 {
636         pte_t *new = pte_alloc_one_kernel(&init_mm, address);
637         if (!new)
638                 return -ENOMEM;
639
640         smp_wmb(); /* See comment in __pte_alloc */
641
642         spin_lock(&init_mm.page_table_lock);
643         if (likely(pmd_none(*pmd))) {   /* Has another populated it ? */
644                 pmd_populate_kernel(&init_mm, pmd, new);
645                 new = NULL;
646         } else
647                 VM_BUG_ON(pmd_trans_splitting(*pmd));
648         spin_unlock(&init_mm.page_table_lock);
649         if (new)
650                 pte_free_kernel(&init_mm, new);
651         return 0;
652 }
653
654 static inline void init_rss_vec(int *rss)
655 {
656         memset(rss, 0, sizeof(int) * NR_MM_COUNTERS);
657 }
658
659 static inline void add_mm_rss_vec(struct mm_struct *mm, int *rss)
660 {
661         int i;
662
663         if (current->mm == mm)
664                 sync_mm_rss(current, mm);
665         for (i = 0; i < NR_MM_COUNTERS; i++)
666                 if (rss[i])
667                         add_mm_counter(mm, i, rss[i]);
668 }
669
670 /*
671  * This function is called to print an error when a bad pte
672  * is found. For example, we might have a PFN-mapped pte in
673  * a region that doesn't allow it.
674  *
675  * The calling function must still handle the error.
676  */
677 static void print_bad_pte(struct vm_area_struct *vma, unsigned long addr,
678                           pte_t pte, struct page *page)
679 {
680         pgd_t *pgd = pgd_offset(vma->vm_mm, addr);
681         pud_t *pud = pud_offset(pgd, addr);
682         pmd_t *pmd = pmd_offset(pud, addr);
683         struct address_space *mapping;
684         pgoff_t index;
685         static unsigned long resume;
686         static unsigned long nr_shown;
687         static unsigned long nr_unshown;
688
689         /*
690          * Allow a burst of 60 reports, then keep quiet for that minute;
691          * or allow a steady drip of one report per second.
692          */
693         if (nr_shown == 60) {
694                 if (time_before(jiffies, resume)) {
695                         nr_unshown++;
696                         return;
697                 }
698                 if (nr_unshown) {
699                         printk(KERN_ALERT
700                                 "BUG: Bad page map: %lu messages suppressed\n",
701                                 nr_unshown);
702                         nr_unshown = 0;
703                 }
704                 nr_shown = 0;
705         }
706         if (nr_shown++ == 0)
707                 resume = jiffies + 60 * HZ;
708
709         mapping = vma->vm_file ? vma->vm_file->f_mapping : NULL;
710         index = linear_page_index(vma, addr);
711
712         printk(KERN_ALERT
713                 "BUG: Bad page map in process %s  pte:%08llx pmd:%08llx\n",
714                 current->comm,
715                 (long long)pte_val(pte), (long long)pmd_val(*pmd));
716         if (page)
717                 dump_page(page);
718         printk(KERN_ALERT
719                 "addr:%p vm_flags:%08lx anon_vma:%p mapping:%p index:%lx\n",
720                 (void *)addr, vma->vm_flags, vma->anon_vma, mapping, index);
721         /*
722          * Choose text because data symbols depend on CONFIG_KALLSYMS_ALL=y
723          */
724         if (vma->vm_ops)
725                 print_symbol(KERN_ALERT "vma->vm_ops->fault: %s\n",
726                                 (unsigned long)vma->vm_ops->fault);
727         if (vma->vm_file && vma->vm_file->f_op)
728                 print_symbol(KERN_ALERT "vma->vm_file->f_op->mmap: %s\n",
729                                 (unsigned long)vma->vm_file->f_op->mmap);
730         dump_stack();
731         add_taint(TAINT_BAD_PAGE);
732 }
733
734 static inline int is_cow_mapping(vm_flags_t flags)
735 {
736         return (flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE;
737 }
738
739 #ifndef is_zero_pfn
740 static inline int is_zero_pfn(unsigned long pfn)
741 {
742         return pfn == zero_pfn;
743 }
744 #endif
745
746 #ifndef my_zero_pfn
747 static inline unsigned long my_zero_pfn(unsigned long addr)
748 {
749         return zero_pfn;
750 }
751 #endif
752
753 /*
754  * vm_normal_page -- This function gets the "struct page" associated with a pte.
755  *
756  * "Special" mappings do not wish to be associated with a "struct page" (either
757  * it doesn't exist, or it exists but they don't want to touch it). In this
758  * case, NULL is returned here. "Normal" mappings do have a struct page.
759  *
760  * There are 2 broad cases. Firstly, an architecture may define a pte_special()
761  * pte bit, in which case this function is trivial. Secondly, an architecture
762  * may not have a spare pte bit, which requires a more complicated scheme,
763  * described below.
764  *
765  * A raw VM_PFNMAP mapping (ie. one that is not COWed) is always considered a
766  * special mapping (even if there are underlying and valid "struct pages").
767  * COWed pages of a VM_PFNMAP are always normal.
768  *
769  * The way we recognize COWed pages within VM_PFNMAP mappings is through the
770  * rules set up by "remap_pfn_range()": the vma will have the VM_PFNMAP bit
771  * set, and the vm_pgoff will point to the first PFN mapped: thus every special
772  * mapping will always honor the rule
773  *
774  *      pfn_of_page == vma->vm_pgoff + ((addr - vma->vm_start) >> PAGE_SHIFT)
775  *
776  * And for normal mappings this is false.
777  *
778  * This restricts such mappings to be a linear translation from virtual address
779  * to pfn. To get around this restriction, we allow arbitrary mappings so long
780  * as the vma is not a COW mapping; in that case, we know that all ptes are
781  * special (because none can have been COWed).
782  *
783  *
784  * In order to support COW of arbitrary special mappings, we have VM_MIXEDMAP.
785  *
786  * VM_MIXEDMAP mappings can likewise contain memory with or without "struct
787  * page" backing, however the difference is that _all_ pages with a struct
788  * page (that is, those where pfn_valid is true) are refcounted and considered
789  * normal pages by the VM. The disadvantage is that pages are refcounted
790  * (which can be slower and simply not an option for some PFNMAP users). The
791  * advantage is that we don't have to follow the strict linearity rule of
792  * PFNMAP mappings in order to support COWable mappings.
793  *
794  */
795 #ifdef __HAVE_ARCH_PTE_SPECIAL
796 # define HAVE_PTE_SPECIAL 1
797 #else
798 # define HAVE_PTE_SPECIAL 0
799 #endif
800 struct page *vm_normal_page(struct vm_area_struct *vma, unsigned long addr,
801                                 pte_t pte)
802 {
803         unsigned long pfn = pte_pfn(pte);
804
805         if (HAVE_PTE_SPECIAL) {
806                 if (likely(!pte_special(pte)))
807                         goto check_pfn;
808                 if (vma->vm_flags & (VM_PFNMAP | VM_MIXEDMAP))
809                         return NULL;
810                 if (!is_zero_pfn(pfn))
811                         print_bad_pte(vma, addr, pte, NULL);
812                 return NULL;
813         }
814
815         /* !HAVE_PTE_SPECIAL case follows: */
816
817         if (unlikely(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP))) {
818                 if (vma->vm_flags & VM_MIXEDMAP) {
819                         if (!pfn_valid(pfn))
820                                 return NULL;
821                         goto out;
822                 } else {
823                         unsigned long off;
824                         off = (addr - vma->vm_start) >> PAGE_SHIFT;
825                         if (pfn == vma->vm_pgoff + off)
826                                 return NULL;
827                         if (!is_cow_mapping(vma->vm_flags))
828                                 return NULL;
829                 }
830         }
831
832         if (is_zero_pfn(pfn))
833                 return NULL;
834 check_pfn:
835         if (unlikely(pfn > highest_memmap_pfn)) {
836                 print_bad_pte(vma, addr, pte, NULL);
837                 return NULL;
838         }
839
840         /*
841          * NOTE! We still have PageReserved() pages in the page tables.
842          * eg. VDSO mappings can cause them to exist.
843          */
844 out:
845         return pfn_to_page(pfn);
846 }
847
848 /*
849  * copy one vm_area from one task to the other. Assumes the page tables
850  * already present in the new task to be cleared in the whole range
851  * covered by this vma.
852  */
853
854 static inline unsigned long
855 copy_one_pte(struct mm_struct *dst_mm, struct mm_struct *src_mm,
856                 pte_t *dst_pte, pte_t *src_pte, struct vm_area_struct *vma,
857                 unsigned long addr, int *rss)
858 {
859         unsigned long vm_flags = vma->vm_flags;
860         pte_t pte = *src_pte;
861         struct page *page;
862
863         /* pte contains position in swap or file, so copy. */
864         if (unlikely(!pte_present(pte))) {
865                 if (!pte_file(pte)) {
866                         swp_entry_t entry = pte_to_swp_entry(pte);
867
868                         if (swap_duplicate(entry) < 0)
869                                 return entry.val;
870
871                         /* make sure dst_mm is on swapoff's mmlist. */
872                         if (unlikely(list_empty(&dst_mm->mmlist))) {
873                                 spin_lock(&mmlist_lock);
874                                 if (list_empty(&dst_mm->mmlist))
875                                         list_add(&dst_mm->mmlist,
876                                                  &src_mm->mmlist);
877                                 spin_unlock(&mmlist_lock);
878                         }
879                         if (likely(!non_swap_entry(entry)))
880                                 rss[MM_SWAPENTS]++;
881                         else if (is_write_migration_entry(entry) &&
882                                         is_cow_mapping(vm_flags)) {
883                                 /*
884                                  * COW mappings require pages in both parent
885                                  * and child to be set to read.
886                                  */
887                                 make_migration_entry_read(&entry);
888                                 pte = swp_entry_to_pte(entry);
889                                 set_pte_at(src_mm, addr, src_pte, pte);
890                         }
891                 }
892                 goto out_set_pte;
893         }
894
895         /*
896          * If it's a COW mapping, write protect it both
897          * in the parent and the child
898          */
899         if (is_cow_mapping(vm_flags)) {
900                 ptep_set_wrprotect(src_mm, addr, src_pte);
901                 pte = pte_wrprotect(pte);
902         }
903
904         /*
905          * If it's a shared mapping, mark it clean in
906          * the child
907          */
908         if (vm_flags & VM_SHARED)
909                 pte = pte_mkclean(pte);
910         pte = pte_mkold(pte);
911
912         page = vm_normal_page(vma, addr, pte);
913         if (page) {
914                 get_page(page);
915                 page_dup_rmap(page);
916                 if (PageAnon(page))
917                         rss[MM_ANONPAGES]++;
918                 else
919                         rss[MM_FILEPAGES]++;
920         }
921
922 out_set_pte:
923         set_pte_at(dst_mm, addr, dst_pte, pte);
924         return 0;
925 }
926
927 int copy_pte_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
928                    pmd_t *dst_pmd, pmd_t *src_pmd, struct vm_area_struct *vma,
929                    unsigned long addr, unsigned long end)
930 {
931         pte_t *orig_src_pte, *orig_dst_pte;
932         pte_t *src_pte, *dst_pte;
933         spinlock_t *src_ptl, *dst_ptl;
934         int progress = 0;
935         int rss[NR_MM_COUNTERS];
936         swp_entry_t entry = (swp_entry_t){0};
937
938 again:
939         init_rss_vec(rss);
940
941         dst_pte = pte_alloc_map_lock(dst_mm, dst_pmd, addr, &dst_ptl);
942         if (!dst_pte)
943                 return -ENOMEM;
944         src_pte = pte_offset_map(src_pmd, addr);
945         src_ptl = pte_lockptr(src_mm, src_pmd);
946         spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
947         orig_src_pte = src_pte;
948         orig_dst_pte = dst_pte;
949         arch_enter_lazy_mmu_mode();
950
951         do {
952                 /*
953                  * We are holding two locks at this point - either of them
954                  * could generate latencies in another task on another CPU.
