tizen 2.4 release
[profile/mobile/platform/kernel/linux-3.10-sc7730.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/delay.h>
53 #include <linux/init.h>
54 #include <linux/writeback.h>
55 #include <linux/memcontrol.h>
56 #include <linux/mmu_notifier.h>
57 #include <linux/kallsyms.h>
58 #include <linux/swapops.h>
59 #include <linux/elf.h>
60 #include <linux/gfp.h>
61 #include <linux/migrate.h>
62 #include <linux/string.h>
63
64 #include <asm/io.h>
65 #include <asm/pgalloc.h>
66 #include <asm/uaccess.h>
67 #include <asm/tlb.h>
68 #include <asm/tlbflush.h>
69 #include <asm/pgtable.h>
70
71 #include "internal.h"
72
73 #ifdef LAST_NID_NOT_IN_PAGE_FLAGS
74 #warning Unfortunate NUMA and NUMA Balancing config, growing page-frame for last_nid.
75 #endif
76
77 #ifndef CONFIG_NEED_MULTIPLE_NODES
78 /* use the per-pgdat data instead for discontigmem - mbligh */
79 unsigned long max_mapnr;
80 struct page *mem_map;
81
82 EXPORT_SYMBOL(max_mapnr);
83 EXPORT_SYMBOL(mem_map);
84 #endif
85
86 unsigned long num_physpages;
87 /*
88  * A number of key systems in x86 including ioremap() rely on the assumption
89  * that high_memory defines the upper bound on direct map memory, then end
90  * of ZONE_NORMAL.  Under CONFIG_DISCONTIG this means that max_low_pfn and
91  * highstart_pfn must be the same; there must be no gap between ZONE_NORMAL
92  * and ZONE_HIGHMEM.
93  */
94 void * high_memory;
95
96 EXPORT_SYMBOL(num_physpages);
97 EXPORT_SYMBOL(high_memory);
98
99 /*
100  * Randomize the address space (stacks, mmaps, brk, etc.).
101  *
102  * ( When CONFIG_COMPAT_BRK=y we exclude brk from randomization,
103  *   as ancient (libc5 based) binaries can segfault. )
104  */
105 int randomize_va_space __read_mostly =
106 #ifdef CONFIG_COMPAT_BRK
107                                         1;
108 #else
109                                         2;
110 #endif
111
112 static int __init disable_randmaps(char *s)
113 {
114         randomize_va_space = 0;
115         return 1;
116 }
117 __setup("norandmaps", disable_randmaps);
118
119 unsigned long zero_pfn __read_mostly;
120 unsigned long highest_memmap_pfn __read_mostly;
121
122 EXPORT_SYMBOL(zero_pfn);
123
124 /*
125  * CONFIG_MMU architectures set up ZERO_PAGE in their paging_init()
126  */
127 static int __init init_zero_pfn(void)
128 {
129         zero_pfn = page_to_pfn(ZERO_PAGE(0));
130         return 0;
131 }
132 core_initcall(init_zero_pfn);
133
134
135 #if defined(SPLIT_RSS_COUNTING)
136
137 void sync_mm_rss(struct mm_struct *mm)
138 {
139         int i;
140
141         for (i = 0; i < NR_MM_COUNTERS; i++) {
142                 if (current->rss_stat.count[i]) {
143                         add_mm_counter(mm, i, current->rss_stat.count[i]);
144                         current->rss_stat.count[i] = 0;
145                 }
146         }
147         current->rss_stat.events = 0;
148 }
149
150 static void add_mm_counter_fast(struct mm_struct *mm, int member, int val)
151 {
152         struct task_struct *task = current;
153
154         if (likely(task->mm == mm))
155                 task->rss_stat.count[member] += val;
156         else
157                 add_mm_counter(mm, member, val);
158 }
159 #define inc_mm_counter_fast(mm, member) add_mm_counter_fast(mm, member, 1)
160 #define dec_mm_counter_fast(mm, member) add_mm_counter_fast(mm, member, -1)
161
162 /* sync counter once per 64 page faults */
163 #define TASK_RSS_EVENTS_THRESH  (64)
164 static void check_sync_rss_stat(struct task_struct *task)
165 {
166         if (unlikely(task != current))
167                 return;
168         if (unlikely(task->rss_stat.events++ > TASK_RSS_EVENTS_THRESH))
169                 sync_mm_rss(task->mm);
170 }
171 #else /* SPLIT_RSS_COUNTING */
172
173 #define inc_mm_counter_fast(mm, member) inc_mm_counter(mm, member)
174 #define dec_mm_counter_fast(mm, member) dec_mm_counter(mm, member)
175
176 static void check_sync_rss_stat(struct task_struct *task)
177 {
178 }
179
180 #endif /* SPLIT_RSS_COUNTING */
181
182 #ifdef HAVE_GENERIC_MMU_GATHER
183
184 static int tlb_next_batch(struct mmu_gather *tlb)
185 {
186         struct mmu_gather_batch *batch;
187
188         batch = tlb->active;
189         if (batch->next) {
190                 tlb->active = batch->next;
191                 return 1;
192         }
193
194         if (tlb->batch_count == MAX_GATHER_BATCH_COUNT)
195                 return 0;
196
197 #ifndef CONFIG_SPRD_PAGERECORDER
198         batch = (void *)__get_free_pages(GFP_NOWAIT | __GFP_NOWARN, 0);
199 #else
200         batch = (void *)__get_free_pages_nopagedebug(GFP_NOWAIT | __GFP_NOWARN, 0);
201 #endif
202         if (!batch)
203                 return 0;
204
205         tlb->batch_count++;
206         batch->next = NULL;
207         batch->nr   = 0;
208         batch->max  = MAX_GATHER_BATCH;
209
210         tlb->active->next = batch;
211         tlb->active = batch;
212
213         return 1;
214 }
215
216 /* tlb_gather_mmu
217  *      Called to initialize an (on-stack) mmu_gather structure for page-table
218  *      tear-down from @mm. The @fullmm argument is used when @mm is without
219  *      users and we're going to destroy the full address space (exit/execve).
220  */
221 void tlb_gather_mmu(struct mmu_gather *tlb, struct mm_struct *mm, unsigned long start, unsigned long end)
222 {
223         tlb->mm = mm;
224
225         /* Is it from 0 to ~0? */
226         tlb->fullmm     = !(start | (end+1));
227         tlb->need_flush_all = 0;
228         tlb->start      = start;
229         tlb->end        = end;
230         tlb->need_flush = 0;
231         tlb->local.next = NULL;
232         tlb->local.nr   = 0;
233         tlb->local.max  = ARRAY_SIZE(tlb->__pages);
234         tlb->active     = &tlb->local;
235         tlb->batch_count = 0;
236
237 #ifdef CONFIG_HAVE_RCU_TABLE_FREE
238         tlb->batch = NULL;
239 #endif
240 }
241
242 void tlb_flush_mmu(struct mmu_gather *tlb)
243 {
244         struct mmu_gather_batch *batch;
245
246         if (!tlb->need_flush)
247                 return;
248         tlb->need_flush = 0;
249         tlb_flush(tlb);
250 #ifdef CONFIG_HAVE_RCU_TABLE_FREE
251         tlb_table_flush(tlb);
252 #endif
253
254         for (batch = &tlb->local; batch; batch = batch->next) {
255                 free_pages_and_swap_cache(batch->pages, batch->nr);
256                 batch->nr = 0;
257         }
258         tlb->active = &tlb->local;
259 }
260
261 /* tlb_finish_mmu
262  *      Called at the end of the shootdown operation to free up any resources
263  *      that were required.
264  */
265 void tlb_finish_mmu(struct mmu_gather *tlb, unsigned long start, unsigned long end)
266 {
267         struct mmu_gather_batch *batch, *next;
268
269         tlb_flush_mmu(tlb);
270
271         /* keep the page table cache within bounds */
272         check_pgt_cache();
273
274         for (batch = tlb->local.next; batch; batch = next) {
275                 next = batch->next;
276                 free_pages((unsigned long)batch, 0);
277         }
278         tlb->local.next = NULL;
279 }
280
281 /* __tlb_remove_page
282  *      Must perform the equivalent to __free_pte(pte_get_and_clear(ptep)), while
283  *      handling the additional races in SMP caused by other CPUs caching valid
284  *      mappings in their TLBs. Returns the number of free page slots left.
285  *      When out of page slots we must call tlb_flush_mmu().
286  */
287 int __tlb_remove_page(struct mmu_gather *tlb, struct page *page)
288 {
289         struct mmu_gather_batch *batch;
290
291         VM_BUG_ON(!tlb->need_flush);
292
293         batch = tlb->active;
294         batch->pages[batch->nr++] = page;
295         if (batch->nr == batch->max) {
296                 if (!tlb_next_batch(tlb))
297                         return 0;
298                 batch = tlb->active;
299         }
300         VM_BUG_ON(batch->nr > batch->max);
301
302         return batch->max - batch->nr;
303 }
304
305 #endif /* HAVE_GENERIC_MMU_GATHER */
306
307 #ifdef CONFIG_HAVE_RCU_TABLE_FREE
308
309 /*
310  * See the comment near struct mmu_table_batch.
311  */
312
313 static void tlb_remove_table_smp_sync(void *arg)
314 {
315         /* Simply deliver the interrupt */
316 }
317
318 static void tlb_remove_table_one(void *table)
319 {
320         /*
321          * This isn't an RCU grace period and hence the page-tables cannot be
322          * assumed to be actually RCU-freed.
323          *
324          * It is however sufficient for software page-table walkers that rely on
325          * IRQ disabling. See the comment near struct mmu_table_batch.
326          */
327         smp_call_function(tlb_remove_table_smp_sync, NULL, 1);
328         __tlb_remove_table(table);
329 }
330
331 static void tlb_remove_table_rcu(struct rcu_head *head)
332 {
333         struct mmu_table_batch *batch;
334         int i;
335
336         batch = container_of(head, struct mmu_table_batch, rcu);
337
338         for (i = 0; i < batch->nr; i++)
339                 __tlb_remove_table(batch->tables[i]);
340
341         free_page((unsigned long)batch);
342 }
343
344 void tlb_table_flush(struct mmu_gather *tlb)
345 {
346         struct mmu_table_batch **batch = &tlb->batch;
347
348         if (*batch) {
349                 call_rcu_sched(&(*batch)->rcu, tlb_remove_table_rcu);
350                 *batch = NULL;
351         }
352 }
353
354 void tlb_remove_table(struct mmu_gather *tlb, void *table)
355 {
356         struct mmu_table_batch **batch = &tlb->batch;
357
358         tlb->need_flush = 1;
359
360         /*
361          * When there's less then two users of this mm there cannot be a
362          * concurrent page-table walk.
363          */
364         if (atomic_read(&tlb->mm->mm_users) < 2) {
365                 __tlb_remove_table(table);
366                 return;
367         }
368
369         if (*batch == NULL) {
370                 *batch = (struct mmu_table_batch *)__get_free_page(GFP_NOWAIT | __GFP_NOWARN);
371                 if (*batch == NULL) {
372                         tlb_remove_table_one(table);
373                         return;
374                 }
375                 (*batch)->nr = 0;
376         }
377         (*batch)->tables[(*batch)->nr++] = table;
378         if ((*batch)->nr == MAX_TABLE_BATCH)
379                 tlb_table_flush(tlb);
380 }
381
382 #endif /* CONFIG_HAVE_RCU_TABLE_FREE */
383
384 /*
385  * If a p?d_bad entry is found while walking page tables, report
386  * the error, before resetting entry to p?d_none.  Usually (but
387  * very seldom) called out from the p?d_none_or_clear_bad macros.
388  */
389
390 void pgd_clear_bad(pgd_t *pgd)
391 {
392         pgd_ERROR(*pgd);
393         pgd_clear(pgd);
394 }
395
396 void pud_clear_bad(pud_t *pud)
397 {
398         pud_ERROR(*pud);
399         pud_clear(pud);
400 }
401
402 void pmd_clear_bad(pmd_t *pmd)
403 {
404         pmd_ERROR(*pmd);
405         pmd_clear(pmd);
406 }
407
408 /*
409  * Note: this doesn't free the actual pages themselves. That
410  * has been handled earlier when unmapping all the memory regions.
411  */
412 static void free_pte_range(struct mmu_gather *tlb, pmd_t *pmd,
413                            unsigned long addr)
414 {
415         pgtable_t token = pmd_pgtable(*pmd);
416         pmd_clear(pmd);
417         pte_free_tlb(tlb, token, addr);
418         tlb->mm->nr_ptes--;
419 }
420
421 static inline void free_pmd_range(struct mmu_gather *tlb, pud_t *pud,
422                                 unsigned long addr, unsigned long end,
423                                 unsigned long floor, unsigned long ceiling)
424 {
425         pmd_t *pmd;
426         unsigned long next;
427         unsigned long start;
428
429         start = addr;
430         pmd = pmd_offset(pud, addr);
431         do {
432                 next = pmd_addr_end(addr, end);
433                 if (pmd_none_or_clear_bad(pmd))
434                         continue;
435                 free_pte_range(tlb, pmd, addr);
436         } while (pmd++, addr = next, addr != end);
437
438         start &= PUD_MASK;
439         if (start < floor)
440                 return;
441         if (ceiling) {
442                 ceiling &= PUD_MASK;
443                 if (!ceiling)
444                         return;
445         }
446         if (end - 1 > ceiling - 1)
447                 return;
448
449         pmd = pmd_offset(pud, start);
450         pud_clear(pud);
451         pmd_free_tlb(tlb, pmd, start);
452 }
453
454 static inline void free_pud_range(struct mmu_gather *tlb, pgd_t *pgd,
455                                 unsigned long addr, unsigned long end,
456                                 unsigned long floor, unsigned long ceiling)
457 {
458         pud_t *pud;
459         unsigned long next;
460         unsigned long start;
461
462         start = addr;
463         pud = pud_offset(pgd, addr);
464         do {
465                 next = pud_addr_end(addr, end);
466                 if (pud_none_or_clear_bad(pud))
467                         continue;
468                 free_pmd_range(tlb, pud, addr, next, floor, ceiling);
469         } while (pud++, addr = next, addr != end);
470
471         start &= PGDIR_MASK;
472         if (start < floor)
473                 return;
474         if (ceiling) {
475                 ceiling &= PGDIR_MASK;
476                 if (!ceiling)
477                         return;
478         }
479         if (end - 1 > ceiling - 1)
480                 return;
481
482         pud = pud_offset(pgd, start);
483         pgd_clear(pgd);
484         pud_free_tlb(tlb, pud, start);
485 }
486
487 /*
488  * This function frees user-level page tables of a process.
489  *
490  * Must be called with pagetable lock held.
491  */
492 void free_pgd_range(struct mmu_gather *tlb,
493                         unsigned long addr, unsigned long end,
494                         unsigned long floor, unsigned long ceiling)
495 {
496         pgd_t *pgd;
497         unsigned long next;
498
499         /*
500          * The next few lines have given us lots of grief...
501          *
502          * Why are we testing PMD* at this top level?  Because often
503          * there will be no work to do at all, and we'd prefer not to
504          * go all the way down to the bottom just to discover that.
505          *
506          * Why all these "- 1"s?  Because 0 represents both the bottom
507          * of the address space and the top of it (using -1 for the
508          * top wouldn't help much: the masks would do the wrong thing).
509          * The rule is that addr 0 and floor 0 refer to the bottom of
510          * the address space, but end 0 and ceiling 0 refer to the top
511          * Comparisons need to use "end - 1" and "ceiling - 1" (though
512          * that end 0 case should be mythical).
513          *
514          * Wherever addr is brought up or ceiling brought down, we must
515          * be careful to reject "the opposite 0" before it confuses the
516          * subsequent tests.  But what about where end is brought down
517          * by PMD_SIZE below? no, end can't go down to 0 there.
518          *
519          * Whereas we round start (addr) and ceiling down, by different
520          * masks at different levels, in order to test whether a table
521          * now has no other vmas using it, so can be freed, we don't
522          * bother to round floor or end up - the tests don't need that.
523          */
524
525         addr &= PMD_MASK;
526         if (addr < floor) {
527                 addr += PMD_SIZE;
528                 if (!addr)
529                         return;
530         }
531         if (ceiling) {
532                 ceiling &= PMD_MASK;
533                 if (!ceiling)
534                         return;
535         }
536         if (end - 1 > ceiling - 1)
537                 end -= PMD_SIZE;
538         if (addr > end - 1)
539                 return;
540
541         pgd = pgd_offset(tlb->mm, addr);
542         do {
543                 next = pgd_addr_end(addr, end);
544                 if (pgd_none_or_clear_bad(pgd))
545                         continue;
546                 free_pud_range(tlb, pgd, addr, next, floor, ceiling);
547         } while (pgd++, addr = next, addr != end);
548 }
549
550 void free_pgtables(struct mmu_gather *tlb, struct vm_area_struct *vma,
551                 unsigned long floor, unsigned long ceiling)
552 {
553         while (vma) {
554                 struct vm_area_struct *next = vma->vm_next;
555                 unsigned long addr = vma->vm_start;
556
557                 /*
558                  * Hide vma from rmap and truncate_pagecache before freeing
559                  * pgtables
560                  */
561                 unlink_anon_vmas(vma);
562                 unlink_file_vma(vma);
563
564                 if (is_vm_hugetlb_page(vma)) {
565                         hugetlb_free_pgd_range(tlb, addr, vma->vm_end,
566                                 floor, next? next->vm_start: ceiling);
567                 } else {
568                         /*
569                          * Optimization: gather nearby vmas into one call down
570                          */
571                         while (next && next->vm_start <= vma->vm_end + PMD_SIZE
572                                && !is_vm_hugetlb_page(next)) {
573                                 vma = next;
574                                 next = vma->vm_next;
575                                 unlink_anon_vmas(vma);
576                                 unlink_file_vma(vma);
577                         }
578                         free_pgd_range(tlb, addr, vma->vm_end,
579                                 floor, next? next->vm_start: ceiling);
580                 }
581                 vma = next;
582         }
583 }
584
585 int __pte_alloc(struct mm_struct *mm, struct vm_area_struct *vma,
586                 pmd_t *pmd, unsigned long address)
587 {
588         pgtable_t new = pte_alloc_one(mm, address);
589         int wait_split_huge_page;
590         if (!new)
591                 return -ENOMEM;
592
593         /*
594          * Ensure all pte setup (eg. pte page lock and page clearing) are
595          * visible before the pte is made visible to other CPUs by being
596          * put into page tables.
597          *
598          * The other side of the story is the pointer chasing in the page
599          * table walking code (when walking the page table without locking;
600          * ie. most of the time). Fortunately, these data accesses consist
601          * of a chain of data-dependent loads, meaning most CPUs (alpha
602          * being the notable exception) will already guarantee loads are
603          * seen in-order. See the alpha page table accessors for the
604          * smp_read_barrier_depends() barriers in page table walking code.
605          */
606         smp_wmb(); /* Could be smp_wmb__xxx(before|after)_spin_lock */
607
608         spin_lock(&mm->page_table_lock);
609         wait_split_huge_page = 0;
610         if (likely(pmd_none(*pmd))) {   /* Has another populated it ? */
611                 mm->nr_ptes++;
612                 pmd_populate(mm, pmd, new);
613                 new = NULL;
614         } else if (unlikely(pmd_trans_splitting(*pmd)))
615                 wait_split_huge_page = 1;
616         spin_unlock(&mm->page_table_lock);
617         if (new)
618                 pte_free(mm, new);
619         if (wait_split_huge_page)
620                 wait_split_huge_page(vma->anon_vma, pmd);
621         return 0;
622 }
623
624 int __pte_alloc_kernel(pmd_t *pmd, unsigned long address)
625 {
626         pte_t *new = pte_alloc_one_kernel(&init_mm, address);
627         if (!new)
628                 return -ENOMEM;
629
630         smp_wmb(); /* See comment in __pte_alloc */
631
632         spin_lock(&init_mm.page_table_lock);
633         if (likely(pmd_none(*pmd))) {   /* Has another populated it ? */
634                 pmd_populate_kernel(&init_mm, pmd, new);
635                 new = NULL;
636         } else
637                 VM_BUG_ON(pmd_trans_splitting(*pmd));
638         spin_unlock(&init_mm.page_table_lock);
639         if (new)
640                 pte_free_kernel(&init_mm, new);
641         return 0;
642 }
643
644 static inline void init_rss_vec(int *rss)
645 {
646         memset(rss, 0, sizeof(int) * NR_MM_COUNTERS);
647 }
648
649 static inline void add_mm_rss_vec(struct mm_struct *mm, int *rss)
650 {
651         int i;
652
653         if (current->mm == mm)
654                 sync_mm_rss(mm);
655         for (i = 0; i < NR_MM_COUNTERS; i++)
656                 if (rss[i])
657                         add_mm_counter(mm, i, rss[i]);
658 }
659
660 /*
661  * This function is called to print an error when a bad pte
662  * is found. For example, we might have a PFN-mapped pte in
663  * a region that doesn't allow it.
664  *
665  * The calling function must still handle the error.
666  */
667 static void print_bad_pte(struct vm_area_struct *vma, unsigned long addr,
668                           pte_t pte, struct page *page)
669 {
670         pgd_t *pgd = pgd_offset(vma->vm_mm, addr);
671         pud_t *pud = pud_offset(pgd, addr);
672         pmd_t *pmd = pmd_offset(pud, addr);
673         struct address_space *mapping;
674         pgoff_t index;
675         static unsigned long resume;
676         static unsigned long nr_shown;
677         static unsigned long nr_unshown;
678
679         /*
680          * Allow a burst of 60 reports, then keep quiet for that minute;
681          * or allow a steady drip of one report per second.
682          */
683         if (nr_shown == 60) {
684                 if (time_before(jiffies, resume)) {
685                         nr_unshown++;
686                         return;
687                 }
688                 if (nr_unshown) {
689                         printk(KERN_ALERT
690                                 "BUG: Bad page map: %lu messages suppressed\n",
691                                 nr_unshown);
692                         nr_unshown = 0;
693                 }
694                 nr_shown = 0;
695         }
696         if (nr_shown++ == 0)
697                 resume = jiffies + 60 * HZ;
698
699         mapping = vma->vm_file ? vma->vm_file->f_mapping : NULL;
700         index = linear_page_index(vma, addr);
701
702         printk(KERN_ALERT
703                 "BUG: Bad page map in process %s  pte:%08llx pmd:%08llx\n",
704                 current->comm,
705                 (long long)pte_val(pte), (long long)pmd_val(*pmd));
706         if (page)
707                 dump_page(page);
708         printk(KERN_ALERT
709                 "addr:%p vm_flags:%08lx anon_vma:%p mapping:%p index:%lx\n",
710                 (void *)addr, vma->vm_flags, vma->anon_vma, mapping, index);
711         /*
712          * Choose text because data symbols depend on CONFIG_KALLSYMS_ALL=y
713          */
714         if (vma->vm_ops)
715                 printk(KERN_ALERT "vma->vm_ops->fault: %pSR\n",
716                        vma->vm_ops->fault);
717         if (vma->vm_file && vma->vm_file->f_op)
718                 printk(KERN_ALERT "vma->vm_file->f_op->mmap: %pSR\n",
719                        vma->vm_file->f_op->mmap);
720         dump_stack();
721         add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
722 }
723
724 static inline bool is_cow_mapping(vm_flags_t flags)
725 {
726         return (flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE;
727 }
728
729 /*
730  * vm_normal_page -- This function gets the "struct page" associated with a pte.
731  *
732  * "Special" mappings do not wish to be associated with a "struct page" (either
733  * it doesn't exist, or it exists but they don't want to touch it). In this
734  * case, NULL is returned here. "Normal" mappings do have a struct page.
735  *
736  * There are 2 broad cases. Firstly, an architecture may define a pte_special()
737  * pte bit, in which case this function is trivial. Secondly, an architecture
738  * may not have a spare pte bit, which requires a more complicated scheme,
739  * described below.
740  *
741  * A raw VM_PFNMAP mapping (ie. one that is not COWed) is always considered a
742  * special mapping (even if there are underlying and valid "struct pages").
743  * COWed pages of a VM_PFNMAP are always normal.
744  *
745  * The way we recognize COWed pages within VM_PFNMAP mappings is through the
746  * rules set up by "remap_pfn_range()": the vma will have the VM_PFNMAP bit
747  * set, and the vm_pgoff will point to the first PFN mapped: thus every special
748  * mapping will always honor the rule
749  *
750  *      pfn_of_page == vma->vm_pgoff + ((addr - vma->vm_start) >> PAGE_SHIFT)
751  *
752  * And for normal mappings this is false.
753  *
754  * This restricts such mappings to be a linear translation from virtual address
755  * to pfn. To get around this restriction, we allow arbitrary mappings so long
756  * as the vma is not a COW mapping; in that case, we know that all ptes are
757  * special (because none can have been COWed).
758  *
759  *
760  * In order to support COW of arbitrary special mappings, we have VM_MIXEDMAP.
761  *
762  * VM_MIXEDMAP mappings can likewise contain memory with or without "struct
763  * page" backing, however the difference is that _all_ pages with a struct
764  * page (that is, those where pfn_valid is true) are refcounted and considered
765  * normal pages by the VM. The disadvantage is that pages are refcounted
766  * (which can be slower and simply not an option for some PFNMAP users). The
767  * advantage is that we don't have to follow the strict linearity rule of
768  * PFNMAP mappings in order to support COWable mappings.
769  *
770  */
771 #ifdef __HAVE_ARCH_PTE_SPECIAL
772 # define HAVE_PTE_SPECIAL 1
773 #else
774 # define HAVE_PTE_SPECIAL 0
775 #endif
776 struct page *vm_normal_page(struct vm_area_struct *vma, unsigned long addr,
777                                 pte_t pte)
778 {
779         unsigned long pfn = pte_pfn(pte);
780
781         if (HAVE_PTE_SPECIAL) {
782                 if (likely(!pte_special(pte)))
783                         goto check_pfn;
784                 if (vma->vm_flags & (VM_PFNMAP | VM_MIXEDMAP))
785                         return NULL;
786                 if (!is_zero_pfn(pfn))
787                         print_bad_pte(vma, addr, pte, NULL);
788                 return NULL;
789         }
790
791         /* !HAVE_PTE_SPECIAL case follows: */
792
793         if (unlikely(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP))) {
794                 if (vma->vm_flags & VM_MIXEDMAP) {
795                         if (!pfn_valid(pfn))
796                                 return NULL;
797                         goto out;
798                 } else {
799                         unsigned long off;
800                         off = (addr - vma->vm_start) >> PAGE_SHIFT;
801                         if (pfn == vma->vm_pgoff + off)
802                                 return NULL;
803                         if (!is_cow_mapping(vma->vm_flags))
804                                 return NULL;
805                 }
806         }
807
808         if (is_zero_pfn(pfn))
809                 return NULL;
810 check_pfn:
811         if (unlikely(pfn > highest_memmap_pfn)) {
812                 print_bad_pte(vma, addr, pte, NULL);
813                 return NULL;
814         }
815
816         /*
817          * NOTE! We still have PageReserved() pages in the page tables.
818          * eg. VDSO mappings can cause them to exist.
819          */
820 out:
821         return pfn_to_page(pfn);
822 }
823
824 /*
825  * copy one vm_area from one task to the other. Assumes the page tables
826  * already present in the new task to be cleared in the whole range
827  * covered by this vma.
828  */
829
830 static inline unsigned long
831 copy_one_pte(struct mm_struct *dst_mm, struct mm_struct *src_mm,
832                 pte_t *dst_pte, pte_t *src_pte, struct vm_area_struct *vma,
833                 unsigned long addr, int *rss)
834 {
835         unsigned long vm_flags = vma->vm_flags;
836         pte_t pte = *src_pte;
837         struct page *page;
838
839         /* pte contains position in swap or file, so copy. */
840         if (unlikely(!pte_present(pte))) {
841                 if (!pte_file(pte)) {
842                         swp_entry_t entry = pte_to_swp_entry(pte);
843
844                         if (likely(!non_swap_entry(entry))) {
845                                 if (swap_duplicate(entry) < 0)
846                                         return entry.val;
847
848                                 /* make sure dst_mm is on swapoff's mmlist. */
849                                 if (unlikely(list_empty(&dst_mm->mmlist))) {
850                                         spin_lock(&mmlist_lock);
851                                         if (list_empty(&dst_mm->mmlist))
852                                                 list_add(&dst_mm->mmlist,
853                                                          &src_mm->mmlist);
854                                         spin_unlock(&mmlist_lock);
855                                 }
856                                 rss[MM_SWAPENTS]++;
857                         } else if (is_migration_entry(entry)) {
858                                 page = migration_entry_to_page(entry);
859
860                                 if (PageAnon(page))
861                                         rss[MM_ANONPAGES]++;
862                                 else
863                                         rss[MM_FILEPAGES]++;
864
865                                 if (is_write_migration_entry(entry) &&
866                                     is_cow_mapping(vm_flags)) {
867                                         /*
868                                          * COW mappings require pages in both
869                                          * parent and child to be set to read.
870                                          */
871                                         make_migration_entry_read(&entry);
872                                         pte = swp_entry_to_pte(entry);
873                                         set_pte_at(src_mm, addr, src_pte, pte);
874                                 }
875                         }
876                 }
877                 goto out_set_pte;
878         }
879
880         /*
881          * If it's a COW mapping, write protect it both
882          * in the parent and the child
883          */
884         if (is_cow_mapping(vm_flags)) {
885                 ptep_set_wrprotect(src_mm, addr, src_pte);
886                 pte = pte_wrprotect(pte);
887         }
888
889         /*
890          * If it's a shared mapping, mark it clean in
891          * the child
892          */
893         if (vm_flags & VM_SHARED)
894                 pte = pte_mkclean(pte);
895         pte = pte_mkold(pte);
896
897         page = vm_normal_page(vma, addr, pte);
898         if (page) {
899                 get_page(page);
900                 page_dup_rmap(page);
901                 if (PageAnon(page))
902                         rss[MM_ANONPAGES]++;
903                 else
904                         rss[MM_FILEPAGES]++;
905         }
906
907 out_set_pte:
908         set_pte_at(dst_mm, addr, dst_pte, pte);
909         return 0;
910 }
911
912 int copy_pte_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
913                    pmd_t *dst_pmd, pmd_t *src_pmd, struct vm_area_struct *vma,
914                    unsigned long addr, unsigned long end)
915 {
916         pte_t *orig_src_pte, *orig_dst_pte;
917         pte_t *src_pte, *dst_pte;
918         spinlock_t *src_ptl, *dst_ptl;
919         int progress = 0;
920         int rss[NR_MM_COUNTERS];
921         swp_entry_t entry = (swp_entry_t){0};
922
923 again:
924         init_rss_vec(rss);
925
926         dst_pte = pte_alloc_map_lock(dst_mm, dst_pmd, addr, &dst_ptl);
927         if (!dst_pte)
928                 return -ENOMEM;
929         src_pte = pte_offset_map(src_pmd, addr);
930         src_ptl = pte_lockptr(src_mm, src_pmd);
931         spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
932         orig_src_pte = src_pte;
933         orig_dst_pte = dst_pte;
934         arch_enter_lazy_mmu_mode();
935
936         do {
937                 /*
938                  * We are holding two locks at this point - either of them
939                  * could generate latencies in another task on another CPU.
940                  */
941                 if (progress >= 32) {
942                         progress = 0;
943                         if (need_resched() ||
944                             spin_needbreak(src_ptl) || spin_needbreak(dst_ptl))
945                                 break;
946                 }
947                 if (pte_none(*src_pte)) {
948                         progress++;
949                         continue;
950                 }
951                 entry.val = copy_one_pte(dst_mm, src_mm, dst_pte, src_pte,
952                                                         vma, addr, rss);
953                 if (entry.val)
954                         break;
955                 progress += 8;
956         } while (dst_pte++, src_pte++, addr += PAGE_SIZE, addr != end);
957
958         arch_leave_lazy_mmu_mode();
959         spin_unlock(src_ptl);
960         pte_unmap(orig_src_pte);
961         add_mm_rss_vec(dst_mm, rss);
962         pte_unmap_unlock(orig_dst_pte, dst_ptl);
963         cond_resched();
964
965         if (entry.val) {
966                 if (add_swap_count_continuation(entry, GFP_KERNEL) < 0)
967                         return -ENOMEM;
968                 progress = 0;
969         }
970         if (addr != end)
971                 goto again;
972         return 0;
973 }
974
975 static inline int copy_pmd_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
976                 pud_t *dst_pud, pud_t *src_pud, struct vm_area_struct *vma,
977                 unsigned long addr, unsigned long end)
978 {
979         pmd_t *src_pmd, *dst_pmd;
980         unsigned long next;
981
982         dst_pmd = pmd_alloc(dst_mm, dst_pud, addr);
983         if (!dst_pmd)
984                 return -ENOMEM;
985         src_pmd = pmd_offset(src_pud, addr);
986         do {
987                 next = pmd_addr_end(addr, end);
988                 if (pmd_trans_huge(*src_pmd)) {
989                         int err;
990                         VM_BUG_ON(next-addr != HPAGE_PMD_SIZE);
991                         err = copy_huge_pmd(dst_mm, src_mm,
992                                             dst_pmd, src_pmd, addr, vma);
993                         if (err == -ENOMEM)
994                                 return -ENOMEM;
995                         if (!err)
996                                 continue;
997                         /* fall through */
998                 }
999                 if (pmd_none_or_clear_bad(src_pmd))
1000                         continue;
1001                 if (copy_pte_range(dst_mm, src_mm, dst_pmd, src_pmd,
1002                                                 vma, addr, next))
1003                         return -ENOMEM;
1004         } while (dst_pmd++, src_pmd++, addr = next, addr != end);
1005         return 0;
1006 }
1007
1008 static inline int copy_pud_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
1009                 pgd_t *dst_pgd, pgd_t *src_pgd, struct vm_area_struct *vma,
1010                 unsigned long addr, unsigned long end)
1011 {
1012         pud_t *src_pud, *dst_pud;
1013         unsigned long next;
1014
1015         dst_pud = pud_alloc(dst_mm, dst_pgd, addr);
1016         if (!dst_pud)
1017                 return -ENOMEM;
1018         src_pud = pud_offset(src_pgd, addr);
1019         do {
1020                 next = pud_addr_end(addr, end);
1021                 if (pud_none_or_clear_bad(src_pud))
1022                         continue;
1023                 if (copy_pmd_range(dst_mm, src_mm, dst_pud, src_pud,
1024                                                 vma, addr, next))
1025                         return -ENOMEM;
1026         } while (dst_pud++, src_pud++, addr = next, addr != end);
1027         return 0;
1028 }
1029
1030 int copy_page_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
1031                 struct vm_area_struct *vma)
1032 {
1033         pgd_t *src_pgd, *dst_pgd;
1034         unsigned long next;
1035         unsigned long addr = vma->vm_start;
1036         unsigned long end = vma->vm_end;
1037         unsigned long mmun_start;       /* For mmu_notifiers */
1038         unsigned long mmun_end;         /* For mmu_notifiers */
1039         bool is_cow;
1040         int ret;
1041
1042         /*
1043          * Don't copy ptes where a page fault will fill them correctly.
1044          * Fork becomes much lighter when there are big shared or private
1045          * readonly mappings. The tradeoff is that copy_page_range is more
1046          * efficient than faulting.
1047          */
1048         if (!(vma->vm_flags & (VM_HUGETLB | VM_NONLINEAR |
1049                                VM_PFNMAP | VM_MIXEDMAP))) {
1050                 if (!vma->anon_vma)
1051                         return 0;
1052         }
1053
1054         if (is_vm_hugetlb_page(vma))
1055                 return copy_hugetlb_page_range(dst_mm, src_mm, vma);
1056
1057         if (unlikely(vma->vm_flags & VM_PFNMAP)) {
1058                 /*
1059                  * We do not free on error cases below as remove_vma
1060                  * gets called on error from higher level routine
1061                  */
1062                 ret = track_pfn_copy(vma);
1063                 if (ret)
1064                         return ret;
1065         }
1066
1067         /*
1068          * We need to invalidate the secondary MMU mappings only when
1069          * there could be a permission downgrade on the ptes of the
1070          * parent mm. And a permission downgrade will only happen if
1071          * is_cow_mapping() returns true.
1072          */
1073         is_cow = is_cow_mapping(vma->vm_flags);
1074         mmun_start = addr;
1075         mmun_end   = end;
1076         if (is_cow)
1077                 mmu_notifier_invalidate_range_start(src_mm, mmun_start,
1078                                                     mmun_end);
1079
1080         ret = 0;
1081         dst_pgd = pgd_offset(dst_mm, addr);
1082         src_pgd = pgd_offset(src_mm, addr);
1083         do {
1084                 next = pgd_addr_end(addr, end);
1085                 if (pgd_none_or_clear_bad(src_pgd))
1086                         continue;
1087                 if (unlikely(copy_pud_range(dst_mm, src_mm, dst_pgd, src_pgd,
1088                                             vma, addr, next))) {
1089                         ret = -ENOMEM;
1090                         break;
1091                 }
1092         } while (dst_pgd++, src_pgd++, addr = next, addr != end);
1093
1094         if (is_cow)
1095                 mmu_notifier_invalidate_range_end(src_mm, mmun_start, mmun_end);
1096         return ret;
1097 }
1098
1099 static unsigned long zap_pte_range(struct mmu_gather *tlb,
1100                                 struct vm_area_struct *vma, pmd_t *pmd,
1101                                 unsigned long addr, unsigned long end,
1102                                 struct zap_details *details)
1103 {
1104         struct mm_struct *mm = tlb->mm;
1105         int force_flush = 0;
1106         int rss[NR_MM_COUNTERS];
1107         spinlock_t *ptl;
1108         pte_t *start_pte;
1109         pte_t *pte;
1110
1111 again:
1112         init_rss_vec(rss);
1113         start_pte = pte_offset_map_lock(mm, pmd, addr, &ptl);
1114         pte = start_pte;
1115         arch_enter_lazy_mmu_mode();
1116         do {
1117                 pte_t ptent = *pte;
1118                 if (pte_none(ptent)) {
1119                         continue;
1120                 }
1121
1122                 if (pte_present(ptent)) {
1123                         struct page *page;
1124
1125                         page = vm_normal_page(vma, addr, ptent);
1126                         if (unlikely(details) && page) {
1127                                 /*
1128                                  * unmap_shared_mapping_pages() wants to
1129                                  * invalidate cache without truncating:
1130                                  * unmap shared but keep private pages.
1131                                  */
1132                                 if (details->check_mapping &&
1133                                     details->check_mapping != page->mapping)
1134                                         continue;
1135                                 /*
1136                                  * Each page->index must be checked when
1137                                  * invalidating or truncating nonlinear.
1138                                  */
1139                                 if (details->nonlinear_vma &&
1140                                     (page->index < details->first_index ||
1141                                      page->index > details->last_index))
1142                                         continue;
1143                         }
1144                         ptent = ptep_get_and_clear_full(mm, addr, pte,
1145                                                         tlb->fullmm);
1146                         tlb_remove_tlb_entry(tlb, pte, addr);
1147                         if (unlikely(!page))
1148                                 continue;
1149                         if (unlikely(details) && details->nonlinear_vma
1150                             && linear_page_index(details->nonlinear_vma,
1151                                                 addr) != page->index)
1152                                 set_pte_at(mm, addr, pte,
1153                                            pgoff_to_pte(page->index));
1154                         if (PageAnon(page))
1155                                 rss[MM_ANONPAGES]--;
1156                         else {
1157                                 if (pte_dirty(ptent))
1158                                         set_page_dirty(page);
1159                                 if (pte_young(ptent) &&
1160                                     likely(!VM_SequentialReadHint(vma)))
1161                                         mark_page_accessed(page);
1162                                 rss[MM_FILEPAGES]--;
1163                         }
1164                         page_remove_rmap(page);
1165                         if (unlikely(page_mapcount(page) < 0))
1166                                 print_bad_pte(vma, addr, ptent, page);
1167                         force_flush = !__tlb_remove_page(tlb, page);
1168                         if (force_flush)
1169                                 break;
1170                         continue;
1171                 }
1172                 /*
1173                  * If details->check_mapping, we leave swap entries;
1174                  * if details->nonlinear_vma, we leave file entries.
1175                  */
1176                 if (unlikely(details))
1177                         continue;
1178                 if (pte_file(ptent)) {
1179                         if (unlikely(!(vma->vm_flags & VM_NONLINEAR)))
1180                                 print_bad_pte(vma, addr, ptent, NULL);
1181                 } else {
1182                         swp_entry_t entry = pte_to_swp_entry(ptent);
1183
1184                         if (!non_swap_entry(entry))
1185                                 rss[MM_SWAPENTS]--;
1186                         else if (is_migration_entry(entry)) {
1187                                 struct page *page;
1188
1189                                 page = migration_entry_to_page(entry);
1190
1191                                 if (PageAnon(page))
1192                                         rss[MM_ANONPAGES]--;
1193                                 else
1194                                         rss[MM_FILEPAGES]--;
1195                         }
1196                         if (unlikely(!free_swap_and_cache(entry)))
1197                                 print_bad_pte(vma, addr, ptent, NULL);
1198                 }
1199                 pte_clear_not_present_full(mm, addr, pte, tlb->fullmm);
1200         } while (pte++, addr += PAGE_SIZE, addr != end);
1201
1202         add_mm_rss_vec(mm, rss);
1203         arch_leave_lazy_mmu_mode();
1204         pte_unmap_unlock(start_pte, ptl);
1205
1206         /*
1207          * mmu_gather ran out of room to batch pages, we break out of
1208          * the PTE lock to avoid doing the potential expensive TLB invalidate
1209          * and page-free while holding it.