955                  */
956                 if (progress >= 32) {
957                         progress = 0;
958                         if (need_resched() ||
959                             spin_needbreak(src_ptl) || spin_needbreak(dst_ptl))
960                                 break;
961                 }
962                 if (pte_none(*src_pte)) {
963                         progress++;
964                         continue;
965                 }
966                 entry.val = copy_one_pte(dst_mm, src_mm, dst_pte, src_pte,
967                                                         vma, addr, rss);
968                 if (entry.val)
969                         break;
970                 progress += 8;
971         } while (dst_pte++, src_pte++, addr += PAGE_SIZE, addr != end);
972
973         arch_leave_lazy_mmu_mode();
974         spin_unlock(src_ptl);
975         pte_unmap(orig_src_pte);
976         add_mm_rss_vec(dst_mm, rss);
977         pte_unmap_unlock(orig_dst_pte, dst_ptl);
978         cond_resched();
979
980         if (entry.val) {
981                 if (add_swap_count_continuation(entry, GFP_KERNEL) < 0)
982                         return -ENOMEM;
983                 progress = 0;
984         }
985         if (addr != end)
986                 goto again;
987         return 0;
988 }
989
990 static inline int copy_pmd_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
991                 pud_t *dst_pud, pud_t *src_pud, struct vm_area_struct *vma,
992                 unsigned long addr, unsigned long end)
993 {
994         pmd_t *src_pmd, *dst_pmd;
995         unsigned long next;
996
997         dst_pmd = pmd_alloc(dst_mm, dst_pud, addr);
998         if (!dst_pmd)
999                 return -ENOMEM;
1000         src_pmd = pmd_offset(src_pud, addr);
1001         do {
1002                 next = pmd_addr_end(addr, end);
1003                 if (pmd_trans_huge(*src_pmd)) {
1004                         int err;
1005                         VM_BUG_ON(next-addr != HPAGE_PMD_SIZE);
1006                         err = copy_huge_pmd(dst_mm, src_mm,
1007                                             dst_pmd, src_pmd, addr, vma);
1008                         if (err == -ENOMEM)
1009                                 return -ENOMEM;
1010                         if (!err)
1011                                 continue;
1012                         /* fall through */
1013                 }
1014                 if (pmd_none_or_clear_bad(src_pmd))
1015                         continue;
1016                 if (copy_pte_range(dst_mm, src_mm, dst_pmd, src_pmd,
1017                                                 vma, addr, next))
1018                         return -ENOMEM;
1019         } while (dst_pmd++, src_pmd++, addr = next, addr != end);
1020         return 0;
1021 }
1022
1023 static inline int copy_pud_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
1024                 pgd_t *dst_pgd, pgd_t *src_pgd, struct vm_area_struct *vma,
1025                 unsigned long addr, unsigned long end)
1026 {
1027         pud_t *src_pud, *dst_pud;
1028         unsigned long next;
1029
1030         dst_pud = pud_alloc(dst_mm, dst_pgd, addr);
1031         if (!dst_pud)
1032                 return -ENOMEM;
1033         src_pud = pud_offset(src_pgd, addr);
1034         do {
1035                 next = pud_addr_end(addr, end);
1036                 if (pud_none_or_clear_bad(src_pud))
1037                         continue;
1038                 if (copy_pmd_range(dst_mm, src_mm, dst_pud, src_pud,
1039                                                 vma, addr, next))
1040                         return -ENOMEM;
1041         } while (dst_pud++, src_pud++, addr = next, addr != end);
1042         return 0;
1043 }
1044
1045 int copy_page_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
1046                 struct vm_area_struct *vma)
1047 {
1048         pgd_t *src_pgd, *dst_pgd;
1049         unsigned long next;
1050         unsigned long addr = vma->vm_start;
1051         unsigned long end = vma->vm_end;
1052         int ret;
1053
1054         /*
1055          * Don't copy ptes where a page fault will fill them correctly.
1056          * Fork becomes much lighter when there are big shared or private
1057          * readonly mappings. The tradeoff is that copy_page_range is more
1058          * efficient than faulting.
1059          */
1060         if (!(vma->vm_flags & (VM_HUGETLB|VM_NONLINEAR|VM_PFNMAP|VM_INSERTPAGE))) {
1061                 if (!vma->anon_vma)
1062                         return 0;
1063         }
1064
1065         if (is_vm_hugetlb_page(vma))
1066                 return copy_hugetlb_page_range(dst_mm, src_mm, vma);
1067
1068         if (unlikely(is_pfn_mapping(vma))) {
1069                 /*
1070                  * We do not free on error cases below as remove_vma
1071                  * gets called on error from higher level routine
1072                  */
1073                 ret = track_pfn_vma_copy(vma);
1074                 if (ret)
1075                         return ret;
1076         }
1077
1078         /*
1079          * We need to invalidate the secondary MMU mappings only when
1080          * there could be a permission downgrade on the ptes of the
1081          * parent mm. And a permission downgrade will only happen if
1082          * is_cow_mapping() returns true.
1083          */
1084         if (is_cow_mapping(vma->vm_flags))
1085                 mmu_notifier_invalidate_range_start(src_mm, addr, end);
1086
1087         ret = 0;
1088         dst_pgd = pgd_offset(dst_mm, addr);
1089         src_pgd = pgd_offset(src_mm, addr);
1090         do {
1091                 next = pgd_addr_end(addr, end);
1092                 if (pgd_none_or_clear_bad(src_pgd))
1093                         continue;
1094                 if (unlikely(copy_pud_range(dst_mm, src_mm, dst_pgd, src_pgd,
1095                                             vma, addr, next))) {
1096                         ret = -ENOMEM;
1097                         break;
1098                 }
1099         } while (dst_pgd++, src_pgd++, addr = next, addr != end);
1100
1101         if (is_cow_mapping(vma->vm_flags))
1102                 mmu_notifier_invalidate_range_end(src_mm,
1103                                                   vma->vm_start, end);
1104         return ret;
1105 }
1106
1107 static unsigned long zap_pte_range(struct mmu_gather *tlb,
1108                                 struct vm_area_struct *vma, pmd_t *pmd,
1109                                 unsigned long addr, unsigned long end,
1110                                 struct zap_details *details)
1111 {
1112         struct mm_struct *mm = tlb->mm;
1113         int force_flush = 0;
1114         int rss[NR_MM_COUNTERS];
1115         spinlock_t *ptl;
1116         pte_t *start_pte;
1117         pte_t *pte;
1118
1119 again:
1120         init_rss_vec(rss);
1121         start_pte = pte_offset_map_lock(mm, pmd, addr, &ptl);
1122         pte = start_pte;
1123         arch_enter_lazy_mmu_mode();
1124         do {
1125                 pte_t ptent = *pte;
1126                 if (pte_none(ptent)) {
1127                         continue;
1128                 }
1129
1130                 if (pte_present(ptent)) {
1131                         struct page *page;
1132
1133                         page = vm_normal_page(vma, addr, ptent);
1134                         if (unlikely(details) && page) {
1135                                 /*
1136                                  * unmap_shared_mapping_pages() wants to
1137                                  * invalidate cache without truncating:
1138                                  * unmap shared but keep private pages.
1139                                  */
1140                                 if (details->check_mapping &&
1141                                     details->check_mapping != page->mapping)
1142                                         continue;
1143                                 /*
1144                                  * Each page->index must be checked when
1145                                  * invalidating or truncating nonlinear.
1146                                  */
1147                                 if (details->nonlinear_vma &&
1148                                     (page->index < details->first_index ||
1149                                      page->index > details->last_index))
1150                                         continue;
1151                         }
1152                         ptent = ptep_get_and_clear_full(mm, addr, pte,
1153                                                         tlb->fullmm);
1154                         tlb_remove_tlb_entry(tlb, pte, addr);
1155                         if (unlikely(!page))
1156                                 continue;
1157                         if (unlikely(details) && details->nonlinear_vma
1158                             && linear_page_index(details->nonlinear_vma,
1159                                                 addr) != page->index)
1160                                 set_pte_at(mm, addr, pte,
1161                                            pgoff_to_pte(page->index));
1162                         if (PageAnon(page))
1163                                 rss[MM_ANONPAGES]--;
1164                         else {
1165                                 if (pte_dirty(ptent))
1166                                         set_page_dirty(page);
1167                                 if (pte_young(ptent) &&
1168                                     likely(!VM_SequentialReadHint(vma)))
1169                                         mark_page_accessed(page);
1170                                 rss[MM_FILEPAGES]--;
1171                         }
1172                         page_remove_rmap(page);
1173                         if (unlikely(page_mapcount(page) < 0))
1174                                 print_bad_pte(vma, addr, ptent, page);
1175                         force_flush = !__tlb_remove_page(tlb, page);
1176                         if (force_flush)
1177                                 break;
1178                         continue;
1179                 }
1180                 /*
1181                  * If details->check_mapping, we leave swap entries;
1182                  * if details->nonlinear_vma, we leave file entries.
1183                  */
1184                 if (unlikely(details))
1185                         continue;
1186                 if (pte_file(ptent)) {
1187                         if (unlikely(!(vma->vm_flags & VM_NONLINEAR)))
1188                                 print_bad_pte(vma, addr, ptent, NULL);
1189                 } else {
1190                         swp_entry_t entry = pte_to_swp_entry(ptent);
1191
1192                         if (!non_swap_entry(entry))
1193                                 rss[MM_SWAPENTS]--;
1194                         if (unlikely(!free_swap_and_cache(entry)))
1195                                 print_bad_pte(vma, addr, ptent, NULL);
1196                 }
1197                 pte_clear_not_present_full(mm, addr, pte, tlb->fullmm);
1198         } while (pte++, addr += PAGE_SIZE, addr != end);
1199
1200         add_mm_rss_vec(mm, rss);
1201         arch_leave_lazy_mmu_mode();
1202         pte_unmap_unlock(start_pte, ptl);
1203
1204         /*
1205          * mmu_gather ran out of room to batch pages, we break out of
1206          * the PTE lock to avoid doing the potential expensive TLB invalidate
1207          * and page-free while holding it.
1208          */
1209         if (force_flush) {
1210                 force_flush = 0;
1211                 tlb_flush_mmu(tlb);
1212                 if (addr != end)
1213                         goto again;
1214         }
1215
1216         return addr;
1217 }
1218
1219 static inline unsigned long zap_pmd_range(struct mmu_gather *tlb,
1220                                 struct vm_area_struct *vma, pud_t *pud,
1221                                 unsigned long addr, unsigned long end,
1222                                 struct zap_details *details)
1223 {
1224         pmd_t *pmd;
1225         unsigned long next;
1226
1227         pmd = pmd_offset(pud, addr);
1228         do {
1229                 next = pmd_addr_end(addr, end);
1230                 if (pmd_trans_huge(*pmd)) {
1231                         if (next - addr != HPAGE_PMD_SIZE) {
1232                                 VM_BUG_ON(!rwsem_is_locked(&tlb->mm->mmap_sem));
1233                                 split_huge_page_pmd(vma->vm_mm, pmd);
1234                         } else if (zap_huge_pmd(tlb, vma, pmd))
1235                                 goto next;
1236                         /* fall through */
1237                 }
1238                 /*
1239                  * Here there can be other concurrent MADV_DONTNEED or
1240                  * trans huge page faults running, and if the pmd is
1241                  * none or trans huge it can change under us. This is
1242                  * because MADV_DONTNEED holds the mmap_sem in read
1243                  * mode.
1244                  */
1245                 if (pmd_none_or_trans_huge_or_clear_bad(pmd))
1246                         goto next;
1247                 next = zap_pte_range(tlb, vma, pmd, addr, next, details);
1248 next:
1249                 cond_resched();
1250         } while (pmd++, addr = next, addr != end);
1251
1252         return addr;
1253 }
1254
1255 static inline unsigned long zap_pud_range(struct mmu_gather *tlb,
1256                                 struct vm_area_struct *vma, pgd_t *pgd,
1257                                 unsigned long addr, unsigned long end,
1258                                 struct zap_details *details)
1259 {
1260         pud_t *pud;
1261         unsigned long next;
1262
1263         pud = pud_offset(pgd, addr);
1264         do {
1265                 next = pud_addr_end(addr, end);
1266                 if (pud_none_or_clear_bad(pud))
1267                         continue;
1268                 next = zap_pmd_range(tlb, vma, pud, addr, next, details);
1269         } while (pud++, addr = next, addr != end);
1270
1271         return addr;
1272 }
1273
1274 static unsigned long unmap_page_range(struct mmu_gather *tlb,
1275                                 struct vm_area_struct *vma,
1276                                 unsigned long addr, unsigned long end,
1277                                 struct zap_details *details)
1278 {
1279         pgd_t *pgd;
1280         unsigned long next;
1281
1282         if (details && !details->check_mapping && !details->nonlinear_vma)
1283                 details = NULL;
1284
1285         BUG_ON(addr >= end);
1286         mem_cgroup_uncharge_start();
1287         tlb_start_vma(tlb, vma);
1288         pgd = pgd_offset(vma->vm_mm, addr);
1289         do {
1290                 next = pgd_addr_end(addr, end);
1291                 if (pgd_none_or_clear_bad(pgd))
1292                         continue;
1293                 next = zap_pud_range(tlb, vma, pgd, addr, next, details);
1294         } while (pgd++, addr = next, addr != end);
1295         tlb_end_vma(tlb, vma);
1296         mem_cgroup_uncharge_end();
1297
1298         return addr;
1299 }
1300
1301 /**
1302  * unmap_vmas - unmap a range of memory covered by a list of vma's
1303  * @tlb: address of the caller's struct mmu_gather
1304  * @vma: the starting vma
1305  * @start_addr: virtual address at which to start unmapping
1306  * @end_addr: virtual address at which to end unmapping
1307  * @nr_accounted: Place number of unmapped pages in vm-accountable vma's here
1308  * @details: details of nonlinear truncation or shared cache invalidation
1309  *
1310  * Returns the end address of the unmapping (restart addr if interrupted).