1210          */
1211         if (force_flush) {
1212                 unsigned long old_end;
1213
1214                 force_flush = 0;
1215
1216                 /*
1217                  * Flush the TLB just for the previous segment,
1218                  * then update the range to be the remaining
1219                  * TLB range.
1220                  */
1221                 old_end = tlb->end;
1222                 tlb->end = addr;
1223
1224                 tlb_flush_mmu(tlb);
1225
1226                 tlb->start = addr;
1227                 tlb->end = old_end;
1228
1229                 if (addr != end)
1230                         goto again;
1231         }
1232
1233         return addr;
1234 }
1235
1236 static inline unsigned long zap_pmd_range(struct mmu_gather *tlb,
1237                                 struct vm_area_struct *vma, pud_t *pud,
1238                                 unsigned long addr, unsigned long end,
1239                                 struct zap_details *details)
1240 {
1241         pmd_t *pmd;
1242         unsigned long next;
1243
1244         pmd = pmd_offset(pud, addr);
1245         do {
1246                 next = pmd_addr_end(addr, end);
1247                 if (pmd_trans_huge(*pmd)) {
1248                         if (next - addr != HPAGE_PMD_SIZE) {
1249 #ifdef CONFIG_DEBUG_VM
1250                                 if (!rwsem_is_locked(&tlb->mm->mmap_sem)) {
1251                                         pr_err("%s: mmap_sem is unlocked! addr=0x%lx end=0x%lx vma->vm_start=0x%lx vma->vm_end=0x%lx\n",
1252                                                 __func__, addr, end,
1253                                                 vma->vm_start,
1254                                                 vma->vm_end);
1255                                         BUG();
1256                                 }
1257 #endif
1258                                 split_huge_page_pmd(vma, addr, pmd);
1259                         } else if (zap_huge_pmd(tlb, vma, pmd, addr))
1260                                 goto next;
1261                         /* fall through */
1262                 }
1263                 /*
1264                  * Here there can be other concurrent MADV_DONTNEED or
1265                  * trans huge page faults running, and if the pmd is
1266                  * none or trans huge it can change under us. This is
1267                  * because MADV_DONTNEED holds the mmap_sem in read
1268                  * mode.
1269                  */
1270                 if (pmd_none_or_trans_huge_or_clear_bad(pmd))
1271                         goto next;
1272                 next = zap_pte_range(tlb, vma, pmd, addr, next, details);
1273 next:
1274                 cond_resched();
1275         } while (pmd++, addr = next, addr != end);
1276
1277         return addr;
1278 }
1279
1280 static inline unsigned long zap_pud_range(struct mmu_gather *tlb,
1281                                 struct vm_area_struct *vma, pgd_t *pgd,
1282                                 unsigned long addr, unsigned long end,
1283                                 struct zap_details *details)
1284 {
1285         pud_t *pud;
1286         unsigned long next;
1287
1288         pud = pud_offset(pgd, addr);
1289         do {
1290                 next = pud_addr_end(addr, end);
1291                 if (pud_none_or_clear_bad(pud))
1292                         continue;
1293                 next = zap_pmd_range(tlb, vma, pud, addr, next, details);
1294         } while (pud++, addr = next, addr != end);
1295
1296         return addr;
1297 }
1298
1299 static void unmap_page_range(struct mmu_gather *tlb,
1300                              struct vm_area_struct *vma,
1301                              unsigned long addr, unsigned long end,
1302                              struct zap_details *details)
1303 {
1304         pgd_t *pgd;
1305         unsigned long next;
1306
1307         if (details && !details->check_mapping && !details->nonlinear_vma)
1308                 details = NULL;
1309
1310         BUG_ON(addr >= end);
1311         mem_cgroup_uncharge_start();
1312         tlb_start_vma(tlb, vma);
1313         pgd = pgd_offset(vma->vm_mm, addr);
1314         do {
1315                 next = pgd_addr_end(addr, end);
1316                 if (pgd_none_or_clear_bad(pgd))
1317                         continue;
1318                 next = zap_pud_range(tlb, vma, pgd, addr, next, details);
1319         } while (pgd++, addr = next, addr != end);
1320         tlb_end_vma(tlb, vma);
1321         mem_cgroup_uncharge_end();
1322 }
1323
1324
1325 static void unmap_single_vma(struct mmu_gather *tlb,
1326                 struct vm_area_struct *vma, unsigned long start_addr,
1327                 unsigned long end_addr,
1328                 struct zap_details *details)
1329 {
1330         unsigned long start = max(vma->vm_start, start_addr);
1331         unsigned long end;
1332
1333         if (start >= vma->vm_end)
1334                 return;
1335         end = min(vma->vm_end, end_addr);
1336         if (end <= vma->vm_start)
1337                 return;
1338
1339         if (vma->vm_file)
1340                 uprobe_munmap(vma, start, end);
1341
1342         if (unlikely(vma->vm_flags & VM_PFNMAP))
1343                 untrack_pfn(vma, 0, 0);
1344
1345         if (start != end) {
1346                 if (unlikely(is_vm_hugetlb_page(vma))) {
1347                         /*
1348                          * It is undesirable to test vma->vm_file as it
1349                          * should be non-null for valid hugetlb area.
1350                          * However, vm_file will be NULL in the error
1351                          * cleanup path of do_mmap_pgoff. When
1352                          * hugetlbfs ->mmap method fails,
1353                          * do_mmap_pgoff() nullifies vma->vm_file
1354                          * before calling this function to clean up.
1355                          * Since no pte has actually been setup, it is
1356                          * safe to do nothing in this case.
1357                          */
1358                         if (vma->vm_file) {
1359                                 mutex_lock(&vma->vm_file->f_mapping->i_mmap_mutex);
1360                                 __unmap_hugepage_range_final(tlb, vma, start, end, NULL);
1361                                 mutex_unlock(&vma->vm_file->f_mapping->i_mmap_mutex);
1362                         }
1363                 } else
1364                         unmap_page_range(tlb, vma, start, end, details);
1365         }
1366 }
1367
1368 /**
1369  * unmap_vmas - unmap a range of memory covered by a list of vma's
1370  * @tlb: address of the caller's struct mmu_gather
1371  * @vma: the starting vma
1372  * @start_addr: virtual address at which to start unmapping
1373  * @end_addr: virtual address at which to end unmapping
1374  *
1375  * Unmap all pages in the vma list.
1376  *
1377  * Only addresses between `start' and `end' will be unmapped.
1378  *
1379  * The VMA list must be sorted in ascending virtual address order.
1380  *
1381  * unmap_vmas() assumes that the caller will flush the whole unmapped address
1382  * range after unmap_vmas() returns.  So the only responsibility here is to
1383  * ensure that any thus-far unmapped pages are flushed before unmap_vmas()
1384  * drops the lock and schedules.
1385  */
1386 void unmap_vmas(struct mmu_gather *tlb,
1387                 struct vm_area_struct *vma, unsigned long start_addr,
1388                 unsigned long end_addr)
1389 {
1390         struct mm_struct *mm = vma->vm_mm;
1391
1392         mmu_notifier_invalidate_range_start(mm, start_addr, end_addr);
1393         for ( ; vma && vma->vm_start < end_addr; vma = vma->vm_next)
1394                 unmap_single_vma(tlb, vma, start_addr, end_addr, NULL);
1395         mmu_notifier_invalidate_range_end(mm, start_addr, end_addr);
1396 }
1397
1398 /**
1399  * zap_page_range - remove user pages in a given range
1400  * @vma: vm_area_struct holding the applicable pages
1401  * @start: starting address of pages to zap
1402  * @size: number of bytes to zap
1403  * @details: details of nonlinear truncation or shared cache invalidation
1404  *
1405  * Caller must protect the VMA list
1406  */
1407 void zap_page_range(struct vm_area_struct *vma, unsigned long start,
1408                 unsigned long size, struct zap_details *details)
1409 {
1410         struct mm_struct *mm = vma->vm_mm;
1411         struct mmu_gather tlb;
1412         unsigned long end = start + size;
1413
1414         lru_add_drain();
1415         tlb_gather_mmu(&tlb, mm, start, end);
1416         update_hiwater_rss(mm);
1417         mmu_notifier_invalidate_range_start(mm, start, end);
1418         for ( ; vma && vma->vm_start < end; vma = vma->vm_next)
1419                 unmap_single_vma(&tlb, vma, start, end, details);
1420         mmu_notifier_invalidate_range_end(mm, start, end);
1421         tlb_finish_mmu(&tlb, start, end);
1422 }
1423
1424 /**
1425  * zap_page_range_single - remove user pages in a given range
1426  * @vma: vm_area_struct holding the applicable pages
1427  * @address: starting address of pages to zap
1428  * @size: number of bytes to zap
1429  * @details: details of nonlinear truncation or shared cache invalidation
1430  *
1431  * The range must fit into one VMA.
1432  */
1433 static void zap_page_range_single(struct vm_area_struct *vma, unsigned long address,
1434                 unsigned long size, struct zap_details *details)
1435 {
1436         struct mm_struct *mm = vma->vm_mm;
1437         struct mmu_gather tlb;
1438         unsigned long end = address + size;
1439
1440         lru_add_drain();
1441         tlb_gather_mmu(&tlb, mm, address, end);
1442         update_hiwater_rss(mm);
1443         mmu_notifier_invalidate_range_start(mm, address, end);
1444         unmap_single_vma(&tlb, vma, address, end, details);
1445         mmu_notifier_invalidate_range_end(mm, address, end);
1446         tlb_finish_mmu(&tlb, address, end);
1447 }
1448
1449 /**
1450  * zap_vma_ptes - remove ptes mapping the vma
1451  * @vma: vm_area_struct holding ptes to be zapped
1452  * @address: starting address of pages to zap
1453  * @size: number of bytes to zap
1454  *
1455  * This function only unmaps ptes assigned to VM_PFNMAP vmas.
1456  *
1457  * The entire address range must be fully contained within the vma.
1458  *
1459  * Returns 0 if successful.
1460  */
1461 int zap_vma_ptes(struct vm_area_struct *vma, unsigned long address,
1462                 unsigned long size)
1463 {
1464         if (address < vma->vm_start || address + size > vma->vm_end ||
1465                         !(vma->vm_flags & VM_PFNMAP))
1466                 return -1;
1467         zap_page_range_single(vma, address, size, NULL);
1468         return 0;
1469 }
1470 EXPORT_SYMBOL_GPL(zap_vma_ptes);
1471
1472 /**
1473  * follow_page_mask - look up a page descriptor from a user-virtual address
1474  * @vma: vm_area_struct mapping @address
1475  * @address: virtual address to look up
1476  * @flags: flags modifying lookup behaviour
1477  * @page_mask: on output, *page_mask is set according to the size of the page
1478  *
1479  * @flags can have FOLL_ flags set, defined in <linux/mm.h>
1480  *
1481  * Returns the mapped (struct page *), %NULL if no mapping exists, or
1482  * an error pointer if there is a mapping to something not represented
1483  * by a page descriptor (see also vm_normal_page()).
1484  */
1485 struct page *follow_page_mask(struct vm_area_struct *vma,
1486                               unsigned long address, unsigned int flags,
1487                               unsigned int *page_mask)
1488 {
1489         pgd_t *pgd;
1490         pud_t *pud;
1491         pmd_t *pmd;
1492         pte_t *ptep, pte;
1493         spinlock_t *ptl;
1494         struct page *page;
1495         struct mm_struct *mm = vma->vm_mm;
1496
1497         *page_mask = 0;
1498
1499         page = follow_huge_addr(mm, address, flags & FOLL_WRITE);
1500         if (!IS_ERR(page)) {
1501                 BUG_ON(flags & FOLL_GET);
1502                 goto out;
1503         }
1504
1505         page = NULL;
1506         pgd = pgd_offset(mm, address);
1507         if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
1508                 goto no_page_table;
1509
1510         pud = pud_offset(pgd, address);
1511         if (pud_none(*pud))
1512                 goto no_page_table;
1513         if (pud_huge(*pud) && vma->vm_flags & VM_HUGETLB) {
1514                 BUG_ON(flags & FOLL_GET);
1515                 page = follow_huge_pud(mm, address, pud, flags & FOLL_WRITE);
1516                 goto out;
1517         }
1518         if (unlikely(pud_bad(*pud)))
1519                 goto no_page_table;
1520
1521         pmd = pmd_offset(pud, address);
1522         if (pmd_none(*pmd))
1523                 goto no_page_table;
1524         if (pmd_huge(*pmd) && vma->vm_flags & VM_HUGETLB) {
1525                 BUG_ON(flags & FOLL_GET);
1526                 page = follow_huge_pmd(mm, address, pmd, flags & FOLL_WRITE);
1527                 goto out;
1528         }
1529         if ((flags & FOLL_NUMA) && pmd_numa(*pmd))
1530                 goto no_page_table;
1531         if (pmd_trans_huge(*pmd)) {
1532                 if (flags & FOLL_SPLIT) {
1533                         split_huge_page_pmd(vma, address, pmd);
1534                         goto split_fallthrough;
1535                 }
1536                 spin_lock(&mm->page_table_lock);
1537                 if (likely(pmd_trans_huge(*pmd))) {
1538                         if (unlikely(pmd_trans_splitting(*pmd))) {
1539                                 spin_unlock(&mm->page_table_lock);
1540                                 wait_split_huge_page(vma->anon_vma, pmd);
1541                         } else {
1542                                 page = follow_trans_huge_pmd(vma, address,
1543                                                              pmd, flags);
1544                                 spin_unlock(&mm->page_table_lock);
1545                                 *page_mask = HPAGE_PMD_NR - 1;
1546                                 goto out;
1547                         }
1548                 } else
1549                         spin_unlock(&mm->page_table_lock);
1550                 /* fall through */
1551         }
1552 split_fallthrough:
1553         if (unlikely(pmd_bad(*pmd)))
1554                 goto no_page_table;
1555
1556         ptep = pte_offset_map_lock(mm, pmd, address, &ptl);
1557
1558         pte = *ptep;
1559         if (!pte_present(pte)) {
1560                 swp_entry_t entry;
1561                 /*
1562                  * KSM's break_ksm() relies upon recognizing a ksm page
1563                  * even while it is being migrated, so for that case we
1564                  * need migration_entry_wait().
1565                  */
1566                 if (likely(!(flags & FOLL_MIGRATION)))
1567                         goto no_page;
1568                 if (pte_none(pte) || pte_file(pte))
1569                         goto no_page;
1570                 entry = pte_to_swp_entry(pte);
1571                 if (!is_migration_entry(entry))
1572                         goto no_page;
1573                 pte_unmap_unlock(ptep, ptl);
1574                 migration_entry_wait(mm, pmd, address);
1575                 goto split_fallthrough;
1576         }
1577         if ((flags & FOLL_NUMA) && pte_numa(pte))
1578                 goto no_page;
1579         if ((flags & FOLL_WRITE) && !pte_write(pte))
1580                 goto unlock;
1581
1582         page = vm_normal_page(vma, address, pte);
1583         if (unlikely(!page)) {
1584                 if ((flags & FOLL_DUMP) ||
1585                     !is_zero_pfn(pte_pfn(pte)))
1586                         goto bad_page;
1587                 page = pte_page(pte);
1588         }
1589
1590         if (flags & FOLL_GET)
1591                 get_page_foll(page);
1592         if (flags & FOLL_TOUCH) {
1593                 if ((flags & FOLL_WRITE) &&
1594                     !pte_dirty(pte) && !PageDirty(page))
1595                         set_page_dirty(page);
1596                 /*
1597                  * pte_mkyoung() would be more correct here, but atomic care
1598                  * is needed to avoid losing the dirty bit: it is easier to use
1599                  * mark_page_accessed().
1600                  */
1601                 mark_page_accessed(page);
1602         }
1603         if ((flags & FOLL_MLOCK) && (vma->vm_flags & VM_LOCKED)) {
1604                 /*
1605                  * The preliminary mapping check is mainly to avoid the
1606                  * pointless overhead of lock_page on the ZERO_PAGE
1607                  * which might bounce very badly if there is contention.
1608                  *
1609                  * If the page is already locked, we don't need to
1610                  * handle it now - vmscan will handle it later if and
1611                  * when it attempts to reclaim the page.
1612                  */
1613                 if (page->mapping && trylock_page(page)) {
1614                         lru_add_drain();  /* push cached pages to LRU */
1615                         /*
1616                          * Because we lock page here, and migration is
1617                          * blocked by the pte's page reference, and we
1618                          * know the page is still mapped, we don't even
1619                          * need to check for file-cache page truncation.
1620                          */
1621                         mlock_vma_page(page);
1622                         unlock_page(page);
1623                 }
1624         }
1625 unlock:
1626         pte_unmap_unlock(ptep, ptl);
1627 out:
1628         return page;
1629
1630 bad_page:
1631         pte_unmap_unlock(ptep, ptl);
1632         return ERR_PTR(-EFAULT);
1633
1634 no_page:
1635         pte_unmap_unlock(ptep, ptl);
1636         if (!pte_none(pte))
1637                 return page;
1638
1639 no_page_table:
1640         /*
1641          * When core dumping an enormous anonymous area that nobody
1642          * has touched so far, we don't want to allocate unnecessary pages or
1643          * page tables.  Return error instead of NULL to skip handle_mm_fault,
1644          * then get_dump_page() will return NULL to leave a hole in the dump.
1645          * But we can only make this optimization where a hole would surely
1646          * be zero-filled if handle_mm_fault() actually did handle it.
1647          */
1648         if ((flags & FOLL_DUMP) &&
1649             (!vma->vm_ops || !vma->vm_ops->fault))
1650                 return ERR_PTR(-EFAULT);
1651         return page;
1652 }
1653
1654 static inline int stack_guard_page(struct vm_area_struct *vma, unsigned long addr)
1655 {
1656         return stack_guard_page_start(vma, addr) ||
1657                stack_guard_page_end(vma, addr+PAGE_SIZE);
1658 }
1659
1660 /**
1661  * __get_user_pages() - pin user pages in memory
1662  * @tsk:        task_struct of target task
1663  * @mm:         mm_struct of target mm
1664  * @start:      starting user address
1665  * @nr_pages:   number of pages from start to pin
1666  * @gup_flags:  flags modifying pin behaviour
1667  * @pages:      array that receives pointers to the pages pinned.
1668  *              Should be at least nr_pages long. Or NULL, if caller
1669  *              only intends to ensure the pages are faulted in.
1670  * @vmas:       array of pointers to vmas corresponding to each page.
1671  *              Or NULL if the caller does not require them.
1672  * @nonblocking: whether waiting for disk IO or mmap_sem contention
1673  *
1674  * Returns number of pages pinned. This may be fewer than the number
1675  * requested. If nr_pages is 0 or negative, returns 0. If no pages
1676  * were pinned, returns -errno. Each page returned must be released
1677  * with a put_page() call when it is finished with. vmas will only
1678  * remain valid while mmap_sem is held.
1679  *
1680  * Must be called with mmap_sem held for read or write.
1681  *
1682  * __get_user_pages walks a process's page tables and takes a reference to
1683  * each struct page that each user address corresponds to at a given
1684  * instant. That is, it takes the page that would be accessed if a user
1685  * thread accesses the given user virtual address at that instant.
1686  *
1687  * This does not guarantee that the page exists in the user mappings when
1688  * __get_user_pages returns, and there may even be a completely different
1689  * page there in some cases (eg. if mmapped pagecache has been invalidated
1690  * and subsequently re faulted). However it does guarantee that the page
1691  * won't be freed completely. And mostly callers simply care that the page
1692  * contains data that was valid *at some point in time*. Typically, an IO
1693  * or similar operation cannot guarantee anything stronger anyway because
1694  * locks can't be held over the syscall boundary.
1695  *
1696  * If @gup_flags & FOLL_WRITE == 0, the page must not be written to. If
1697  * the page is written to, set_page_dirty (or set_page_dirty_lock, as
1698  * appropriate) must be called after the page is finished with, and
1699  * before put_page is called.
1700  *
1701  * If @nonblocking != NULL, __get_user_pages will not wait for disk IO
1702  * or mmap_sem contention, and if waiting is needed to pin all pages,
1703  * *@nonblocking will be set to 0.
1704  *
1705  * In most cases, get_user_pages or get_user_pages_fast should be used
1706  * instead of __get_user_pages. __get_user_pages should be used only if
1707  * you need some special @gup_flags.
1708  */
1709 long __get_user_pages(struct task_struct *tsk, struct mm_struct *mm,
1710                 unsigned long start, unsigned long nr_pages,
1711                 unsigned int gup_flags, struct page **pages,
1712                 struct vm_area_struct **vmas, int *nonblocking)
1713 {
1714         long i;
1715         unsigned long vm_flags;
1716         unsigned int page_mask;
1717
1718         if (!nr_pages)
1719                 return 0;
1720
1721         VM_BUG_ON(!!pages != !!(gup_flags & FOLL_GET));
1722
1723         /* 
1724          * Require read or write permissions.
1725          * If FOLL_FORCE is set, we only require the "MAY" flags.
1726          */
1727         vm_flags  = (gup_flags & FOLL_WRITE) ?
1728                         (VM_WRITE | VM_MAYWRITE) : (VM_READ | VM_MAYREAD);
1729         vm_flags &= (gup_flags & FOLL_FORCE) ?
1730                         (VM_MAYREAD | VM_MAYWRITE) : (VM_READ | VM_WRITE);
1731
1732         /*
1733          * If FOLL_FORCE and FOLL_NUMA are both set, handle_mm_fault
1734          * would be called on PROT_NONE ranges. We must never invoke
1735          * handle_mm_fault on PROT_NONE ranges or the NUMA hinting
1736          * page faults would unprotect the PROT_NONE ranges if
1737          * _PAGE_NUMA and _PAGE_PROTNONE are sharing the same pte/pmd
1738          * bitflag. So to avoid that, don't set FOLL_NUMA if
1739          * FOLL_FORCE is set.
1740          */
1741         if (!(gup_flags & FOLL_FORCE))
1742                 gup_flags |= FOLL_NUMA;
1743
1744         i = 0;
1745
1746         do {
1747                 struct vm_area_struct *vma;
1748
1749                 vma = find_extend_vma(mm, start);
1750                 if (!vma && in_gate_area(mm, start)) {
1751                         unsigned long pg = start & PAGE_MASK;
1752                         pgd_t *pgd;
1753                         pud_t *pud;
1754                         pmd_t *pmd;
1755                         pte_t *pte;
1756
1757                         /* user gate pages are read-only */
1758                         if (gup_flags & FOLL_WRITE)
1759                                 return i ? : -EFAULT;
1760                         if (pg > TASK_SIZE)
1761                                 pgd = pgd_offset_k(pg);
1762                         else
1763                                 pgd = pgd_offset_gate(mm, pg);
1764                         BUG_ON(pgd_none(*pgd));
1765                         pud = pud_offset(pgd, pg);
1766                         BUG_ON(pud_none(*pud));
1767                         pmd = pmd_offset(pud, pg);
1768                         if (pmd_none(*pmd))
1769                                 return i ? : -EFAULT;
1770                         VM_BUG_ON(pmd_trans_huge(*pmd));
1771                         pte = pte_offset_map(pmd, pg);
1772                         if (pte_none(*pte)) {
1773                                 pte_unmap(pte);
1774                                 return i ? : -EFAULT;
1775                         }
1776                         vma = get_gate_vma(mm);
1777                         if (pages) {
1778                                 struct page *page;
1779
1780                                 page = vm_normal_page(vma, start, *pte);
1781                                 if (!page) {
1782                                         if (!(gup_flags & FOLL_DUMP) &&
1783                                              is_zero_pfn(pte_pfn(*pte)))
1784                                                 page = pte_page(*pte);
1785                                         else {
1786                                                 pte_unmap(pte);
1787                                                 return i ? : -EFAULT;
1788                                         }
1789                                 }
1790                                 pages[i] = page;
1791                                 get_page(page);
1792                         }
1793                         pte_unmap(pte);
1794                         page_mask = 0;
1795                         goto next_page;
1796                 }
1797
1798                 if (!vma ||
1799                     (vma->vm_flags & (VM_IO | VM_PFNMAP)) ||
1800                     !(vm_flags & vma->vm_flags))
1801                         return i ? : -EFAULT;
1802
1803                 if (is_vm_hugetlb_page(vma)) {
1804                         i = follow_hugetlb_page(mm, vma, pages, vmas,
1805                                         &start, &nr_pages, i, gup_flags);
1806                         continue;
1807                 }
1808
1809                 do {
1810                         struct page *page;
1811                         unsigned int foll_flags = gup_flags;
1812                         unsigned int page_increm;
1813
1814                         /*
1815                          * If we have a pending SIGKILL, don't keep faulting
1816                          * pages and potentially allocating memory.
1817                          */
1818                         if (unlikely(fatal_signal_pending(current)))
1819                                 return i ? i : -ERESTARTSYS;
1820
1821                         cond_resched();
1822                         while (!(page = follow_page_mask(vma, start,
1823                                                 foll_flags, &page_mask))) {
1824                                 int ret;
1825                                 unsigned int fault_flags = 0;
1826
1827                                 /* For mlock, just skip the stack guard page. */
1828                                 if (foll_flags & FOLL_MLOCK) {
1829                                         if (stack_guard_page(vma, start))
1830                                                 goto next_page;
1831                                 }
1832                                 if (foll_flags & FOLL_WRITE)
1833                                         fault_flags |= FAULT_FLAG_WRITE;
1834                                 if (nonblocking)
1835                                         fault_flags |= FAULT_FLAG_ALLOW_RETRY;
1836                                 if (foll_flags & FOLL_NOWAIT)
1837                                         fault_flags |= (FAULT_FLAG_ALLOW_RETRY | FAULT_FLAG_RETRY_NOWAIT);
1838
1839                                 ret = handle_mm_fault(mm, vma, start,
1840                                                         fault_flags);
1841
1842                                 if (ret & VM_FAULT_ERROR) {
1843                                         if (ret & VM_FAULT_OOM)
1844                                                 return i ? i : -ENOMEM;
1845                                         if (ret & (VM_FAULT_HWPOISON |
1846                                                    VM_FAULT_HWPOISON_LARGE)) {
1847                                                 if (i)
1848                                                         return i;
1849                                                 else if (gup_flags & FOLL_HWPOISON)
1850                                                         return -EHWPOISON;
1851                                                 else
1852                                                         return -EFAULT;
1853                                         }
1854                                         if (ret & VM_FAULT_SIGBUS)
1855                                                 return i ? i : -EFAULT;
1856                                         BUG();
1857                                 }
1858
1859                                 if (tsk) {
1860                                         if (ret & VM_FAULT_MAJOR)
1861                                                 tsk->maj_flt++;
1862                                         else
1863                                                 tsk->min_flt++;
1864                                 }
1865
1866                                 if (ret & VM_FAULT_RETRY) {
1867                                         if (nonblocking)
1868                                                 *nonblocking = 0;
1869                                         return i;
1870                                 }
1871
1872                                 /*
1873                                  * The VM_FAULT_WRITE bit tells us that
1874                                  * do_wp_page has broken COW when necessary,
1875                                  * even if maybe_mkwrite decided not to set
1876                                  * pte_write. We can thus safely do subsequent
1877                                  * page lookups as if they were reads. But only
1878                                  * do so when looping for pte_write is futile:
1879                                  * in some cases userspace may also be wanting
1880                                  * to write to the gotten user page, which a
1881                                  * read fault here might prevent (a readonly
1882                                  * page might get reCOWed by userspace write).
1883                                  */
1884                                 if ((ret & VM_FAULT_WRITE) &&
1885                                     !(vma->vm_flags & VM_WRITE))
1886                                         foll_flags &= ~FOLL_WRITE;
1887
1888                                 cond_resched();
1889                         }
1890                         if (IS_ERR(page))
1891                                 return i ? i : PTR_ERR(page);
1892                         if (pages) {
1893                                 pages[i] = page;
1894
1895                                 flush_anon_page(vma, page, start);
1896                                 flush_dcache_page(page);
1897                                 page_mask = 0;
1898                         }
1899 next_page:
1900                         if (vmas) {
1901                                 vmas[i] = vma;
1902                                 page_mask = 0;
1903                         }
1904                         page_increm = 1 + (~(start >> PAGE_SHIFT) & page_mask);
1905                         if (page_increm > nr_pages)
1906                                 page_increm = nr_pages;
1907                         i += page_increm;
1908                         start += page_increm * PAGE_SIZE;
1909                         nr_pages -= page_increm;
1910                 } while (nr_pages && start < vma->vm_end);
1911         } while (nr_pages);
1912         return i;
1913 }
1914 EXPORT_SYMBOL(__get_user_pages);
1915
1916 /*
1917  * fixup_user_fault() - manually resolve a user page fault
1918  * @tsk:        the task_struct to use for page fault accounting, or
1919  *              NULL if faults are not to be recorded.
1920  * @mm:         mm_struct of target mm
1921  * @address:    user address
1922  * @fault_flags:flags to pass down to handle_mm_fault()
1923  *
1924  * This is meant to be called in the specific scenario where for locking reasons
1925  * we try to access user memory in atomic context (within a pagefault_disable()
1926  * section), this returns -EFAULT, and we want to resolve the user fault before
1927  * trying again.
1928  *
1929  * Typically this is meant to be used by the futex code.
1930  *
1931  * The main difference with get_user_pages() is that this function will
1932  * unconditionally call handle_mm_fault() which will in turn perform all the
1933  * necessary SW fixup of the dirty and young bits in the PTE, while
1934  * handle_mm_fault() only guarantees to update these in the struct page.
1935  *
1936  * This is important for some architectures where those bits also gate the
1937  * access permission to the page because they are maintained in software.  On
1938  * such architectures, gup() will not be enough to make a subsequent access
1939  * succeed.
1940  *
1941  * This should be called with the mm_sem held for read.
1942  */
1943 int fixup_user_fault(struct task_struct *tsk, struct mm_struct *mm,
1944                      unsigned long address, unsigned int fault_flags)
1945 {
1946         struct vm_area_struct *vma;
1947         vm_flags_t vm_flags;
1948         int ret;
1949
1950         vma = find_extend_vma(mm, address);
1951         if (!vma || address < vma->vm_start)
1952                 return -EFAULT;
1953
1954         vm_flags = (fault_flags & FAULT_FLAG_WRITE) ? VM_WRITE : VM_READ;
1955         if (!(vm_flags & vma->vm_flags))
1956                 return -EFAULT;
1957
1958         ret = handle_mm_fault(mm, vma, address, fault_flags);
1959         if (ret & VM_FAULT_ERROR) {
1960                 if (ret & VM_FAULT_OOM)
1961                         return -ENOMEM;
1962                 if (ret & (VM_FAULT_HWPOISON | VM_FAULT_HWPOISON_LARGE))
1963                         return -EHWPOISON;
1964                 if (ret & VM_FAULT_SIGBUS)
1965                         return -EFAULT;
1966                 BUG();
1967         }
1968         if (tsk) {
1969                 if (ret & VM_FAULT_MAJOR)
1970                         tsk->maj_flt++;
1971                 else
1972                         tsk->min_flt++;
1973         }
1974         return 0;
1975 }
1976
1977 /*
1978  * get_user_pages() - pin user pages in memory
1979  * @tsk:        the task_struct to use for page fault accounting, or
1980  *              NULL if faults are not to be recorded.
1981  * @mm:         mm_struct of target mm
1982  * @start:      starting user address
1983  * @nr_pages:   number of pages from start to pin
1984  * @write:      whether pages will be written to by the caller
1985  * @force:      whether to force write access even if user mapping is
1986  *              readonly. This will result in the page being COWed even
1987  *              in MAP_SHARED mappings. You do not want this.
1988  * @pages:      array that receives pointers to the pages pinned.
1989  *              Should be at least nr_pages long. Or NULL, if caller
1990  *              only intends to ensure the pages are faulted in.
1991  * @vmas:       array of pointers to vmas corresponding to each page.
1992  *              Or NULL if the caller does not require them.
1993  *
1994  * Returns number of pages pinned. This may be fewer than the number
1995  * requested. If nr_pages is 0 or negative, returns 0. If no pages
1996  * were pinned, returns -errno. Each page returned must be released
1997  * with a put_page() call when it is finished with. vmas will only
1998  * remain valid while mmap_sem is held.
1999  *
2000  * Must be called with mmap_sem held for read or write.
2001  *
2002  * get_user_pages walks a process's page tables and takes a reference to
2003  * each struct page that each user address corresponds to at a given
2004  * instant. That is, it takes the page that would be accessed if a user
2005  * thread accesses the given user virtual address at that instant.
2006  *
2007  * This does not guarantee that the page exists in the user mappings when
2008  * get_user_pages returns, and there may even be a completely different
2009  * page there in some cases (eg. if mmapped pagecache has been invalidated
2010  * and subsequently re faulted). However it does guarantee that the page
2011  * won't be freed completely. And mostly callers simply care that the page
2012  * contains data that was valid *at some point in time*. Typically, an IO
2013  * or similar operation cannot guarantee anything stronger anyway because
2014  * locks can't be held over the syscall boundary.
2015  *
2016  * If write=0, the page must not be written to. If the page is written to,
2017  * set_page_dirty (or set_page_dirty_lock, as appropriate) must be called
2018  * after the page is finished with, and before put_page is called.
2019  *
2020  * get_user_pages is typically used for fewer-copy IO operations, to get a
2021  * handle on the memory by some means other than accesses via the user virtual
2022  * addresses. The pages may be submitted for DMA to devices or accessed via
2023  * their kernel linear mapping (via the kmap APIs). Care should be taken to
2024  * use the correct cache flushing APIs.
2025  *
2026  * See also get_user_pages_fast, for performance critical applications.
2027  */
2028 long get_user_pages(struct task_struct *tsk, struct mm_struct *mm,
2029                 unsigned long start, unsigned long nr_pages, int write,
2030                 int force, struct page **pages, struct vm_area_struct **vmas)
2031 {
2032         int flags = FOLL_TOUCH;
2033
2034         if (pages)
2035                 flags |= FOLL_GET;
2036         if (write)
2037                 flags |= FOLL_WRITE;
2038         if (force)
2039                 flags |= FOLL_FORCE;
2040
2041         return __get_user_pages(tsk, mm, start, nr_pages, flags, pages, vmas,
2042                                 NULL);
2043 }
2044 EXPORT_SYMBOL(get_user_pages);
2045
2046 /**
2047  * get_dump_page() - pin user page in memory while writing it to core dump
2048  * @addr: user address
2049  *
2050  * Returns struct page pointer of user page pinned for dump,
2051  * to be freed afterwards by page_cache_release() or put_page().
2052  *
2053  * Returns NULL on any kind of failure - a hole must then be inserted into
2054  * the corefile, to preserve alignment with its headers; and also returns
2055  * NULL wherever the ZERO_PAGE, or an anonymous pte_none, has been found -
2056  * allowing a hole to be left in the corefile to save diskspace.
2057  *
2058  * Called without mmap_sem, but after all other threads have been killed.
2059  */
2060 #ifdef CONFIG_ELF_CORE
2061 struct page *get_dump_page(unsigned long addr)
2062 {
2063         struct vm_area_struct *vma;
2064         struct page *page;
2065
2066         if (__get_user_pages(current, current->mm, addr, 1,
2067                              FOLL_FORCE | FOLL_DUMP | FOLL_GET, &page, &vma,
2068                              NULL) < 1)
2069                 return NULL;
2070         flush_cache_page(vma, addr, page_to_pfn(page));
2071         return page;
2072 }
2073 #endif /* CONFIG_ELF_CORE */
2074
2075 pte_t *__get_locked_pte(struct mm_struct *mm, unsigned long addr,
2076                         spinlock_t **ptl)
2077 {
2078         pgd_t * pgd = pgd_offset(mm, addr);
2079         pud_t * pud = pud_alloc(mm, pgd, addr);
2080         if (pud) {
2081                 pmd_t * pmd = pmd_alloc(mm, pud, addr);
2082                 if (pmd) {
2083                         VM_BUG_ON(pmd_trans_huge(*pmd));
2084                         return pte_alloc_map_lock(mm, pmd, addr, ptl);
2085                 }
2086         }
2087         return NULL;
2088 }
2089
2090 /*
2091  * This is the old fallback for page remapping.
2092  *
2093  * For historical reasons, it only allows reserved pages. Only
2094  * old drivers should use this, and they needed to mark their
2095  * pages reserved for the old functions anyway.
2096  */
2097 static int insert_page(struct vm_area_struct *vma, unsigned long addr,
2098                         struct page *page, pgprot_t prot)
2099 {
2100         struct mm_struct *mm = vma->vm_mm;
2101         int retval;
2102         pte_t *pte;
2103         spinlock_t *ptl;
2104
2105         retval = -EINVAL;
2106         if (PageAnon(page))
2107                 goto out;
2108         retval = -ENOMEM;
2109         flush_dcache_page(page);
2110         pte = get_locked_pte(mm, addr, &ptl);
2111         if (!pte)
2112                 goto out;
2113         retval = -EBUSY;
2114         if (!pte_none(*pte))
2115                 goto out_unlock;
2116
2117         /* Ok, finally just insert the thing.. */
2118         get_page(page);
2119         inc_mm_counter_fast(mm, MM_FILEPAGES);
2120         page_add_file_rmap(page);
2121         set_pte_at(mm, addr, pte, mk_pte(page, prot));
2122
2123         retval = 0;
2124         pte_unmap_unlock(pte, ptl);
2125         return retval;
2126 out_unlock:
2127         pte_unmap_unlock(pte, ptl);
2128 out:
2129         return retval;
2130 }
2131
2132 /**
2133  * vm_insert_page - insert single page into user vma
2134  * @vma: user vma to map to
2135  * @addr: target user address of this page
2136  * @page: source kernel page
2137  *
2138  * This allows drivers to insert individual pages they've allocated
2139  * into a user vma.
2140  *
2141  * The page has to be a nice clean _individual_ kernel allocation.
2142  * If you allocate a compound page, you need to have marked it as
2143  * such (__GFP_COMP), or manually just split the page up yourself
2144  * (see split_page()).
2145  *
2146  * NOTE! Traditionally this was done with "remap_pfn_range()" which
2147  * took an arbitrary page protection parameter. This doesn't allow
2148  * that. Your vma protection will have to be set up correctly, which
2149  * means that if you want a shared writable mapping, you'd better
2150  * ask for a shared writable mapping!
2151  *
2152  * The page does not need to be reserved.
2153  *
2154  * Usually this function is called from f_op->mmap() handler
2155  * under mm->mmap_sem write-lock, so it can change vma->vm_flags.
2156  * Caller must set VM_MIXEDMAP on vma if it wants to call this
2157  * function from other places, for example from page-fault handler.
2158  */
2159 int vm_insert_page(struct vm_area_struct *vma, unsigned long addr,
2160                         struct page *page)
2161 {
2162         if (addr < vma->vm_start || addr >= vma->vm_end)
2163                 return -EFAULT;
2164         if (!page_count(page))
2165                 return -EINVAL;
2166         if (!(vma->vm_flags & VM_MIXEDMAP)) {
2167                 BUG_ON(down_read_trylock(&vma->vm_mm->mmap_sem));
2168                 BUG_ON(vma->vm_flags & VM_PFNMAP);
2169                 vma->vm_flags |= VM_MIXEDMAP;
2170         }
2171         return insert_page(vma, addr, page, vma->vm_page_prot);
2172 }
2173 EXPORT_SYMBOL(vm_insert_page);
2174
2175 static int insert_pfn(struct vm_area_struct *vma, unsigned long addr,
2176                         unsigned long pfn, pgprot_t prot)
2177 {
2178         struct mm_struct *mm = vma->vm_mm;
2179         int retval;
2180         pte_t *pte, entry;
2181         spinlock_t *ptl;
2182
2183         retval = -ENOMEM;
2184         pte = get_locked_pte(mm, addr, &ptl);
2185         if (!pte)
2186                 goto out;
2187         retval = -EBUSY;
2188         if (!pte_none(*pte))
2189                 goto out_unlock;
2190
2191         /* Ok, finally just insert the thing.. */
2192         entry = pte_mkspecial(pfn_pte(pfn, prot));
2193         set_pte_at(mm, addr, pte, entry);
2194         update_mmu_cache(vma, addr, pte); /* XXX: why not for insert_page? */
2195
2196         retval = 0;
2197 out_unlock:
2198         pte_unmap_unlock(pte, ptl);
2199 out:
2200         return retval;
2201 }
2202
2203 /**
2204  * vm_insert_pfn - insert single pfn into user vma
2205  * @vma: user vma to map to
2206  * @addr: target user address of this page
2207  * @pfn: source kernel pfn
2208  *
2209  * Similar to vm_insert_page, this allows drivers to insert individual pages
2210  * they've allocated into a user vma. Same comments apply.
2211  *
2212  * This function should only be called from a vm_ops->fault handler, and
2213  * in that case the handler should return NULL.