1311  *
1312  * Unmap all pages in the vma list.
1313  *
1314  * Only addresses between `start' and `end' will be unmapped.
1315  *
1316  * The VMA list must be sorted in ascending virtual address order.
1317  *
1318  * unmap_vmas() assumes that the caller will flush the whole unmapped address
1319  * range after unmap_vmas() returns.  So the only responsibility here is to
1320  * ensure that any thus-far unmapped pages are flushed before unmap_vmas()
1321  * drops the lock and schedules.
1322  */
1323 unsigned long unmap_vmas(struct mmu_gather *tlb,
1324                 struct vm_area_struct *vma, unsigned long start_addr,
1325                 unsigned long end_addr, unsigned long *nr_accounted,
1326                 struct zap_details *details)
1327 {
1328         unsigned long start = start_addr;
1329         struct mm_struct *mm = vma->vm_mm;
1330
1331         mmu_notifier_invalidate_range_start(mm, start_addr, end_addr);
1332         for ( ; vma && vma->vm_start < end_addr; vma = vma->vm_next) {
1333                 unsigned long end;
1334
1335                 start = max(vma->vm_start, start_addr);
1336                 if (start >= vma->vm_end)
1337                         continue;
1338                 end = min(vma->vm_end, end_addr);
1339                 if (end <= vma->vm_start)
1340                         continue;
1341
1342                 if (vma->vm_flags & VM_ACCOUNT)
1343                         *nr_accounted += (end - start) >> PAGE_SHIFT;
1344
1345                 if (unlikely(is_pfn_mapping(vma)))
1346                         untrack_pfn_vma(vma, 0, 0);
1347
1348                 while (start != end) {
1349                         if (unlikely(is_vm_hugetlb_page(vma))) {
1350                                 /*
1351                                  * It is undesirable to test vma->vm_file as it
1352                                  * should be non-null for valid hugetlb area.
1353                                  * However, vm_file will be NULL in the error
1354                                  * cleanup path of do_mmap_pgoff. When
1355                                  * hugetlbfs ->mmap method fails,
1356                                  * do_mmap_pgoff() nullifies vma->vm_file
1357                                  * before calling this function to clean up.
1358                                  * Since no pte has actually been setup, it is
1359                                  * safe to do nothing in this case.
1360                                  */
1361                                 if (vma->vm_file) {
1362                                         mutex_lock(&vma->vm_file->f_mapping->i_mmap_mutex);
1363                                         __unmap_hugepage_range_final(vma, start, end, NULL);
1364                                         mutex_unlock(&vma->vm_file->f_mapping->i_mmap_mutex);
1365                                 }
1366
1367                                 start = end;
1368                         } else
1369                                 start = unmap_page_range(tlb, vma, start, end, details);
1370                 }
1371         }
1372
1373         mmu_notifier_invalidate_range_end(mm, start_addr, end_addr);
1374         return start;   /* which is now the end (or restart) address */
1375 }
1376
1377 /**
1378  * zap_page_range - remove user pages in a given range
1379  * @vma: vm_area_struct holding the applicable pages
1380  * @address: starting address of pages to zap
1381  * @size: number of bytes to zap
1382  * @details: details of nonlinear truncation or shared cache invalidation
1383  */
1384 unsigned long zap_page_range(struct vm_area_struct *vma, unsigned long address,
1385                 unsigned long size, struct zap_details *details)
1386 {
1387         struct mm_struct *mm = vma->vm_mm;
1388         struct mmu_gather tlb;
1389         unsigned long end = address + size;
1390         unsigned long nr_accounted = 0;
1391
1392         lru_add_drain();
1393         tlb_gather_mmu(&tlb, mm, 0);
1394         update_hiwater_rss(mm);
1395         end = unmap_vmas(&tlb, vma, address, end, &nr_accounted, details);
1396         tlb_finish_mmu(&tlb, address, end);
1397         return end;
1398 }
1399 EXPORT_SYMBOL_GPL(zap_page_range);
1400
1401 /**
1402  * zap_vma_ptes - remove ptes mapping the vma
1403  * @vma: vm_area_struct holding ptes to be zapped
1404  * @address: starting address of pages to zap
1405  * @size: number of bytes to zap
1406  *
1407  * This function only unmaps ptes assigned to VM_PFNMAP vmas.
1408  *
1409  * The entire address range must be fully contained within the vma.
1410  *
1411  * Returns 0 if successful.
1412  */
1413 int zap_vma_ptes(struct vm_area_struct *vma, unsigned long address,
1414                 unsigned long size)
1415 {
1416         if (address < vma->vm_start || address + size > vma->vm_end ||
1417                         !(vma->vm_flags & VM_PFNMAP))
1418                 return -1;
1419         zap_page_range(vma, address, size, NULL);
1420         return 0;
1421 }
1422 EXPORT_SYMBOL_GPL(zap_vma_ptes);
1423
1424 /**
1425  * follow_page - look up a page descriptor from a user-virtual address
1426  * @vma: vm_area_struct mapping @address
1427  * @address: virtual address to look up
1428  * @flags: flags modifying lookup behaviour
1429  *
1430  * @flags can have FOLL_ flags set, defined in <linux/mm.h>
1431  *
1432  * Returns the mapped (struct page *), %NULL if no mapping exists, or
1433  * an error pointer if there is a mapping to something not represented
1434  * by a page descriptor (see also vm_normal_page()).
1435  */
1436 struct page *follow_page(struct vm_area_struct *vma, unsigned long address,
1437                         unsigned int flags)
1438 {
1439         pgd_t *pgd;
1440         pud_t *pud;
1441         pmd_t *pmd;
1442         pte_t *ptep, pte;
1443         spinlock_t *ptl;
1444         struct page *page;
1445         struct mm_struct *mm = vma->vm_mm;
1446
1447         page = follow_huge_addr(mm, address, flags & FOLL_WRITE);
1448         if (!IS_ERR(page)) {
1449                 BUG_ON(flags & FOLL_GET);
1450                 goto out;
1451         }
1452
1453         page = NULL;
1454         pgd = pgd_offset(mm, address);
1455         if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
1456                 goto no_page_table;
1457
1458         pud = pud_offset(pgd, address);
1459         if (pud_none(*pud))
1460                 goto no_page_table;
1461         if (pud_huge(*pud) && vma->vm_flags & VM_HUGETLB) {
1462                 BUG_ON(flags & FOLL_GET);
1463                 page = follow_huge_pud(mm, address, pud, flags & FOLL_WRITE);
1464                 goto out;
1465         }
1466         if (unlikely(pud_bad(*pud)))
1467                 goto no_page_table;
1468
1469         pmd = pmd_offset(pud, address);
1470         if (pmd_none(*pmd))
1471                 goto no_page_table;
1472         if (pmd_huge(*pmd) && vma->vm_flags & VM_HUGETLB) {
1473                 BUG_ON(flags & FOLL_GET);
1474                 page = follow_huge_pmd(mm, address, pmd, flags & FOLL_WRITE);
1475                 goto out;
1476         }
1477         if (pmd_trans_huge(*pmd)) {
1478                 if (flags & FOLL_SPLIT) {
1479                         split_huge_page_pmd(mm, pmd);
1480                         goto split_fallthrough;
1481                 }
1482                 spin_lock(&mm->page_table_lock);
1483                 if (likely(pmd_trans_huge(*pmd))) {
1484                         if (unlikely(pmd_trans_splitting(*pmd))) {
1485                                 spin_unlock(&mm->page_table_lock);
1486                                 wait_split_huge_page(vma->anon_vma, pmd);
1487                         } else {
1488                                 page = follow_trans_huge_pmd(mm, address,
1489                                                              pmd, flags);
1490                                 spin_unlock(&mm->page_table_lock);
1491                                 goto out;
1492                         }
1493                 } else
1494                         spin_unlock(&mm->page_table_lock);
1495                 /* fall through */
1496         }
1497 split_fallthrough:
1498         if (unlikely(pmd_bad(*pmd)))
1499                 goto no_page_table;
1500
1501         ptep = pte_offset_map_lock(mm, pmd, address, &ptl);
1502
1503         pte = *ptep;
1504         if (!pte_present(pte))
1505                 goto no_page;
1506         if ((flags & FOLL_WRITE) && !pte_write(pte))
1507                 goto unlock;
1508
1509         page = vm_normal_page(vma, address, pte);
1510         if (unlikely(!page)) {
1511                 if ((flags & FOLL_DUMP) ||
1512                     !is_zero_pfn(pte_pfn(pte)))
1513                         goto bad_page;
1514                 page = pte_page(pte);
1515         }
1516
1517         if (flags & FOLL_GET)
1518                 get_page_foll(page);
1519         if (flags & FOLL_TOUCH) {
1520                 if ((flags & FOLL_WRITE) &&
1521                     !pte_dirty(pte) && !PageDirty(page))
1522                         set_page_dirty(page);
1523                 /*
1524                  * pte_mkyoung() would be more correct here, but atomic care
1525                  * is needed to avoid losing the dirty bit: it is easier to use
1526                  * mark_page_accessed().
1527                  */
1528                 mark_page_accessed(page);
1529         }
1530         if ((flags & FOLL_MLOCK) && (vma->vm_flags & VM_LOCKED)) {
1531                 /*
1532                  * The preliminary mapping check is mainly to avoid the
1533                  * pointless overhead of lock_page on the ZERO_PAGE
1534                  * which might bounce very badly if there is contention.
1535                  *
1536                  * If the page is already locked, we don't need to
1537                  * handle it now - vmscan will handle it later if and
1538                  * when it attempts to reclaim the page.
1539                  */
1540                 if (page->mapping && trylock_page(page)) {
1541                         lru_add_drain();  /* push cached pages to LRU */
1542                         /*
1543                          * Because we lock page here and migration is
1544                          * blocked by the pte's page reference, we need
1545                          * only check for file-cache page truncation.
1546                          */
1547                         if (page->mapping)
1548                                 mlock_vma_page(page);
1549                         unlock_page(page);
1550                 }
1551         }
1552 unlock:
1553         pte_unmap_unlock(ptep, ptl);
1554 out:
1555         return page;
1556
1557 bad_page:
1558         pte_unmap_unlock(ptep, ptl);
1559         return ERR_PTR(-EFAULT);
1560
1561 no_page:
1562         pte_unmap_unlock(ptep, ptl);
1563         if (!pte_none(pte))
1564                 return page;
1565
1566 no_page_table:
1567         /*
1568          * When core dumping an enormous anonymous area that nobody
1569          * has touched so far, we don't want to allocate unnecessary pages or
1570          * page tables.  Return error instead of NULL to skip handle_mm_fault,
1571          * then get_dump_page() will return NULL to leave a hole in the dump.
1572          * But we can only make this optimization where a hole would surely
1573          * be zero-filled if handle_mm_fault() actually did handle it.
1574          */
1575         if ((flags & FOLL_DUMP) &&
1576             (!vma->vm_ops || !vma->vm_ops->fault))
1577                 return ERR_PTR(-EFAULT);
1578         return page;
1579 }
1580
1581 static inline int stack_guard_page(struct vm_area_struct *vma, unsigned long addr)
1582 {
1583         return stack_guard_page_start(vma, addr) ||
1584                stack_guard_page_end(vma, addr+PAGE_SIZE);
1585 }
1586
1587 /**
1588  * __get_user_pages() - pin user pages in memory
1589  * @tsk:        task_struct of target task
1590  * @mm:         mm_struct of target mm
1591  * @start:      starting user address
1592  * @nr_pages:   number of pages from start to pin
1593  * @gup_flags:  flags modifying pin behaviour
1594  * @pages:      array that receives pointers to the pages pinned.
1595  *              Should be at least nr_pages long. Or NULL, if caller
1596  *              only intends to ensure the pages are faulted in.
1597  * @vmas:       array of pointers to vmas corresponding to each page.
1598  *              Or NULL if the caller does not require them.
1599  * @nonblocking: whether waiting for disk IO or mmap_sem contention
1600  *
1601  * Returns number of pages pinned. This may be fewer than the number
1602  * requested. If nr_pages is 0 or negative, returns 0. If no pages
1603  * were pinned, returns -errno. Each page returned must be released
1604  * with a put_page() call when it is finished with. vmas will only
1605  * remain valid while mmap_sem is held.