2214  *
2215  * vma cannot be a COW mapping.
2216  *
2217  * As this is called only for pages that do not currently exist, we
2218  * do not need to flush old virtual caches or the TLB.
2219  */
2220 int vm_insert_pfn(struct vm_area_struct *vma, unsigned long addr,
2221                         unsigned long pfn)
2222 {
2223         int ret;
2224         pgprot_t pgprot = vma->vm_page_prot;
2225         /*
2226          * Technically, architectures with pte_special can avoid all these
2227          * restrictions (same for remap_pfn_range).  However we would like
2228          * consistency in testing and feature parity among all, so we should
2229          * try to keep these invariants in place for everybody.
2230          */
2231         BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)));
2232         BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) ==
2233                                                 (VM_PFNMAP|VM_MIXEDMAP));
2234         BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags));
2235         BUG_ON((vma->vm_flags & VM_MIXEDMAP) && pfn_valid(pfn));
2236
2237         if (addr < vma->vm_start || addr >= vma->vm_end)
2238                 return -EFAULT;
2239         if (track_pfn_insert(vma, &pgprot, pfn))
2240                 return -EINVAL;
2241
2242         ret = insert_pfn(vma, addr, pfn, pgprot);
2243
2244         return ret;
2245 }
2246 EXPORT_SYMBOL(vm_insert_pfn);
2247
2248 int vm_insert_mixed(struct vm_area_struct *vma, unsigned long addr,
2249                         unsigned long pfn)
2250 {
2251         BUG_ON(!(vma->vm_flags & VM_MIXEDMAP));
2252
2253         if (addr < vma->vm_start || addr >= vma->vm_end)
2254                 return -EFAULT;
2255
2256         /*
2257          * If we don't have pte special, then we have to use the pfn_valid()
2258          * based VM_MIXEDMAP scheme (see vm_normal_page), and thus we *must*
2259          * refcount the page if pfn_valid is true (hence insert_page rather
2260          * than insert_pfn).  If a zero_pfn were inserted into a VM_MIXEDMAP
2261          * without pte special, it would there be refcounted as a normal page.
2262          */
2263         if (!HAVE_PTE_SPECIAL && pfn_valid(pfn)) {
2264                 struct page *page;
2265
2266                 page = pfn_to_page(pfn);
2267                 return insert_page(vma, addr, page, vma->vm_page_prot);
2268         }
2269         return insert_pfn(vma, addr, pfn, vma->vm_page_prot);
2270 }
2271 EXPORT_SYMBOL(vm_insert_mixed);
2272
2273 /*
2274  * maps a range of physical memory into the requested pages. the old
2275  * mappings are removed. any references to nonexistent pages results
2276  * in null mappings (currently treated as "copy-on-access")
2277  */
2278 static int remap_pte_range(struct mm_struct *mm, pmd_t *pmd,
2279                         unsigned long addr, unsigned long end,
2280                         unsigned long pfn, pgprot_t prot)
2281 {
2282         pte_t *pte;
2283         spinlock_t *ptl;
2284
2285         pte = pte_alloc_map_lock(mm, pmd, addr, &ptl);
2286         if (!pte)
2287                 return -ENOMEM;
2288         arch_enter_lazy_mmu_mode();
2289         do {
2290                 BUG_ON(!pte_none(*pte));
2291                 set_pte_at(mm, addr, pte, pte_mkspecial(pfn_pte(pfn, prot)));
2292                 pfn++;
2293         } while (pte++, addr += PAGE_SIZE, addr != end);
2294         arch_leave_lazy_mmu_mode();
2295         pte_unmap_unlock(pte - 1, ptl);
2296         return 0;
2297 }
2298
2299 static inline int remap_pmd_range(struct mm_struct *mm, pud_t *pud,
2300                         unsigned long addr, unsigned long end,
2301                         unsigned long pfn, pgprot_t prot)
2302 {
2303         pmd_t *pmd;
2304         unsigned long next;
2305
2306         pfn -= addr >> PAGE_SHIFT;
2307         pmd = pmd_alloc(mm, pud, addr);
2308         if (!pmd)
2309                 return -ENOMEM;
2310         VM_BUG_ON(pmd_trans_huge(*pmd));
2311         do {
2312                 next = pmd_addr_end(addr, end);
2313                 if (remap_pte_range(mm, pmd, addr, next,
2314                                 pfn + (addr >> PAGE_SHIFT), prot))
2315                         return -ENOMEM;
2316         } while (pmd++, addr = next, addr != end);
2317         return 0;
2318 }
2319
2320 static inline int remap_pud_range(struct mm_struct *mm, pgd_t *pgd,
2321                         unsigned long addr, unsigned long end,
2322                         unsigned long pfn, pgprot_t prot)
2323 {
2324         pud_t *pud;
2325         unsigned long next;
2326
2327         pfn -= addr >> PAGE_SHIFT;
2328         pud = pud_alloc(mm, pgd, addr);
2329         if (!pud)
2330                 return -ENOMEM;
2331         do {
2332                 next = pud_addr_end(addr, end);
2333                 if (remap_pmd_range(mm, pud, addr, next,
2334                                 pfn + (addr >> PAGE_SHIFT), prot))
2335                         return -ENOMEM;
2336         } while (pud++, addr = next, addr != end);
2337         return 0;
2338 }
2339
2340 /**
2341  * remap_pfn_range - remap kernel memory to userspace
2342  * @vma: user vma to map to
2343  * @addr: target user address to start at
2344  * @pfn: physical address of kernel memory
2345  * @size: size of map area
2346  * @prot: page protection flags for this mapping
2347  *
2348  *  Note: this is only safe if the mm semaphore is held when called.
2349  */
2350 int remap_pfn_range(struct vm_area_struct *vma, unsigned long addr,
2351                     unsigned long pfn, unsigned long size, pgprot_t prot)
2352 {
2353         pgd_t *pgd;
2354         unsigned long next;
2355         unsigned long end = addr + PAGE_ALIGN(size);
2356         struct mm_struct *mm = vma->vm_mm;
2357         int err;
2358
2359         /*
2360          * Physically remapped pages are special. Tell the
2361          * rest of the world about it:
2362          *   VM_IO tells people not to look at these pages
2363          *      (accesses can have side effects).
2364          *   VM_PFNMAP tells the core MM that the base pages are just
2365          *      raw PFN mappings, and do not have a "struct page" associated
2366          *      with them.
2367          *   VM_DONTEXPAND
2368          *      Disable vma merging and expanding with mremap().
2369          *   VM_DONTDUMP
2370          *      Omit vma from core dump, even when VM_IO turned off.
2371          *
2372          * There's a horrible special case to handle copy-on-write
2373          * behaviour that some programs depend on. We mark the "original"
2374          * un-COW'ed pages by matching them up with "vma->vm_pgoff".
2375          * See vm_normal_page() for details.
2376          */
2377         if (is_cow_mapping(vma->vm_flags)) {
2378                 if (addr != vma->vm_start || end != vma->vm_end)
2379                         return -EINVAL;
2380                 vma->vm_pgoff = pfn;
2381         }
2382
2383         err = track_pfn_remap(vma, &prot, pfn, addr, PAGE_ALIGN(size));
2384         if (err)
2385                 return -EINVAL;
2386
2387         vma->vm_flags |= VM_IO | VM_PFNMAP | VM_DONTEXPAND | VM_DONTDUMP;
2388
2389         BUG_ON(addr >= end);
2390         pfn -= addr >> PAGE_SHIFT;
2391         pgd = pgd_offset(mm, addr);
2392         flush_cache_range(vma, addr, end);
2393         do {
2394                 next = pgd_addr_end(addr, end);
2395                 err = remap_pud_range(mm, pgd, addr, next,
2396                                 pfn + (addr >> PAGE_SHIFT), prot);
2397                 if (err)
2398                         break;
2399         } while (pgd++, addr = next, addr != end);
2400
2401         if (err)
2402                 untrack_pfn(vma, pfn, PAGE_ALIGN(size));
2403
2404         return err;
2405 }
2406 EXPORT_SYMBOL(remap_pfn_range);
2407
2408 /**
2409  * vm_iomap_memory - remap memory to userspace
2410  * @vma: user vma to map to
2411  * @start: start of area
2412  * @len: size of area
2413  *
2414  * This is a simplified io_remap_pfn_range() for common driver use. The
2415  * driver just needs to give us the physical memory range to be mapped,
2416  * we'll figure out the rest from the vma information.
2417  *
2418  * NOTE! Some drivers might want to tweak vma->vm_page_prot first to get
2419  * whatever write-combining details or similar.
2420  */
2421 int vm_iomap_memory(struct vm_area_struct *vma, phys_addr_t start, unsigned long len)
2422 {
2423         unsigned long vm_len, pfn, pages;
2424
2425         /* Check that the physical memory area passed in looks valid */
2426         if (start + len < start)
2427                 return -EINVAL;
2428         /*
2429          * You *really* shouldn't map things that aren't page-aligned,
2430          * but we've historically allowed it because IO memory might
2431          * just have smaller alignment.
2432          */
2433         len += start & ~PAGE_MASK;
2434         pfn = start >> PAGE_SHIFT;
2435         pages = (len + ~PAGE_MASK) >> PAGE_SHIFT;
2436         if (pfn + pages < pfn)
2437                 return -EINVAL;
2438
2439         /* We start the mapping 'vm_pgoff' pages into the area */
2440         if (vma->vm_pgoff > pages)
2441                 return -EINVAL;
2442         pfn += vma->vm_pgoff;
2443         pages -= vma->vm_pgoff;
2444
2445         /* Can we fit all of the mapping? */
2446         vm_len = vma->vm_end - vma->vm_start;
2447         if (vm_len >> PAGE_SHIFT > pages)
2448                 return -EINVAL;
2449
2450         /* Ok, let it rip */
2451         return io_remap_pfn_range(vma, vma->vm_start, pfn, vm_len, vma->vm_page_prot);
2452 }
2453 EXPORT_SYMBOL(vm_iomap_memory);
2454
2455 static int apply_to_pte_range(struct mm_struct *mm, pmd_t *pmd,
2456                                      unsigned long addr, unsigned long end,
2457                                      pte_fn_t fn, void *data)
2458 {
2459         pte_t *pte;
2460         int err;
2461         pgtable_t token;
2462         spinlock_t *uninitialized_var(ptl);
2463
2464         pte = (mm == &init_mm) ?
2465                 pte_alloc_kernel(pmd, addr) :
2466                 pte_alloc_map_lock(mm, pmd, addr, &ptl);
2467         if (!pte)
2468                 return -ENOMEM;
2469
2470         BUG_ON(pmd_huge(*pmd));
2471
2472         arch_enter_lazy_mmu_mode();
2473
2474         token = pmd_pgtable(*pmd);
2475
2476         do {
2477                 err = fn(pte++, token, addr, data);
2478                 if (err)
2479                         break;
2480         } while (addr += PAGE_SIZE, addr != end);
2481
2482         arch_leave_lazy_mmu_mode();
2483
2484         if (mm != &init_mm)
2485                 pte_unmap_unlock(pte-1, ptl);
2486         return err;
2487 }
2488
2489 static int apply_to_pmd_range(struct mm_struct *mm, pud_t *pud,
2490                                      unsigned long addr, unsigned long end,
2491                                      pte_fn_t fn, void *data)
2492 {
2493         pmd_t *pmd;
2494         unsigned long next;
2495         int err;
2496
2497         BUG_ON(pud_huge(*pud));
2498
2499         pmd = pmd_alloc(mm, pud, addr);
2500         if (!pmd)
2501                 return -ENOMEM;
2502         do {
2503                 next = pmd_addr_end(addr, end);
2504                 err = apply_to_pte_range(mm, pmd, addr, next, fn, data);
2505                 if (err)
2506                         break;
2507         } while (pmd++, addr = next, addr != end);
2508         return err;
2509 }
2510
2511 static int apply_to_pud_range(struct mm_struct *mm, pgd_t *pgd,
2512                                      unsigned long addr, unsigned long end,
2513                                      pte_fn_t fn, void *data)
2514 {
2515         pud_t *pud;
2516         unsigned long next;
2517         int err;
2518
2519         pud = pud_alloc(mm, pgd, addr);
2520         if (!pud)
2521                 return -ENOMEM;
2522         do {
2523                 next = pud_addr_end(addr, end);
2524                 err = apply_to_pmd_range(mm, pud, addr, next, fn, data);
2525                 if (err)
2526                         break;
2527         } while (pud++, addr = next, addr != end);
2528         return err;
2529 }
2530
2531 /*
2532  * Scan a region of virtual memory, filling in page tables as necessary
2533  * and calling a provided function on each leaf page table.
2534  */
2535 int apply_to_page_range(struct mm_struct *mm, unsigned long addr,
2536                         unsigned long size, pte_fn_t fn, void *data)
2537 {
2538         pgd_t *pgd;
2539         unsigned long next;
2540         unsigned long end = addr + size;
2541         int err;
2542
2543         BUG_ON(addr >= end);
2544         pgd = pgd_offset(mm, addr);
2545         do {
2546                 next = pgd_addr_end(addr, end);
2547                 err = apply_to_pud_range(mm, pgd, addr, next, fn, data);
2548                 if (err)
2549                         break;
2550         } while (pgd++, addr = next, addr != end);
2551
2552         return err;
2553 }
2554 EXPORT_SYMBOL_GPL(apply_to_page_range);
2555
2556 /*
2557  * handle_pte_fault chooses page fault handler according to an entry
2558  * which was read non-atomically.  Before making any commitment, on
2559  * those architectures or configurations (e.g. i386 with PAE) which
2560  * might give a mix of unmatched parts, do_swap_page and do_nonlinear_fault
2561  * must check under lock before unmapping the pte and proceeding
2562  * (but do_wp_page is only called after already making such a check;
2563  * and do_anonymous_page can safely check later on).
2564  */
2565 static inline int pte_unmap_same(struct mm_struct *mm, pmd_t *pmd,
2566                                 pte_t *page_table, pte_t orig_pte)
2567 {
2568         int same = 1;
2569 #if defined(CONFIG_SMP) || defined(CONFIG_PREEMPT)
2570         if (sizeof(pte_t) > sizeof(unsigned long)) {
2571                 spinlock_t *ptl = pte_lockptr(mm, pmd);
2572                 spin_lock(ptl);
2573                 same = pte_same(*page_table, orig_pte);
2574                 spin_unlock(ptl);
2575         }
2576 #endif
2577         pte_unmap(page_table);
2578         return same;
2579 }
2580
2581 static inline void cow_user_page(struct page *dst, struct page *src, unsigned long va, struct vm_area_struct *vma)
2582 {
2583         /*
2584          * If the source page was a PFN mapping, we don't have
2585          * a "struct page" for it. We do a best-effort copy by
2586          * just copying from the original user address. If that
2587          * fails, we just zero-fill it. Live with it.
2588          */
2589         if (unlikely(!src)) {
2590                 void *kaddr = kmap_atomic(dst);
2591                 void __user *uaddr = (void __user *)(va & PAGE_MASK);
2592
2593                 /*
2594                  * This really shouldn't fail, because the page is there
2595                  * in the page tables. But it might just be unreadable,
2596                  * in which case we just give up and fill the result with
2597                  * zeroes.
2598                  */
2599                 if (__copy_from_user_inatomic(kaddr, uaddr, PAGE_SIZE))
2600                         clear_page(kaddr);
2601                 kunmap_atomic(kaddr);
2602                 flush_dcache_page(dst);
2603         } else
2604                 copy_user_highpage(dst, src, va, vma);
2605 }
2606
2607 /*
2608  * This routine handles present pages, when users try to write
2609  * to a shared page. It is done by copying the page to a new address
2610  * and decrementing the shared-page counter for the old page.
2611  *
2612  * Note that this routine assumes that the protection checks have been
2613  * done by the caller (the low-level page fault routine in most cases).
2614  * Thus we can safely just mark it writable once we've done any necessary
2615  * COW.
2616  *
2617  * We also mark the page dirty at this point even though the page will
2618  * change only once the write actually happens. This avoids a few races,
2619  * and potentially makes it more efficient.
2620  *
2621  * We enter with non-exclusive mmap_sem (to exclude vma changes,
2622  * but allow concurrent faults), with pte both mapped and locked.
2623  * We return with mmap_sem still held, but pte unmapped and unlocked.
2624  */
2625 static int do_wp_page(struct mm_struct *mm, struct vm_area_struct *vma,
2626                 unsigned long address, pte_t *page_table, pmd_t *pmd,
2627                 spinlock_t *ptl, pte_t orig_pte)
2628         __releases(ptl)
2629 {
2630         struct page *old_page, *new_page = NULL;
2631         pte_t entry;
2632         int ret = 0;
2633         int page_mkwrite = 0;
2634         struct page *dirty_page = NULL;
2635         unsigned long mmun_start = 0;   /* For mmu_notifiers */
2636         unsigned long mmun_end = 0;     /* For mmu_notifiers */
2637
2638         old_page = vm_normal_page(vma, address, orig_pte);
2639         if (!old_page) {
2640                 /*
2641                  * VM_MIXEDMAP !pfn_valid() case
2642                  *
2643                  * We should not cow pages in a shared writeable mapping.
2644                  * Just mark the pages writable as we can't do any dirty
2645                  * accounting on raw pfn maps.
2646                  */
2647                 if ((vma->vm_flags & (VM_WRITE|VM_SHARED)) ==
2648                                      (VM_WRITE|VM_SHARED))
2649                         goto reuse;
2650                 goto gotten;
2651         }
2652
2653         /*
2654          * Take out anonymous pages first, anonymous shared vmas are
2655          * not dirty accountable.
2656          */
2657         if (PageAnon(old_page) && !PageKsm(old_page)) {
2658                 if (!trylock_page(old_page)) {
2659                         page_cache_get(old_page);
2660                         pte_unmap_unlock(page_table, ptl);
2661                         lock_page(old_page);
2662                         page_table = pte_offset_map_lock(mm, pmd, address,
2663                                                          &ptl);
2664                         if (!pte_same(*page_table, orig_pte)) {
2665                                 unlock_page(old_page);
2666                                 goto unlock;
2667                         }
2668                         page_cache_release(old_page);
2669                 }
2670                 if (reuse_swap_page(old_page)) {
2671                         /*
2672                          * The page is all ours.  Move it to our anon_vma so
2673                          * the rmap code will not search our parent or siblings.
2674                          * Protected against the rmap code by the page lock.
2675                          */
2676                         page_move_anon_rmap(old_page, vma, address);
2677                         unlock_page(old_page);
2678                         goto reuse;
2679                 }
2680                 unlock_page(old_page);
2681         } else if (unlikely((vma->vm_flags & (VM_WRITE|VM_SHARED)) ==
2682                                         (VM_WRITE|VM_SHARED))) {
2683                 /*
2684                  * Only catch write-faults on shared writable pages,
2685                  * read-only shared pages can get COWed by
2686                  * get_user_pages(.write=1, .force=1).
2687                  */
2688                 if (vma->vm_ops && vma->vm_ops->page_mkwrite) {
2689                         struct vm_fault vmf;
2690                         int tmp;
2691
2692                         vmf.virtual_address = (void __user *)(address &
2693                                                                 PAGE_MASK);
2694                         vmf.pgoff = old_page->index;
2695                         vmf.flags = FAULT_FLAG_WRITE|FAULT_FLAG_MKWRITE;
2696                         vmf.page = old_page;
2697
2698                         /*
2699                          * Notify the address space that the page is about to
2700                          * become writable so that it can prohibit this or wait
2701                          * for the page to get into an appropriate state.
2702                          *
2703                          * We do this without the lock held, so that it can
2704                          * sleep if it needs to.
2705                          */
2706                         page_cache_get(old_page);
2707                         pte_unmap_unlock(page_table, ptl);
2708
2709                         tmp = vma->vm_ops->page_mkwrite(vma, &vmf);
2710                         if (unlikely(tmp &
2711                                         (VM_FAULT_ERROR | VM_FAULT_NOPAGE))) {
2712                                 ret = tmp;
2713                                 goto unwritable_page;
2714                         }
2715                         if (unlikely(!(tmp & VM_FAULT_LOCKED))) {
2716                                 lock_page(old_page);
2717                                 if (!old_page->mapping) {
2718                                         ret = 0; /* retry the fault */
2719                                         unlock_page(old_page);
2720                                         goto unwritable_page;
2721                                 }
2722                         } else
2723                                 VM_BUG_ON(!PageLocked(old_page));
2724
2725                         /*
2726                          * Since we dropped the lock we need to revalidate
2727                          * the PTE as someone else may have changed it.  If
2728                          * they did, we just return, as we can count on the
2729                          * MMU to tell us if they didn't also make it writable.