1606  *
1607  * Must be called with mmap_sem held for read or write.
1608  *
1609  * __get_user_pages walks a process's page tables and takes a reference to
1610  * each struct page that each user address corresponds to at a given
1611  * instant. That is, it takes the page that would be accessed if a user
1612  * thread accesses the given user virtual address at that instant.
1613  *
1614  * This does not guarantee that the page exists in the user mappings when
1615  * __get_user_pages returns, and there may even be a completely different
1616  * page there in some cases (eg. if mmapped pagecache has been invalidated
1617  * and subsequently re faulted). However it does guarantee that the page
1618  * won't be freed completely. And mostly callers simply care that the page
1619  * contains data that was valid *at some point in time*. Typically, an IO
1620  * or similar operation cannot guarantee anything stronger anyway because
1621  * locks can't be held over the syscall boundary.
1622  *
1623  * If @gup_flags & FOLL_WRITE == 0, the page must not be written to. If
1624  * the page is written to, set_page_dirty (or set_page_dirty_lock, as
1625  * appropriate) must be called after the page is finished with, and
1626  * before put_page is called.
1627  *
1628  * If @nonblocking != NULL, __get_user_pages will not wait for disk IO
1629  * or mmap_sem contention, and if waiting is needed to pin all pages,
1630  * *@nonblocking will be set to 0.
1631  *
1632  * In most cases, get_user_pages or get_user_pages_fast should be used
1633  * instead of __get_user_pages. __get_user_pages should be used only if
1634  * you need some special @gup_flags.
1635  */
1636 int __get_user_pages(struct task_struct *tsk, struct mm_struct *mm,
1637                      unsigned long start, int nr_pages, unsigned int gup_flags,
1638                      struct page **pages, struct vm_area_struct **vmas,
1639                      int *nonblocking)
1640 {
1641         int i;
1642         unsigned long vm_flags;
1643
1644         if (nr_pages <= 0)
1645                 return 0;
1646
1647         VM_BUG_ON(!!pages != !!(gup_flags & FOLL_GET));
1648
1649         /* 
1650          * Require read or write permissions.
1651          * If FOLL_FORCE is set, we only require the "MAY" flags.
1652          */
1653         vm_flags  = (gup_flags & FOLL_WRITE) ?
1654                         (VM_WRITE | VM_MAYWRITE) : (VM_READ | VM_MAYREAD);
1655         vm_flags &= (gup_flags & FOLL_FORCE) ?
1656                         (VM_MAYREAD | VM_MAYWRITE) : (VM_READ | VM_WRITE);
1657         i = 0;
1658
1659         do {
1660                 struct vm_area_struct *vma;
1661
1662                 vma = find_extend_vma(mm, start);
1663                 if (!vma && in_gate_area(mm, start)) {
1664                         unsigned long pg = start & PAGE_MASK;
1665                         pgd_t *pgd;
1666                         pud_t *pud;
1667                         pmd_t *pmd;
1668                         pte_t *pte;
1669
1670                         /* user gate pages are read-only */
1671                         if (gup_flags & FOLL_WRITE)
1672                                 return i ? : -EFAULT;
1673                         if (pg > TASK_SIZE)
1674                                 pgd = pgd_offset_k(pg);
1675                         else
1676                                 pgd = pgd_offset_gate(mm, pg);
1677                         BUG_ON(pgd_none(*pgd));
1678                         pud = pud_offset(pgd, pg);
1679                         BUG_ON(pud_none(*pud));
1680                         pmd = pmd_offset(pud, pg);
1681                         if (pmd_none(*pmd))
1682                                 return i ? : -EFAULT;
1683                         VM_BUG_ON(pmd_trans_huge(*pmd));
1684                         pte = pte_offset_map(pmd, pg);
1685                         if (pte_none(*pte)) {
1686                                 pte_unmap(pte);
1687                                 return i ? : -EFAULT;
1688                         }
1689                         vma = get_gate_vma(mm);
1690                         if (pages) {
1691                                 struct page *page;
1692
1693                                 page = vm_normal_page(vma, start, *pte);
1694                                 if (!page) {
1695                                         if (!(gup_flags & FOLL_DUMP) &&
1696                                              is_zero_pfn(pte_pfn(*pte)))
1697                                                 page = pte_page(*pte);
1698                                         else {
1699                                                 pte_unmap(pte);
1700                                                 return i ? : -EFAULT;
1701                                         }
1702                                 }
1703                                 pages[i] = page;
1704                                 get_page(page);
1705                         }
1706                         pte_unmap(pte);
1707                         goto next_page;
1708                 }
1709
1710                 if (!vma ||
1711                     (vma->vm_flags & (VM_IO | VM_PFNMAP)) ||
1712                     !(vm_flags & vma->vm_flags))
1713                         return i ? : -EFAULT;
1714
1715                 if (is_vm_hugetlb_page(vma)) {
1716                         i = follow_hugetlb_page(mm, vma, pages, vmas,
1717                                         &start, &nr_pages, i, gup_flags);
1718                         continue;
1719                 }
1720
1721                 do {
1722                         struct page *page;
1723                         unsigned int foll_flags = gup_flags;
1724
1725                         /*
1726                          * If we have a pending SIGKILL, don't keep faulting
1727                          * pages and potentially allocating memory.
1728                          */
1729                         if (unlikely(fatal_signal_pending(current)))
1730                                 return i ? i : -ERESTARTSYS;
1731
1732                         cond_resched();
1733                         while (!(page = follow_page(vma, start, foll_flags))) {
1734                                 int ret;
1735                                 unsigned int fault_flags = 0;
1736
1737                                 /* For mlock, just skip the stack guard page. */
1738                                 if (foll_flags & FOLL_MLOCK) {
1739                                         if (stack_guard_page(vma, start))
1740                                                 goto next_page;
1741                                 }
1742                                 if (foll_flags & FOLL_WRITE)
1743                                         fault_flags |= FAULT_FLAG_WRITE;
1744                                 if (nonblocking)
1745                                         fault_flags |= FAULT_FLAG_ALLOW_RETRY;
1746                                 if (foll_flags & FOLL_NOWAIT)
1747                                         fault_flags |= (FAULT_FLAG_ALLOW_RETRY | FAULT_FLAG_RETRY_NOWAIT);
1748
1749                                 ret = handle_mm_fault(mm, vma, start,
1750                                                         fault_flags);
1751
1752                                 if (ret & VM_FAULT_ERROR) {
1753                                         if (ret & VM_FAULT_OOM)
1754                                                 return i ? i : -ENOMEM;
1755                                         if (ret & (VM_FAULT_HWPOISON |
1756                                                    VM_FAULT_HWPOISON_LARGE)) {
1757                                                 if (i)
1758                                                         return i;
1759                                                 else if (gup_flags & FOLL_HWPOISON)
1760                                                         return -EHWPOISON;
1761                                                 else
1762                                                         return -EFAULT;
1763                                         }
1764                                         if (ret & VM_FAULT_SIGBUS)
1765                                                 return i ? i : -EFAULT;
1766                                         BUG();
1767                                 }
1768
1769                                 if (tsk) {
1770                                         if (ret & VM_FAULT_MAJOR)
1771                                                 tsk->maj_flt++;
1772                                         else
1773                                                 tsk->min_flt++;
1774                                 }
1775
1776                                 if (ret & VM_FAULT_RETRY) {
1777                                         if (nonblocking)
1778                                                 *nonblocking = 0;
1779                                         return i;
1780                                 }
1781
1782                                 /*
1783                                  * The VM_FAULT_WRITE bit tells us that
1784                                  * do_wp_page has broken COW when necessary,
1785                                  * even if maybe_mkwrite decided not to set
1786                                  * pte_write. We can thus safely do subsequent
1787                                  * page lookups as if they were reads. But only
1788                                  * do so when looping for pte_write is futile:
1789                                  * in some cases userspace may also be wanting
1790                                  * to write to the gotten user page, which a
1791                                  * read fault here might prevent (a readonly
1792                                  * page might get reCOWed by userspace write).
1793                                  */
1794                                 if ((ret & VM_FAULT_WRITE) &&
1795                                     !(vma->vm_flags & VM_WRITE))
1796                                         foll_flags &= ~FOLL_WRITE;
1797
1798                                 cond_resched();
1799                         }
1800                         if (IS_ERR(page))
1801                                 return i ? i : PTR_ERR(page);
1802                         if (pages) {
1803                                 pages[i] = page;
1804
1805                                 flush_anon_page(vma, page, start);
1806                                 flush_dcache_page(page);
1807                         }
1808 next_page:
1809                         if (vmas)
1810                                 vmas[i] = vma;
1811                         i++;
1812                         start += PAGE_SIZE;
1813                         nr_pages--;
1814                 } while (nr_pages && start < vma->vm_end);
1815         } while (nr_pages);
1816         return i;
1817 }
1818 EXPORT_SYMBOL(__get_user_pages);
1819
1820 /*
1821  * fixup_user_fault() - manually resolve a user page fault
1822  * @tsk:        the task_struct to use for page fault accounting, or
1823  *              NULL if faults are not to be recorded.
1824  * @mm:         mm_struct of target mm
1825  * @address:    user address
1826  * @fault_flags:flags to pass down to handle_mm_fault()
1827  *
1828  * This is meant to be called in the specific scenario where for locking reasons
1829  * we try to access user memory in atomic context (within a pagefault_disable()
1830  * section), this returns -EFAULT, and we want to resolve the user fault before
1831  * trying again.
1832  *
1833  * Typically this is meant to be used by the futex code.
1834  *
1835  * The main difference with get_user_pages() is that this function will
1836  * unconditionally call handle_mm_fault() which will in turn perform all the
1837  * necessary SW fixup of the dirty and young bits in the PTE, while
1838  * handle_mm_fault() only guarantees to update these in the struct page.
1839  *
1840  * This is important for some architectures where those bits also gate the
1841  * access permission to the page because they are maintained in software.  On
1842  * such architectures, gup() will not be enough to make a subsequent access
1843  * succeed.
1844  *
1845  * This should be called with the mm_sem held for read.
1846  */
1847 int fixup_user_fault(struct task_struct *tsk, struct mm_struct *mm,
1848                      unsigned long address, unsigned int fault_flags)
1849 {
1850         struct vm_area_struct *vma;
1851         int ret;
1852
1853         vma = find_extend_vma(mm, address);
1854         if (!vma || address < vma->vm_start)
1855                 return -EFAULT;
1856
1857         ret = handle_mm_fault(mm, vma, address, fault_flags);
1858         if (ret & VM_FAULT_ERROR) {
1859                 if (ret & VM_FAULT_OOM)
1860                         return -ENOMEM;
1861                 if (ret & (VM_FAULT_HWPOISON | VM_FAULT_HWPOISON_LARGE))
1862                         return -EHWPOISON;
1863                 if (ret & VM_FAULT_SIGBUS)
1864                         return -EFAULT;
1865                 BUG();
1866         }
1867         if (tsk) {
1868                 if (ret & VM_FAULT_MAJOR)
1869                         tsk->maj_flt++;
1870                 else
1871                         tsk->min_flt++;
1872         }
1873         return 0;
1874 }
1875
1876 /*
1877  * get_user_pages() - pin user pages in memory
1878  * @tsk:        the task_struct to use for page fault accounting, or
1879  *              NULL if faults are not to be recorded.
1880  * @mm:         mm_struct of target mm
1881  * @start:      starting user address
1882  * @nr_pages:   number of pages from start to pin
1883  * @write:      whether pages will be written to by the caller
1884  * @force:      whether to force write access even if user mapping is
1885  *              readonly. This will result in the page being COWed even
1886  *              in MAP_SHARED mappings. You do not want this.
1887  * @pages:      array that receives pointers to the pages pinned.
1888  *              Should be at least nr_pages long. Or NULL, if caller
1889  *              only intends to ensure the pages are faulted in.
1890  * @vmas:       array of pointers to vmas corresponding to each page.
1891  *              Or NULL if the caller does not require them.
1892  *
1893  * Returns number of pages pinned. This may be fewer than the number
1894  * requested. If nr_pages is 0 or negative, returns 0. If no pages
1895  * were pinned, returns -errno. Each page returned must be released
1896  * with a put_page() call when it is finished with. vmas will only
1897  * remain valid while mmap_sem is held.
1898  *
1899  * Must be called with mmap_sem held for read or write.
1900  *
1901  * get_user_pages walks a process's page tables and takes a reference to
1902  * each struct page that each user address corresponds to at a given
1903  * instant. That is, it takes the page that would be accessed if a user
1904  * thread accesses the given user virtual address at that instant.