2730                          */
2731                         page_table = pte_offset_map_lock(mm, pmd, address,
2732                                                          &ptl);
2733                         if (!pte_same(*page_table, orig_pte)) {
2734                                 unlock_page(old_page);
2735                                 goto unlock;
2736                         }
2737
2738                         page_mkwrite = 1;
2739                 }
2740                 dirty_page = old_page;
2741                 get_page(dirty_page);
2742
2743 reuse:
2744                 flush_cache_page(vma, address, pte_pfn(orig_pte));
2745                 entry = pte_mkyoung(orig_pte);
2746                 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2747                 if (ptep_set_access_flags(vma, address, page_table, entry,1))
2748                         update_mmu_cache(vma, address, page_table);
2749                 pte_unmap_unlock(page_table, ptl);
2750                 ret |= VM_FAULT_WRITE;
2751
2752                 if (!dirty_page)
2753                         return ret;
2754
2755                 /*
2756                  * Yes, Virginia, this is actually required to prevent a race
2757                  * with clear_page_dirty_for_io() from clearing the page dirty
2758                  * bit after it clear all dirty ptes, but before a racing
2759                  * do_wp_page installs a dirty pte.
2760                  *
2761                  * __do_fault is protected similarly.
2762                  */
2763                 if (!page_mkwrite) {
2764                         wait_on_page_locked(dirty_page);
2765                         set_page_dirty_balance(dirty_page, page_mkwrite);
2766                         /* file_update_time outside page_lock */
2767                         if (vma->vm_file)
2768                                 file_update_time(vma->vm_file);
2769                 }
2770                 put_page(dirty_page);
2771                 if (page_mkwrite) {
2772                         struct address_space *mapping = dirty_page->mapping;
2773
2774                         set_page_dirty(dirty_page);
2775                         unlock_page(dirty_page);
2776                         page_cache_release(dirty_page);
2777                         if (mapping)    {
2778                                 /*
2779                                  * Some device drivers do not set page.mapping
2780                                  * but still dirty their pages
2781                                  */
2782                                 balance_dirty_pages_ratelimited(mapping);
2783                         }
2784                 }
2785
2786                 return ret;
2787         }
2788
2789         /*
2790          * Ok, we need to copy. Oh, well..
2791          */
2792         page_cache_get(old_page);
2793 gotten:
2794         pte_unmap_unlock(page_table, ptl);
2795
2796         if (unlikely(anon_vma_prepare(vma)))
2797                 goto oom;
2798
2799         if (is_zero_pfn(pte_pfn(orig_pte))) {
2800                 new_page = alloc_zeroed_user_highpage_movable(vma, address);
2801                 if (!new_page)
2802                         goto oom;
2803         } else {
2804                 new_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, address);
2805                 if (!new_page)
2806                         goto oom;
2807                 cow_user_page(new_page, old_page, address, vma);
2808         }
2809         __SetPageUptodate(new_page);
2810
2811         if (mem_cgroup_newpage_charge(new_page, mm, GFP_KERNEL))
2812                 goto oom_free_new;
2813
2814         mmun_start  = address & PAGE_MASK;
2815         mmun_end    = mmun_start + PAGE_SIZE;
2816         mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
2817
2818         /*
2819          * Re-check the pte - we dropped the lock
2820          */
2821         page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2822         if (likely(pte_same(*page_table, orig_pte))) {
2823                 if (old_page) {
2824                         if (!PageAnon(old_page)) {
2825                                 dec_mm_counter_fast(mm, MM_FILEPAGES);
2826                                 inc_mm_counter_fast(mm, MM_ANONPAGES);
2827                         }
2828                 } else
2829                         inc_mm_counter_fast(mm, MM_ANONPAGES);
2830                 flush_cache_page(vma, address, pte_pfn(orig_pte));
2831                 entry = mk_pte(new_page, vma->vm_page_prot);
2832                 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2833                 /*
2834                  * Clear the pte entry and flush it first, before updating the
2835                  * pte with the new entry. This will avoid a race condition
2836                  * seen in the presence of one thread doing SMC and another
2837                  * thread doing COW.
2838                  */
2839                 ptep_clear_flush(vma, address, page_table);
2840                 page_add_new_anon_rmap(new_page, vma, address);
2841                 /*
2842                  * We call the notify macro here because, when using secondary
2843                  * mmu page tables (such as kvm shadow page tables), we want the
2844                  * new page to be mapped directly into the secondary page table.
2845                  */
2846                 set_pte_at_notify(mm, address, page_table, entry);
2847                 update_mmu_cache(vma, address, page_table);
2848                 if (old_page) {
2849                         /*
2850                          * Only after switching the pte to the new page may
2851                          * we remove the mapcount here. Otherwise another
2852                          * process may come and find the rmap count decremented
2853                          * before the pte is switched to the new page, and
2854                          * "reuse" the old page writing into it while our pte
2855                          * here still points into it and can be read by other
2856                          * threads.
2857                          *
2858                          * The critical issue is to order this
2859                          * page_remove_rmap with the ptp_clear_flush above.
2860                          * Those stores are ordered by (if nothing else,)
2861                          * the barrier present in the atomic_add_negative
2862                          * in page_remove_rmap.
2863                          *
2864                          * Then the TLB flush in ptep_clear_flush ensures that
2865                          * no process can access the old page before the
2866                          * decremented mapcount is visible. And the old page
2867                          * cannot be reused until after the decremented
2868                          * mapcount is visible. So transitively, TLBs to
2869                          * old page will be flushed before it can be reused.
2870                          */
2871                         page_remove_rmap(old_page);
2872                 }
2873
2874                 /* Free the old page.. */
2875                 new_page = old_page;
2876                 ret |= VM_FAULT_WRITE;
2877         } else
2878                 mem_cgroup_uncharge_page(new_page);
2879
2880         if (new_page)
2881                 page_cache_release(new_page);
2882 unlock:
2883         pte_unmap_unlock(page_table, ptl);
2884         if (mmun_end > mmun_start)
2885                 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
2886         if (old_page) {
2887                 /*
2888                  * Don't let another task, with possibly unlocked vma,
2889                  * keep the mlocked page.
2890                  */
2891                 if ((ret & VM_FAULT_WRITE) && (vma->vm_flags & VM_LOCKED)) {
2892                         lock_page(old_page);    /* LRU manipulation */
2893                         munlock_vma_page(old_page);
2894                         unlock_page(old_page);
2895                 }
2896                 page_cache_release(old_page);
2897         }
2898         return ret;
2899 oom_free_new:
2900         page_cache_release(new_page);
2901 oom:
2902         if (old_page)
2903                 page_cache_release(old_page);
2904         return VM_FAULT_OOM;
2905
2906 unwritable_page:
2907         page_cache_release(old_page);
2908         return ret;
2909 }
2910
2911 static void unmap_mapping_range_vma(struct vm_area_struct *vma,
2912                 unsigned long start_addr, unsigned long end_addr,
2913                 struct zap_details *details)
2914 {
2915         zap_page_range_single(vma, start_addr, end_addr - start_addr, details);
2916 }
2917
2918 static inline void unmap_mapping_range_tree(struct rb_root *root,
2919                                             struct zap_details *details)
2920 {
2921         struct vm_area_struct *vma;
2922         pgoff_t vba, vea, zba, zea;
2923
2924         vma_interval_tree_foreach(vma, root,
2925                         details->first_index, details->last_index) {
2926
2927                 vba = vma->vm_pgoff;
2928                 vea = vba + ((vma->vm_end - vma->vm_start) >> PAGE_SHIFT) - 1;
2929                 /* Assume for now that PAGE_CACHE_SHIFT == PAGE_SHIFT */
2930                 zba = details->first_index;
2931                 if (zba < vba)
2932                         zba = vba;
2933                 zea = details->last_index;
2934                 if (zea > vea)
2935                         zea = vea;
2936
2937                 unmap_mapping_range_vma(vma,
2938                         ((zba - vba) << PAGE_SHIFT) + vma->vm_start,
2939                         ((zea - vba + 1) << PAGE_SHIFT) + vma->vm_start,
2940                                 details);
2941         }
2942 }
2943
2944 static inline void unmap_mapping_range_list(struct list_head *head,
2945                                             struct zap_details *details)
2946 {
2947         struct vm_area_struct *vma;
2948
2949         /*
2950          * In nonlinear VMAs there is no correspondence between virtual address
2951          * offset and file offset.  So we must perform an exhaustive search
2952          * across *all* the pages in each nonlinear VMA, not just the pages
2953          * whose virtual address lies outside the file truncation point.
2954          */
2955         list_for_each_entry(vma, head, shared.nonlinear) {
2956                 details->nonlinear_vma = vma;
2957                 unmap_mapping_range_vma(vma, vma->vm_start, vma->vm_end, details);
2958         }
2959 }
2960
2961 /**
2962  * unmap_mapping_range - unmap the portion of all mmaps in the specified address_space corresponding to the specified page range in the underlying file.
2963  * @mapping: the address space containing mmaps to be unmapped.
2964  * @holebegin: byte in first page to unmap, relative to the start of
2965  * the underlying file.  This will be rounded down to a PAGE_SIZE
2966  * boundary.  Note that this is different from truncate_pagecache(), which
2967  * must keep the partial page.  In contrast, we must get rid of
2968  * partial pages.
2969  * @holelen: size of prospective hole in bytes.  This will be rounded
2970  * up to a PAGE_SIZE boundary.  A holelen of zero truncates to the
2971  * end of the file.
2972  * @even_cows: 1 when truncating a file, unmap even private COWed pages;
2973  * but 0 when invalidating pagecache, don't throw away private data.
2974  */
2975 void unmap_mapping_range(struct address_space *mapping,
2976                 loff_t const holebegin, loff_t const holelen, int even_cows)
2977 {
2978         struct zap_details details;
2979         pgoff_t hba = holebegin >> PAGE_SHIFT;
2980         pgoff_t hlen = (holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
2981
2982         /* Check for overflow. */
2983         if (sizeof(holelen) > sizeof(hlen)) {
2984                 long long holeend =
2985                         (holebegin + holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
2986                 if (holeend & ~(long long)ULONG_MAX)
2987                         hlen = ULONG_MAX - hba + 1;
2988         }
2989
2990         details.check_mapping = even_cows? NULL: mapping;
2991         details.nonlinear_vma = NULL;
2992         details.first_index = hba;
2993         details.last_index = hba + hlen - 1;
2994         if (details.last_index < details.first_index)
2995                 details.last_index = ULONG_MAX;
2996
2997
2998         mutex_lock(&mapping->i_mmap_mutex);
2999         if (unlikely(!RB_EMPTY_ROOT(&mapping->i_mmap)))
3000                 unmap_mapping_range_tree(&mapping->i_mmap, &details);
3001         if (unlikely(!list_empty(&mapping->i_mmap_nonlinear)))
3002                 unmap_mapping_range_list(&mapping->i_mmap_nonlinear, &details);
3003         mutex_unlock(&mapping->i_mmap_mutex);
3004 }
3005 EXPORT_SYMBOL(unmap_mapping_range);
3006
3007 #ifdef CONFIG_ZRAM
3008 #include <linux/module.h>
3009 // always try to free swap in zram swap case
3010 static int keep_to_free_swap = 1;
3011 module_param_named(keep_to_free_swap,keep_to_free_swap, int, S_IRUGO | S_IWUSR);
3012 #endif
3013
3014
3015 /*
3016  * We enter with non-exclusive mmap_sem (to exclude vma changes,
3017  * but allow concurrent faults), and pte mapped but not yet locked.
3018  * We return with mmap_sem still held, but pte unmapped and unlocked.
3019  */
3020 static int do_swap_page(struct mm_struct *mm, struct vm_area_struct *vma,
3021                 unsigned long address, pte_t *page_table, pmd_t *pmd,
3022                 unsigned int flags, pte_t orig_pte)
3023 {
3024         spinlock_t *ptl;
3025         struct page *page, *swapcache;
3026         swp_entry_t entry;
3027         pte_t pte;
3028         int locked;
3029         struct mem_cgroup *ptr;
3030         int exclusive = 0;
3031         int ret = 0;
3032
3033         if (!pte_unmap_same(mm, pmd, page_table, orig_pte))
3034                 goto out;
3035
3036         entry = pte_to_swp_entry(orig_pte);
3037         if (unlikely(non_swap_entry(entry))) {
3038                 if (is_migration_entry(entry)) {
3039 #ifdef CONFIG_CMA
3040                         /*
3041                          * FIXME: mszyprow: cruel, brute-force method for
3042                          * letting cma/migration to finish it's job without
3043                          * stealing the lock migration_entry_wait() and creating
3044                          * a live-lock on the faulted page
3045                          * (page->_count == 2 migration failure issue)
3046                          */
3047                         mdelay(10);
3048 #endif
3049                         migration_entry_wait(mm, pmd, address);
3050                 } else if (is_hwpoison_entry(entry)) {
3051                         ret = VM_FAULT_HWPOISON;
3052                 } else {
3053                         print_bad_pte(vma, address, orig_pte, NULL);
3054                         ret = VM_FAULT_SIGBUS;
3055                 }
3056                 goto out;
3057         }
3058         delayacct_set_flag(DELAYACCT_PF_SWAPIN);
3059         page = lookup_swap_cache(entry);
3060         if (!page) {
3061                 page = swapin_readahead(entry,
3062                                         GFP_HIGHUSER_MOVABLE, vma, address);
3063                 if (!page) {
3064                         /*
3065                          * Back out if somebody else faulted in this pte
3066                          * while we released the pte lock.
3067                          */
3068                         page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
3069                         if (likely(pte_same(*page_table, orig_pte)))
3070                                 ret = VM_FAULT_OOM;
3071                         delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
3072                         goto unlock;
3073                 }
3074
3075                 /* Had to read the page from swap area: Major fault */
3076                 ret = VM_FAULT_MAJOR;
3077                 count_vm_event(PGMAJFAULT);
3078                 mem_cgroup_count_vm_event(mm, PGMAJFAULT);
3079         } else if (PageHWPoison(page)) {
3080                 /*
3081                  * hwpoisoned dirty swapcache pages are kept for killing
3082                  * owner processes (which may be unknown at hwpoison time)
3083                  */
3084                 ret = VM_FAULT_HWPOISON;
3085                 delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
3086                 swapcache = page;
3087                 goto out_release;
3088         }
3089
3090         swapcache = page;
3091         locked = lock_page_or_retry(page, mm, flags);
3092
3093         delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
3094         if (!locked) {
3095                 ret |= VM_FAULT_RETRY;
3096                 goto out_release;
3097         }
3098
3099         /*
3100          * Make sure try_to_free_swap or reuse_swap_page or swapoff did not
3101          * release the swapcache from under us.  The page pin, and pte_same
3102          * test below, are not enough to exclude that.  Even if it is still
3103          * swapcache, we need to check that the page's swap has not changed.
3104          */
3105         if (unlikely(!PageSwapCache(page) || page_private(page) != entry.val))
3106                 goto out_page;
3107
3108         page = ksm_might_need_to_copy(page, vma, address);
3109         if (unlikely(!page)) {
3110                 ret = VM_FAULT_OOM;
3111                 page = swapcache;
3112                 goto out_page;
3113         }
3114
3115         if (mem_cgroup_try_charge_swapin(mm, page, GFP_KERNEL, &ptr)) {
3116                 ret = VM_FAULT_OOM;
3117                 goto out_page;
3118         }
3119
3120         /*
3121          * Back out if somebody else already faulted in this pte.
3122          */
3123         page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
3124         if (unlikely(!pte_same(*page_table, orig_pte)))
3125                 goto out_nomap;
3126
3127         if (unlikely(!PageUptodate(page))) {
3128                 ret = VM_FAULT_SIGBUS;
3129                 goto out_nomap;
3130         }
3131
3132         /*
3133          * The page isn't present yet, go ahead with the fault.
3134          *
3135          * Be careful about the sequence of operations here.
3136          * To get its accounting right, reuse_swap_page() must be called
3137          * while the page is counted on swap but not yet in mapcount i.e.
3138          * before page_add_anon_rmap() and swap_free(); try_to_free_swap()
3139          * must be called after the swap_free(), or it will never succeed.
3140          * Because delete_from_swap_page() may be called by reuse_swap_page(),
3141          * mem_cgroup_commit_charge_swapin() may not be able to find swp_entry
3142          * in page->private. In this case, a record in swap_cgroup  is silently
3143          * discarded at swap_free().
3144          */
3145
3146         inc_mm_counter_fast(mm, MM_ANONPAGES);
3147         dec_mm_counter_fast(mm, MM_SWAPENTS);
3148         pte = mk_pte(page, vma->vm_page_prot);
3149         if ((flags & FAULT_FLAG_WRITE) && reuse_swap_page(page)) {
3150                 pte = maybe_mkwrite(pte_mkdirty(pte), vma);
3151                 flags &= ~FAULT_FLAG_WRITE;
3152                 ret |= VM_FAULT_WRITE;
3153                 exclusive = 1;
3154         }
3155         flush_icache_page(vma, page);
3156         set_pte_at(mm, address, page_table, pte);
3157         if (page == swapcache)
3158                 do_page_add_anon_rmap(page, vma, address, exclusive);
3159         else /* ksm created a completely new copy */
3160                 page_add_new_anon_rmap(page, vma, address);
3161         /* It's better to call commit-charge after rmap is established */
3162         mem_cgroup_commit_charge_swapin(page, ptr);
3163
3164         swap_free(entry);
3165
3166 #ifdef CONFIG_ZRAM
3167         if (keep_to_free_swap || vm_swap_full() || (vma->vm_flags & VM_LOCKED) || PageMlocked(page))
3168 #else
3169         if (vm_swap_full() || (vma->vm_flags & VM_LOCKED) || PageMlocked(page))
3170 #endif
3171                 try_to_free_swap(page);
3172         unlock_page(page);
3173         if (page != swapcache) {
3174                 /*
3175                  * Hold the lock to avoid the swap entry to be reused
3176                  * until we take the PT lock for the pte_same() check
3177                  * (to avoid false positives from pte_same). For
3178                  * further safety release the lock after the swap_free
3179                  * so that the swap count won't change under a
3180                  * parallel locked swapcache.
3181                  */
3182                 unlock_page(swapcache);
3183                 page_cache_release(swapcache);
3184         }
3185
3186         if (flags & FAULT_FLAG_WRITE) {
3187                 ret |= do_wp_page(mm, vma, address, page_table, pmd, ptl, pte);
3188                 if (ret & VM_FAULT_ERROR)
3189                         ret &= VM_FAULT_ERROR;
3190                 goto out;
3191         }
3192
3193         /* No need to invalidate - it was non-present before */
3194         update_mmu_cache(vma, address, page_table);
3195 unlock:
3196         pte_unmap_unlock(page_table, ptl);
3197 out:
3198         return ret;
3199 out_nomap:
3200         mem_cgroup_cancel_charge_swapin(ptr);
3201         pte_unmap_unlock(page_table, ptl);
3202 out_page:
3203         unlock_page(page);
3204 out_release:
3205         page_cache_release(page);
3206         if (page != swapcache) {
3207                 unlock_page(swapcache);
3208                 page_cache_release(swapcache);
3209         }
3210         return ret;
3211 }
3212
3213 /*
3214  * This is like a special single-page "expand_{down|up}wards()",
3215  * except we must first make sure that 'address{-|+}PAGE_SIZE'
3216  * doesn't hit another vma.
3217  */
3218 static inline int check_stack_guard_page(struct vm_area_struct *vma, unsigned long address)
3219 {
3220         address &= PAGE_MASK;
3221         if ((vma->vm_flags & VM_GROWSDOWN) && address == vma->vm_start) {
3222                 struct vm_area_struct *prev = vma->vm_prev;
3223
3224                 /*
3225                  * Is there a mapping abutting this one below?
3226                  *
3227                  * That's only ok if it's the same stack mapping
3228                  * that has gotten split..
3229                  */
3230                 if (prev && prev->vm_end == address)
3231                         return prev->vm_flags & VM_GROWSDOWN ? 0 : -ENOMEM;
3232
3233                 return expand_downwards(vma, address - PAGE_SIZE);
3234         }
3235         if ((vma->vm_flags & VM_GROWSUP) && address + PAGE_SIZE == vma->vm_end) {
3236                 struct vm_area_struct *next = vma->vm_next;
3237
3238                 /* As VM_GROWSDOWN but s/below/above/ */
3239                 if (next && next->vm_start == address + PAGE_SIZE)
3240                         return next->vm_flags & VM_GROWSUP ? 0 : -ENOMEM;
3241
3242                 return expand_upwards(vma, address + PAGE_SIZE);
3243         }
3244         return 0;
3245 }
3246
3247 /*
3248  * We enter with non-exclusive mmap_sem (to exclude vma changes,
3249  * but allow concurrent faults), and pte mapped but not yet locked.
3250  * We return with mmap_sem still held, but pte unmapped and unlocked.