1905  *
1906  * This does not guarantee that the page exists in the user mappings when
1907  * get_user_pages returns, and there may even be a completely different
1908  * page there in some cases (eg. if mmapped pagecache has been invalidated
1909  * and subsequently re faulted). However it does guarantee that the page
1910  * won't be freed completely. And mostly callers simply care that the page
1911  * contains data that was valid *at some point in time*. Typically, an IO
1912  * or similar operation cannot guarantee anything stronger anyway because
1913  * locks can't be held over the syscall boundary.
1914  *
1915  * If write=0, the page must not be written to. If the page is written to,
1916  * set_page_dirty (or set_page_dirty_lock, as appropriate) must be called
1917  * after the page is finished with, and before put_page is called.
1918  *
1919  * get_user_pages is typically used for fewer-copy IO operations, to get a
1920  * handle on the memory by some means other than accesses via the user virtual
1921  * addresses. The pages may be submitted for DMA to devices or accessed via
1922  * their kernel linear mapping (via the kmap APIs). Care should be taken to
1923  * use the correct cache flushing APIs.
1924  *
1925  * See also get_user_pages_fast, for performance critical applications.
1926  */
1927 int get_user_pages(struct task_struct *tsk, struct mm_struct *mm,
1928                 unsigned long start, int nr_pages, int write, int force,
1929                 struct page **pages, struct vm_area_struct **vmas)
1930 {
1931         int flags = FOLL_TOUCH;
1932
1933         if (pages)
1934                 flags |= FOLL_GET;
1935         if (write)
1936                 flags |= FOLL_WRITE;
1937         if (force)
1938                 flags |= FOLL_FORCE;
1939
1940         return __get_user_pages(tsk, mm, start, nr_pages, flags, pages, vmas,
1941                                 NULL);
1942 }
1943 EXPORT_SYMBOL(get_user_pages);
1944
1945 /**
1946  * get_dump_page() - pin user page in memory while writing it to core dump
1947  * @addr: user address
1948  *
1949  * Returns struct page pointer of user page pinned for dump,
1950  * to be freed afterwards by page_cache_release() or put_page().
1951  *
1952  * Returns NULL on any kind of failure - a hole must then be inserted into
1953  * the corefile, to preserve alignment with its headers; and also returns
1954  * NULL wherever the ZERO_PAGE, or an anonymous pte_none, has been found -
1955  * allowing a hole to be left in the corefile to save diskspace.
1956  *
1957  * Called without mmap_sem, but after all other threads have been killed.
1958  */
1959 #ifdef CONFIG_ELF_CORE
1960 struct page *get_dump_page(unsigned long addr)
1961 {
1962         struct vm_area_struct *vma;
1963         struct page *page;
1964
1965         if (__get_user_pages(current, current->mm, addr, 1,
1966                              FOLL_FORCE | FOLL_DUMP | FOLL_GET, &page, &vma,
1967                              NULL) < 1)
1968                 return NULL;
1969         flush_cache_page(vma, addr, page_to_pfn(page));
1970         return page;
1971 }
1972 #endif /* CONFIG_ELF_CORE */
1973
1974 pte_t *__get_locked_pte(struct mm_struct *mm, unsigned long addr,
1975                         spinlock_t **ptl)
1976 {
1977         pgd_t * pgd = pgd_offset(mm, addr);
1978         pud_t * pud = pud_alloc(mm, pgd, addr);
1979         if (pud) {
1980                 pmd_t * pmd = pmd_alloc(mm, pud, addr);
1981                 if (pmd) {
1982                         VM_BUG_ON(pmd_trans_huge(*pmd));
1983                         return pte_alloc_map_lock(mm, pmd, addr, ptl);
1984                 }
1985         }
1986         return NULL;
1987 }
1988
1989 /*
1990  * This is the old fallback for page remapping.
1991  *
1992  * For historical reasons, it only allows reserved pages. Only
1993  * old drivers should use this, and they needed to mark their
1994  * pages reserved for the old functions anyway.
1995  */
1996 static int insert_page(struct vm_area_struct *vma, unsigned long addr,
1997                         struct page *page, pgprot_t prot)
1998 {
1999         struct mm_struct *mm = vma->vm_mm;
2000         int retval;
2001         pte_t *pte;
2002         spinlock_t *ptl;
2003
2004         retval = -EINVAL;
2005         if (PageAnon(page))
2006                 goto out;
2007         retval = -ENOMEM;
2008         flush_dcache_page(page);
2009         pte = get_locked_pte(mm, addr, &ptl);
2010         if (!pte)
2011                 goto out;
2012         retval = -EBUSY;
2013         if (!pte_none(*pte))
2014                 goto out_unlock;
2015
2016         /* Ok, finally just insert the thing.. */
2017         get_page(page);
2018         inc_mm_counter_fast(mm, MM_FILEPAGES);
2019         page_add_file_rmap(page);
2020         set_pte_at(mm, addr, pte, mk_pte(page, prot));
2021
2022         retval = 0;
2023         pte_unmap_unlock(pte, ptl);
2024         return retval;
2025 out_unlock:
2026         pte_unmap_unlock(pte, ptl);
2027 out:
2028         return retval;
2029 }
2030
2031 /**
2032  * vm_insert_page - insert single page into user vma
2033  * @vma: user vma to map to
2034  * @addr: target user address of this page
2035  * @page: source kernel page
2036  *
2037  * This allows drivers to insert individual pages they've allocated
2038  * into a user vma.
2039  *
2040  * The page has to be a nice clean _individual_ kernel allocation.
2041  * If you allocate a compound page, you need to have marked it as
2042  * such (__GFP_COMP), or manually just split the page up yourself
2043  * (see split_page()).
2044  *
2045  * NOTE! Traditionally this was done with "remap_pfn_range()" which
2046  * took an arbitrary page protection parameter. This doesn't allow
2047  * that. Your vma protection will have to be set up correctly, which
2048  * means that if you want a shared writable mapping, you'd better
2049  * ask for a shared writable mapping!
2050  *
2051  * The page does not need to be reserved.
2052  */
2053 int vm_insert_page(struct vm_area_struct *vma, unsigned long addr,
2054                         struct page *page)
2055 {
2056         if (addr < vma->vm_start || addr >= vma->vm_end)
2057                 return -EFAULT;
2058         if (!page_count(page))
2059                 return -EINVAL;
2060         vma->vm_flags |= VM_INSERTPAGE;
2061         return insert_page(vma, addr, page, vma->vm_page_prot);
2062 }
2063 EXPORT_SYMBOL(vm_insert_page);
2064
2065 static int insert_pfn(struct vm_area_struct *vma, unsigned long addr,
2066                         unsigned long pfn, pgprot_t prot)
2067 {
2068         struct mm_struct *mm = vma->vm_mm;
2069         int retval;
2070         pte_t *pte, entry;
2071         spinlock_t *ptl;
2072
2073         retval = -ENOMEM;
2074         pte = get_locked_pte(mm, addr, &ptl);
2075         if (!pte)
2076                 goto out;
2077         retval = -EBUSY;
2078         if (!pte_none(*pte))
2079                 goto out_unlock;
2080
2081         /* Ok, finally just insert the thing.. */
2082         entry = pte_mkspecial(pfn_pte(pfn, prot));
2083         set_pte_at(mm, addr, pte, entry);
2084         update_mmu_cache(vma, addr, pte); /* XXX: why not for insert_page? */
2085
2086         retval = 0;
2087 out_unlock:
2088         pte_unmap_unlock(pte, ptl);
2089 out:
2090         return retval;
2091 }
2092
2093 /**
2094  * vm_insert_pfn - insert single pfn into user vma
2095  * @vma: user vma to map to
2096  * @addr: target user address of this page
2097  * @pfn: source kernel pfn
2098  *
2099  * Similar to vm_inert_page, this allows drivers to insert individual pages
2100  * they've allocated into a user vma. Same comments apply.
2101  *
2102  * This function should only be called from a vm_ops->fault handler, and
2103  * in that case the handler should return NULL.
2104  *
2105  * vma cannot be a COW mapping.
2106  *
2107  * As this is called only for pages that do not currently exist, we
2108  * do not need to flush old virtual caches or the TLB.
2109  */
2110 int vm_insert_pfn(struct vm_area_struct *vma, unsigned long addr,
2111                         unsigned long pfn)
2112 {
2113         int ret;
2114         pgprot_t pgprot = vma->vm_page_prot;
2115         /*
2116          * Technically, architectures with pte_special can avoid all these
2117          * restrictions (same for remap_pfn_range).  However we would like
2118          * consistency in testing and feature parity among all, so we should
2119          * try to keep these invariants in place for everybody.
2120          */
2121         BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)));
2122         BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) ==
2123                                                 (VM_PFNMAP|VM_MIXEDMAP));
2124         BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags));
2125         BUG_ON((vma->vm_flags & VM_MIXEDMAP) && pfn_valid(pfn));
2126
2127         if (addr < vma->vm_start || addr >= vma->vm_end)
2128                 return -EFAULT;
2129         if (track_pfn_vma_new(vma, &pgprot, pfn, PAGE_SIZE))
2130                 return -EINVAL;
2131
2132         ret = insert_pfn(vma, addr, pfn, pgprot);
2133
2134         if (ret)
2135                 untrack_pfn_vma(vma, pfn, PAGE_SIZE);
2136
2137         return ret;
2138 }
2139 EXPORT_SYMBOL(vm_insert_pfn);
2140
2141 int vm_insert_mixed(struct vm_area_struct *vma, unsigned long addr,
2142                         unsigned long pfn)
2143 {
2144         BUG_ON(!(vma->vm_flags & VM_MIXEDMAP));
2145
2146         if (addr < vma->vm_start || addr >= vma->vm_end)
2147                 return -EFAULT;
2148
2149         /*
2150          * If we don't have pte special, then we have to use the pfn_valid()
2151          * based VM_MIXEDMAP scheme (see vm_normal_page), and thus we *must*
2152          * refcount the page if pfn_valid is true (hence insert_page rather
2153          * than insert_pfn).  If a zero_pfn were inserted into a VM_MIXEDMAP
2154          * without pte special, it would there be refcounted as a normal page.
2155          */
2156         if (!HAVE_PTE_SPECIAL && pfn_valid(pfn)) {
2157                 struct page *page;
2158
2159                 page = pfn_to_page(pfn);
2160                 return insert_page(vma, addr, page, vma->vm_page_prot);
2161         }
2162         return insert_pfn(vma, addr, pfn, vma->vm_page_prot);
2163 }
2164 EXPORT_SYMBOL(vm_insert_mixed);
2165
2166 /*
2167  * maps a range of physical memory into the requested pages. the old
2168  * mappings are removed. any references to nonexistent pages results
2169  * in null mappings (currently treated as "copy-on-access")
2170  */
2171 static int remap_pte_range(struct mm_struct *mm, pmd_t *pmd,
2172                         unsigned long addr, unsigned long end,
2173                         unsigned long pfn, pgprot_t prot)
2174 {
2175         pte_t *pte;
2176         spinlock_t *ptl;
2177
2178         pte = pte_alloc_map_lock(mm, pmd, addr, &ptl);
2179         if (!pte)
2180                 return -ENOMEM;
2181         arch_enter_lazy_mmu_mode();
2182         do {
2183                 BUG_ON(!pte_none(*pte));
2184                 set_pte_at(mm, addr, pte, pte_mkspecial(pfn_pte(pfn, prot)));
2185                 pfn++;
2186         } while (pte++, addr += PAGE_SIZE, addr != end);
2187         arch_leave_lazy_mmu_mode();
2188         pte_unmap_unlock(pte - 1, ptl);
2189         return 0;
2190 }
2191
2192 static inline int remap_pmd_range(struct mm_struct *mm, pud_t *pud,
2193                         unsigned long addr, unsigned long end,
2194                         unsigned long pfn, pgprot_t prot)
2195 {
2196         pmd_t *pmd;
2197         unsigned long next;
2198
2199         pfn -= addr >> PAGE_SHIFT;
2200         pmd = pmd_alloc(mm, pud, addr);
2201         if (!pmd)
2202                 return -ENOMEM;
2203         VM_BUG_ON(pmd_trans_huge(*pmd));
2204         do {
2205                 next = pmd_addr_end(addr, end);
2206                 if (remap_pte_range(mm, pmd, addr, next,
2207                                 pfn + (addr >> PAGE_SHIFT), prot))
2208                         return -ENOMEM;
2209         } while (pmd++, addr = next, addr != end);
2210         return 0;
2211 }
2212
2213 static inline int remap_pud_range(struct mm_struct *mm, pgd_t *pgd,
2214                         unsigned long addr, unsigned long end,
2215                         unsigned long pfn, pgprot_t prot)
2216 {
2217         pud_t *pud;
2218         unsigned long next;
2219
2220         pfn -= addr >> PAGE_SHIFT;
2221         pud = pud_alloc(mm, pgd, addr);
2222         if (!pud)
2223                 return -ENOMEM;
2224         do {
2225                 next = pud_addr_end(addr, end);
2226                 if (remap_pmd_range(mm, pud, addr, next,
2227                                 pfn + (addr >> PAGE_SHIFT), prot))
2228                         return -ENOMEM;
2229         } while (pud++, addr = next, addr != end);
2230         return 0;
2231 }
2232
2233 /**
2234  * remap_pfn_range - remap kernel memory to userspace
2235  * @vma: user vma to map to
2236  * @addr: target user address to start at
2237  * @pfn: physical address of kernel memory
2238  * @size: size of map area
2239  * @prot: page protection flags for this mapping
2240  *
2241  *  Note: this is only safe if the mm semaphore is held when called.
2242  */
2243 int remap_pfn_range(struct vm_area_struct *vma, unsigned long addr,
2244                     unsigned long pfn, unsigned long size, pgprot_t prot)
2245 {
2246         pgd_t *pgd;
2247         unsigned long next;
2248         unsigned long end = addr + PAGE_ALIGN(size);
2249         struct mm_struct *mm = vma->vm_mm;
2250         int err;
2251
2252         /*
2253          * Physically remapped pages are special. Tell the
2254          * rest of the world about it:
2255          *   VM_IO tells people not to look at these pages
2256          *      (accesses can have side effects).
2257          *   VM_RESERVED is specified all over the place, because
2258          *      in 2.4 it kept swapout's vma scan off this vma; but
2259          *      in 2.6 the LRU scan won't even find its pages, so this
2260          *      flag means no more than count its pages in reserved_vm,
2261          *      and omit it from core dump, even when VM_IO turned off.
2262          *   VM_PFNMAP tells the core MM that the base pages are just
2263          *      raw PFN mappings, and do not have a "struct page" associated
2264          *      with them.
2265          *
2266          * There's a horrible special case to handle copy-on-write
2267          * behaviour that some programs depend on. We mark the "original"
2268          * un-COW'ed pages by matching them up with "vma->vm_pgoff".
2269          */
2270         if (addr == vma->vm_start && end == vma->vm_end) {
2271                 vma->vm_pgoff = pfn;
2272                 vma->vm_flags |= VM_PFN_AT_MMAP;
2273         } else if (is_cow_mapping(vma->vm_flags))
2274                 return -EINVAL;
2275
2276         vma->vm_flags |= VM_IO | VM_RESERVED | VM_PFNMAP;
2277
2278         err = track_pfn_vma_new(vma, &prot, pfn, PAGE_ALIGN(size));
2279         if (err) {
2280                 /*
2281                  * To indicate that track_pfn related cleanup is not
2282                  * needed from higher level routine calling unmap_vmas
2283                  */
2284                 vma->vm_flags &= ~(VM_IO | VM_RESERVED | VM_PFNMAP);
2285                 vma->vm_flags &= ~VM_PFN_AT_MMAP;
2286                 return -EINVAL;
2287         }
2288
2289         BUG_ON(addr >= end);
2290         pfn -= addr >> PAGE_SHIFT;
2291         pgd = pgd_offset(mm, addr);
2292         flush_cache_range(vma, addr, end);
2293         do {
2294                 next = pgd_addr_end(addr, end);
2295                 err = remap_pud_range(mm, pgd, addr, next,
2296                                 pfn + (addr >> PAGE_SHIFT), prot);
2297                 if (err)
2298                         break;
2299         } while (pgd++, addr = next, addr != end);
2300
2301         if (err)
2302                 untrack_pfn_vma(vma, pfn, PAGE_ALIGN(size));
2303
2304         return err;
2305 }
2306 EXPORT_SYMBOL(remap_pfn_range);
2307
2308 static int apply_to_pte_range(struct mm_struct *mm, pmd_t *pmd,
2309                                      unsigned long addr, unsigned long end,
2310                                      pte_fn_t fn, void *data)
2311 {
2312         pte_t *pte;
2313         int err;
2314         pgtable_t token;
2315         spinlock_t *uninitialized_var(ptl);
2316
2317         pte = (mm == &init_mm) ?
2318                 pte_alloc_kernel(pmd, addr) :
2319                 pte_alloc_map_lock(mm, pmd, addr, &ptl);
2320         if (!pte)
2321                 return -ENOMEM;
2322
2323         BUG_ON(pmd_huge(*pmd));
2324
2325         arch_enter_lazy_mmu_mode();
2326
2327         token = pmd_pgtable(*pmd);
2328
2329         do {
2330                 err = fn(pte++, token, addr, data);
2331                 if (err)
2332                         break;
2333         } while (addr += PAGE_SIZE, addr != end);
2334
2335         arch_leave_lazy_mmu_mode();
2336
2337         if (mm != &init_mm)
2338                 pte_unmap_unlock(pte-1, ptl);
2339         return err;
2340 }
2341
2342 static int apply_to_pmd_range(struct mm_struct *mm, pud_t *pud,
2343                                      unsigned long addr, unsigned long end,
2344                                      pte_fn_t fn, void *data)
2345 {
2346         pmd_t *pmd;
2347         unsigned long next;
2348         int err;
2349
2350         BUG_ON(pud_huge(*pud));
2351
2352         pmd = pmd_alloc(mm, pud, addr);
2353         if (!pmd)
2354                 return -ENOMEM;
2355         do {
2356                 next = pmd_addr_end(addr, end);
2357                 err = apply_to_pte_range(mm, pmd, addr, next, fn, data);
2358                 if (err)
2359                         break;
2360         } while (pmd++, addr = next, addr != end);
2361         return err;
2362 }
2363
2364 static int apply_to_pud_range(struct mm_struct *mm, pgd_t *pgd,
2365                                      unsigned long addr, unsigned long end,
2366                                      pte_fn_t fn, void *data)
2367 {
2368         pud_t *pud;
2369         unsigned long next;
2370         int err;
2371
2372         pud = pud_alloc(mm, pgd, addr);
2373         if (!pud)
2374                 return -ENOMEM;
2375         do {
2376                 next = pud_addr_end(addr, end);
2377                 err = apply_to_pmd_range(mm, pud, addr, next, fn, data);
2378                 if (err)
2379                         break;
2380         } while (pud++, addr = next, addr != end);
2381         return err;
2382 }
2383
2384 /*
2385  * Scan a region of virtual memory, filling in page tables as necessary
2386  * and calling a provided function on each leaf page table.
2387  */
2388 int apply_to_page_range(struct mm_struct *mm, unsigned long addr,
2389                         unsigned long size, pte_fn_t fn, void *data)
2390 {
2391         pgd_t *pgd;
2392         unsigned long next;
2393         unsigned long end = addr + size;
2394         int err;
2395
2396         BUG_ON(addr >= end);
2397         pgd = pgd_offset(mm, addr);
2398         do {
2399                 next = pgd_addr_end(addr, end);
2400                 err = apply_to_pud_range(mm, pgd, addr, next, fn, data);
2401                 if (err)
2402                         break;
2403         } while (pgd++, addr = next, addr != end);
2404
2405         return err;
2406 }
2407 EXPORT_SYMBOL_GPL(apply_to_page_range);
2408
2409 /*
2410  * handle_pte_fault chooses page fault handler according to an entry
2411  * which was read non-atomically.  Before making any commitment, on
2412  * those architectures or configurations (e.g. i386 with PAE) which
2413  * might give a mix of unmatched parts, do_swap_page and do_nonlinear_fault
2414  * must check under lock before unmapping the pte and proceeding
2415  * (but do_wp_page is only called after already making such a check;
2416  * and do_anonymous_page can safely check later on).
2417  */
2418 static inline int pte_unmap_same(struct mm_struct *mm, pmd_t *pmd,
2419                                 pte_t *page_table, pte_t orig_pte)
2420 {
2421         int same = 1;
2422 #if defined(CONFIG_SMP) || defined(CONFIG_PREEMPT)
2423         if (sizeof(pte_t) > sizeof(unsigned long)) {
2424                 spinlock_t *ptl = pte_lockptr(mm, pmd);
2425                 spin_lock(ptl);
2426                 same = pte_same(*page_table, orig_pte);
2427                 spin_unlock(ptl);
2428         }
2429 #endif
2430         pte_unmap(page_table);
2431         return same;
2432 }
2433
2434 static inline void cow_user_page(struct page *dst, struct page *src, unsigned long va, struct vm_area_struct *vma)
2435 {
2436         /*
2437          * If the source page was a PFN mapping, we don't have
2438          * a "struct page" for it. We do a best-effort copy by
2439          * just copying from the original user address. If that
2440          * fails, we just zero-fill it. Live with it.
2441          */
2442         if (unlikely(!src)) {
2443                 void *kaddr = kmap_atomic(dst, KM_USER0);
2444                 void __user *uaddr = (void __user *)(va & PAGE_MASK);
2445
2446                 /*
2447                  * This really shouldn't fail, because the page is there
2448                  * in the page tables. But it might just be unreadable,
2449                  * in which case we just give up and fill the result with
2450                  * zeroes.
2451                  */
2452                 if (__copy_from_user_inatomic(kaddr, uaddr, PAGE_SIZE))
2453                         clear_page(kaddr);
2454                 kunmap_atomic(kaddr, KM_USER0);
2455                 flush_dcache_page(dst);
2456         } else
2457                 copy_user_highpage(dst, src, va, vma);
2458 }
2459
2460 /*
2461  * This routine handles present pages, when users try to write
2462  * to a shared page. It is done by copying the page to a new address
2463  * and decrementing the shared-page counter for the old page.
2464  *
2465  * Note that this routine assumes that the protection checks have been
2466  * done by the caller (the low-level page fault routine in most cases).
2467  * Thus we can safely just mark it writable once we've done any necessary
2468  * COW.
2469  *
2470  * We also mark the page dirty at this point even though the page will
2471  * change only once the write actually happens. This avoids a few races,
2472  * and potentially makes it more efficient.
2473  *
2474  * We enter with non-exclusive mmap_sem (to exclude vma changes,
2475  * but allow concurrent faults), with pte both mapped and locked.
2476  * We return with mmap_sem still held, but pte unmapped and unlocked.
2477  */
2478 static int do_wp_page(struct mm_struct *mm, struct vm_area_struct *vma,
2479                 unsigned long address, pte_t *page_table, pmd_t *pmd,
2480                 spinlock_t *ptl, pte_t orig_pte)
2481         __releases(ptl)
2482 {
2483         struct page *old_page, *new_page;
2484         pte_t entry;
2485         int ret = 0;
2486         int page_mkwrite = 0;
2487         struct page *dirty_page = NULL;
2488
2489         old_page = vm_normal_page(vma, address, orig_pte);
2490         if (!old_page) {
2491                 /*
2492                  * VM_MIXEDMAP !pfn_valid() case
2493                  *
2494                  * We should not cow pages in a shared writeable mapping.
2495                  * Just mark the pages writable as we can't do any dirty
2496                  * accounting on raw pfn maps.
2497                  */
2498                 if ((vma->vm_flags & (VM_WRITE|VM_SHARED)) ==
2499                                      (VM_WRITE|VM_SHARED))
2500                         goto reuse;
2501                 goto gotten;
2502         }
2503
2504         /*
2505          * Take out anonymous pages first, anonymous shared vmas are
2506          * not dirty accountable.
2507          */
2508         if (PageAnon(old_page) && !PageKsm(old_page)) {
2509                 if (!trylock_page(old_page)) {
2510                         page_cache_get(old_page);
2511                         pte_unmap_unlock(page_table, ptl);
2512                         lock_page(old_page);
2513                         page_table = pte_offset_map_lock(mm, pmd, address,
2514                                                          &ptl);
2515                         if (!pte_same(*page_table, orig_pte)) {
2516                                 unlock_page(old_page);
2517                                 goto unlock;
2518                         }
2519                         page_cache_release(old_page);
2520                 }
2521                 if (reuse_swap_page(old_page)) {
2522                         /*
2523                          * The page is all ours.  Move it to our anon_vma so
2524                          * the rmap code will not search our parent or siblings.
2525                          * Protected against the rmap code by the page lock.
2526                          */
2527                         page_move_anon_rmap(old_page, vma, address);
2528                         unlock_page(old_page);
2529                         goto reuse;
2530                 }
2531                 unlock_page(old_page);
2532         } else if (unlikely((vma->vm_flags & (VM_WRITE|VM_SHARED)) ==
2533                                         (VM_WRITE|VM_SHARED))) {
2534                 /*
2535                  * Only catch write-faults on shared writable pages,
2536                  * read-only shared pages can get COWed by
2537                  * get_user_pages(.write=1, .force=1).