3251  */
3252 static int do_anonymous_page(struct mm_struct *mm, struct vm_area_struct *vma,
3253                 unsigned long address, pte_t *page_table, pmd_t *pmd,
3254                 unsigned int flags)
3255 {
3256         struct page *page;
3257         spinlock_t *ptl;
3258         pte_t entry;
3259
3260         pte_unmap(page_table);
3261
3262         /* Check if we need to add a guard page to the stack */
3263         if (check_stack_guard_page(vma, address) < 0)
3264                 return VM_FAULT_SIGBUS;
3265
3266         /* Use the zero-page for reads */
3267         if (!(flags & FAULT_FLAG_WRITE)) {
3268                 entry = pte_mkspecial(pfn_pte(my_zero_pfn(address),
3269                                                 vma->vm_page_prot));
3270                 page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
3271                 if (!pte_none(*page_table))
3272                         goto unlock;
3273                 goto setpte;
3274         }
3275
3276         /* Allocate our own private page. */
3277         if (unlikely(anon_vma_prepare(vma)))
3278                 goto oom;
3279         page = alloc_zeroed_user_highpage_movable(vma, address);
3280         if (!page)
3281                 goto oom;
3282         /*
3283          * The memory barrier inside __SetPageUptodate makes sure that
3284          * preceeding stores to the page contents become visible before
3285          * the set_pte_at() write.
3286          */
3287         __SetPageUptodate(page);
3288
3289         if (mem_cgroup_newpage_charge(page, mm, GFP_KERNEL))
3290                 goto oom_free_page;
3291
3292         entry = mk_pte(page, vma->vm_page_prot);
3293         if (vma->vm_flags & VM_WRITE)
3294                 entry = pte_mkwrite(pte_mkdirty(entry));
3295
3296         page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
3297         if (!pte_none(*page_table))
3298                 goto release;
3299
3300         inc_mm_counter_fast(mm, MM_ANONPAGES);
3301         page_add_new_anon_rmap(page, vma, address);
3302 setpte:
3303         set_pte_at(mm, address, page_table, entry);
3304
3305         /* No need to invalidate - it was non-present before */
3306         update_mmu_cache(vma, address, page_table);
3307 unlock:
3308         pte_unmap_unlock(page_table, ptl);
3309         return 0;
3310 release:
3311         mem_cgroup_uncharge_page(page);
3312         page_cache_release(page);
3313         goto unlock;
3314 oom_free_page:
3315         page_cache_release(page);
3316 oom:
3317         return VM_FAULT_OOM;
3318 }
3319
3320 /*
3321  * __do_fault() tries to create a new page mapping. It aggressively
3322  * tries to share with existing pages, but makes a separate copy if
3323  * the FAULT_FLAG_WRITE is set in the flags parameter in order to avoid
3324  * the next page fault.
3325  *
3326  * As this is called only for pages that do not currently exist, we
3327  * do not need to flush old virtual caches or the TLB.
3328  *
3329  * We enter with non-exclusive mmap_sem (to exclude vma changes,
3330  * but allow concurrent faults), and pte neither mapped nor locked.
3331  * We return with mmap_sem still held, but pte unmapped and unlocked.
3332  */
3333 static int __do_fault(struct mm_struct *mm, struct vm_area_struct *vma,
3334                 unsigned long address, pmd_t *pmd,
3335                 pgoff_t pgoff, unsigned int flags, pte_t orig_pte)
3336 {
3337         pte_t *page_table;
3338         spinlock_t *ptl;
3339         struct page *page;
3340         struct page *cow_page;
3341         pte_t entry;
3342         int anon = 0;
3343         struct page *dirty_page = NULL;
3344         struct vm_fault vmf;
3345         int ret;
3346         int page_mkwrite = 0;
3347
3348         /*
3349          * If we do COW later, allocate page befor taking lock_page()
3350          * on the file cache page. This will reduce lock holding time.
3351          */
3352         if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) {
3353
3354                 if (unlikely(anon_vma_prepare(vma)))
3355                         return VM_FAULT_OOM;
3356
3357                 cow_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, address);
3358                 if (!cow_page)
3359                         return VM_FAULT_OOM;
3360
3361                 if (mem_cgroup_newpage_charge(cow_page, mm, GFP_KERNEL)) {
3362                         page_cache_release(cow_page);
3363                         return VM_FAULT_OOM;
3364                 }
3365         } else
3366                 cow_page = NULL;
3367
3368         vmf.virtual_address = (void __user *)(address & PAGE_MASK);
3369         vmf.pgoff = pgoff;
3370         vmf.flags = flags;
3371         vmf.page = NULL;
3372
3373         ret = vma->vm_ops->fault(vma, &vmf);
3374         if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE |
3375                             VM_FAULT_RETRY)))
3376                 goto uncharge_out;
3377
3378         if (unlikely(PageHWPoison(vmf.page))) {
3379                 if (ret & VM_FAULT_LOCKED)
3380                         unlock_page(vmf.page);
3381                 ret = VM_FAULT_HWPOISON;
3382                 goto uncharge_out;
3383         }
3384
3385         /*
3386          * For consistency in subsequent calls, make the faulted page always
3387          * locked.
3388          */
3389         if (unlikely(!(ret & VM_FAULT_LOCKED)))
3390                 lock_page(vmf.page);
3391         else
3392                 VM_BUG_ON(!PageLocked(vmf.page));
3393
3394         /*
3395          * Should we do an early C-O-W break?
3396          */
3397         page = vmf.page;
3398         if (flags & FAULT_FLAG_WRITE) {
3399                 if (!(vma->vm_flags & VM_SHARED)) {
3400                         page = cow_page;
3401                         anon = 1;
3402                         copy_user_highpage(page, vmf.page, address, vma);
3403                         __SetPageUptodate(page);
3404                 } else {
3405                         /*
3406                          * If the page will be shareable, see if the backing
3407                          * address space wants to know that the page is about
3408                          * to become writable
3409                          */
3410                         if (vma->vm_ops->page_mkwrite) {
3411                                 int tmp;
3412
3413                                 unlock_page(page);
3414                                 vmf.flags = FAULT_FLAG_WRITE|FAULT_FLAG_MKWRITE;
3415                                 tmp = vma->vm_ops->page_mkwrite(vma, &vmf);
3416                                 if (unlikely(tmp &
3417                                           (VM_FAULT_ERROR | VM_FAULT_NOPAGE))) {
3418                                         ret = tmp;
3419                                         goto unwritable_page;
3420                                 }
3421                                 if (unlikely(!(tmp & VM_FAULT_LOCKED))) {
3422                                         lock_page(page);
3423                                         if (!page->mapping) {
3424                                                 ret = 0; /* retry the fault */
3425                                                 unlock_page(page);
3426                                                 goto unwritable_page;
3427                                         }
3428                                 } else
3429                                         VM_BUG_ON(!PageLocked(page));
3430                                 page_mkwrite = 1;
3431                         }
3432                 }
3433
3434         }
3435
3436         page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
3437
3438         /*
3439          * This silly early PAGE_DIRTY setting removes a race
3440          * due to the bad i386 page protection. But it's valid
3441          * for other architectures too.
3442          *
3443          * Note that if FAULT_FLAG_WRITE is set, we either now have
3444          * an exclusive copy of the page, or this is a shared mapping,
3445          * so we can make it writable and dirty to avoid having to
3446          * handle that later.
3447          */
3448         /* Only go through if we didn't race with anybody else... */
3449         if (likely(pte_same(*page_table, orig_pte))) {
3450                 flush_icache_page(vma, page);
3451                 entry = mk_pte(page, vma->vm_page_prot);
3452                 if (flags & FAULT_FLAG_WRITE)
3453                         entry = maybe_mkwrite(pte_mkdirty(entry), vma);
3454                 if (anon) {
3455                         inc_mm_counter_fast(mm, MM_ANONPAGES);
3456                         page_add_new_anon_rmap(page, vma, address);
3457                 } else {
3458                         inc_mm_counter_fast(mm, MM_FILEPAGES);
3459                         page_add_file_rmap(page);
3460                         if (flags & FAULT_FLAG_WRITE) {
3461                                 dirty_page = page;
3462                                 get_page(dirty_page);
3463                         }
3464                 }
3465                 set_pte_at(mm, address, page_table, entry);
3466
3467                 /* no need to invalidate: a not-present page won't be cached */
3468                 update_mmu_cache(vma, address, page_table);
3469         } else {
3470                 if (cow_page)
3471                         mem_cgroup_uncharge_page(cow_page);
3472                 if (anon)
3473                         page_cache_release(page);
3474                 else
3475                         anon = 1; /* no anon but release faulted_page */
3476         }
3477
3478         pte_unmap_unlock(page_table, ptl);
3479
3480         if (dirty_page) {
3481                 struct address_space *mapping = page->mapping;
3482                 int dirtied = 0;
3483
3484                 if (set_page_dirty(dirty_page))
3485                         dirtied = 1;
3486                 unlock_page(dirty_page);
3487                 put_page(dirty_page);
3488                 if ((dirtied || page_mkwrite) && mapping) {
3489                         /*
3490                          * Some device drivers do not set page.mapping but still
3491                          * dirty their pages
3492                          */
3493                         balance_dirty_pages_ratelimited(mapping);
3494                 }
3495
3496                 /* file_update_time outside page_lock */
3497                 if (vma->vm_file && !page_mkwrite)
3498                         file_update_time(vma->vm_file);
3499         } else {
3500                 unlock_page(vmf.page);
3501                 if (anon)
3502                         page_cache_release(vmf.page);
3503         }
3504
3505         return ret;
3506
3507 unwritable_page:
3508         page_cache_release(page);
3509         return ret;
3510 uncharge_out:
3511         /* fs's fault handler get error */
3512         if (cow_page) {
3513                 mem_cgroup_uncharge_page(cow_page);
3514                 page_cache_release(cow_page);
3515         }
3516         return ret;
3517 }
3518
3519 static int do_linear_fault(struct mm_struct *mm, struct vm_area_struct *vma,
3520                 unsigned long address, pte_t *page_table, pmd_t *pmd,
3521                 unsigned int flags, pte_t orig_pte)
3522 {
3523         pgoff_t pgoff = (((address & PAGE_MASK)
3524                         - vma->vm_start) >> PAGE_SHIFT) + vma->vm_pgoff;
3525
3526         pte_unmap(page_table);
3527         return __do_fault(mm, vma, address, pmd, pgoff, flags, orig_pte);
3528 }
3529
3530 /*
3531  * Fault of a previously existing named mapping. Repopulate the pte
3532  * from the encoded file_pte if possible. This enables swappable
3533  * nonlinear vmas.
3534  *
3535  * We enter with non-exclusive mmap_sem (to exclude vma changes,
3536  * but allow concurrent faults), and pte mapped but not yet locked.
3537  * We return with mmap_sem still held, but pte unmapped and unlocked.
3538  */
3539 static int do_nonlinear_fault(struct mm_struct *mm, struct vm_area_struct *vma,
3540                 unsigned long address, pte_t *page_table, pmd_t *pmd,
3541                 unsigned int flags, pte_t orig_pte)
3542 {
3543         pgoff_t pgoff;
3544
3545         flags |= FAULT_FLAG_NONLINEAR;
3546
3547         if (!pte_unmap_same(mm, pmd, page_table, orig_pte))
3548                 return 0;
3549
3550         if (unlikely(!(vma->vm_flags & VM_NONLINEAR))) {
3551                 /*
3552                  * Page table corrupted: show pte and kill process.
3553                  */
3554                 print_bad_pte(vma, address, orig_pte, NULL);
3555                 return VM_FAULT_SIGBUS;
3556         }
3557
3558         pgoff = pte_to_pgoff(orig_pte);
3559         return __do_fault(mm, vma, address, pmd, pgoff, flags, orig_pte);
3560 }
3561
3562 int numa_migrate_prep(struct page *page, struct vm_area_struct *vma,
3563                                 unsigned long addr, int page_nid)
3564 {
3565         get_page(page);
3566
3567         count_vm_numa_event(NUMA_HINT_FAULTS);
3568         if (page_nid == numa_node_id())
3569                 count_vm_numa_event(NUMA_HINT_FAULTS_LOCAL);
3570
3571         return mpol_misplaced(page, vma, addr);
3572 }
3573
3574 int do_numa_page(struct mm_struct *mm, struct vm_area_struct *vma,
3575                    unsigned long addr, pte_t pte, pte_t *ptep, pmd_t *pmd)
3576 {
3577         struct page *page = NULL;
3578         spinlock_t *ptl;
3579         int page_nid = -1;
3580         int target_nid;
3581         bool migrated = false;
3582
3583         /*
3584         * The "pte" at this point cannot be used safely without
3585         * validation through pte_unmap_same(). It's of NUMA type but
3586         * the pfn may be screwed if the read is non atomic.
3587         *
3588         * ptep_modify_prot_start is not called as this is clearing
3589         * the _PAGE_NUMA bit and it is not really expected that there
3590         * would be concurrent hardware modifications to the PTE.
3591         */
3592         ptl = pte_lockptr(mm, pmd);
3593         spin_lock(ptl);
3594         if (unlikely(!pte_same(*ptep, pte))) {
3595                 pte_unmap_unlock(ptep, ptl);
3596                 goto out;
3597         }
3598
3599         pte = pte_mknonnuma(pte);
3600         set_pte_at(mm, addr, ptep, pte);
3601         update_mmu_cache(vma, addr, ptep);
3602
3603         page = vm_normal_page(vma, addr, pte);
3604         if (!page) {
3605                 pte_unmap_unlock(ptep, ptl);
3606                 return 0;
3607         }
3608
3609         page_nid = page_to_nid(page);
3610         target_nid = numa_migrate_prep(page, vma, addr, page_nid);
3611         pte_unmap_unlock(ptep, ptl);
3612         if (target_nid == -1) {
3613                 put_page(page);
3614                 goto out;
3615         }
3616
3617         /* Migrate to the requested node */
3618         migrated = migrate_misplaced_page(page, target_nid);
3619         if (migrated)
3620                 page_nid = target_nid;
3621
3622 out:
3623         if (page_nid != -1)
3624                 task_numa_fault(page_nid, 1, migrated);
3625         return 0;
3626 }
3627
3628 /* NUMA hinting page fault entry point for regular pmds */
3629 #ifdef CONFIG_NUMA_BALANCING
3630 static int do_pmd_numa_page(struct mm_struct *mm, struct vm_area_struct *vma,
3631                      unsigned long addr, pmd_t *pmdp)
3632 {
3633         pmd_t pmd;
3634         pte_t *pte, *orig_pte;
3635         unsigned long _addr = addr & PMD_MASK;
3636         unsigned long offset;
3637         spinlock_t *ptl;
3638         bool numa = false;
3639
3640         spin_lock(&mm->page_table_lock);
3641         pmd = *pmdp;
3642         if (pmd_numa(pmd)) {
3643                 set_pmd_at(mm, _addr, pmdp, pmd_mknonnuma(pmd));
3644                 numa = true;
3645         }
3646         spin_unlock(&mm->page_table_lock);
3647
3648         if (!numa)
3649                 return 0;
3650
3651         /* we're in a page fault so some vma must be in the range */
3652         BUG_ON(!vma);
3653         BUG_ON(vma->vm_start >= _addr + PMD_SIZE);
3654         offset = max(_addr, vma->vm_start) & ~PMD_MASK;
3655         VM_BUG_ON(offset >= PMD_SIZE);
3656         orig_pte = pte = pte_offset_map_lock(mm, pmdp, _addr, &ptl);
3657         pte += offset >> PAGE_SHIFT;
3658         for (addr = _addr + offset; addr < _addr + PMD_SIZE; pte++, addr += PAGE_SIZE) {
3659                 pte_t pteval = *pte;
3660                 struct page *page;
3661                 int page_nid = -1;
3662                 int target_nid;
3663                 bool migrated = false;
3664
3665                 if (!pte_present(pteval))
3666                         continue;
3667                 if (!pte_numa(pteval))
3668                         continue;
3669                 if (addr >= vma->vm_end) {
3670                         vma = find_vma(mm, addr);
3671                         /* there's a pte present so there must be a vma */
3672                         BUG_ON(!vma);
3673                         BUG_ON(addr < vma->vm_start);
3674                 }
3675                 if (pte_numa(pteval)) {
3676                         pteval = pte_mknonnuma(pteval);
3677                         set_pte_at(mm, addr, pte, pteval);
3678                 }
3679                 page = vm_normal_page(vma, addr, pteval);
3680                 if (unlikely(!page))
3681                         continue;
3682                 /* only check non-shared pages */
3683                 if (unlikely(page_mapcount(page) != 1))
3684                         continue;
3685
3686                 page_nid = page_to_nid(page);
3687                 target_nid = numa_migrate_prep(page, vma, addr, page_nid);
3688                 pte_unmap_unlock(pte, ptl);
3689                 if (target_nid != -1) {
3690                         migrated = migrate_misplaced_page(page, target_nid);
3691                         if (migrated)
3692                                 page_nid = target_nid;
3693                 } else {
3694                         put_page(page);
3695                 }
3696
3697                 if (page_nid != -1)
3698                         task_numa_fault(page_nid, 1, migrated);
3699
3700                 pte = pte_offset_map_lock(mm, pmdp, addr, &ptl);
3701         }
3702         pte_unmap_unlock(orig_pte, ptl);
3703
3704         return 0;
3705 }
3706 #else
3707 static int do_pmd_numa_page(struct mm_struct *mm, struct vm_area_struct *vma,
3708                      unsigned long addr, pmd_t *pmdp)
3709 {
3710         BUG();
3711         return 0;
3712 }
3713 #endif /* CONFIG_NUMA_BALANCING */
3714
3715 /*
3716  * These routines also need to handle stuff like marking pages dirty
3717  * and/or accessed for architectures that don't do it in hardware (most
3718  * RISC architectures).  The early dirtying is also good on the i386.
3719  *
3720  * There is also a hook called "update_mmu_cache()" that architectures
3721  * with external mmu caches can use to update those (ie the Sparc or
3722  * PowerPC hashed page tables that act as extended TLBs).
3723  *
3724  * We enter with non-exclusive mmap_sem (to exclude vma changes,
3725  * but allow concurrent faults), and pte mapped but not yet locked.
3726  * We return with mmap_sem still held, but pte unmapped and unlocked.
3727  */
3728 int handle_pte_fault(struct mm_struct *mm,
3729                      struct vm_area_struct *vma, unsigned long address,
3730                      pte_t *pte, pmd_t *pmd, unsigned int flags)
3731 {
3732         pte_t entry;
3733         spinlock_t *ptl;
3734
3735         entry = *pte;
3736         if (!pte_present(entry)) {
3737                 if (pte_none(entry)) {
3738                         if (vma->vm_ops) {
3739                                 if (likely(vma->vm_ops->fault))
3740                                         return do_linear_fault(mm, vma, address,
3741                                                 pte, pmd, flags, entry);
3742                         }
3743                         return do_anonymous_page(mm, vma, address,
3744                                                  pte, pmd, flags);
3745                 }
3746                 if (pte_file(entry))
3747                         return do_nonlinear_fault(mm, vma, address,
3748                                         pte, pmd, flags, entry);
3749                 return do_swap_page(mm, vma, address,
3750                                         pte, pmd, flags, entry);
3751         }
3752
3753         if (pte_numa(entry))
3754                 return do_numa_page(mm, vma, address, entry, pte, pmd);
3755
3756         ptl = pte_lockptr(mm, pmd);
3757         spin_lock(ptl);
3758         if (unlikely(!pte_same(*pte, entry)))
3759                 goto unlock;
3760         if (flags & FAULT_FLAG_WRITE) {
3761                 if (!pte_write(entry))
3762                         return do_wp_page(mm, vma, address,
3763                                         pte, pmd, ptl, entry);
3764                 entry = pte_mkdirty(entry);
3765         }
3766         entry = pte_mkyoung(entry);
3767         if (ptep_set_access_flags(vma, address, pte, entry, flags & FAULT_FLAG_WRITE)) {
3768                 update_mmu_cache(vma, address, pte);
3769         } else {
3770                 /*
3771                  * This is needed only for protection faults but the arch code
3772                  * is not yet telling us if this is a protection fault or not.
3773                  * This still avoids useless tlb flushes for .text page faults
3774                  * with threads.