2538                  */
2539                 if (vma->vm_ops && vma->vm_ops->page_mkwrite) {
2540                         struct vm_fault vmf;
2541                         int tmp;
2542
2543                         vmf.virtual_address = (void __user *)(address &
2544                                                                 PAGE_MASK);
2545                         vmf.pgoff = old_page->index;
2546                         vmf.flags = FAULT_FLAG_WRITE|FAULT_FLAG_MKWRITE;
2547                         vmf.page = old_page;
2548
2549                         /*
2550                          * Notify the address space that the page is about to
2551                          * become writable so that it can prohibit this or wait
2552                          * for the page to get into an appropriate state.
2553                          *
2554                          * We do this without the lock held, so that it can
2555                          * sleep if it needs to.
2556                          */
2557                         page_cache_get(old_page);
2558                         pte_unmap_unlock(page_table, ptl);
2559
2560                         tmp = vma->vm_ops->page_mkwrite(vma, &vmf);
2561                         if (unlikely(tmp &
2562                                         (VM_FAULT_ERROR | VM_FAULT_NOPAGE))) {
2563                                 ret = tmp;
2564                                 goto unwritable_page;
2565                         }
2566                         if (unlikely(!(tmp & VM_FAULT_LOCKED))) {
2567                                 lock_page(old_page);
2568                                 if (!old_page->mapping) {
2569                                         ret = 0; /* retry the fault */
2570                                         unlock_page(old_page);
2571                                         goto unwritable_page;
2572                                 }
2573                         } else
2574                                 VM_BUG_ON(!PageLocked(old_page));
2575
2576                         /*
2577                          * Since we dropped the lock we need to revalidate
2578                          * the PTE as someone else may have changed it.  If
2579                          * they did, we just return, as we can count on the
2580                          * MMU to tell us if they didn't also make it writable.
2581                          */
2582                         page_table = pte_offset_map_lock(mm, pmd, address,
2583                                                          &ptl);
2584                         if (!pte_same(*page_table, orig_pte)) {
2585                                 unlock_page(old_page);
2586                                 goto unlock;
2587                         }
2588
2589                         page_mkwrite = 1;
2590                 }
2591                 dirty_page = old_page;
2592                 get_page(dirty_page);
2593
2594 reuse:
2595                 flush_cache_page(vma, address, pte_pfn(orig_pte));
2596                 entry = pte_mkyoung(orig_pte);
2597                 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2598                 if (ptep_set_access_flags(vma, address, page_table, entry,1))
2599                         update_mmu_cache(vma, address, page_table);
2600                 pte_unmap_unlock(page_table, ptl);
2601                 ret |= VM_FAULT_WRITE;
2602
2603                 if (!dirty_page)
2604                         return ret;
2605
2606                 /*
2607                  * Yes, Virginia, this is actually required to prevent a race
2608                  * with clear_page_dirty_for_io() from clearing the page dirty
2609                  * bit after it clear all dirty ptes, but before a racing
2610                  * do_wp_page installs a dirty pte.
2611                  *
2612                  * __do_fault is protected similarly.
2613                  */
2614                 if (!page_mkwrite) {
2615                         wait_on_page_locked(dirty_page);
2616                         set_page_dirty_balance(dirty_page, page_mkwrite);
2617                 }
2618                 put_page(dirty_page);
2619                 if (page_mkwrite) {
2620                         struct address_space *mapping = dirty_page->mapping;
2621
2622                         set_page_dirty(dirty_page);
2623                         unlock_page(dirty_page);
2624                         page_cache_release(dirty_page);
2625                         if (mapping)    {
2626                                 /*
2627                                  * Some device drivers do not set page.mapping
2628                                  * but still dirty their pages
2629                                  */
2630                                 balance_dirty_pages_ratelimited(mapping);
2631                         }
2632                 }
2633
2634                 /* file_update_time outside page_lock */
2635                 if (vma->vm_file)
2636                         file_update_time(vma->vm_file);
2637
2638                 return ret;
2639         }
2640
2641         /*
2642          * Ok, we need to copy. Oh, well..
2643          */
2644         page_cache_get(old_page);
2645 gotten:
2646         pte_unmap_unlock(page_table, ptl);
2647
2648         if (unlikely(anon_vma_prepare(vma)))
2649                 goto oom;
2650
2651         if (is_zero_pfn(pte_pfn(orig_pte))) {
2652                 new_page = alloc_zeroed_user_highpage_movable(vma, address);
2653                 if (!new_page)
2654                         goto oom;
2655         } else {
2656                 new_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, address);
2657                 if (!new_page)
2658                         goto oom;
2659                 cow_user_page(new_page, old_page, address, vma);
2660         }
2661         __SetPageUptodate(new_page);
2662
2663         if (mem_cgroup_newpage_charge(new_page, mm, GFP_KERNEL))
2664                 goto oom_free_new;
2665
2666         /*
2667          * Re-check the pte - we dropped the lock
2668          */
2669         page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2670         if (likely(pte_same(*page_table, orig_pte))) {
2671                 if (old_page) {
2672                         if (!PageAnon(old_page)) {
2673                                 dec_mm_counter_fast(mm, MM_FILEPAGES);
2674                                 inc_mm_counter_fast(mm, MM_ANONPAGES);
2675                         }
2676                 } else
2677                         inc_mm_counter_fast(mm, MM_ANONPAGES);
2678                 flush_cache_page(vma, address, pte_pfn(orig_pte));
2679                 entry = mk_pte(new_page, vma->vm_page_prot);
2680                 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2681                 /*
2682                  * Clear the pte entry and flush it first, before updating the
2683                  * pte with the new entry. This will avoid a race condition
2684                  * seen in the presence of one thread doing SMC and another
2685                  * thread doing COW.
2686                  */
2687                 ptep_clear_flush(vma, address, page_table);
2688                 page_add_new_anon_rmap(new_page, vma, address);
2689                 /*
2690                  * We call the notify macro here because, when using secondary
2691                  * mmu page tables (such as kvm shadow page tables), we want the
2692                  * new page to be mapped directly into the secondary page table.
2693                  */
2694                 set_pte_at_notify(mm, address, page_table, entry);
2695                 update_mmu_cache(vma, address, page_table);
2696                 if (old_page) {
2697                         /*
2698                          * Only after switching the pte to the new page may
2699                          * we remove the mapcount here. Otherwise another
2700                          * process may come and find the rmap count decremented
2701                          * before the pte is switched to the new page, and
2702                          * "reuse" the old page writing into it while our pte
2703                          * here still points into it and can be read by other
2704                          * threads.
2705                          *
2706                          * The critical issue is to order this
2707                          * page_remove_rmap with the ptp_clear_flush above.
2708                          * Those stores are ordered by (if nothing else,)
2709                          * the barrier present in the atomic_add_negative
2710                          * in page_remove_rmap.
2711                          *
2712                          * Then the TLB flush in ptep_clear_flush ensures that
2713                          * no process can access the old page before the
2714                          * decremented mapcount is visible. And the old page
2715                          * cannot be reused until after the decremented
2716                          * mapcount is visible. So transitively, TLBs to
2717                          * old page will be flushed before it can be reused.
2718                          */
2719                         page_remove_rmap(old_page);
2720                 }
2721
2722                 /* Free the old page.. */
2723                 new_page = old_page;
2724                 ret |= VM_FAULT_WRITE;
2725         } else
2726                 mem_cgroup_uncharge_page(new_page);
2727
2728         if (new_page)
2729                 page_cache_release(new_page);
2730 unlock:
2731         pte_unmap_unlock(page_table, ptl);
2732         if (old_page) {
2733                 /*
2734                  * Don't let another task, with possibly unlocked vma,
2735                  * keep the mlocked page.
2736                  */
2737                 if ((ret & VM_FAULT_WRITE) && (vma->vm_flags & VM_LOCKED)) {
2738                         lock_page(old_page);    /* LRU manipulation */
2739                         munlock_vma_page(old_page);
2740                         unlock_page(old_page);
2741                 }
2742                 page_cache_release(old_page);
2743         }
2744         return ret;
2745 oom_free_new:
2746         page_cache_release(new_page);
2747 oom:
2748         if (old_page) {
2749                 if (page_mkwrite) {
2750                         unlock_page(old_page);
2751                         page_cache_release(old_page);
2752                 }
2753                 page_cache_release(old_page);
2754         }
2755         return VM_FAULT_OOM;
2756
2757 unwritable_page:
2758         page_cache_release(old_page);
2759         return ret;
2760 }
2761
2762 static void unmap_mapping_range_vma(struct vm_area_struct *vma,
2763                 unsigned long start_addr, unsigned long end_addr,
2764                 struct zap_details *details)
2765 {
2766         zap_page_range(vma, start_addr, end_addr - start_addr, details);
2767 }
2768
2769 static inline void unmap_mapping_range_tree(struct prio_tree_root *root,
2770                                             struct zap_details *details)
2771 {
2772         struct vm_area_struct *vma;
2773         struct prio_tree_iter iter;
2774         pgoff_t vba, vea, zba, zea;
2775
2776         vma_prio_tree_foreach(vma, &iter, root,
2777                         details->first_index, details->last_index) {
2778
2779                 vba = vma->vm_pgoff;
2780                 vea = vba + ((vma->vm_end - vma->vm_start) >> PAGE_SHIFT) - 1;
2781                 /* Assume for now that PAGE_CACHE_SHIFT == PAGE_SHIFT */
2782                 zba = details->first_index;
2783                 if (zba < vba)
2784                         zba = vba;
2785                 zea = details->last_index;
2786                 if (zea > vea)
2787                         zea = vea;
2788
2789                 unmap_mapping_range_vma(vma,
2790                         ((zba - vba) << PAGE_SHIFT) + vma->vm_start,
2791                         ((zea - vba + 1) << PAGE_SHIFT) + vma->vm_start,
2792                                 details);
2793         }
2794 }
2795
2796 static inline void unmap_mapping_range_list(struct list_head *head,
2797                                             struct zap_details *details)
2798 {
2799         struct vm_area_struct *vma;
2800
2801         /*
2802          * In nonlinear VMAs there is no correspondence between virtual address
2803          * offset and file offset.  So we must perform an exhaustive search
2804          * across *all* the pages in each nonlinear VMA, not just the pages
2805          * whose virtual address lies outside the file truncation point.
2806          */
2807         list_for_each_entry(vma, head, shared.vm_set.list) {
2808                 details->nonlinear_vma = vma;
2809                 unmap_mapping_range_vma(vma, vma->vm_start, vma->vm_end, details);
2810         }
2811 }
2812
2813 /**
2814  * unmap_mapping_range - unmap the portion of all mmaps in the specified address_space corresponding to the specified page range in the underlying file.
2815  * @mapping: the address space containing mmaps to be unmapped.
2816  * @holebegin: byte in first page to unmap, relative to the start of
2817  * the underlying file.  This will be rounded down to a PAGE_SIZE
2818  * boundary.  Note that this is different from truncate_pagecache(), which
2819  * must keep the partial page.  In contrast, we must get rid of
2820  * partial pages.
2821  * @holelen: size of prospective hole in bytes.  This will be rounded
2822  * up to a PAGE_SIZE boundary.  A holelen of zero truncates to the
2823  * end of the file.
2824  * @even_cows: 1 when truncating a file, unmap even private COWed pages;
2825  * but 0 when invalidating pagecache, don't throw away private data.
2826  */
2827 void unmap_mapping_range(struct address_space *mapping,
2828                 loff_t const holebegin, loff_t const holelen, int even_cows)
2829 {
2830         struct zap_details details;
2831         pgoff_t hba = holebegin >> PAGE_SHIFT;
2832         pgoff_t hlen = (holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
2833
2834         /* Check for overflow. */
2835         if (sizeof(holelen) > sizeof(hlen)) {
2836                 long long holeend =
2837                         (holebegin + holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
2838                 if (holeend & ~(long long)ULONG_MAX)
2839                         hlen = ULONG_MAX - hba + 1;
2840         }
2841
2842         details.check_mapping = even_cows? NULL: mapping;
2843         details.nonlinear_vma = NULL;
2844         details.first_index = hba;
2845         details.last_index = hba + hlen - 1;
2846         if (details.last_index < details.first_index)
2847                 details.last_index = ULONG_MAX;
2848
2849
2850         mutex_lock(&mapping->i_mmap_mutex);
2851         if (unlikely(!prio_tree_empty(&mapping->i_mmap)))
2852                 unmap_mapping_range_tree(&mapping->i_mmap, &details);
2853         if (unlikely(!list_empty(&mapping->i_mmap_nonlinear)))
2854                 unmap_mapping_range_list(&mapping->i_mmap_nonlinear, &details);
2855         mutex_unlock(&mapping->i_mmap_mutex);
2856 }
2857 EXPORT_SYMBOL(unmap_mapping_range);
2858
2859 /*
2860  * We enter with non-exclusive mmap_sem (to exclude vma changes,
2861  * but allow concurrent faults), and pte mapped but not yet locked.