3775                  */
3776                 if (flags & FAULT_FLAG_WRITE)
3777                         flush_tlb_fix_spurious_fault(vma, address);
3778         }
3779 unlock:
3780         pte_unmap_unlock(pte, ptl);
3781         return 0;
3782 }
3783
3784 /*
3785  * By the time we get here, we already hold the mm semaphore
3786  */
3787 static int __handle_mm_fault(struct mm_struct *mm, struct vm_area_struct *vma,
3788                              unsigned long address, unsigned int flags)
3789 {
3790         pgd_t *pgd;
3791         pud_t *pud;
3792         pmd_t *pmd;
3793         pte_t *pte;
3794
3795         if (unlikely(is_vm_hugetlb_page(vma)))
3796                 return hugetlb_fault(mm, vma, address, flags);
3797
3798 retry:
3799         pgd = pgd_offset(mm, address);
3800         pud = pud_alloc(mm, pgd, address);
3801         if (!pud)
3802                 return VM_FAULT_OOM;
3803         pmd = pmd_alloc(mm, pud, address);
3804         if (!pmd)
3805                 return VM_FAULT_OOM;
3806         if (pmd_none(*pmd) && transparent_hugepage_enabled(vma)) {
3807                 if (!vma->vm_ops)
3808                         return do_huge_pmd_anonymous_page(mm, vma, address,
3809                                                           pmd, flags);
3810         } else {
3811                 pmd_t orig_pmd = *pmd;
3812                 int ret;
3813
3814                 barrier();
3815                 if (pmd_trans_huge(orig_pmd)) {
3816                         unsigned int dirty = flags & FAULT_FLAG_WRITE;
3817
3818                         /*
3819                          * If the pmd is splitting, return and retry the
3820                          * the fault.  Alternative: wait until the split
3821                          * is done, and goto retry.
3822                          */
3823                         if (pmd_trans_splitting(orig_pmd))
3824                                 return 0;
3825
3826                         if (pmd_numa(orig_pmd))
3827                                 return do_huge_pmd_numa_page(mm, vma, address,
3828                                                              orig_pmd, pmd);
3829
3830                         if (dirty && !pmd_write(orig_pmd)) {
3831                                 ret = do_huge_pmd_wp_page(mm, vma, address, pmd,
3832                                                           orig_pmd);
3833                                 /*
3834                                  * If COW results in an oom, the huge pmd will
3835                                  * have been split, so retry the fault on the
3836                                  * pte for a smaller charge.
3837                                  */
3838                                 if (unlikely(ret & VM_FAULT_OOM))
3839                                         goto retry;
3840                                 return ret;
3841                         } else {
3842                                 huge_pmd_set_accessed(mm, vma, address, pmd,
3843                                                       orig_pmd, dirty);
3844                         }
3845
3846                         return 0;
3847                 }
3848         }
3849
3850         if (pmd_numa(*pmd))
3851                 return do_pmd_numa_page(mm, vma, address, pmd);
3852
3853         /*
3854          * Use __pte_alloc instead of pte_alloc_map, because we can't
3855          * run pte_offset_map on the pmd, if an huge pmd could
3856          * materialize from under us from a different thread.
3857          */
3858         if (unlikely(pmd_none(*pmd)) &&
3859             unlikely(__pte_alloc(mm, vma, pmd, address)))
3860                 return VM_FAULT_OOM;
3861         /* if an huge pmd materialized from under us just retry later */
3862         if (unlikely(pmd_trans_huge(*pmd)))
3863                 return 0;
3864         /*
3865          * A regular pmd is established and it can't morph into a huge pmd
3866          * from under us anymore at this point because we hold the mmap_sem
3867          * read mode and khugepaged takes it in write mode. So now it's
3868          * safe to run pte_offset_map().
3869          */
3870         pte = pte_offset_map(pmd, address);
3871
3872         return handle_pte_fault(mm, vma, address, pte, pmd, flags);
3873 }
3874
3875 int handle_mm_fault(struct mm_struct *mm, struct vm_area_struct *vma,
3876                     unsigned long address, unsigned int flags)
3877 {
3878         int ret;
3879
3880         __set_current_state(TASK_RUNNING);
3881
3882         count_vm_event(PGFAULT);
3883         mem_cgroup_count_vm_event(mm, PGFAULT);
3884
3885         /* do counter updates before entering really critical section. */
3886         check_sync_rss_stat(current);
3887
3888         /*
3889          * Enable the memcg OOM handling for faults triggered in user
3890          * space.  Kernel faults are handled more gracefully.
3891          */
3892         if (flags & FAULT_FLAG_USER)
3893                 mem_cgroup_oom_enable();
3894
3895         ret = __handle_mm_fault(mm, vma, address, flags);
3896
3897         if (flags & FAULT_FLAG_USER) {
3898                 mem_cgroup_oom_disable();
3899                 /*
3900                  * The task may have entered a memcg OOM situation but
3901                  * if the allocation error was handled gracefully (no
3902                  * VM_FAULT_OOM), there is no need to kill anything.
3903                  * Just clean up the OOM state peacefully.
3904                  */
3905                 if (task_in_memcg_oom(current) && !(ret & VM_FAULT_OOM))
3906                         mem_cgroup_oom_synchronize(false);
3907         }
3908
3909         return ret;
3910 }
3911
3912 #ifndef __PAGETABLE_PUD_FOLDED
3913 /*
3914  * Allocate page upper directory.
3915  * We've already handled the fast-path in-line.
3916  */
3917 int __pud_alloc(struct mm_struct *mm, pgd_t *pgd, unsigned long address)
3918 {
3919         pud_t *new = pud_alloc_one(mm, address);
3920         if (!new)
3921                 return -ENOMEM;
3922
3923         smp_wmb(); /* See comment in __pte_alloc */
3924
3925         spin_lock(&mm->page_table_lock);
3926         if (pgd_present(*pgd))          /* Another has populated it */
3927                 pud_free(mm, new);
3928         else
3929                 pgd_populate(mm, pgd, new);
3930         spin_unlock(&mm->page_table_lock);
3931         return 0;
3932 }
3933 #endif /* __PAGETABLE_PUD_FOLDED */
3934
3935 #ifndef __PAGETABLE_PMD_FOLDED
3936 /*
3937  * Allocate page middle directory.
3938  * We've already handled the fast-path in-line.
3939  */
3940 int __pmd_alloc(struct mm_struct *mm, pud_t *pud, unsigned long address)
3941 {
3942         pmd_t *new = pmd_alloc_one(mm, address);
3943         if (!new)
3944                 return -ENOMEM;
3945
3946         smp_wmb(); /* See comment in __pte_alloc */
3947
3948         spin_lock(&mm->page_table_lock);
3949 #ifndef __ARCH_HAS_4LEVEL_HACK
3950         if (pud_present(*pud))          /* Another has populated it */
3951                 pmd_free(mm, new);
3952         else
3953                 pud_populate(mm, pud, new);
3954 #else
3955         if (pgd_present(*pud))          /* Another has populated it */
3956                 pmd_free(mm, new);
3957         else
3958                 pgd_populate(mm, pud, new);
3959 #endif /* __ARCH_HAS_4LEVEL_HACK */
3960         spin_unlock(&mm->page_table_lock);
3961         return 0;
3962 }
3963 #endif /* __PAGETABLE_PMD_FOLDED */
3964
3965 #if !defined(__HAVE_ARCH_GATE_AREA)
3966
3967 #if defined(AT_SYSINFO_EHDR)
3968 static struct vm_area_struct gate_vma;
3969
3970 static int __init gate_vma_init(void)
3971 {
3972         gate_vma.vm_mm = NULL;
3973         gate_vma.vm_start = FIXADDR_USER_START;
3974         gate_vma.vm_end = FIXADDR_USER_END;
3975         gate_vma.vm_flags = VM_READ | VM_MAYREAD | VM_EXEC | VM_MAYEXEC;
3976         gate_vma.vm_page_prot = __P101;
3977
3978         return 0;
3979 }
3980 __initcall(gate_vma_init);
3981 #endif
3982
3983 struct vm_area_struct *get_gate_vma(struct mm_struct *mm)
3984 {
3985 #ifdef AT_SYSINFO_EHDR
3986         return &gate_vma;
3987 #else
3988         return NULL;
3989 #endif
3990 }
3991
3992 int in_gate_area_no_mm(unsigned long addr)
3993 {
3994 #ifdef AT_SYSINFO_EHDR
3995         if ((addr >= FIXADDR_USER_START) && (addr < FIXADDR_USER_END))
3996                 return 1;
3997 #endif
3998         return 0;
3999 }
4000
4001 #endif  /* __HAVE_ARCH_GATE_AREA */
4002
4003 static int __follow_pte(struct mm_struct *mm, unsigned long address,
4004                 pte_t **ptepp, spinlock_t **ptlp)
4005 {
4006         pgd_t *pgd;
4007         pud_t *pud;
4008         pmd_t *pmd;
4009         pte_t *ptep;
4010
4011         pgd = pgd_offset(mm, address);
4012         if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
4013                 goto out;
4014
4015         pud = pud_offset(pgd, address);
4016         if (pud_none(*pud) || unlikely(pud_bad(*pud)))
4017                 goto out;
4018
4019         pmd = pmd_offset(pud, address);
4020         VM_BUG_ON(pmd_trans_huge(*pmd));
4021         if (pmd_none(*pmd) || unlikely(pmd_bad(*pmd)))
4022                 goto out;
4023
4024         /* We cannot handle huge page PFN maps. Luckily they don't exist. */
4025         if (pmd_huge(*pmd))
4026                 goto out;
4027
4028         ptep = pte_offset_map_lock(mm, pmd, address, ptlp);
4029         if (!ptep)
4030                 goto out;
4031         if (!pte_present(*ptep))
4032                 goto unlock;
4033         *ptepp = ptep;
4034         return 0;
4035 unlock:
4036         pte_unmap_unlock(ptep, *ptlp);
4037 out:
4038         return -EINVAL;
4039 }
4040
4041 static inline int follow_pte(struct mm_struct *mm, unsigned long address,
4042                              pte_t **ptepp, spinlock_t **ptlp)
4043 {
4044         int res;
4045
4046         /* (void) is needed to make gcc happy */
4047         (void) __cond_lock(*ptlp,
4048                            !(res = __follow_pte(mm, address, ptepp, ptlp)));
4049         return res;
4050 }
4051
4052 /**
4053  * follow_pfn - look up PFN at a user virtual address
4054  * @vma: memory mapping
4055  * @address: user virtual address
4056  * @pfn: location to store found PFN
4057  *
4058  * Only IO mappings and raw PFN mappings are allowed.
4059  *
4060  * Returns zero and the pfn at @pfn on success, -ve otherwise.
4061  */
4062 int follow_pfn(struct vm_area_struct *vma, unsigned long address,
4063         unsigned long *pfn)
4064 {
4065         int ret = -EINVAL;
4066         spinlock_t *ptl;
4067         pte_t *ptep;
4068
4069         if (!(vma->vm_flags & (VM_IO | VM_PFNMAP)))
4070                 return ret;
4071
4072         ret = follow_pte(vma->vm_mm, address, &ptep, &ptl);
4073         if (ret)
4074                 return ret;
4075         *pfn = pte_pfn(*ptep);
4076         pte_unmap_unlock(ptep, ptl);
4077         return 0;
4078 }
4079 EXPORT_SYMBOL(follow_pfn);
4080
4081 #ifdef CONFIG_HAVE_IOREMAP_PROT
4082 int follow_phys(struct vm_area_struct *vma,
4083                 unsigned long address, unsigned int flags,
4084                 unsigned long *prot, resource_size_t *phys)
4085 {
4086         int ret = -EINVAL;
4087         pte_t *ptep, pte;
4088         spinlock_t *ptl;
4089
4090         if (!(vma->vm_flags & (VM_IO | VM_PFNMAP)))
4091                 goto out;
4092
4093         if (follow_pte(vma->vm_mm, address, &ptep, &ptl))
4094                 goto out;
4095         pte = *ptep;
4096
4097         if ((flags & FOLL_WRITE) && !pte_write(pte))
4098                 goto unlock;
4099
4100         *prot = pgprot_val(pte_pgprot(pte));
4101         *phys = (resource_size_t)pte_pfn(pte) << PAGE_SHIFT;
4102
4103         ret = 0;
4104 unlock:
4105         pte_unmap_unlock(ptep, ptl);
4106 out:
4107         return ret;
4108 }
4109
4110 int generic_access_phys(struct vm_area_struct *vma, unsigned long addr,
4111                         void *buf, int len, int write)
4112 {
4113         resource_size_t phys_addr;
4114         unsigned long prot = 0;
4115         void __iomem *maddr;
4116         int offset = addr & (PAGE_SIZE-1);
4117
4118         if (follow_phys(vma, addr, write, &prot, &phys_addr))
4119                 return -EINVAL;
4120
4121         maddr = ioremap_prot(phys_addr, PAGE_ALIGN(len + offset), prot);
4122         if (write)
4123                 memcpy_toio(maddr + offset, buf, len);
4124         else
4125                 memcpy_fromio(buf, maddr + offset, len);
4126         iounmap(maddr);
4127
4128         return len;
4129 }
4130 EXPORT_SYMBOL_GPL(generic_access_phys);
4131 #endif
4132
4133 /*
4134  * Access another process' address space as given in mm.  If non-NULL, use the
4135  * given task for page fault accounting.
4136  */
4137 static int __access_remote_vm(struct task_struct *tsk, struct mm_struct *mm,
4138                 unsigned long addr, void *buf, int len, int write)
4139 {
4140         struct vm_area_struct *vma;
4141         void *old_buf = buf;
4142
4143         down_read(&mm->mmap_sem);
4144         /* ignore errors, just check how much was successfully transferred */
4145         while (len) {
4146                 int bytes, ret, offset;
4147                 void *maddr;
4148                 struct page *page = NULL;
4149
4150                 ret = get_user_pages(tsk, mm, addr, 1,
4151                                 write, 1, &page, &vma);
4152                 if (ret <= 0) {
4153                         /*
4154                          * Check if this is a VM_IO | VM_PFNMAP VMA, which
4155                          * we can access using slightly different code.
4156                          */
4157 #ifdef CONFIG_HAVE_IOREMAP_PROT
4158                         vma = find_vma(mm, addr);
4159                         if (!vma || vma->vm_start > addr)
4160                                 break;
4161                         if (vma->vm_ops && vma->vm_ops->access)
4162                                 ret = vma->vm_ops->access(vma, addr, buf,
4163                                                           len, write);
4164                         if (ret <= 0)
4165 #endif
4166                                 break;
4167                         bytes = ret;
4168                 } else {
4169                         bytes = len;
4170                         offset = addr & (PAGE_SIZE-1);
4171                         if (bytes > PAGE_SIZE-offset)
4172                                 bytes = PAGE_SIZE-offset;
4173
4174                         maddr = kmap(page);
4175                         if (write) {
4176                                 copy_to_user_page(vma, page, addr,
4177                                                   maddr + offset, buf, bytes);
4178                                 set_page_dirty_lock(page);
4179                         } else {
4180                                 copy_from_user_page(vma, page, addr,
4181                                                     buf, maddr + offset, bytes);
4182                         }
4183                         kunmap(page);
4184                         page_cache_release(page);
4185                 }
4186                 len -= bytes;
4187                 buf += bytes;
4188                 addr += bytes;
4189         }
4190         up_read(&mm->mmap_sem);
4191
4192         return buf - old_buf;
4193 }
4194
4195 /**
4196  * access_remote_vm - access another process' address space
4197  * @mm:         the mm_struct of the target address space
4198  * @addr:       start address to access
4199  * @buf:        source or destination buffer
4200  * @len:        number of bytes to transfer
4201  * @write:      whether the access is a write
4202  *
4203  * The caller must hold a reference on @mm.
4204  */
4205 int access_remote_vm(struct mm_struct *mm, unsigned long addr,
4206                 void *buf, int len, int write)
4207 {
4208         return __access_remote_vm(NULL, mm, addr, buf, len, write);
4209 }
4210
4211 /*
4212  * Access another process' address space.
4213  * Source/target buffer must be kernel space,
4214  * Do not walk the page table directly, use get_user_pages
4215  */
4216 int access_process_vm(struct task_struct *tsk, unsigned long addr,
4217                 void *buf, int len, int write)
4218 {
4219         struct mm_struct *mm;
4220         int ret;
4221
4222         mm = get_task_mm(tsk);
4223         if (!mm)
4224                 return 0;
4225
4226         ret = __access_remote_vm(tsk, mm, addr, buf, len, write);
4227         mmput(mm);
4228
4229         return ret;
4230 }
4231
4232 /*
4233  * Print the name of a VMA.
4234  */
4235 void print_vma_addr(char *prefix, unsigned long ip)
4236 {
4237         struct mm_struct *mm = current->mm;
4238         struct vm_area_struct *vma;
4239
4240         /*
4241          * Do not print if we are in atomic
4242          * contexts (in exception stacks, etc.):
4243          */
4244         if (preempt_count())
4245                 return;
4246
4247         down_read(&mm->mmap_sem);
4248         vma = find_vma(mm, ip);
4249         if (vma && vma->vm_file) {
4250                 struct file *f = vma->vm_file;
4251                 char *buf = (char *)__get_free_page(GFP_KERNEL);
4252                 if (buf) {
4253                         char *p;
4254
4255                         p = d_path(&f->f_path, buf, PAGE_SIZE);
4256                         if (IS_ERR(p))
4257                                 p = "?";
4258                         printk("%s%s[%lx+%lx]", prefix, kbasename(p),
4259                                         vma->vm_start,
4260                                         vma->vm_end - vma->vm_start);
4261                         free_page((unsigned long)buf);
4262                 }
4263         }
4264         up_read(&mm->mmap_sem);
4265 }
4266
4267 #ifdef CONFIG_PROVE_LOCKING
4268 void might_fault(void)
4269 {
4270         /*
4271          * Some code (nfs/sunrpc) uses socket ops on kernel memory while
4272          * holding the mmap_sem, this is safe because kernel memory doesn't
4273          * get paged out, therefore we'll never actually fault, and the
4274          * below annotations will generate false positives.
4275          */
4276         if (segment_eq(get_fs(), KERNEL_DS))
4277                 return;
4278
4279         might_sleep();
4280         /*
4281          * it would be nicer only to annotate paths which are not under
4282          * pagefault_disable, however that requires a larger audit and
4283          * providing helpers like get_user_atomic.
4284          */
4285         if (!in_atomic() && current->mm)
4286                 might_lock_read(&current->mm->mmap_sem);
4287 }
4288 EXPORT_SYMBOL(might_fault);
4289 #endif
4290
4291 #if defined(CONFIG_TRANSPARENT_HUGEPAGE) || defined(CONFIG_HUGETLBFS)
4292 static void clear_gigantic_page(struct page *page,
4293                                 unsigned long addr,
4294                                 unsigned int pages_per_huge_page)
4295 {
4296         int i;
4297         struct page *p = page;
4298
4299         might_sleep();
4300         for (i = 0; i < pages_per_huge_page;
4301              i++, p = mem_map_next(p, page, i)) {
4302                 cond_resched();
4303                 clear_user_highpage(p, addr + i * PAGE_SIZE);
4304         }
4305 }
4306 void clear_huge_page(struct page *page,
4307                      unsigned long addr, unsigned int pages_per_huge_page)
4308 {
4309         int i;
4310
4311         if (unlikely(pages_per_huge_page > MAX_ORDER_NR_PAGES)) {
4312                 clear_gigantic_page(page, addr, pages_per_huge_page);
4313                 return;
4314         }
4315
4316         might_sleep();
4317         for (i = 0; i < pages_per_huge_page; i++) {
4318                 cond_resched();
4319                 clear_user_highpage(page + i, addr + i * PAGE_SIZE);
4320         }
4321 }
4322
4323 static void copy_user_gigantic_page(struct page *dst, struct page *src,
4324                                     unsigned long addr,
4325                                     struct vm_area_struct *vma,
4326                                     unsigned int pages_per_huge_page)
4327 {
4328         int i;
4329         struct page *dst_base = dst;
4330         struct page *src_base = src;
4331
4332         for (i = 0; i < pages_per_huge_page; ) {
4333                 cond_resched();
4334                 copy_user_highpage(dst, src, addr + i*PAGE_SIZE, vma);
4335
4336                 i++;
4337                 dst = mem_map_next(dst, dst_base, i);
4338                 src = mem_map_next(src, src_base, i);
4339         }
4340 }
4341
4342 void copy_user_huge_page(struct page *dst, struct page *src,
4343                          unsigned long addr, struct vm_area_struct *vma,
4344                          unsigned int pages_per_huge_page)
4345 {
4346         int i;
4347
4348         if (unlikely(pages_per_huge_page > MAX_ORDER_NR_PAGES)) {
4349                 copy_user_gigantic_page(dst, src, addr, vma,
4350                                         pages_per_huge_page);
4351                 return;
4352         }
4353
4354         might_sleep();
4355         for (i = 0; i < pages_per_huge_page; i++) {
4356                 cond_resched();
4357                 copy_user_highpage(dst + i, src + i, addr + i*PAGE_SIZE, vma);
4358         }
4359 }
4360 #endif /* CONFIG_TRANSPARENT_HUGEPAGE || CONFIG_HUGETLBFS */