2862  * We return with mmap_sem still held, but pte unmapped and unlocked.
2863  */
2864 static int do_swap_page(struct mm_struct *mm, struct vm_area_struct *vma,
2865                 unsigned long address, pte_t *page_table, pmd_t *pmd,
2866                 unsigned int flags, pte_t orig_pte)
2867 {
2868         spinlock_t *ptl;
2869         struct page *page, *swapcache = NULL;
2870         swp_entry_t entry;
2871         pte_t pte;
2872         int locked;
2873         struct mem_cgroup *ptr;
2874         int exclusive = 0;
2875         int ret = 0;
2876
2877         if (!pte_unmap_same(mm, pmd, page_table, orig_pte))
2878                 goto out;
2879
2880         entry = pte_to_swp_entry(orig_pte);
2881         if (unlikely(non_swap_entry(entry))) {
2882                 if (is_migration_entry(entry)) {
2883                         migration_entry_wait(mm, pmd, address);
2884                 } else if (is_hwpoison_entry(entry)) {
2885                         ret = VM_FAULT_HWPOISON;
2886                 } else {
2887                         print_bad_pte(vma, address, orig_pte, NULL);
2888                         ret = VM_FAULT_SIGBUS;
2889                 }
2890                 goto out;
2891         }
2892         delayacct_set_flag(DELAYACCT_PF_SWAPIN);
2893         page = lookup_swap_cache(entry);
2894         if (!page) {
2895                 grab_swap_token(mm); /* Contend for token _before_ read-in */
2896                 page = swapin_readahead(entry,
2897                                         GFP_HIGHUSER_MOVABLE, vma, address);
2898                 if (!page) {
2899                         /*
2900                          * Back out if somebody else faulted in this pte
2901                          * while we released the pte lock.
2902                          */
2903                         page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2904                         if (likely(pte_same(*page_table, orig_pte)))
2905                                 ret = VM_FAULT_OOM;
2906                         delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
2907                         goto unlock;
2908                 }
2909
2910                 /* Had to read the page from swap area: Major fault */
2911                 ret = VM_FAULT_MAJOR;
2912                 count_vm_event(PGMAJFAULT);
2913                 mem_cgroup_count_vm_event(mm, PGMAJFAULT);
2914         } else if (PageHWPoison(page)) {
2915                 /*
2916                  * hwpoisoned dirty swapcache pages are kept for killing
2917                  * owner processes (which may be unknown at hwpoison time)
2918                  */
2919                 ret = VM_FAULT_HWPOISON;
2920                 delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
2921                 goto out_release;
2922         }
2923
2924         locked = lock_page_or_retry(page, mm, flags);
2925         delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
2926         if (!locked) {
2927                 ret |= VM_FAULT_RETRY;
2928                 goto out_release;
2929         }
2930
2931         /*
2932          * Make sure try_to_free_swap or reuse_swap_page or swapoff did not
2933          * release the swapcache from under us.  The page pin, and pte_same
2934          * test below, are not enough to exclude that.  Even if it is still
2935          * swapcache, we need to check that the page's swap has not changed.
2936          */
2937         if (unlikely(!PageSwapCache(page) || page_private(page) != entry.val))
2938                 goto out_page;
2939
2940         if (ksm_might_need_to_copy(page, vma, address)) {
2941                 swapcache = page;
2942                 page = ksm_does_need_to_copy(page, vma, address);
2943
2944                 if (unlikely(!page)) {
2945                         ret = VM_FAULT_OOM;
2946                         page = swapcache;
2947                         swapcache = NULL;
2948                         goto out_page;
2949                 }
2950         }
2951
2952         if (mem_cgroup_try_charge_swapin(mm, page, GFP_KERNEL, &ptr)) {
2953                 ret = VM_FAULT_OOM;
2954                 goto out_page;
2955         }
2956
2957         /*
2958          * Back out if somebody else already faulted in this pte.
2959          */
2960         page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2961         if (unlikely(!pte_same(*page_table, orig_pte)))
2962                 goto out_nomap;
2963
2964         if (unlikely(!PageUptodate(page))) {
2965                 ret = VM_FAULT_SIGBUS;
2966                 goto out_nomap;
2967         }
2968
2969         /*
2970          * The page isn't present yet, go ahead with the fault.
2971          *
2972          * Be careful about the sequence of operations here.
2973          * To get its accounting right, reuse_swap_page() must be called
2974          * while the page is counted on swap but not yet in mapcount i.e.
2975          * before page_add_anon_rmap() and swap_free(); try_to_free_swap()
2976          * must be called after the swap_free(), or it will never succeed.
2977          * Because delete_from_swap_page() may be called by reuse_swap_page(),
2978          * mem_cgroup_commit_charge_swapin() may not be able to find swp_entry
2979          * in page->private. In this case, a record in swap_cgroup  is silently
2980          * discarded at swap_free().
2981          */
2982
2983         inc_mm_counter_fast(mm, MM_ANONPAGES);
2984         dec_mm_counter_fast(mm, MM_SWAPENTS);
2985         pte = mk_pte(page, vma->vm_page_prot);
2986         if ((flags & FAULT_FLAG_WRITE) && reuse_swap_page(page)) {
2987                 pte = maybe_mkwrite(pte_mkdirty(pte), vma);
2988                 flags &= ~FAULT_FLAG_WRITE;
2989                 ret |= VM_FAULT_WRITE;
2990                 exclusive = 1;
2991         }
2992         flush_icache_page(vma, page);
2993         set_pte_at(mm, address, page_table, pte);
2994         do_page_add_anon_rmap(page, vma, address, exclusive);
2995         /* It's better to call commit-charge after rmap is established */
2996         mem_cgroup_commit_charge_swapin(page, ptr);
2997
2998         swap_free(entry);
2999         if (vm_swap_full() || (vma->vm_flags & VM_LOCKED) || PageMlocked(page))
3000                 try_to_free_swap(page);
3001         unlock_page(page);
3002         if (swapcache) {
3003                 /*
3004                  * Hold the lock to avoid the swap entry to be reused
3005                  * until we take the PT lock for the pte_same() check
3006                  * (to avoid false positives from pte_same). For
3007                  * further safety release the lock after the swap_free
3008                  * so that the swap count won't change under a
3009                  * parallel locked swapcache.
3010                  */
3011                 unlock_page(swapcache);
3012                 page_cache_release(swapcache);
3013         }
3014
3015         if (flags & FAULT_FLAG_WRITE) {
3016                 ret |= do_wp_page(mm, vma, address, page_table, pmd, ptl, pte);
3017                 if (ret & VM_FAULT_ERROR)
3018                         ret &= VM_FAULT_ERROR;
3019                 goto out;
3020         }
3021
3022         /* No need to invalidate - it was non-present before */
3023         update_mmu_cache(vma, address, page_table);
3024 unlock:
3025         pte_unmap_unlock(page_table, ptl);
3026 out:
3027         return ret;
3028 out_nomap:
3029         mem_cgroup_cancel_charge_swapin(ptr);
3030         pte_unmap_unlock(page_table, ptl);
3031 out_page:
3032         unlock_page(page);
3033 out_release:
3034         page_cache_release(page);
3035         if (swapcache) {
3036                 unlock_page(swapcache);
3037                 page_cache_release(swapcache);
3038         }
3039         return ret;
3040 }
3041
3042 /*
3043  * This is like a special single-page "expand_{down|up}wards()",
3044  * except we must first make sure that 'address{-|+}PAGE_SIZE'
3045  * doesn't hit another vma.
3046  */
3047 static inline int check_stack_guard_page(struct vm_area_struct *vma, unsigned long address)
3048 {
3049         address &= PAGE_MASK;
3050         if ((vma->vm_flags & VM_GROWSDOWN) && address == vma->vm_start) {
3051                 struct vm_area_struct *prev = vma->vm_prev;
3052
3053                 /*
3054                  * Is there a mapping abutting this one below?
3055                  *
3056                  * That's only ok if it's the same stack mapping
3057                  * that has gotten split..
3058                  */
3059                 if (prev && prev->vm_end == address)
3060                         return prev->vm_flags & VM_GROWSDOWN ? 0 : -ENOMEM;
3061
3062                 expand_downwards(vma, address - PAGE_SIZE);
3063         }
3064         if ((vma->vm_flags & VM_GROWSUP) && address + PAGE_SIZE == vma->vm_end) {
3065                 struct vm_area_struct *next = vma->vm_next;
3066
3067                 /* As VM_GROWSDOWN but s/below/above/ */
3068                 if (next && next->vm_start == address + PAGE_SIZE)
3069                         return next->vm_flags & VM_GROWSUP ? 0 : -ENOMEM;
3070
3071                 expand_upwards(vma, address + PAGE_SIZE);
3072         }
3073         return 0;
3074 }
3075
3076 /*
3077  * We enter with non-exclusive mmap_sem (to exclude vma changes,
3078  * but allow concurrent faults), and pte mapped but not yet locked.
3079  * We return with mmap_sem still held, but pte unmapped and unlocked.
3080  */
3081 static int do_anonymous_page(struct mm_struct *mm, struct vm_area_struct *vma,
3082                 unsigned long address, pte_t *page_table, pmd_t *pmd,
3083                 unsigned int flags)
3084 {
3085         struct page *page;
3086         spinlock_t *ptl;
3087         pte_t entry;
3088
3089         pte_unmap(page_table);
3090
3091         /* Check if we need to add a guard page to the stack */
3092         if (check_stack_guard_page(vma, address) < 0)
3093                 return VM_FAULT_SIGBUS;
3094
3095         /* Use the zero-page for reads */
3096         if (!(flags & FAULT_FLAG_WRITE)) {
3097                 entry = pte_mkspecial(pfn_pte(my_zero_pfn(address),
3098                                                 vma->vm_page_prot));
3099                 page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
3100                 if (!pte_none(*page_table))
3101                         goto unlock;
3102                 goto setpte;
3103         }
3104
3105         /* Allocate our own private page. */
3106         if (unlikely(anon_vma_prepare(vma)))
3107                 goto oom;
3108         page = alloc_zeroed_user_highpage_movable(vma, address);
3109         if (!page)
3110                 goto oom;
3111         __SetPageUptodate(page);
3112
3113         if (mem_cgroup_newpage_charge(page, mm, GFP_KERNEL))
3114                 goto oom_free_page;
3115
3116         entry = mk_pte(page, vma->vm_page_prot);
3117         if (vma->vm_flags & VM_WRITE)
3118                 entry = pte_mkwrite(pte_mkdirty(entry));
3119
3120         page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
3121         if (!pte_none(*page_table))
3122                 goto release;
3123
3124         inc_mm_counter_fast(mm, MM_ANONPAGES);
3125         page_add_new_anon_rmap(page, vma, address);
3126 setpte:
3127         set_pte_at(mm, address, page_table, entry);
3128
3129         /* No need to invalidate - it was non-present before */
3130         update_mmu_cache(vma, address, page_table);
3131 unlock:
3132         pte_unmap_unlock(page_table, ptl);
3133         return 0;
3134 release:
3135         mem_cgroup_uncharge_page(page);
3136         page_cache_release(page);
3137         goto unlock;
3138 oom_free_page:
3139         page_cache_release(page);
3140 oom:
3141         return VM_FAULT_OOM;
3142 }
3143
3144 /*
3145  * __do_fault() tries to create a new page mapping. It aggressively
3146  * tries to share with existing pages, but makes a separate copy if
3147  * the FAULT_FLAG_WRITE is set in the flags parameter in order to avoid
3148  * the next page fault.
3149  *
3150  * As this is called only for pages that do not currently exist, we
3151  * do not need to flush old virtual caches or the TLB.
3152  *
3153  * We enter with non-exclusive mmap_sem (to exclude vma changes,
3154  * but allow concurrent faults), and pte neither mapped nor locked.
3155  * We return with mmap_sem still held, but pte unmapped and unlocked.
3156  */
3157 static int __do_fault(struct mm_struct *mm, struct vm_area_struct *vma,
3158                 unsigned long address, pmd_t *pmd,
3159                 pgoff_t pgoff, unsigned int flags, pte_t orig_pte)
3160 {
3161         pte_t *page_table;
3162         spinlock_t *ptl;
3163         struct page *page;
3164         struct page *cow_page;
3165         pte_t entry;
3166         int anon = 0;
3167         struct page *dirty_page = NULL;
3168         struct vm_fault vmf;
3169         int ret;
3170         int page_mkwrite = 0;
3171
3172         /*
3173          * If we do COW later, allocate page befor taking lock_page()
3174          * on the file cache page. This will reduce lock holding time.
3175          */
3176         if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) {
3177
3178                 if (unlikely(anon_vma_prepare(vma)))
3179                &nb